Full download ebook at ebookgate.

com

Nonlinear Computational Geometry 2010th

Edition Ioannis Z Emiris

https://ebookgate.com/product/nonlinear-
computational-geometry-2010th-edition-ioannis-z-
emiris/

Download more ebook from https://ebookgate.com

More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Computational Intelligence in Games 1st Edition Z. Chen

https://ebookgate.com/product/computational-intelligence-in-
games-1st-edition-z-chen/

Computational Algebraic and Analytic Geometry AMS

Special Sessions on Computational Algebrac and Analytic
Geometry for Low Dimensional Varieties Mika Seppala

https://ebookgate.com/product/computational-algebraic-and-
analytic-geometry-ams-special-sessions-on-computational-algebrac-
and-analytic-geometry-for-low-dimensional-varieties-mika-seppala/

Analysis Geometry Nonlinear Optimization and

Applications 1st Edition Panos M. Pardalos

https://ebookgate.com/product/analysis-geometry-nonlinear-
optimization-and-applications-1st-edition-panos-m-pardalos/

Handbook of Discrete and Computational Geometry 2nd

Edition Rajeev Bansal

https://ebookgate.com/product/handbook-of-discrete-and-
computational-geometry-2nd-edition-rajeev-bansal/
31st International Symposium on Computational Geometry
SoCG 2015 1. Aufl Edition Strupat

https://ebookgate.com/product/31st-international-symposium-on-
computational-geometry-socg-2015-1-aufl-edition-strupat/

Computational Methods for Modelling of Nonlinear

Systems 1st Edition A. Torokhti And P. Howlett (Eds.)

https://ebookgate.com/product/computational-methods-for-
modelling-of-nonlinear-systems-1st-edition-a-torokhti-and-p-
howlett-eds/

Elasticity and Geometry From hair curls to the

nonlinear response of shells 1st Edition Basile Audoly

https://ebookgate.com/product/elasticity-and-geometry-from-hair-
curls-to-the-nonlinear-response-of-shells-1st-edition-basile-
audoly/

Nonlinear Dynamic Modeling of Physiological Systems

IEEE Press Series on Biomedical Engineering 1st Edition
Professor Vasilis Z. Marmarelis

https://ebookgate.com/product/nonlinear-dynamic-modeling-of-
physiological-systems-ieee-press-series-on-biomedical-
engineering-1st-edition-professor-vasilis-z-marmarelis/

Improper Riemann Integrals 2nd Edition Ioannis M.

Roussos

https://ebookgate.com/product/improper-riemann-integrals-2nd-
edition-ioannis-m-roussos/
 The IMA Volumes
in Mathematics
and its Applications

Volume 151

Series Editors
Fadil Santosa Markus Keel

For other titles published in this series, go to

www.springer.com/series/811
 Ioannis Z. Emiris Frank Sottile
Thorsten Theobald
Editors

Nonlinear Computational
Geometry
Editors
Ioannis Z. Emiris Frank Sottile
Lab of Geometric & Algebraic Algorithms Department of Mathematics
Department of Informatics Texas A&M University
& Telecommunications College Station, TX 77843
National and Kapodistrian University USA
of Athens http://www.math.tamu.edu/~sottile
Panepistimiopolis, 15784
Greece
http://cgi.di.uoa.gr/~emiris/index-eng.html

Thorsten Theobald
FB 12 - Institut für Mathematik
Johann Wolfgang Goethe-Universität
Robert-Mayer-Str. 10
D-60325 Frankfurt am Main
Germany
http://www.math.uni-frankfurt.de/~theobald/

Series Editors
Fadil Santosa
Markus Keel
Institute for Mathematics
& its Applications
University of Minnesota
Minneapolis, MN 55455
USA

ISSN 0940-6573
ISBN 978-1-4419-0998-5 e-ISBN 978-1-4419-0999-2
DOI 10.1007/978-1-4419-0999-2
Springer New York Dordrecht Heidelberg London

Library of Congress Control Number: 200931559

Mathematics Subject Classification (2000): 14Q, 68U05, 68W30, 65D, 52A35

© Springer Science + Business Media, LLC 2010

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science + Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they
are not identified as such, is not to be taken as an expression of opinion as to whether or not they are
subject to proprietary rights.

Camera-ready copy provided by the IMA.

Springer is part of Springer Science+Business Media (www.springer.com)

 FOREWORD

This IMA Volume in Mathematics and its Applications

NONLINEAR COMPUTATIONAL GEOMETRY

contains papers presented at a highly successful one-week workshop held

on May 29–June 2, 2007 on the same title. The event was an integral
part of the 2006–2007 IMA Thematic Year on “Applications of Algebraic
Geometry.” We are grateful to all the participants for making this workshop
a very productive and stimulating event.
We owe special thanks to Ioannis Z. Emiris (Department of Infor-
matics and Telecommunications, National and Kapodistrian University of
Athens), Frank Sottile (Department of Mathematics, Texas A&M Univer-
sity), and Thorsten Theobald (Institut für Mathematik, Johann Wolfgang
Goethe-Universität) for their superb role as workshop organizers and ed-
itors of these proceedings. We also thank Ron Goldman (Department of
Computer Science, Rice University) for his valuable contribution in orga-
nizing the workshop.
We take this opportunity to thank the National Science Foundation
for its support of the IMA.

Series Editors
Fadil Santosa, Director of the IMA
Markus Keel, Deputy Director of the IMA

v
 PREFACE

An original goal of algebraic geometry was to understand curves and

surfaces in three dimensions. From these roots, algebraic geometry has
grown into a theoretically deep and technically sophisticated field. Re-
cently, questions from robotics, computer vision, computer-aided geometric
design and molecular biology, together with the development of computa-
tional methods, have brought these original questions back to the forefront
of research. The implicitization of parametric surfaces, the geometry of
molecules, mechanical design and computer vision all lead to problems that
are challenging from the perspective of computational algebraic geometry.
In recent decades, computational geometry has developed as a disci-
pline at the intersection of mathematics and computer science that pro-
vides effective and algorithmic methods for treating geometric problems.
For natural reasons, the primary focus in computational geometry has been
on polyhedral (piecewise-linear) problems.
The challenge has arisen to combine the applicable methods of alge-
braic geometry with the proven techniques of piecewise-linear computa-
tional geometry (such as Voronoi diagrams and hyperplane arrangements)
to generalize the latter to curved objects. These research efforts may be
summarized under the term nonlinear computational geometry. In this
area, the development of reliable and practical algorithms is often based
on interrelated techniques that incorporate both symbolic and numerical
elements.
Within the thematic year Applications of Algebraic Geometry 2006/
2007 at the Institute for Mathematics and its Applications in Minneapolis,
a week-long workshop was devoted to this topic of nonlinear computational
geometry. This workshop took place from May 29 to June 2, 2007, and was
organized by I.Z. Emiris, R. Goldman, F. Sottile, and T. Theobald. Around
100 experts in this emerging field attended.
The present volume is comprised of nine contributions covering the
spectrum of topics from the workshop. Its purpose is to establish a col-
lection of research and expository articles describing the state-of-the-art in
nonlinear computational geometry and the challenges it poses for compu-
tational geometry, algebraic geometry, and geometric modeling.
In the first chapter, Spectral techniques to explore point clouds in
Euclidean space with applications in structural biology, Frédéric Cazals,
Frédéric Chazal and Joachim Giesen survey recent progress in spectral
techniques for machine learning, such as the principal component analysis
(PCA) and multi-dimensional scaling (MDS), in order to identify structure
in point clouds. The authors then offer an overview of such methods ap-
plied to understanding the geometry of large molecules and, in particular,
the important open problem of protein folding in bioinformatics.
vii
viii PREFACE

Algebraic hypersurfaces in space may have dual descriptions in terms

of an implicit equation or via a parametric representation. Transform-
ing one of these representations into the other and understanding their
connection is an ubiquitous task in nonlinear geometric computations. In
the chapter Rational parameterizations, intersection theory and Newton
polytopes, Carlos D’Andrea and Martı́n Sombra exploit recent advances
in algebraic geometry and combinatorial geometry to describe the New-
ton polytope and the support of the implicit equation of a given rational
parametric hypersurface, by means of invariants of its parameterization.
When transferring classical problems from the space Rn to the space of
lines, i.e., to the Grassmannian manifold, problems and structures obtain
an additional nonlinear component. In the chapter Some discrete properties
of the space of line transversals to disjoint balls, Xavier Goaoc describes the
state-of-the-art in generalizing classical statements from convex geometry
(in particular Helly’s Theorem and geometric transversals) to the space of
lines, and shows how these can lead to algorithmic applications.
Many problems in kinematics have natural formulations in terms of
polynomial equations, and already more than a century ago these formu-
lations were studied and applied to kinematics. In Algebraic geometry and
kinematics, Manfred Husty and Peter Schröcker explain how the classical
Study mapping can be used to transform kinematic problems into algebraic-
geometric problems and discuss the analysis of mechanisms from a modern
point of view.
In geometric modeling of real-world phenomena, offsets—surfaces at a
constant normal distance to a given surface—frequently arise. Even when
the original surface is rational, its offsets often are not, and it is a challenge
to understand those surfaces with rational offsets. Rimvydas Krasauskas
and Martin Peternell survey this in Rational offset surfaces and their mod-
eling applications. A particular focus is given on Pythagorean normal sur-
faces as well as on the viewpoint afforded by Laguerre geometry.
In his contribution A list of challenges for real algebraic plane curve vi-
sualization software, Oliver Labs discusses the issue of correctly visualizing
a real plane algebraic curve. The occurrence of singularities can make this
task quite challenging. Particularly difficult are those curves which contain
“complicated” singularities such as high tangencies or many halfbranches.
The exposition describes several classes of curves to serve as benchmarks
for future visualization software.
In the chapter Subdivision method for arrangement computation of
semi-algebraic curves, Bernard Mourrain and Julien Wintz present a syn-
thesis of existing approaches, enhanced with new tools, to compute ar-
rangements of semi-algebraic sets in a certified and efficient manner. Their
approach is to apply a subdivision technique and to analyze the topological
structure through this process. These concepts and methods are illustrated
by an implementation in the geometric modeler Axel.
Solving problems in nonlinear computational geometry exactly often
 PREFACE ix

leads to geometric predicates of high algebraic degree. In the chapter

Invariant-based characterization of the relative position of two projective
conics, Sylvain Petitjean studies these predicates for the fundamental prob-
lem of characterizing the relative position of two given conics. For his
bounds, he applies methods of classical invariant theory.
In the architectural design of glass/steel panel structures, free-form
surfaces may be approximated by polyhedral surfaces with hexagonal facets.
The final chapter, A note on planar hexagonal meshes, Wenping Wang and
Yang Liu first describes the applications and the theory of such surfaces.
It then discusses different algorithms for generating hexagonal meshes, in-
cluding a new one proposed by the authors.
We hope that the reader will enjoy the articles in the volume and that
the articles offer the reader an overview of the current developments in
nonlinear computational geometry.

Acknowledgments: The editors wish to thank the Institute for Mathe-

matics and its Applications for providing an inspiring and fruitful workshop
atmosphere as well as for their help in editing the volume.

Ioannis Z. Emiris
Department of Informatics & Telecommunications
National and Kapodistrian University of Athens
http://cgi.di.uoa.gr/∼ emiris/index-eng.html

Frank Sottile
Department of Mathematics
Texas A&M University
http://www.math.tamu.edu/∼ sottile

Thorsten Theobald
Institut für Mathematik
Johann Wolfgang Goethe-Universität
http://www.math.uni-frankfurt.de/∼ theobald/
 CONTENTS

Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Spectral techniques to explore point clouds

in Euclidean space, with applications to
collective coordinates in structural biology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Frédéric Cazals, Frédéric Chazal,
and Joachim Giesen

Rational parametrizations, intersection theory,

and Newton polytopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Carlos D’Andrea and Martı́n Sombra

Some discrete properties of the space of line

transversals to disjoint balls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Xavier Goaoc

Algebraic geometry and kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

Manfred L. Husty and Hans-Peter Schröcker

Rational offset surfaces and their modeling

applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Rimvydas Krasauskas and Martin Peternell

A list of challenges for real algebraic plane

curve visualization software. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Oliver Labs

A subdivision method for arrangement

computation of semi-algebraic curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Bernard Mourrain and Julien Wintz

Invariant-based characterization of the

relative position of two projective conics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Sylvain Petitjean

A note on planar hexagonal meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

Wenping Wang and Yang Liu

List of workshop participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

xi
SPECTRAL TECHNIQUES TO EXPLORE POINT CLOUDS
IN EUCLIDEAN SPACE, WITH APPLICATIONS TO
COLLECTIVE COORDINATES IN STRUCTURAL BIOLOGY
FRÉDÉRIC CAZALS∗ , FRÉDÉRIC CHAZAL† , AND JOACHIM GIESEN‡

Abstract. Life sciences, engineering, or telecommunications provide numerous sys-

tems whose description requires a large number of variables. Developing insights into
such systems, forecasting their evolution, or monitoring them is often based on the in-
ference of correlations between these variables. Given a collection of points describing
states of the system, questions such as inferring the effective number of independent
parameters of the system (its intrinsic dimensionality) and the way these are coupled
are paramount to develop models. In this context, this paper makes two contributions.
First, we review recent work on spectral techniques to organize point clouds in
Euclidean space, with emphasis on the main difficulties faced. Second, after a care-
ful presentation of the bio-physical context, we present applications of dimensionality
reduction techniques to a core problem in structural biology, namely protein folding.
Both from the computer science and the structural biology perspective, we expect
this survey to shed new light on the importance of non linear computational geometry
in geometric data analysis in general, and for protein folding in particular.

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1 Geometric data analysis and spectral point cloud processing . . . . . . . . . . . . . . 2
1.2 Spectral methods and alternatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 An application in structural biology: Protein folding . . . . . . . . . . . . . . . . . . 5
1.4 Notations and paper overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 PCA and MDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 MDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1 Neighborhood criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Dimension detection using PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4 Turning non-linear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.1 Maximum variance unfolding (MVU) . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.2 Locally linear embedding (LLE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.3 ISOMAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.4 Laplacian eigenmaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.5 Hessian eigenmaps (HLLE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.6 Diffusion maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
5 Applications in structural biology: the folding problem . . . . . . . . . . . . . . . . . . . . . 15
5.1 Folding: from experiments to modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 16
5.2 Energy landscapes and dimensionality reduction . . . . . . . . . . . . . . . . . . . . . 16
5.2.1 Potential and free energy landscape . . . . . . . . . . . . . . . . . . . . . . 16
5.2.2 Enthalpy-entropy compensation, energy funnel, ruggedness and frustration . . 16
5.2.3 Cooperativity and correlated motions . . . . . . . . . . . . . . . . . . . . . 18
5.3 Bio-physics: Pre-requisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.3.1 Molecular dynamics simulations . . . . . . . . . . . . . . . . . . . . . . . . 19
5.3.2 Models, potential energy landscapes and their ruggedness . . . . . . . . . . . 19
5.3.3 Morse theory and singularity theory . . . . . . . . . . . . . . . . . . . . . . 20
5.3.4 Free energy landscapes and reaction coordinates . . . . . . . . . . . . . . . . 20
5.3.5 Folding probability pf old . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.4 Inferring reaction coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.4.1 Reaction coordinates? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.4.2 Contacts based analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.4.3 Dimension reduction based analysis . . . . . . . . . . . . . . . . . . . . . . 26
5.4.4 Morse theory related analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 27
6 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

∗ INRIA Sophia-Antipolis, 2004 route des Lucioles, BP 93, 06902 Sophia Antipolis,

France ([email protected]).
† INRIA Saclay, Parc Orsay Université, 4 rue Jacques Monod, 91893 Orsay Cedex,

France ([email protected]).
‡ Institut fuer Informatik, Ernst-Abbe-Platz 2, D-07743 Jena, Germany
([email protected]).

I.Z. Emiris et al. (eds.), Nonlinear Computational Geometry, The IMA Volumes in 1
Mathematics and its Applications 151, DOI 10.1007/978-1-4419-0999-2_1,
© Springer Science + Business Media, LLC 2010
2 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

1. Introduction.

1.1. Geometric data analysis and spectral point cloud pro-

cessing. Modeling the climate, understanding the interplay between pro-
teins, metabolites and nucleic acids making up a regulation network within
a cell, or unraveling the connexions between spiking neurons are example
problems where a large number of variables interplay in a complex non lin-
ear way. Developing insights into such systems, forecasting their evolution,
or monitoring them is often based on the inference of correlations between
these variables. More precisely, learning such correlations from experiments
is paramount to model development, as theory and experimental inference
are tightly coupled.
Consider a complex system, and assume we are given a number of
observations describing different states of the system. In such a setting,
we are interested in the question of inferring the effective number of in-
dependent parameters of the system (its intrinsic dimensionality) and the
way these are coupled. To meet these challenges, a set of new geometric
methods, known as manifold learning, have been developed in the machine
learning community mainly over the past decade. These methods are based
upon the assumption that the observed data –a point cloud in some n di-
mensional space, lie on or are close to a submanifold M in Rd .
Naturally, given the variety of situations, one cannot expect a single
method to meet all needs. Nevertheless, many of the most popular ap-
proaches boil down to spectral methods. Note that the term spectral method
is ambiguous and used differently within different communities, e.g., in nu-
merical methods for partial differential equations it often involves the use
of the fast Fourier transform. Here we want to use the term in the sense
of data analysis similar as van der Maaten et al. did [51]. That is, for us
in a spectral method, a symmetric matrix is derived from the point cloud
data and the solution to a given optimization problem can be obtained
from the eigenvectors of this matrix. We should mention that the term
spectral method is also used in mesh processing in the geometric modeling
community where the symmetric matrix is obtained from the connectivity
of the mesh, see [58] for an overview. The geometric optimization problems
that lead to a spectral technique are mostly of a least squares nature and
include the following classical (and archetypical) problems:
(1) Find the k-dimensional subspace that approximates the point cloud
best in a least squares sense.
(2) Find the embedding of the point cloud in k-dimensional space that
preserves the distances between the points best possible in a least
squares sense.
The first problem is called principal component analysis (PCA) as it
asks for the principal directions (components) of the data. It essentially
is a data quantization technique: every data point gets replaces by its
projection onto the best approximating k-dimensional subspace. The loss
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 3

incurred by the quantization is the variance of the data in the directions

orthogonal the best approximating k-dimensional subspace. As long as
this variance is small PCA can also be seen as denoising the original data.
Many machine learning techniques including clustering, classification and
semi-supervised learning [32], but also near neighbor indexing and search
can benefit from such a denoising.
The second problem is called multi-dimensional scaling (MDS). An
important application of MDS is visualization of the point cloud data: the
data points get embedded into two- or three-dimensional space, where they
can be directly visualized. The main purpose of visualization is to use the
human visual system to get insights into the structure of the point cloud
data, e.g., the existence of clusters or—for data points labeled with discrete
attributes—relations between this attributes. MDS visualization remains
to be a popular tool for point cloud data analysis, but of course a lot of
information will get lost (and in general cannot be restored by the human
visual system) if the intrinsic dimension of the data points is larger than
three.
Recently the focus in point cloud data analysis shifted: more emphasis
is put on detecting non-linear features in the data, although processing
the data for visual inspection still is important. What drives this shift
in focus is the insight that most features are based on local correlations
of the data points, but PCA and MDS both have only a global view on
the point cloud data. The shift towards local correlations was pioneered
by two techniques called ISOMAP [48, 21] and Locally Linear Embedding
(LLE) [45, 46]. It is important to note that focusing on local correlations
does not mean that one loses the global picture: for example the global
intrinsic dimension of the data can be estimated from local information,
whereas it is often (when the data are embedded non-linearly) not possible
to derive this information from a purely global analysis. ISOMAP and LLE
and their successors (some of which we will also discuss here) can be used
both for the traditional purposes data quantization and data visualization.
In general they preserve more information of the data (than PCA and
MDS) while achieving a similar quantization error or targeting the same
embedding dimension for data visualization, respectively.

1.2. Spectral methods and alternatives.

Advantages of spectral methods. Consider a point cloud P sampled
from a manifold M embedded in Rd . In this survey, we focus on a set of
quite famous methods following a common thread, as they ultimately re-
sort to spectral analysis. They all intend to find the best embedding of the
dataset P into an Euclidean space Rk with respect to some quadratic con-
straint reflecting different geometric properties of the underlying manifold
M . The embedding of the data that minimizes the quadratic constraint can
then be interpreted as the best k-dimensional embedding of the data with
respect to the geometric property we aim to preserve. In most cases, the
4 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

quadratic minimization problem boils down to a general eigenvalue prob-

lem ensuring to find a global minimum. Moreover, the embedding can be
found by easy-to-implement polynomial time algorithms.
This provides a substantial advantage over iterative or greedy methods
based upon Expectation-Maximization like algorithms that do not provide
guarantees of global optimality. In particular, for quite large data sets,
the methods we consider still provide results when iterative and greedy
methods fail due to complexity issues. Another advantage of “spectral
methods” is that the quadratic constraint leads to a measurement of the
quality of the embedding1 . At last, “spectral methods” have been widely
used and studied in many applications areas (graph theory, mesh processing
[58],...) giving rise to a large amount of efficient theoretical and algorithmic
tools that can be used for dimensionality reduction.
Approaches not covered. As our focus is on spectral techniques, a num-
ber of dimensionality reduction techniques are not covered in this paper.
While the reader might consult [51] for a rather exhaustive catalog, the
following comments are in order about the missing classes:
• EM-based methods: a large set of manifold learning algorithms
developed in the machine learning community adopt a probabilistic
point of view, so as to maximize a likelihood (Self Organizing Maps,
Generative Topographic Mapping, Principal curves, etc. See [8] for
example.). Some of them, like principal curves [31] or generative
topographic mapping [9] for example, aim to fit the data set by a
parameterized low dimensional (in general 1 or 2) manifold. They
usually assume that the topology of the manifold is known and
simple (simple curves, planes, discs) and do not allow to deal with
data sampled from more complicated shapes.
• Methods related to the Johnson-Lindenstrauss lemma: the
Johnson-Lindenstrauss lemma addresses the dimensionality reduc-
tion problem of a general point cloud (not necessarily sampled
around a low dimensional manifold) in the perspective of preserv-
ing the pairwise distances between the points. An extension to
points and flats and algebraic surfaces has been proposed in [1].
• Kernel methods: a number of methods, including some of the
methods we shall discuss, can be interpreted in the framework of
kernel methods. See [29, 53] for example.
• Methods targeting non manifold shapes: more recently, some ge-
ometric inference methods have been developed in the case where
the shape underlying the data is not assumed to be a smooth man-
1 For example, in [48], the quality of the embedding obtained is assessed resorting to

the residual variance σk (k, d) defined by:

σk (k, d) = 1 − R2 (D̂k , Dd ) (1)

with R(D̂k , Dd ) the correlation coefficient taken over all entries of matrices D̂k and Dd .
The closer to zero this variance, the better the approximation.
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 5

ifold. They lead to promising but preliminary results for dimen-

sionality reduction of general shapes [11, 12].
1.3. An application in structural biology: Protein folding. As
an application of dimensionality reduction techniques in general, and of
spectral methods in particular, we shall give a detailed account of one of
the core open problems in structural biology, namely protein folding: how
does a protein reach its folded-, i.e., its biologically active state, from the
unfolded one? As of October 2007, about 1,000 genomes have been fully
sequenced or are about to be so, while the Protein Data Bank contains (a
mere) 40,000 structures. The question of understanding folding so as to
predict the structure of a protein from its sequence is therefore central2 ,
to foster the understanding of central mechanisms in the cell, but also to
perform protein engineering with applications ranging from drug design to
bio-technologies.
Aside these general incentives, a number of technical ones advocate
this particular problem.
First, the question of folding is closely related to a specific d-dimensional
manifold which associates an energy to a conformation (the energy land-
scape), on which point clouds are sampled thanks to simulations techniques
like the prototypical molecular dynamics method. Thus, the underlying
mathematical structure is a manifold and not a (stratified) complex of ar-
bitrary topology.
Second, assuming the folded and unfolded conformations correspond to
(significant) local minima of the energy landscape, the problem is tanta-
mount to understand transitions on this landscape, i.e. paths joining these
minima. The difficulty of the problem is rooted in two facts: the high-
dimensionality of the landscape (d = 3n or d = 6n as argued below, with n
the number of atoms), and its complex topography which reflects the com-
plex interactions (forces) between atoms. These intrinsic difficulties call
for dimensionality reduction techniques, so as to exhibit a small number
of new variables (typically one or two), called the reaction coordinates, ac-
counting for the transition. These coordinates should match the effective
large amplitude - slow frequency degrees of freedom of the system, thus
providing a simplified view of the process, and easing its interpretation.
Thus in essence, one wishes to quantize information located on a non lin-
ear manifold, while retaining the essential features.
Third, as opposed to a large number of multi-dimensional data sets, folding
features a stimulating interplay between modeling and experiments. The
point clouds studied in folding are indeed closely related to a number of
experiments in bio-physics, so that one can precisely assess the quality and
the interest of dimensionality reduction procedures. Example such experi-
mental methods are dynamic NMR, protein engineering (φ-value analysis),
2 At least for proteins consisting of a single polypeptidic chain, as the formation of

multimers also poses docking questions.

6 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

laser initiated folding, etc. Describing these procedures is clearly beyond

the scope of this survey, and the reader is referred to [84, 75] for starting
pointers.
1.4. Notations and paper overview. Throughout this paper we
will be using the following notations:
P point cloud
n number of point is P
d dimension of the Euclidean space form which the points
in P are drawn
k target dimension.
The paper is organized as follows. Section 2 presents the two archetyp-
ical spectral methods used to explore point clouds, namely PCA and MDS.
The question of localizing neighborhoods is discussed in Section 3, while
methods meant to accommodate non linear geometries are discussed in Sec-
tion 4. The application of dimensionality reduction techniques to protein
folding is discussed in Section 5. To conclude, Section 6 discusses a number
of research challenges.
2. PCA and MDS. In the P following we assume that the points in
n
P are centered at the origin, i.e., i=1 pi =P 0. Note that this can always
1 n 0
achieved
Pn by a simple translation: let p̄ = n i=1 pi and pi = pi − p̄, then
0
i=1 pi = 0.
Principal component analysis (PCA) asks for the k-dimensional sub-
space of Rd that approximates the point set P best possible in a
least squares sense and projects P onto that subspace, whereas multi-
dimensional scaling (MDS) in its basic form aims for the k-dimensional
embedding of P that preserves the pairwise inner products of the points in
P best possible in a least squares sense. In both cases k can range from 1
to d − 1.
Though different in their motivation and objective, PCA and MDS are
almost identical in a technical sense: both can be formulated in terms of
eigenvectors of some positive semi-definite matrix derived from the point
set P , which itself can be written as a (d × n)-matrix as follows:
 
p11 ... pn1
 .. ..  ,
P = . . 
p1d ... pnd

where pij is the j’th component of the point pi ∈ P . From the matrix P
one canonically derives two positive semi-definite matrices,
(1) the covariance matrix C = P P T , and
(2) the Gram matrix G = P T P .
The
Pn covariance matrix is a (d × d)-matrix and can also be written as C =
T
i=1 pi pi , whereas the Gram matrix has dimension n × n and can also be
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 7


written as G = pTi pj . Both matrices are intimately linked also via their
eigenvectors and eigenvalues. We have the following observation.
Observation 1. The matrices C and G have the same non-zero
(positive) eigenvalues (and thus the same rank).
Proof. Let v ∈ Rd be an eigenvector of C with eigenvalue λ > 0, then
T
P v is an eigenvector of G also with eigenvalue λ as can be seen from the
following simple calculation:
GP T v = P T P P T v = P T Cv = λP T v.
Similarly, if u ∈ Rn is an eigenvector of G with eigenvalue µ > 0, then P u
is an eigenvector of C with eigenvalue µ.

One important issue with both PCA and MDS is how to

choose/determine k (the intrinsic dimensionality of the point cloud data).
Sometimes there is a “large” gap in the eigenvalue spectrum of C or G,
respectively, and k is then often chosen as the number of eigenvalues above
this gap.
2.1. PCA. As mentioned earlier PCA asks for the k-dimensional sub-
space of Rd that approximates the point set P best possible in a least
squares sense. Let us discuss this for the case k = d − 1 first. In this
case we are looking for a unit vector v ∈ Rd such that the sum of the
squared lengths of the projections (v T pi )v is minimized. Formally this can
be written as
min v T P P T v
s.t. kvk2 = 1.
From the Lagrange multiplier theorem one derives the following condition
for an optimal solution to this optimization problem: λv = P P T v = Cv.
That is, an optimal solution is the subspace orthogonal to an eigenvector
of the covariance matrix C and the value of the optimization problem at
an optimal solution is v T P P T v = v T Cv = λkvk2 = λ. Hence we are
looking for an eigenvector associated to the smallest eigenvalue of C and
the optimal solution is spanned by all eigenvectors of the covariance matrix
C except the one corresponding to the smallest eigenvalue.
Let λ1 ≥ . . . ≥ λd ≥ 0 be the eigenvalues P of C, v1 , . . . , vd ∈ Rd a cor-
k
responding orthonormal eigenbasis and Pk = i=1 vi viT , k = 1, . . . , d, the
projector on the k’th invariant eigenspace, i.e., the eigenspace spanned by
the first k eigenvectors. Iteratively it follows that the best approximating k-
dimensional subspace of Rd in a least square sense is spanned by v1 , . . . , vk .
The k’th order PCA is then given as the following transformation:
pi 7→ Pk pi = pi − (I − Pk )pi .
In a way Pk pi is seen as the signal conveyed with the point pi and (I−Pk )pi
is seen as noise.
8 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

2.2. MDS. Multi-dimensional scaling is aiming for a k-dimensional

embedding of P that preserves the pairwise inner products pTi pj as well as
possible in a least squares sense3 . Note that all inner products are stored as
entries in the Gram matrix G. Let µ1 ≥ . . . ≥ µn ≥ 0 be the eigenvalues of
u1 , . . . , un ∈ Rn be a corresponding orthonormal eigenbasis and let
G, let P
k
Qk = i=1 ui uTi , for k = 1, . . . , n, be the projector on the k’th invariant
eigenspace. We have the following observation:
Observation 2. The matrix Qk G is the best rank k approximation
of the Gram matrix G in the sense that

kQk G − Gk2 = argmin Q: (n×n)-matrix of rank k kQG − Gk2 .

The matrix Qk G can also be interpreted as a matrix of inner products.
To see this we use (a) the projector property Q2k = Qk , (b) symmetry
QTk = Qk , and (c) the commutator property Qk G = GQk , and get

Qk G = Q2k G = Qk GQk = Qk P T P Qk = QTk P T P Qk = (P Qk )T P Qk ,

which shows that Qk G is the matrix of inner products of the columns

of P Qk = (Qk P T )T . Here the (n × d)-matrix Qk P T is the projection
of the rows of P onto the space spanned by u1 , . . . , un . The k’th order
MDS maps the point pi to the i’th column Qk P T , i.e., the i’th column of
P Qk . This column is uniquely specified by its coefficients α1i , . . . , αki in the
orthonormal basis u1 , . . . , uk . Representing the points pi by (α1i , . . . , αki )
gives the thought for least squares optimal k-dimensional embedding of the
point set P .
3. Localization.
3.1. Neighborhood criteria. In using PCA and MDS, feature pre-
serving data quantization and visualization can be enhanced by taking only
local relations among all the data points into account. Localization the rela-
tions means choosing neighborhoods for each data point, i.e., building a (in
general directed) neighborhood graph on the data points. The right choice
of neighborhood is crucial for the localized version of PCA and MDS to
work properly. Commonly used methods to define the neighborhoods are:
(1) κ nearest neighbors: connect every pi to its κ nearest neighbors
(in terms of Euclidean distance) in P .
(2) symmetric κ nearest neighbors: connect pi to its κ nearest neigh-
bors and all points in this neighborhood to each other.
(3) fixed neighborhood: given ε > 0, connect every pi to all points in
P that have distance less than ε to pi .
(4) relative neighborhood: given ρ > 1, connect every pi to all neigh-
bors at distance ρ times the distance of pi to its nearest neighbor.
3 Observe that completely preserving the inner products allows us to recover P up

to a rotation, i.e., completely preserving the pairwise inner products also preserves the
pairwise distances kpi − pj k.
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 9

An important observation is that (1) and (2), i.e., κ nearest neigh-

bors and symmetric κ nearest neighbors, respectively, do not automatically
adapt to the intrinsic dimension of the point cloud data. Intuitively, if the
intrinsic dimension is large also κ needs to be large in order to cover a mean-
ingful neighborhood for a data point (we expect this neighborhood to grow
exponentially in the intrinsic dimension), whereas if the intrinsic dimension
is small, for the same value of κ one already covers data points quite far
away. Methods (3) and (4), fixed- and relative neighborhood, both auto-
matically adapt to the intrinsic dimension, but cannot—in contrast to the
κ nearest neighbor methods—adapt to non-uniform or anisotropic spacing
of the data points. In practice a good choice for the value of the parameter
ρ of (4) may be easier to find than for the value of ε in (3).
More neighborhood graphs are discussed by Yang [56] who also pro-
vides experimental results.
3.2. Dimension detection using PCA. We have seen above that
knowing the local dimension at a data point can guide the right choice of
parameter κ when computing the κ nearest neighbors neighborhood. On
the other hand, using that given p ∈ M , there exists a small neighborhood
of p in which M is close to its tangent space at p, it is appealing to use
localized versions of PCA to infer the local intrinsic dimension of M at p
from the point cloud data P . With a good neighborhood N (p) ⊂ P of
p ∈ P one can estimate the intrinsic dimension at p by a localized version
of PCA. The localized version uses the local covariance matrix Cp of the
points
1 X
p0i = (pi − p) − (pi − p) for pi ∈ N (p).
n
pi ∈N (p)

Intuitively, if the local dimension at p is k, then we expect a gap in the

eigenvalue spectrum of Cp in the sense that k’th largest eigenvalue is much
larger than the (k+1)’st eigenvalue and the k largest eigenvalues are roughly
of the same magnitude. That is, we expect a threshold θ such that
λj λj
≥ θ for j ≤ k and ≤ θ for j > k.
λ1 λ1
Indeed, Cheng, Wang and Wu [13] were able to prove the existence of such a
threshold θ under the assumption that the data are sampled from a smooth
manifold and obey a sampling condition. The sampling condition rules out
locally non-uniform or anisotropic spacing of the sample points. Under
this assumption fixed- and relative neighborhoods should work. Cheng et.
al use the relative neighborhood for their proof. Though their threshold
parameter θ depends on parameters of the sampling condition they report
good results in practice using a threshold of θ = 1/4.
It is important to remark that when the sampling conditions are not
fullfilled or when the size of the neighborhoods are not well-choosen, the
10 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

previous method usually leads to unclear and confusing estimations. In

particular the dimension estimation may depend on a “scale” (in the pre-
vious case the size of the neighborhoods) at which the data is considered:
assume that P samples a planar spiral with gaussian noise in the normal
direction to the curve. At a “microscopic” scale, P just looks like a finite
set of points and its dimension is 0. At a scale of the size of the standard
deviation of the noise, P seems to locally sample the ambient space and the
localized PCA method will probably estimate M to be 2-dimensional. At a
higher, but not too big, scale the localized PCA will provide the right esti-
mation and at large scales, it will again provide a 2-dimensional estimation.
Various notions of dimension (q-dimension, capacity dimension, correlation
dimension, etc...) have been introduced to define the intrinsic dimension of
general shapes (including non smooth shapes and fractal sets). They give
rise to algorithmically simple methods that simultaneously provide dimen-
sion estimations at different explicit scales allowing the user to select the
one which is most relevant for his purpose. An introduction to this subject
may be found in [39].
4. Turning non-linear. The linear and global aspects of PCA and
MDS make them inefficient when the underlying manifold M is highly non
linear. Designing non-linear dimensionality reduction methods that lead
to good results for non linear smooth manifolds is an active research area
that gave rise to a big amount of literature during the last decade. In
this section, we quickly present a set of quite famous dimension reduction
methods that take advantage of the localization techniques presented in the
previous section and that have interesting geometric interpretations. They
also have the advantage of leading to easy to implement polynomial time
algorithms that prove more efficient with larger data sets than the ones
usually involved in iterative or greedy methods (like e.g. the ones involving
EM or EM-like algorithms). We also discuss the guarantees provided by
these methods.
Recall that in the following the considered data sets P ⊂ Rd are as-
sumed to be sampled on/around a possibly unknown smooth manifold M
of dimension k. The common thread of the few methods presented below
is that they all aim to find a projection P̂ ⊂ Rk of the data set minimizing
a quadratic functional φ(P̂ ) that intends to preserve (local) neighborhood
information between the sample points.
4.1. Maximum variance unfolding (MVU). PCA and MDS per-
form poorly when data points are not close to an affine subspace, i.e., they
are both based on an inherent linearity assumption. Especially, both meth-
ods fail when the data points are close to a “curled up” linear space—the
most famous example is the so called Swiss roll data set, points sampled
densely from a curled up planar rectangle in R3 . The idea behind maximum
variance unfolding (MVU), introduced by Weinberger and Saul [52, 55, 54],
is to unfold the data, i.e., to transform the data set to a locally isometric
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 11

data set, that is closer to an affine subspace. The unfolding aims at max-
imizing the distance between non-neighboring points (after some choice of
neighborhood) while preserving the distances between neighboring points.
Technically MVU proceeds as follows: let D = (dij = kp− pj k2 ) be the
symmetric (n × n)-matrix of pairwise distances. Choose a suited neigh-
borhood for each point in P (Weinberger and Saul choose the symmetric
κ-nearest neighbors) and let the indicator variable nij be 1 if either pi is in
the neighborhood of pj or pj is in the neighborhood of pi , and 0 otherwise.
From D an unfolding, a positive semi-definite (n × n)-matrix K = (kij )
(interpreted as the Gram matrix of the unfolded point set) is computed
through the following semi-definite program (SDP)
Maximize the trace of K subject to
(1) K is positive semi-definite
Xn
(2) kij = 0
i,j=1
(3) kii − 2kij + kjj = dij for all (i, j) with nij = 1.
From K a lower dimensional embedding can be computed as described for
MDS.
4.2. Locally linear embedding (LLE). LLE is a method intro-
duced in [45, 46] that intends to take into account the local linearity of the
underlying manifold M to perform the reduction of dimension. In a first
step, LLE discards pairwise distances between widely separated points by
building a neighborhood graph G (see Section 3). The goal of this first step
is to connect only close points of P so that the neighbors of each vertex
pi in G are contained in a small neighborhood of pi which is close to the
tangent space of the underlying manifold M at pi . To take this local lin-
earity into account, LLE computes for each vertex pi of the graph its best
approximation as a linear combination of its neighbors. More precisely,
one computes a sparse matrix of weights Wi,j that minimize the quadratic
error
n
X X
ε(W ) = kpi − Wi,j pj k2
i=1 j∈N (pi )

where N (pi ) is the set of the vertices that are connected to pi in G. This
is a simple least square problem. Solving it with the additional constraint
X
∀i, Wi,j = 1
j∈N gb(pi )

makes the weights invariant to rescaling, rotations and translations of the

data (the weights thus characterize intrinsic properties of the data). The
weights matrix is then used to perform the dimensionality reduction: given
k < d, the points pi are mapped to the points p̂i ∈ Rk that minimize the
quadratic function
12 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

X X
Φ(p̂i ) = kp̂i − Wi,j p̂j k2
i j

This quadratic minimization problem classically reduces to solving a sparse

n×n eigenvalue problem. As for MDS, the LLE algorithm projects the data
in a low dimensional space, no matter what the mapping is. To provide
satisfactory result, the data have to be sufficiently dense to insure that the
neighbors of a given point provide a good approximation of the tangent
space of M . Moreover, even if the data are dense enough, the choice of
the neighbors may also be awkward: choosing a too small or too large
neighborhood may lead to very bad estimates of the tangent space.
4.3. ISOMAP. ISOMAP is a version of MDS introduced in [48, 21],
where the matrix of inner products or Euclidean distances, respectively, is
replaced by the matrix of the geodesic distances between data points on
M . In a first step, ISOMAP builds a neighborhood graph such that the
distances between points of P in the graph are close to the geodesic dis-
tances on M . Once the geodesic distance matrix has been built, ISOMAP
proceeds like classical MDS to project P in Rk .
One of the advantage of ISOMAP is that it provides convergence guar-
antees. First, it can be proven that if the data are sufficiently densely
sampled on M , the distance on the neighbor graph is close to the one on
M [20, 44, 26]. Nevertheless, in practice robust estimation of geodesic dis-
tances on a manifold is an awkward problem that requires rather restrictive
assumptions on the sampling. Second, since the MDS step in the ISOMAP
algorithm intends to preserve the geodesic distances between points, it pro-
vides a correct embedding if M is isometric to a convex open set of Rk .
The convexity constraint comes from the following remark: if M is an open
subset of Rk which is not convex, then there exist a pair of points that can-
not be joined by a straight line contained in M . As a consequence, their
geodesic distance cannot be equal to the Euclidean distance. It appears
that ISOMAP is not well-suited to deal with data on manifolds M that do
not fulfill this hypothesis. Nevertheless some variants (conformal ISOMAP
[21]) have been proposed to overcome this issue. Note also that ISOMAP
is a non local method since all geodesic distances between pairs of points
are taken into account. As a consequence ISOMAP involves a non-sparse
eigenvalue problem which is a main drawback of this method. To partly
overcome this difficulty some variant of the algorithm using landmarks have
been proposed in [21].
4.4. Laplacian eigenmaps. This method introduced in [4, 3] follows
the following general scheme: first a weighted graph G with weights Wi,j is
built from the data. Here the weights measure closeness between the points:
intuitively the bigger Wi,j is, the closer pi and pj are. A classical choice for
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 13

kp −p k2
the weights is given by the Gaussian kernel Wi,j = exp(− i 4σ j ), where
σ is a user-defined parameter4 . Second the graph G is embedded into Rk in
such a way that the close connected points stay as close as possible. More
precisely the points pi are mapped to the points p̂i ∈ Rk that minimize
X
φ(P̂ ) = kp̂i − p̂j k2 Wi,j .
i,j

There is an interesting and fundamental analogy between this discrete min-

imization problem on the graph G and a continuous minimization problem
on M . Indeed, it can be seen that minimizing φ on the functions defined
on the
R vertices of G corresponds (in a discretized version) to minimiz-
2
ing M k∇f k on the space of functions f defined on M with L2 norm
2 2
R
R kL2 = M kf k = 1. From the Stokes formula, this integral is equal to
kf
M
L(f )f , where L is the Laplace-Beltrami operator on M and its minimum
is realized for eigenfunctions of L. Similarly the minimization problem on
G boils down to a general eigenvector problem involving the Laplacian of
the graph. Indeed the Laplace operator Pon G is the matrix L = D − W ,
where D is the diagonal matrix Di,i = j Wi,j . It can be seen as an oper-
ator acting on the functions f defined on the vertices of G by subtracting
from f (pi ) the weighted mean value of f on the neighbors of pi . By a
classical computation, one can see that φ(P̂ ) = tr(P̂ T LP̂ ), where P̂ is the
n × k matrix with i-th row given by the coordinates of p̂i . It follows that,
given k > 0, the minimum of φ is deduced from the computation of the
k + 1 smallest eigenvalues of the equation Ly = λDy (the smallest one cor-
responding to the eigenvalue 0 has to be removed). The analogy between
the discrete and continuous setting extends to the choice of the weights of
kp −p k2
G: choosing Wi,j = exp(− i 4σ j ), where σ is a user-defined parameter,
allows to interpret the weights as a discretization of the heat kernel on M
[4]. From the side of the guarantees, the Laplacian eigenmaps only involve
intrinsic properties of G so they are robust to isometric perturbations of
the data. Moreover, the relationship with the Laplacian operator on M
provides a framework leading to convergence results of L to the Laplace
operator on M [3].
4.5. Hessian eigenmaps (HLLE). ISOMAP provides guarantees
when the unknown manifold M is isometric to a convex open subset of Rk .
Although the hypothesis of being isometric to an open subset of Rk seems
to be rather reasonable in several practical applications, the convexity hy-
pothesis appears to be often too restrictive. HLLE is a method introduced
in [24] intending to overcome this convexity constraint. The motivation of
HLLE comes from a rather elementary result stating that if M is isomet-
ric to a connected open subset of Rk then the null-space of the operator
defined on the space of C 2 -functions on M by
4 To obtain a sparse matrix W the values of W
i,j that are smaller than some fixed
small threshold are usually set to 0.
14 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN
Z
H:f → kHessf (m)k2 dm
M

where Hessf is the Hessian of f , is a (k + 1)-dimensional space spanned

by the constant functions and the “isometric coordinates” of M . More
precisely, if there exists an open set U in Rk and an isometric embedding
φ : M → U then it can be proven that the constant functions and the func-
tions φ1 , · · · φk , where φi is the i-th coordinate of the map φ, are contained
in the null-space of H. Moreover, the constant functions span one dimen-
sion of this null-space and the k functions φi span the k other dimensions.
It is thus appealing to estimate this null space in order to recover these
isometric coordinates to map M isometrically on an open subset of Rk . To
do this the algorithm follows the same scheme as LLE and the estimation
of the null-space of H reduces to an eigenvalue computation of a sparse
n × n matrix. As a consequence HLLE allows to process dimensionality re-
duction for a larger class of manifolds M than ISOMAP. The quality of the
reduction is obviously closely related to the quality of the approximation
of the kernel of the operator H. Nevertheless, it is important to notice that
the algorithm involves the estimation of second order differential quantities
for the computation of the Hessian while LLE requires only first order ones
to approximate the tangent space of M . To be done efficiently this usually
needs a very dense sampling of M . At last, note that HLLE is the same
as Laplacian Eigenmaps where the Laplacian operator has been replaced
by H.
4.6. Diffusion maps. Diffusion maps [15] provide a method for di-
mensionality reduction based upon Markov random walks on a weighted
graph G reflecting the local geometry of P . The graph G is built in a
similar way as for Laplacian Eigenmaps: the larger is the weight of an
edge, the “closer” are its endpoints. In particular G can be built using the
discretization of the heat kernel on M (see Section 4.4). From the weights
matrix W one constructs a Markov transition matrix Π by normalizing the
rows of W
Wi,j X
Πi,j = where d(pi ) = Wi,k is the degree of the vertex pi
d(pi )
k

Πi,j can be interpreted as the probability of transition from pi to pj in one

time step. The term Πt (i, j) of the successive powers Πt of Π represent the
probability Πt (pi , pj ) of going from pi to pj in t steps. The matrix Π can
be seen as an operator acting on the probability distributions supported
on the vertices of G. It admits an invariant distribution φ0 defined by
φ0 (pi ) = Pd(pd(p
i)
j)
. The idea of diffusion maps is thus to define a metric
j
between the points of P which is such that at a given t > 0 two points
pi and pj are close if the conditional distributions of probability Πt (pi , .)
and Πt (pj , .) are close. The choice of a weighted L2 metric between the
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 15

Fig. 1. An example of a graph G (the weights are given by the heat kernel ap-
proximation, see text) with points that are close or far to each other with respect to the
diffusion metric: the points x and y are close to each other while the points x and z are
far away because G is “pinched” between the two parts containing x and z. So there
are few paths connecting x to z.

conditional distributions allows to define a diffusion metric between the

points of P

X (Πt (pi , pk ) − Πt (pj , pk ))2

Dt2 (pi , pj ) =
φ0 (pk )
k

which is closely related to the spectral properties of the random walk on

G given by Π. Intuitively, two points pi and pj are close if there are
many paths connecting them in G as illustrated on Fig. 1. Note that the
parameter t representing the “duration” of the diffusion process may be
interpreted as a scale parameter in the analysis. Given k and t > 0, the
diffusion map provides a parameterization and a projection of the data set
which performs a dimensionality reduction that minimizes the distortion
between the Euclidean distance in Rk and the diffusion distance Dt . The
diffusion map is obtained from the eigenvectors of the transition matrix Π
and the eigenvalues to the power t of the transition matrix. The diffusion
maps framework reveals deep connections with other areas (such as spectral
clustering, spectral analysis on manifolds,...) that open many questions
and make it an active research area. For a more detailed presentation
of diffusion maps and its further developments the reader is referred to
[15, 16, 37].

5. Applications in structural biology: the folding problem. In

this section, we first recall the intrinsic difficulty of folding proteins on a
computer –Section 5.1, and bridge the gap between folding and dimension-
ality reduction –Section 5.2. We then proceed with a detailed account of
the bio-physical context by discussing the question of cooperative motions
within a protein –Section 5.3, and make the connexion to Morse theory and
singularity theory along the way. Finally, we review techniques to derive
meaningful so-called reaction coordinates –Section 5.4.
16 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

5.1. Folding: from experiments to modeling. Anfinsen was

awarded the 1972 Nobel prize in chemistry for his work on ribonuclease,
especially concerning the connection between the amino acid sequence and
the biologically active conformation5 . Since then, Anfinsen’s dogma states
that for (small globular) proteins, the sequence of amino-acids contains
the information that allows the protein to fold i.e. to adopts its (essen-
tially unique) native structure, or phrased differently, the 3d structure that
accounts for its function. At room temperature, the folding of a protein
typically requires from millisecond to seconds, while the time-scale of the
finest (Newtonian) physical phenomena involved is the femtosecond.
When compared to femtoseconds, folding times are rather slow, which
points towards a process more complex than a mere descent towards a
minimum of energy. On the other hand, such folding times are definitely
too fast to be compatible with a uniform exploration of an exponential
number conformations6 . This observation is known as Levinthal’s paradox
[90], and scales the difficulty of folding from a computational perspective.
5.2. Energy landscapes and dimensionality reduction.
5.2.1. Potential and free energy landscape. Consider a system
consisting of a protein and the surrounding solvent, for a total of n atoms.
Each atom is described by 3 parameters for the position and 3 for its
velocity (momentum). In the following, depending on the context, we shall
be interested in a parameter space of dimension d = 3n (positions) or
d = 6n (positions+velocities), the latter being called the phase space. As
the system is invariant upon rigid motions, one could work with d − 6
degrees of freedom, but we skip this subtlety in the following. From this d-
dimensional parameter space, one defines the the energy landscape [99], i.e.
the d dimensional manifold obtained by associating to each conformation
of the system an energy (potential energy or free energy7 ). Since the water
molecules are critical to model appropriately the electrostatic interactions,
n typically lies in the range 104 to 105 for a system consisting of a protein
and its aqueous environment.
5.2.2. Enthalpy-entropy compensation, energy funnel,
ruggedness and frustration.
Energy funnels. Folding may be seen as the process driving a het-
erogeneous ensemble of conformations populating the unfolded state to a
homogeneous ensemble of conformations populating the folded or native
state. To intuitively capture one major subtlety of folding, it is instructive
to examine the variation of the enthalpy H and entropy S of the system
5 See http://nobelprize.org/nobel prizes/chemistry/laureates/1972/index.
html
6 Recall that the side-chains of the amino-acids take conformations within finite sets

–the so-called rotamers [86, 73], whence a priori an exponential number of conformations.
7 As will be shown with Eq. (2), a free energy landscape is obtained from the potential

energy landscape by projecting onto selected coordinates.

 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 17
Energy
Entropy

Unfolded
conformations
Similarity
to native state

Partially folded
conformations

Folded
ie native state

Coordinates
(cartesian, internal)

Fig. 2. Folding funnel.The variability of conformations encodes the entropy of the

system, while its energy level encodes the proximity to nativeness. Adapted from [67].

protein+solvent. While the protein folds, more native contacts between

atoms get formed, whence an enthalpy decrease. On the other hand, two
phenomena account for an entropic drop down: first, the conformational
variability of the protein decreases; second, the structure of the solvent
around the protein changes. This latter re-organization, known as the hy-
drophobic effect, corresponds to the fact that water molecules line-up along
the hydrophobic wall formed by the molecular surface of the protein. Over-
all, the variations of the enthalpy and entropy almost cancel out, resulting
in a small variation of the free energy G = H − T S of the system. This
phenomenon is known as the enthalpy-entropy compensation [75], and can
be illustrated using energy landscapes, as seen from Fig. 2. On this fig-
ure, the vertical axis features the free energy, and the horizontal one the
entropy: while the folding process progresses, the free energy (slightly) de-
creases and the landscape becomes narrower—the entropy decreases. Such
a landscape is generally called a folding funnel [67].
While the previous discussion provides a thermodynamic overview of
the folding process, Levinthal’s paradox deals with a kinetic problem—how
come the folding process is so fast? Travelling down the folding funnel8 pro-
vides an intuitive explanation: the protein is driven towards the minimum
of energy corresponding to the native state by a steep gradient along the

8 The fact that the kinetic pathway follows the thermodynamic one is non trivial,

and in general unwarranted, see [75, Chapter 19].

18 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

energy surface. This intuitive simplified view, however, must be amended

in several directions.
Ruggedness and frustration. Two important concepts which help to
describe landscapes are ruggedness or roughness and frustration. Rugged-
ness refers to the presence of local minima, which in a folding process may
correspond to partially folded states. Frustration refers to the presence of
several equally deep minima separated by significant barriers, which may
prevent the system from reaching the deepest one. The fact that most
proteins seem to have a single native state9 seems to advocate a minimal
frustration principle. Yet, even for non frustrated landscapes, several levels
of ruggedness may exist. In particular, on the easy side of the spectrum,
one finds proteins folding with a two-states kinetics, i.e. without any in-
termediates [85].
Ruggedness / frustration may actually come from two sources, namely
from the interaction energy between atoms of the protein, and/or from the
conformational entropy [68]. The enthalpic frustration comes from local
minima of the interaction potential energy. For the entropic frustration,
observe that the folding process is accompanied by a loss of conformational
entropy (of the protein). If this loss is heterogeneous and larger than the
energetic heterogeneity, the corresponding free energetic landscape becomes
frustrated.
5.2.3. Cooperativity and correlated motions. Another concept
related to minimally frustrated folding funnels is that of cooperative motions
between atoms. Cooperativity stipulates that when one atom is moving,
atoms nearby must move in a coherent fashion. This is rather intuitive
for condensed states where local forces (repulsion forces as atoms cannot
inter-penetrate, hydrogen bonding) are prominent. At a more global scale,
cooperation is likely to also be important, e.g. due to electrostatic inter-
actions. From a technical point of view, simple illustrations of correlated
motions are provided by normal modes studies10 , as well as correlations
between positional fluctuations11 .
Having mentioned correlated motions of atoms, the fact that dimen-
sionality reduction techniques play a key role in modeling folding (and more
generally the behavior of macro-molecular systems) is expected. First, the
9 As opposed to many polymers which exist under a number of energetically equiva-

lent inter-convertible states.

10 Assuming the system is at a minimum of its potential energy V , the dominant term

in the Taylor series expansion of V is the quadratic one. Diagonalizing the corresponding
quadratic form yields the so-called normal modes, whose associated eigenvectors are
collective coordinates. See for example [97].
11 Given a molecular dynamics simulation, one may investigate the correlations be-

tween the atomic fluctuations —with respect to a reference conformation. Both PCA
and MDS have been used for this problem: in [78, 59], the average covariance matrix of
the positional fluctuations is resorted to, while [83] computes the average Gram matrix.
See also [89] for a characterization of pairwise atomic correlations based of Pearson’s
coefficients and relatives.
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 19

d degrees of freedom are certainly not equivalent, as different time-scales

are clearly involved: from small amplitude - high frequency vibrations apart
from chemical bonds, to large amplitude - slow frequency deformations of
the protein. Second, the constraints inherent to the large amplitude mo-
tions are such that one expects the effective parameters to lie on some lower-
dimensional manifold representing the effective energy landscape, that is
the one accounting for transitions.

5.3. Bio-physics: Pre-requisites.

5.3.1. Molecular dynamics simulations. The simulation data we

shall be concerned with are essentially molecular dynamics (MD) data.
(The reader is referred to [77] for alternate simulation methods, such as
Monte Carlo simulations or Langevin dynamics.) A MD simulation is a
deterministic process which evolves a system according to the Newtonian
laws of motion. Central in the process is the force field associated to the
system, or equivalently the potential energy stemming from the interac-
tions between atoms. A typical potential energy involves bonded terms
(energies associated to covalent bonds), and non bonded terms (Van der
Waals interactions and electrostatic interactions). From the potential en-
ergy V associated to two atoms, one derives an associated force. Given
these forces, together with the positions and momenta of the atoms, one
determines the configuration of the system at time t + ∆t. Practically, ∆t
is of the order of the femtosecond, so that in retaining one conformation
every 10, long simulations (beyond the nanosecond) result in a number of
conformations > 100, 000.

5.3.2. Models, potential energy landscapes and their rugged-

ness. As exploring exhaustively the energy landscape of large atomic mod-
els is not possible, a number of coarse models mimicking the properties of
all atoms models have been developed. We may cite the united residues
model [74]; the BLN model [63] which features three types of beads only
(hydrophilic, hydrophobic, neutral); the 20 colors beads model [71], which
accommodates anisotropic interactions between residues so as to maximize
packing of side chains.
Such coarse models deserve a comment about the ruggedness of po-
tential energy landscapes. Ruggedness and frustration are indeed clearly
related to the complexity of the force field governing the system, since non
local interactions between atoms are likely to yield local minima of the
landscape —cf the Go models thereafter. On the other hand, non local
interactions are likely to help the protein to overcome local energy barri-
ers (to escape the local minima of the rugged landscape) due to solvent
collisions, non-native contacts, etc. See for example [92].
Having mentioned energy landscapes and MD simulations, a crucial
remark is in order. Following the gradient vector field of the energy on
the potential energy surface amounts to a mere energy minimization. But
20 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

MD simulations are more powerful, since a system evolved by a MD can

cross energy barriers thanks to its kinetic energy12 . Another way to cross
barriers is to resort to a Monte Carlo simulation [77].

5.3.3. Morse theory and singularity theory. As outlined above,

the properties of a system are described by its energy landscape. To investi-
gate transitions of our macro-molecular system, the topographical features
of the landscape i.e. its minima, maxima, and passes are of utmost impor-
tance [99]. These features are best described in terms of Morse theory [91]
as well as singularity theory [23], which in our setting amounts to studying
the gradient vector of the energy function on the manifold.
Following classical terminology, a critical point of a differentiable func-
tion is a point where the gradient of the function vanishes, and the func-
tion is called a Morse function if its critical points are isolated and non-
degenerate. For a critical point p of such a function, the stable (unstable)
manifold W s (p) (W u (p)) is the union of all integral curves associated to the
gradient of the function, and respectively ending (originating) at p. Locally
about a critical point of index i (the Hessian has i negative eigenvalues), the
(un-)stable manifold is a topological disk of dimension i (d − i). The stable
and unstable manifolds are also called the separatrices, as they partition
the manifold into integral curves having the same origin and endpoint. In a
more prosaic language, they are also called watersheds, by analogy with wa-
ter drainage. In particular, under mild non degeneracy assumptions of the
energy landscape, a transition between two adjacent minima is expected to
correspond to the stable manifold of index one saddle joining the minima
–a result known as the Murrell-Laidler theorem in bio-physics [99].
If Morse theory provides a powerful framework to describe energy land-
scapes, the pieces of information provided should be mitigated by the rel-
ative energies associated to critical points of various indices. As already
noticed at the end of Section 5.3.2, the thermal energy of the system indeed
allows barrier crossing.

5.3.4. Free energy landscapes and reaction coordinates. In

classical chemistry, a chemical systems moves from one minimum of energy
to another following the minimum energy path, which, as just discussed is
expected to go through index one saddles and intermediate minima. For
complex systems such as a protein in its aqueous solution, things are more
involved [72, 65, 69]. The different parameters have different relaxation
times: fast parameters are those describing the solvent, as well as the vari-
ables accounting for the fast vibrations apart from covalent bonds of the
protein; slow ones account for the large amplitude motions of the protein.
Because the system equilibrates faster for some coordinates than others,

12 If the internal (potential+kinetics) energy remains constant along the MD simula-

tion, the system is Hamiltonian, and a large number of mathematical results apply [94].
We shall get back on this issue in the outlook.
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 21

we may partition the parameters as x = (q, s). Denote V (x) the poten-
tial energy of the system. By focusing on q and averaging out the other
parameters, one defines the free energy landscape, which is the kinetically
relevant one, by:
Z  
V (q, s)
W (q) = −kT ln exp − ds. (2)
kT

Coordinates q which provide kinetically relevant informations on tran-

sitions are called reactions coordinates. Finding such coordinates is chal-
lenging, even on simple systems. We illustrate these difficulties with a two
dimensional system corresponding to a two states folding protein, whose
unfolded and folded states are respectively denoted A and B. If q is the
reaction coordinate sought, obvious requirements are (i) q takes different
values qA and qB for these states, and (ii) q is such that the free energy
W has a maximum at some value q ∗ in-between qA and qB . When these
conditions are met, q is called an order parameter. If q provides in addition
informations about the kinetics of the transitions, it is called a reaction
coordinate. As illustrated on Fig. 3(a,b), these are different notions. In
particular, Fig. 3(b) features a parameter q which is a good order param-
eter but not a reaction coordinate. For example, the dashed trajectory
passes through q ∗ but does not correspond to a transition. In the ideal
setting, for a reaction coordinate, the unstable manifold of the index one
saddle joining the two minima separates the points which are committed to
one state or the other, and thus determines the so-called Transition State
Ensemble (TSE).
Practically, dealing with reaction coordinates poses several problems.
First, for a system such as a protein and its solvent, one does not know a
priori which variables are the slow ones. This issue is further discussed at
the end of Section 5.4.3. Second, if there is not a unique coordinate which
is slower than the remaining ones, a multi-dimensional analysis must be
carried out. Third, computing a free energy profile from Eq. (2) requires
the coordinates over which the integration is performed to be equilibrated.
5.3.5. Folding probability pf old . To probe the relevance of a pa-
rameter as a reaction coordinate, one resorts to the committor probability,
i.e. the probability of being committed to a given state [72]. More precisely,
for any state x in the system, this is the probability of arriving say at B
before before arriving at A within some time ts . If the potential energy de-
pends on positions and momenta, averaging is understood w.r.t. momenta.
By studying this probability along a given path, one locates points near
the TSE, since such points are equally committed to both states. Denote
Dirac’s delta function δ, and let < z >E the average of quantity z over
an ensemble E. To probe the interest of an order parameter as a reaction
coordinate, one studies the distribution of the committor probability at
q = q ∗ , that is
22 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

(a) (b)
0
q q0

q q

W (q) W (q)

qA q∗ qB q qA q∗ qB q

Fig. 3. Potential energy landscape, with separatrices of the saddle in red: an order
parameter may not be a good reaction coordinate. Adapted from [65].

P (pB ) =< δ[pB (x, ts ) − pB ] >q∗ , pB ∈ [0, 1].

For a good reaction coordinate, one expects P (pB ) to be sharply peaked

at pB = 1/2. This is the case on Fig. 4(a), but not on Fig. 4(b) where
P (pB ) is bimodal, meaning that the orthogonal coordinates are such that
commitment to the two states occurs. The reader is referred to [72, 66, 65]
for example physical systems featuring various committor’s distributions.
The notion of transition state is also closely related to that of transition
path [82, 64]. Define a transition path T P as a path in phase space that
exits a region about the unfolded state, and reaches a region about the
folded state. A collection of transition paths determines a conditional phase
space density p(x | T P ), and one has
p(x | T P )p(T P )
p(T P | x) = , (3)
peq (x)

with peq (x) the equilibrium probability of state x and p(T P ) the fraction
of time spent on transition paths. Transition states are naturally defined
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 23

(a)
F

P
A B

q∗ q PB

(b) F

B s

q∗ q
PB

Fig. 4. Probing a reaction coordinate by computing the committor probabilities pB .

Adapted from [66].

as points maximizing p(T P | x). Moreover, denoting x a point with same

position and reversed momentum, and pA (x) the probability of reaching
state A before state B from x, it can be shown [82] that

p(T P | x) = pA (x)pB (x) + pA (x)pB (x). (4)

An important property of this equation is that one can project x onto

a lower dimensional space –see Section 5.4. Denoting r = r(x) such a
coordinate, it can be shown [82] that
p(r | T P )p(T P )
p(T P | r) = . (5)
peq (r)
24 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

Free energy

Reaction coordinate #2

Transition State

F
U

Reaction coordinate #1

Fig. 5. Crossing the Transition State on a rugged energy landscape: the system
moves from one watershed (state U ) to a neighboring watershed (state F ) by crossing
the energy landscape pass. Adapted from [69].

Practically, the difficulty with pf old and related quantities are several
[69]. First, the concept is bound to simple landscapes corresponding to two
states folding processes. Most importantly, estimating pf old requires sam-
pling the TSE, which either requires long simulations –usually out of reach,
or some form of importance sampling to favor the rare events corresponding
to crossings of the TSE.
5.4. Inferring reaction coordinates. In the following, we review
some of the most successful techniques to analyze transitions. We focus on
the methodological aspects, and refer the reader to the original papers for
a discussion of the insights gained, including connexions with experimental
facts. As it can be seen from [69] for example, assessing the relevance of a
particular coordinate can be rather controversial.
5.4.1. Reaction coordinates? In order to prove efficient to investi-
gate folding, funnels such as that of Fig. 2 must be made quantitative, that
is, one needs to specify what the axis account for. The variables parame-
terizing the axis are called reaction coordinates, and a quantitative energy
landscape is displayed on Fig. 5. We now discuss several ways to design
such coordinates.
5.4.2. Contacts based analysis. Following the work of Go [80], a
natural way to tackle Levinthal’s paradox consists of introducing a bias in
the energy function towards native contacts, i.e. contacts observed in the
folded state. More precisely, two residues which are not adjacent along the
primary sequence of the protein form a native contact if they are spatially
close in the protein’s native state. Such pairs of residues are associated
a favorable interaction energy, while the remaining ones are associated a
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 25

Q Pf old = 0.01 Pf old = 0.03 Pf old = 0.02 Pf old = 0.015 F (Q)

0.1 0

Pf old = 0.09 Pf old = 0.07

0.3 1

0.5 3

Pf old = 1.0

0.8 1

Fig. 6. Folding process down a folding funnel: fraction of native contacts Q in-
creases, free energy F (Q) crosses a barrier, pf old increases. Adapted From [69].

repulsive, neutral or less attractive interaction energy. Figure 6 illustrates

a folding process down a funnel, described using the fraction of native
contacts. On one hand, energy landscapes obtained with Go models are
generally minimally frustrated. On the other hand, as discussed in Section
5.2, removing non local contacts may impair the folding process. At any
rate and regardless of the energy model used, the fraction of native con-
tacts Q can be used as reaction coordinate. Alternative empirical reaction
coordinates, also exploiting the resemblance of a particular conformation
with the native state, are being used: the radius of gyration (the root mean
square distance of the collection of atoms from their center of mass), the
effective loop length and the partial contact order [68]. In particular, the
latter two coordinates are used in [68] to measure the fraction of confor-
mations that are actually accessible amongst the conformations with the
same degree of nativeness Q. Such measures are directly related to the
entropy of the system along the folding route, and thus allow one to assess
the entropic ruggedness of the free energy landscape.
The native contacts can be used in a more elaborate fashion. Following
[64], denote Q the matrix such that Qij = 1 if the distance between residues
i and j is less than some cutoff (e.g. 12Å), and 0 otherwise. Using a
26 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

weight matrix W = (wij ),Pthe contact matrix is projected onto a reaction

coordinate defined by r = ij wij qij . Starting from a random initialization
of matrix W , the weights are optimized so as to maximize a Gaussian fit
of p(T P | r) –see Eq. (5). In doing so, one ensures that all reactive
configurations are condensed in a single peak.
5.4.3. Dimension reduction based analysis. If one discards the
momenta of the points, an important question is to come up with a sim-
plified representation of the 3n dimensional energy landscape. Not sur-
prisingly, PCA and MDS have been used for this purpose13 . A typical
illustration is provided by [62], where a PCA analysis of the conforma-
tions is first performed. Using the two most informative eigenvectors, an
approximation of the landscape termed the energy envelope is computed.
Fine informations on barriers between watersheds of minima might be lost
–the ruggedness observed on a landscape computed from two PCA coordi-
nates is at best questionable, but one expects to retain the overall shape
of the watersheds. In [87], a PCA analysis is carried out on the critical
points of an energy landscape, rather than on the whole point cloud. This
analysis yields new coordinates, which can be plugged into the potential
energy function.
One step towards a finer analysis is made in [70], where an adaptation
of ISOMAP is used to derive new coordinates. The adaptions w.r.t. the
standard ISOMAP algorithm are threefold. First, the computation of the
nearest neighbors is done resorting to the least RMSD (lRMSD)14 . Sec-
ond, following [21], landmarks are used to alleviate the pairwise geodesic
distance calculations. Third, the point cloud is trimmed to get rid of redun-
dancies, which are expected in particular near the minima of potential en-
ergy. These conformations are later re-introduced into the low-dimensional
embedding, which is important in particular to recover statistical averages.
To assess the performance of the dimensionality reduction, a residual vari-
ance calculation is performed. For a two states folding protein, the tran-
sition state identified from the maximum of the free energy profile W (x1 )
associated to the first embedding coordinate x1 is in full agreement with
pf old . (A result also holding for the reaction coordinate Q in this case.)
Motivated by the fact that 95% of the running time is devoted to the cal-
culation of nearest neighbors, a further improvement is proposed in [95].
Assume m landmark conformations have been selected. Following the strat-
egy used by the General Positioning System, each conformation (a point
R3n ) is represented as a m-dimensional point whose coordinates are the
lRMSD distances to the m landmarks. In the corresponding m-dimensional
space, the l > k nearest neighbors of a point can be computed using the

13 Notice this analysis is different from the investigation of positional fluctuations

mentioned in Section 5.2.

14 The Root Mean Square Deviation computed once the two structures have been

aligned.
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 27

Euclidean distance, from which the k nearest ones according to the lRMSD
are selected.
To finish up this review, one should mention methods which do not
provide a simplified embedding of the landscape, but resort instead to a
clustering of the nodes in parameter space [81, 79]. Nodes within the same
watershed should belong to the same cluster, from which a Configuration
Space Network (CSN) can be built. In some cases, quantitative informa-
tions (e.g. free energies) can even be retrieved.
Remark 5.1. Having discussed dimensionality reduction techniques,
one comment is in order. If one does not know a priory which are the slow
variables, integrating Eq. (2) is not possible. This accounts for a three-
stage strategy which consists of performing a simulation, performing a di-
mensionality reduction to infer candidate reaction coordinates, and probing
them using pf old .
5.4.4. Morse theory related analysis. Energy landscapes govern
the folding process of proteins, but also the behavior of a number of physical
systems such as clusters of atoms, ions or simple molecules [67, 99]. For
some of these systems which exhibit a small number of stable crystalline
geometries and a large number of amorphous forms, exploring the landscape
exhaustively is impossible. Yet, a qualitative analysis can be carried out
by focusing on selected critical points. In [88, 60, 63], sequences of triples
minimum - saddle - minimum are sought, and super-basins are built from
their concatenation. In a related vein, the relative accessibility of potential
energy basins associated to minima is investigated in [61], so as to define the
so-called disconnectivity graph (DG). More precisely, two constructions are
performed in [61]. The first one, based on the canonical mapping, focuses
on the relative height of energy barriers, which governs transitions between
states, thus encoding the kinetic behavior of the system. The second one,
based on the canonical mapping, probes the potential energy surface at pre-
defined values of the energy, thus encoding global topological properties of
the landscape. Mathematically, constructing either DG is tantamount to
tracking the topological changes of the set V −1 (]∞, v]) when increasing v.
As such changes occur at critical values only [91], the graph built when all
critical values are available is called the contour tree15 . In [61], a discrete
set of energies are used to probe the topological changes of the level sets,
though.
If one focuses on the relative accessibility of basins, one problem aris-
ing is that the DG built does not have any privileged embedding —the
vertical axis encodes an energy, but the horizontal one does is meaning-
less. To complement the topological information by a geometric one, the

15 Consider the level sets of a Morse function f , and call a connected component of

a level set f −1 (h) a contour. Further contract every contour to a point. The graph
encoding the merge/split events between these contours is called the Reeb graph, or the
contour tree if the domain is simply connected [76].
28 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

following is carried out in [96]: first, similarly to [87], a PCA of critical

points is carried out, from which a two-dimensional embedding of these
critical points is derived; next, the DG is rendered as a three-dimensional
tree, the z coordinate corresponding to the potential energy. Interestingly,
such representations convey the (lack of) frustration of BLN models [63],
depending on the interaction energy used.
6. Outlook.
Algorithms. Exploring a high-dimensional point cloud with the meth-
ods discussed and mentioned raises critical issues which should be kept in
mind. First, it is usually assumed that the data points lie on a manifold.
But for complex data corresponding e.g. to physical phenomena featuring
bifurcations, a stratified complex might actually be the true underlying
structure. Even in the manifold case, since the underlying manifold M is
unknown, the geometric quantities we aim to preserve have to be estimated
from the data set. Coming up with robust estimators poses difficult ques-
tions, especially since noisy data (i.e. not exactly sampled on M ) has to
be accommodated from a practical standpoint. Worse, the sampling con-
ditions insuring that the geometry can be correctly inferred from the data
usually depend on some assumptions on M ... which is unknown! These
questions have been widely studied in computational geometry, in partic-
ular for the three dimensional surface reconstruction problem, but remain
largely open in a broader setting.
Closely related to the previous questions are those concerning the con-
vergence and theoretical guarantees. As discussed earlier, dimensionality
reductions methods are not well suited for all situations. It is thus impor-
tant to identify the necessary assumptions on M so as to ensure satisfactory
results. We have seen in Section 4 that one can answer this question for
some of the methods (ISOMAP, HLLE). It is also interesting to have in-
formations on the asymptotic behavior of the considered methods when
the samples become denser and denser and converge to M . In this way,
Hessian eigenmaps and diffusion maps reveal interesting asymptotic con-
nections with classical operators defined on the underlying manifold M
that need to be further explored.
Protein folding. In spite of three decades of intense research, the prob-
lem of protein folding is still open. In the context of energy landscapes and
dimensionality reduction, a number of further developments are called for.
A variety of (molecular dynamics) simulations are being performed:
depending on the system studied (all atoms/coarse, explicit/implicit/no
solvent), either the temperature, the pressure or the internal energy of the
system are kept constant. For example, if the temperature is held constant
using a thermostat —for example the Nose-Hoover, part of the internal
energy of the system is dissipated into the thermostat. If the internal
energy of the system is conserved, then, the system is Hamiltonian.
For Hamiltonian systems, a large number of mathematical results ex-
ist. For example, using the geometrization of Hamiltonian dynamics, a
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 29

trajectory of the system corresponds to a geodesic of a suitable Rieman-

nian manifold [94]. This point of view is not really used in recent folding
studies, which focus on Morse related analysis of potential and free energy
surfaces. The study of the relationship between folding properties inferred
from energy landscapes on the one hand, and from Hamiltonian dynamics
on the other hand deserves further scrutiny.
Practically, one or two reaction coordinates are usually dealt with,
a rather stringent limitation. Methods based on manifold learning are
appealing in this perspective, since the dimensionality of the embedding
can be estimated. But a critical step for these methods is that of the
neighborhood selection. On one hand, the samples are generally processed
in a uniform way since the same number of neighbors is used for all points.
On the other hand, Morse theory tells us that the local density of samples
about a critical point is related to its index. Therefore, a segmentation
of the point cloud might be in order before resorting to dimensionality
reduction techniques. Doing so might allow one to bridge the gap with
Morse theory related methods, whose focus has been on the decomposition
of the landscape into basins –as opposed to the design of new coordinates
accounting for transitions.
Another key problem is that of stability, in the context of rugged /
frustrated landscapes. Ideally, multi-scale analysis of landscapes should be
developed, so as to asses what is significant and what is not at a given
scale. Topological persistence and more generally tools developed in com-
putational topology might be helpful here. Such analysis might also allow
one to spot cascades of minor events in the folding process, such cascades
triggering major events –cf phase transitions.
Finally, an improved analysis of landscapes would have another dra-
matic impact, namely on the simulation processes themselves. Should a
finer understanding of cooperative motions be available, steered simula-
tions favoring these coordinates should allow one to move faster along a
(rugged) landscape.
Hopefully, a finer geometric and topological analysis of non linearities
arising on energy landscapes will help in making simulation able to cope
with biologically relevant time scales.

Acknowledgments. F. Cazals wishes to acknoledge Benjamin Bou-

vier, Ricardo Lima, Marco Pettini and Charles Robert for stimulating dis-
cussions.

REFERENCES

[1] P.K. Agarwal, S. Har-Peled, and H. Yu. Embeddings of surfaces, curves, and
moving points in euclidean space. In ACM SoCG, 2007.
[2] D. Agrafiotus and H. Xu. A self-organizing principle for learning nonlinear
manifolds. PNAS.
30 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

[3] M. Belkin and P. Niyogi. Towards a theoretical foundation for laplacian-based

manifold methods. In COLT 2005.
[4] M. Belkin and P. Niyogi. Laplacian eigenmaps for dimensionality reduction
and data representation. Neural Computation, 15(6):1373–1396, 2003.
[5] M. Belkin and P. Niyogi. Semi-supervised learning on riemannian manifolds.
Machine Learning, Invited, Special Issue on Clustering, pages 209–234, 2004.
[6] Y. Bengio, M.Monperrus, and H. Larochelle. Nonlocal estimation of mani-
fold structure. Neural Computation, 18, 2006.
[7] Y. Bengio, J.-F. Paiement, P. Vincent, O. Delalleau, N. Le Roux, and
M. Ouimet. Out-of-sample extensions for lle, isomap, mds, eigenmaps, and
spectral clustering. In NIPS, 2004.
[8] C.M. Bishop. Pattern Recognition and Machine Learning. Springer, 2007.
[9] C.M. Bishop, M. Svensen, and C.K.I. Williams. Gtm: The generative topo-
graphic mapping. Neural Computation, 10:215–234, 1998.
[10] M. Brand. Charting a manifold. In Advances in Neural Information Processing
Systems 15. MIT Press, Cambridge, MA, 2003.
[11] F. Chazal, D. Cohen-Steiner, and A. Lieutier. A sampling theory for com-
pact sets in euclidean space. In Proceedings of the 22nd ACM Symposium
on Computational Geometry, 2006.
[12] F. Chazal, D. Cohen-Steiner, and Q. Mérigot. Stability of boundary mea-
sures. 2007.
[13] Siu-Wing Cheng, Yajun Wang, and Zhuangzhi Wu. Provable dimension de-
tection using principal component analysis. In Symposium on Computational
Geometry, pp. 208–217, 2005.
[14] B. Christiansen. The shortcomings of nlpca in identifying circulation regimes.
J. Climate, 18:4814–4823, 2005.
[15] R.R. Coifman, S. Lafon, A. Lee, M. Maggioni, B. Nadler, F. Warner, and
S. Zucker. Geometric diffusions as a tool for harmonic analysis and structure
definition of data: Diffusion maps. Proc. of Nat. Acad. Sci., 102:7426–7431,
2005.
[16] R.R. Coifman, S. Lafon, A. Lee, M. Maggioni, B. Nadler, F. Warner,
and S. Zucker. Geometric diffusions as a tool for harmonic analysis and
structure definition of data: Multiscale methods. Proc. of Nat. Acad. Sci.,
102:7432–7437, 2005.
[17] J.A. Costa and A.O. Hero. Geodesic entropic graphs for dimension and entropy
estimation in manifold learning. IEEE Trans. on Signal Processing, 52(8),
2004.
[18] T.F. Cox and M.A. Cox. Multidimensional Scaling. Chapman Hall, 1994.
[19] V. de Silva and G. Carlsson. Topological estimation using witness complexes.
In Eurographics Symposium on Point-BasedGraphics, ETH, Switzerland,
2004.
[20] V. de Silva, J.C. Langford, and J.B. Tenenbaum. Graph approximations to
geodesics on embedded manifolds. 2000.
[21] V. de Silva and J.B. Tenenbaum. Global versus local methods in nonlinear
dimensionality reduction. In Advances in Neural Information Processing
Systems 15. MIT Press, Cambridge, MA, 2003.
[22] M. Dellnitz, M. Hessel von Molo, P. Metzner, R. Preiss, and C. Schutte.
Graph algorithms for dynamical systems. In A. Mielke, editor, Analysis,
Modeling and Simulation of Multiscale Problems. Springer, 2006.
[23] M. Demazure. Bifurcations and Catastrophes: Geometry of Solutions to Non-
linear Problems. Springer, 1898.
[24] D. Donoho and C. Grimes. Hessian eigenmaps: Locally linear embedding tech-
niques for high-dimensional data. Proceedings of the National Academy of
Sciences, 100(10):5591–5596, 2003.
[25] Y. Bengio et al. Learning eigenfunctions links spectral embedding and kernel
pca. Neural compuation, 16(10), 2004.
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 31

[26] J. Giesen and U. Wagner. Shape dimansion and intrinsic metric from samples
of manifolds with high co-dimension. In Proc. of the 19th Annual symp.
Computational Geometry, pp. 329–337, 2003.
[27] D. Givon, R. Kupferman, and A. Stuart. Extracting macroscopic dymamics.
Nonlinearity, 17:R55–R127, 2004.
[28] A. Globerson and S. Roweis. Metric learning by collapsing classes. In NIPS,
2005.
[29] Jihun Ham, Daniel D. Lee, Sebastian Mika, and Bernhard Schölkopf. A
kernel view of the dimensionality reduction of manifolds. In ICML ’04: Pro-
ceedings of the twenty-first international conference on Machine learning,
p. 47, New York, NY, USA, 2004. ACM.
[30] Gloria Haro, Gregory Randall, and Guillermo Sapiro. Stratification learn-
ing: Detecting mixed density and dimensionality in high dimensional point
clouds. In B. Schölkopf, J. Platt, and T. Hoffman, editors, Advances in Neu-
ral Information Processing Systems 19, pp. 553–560. MIT Press, Cambridge,
MA, 2007.
[31] T. Hastie and W. Stuetzle. Principal curves. J. Amer. Stat. Assoc., 84:502–
516, 1989.
[32] Matthias Hein and Markus Maier. Manifold denoising. In NIPS, pp. 561–568,
2006.
[33] Matthias Hein and Markus Maier. Manifold denoising. In B. Schölkopf,
J. Platt, and T. Hoffman, editors, Advances in Neural Information Process-
ing Systems 19, pp. 561–568. MIT Press, Cambridge, MA, 2007.
[34] I. Horenko, J. Schmidt-Ehrenberg, and C. Schutte. Set-oriented dimension
reduction: localizing principal component analysis vie hidden markov models.
In LNBS in Bio-informatics. 2006.
[35] B. Kégl. Intrinsic dimension estimation using packing numbers. In Advances
in Neural Information Processing Systems 17. MIT Press, Cambridge, MA,
2002.
[36] R.I. Kondor and J. Lafferty. Diffusion kernels on graphs and other discrete
structures.
[37] S. Lafon and A.B. Lee. Diffusion maps and coarse-graining: A unified frame-
work for dimensionality reduction, graph partitioning and data set parame-
terization. IEEE PAMI, 28(9):1393–1403, 2006.
[38] M.C. Law and A.K. Jain. Incremental nonlinear dimensionality reduction by
manifold learning. IEEE Trans. on pattern analysis and machine intelli-
gence, 28(3), 2006.
[39] J.A. Lee and M. Verleysen. Nonlinear Dimensionality Reduction. Springer,
2007.
[40] Elizaveta Levina and Peter J. Bickel. Maximum likelihood estimation of
intrinsic dimension. In Lawrence K. Saul, Yair Weiss, and Léon Bottou,
editors, Advances in Neural Information Processing Systems 17, pp. 777–
784. MIT Press, Cambridge, MA, 2005.
[41] Nathan Linial, Eran London, and Yuri Rabinovich. The geometry of graphs
and some of its algorithmic applications. In IEEE Symposium on Founda-
tions of Computer Science, pp. 577–591, 1994.
[42] J. Mao and A.K. Jain. Artificial neural networks for feature extraction and
multivariate data projection. IEEE Trans. Neural Networks, 6(2), 1995.
[43] E. Meerbach, E. Dittmer, I. Horenko, and C. Schutte. Multiscale modelling
in molecular dynamics : Biomolecular conformations as metastable states.
Lecture notes in physics, 703, 2006.
[44] F. Memoli and G. Sapiro. Distance functions and geodesics on point clouds,
2005.
[45] S.T. Roweis and L.K. Saul. Non linear dimensionality reduction by locally
linear embedding. Science, 290:2323–2326, 2000.
32 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

[46] S.T. Roweis and L.K. Saul. Think globally, fit locally: Unsupervised learning
of low dimensional manifolds. Journal of Machine Learning Research, 4:119–
155, 2003.
[47] J.B. Tenenbaum and V. de Silva. Sparse multi-dimensional scaling using land-
mark points. In preparation.
[48] J.B. Tenenbaum, V. de Silva, and J.C. Langford. A global geometric frame-
work for nonlinear dimensionality reduction. Science, 290:2319–2323, 2000.
[49] R. Tibshirani. Principal curves revisited. Statistics and Computing, 2:183–190,
1992.
[50] M. Trosset. Applications of multidimensional scaling to molecular conformation.
Computing Science and Statistics, (29):148–152, 1998.
[51] L.J.P. van der Maaten, E.O. Postma, and H.J. van den Herik. Dimension-
ality reduction: a comparative review. 2007.
[52] Kilian Q. Weinberger and Lawrence K. Saul. Unsupervised learning of image
manifolds by semidefinite programming. In CVPR (2), pp. 988–995, 2004.
[53] Kilian Q. Weinberger, Fei Sha, and Lawrence K. Saul. Learning a kernel
matrix for nonlinear dimensionality reduction. In ICML ’04: Proceedings of
the twenty-first international conference on Machine learning, p. 106, New
York, NY, USA, 2004. ACM.
[54] K.Q. Weinberger and L.K. Saul. An introduction to nonlinear dimensionality
reduction by maximum variance unfolding. In AAAI, 2006.
[55] K.Q. Weinberger and L.K. Saul. Unsupervised learning of image manifolds
by semidefinite programming. International Journal of Computer Vision,
70(1):77–90, 2006.
[56] Li Yang. Building connected neighborhood graphs for isometric data embedding.
In KDD, pp. 722–728, 2005.
[57] P. Zhand, Y. Huang, S. Shekhar, and V. Kumar. Correlation analysis of
spatial time series datasets. In Pacific Asia Conf. on Knowledge Discovery
and Data Mining, 2003.
[58] Hao Zhang, Oliver van Kaick, and Ramsay Dyer. Spectral mesh processing.
Computer Graphics Forum (to appear), 2008.
[59] A. Amadei, A.B.M. Linssen, and H.J.C. Berendsen. Essential dynamics of
proteins. Proteins: Structure, Function, and Genetics, 17(4):412–425, 1993.
[60] K.D. Ball, R.S. Berry, R.Kunz, F-Y. Li, A. Proykova, and D.J. Wales.
From topographies to dynamics on multidimensional potential energy sur-
faces of atomic clusters. Science, 271(5251):963 – 966, 1996.
[61] O. Becker and M. Karplus. The topology of multidimensional potential energy
surfaces: Theory and application to peptide structure and kinetics. The
Journal of Chemical Physics, 106(4):1495–1517, 1997.
[62] O.M. Becker. Principal coordinate maps of molecular potential energy surfaces.
J. of Comp. Chem., 19(11):1255–1267, 1998.
[63] R. Stephen Berry, Nuran Elmaci, John P. Rose, and Benjamin Vekhter.
Linking topography of its potential surface with the dynamics of fold-
ing of a proteinmodel. Proceedings of the National Academy of Sciences,
94(18):9520–9524, 1997.
[64] Robert B. Best and Gerhard Hummer. Chemical Theory and Computation
Special Feature: Reaction coordinates and rates from transition paths. Pro-
ceedings of the National Academy of Sciences, 102(19):6732–6737, 2005.
 TECHNIQUES TO EXPLORE POINT CLOUDS IN EUCLIDEAN SPACE 33

[65] P.G. Bolhuis, D. Chandler, C. Dellago, and P.L. Geissler. Transition path
sampling: Throwing ropes over rough mountain passes, in the dark. Annual
review of physical chemistry, 53:291–318, 2002.
[66] P.G. Bolhuisdagger, C. Dellago, and D. Chandler. Reaction coordinates of
biomolecular isomerization. PNAS, 97(11):5877–5882, 2000.
[67] C.L. Brooks, J. Onuchic, and D.J. Wales. Statistical thermodynamics: taking
a walk on a landscape. Science, 293(5530):612 – 613, 2001.
[68] L. Chavez, J.N. Onuchic, and C. Clementi. Quantifying the roughness on
the free energy landscape: Entropic bottlenecks and protein folding rates. J.
Am. Chem. Soc., 126(27):8426–8432, 2004.
[69] Samuel S. Cho, Yaakov Levy, and Peter G. Wolynes. P versus Q: Structural
reaction coordinates capture protein folding on smooth landscapes. Proceed-
ings of the National Academy of Sciences, 103(3):586–591, 2006.
[70] P. Das, M. Moll, H. Stamati, L. Kavraki, and C. Clementi. Low-
dimensional, free-energy landscapes of protein-folding reactions by nonlinear
dimensionality reduction. PNAS, 103(26):9885–9890, 2006.
[71] Payel Das, Corey J. Wilson, Giovanni Fossati, Pernilla Wittung-
Stafshede, Kathleen S. Matthews, and Cecilia Clementi. Characteri-
zation of the folding landscape of monomeric lactose repressor: Quantitative
comparison of theory and experiment. Proceedings of the National Academy
of Sciences, 102(41):14569–14574, 2005.
[72] R. Du, V. Pande, A.Y. Grosberg, T. Tanaka, and E.I. Shakhnovich. On the
transition coordinate for protein folding. J. Chem. Phys., 108(1):334–350,
1998.
[73] R.L. Dunbrack. Rotamer libraries in the 21st century. Curr. Opin. Struct. Biol.,
12(4):431–440, 2002.
[74] H.A. Scheraga et al. A united-residue force field for off-lattice protein-structure
simulations. i. functional forms and parameters of long-range side-chain inter-
action potentials from protein crystal data. J. of Computational Chemistry,
18(7):849–873, 1997.
[75] A. Fersht. Structure and Mechanism in Protein Science: A Guide to Enzyme
Catalysis and Protein Folding. 1999.
[76] A.T. Fomenko and T.L. Kunii. Topological Modeling for visualization. Springer,
1997.
[77] D. Frenkel and B. Smit. Understanding molecular simulation. Academic Press,
2002.
[78] A.E. Garcia. Large-amplitude nonlinear motions in proteins. Physical Review
Letters, 68(17):2696–2699, 1992.
[79] D. Gfeller, P. De Los Rios, A. Caflisch, and F. Rao. Complex network
analysis of free-energy landscapes. Proceedings of the National Academy of
Sciences, 104(6):1817–1822, 2007.
[80] Nobuhiro Go and Hiroshi Taketomi. Respective Roles of Short- and Long-
Range Interactions in Protein Folding. Proceedings of the National Academy
of Sciences, 75(2):559–563, 1978.
[81] Isaac A. Hubner, Eric J. Deeds, and Eugene I. Shakhnovich. Understand-
ing ensemble protein folding at atomic detail. Proceedings of the National
Academy of Sciences, 103(47):17747–17752, 2006.
[82] G. Hummer. From transition paths to transition states and rate coefficients. J.
Chemical Physics, 120(2), 2004.
[83] T. Ichiye and M. Karplus. Collective motions in proteins: A covariance analysis
of atomic fluctuations in molecular dynamics and normal mode simulations.
Proteins: Structure, Function, and Genetics, 11(3):205–217, 1991.
[84] C.L Brooks III, M. Gruebele, J. Onuchic, and P. Wolynes. Chemical
physics of protein folding. Proceedings of the National Academy of Sciences,
95(19):11037–11038, 1998.
34 FRÉDÉRIC CAZALS, FRÉDÉRIC CHAZAL, AND JOACHIM GIESEN

[85] S.E. Jackson. How do small single-domain proteins fold? Fold Des., 3(4):R81–
91, 1998.
[86] J. Janin, S. Wodak, M. Levitt, and B. Maigret. Conformations of amino
acid side chains in proteins. J. Mol. Biol., 125:357–386, 1978.
[87] T. Komatsuzaki, K. Hoshino, Y. Matsunaga, G.J. Rylance, R.L. Johnston,
and D. Wales. How many dimensions are required to approximate the po-
tential energy landscape of a model protein? J. Chem. Phys., 122, February
2005.
[88] R.E. Kunz and R.S. Berry. Statistical interpretation of topographies and dy-
namics of multidimensional potentials. J. Chem. Phys., 103:1904–1912, Au-
gust 1995.
[89] O.F. Lange and H Grubmller. Generalized correlation for biomolecular dy-
namics. Proteins, 62:1053–1061, 2006.
[90] C. Levinthal. Are there pathways for protein folding? Journal de Chimie
Physique et de Physico-Chimie Biologique, 65:44–45, 1968.
[91] John W. Milnor. Morse Theory. Princeton University Press, Princeton, NJ,
1963.
[92] E. Paci, M. Vendruscolo, and M. Karplus. Native and non-native interactions
along protein folding and unfolding pathways. Proteins, 47(3):379–392, 2002.
[93] J. Palis and W. de Melo. Geometric Theory of Dynamical Systems. Springer,
1982.
[94] M. Pettini. Geometry and Topology in Hamiltonian Dynamics and Statistical
Mechanics. Springer, 2007.
[95] E. Plaku, H. Stamati, C. Clementi, and L.E. Kavraki. Fast and reliable
analysis of molecular motion using proximity relations and dimensionality
reduction. Proteins: Structure, Function, and Bioinformatics, 67(4):897–
907, 2007.
[96] G. Rylance, R. Johnston, Y. Matsunaga, C-B Li A. Baba, and T. Komat-
suzaki. Topographical complexity of multidimensional energy landscapes.
PNAS, 103(49):18551–18555, 2006.
[97] M. Tirion. Large amplitude elastic motions in proteins from a single-parameter,
atomic analysis. Phys. Rev. Lett., 77:1905–1908, 1996.
[98] Monique M. Tirion. Large amplitude elastic motions in proteins from a single-
parameter, atomic analysis. Phys. Rev. Lett., 77(9):1905–1908, Aug 1996.
[99] D.J. Wales. Energy Landscapes. Cambridge University Press, 2003.
[100] L. Yang, G. Song, and R. Jernigan. Comparison of experimental and com-
puted protein anisotropic temperature factors. In IEEE Bioinformactics and
biomedecine workshop, 2007.
 RATIONAL PARAMETRIZATIONS,
INTERSECTION THEORY, AND NEWTON POLYTOPES∗
CARLOS D’ANDREA† AND MARTÍN SOMBRA‡

Abstract. The study of the Newton polytope of a parametric hypersurface is

currently receiving a lot of attention both because of its computational interest and
its connections with Tropical Geometry, Singularity Theory, Intersection Theory and
Combinatorics. We introduce the problem and survey the recent progress on it, with
emphasis in the case of curves.

Key words. Parametric curve, implicit equation, Newton polytope, tropical geom-
etry, Intersection Theory.

AMS(MOS) 2000 subject classifications. Primary 14Q05; Secondary 12Y05,

52B20, 14C17.

1. Introduction. Parametric curves and surfaces play a central role

in Computer Aided Geometric Design (CAGD) because they provide
shapes which are easy to plot. Indeed, a rational parametrization allows to
produce many points in the variety using only the elementary operations
(±, ×, ÷) of the base field.
For instance, consider the folium of Descartes (Figure 1). This plane
curve can be defined either by the equation x3 + y 2 − 3xy = 0 or as the
image of the rational map

3t2
 
3t
C 99K C , t 7→ , . (1)
1 + t3 1 + t3

The parametric representation is certainly more suitable for plotting

the curve. If instead we plot it using only its implicit equation, the result
is bound to be poor, specially around the singular point (0, 0) (Figure 2).
This is because in order to produce many points in the folium in this
way, we have to solve as many cubic equations. This is certainly more
expensive than evaluating the parametrization and moreover, the resulting
points are typically not rational but live in different cubic extensions of Q.
On the other hand, if we are to decide whether a given point lies in
the folium or not, it is better to use the implicit equation. For instance, it
is straightforward to conclude that (−2, 1) does not belong to the folium

∗ D’Andrea is partially supported by the Programa Ramón y Cajal of the Ministerio

de Ciencia e Innovación (Spain) and by the research project MTM2007–67493. Sombra

is partially supported by the research project MTM2006-14234.
† Universitat de Barcelona, Departament d’Àlgebra i Geometria, Gran Via 585, 08007

Barcelona, Spain ([email protected], http://carlos.dandrea.name).

‡ Université de Bordeaux 1, Institut de Mathématiques de Bordeaux, 351 cours

de la Libération, 33405 Talence Cedex, France ([email protected],

http://www.math.u-bordeaux1.fr/∼sombra).

I.Z. Emiris et al. (eds.), Nonlinear Computational Geometry, The IMA Volumes in 35
Mathematics and its Applications 151, DOI 10.1007/978-1-4419-0999-2_2,
© Springer Science + Business Media, LLC 2010
36 CARLOS D’ANDREA AND MARTÍN SOMBRA

1,5

1,0

0,5

−2 −1 0 1
0,0

−0,5

−1,0

−1,5

Fig. 1. The folium of Descartes.

Fig. 2. The folium of Descartes according to the Maple command implicitplot.

by evaluating the equation: (−2)3 + 13 − 3(−2) = −1 6= 0. If we were to

find that out from the parametrization (1), we would have to determine if
the system of equations

3t 3t2
−2 = , 1=
1 + t3 1 + t3
admits a solution for t ∈ C or not, which is a harder task.
Depending on which kind of operation one needs to perform on a cer-
tain parametric variety, it may be convenient to dispose of the parametric
representation or of the implicit one. Efficiently performing the passage
from one representation to the other is one of the central problems of
Computational Algebraic Geometry. In the present text we will mostly
concentrate in one these directions: the implicitization problem, consisting
in computing equations for an algebraic variety given in parametric form.
 INTERSECTION THEORY AND NEWTON POLYTOPES 37

In precise terms, the implicitization problem is: let ρ1 , . . . , ρn ∈

C(t1 , . . . , tn−1 ) be a family of rational functions and consider the map

ρ : Cn−1 99K Cn , t = (t1 , . . . , tn−1 ) 7→ (ρ1 (t), . . . , ρn (t)). (2)

Suppose that the Zariski closure Im(ρ) of the image of this map is a hyper-
surface or equivalently, that the Jacobian matrix ( ∂ρ∂tj (t))i,j has maximal
i

n−1
rank n − 1 for generic t ∈ C . The ideal of this parametric (or unira-
tional) hypersurface is generated by a single irreducible polynomial and the
problem consists in computing this “implicit equation”.
This problem is equivalent to the elimination of the parameter vari-
ables from some system of equations. For instance, to compute the implicit
equation of the folium from the parametrization (1), one should eliminate
the variable t from the system of equations

(1 + t3 ) x − 3t = 0 , (1 + t3 ) y − 3t2 = 0, (3)

that is, we have to find the irreducible polynomial in C[x, y] vanishing at

the points (x, y, t) satisfying (3) for some t ∈ C.
The same procedure works in general. For a parametrization like in (2),
(t)
write ρi (t) = pqii(t) for some coprime polynomials pi , qi for 1 ≤ i ≤ n.
The implicit equation of the hypersurface Im(ρ) can then be obtained by
eliminating the variables t1 , . . . , tn−1 from the system of equations

q1 (t)x1 − p1 (t) = 0, . . . , qn (t)xn − pn (t) = 0.

This elimination task can be effectively done either with Gröbner bases or
with resultants [3] but in practice, this can be too expensive. For instance,
for a ∈ N consider the parametrization

t(t − 1)a (t + 1)a

 
2
ρ : C → C , t 7→ , . (4)
(t + 1)a+1 t(t − 1)a−1

It is not hard to check by hand that the implicit equation of the image of
this map is

2 − xa−1 y a − xa y a+1 = 0.

However, all current implementations of Gröbner bases and resultant algo-

rithms fail to solve the problem for moderately large values of a, because
of the increasing number of intermediate computations involved.
2. The Newton polytope of the implicit equation. Instead of
trying to compute the implicit equation of a parametric hypersurface, we
will focus in the problem of determining its Newton polytope. We will work
with Laurent polynomials, that is expressions of the form x−1 −2
2 +x1 x2 where
the exponents can be any integer numbers.
38 CARLOS D’ANDREA AND MARTÍN SOMBRA

Fig. 3. The Newton polygon of the folium of Descartes.

Definition 2.1. The Newton polytope N(F ) ⊂ Rn of a Laurent poly-

nomial F ∈ C[x±1 ±1
1 , . . . , xn ] is the convex hull of the exponents in its mono-
mial expansion.
This notion readily extends to hypersurfaces: we define the Newton
polytope N(Z) of a hypersurface Z ⊂ Cn as the Newton polytope of its
implicit equation; this polytope is well defined because the equation is
unique up to a scalar factor. For the case n = 2 we will apply the more
usual terminology of “polygon” instead of polytope. For instance, the
Newton polygon of the
folium x31 + x32 − 3x1 x2 = 0 is the convex hull
Conv (1, 1), (3, 0), (0, 3) (Figure 3).
The Newton polytope tells us which are the possible exponents occur-
ing in a given Laurent polynomial: if the polytope is small, then the poly-
nomial is sparse, in the sense that it has few monomials. It is an important
refinement of the notion of degree: if we denote by S := Conv(0, e1 , . . . , en )
the standard simplex of Rn , the degree of a polynomial F ∈ C[x1 , . . . , xn ]
is the least integer d such that N(F ) ⊂ dS. Note that the Newton polytope
of a polynomial (and a fortiori that of a hypersurface) is always contained
in the octant (R≥0 )n .
From now on, we will focus in the following problem: determine the
Newton polytope of a hypersurface given by a rational map ρ : Cn−1 99K
Cn . This problem is currently receiving a lot of attention because of its con-
nections with Tropical Geometry, Singularity Theory, Intersection Theory
and Combinatorics. The Newton polytope does not characterize the hy-
persurface but retains a lot of relevant information and as a consequence of
the research done during the last years, we now know that in plenty of cases
its computation is much simpler than that of the full implicit equation.
A preliminary version of this question was first posed by B. Sturmfels
and J.-T. Yu. In the context of the sparse elimination theory, their question
can be resumed in: “can I predict the Newton polytope of the implicit
equation from the Newton polytopes of the input parametrization?” In
precise terms:
 INTERSECTION THEORY AND NEWTON POLYTOPES 39

60

50

40

30

20

10

10 20 30

Fig. 4. The Newton polygon of the implicit equation of (5).

Problem 2.2. Let P1 , . . . , Pn ⊂ Rn−1 be lattice polytopes with non-

empty interior and consider the family of n Laurent polynomials in n − 1
variables
h i
an−1
X
ρi = λi,a ta1 1 · · · tn−1 ∈ C t±1
1 , . . . , t±1
n−1
a∈Pi ∩Zn−1

for 1 ≤ i ≤ n and λi,a ∈ C generic. Determine the Newton polytope of the

image of the parametrization t 7→ (ρ1 (t), . . . , ρn (t)).
A lattice polytope in Rn−1 is a polytope whose vertices lie in Zn−1 . The
hypothesis that the Pi ’s have non empty interior ensures that the image of
the parametrization is a hypersurface for a generic choice of the coefficients
λi,a (that is, in some non empty open set of the space of parameters).
The Newton polytope of this hypersurface does not depend on this generic
choice although the equation itself does.
As an example, let us consider the parametrization proposed by
A. Dickenstein and R. Fröberg [8]:

ρ : C → C2 , t 7→ t48 − t56 − t60 − t62 − t63 , t32 .

(5)

The Newton polytopes of the defining polynomials are relatively small: the
real interval [48, 63] and the singleton {32}. The exponents are rather
large, but in any case the implicit equation can be computed via the
Sylvester resultant. It’s Newton polygon is the triangle with vertices
(32, 0), (0, 48), (0, 63).
This example was studied by I. Emiris and I. Kotsireas, who succeeded
in determining the polygon by analysing the behavior of the resultant under
specialization, thus showing that it is sometimes possible to access to the
Newton polytope without computing the implicit equation [8]; see also [7]
for further applications of this method.
The recent irruption of Tropical Geometry in the mathematical
panorama has boosted the interest in the problem. The tropical variety
40 CARLOS D’ANDREA AND MARTÍN SOMBRA

associated to an affine hypersurface is a polyhedral object, equivalent to

its Newton polytope in the sense that one can be recovered from the other
and viceversa. In this direction, Sturmfels, J. Tevelev and J. Yu succeeded
in determining the tropical variety (and thus the Newton polytope) of a
hypersurface parametrized by generic Laurent polynomials [15, 16] and im-
plemented the resulting algorithm [17, 18].
From another point of view, A. Esterov and A. Khovanskiı̆ have shown
that the Newton polytope of the implicit equation of a generic parametriza-
tion can be identified with the mixed fiber polytope in the sense of P. Mc-
Mullen, hence providing a different characterization of this object [9].

2.1. The Newton polygon of a parametric curve. If one wants

to determine the Newton polytope in all cases and not just in the generic
ones, it is clear that finer invariants of the parametrization must be taken
into account.
In this section we will focus in the case of parametric plane curves,
which has been recently solved in the papers [4, 5, 15, 20]. In this case, the
Newton polygon is determined by the multiplicities of the parametrization.
Let ρ : C 99K C2 be a map given by rational function f, g ∈ C(t) \ C. For
a point v in the projective line P1 , the multiplicity of ρ in v is

ordv (ρ(t)) := ordv (f (t)), ordv (g(t)) ∈ Z2 ,

where ordv (f ) denotes the order of vanishing of f at v. Recall that the

order of vanishing at v = ∞ of a rational function pq ∈ C(t) (p, q ∈ C[t])
equals deg(q) − deg(p).
The basic properties of these multiplicities are:
ordv (ρ) = (0, 0) except for a finite number of v ∈ P1 and
• P
• v∈P1 ordv (ρ) = (0, 0).

We next define an auxiliary operation which produces a convex lattice

polygon from a balanced family of vectors of the plane. Let B ⊂ Z2 be
a family ofPvectors which are zero except for a finite number of them and
such that b∈B b = (0, 0). We denote by P(B) ⊂ (R≥0 )2 the (unique)
convex polygon obtained by: 1) rotating −90◦ the non-zero vectors of B,
2) concatenating them following their directions counterclockwise and 3)
translating the resulting polygon to the first quadrant (R≥0 )2 in such a way
that it “touches” the coordinate axes (Figure 5). The zero-sum condition
warrants that the polygonal line closes at the end of the concatenation step.
The tracing index (or degree) ind(ρ) ≥ 1 is the number of times the
parametrization ρ runs over the curve when t runs over C. When ind(ρ) =
1, we say that the parametrization is birational.
The solution to the problem of the computation of the Newton polygon
of a parametric plane curve can be found in the papers of Dickenstein, E.-
M. Feichtner, Sturmfels and Tevelev [5, 15, 20] and also in ours [4].
 INTERSECTION THEORY AND NEWTON POLYTOPES 41

B ⊂ Z2 1)

×2

×2

2) 3)

P(B)

Fig. 5. The operation P(B).

Theorem 2.1. Let ρ : C 99K C2 be a rational map and set C := Im(ρ),

then
1

N(C) = P (ordv (ρ))v∈P1 . (6)
ind(ρ)

Example 1. Consider the parametrization

 1 t2 − 5t + 2 
ρ : t 7→ , .
t(t − 1) t
Its multiplicities are

ord0 (ρ) = (−1, −1) , ord1 (ρ) = (−1, 0) , ord∞ (ρ) = (2, −1)

and ordvi (ρ) = (0, 1) for each of the two zeros v1 , v2 of the equation
t2 − 5t + 2 = 0, while ordv (ρ) = (0, 0) for v 6= 0, ±1, ∞, v1 , v2 . Figure 5
illustrates the family B and the associated polygon P(B).
Theorem 2.1 tells us that this polygon is ind(ρ) times the actual New-
ton polygon of the curve. It is easy to check that the constructed polygon
is non contractible, in the sense that it is not a non trivial integer multiple
of another lattice polygon. We conclude that the parametrization is bira-
tional (that is, ind(ρ) = 1) and that N(C) = P(B). These results can be
Another random document with
no related content on Scribd:
 PLATE CLXXXVIII.

GLADIOLUS CAMPANULATUS.

Bell-flowered Gladiolus.

CLASS III. ORDER I.

TRIANDRIA MONOGYNIA. Three Chives. One Pointal.

ESSENTIAL GENERIC CHARACTER.

Corolla sexpartita, ringens.

Stamina adscendentia.
Blossom six divisions, gaping.
Chives ascending.
See Pl. XI. Vol. I. Gladiolus Roseus.

SPECIFIC CHARACTER.

Gladiolus foliis lanceolatis, nervosis, glabris; scapo subtrifloro, foliis

longior; corolla sub-campanulata, palidè purpurea, laciniis sub-æqualibus;
stigmatibus bifidis.
Gladiolus with lance-shaped leaves, nerved and smooth; flower-stem
mostly three-flowered, longer than the leaves; blossom rather bell-shaped, of
a pale purple, the segments nearly equal, with the summits two-cleft.

REFERENCE TO THE PLATE.

1. The Sheaths of the Empalement.

2. A Flower spread open, with the Chives attached.
3. The Seed-bud, Shaft, and Summits, one Summit detached and
magnified.
The Bell-flowered Gladiolus, was amongst the number of those imported
from Holland, in the year 1794, by Messrs. Lee and Kennedy,
Hammersmith; when they partook of that large collection, brought to
Haarlem by a Frenchman; who had been long resident at the Cape of Good
Hope, where he had cultivated most of the bulbs prior to his bringing them
to Europe. Nothing particular is required for the management of this, more
than the most common of the Genus, from the Cape. It flowers in May, and
increases by the root; the seeds rarely ripen.
 PLATE CLXXXIX.

ZINNIA VERTICILLATA.

Double Zinnia.

CLASS XIX. ORDER II.

SYNGENESIA POLYGAMIA SUPERFLUA. Tips united. Superfluous
Pointals.

ESSENTIAL GENERIC CHARACTER.

Receptaculum paleaceum. Pappus aristis 2 erectis. Calyx ovato-

cylindricus, imbricatus. Flosculi radii 5, persistentes, integri.
Receptacle chaffy. Feather with 2 upright awns. Empalement
cylindrical-egg-shaped, and tiled. Florets of the ray 5, remaining and entire.
See Zinnia Violacea. Pl. LV. Vol. I.

SPECIFIC CHARACTER.

Zinnia foliis verticillatis, sessilibus; floribus pedunculatis; flosculi radii

sæpe tria series.
Zinnia with leaves growing in whorls without foot-stalks close to the
stem; flowers with foot-stalks; the florets of the ray often three rows.

REFERENCE TO THE PLATE.

1. An outer female Floret of the ray, the seed attached, a little

larger than nature.
2. An inner hermaphrodite Floret of the disk, with its seed and
skinny chaff, magnified.
3. The Chives, Pointal, and Seed of an hermaphrodite Floret,
divested of its corolla, magnified.
The English specific title to this plant, should seem to imply, that the flowers
are such, as should not come into our arrangement; but, as the character is
not constant in all the flowers, even on the same plant, it cannot be
considered but as a specific character in this particular species, though the
name has its proper force, in contradistinction to its congeners, in our
language. It is a native of Mexico, South America; and was introduced to our
gardens about the year 1789, by Monsʳ Richard, from the Paris gardens, at
the same time with the Virgilia; a most beautiful annual, of the habit of
Arctotis, now lost in both countries from the difficulty of procuring ripened
seeds. It is to be raised in the same manner as the other species, on a gentle
hot-bed, in March, and planted out the beginning of May. The flowers make
their appearance about the beginning of August, and continue, in succession,
till they are destroyed by the frost. To be certain of the seed, the heads must
be taken from the plant, whilst they appear yet fresh; as the petals are
persistent, and have not the appearance of entire decay, though the seed is
nearly ripe; for if the receptacle once begins to rot, (which it is very subject
to,) the seeds are immediately contaminated and spoilt. Our figure was
taken, this year, at the Hammersmith Nursery, where, it was grown first in
this kingdom.
 PLATE CXC.

GERANIUM ASTRAGALIFOLIUM.

Astragalus-leaved Geranium.

CLASS XVI. ORDER IV.

MONADELPHIA DECANDRIA. Threads united. Ten Chives.

ESSENTIAL GENERIC CHARACTER.

Monogyna. Stigmata quinque.

Fructus rostratus, penta-coccus.
One Pointal. Five Summits.
Fruit furnished with long awns; five dry berries.
See Geranium Grandiflorum, Pl. XII. Vol. I.

SPECIFIC CHARACTER.

Geranium foliis pinnatis, hirsutis, foliolis rotundato-ovatis; calycibus

monophyllis; petalis undulatis ad basin tortis; staminibus quinque fertilibus;
radice tuberosa.
Geranium with winged, hairy leaves; leaflets of a roundish-oval shape;
cups one-leaved; petals waved, twisted at the base; five fertile chives; root
tuberous.

REFERENCE TO THE PLATE.

1. The Empalement.
2. The Chives and Pointal.
3. The Chives spread open and magnified.
4. The Seed-bud, Shaft, and Summits, magnified.
 By the Kew Catalogue we are informed, that this species of Geranium,
was introduced to this country, in the year 1788, by Mr. F. Masson. It is, like
many of this branch of the extended family of Geranium, rather a tender
Green-house plant; and will not flower, in perfection, without the assistance
of the Hot-house. It loses its foliage after flowering, and remains in a state of
inaction for at least three months; during which period, it should be watered
but seldom, and that sparingly. To propagate it, the only mode is, by cutting
small portions of the root off, and putting them into the strong heat of a hot-
bed, about the month of March; as hitherto, it has not perfected any seeds
with us, and the plant produces no branch, except the flower-stem may be so
denominated. Our drawing was made from the Clapham Collection, in July,
this year. This species has been considered by Professor Martyn, (see his
edition of Mill. Dict. article Pelargonium 2.) as the same with G. pinnatum,
and G. prolificum of Linn. Sp. Plan. But, however, the specific characters in
Linnæus, of those species, may agree with our figure, the G. Astragalifolium
of Jacquin and Cavanilles, they are, unquestionably, all different plants;
drawings of the two former we have, and will be given in due course.
 PLATE CXCI.

PLATYLOBIUM SCOLOPENDRUM.

Scolopendra-like stemmed Flat-Pea.

CLASS XVII. ORDER IV.

DIADELPHIA DECANDRIA. Threads in two Sets. Ten Chives.

GENERIC CHARACTER.

Calyx. Perianthium, campanulatum quinquedentatum; laciniis tribus

inferioribus acutis, patentibus; duabus supremis maximis, obtusis, obovatis,
vexillo adpressis.
Corolla papilionacea.
Vexillum, obcordatum, emarginatum, erectum, maximum.
Alæ vexillo breviores, obtusæ, semi-obcordatæ, basi denticulatæ.
Carina obtusa, compressa, longitudine et figura alarum.
Stamina filamenta decem, coalita in vaginam, supra semifissam, apice
libera, æqualia, assurentia. Antheræ subrotundæ, versatiles.
Pistillum. Germen lineare, pilosum. Stylus incurvatus, glaber. Stigma
simplex.
Pericarpium. Legumen pedicellatum, compressum, obtusum,
mucronatum, uniloculare, dorso alatum.
Semina, plurima, compressa, reniformia.
Empalement. Cup bell-shaped, five-toothed; the three lower segments
pointed, spreading; the two upper very large, obtuse, pressed to the standard.
Blossom butterfly-shaped.
Standard, inversely heart-shaped, notched at the end, upright, very large.
Wings shorter than the standard, obtuse, half inversely heart-shaped,
toothed at the base.
Keel, obtuse, flattened, the length and shape of the wings.
 Chives. Ten threads, united into a sheath, half cleft on the upper side,
separate at the top, equal and turned upwards. Tips roundish, versatile.
Pointal. Seed-bud linear, hairy. Shaft turned inwards, smooth. Summit
simple.
Seed-vessel. Pod with a footstalk, flattened, obtuse, with a small point,
one-celled, winged along the back.
Seeds many, flattened, kidney-shape.

SPECIFIC CHARACTER.

Platylobium foliis ovatis, glabris; ramis ramulisque compressis, alatis,

margine, cicatrisatis, floribus solitariis.
Flat-pea with egg-shaped smooth leaves, larger and smaller branches flat,
winged and hatched at the edges; flowers solitary.

REFERENCE TO THE PLATE.

1. The Empalement, natural size.

2. The Standard of the blossom.
3. One of the Wings of the blossom.
4. The two petals of the Keel.
5. The Chives and Pointal, with part of the cup, magnified.
6. The Seed-bud magnified.
This Genus of plants was first named by Dr. Smith, in the Linn. Trans. Vol.
II. 350, from the P. formosum, which he afterwards figured in the New-
Holland specimens, Tab. VI. Our species was introduced, to Britain, in the
year 1792, by Messrs. Lee and Kennedy. It is a hardy greenhouse plant; but
has not, hitherto, been increased in this country. It must be planted in very
sandy peat earth, and not much watered, in winter, as too much wet is apt to
destroy it. The young branches, which in the old plant appear much more
like leaves, (as seldom any leaves are produced from the upper part of the
plant, after a certain age,) are very tender; but in time become as tough as
leather, and are almost equally pliable. Our drawing was taken in May 1799,
from a plant, we believe, the first that flowered in England, in the Hibbertian
Collection.
 PLATE CXCII.

ANTHOLYZA FULGENS.

Refulgent-flowered Antholyza.

CLASS III. ORDER I.

TRIANDRIA MONOGYNIA. Three Chives. One Pointal.

ESSENTIAL GENERIC CHARACTER.

Corolla tubulosa, irregularis, recurvata. Capsula infera.

Blossom tubular, irregular and bent backward. Capsule beneath.
See Antholyza ringens, Pl. XXXII. Vol. I.

SPECIFIC CHARACTER.

Antholyza floribus tubiformibus, curvatis, coccineis, fulgentibus; laciniis

corollæ maximis, patentibus; foliis longissimis, glabris, basi attenuatis.
Antholyza with trumpet-shaped flowers, curved, scarlet, and refulgent;
the segments of the blossom very large, spreading; leaves very long, smooth,
and tapered at the base.

REFERENCE TO THE PLATE.

1. Part of a Leaf, cut from the upper part.

2. The two sheaths of the Empalement.
3. The Flower cut open, with the chives attached.
4. The Pointal and Seed-bud; one of the summits detached and
magnified.
This most beautiful genus does not possess amongst its numerous species,
(drawings of twenty-two of which we have) a rival to A. fulgens; whether,
for the size of the plant, which grows to the height of three feet, or the
extreme brilliancy of its blossoms, which frequently make a spike near a foot
in length. The roots should not be taken from the pots, but shifted into fresh
earth annually, which may be a composition of half sandy peat, and half
loam, as the leaves do not decay, until fresh ones are produced. Our figure
was taken at the Hammersmith Nursery, in May 1800, to which it was first
brought, from the Cape of Good Hope, in 1792. It increases by the root.
In a cotemporary, and something similar publication to our own, we were
sorry to observe, a rising itch to do away, what, under the conduct of its
original scientific proprietor, was allowed by all, to constitute its chief merit
and utility; especially to those, “who wish to become scientifically
acquainted with the plants they cultivate.” The late Mr. Curtis, pursuing the
path he planned, with rigour, to prevent confusion, and avoid as much as
possible the greatest difficulty of the science; seldom altered a commonly
known, or established name; unless absolutely necessary to systematic
arrangement. We were naturally led to these obvious observations, from the
hints thrown out in the last Number of the Bot. Mag. in which, the A.
tubulosa of all the collections, which possess the plant, and so named and
figured by us, in the preceding Number of the Botanists Repository, has a
new generic and specific title; and in which a gentleman “with INFINITE
skill” of the name of Gawler, the acknowledged father of the innovation, is
spoken of as qualified to scrutinize and rectify the “errors, false synonims,
and blunders upon blunders, which have from carelessness, &c.” crept into
the, of course, insignificant labours of a Linnæus, a Jacquin, a Thunberg, a
Willdenow, or a Curtis. It may perhaps be an acquisition to the science, that,
since such confusion prevails amongst “the most learned Botanists,” from
their “acknowledged inability to determine those plants;” which,
nevertheless, they have all foolishly attempted to do, we have one at last,
whose “scrutinizing” eye “has been able to make out all Linnæus’s and even
Thunberg’s species.” This elucidation, of so intricate a subject, by a person
whose knowledge of living plants, we fear, does not lead him, scarcely, to an
acquaintanceship with the difference of face in a Plane from a Poplar, must
be matter of infinite moment, to those, “who wish to become acquainted with
the plants they cultivate;” and the small trouble, to most persons, of learning
new, and ousting the old names for plants, which have been long rivetted to
the memory by habitual use, will be amply compensated, by the pleasure of
novelty, which must necessarily result, from the certain alteration in some
part of the title, of every plant which has hitherto, or is to come under, this
learned judge’s dictatorial fiat. Our opinions, as do our labours, run counter
to these new fashions, of rendering a difficult science easy; and our road
must still be in the old track of the trifler Linnæus.
 PLATE CXCIII.

GERANIUM LINEARE.

Linear-petalled Geranium.

CLASS XVI. ORDER IV.

MONADELPHIA DECANDRIA. Threads united. Ten Chives.

ESSENTIAL GENERIC CHARACTER.

Monogyna. Stigmata quinque.

Fructus rostratus, 5-coccus.
One Pointal. Five Summits.
Fruit furnished with long awns; five dry berries.
See Geranium Grandiflorum, Pl. XII. Vol. I.

SPECIFIC CHARACTER.

Geranium foliis lanceolatis, obtusis, subsinuatis; petalis subæqualibus,

linearibus; floribus pentandris; radice tuberosa.
Geranium with leaves lance-shaped, obtuse, and a little scolloped at the
edges; petals nearly equal, linear; flowers with five fertile chives; root
tuberous.

REFERENCE TO THE PLATE.

1. The Empalement cut open, to shew its hollow structure.

2. The Chives and Pointal natural size.
3. The Chives spread open, magnified.
4. The Pointal, magnified.
This is another of those curious tuberous Geraniums, which have been
introduced to this country, by Mr. Niven; who was sent to the Cape of Good
Hope by G. Hibbert, Esq. for the sole purpose of enriching his Gardens and
Herbarium, (now, we presume, the first in Europe) with the vegetable
productions of that country. It has no apparent difference, in habit, to require
any other treatment than has been mentioned in the former part of this work,
as necessary to the rest of its congeners. Our drawing was taken, from the
Clapham Collection, in July 1801; the roots having been received the
preceding autumn.
 PLATE CXCIV.

HEMEROCALLIS ALBA.

White Day-Lily.

CLASS VI. ORDER I.

HEXANDRIA MONOGYNIA. Six Chives. One Pointal.

ESSENTIAL GENERIC CHARACTER.

Corolla campanulata; tubo cylindrico. Stamina declinata.

Blossom bell-shaped; tube cylindrical. Chives declining.
See Hemerocallis cærulea, Pl. VI. Vol. I.

SPECIFIC CHARACTER.

Hemerocallis foliis cordatis, petiolatis; corolla alba, tubo longissimo.

Day-Lily with heart-shaped leaves that have foot-stalks; blossom white,
tube very long.

REFERENCE TO THE PLATE.

1. The Chives and Pointal, as they are placed in the flower.

2. The Seed-bud, Shaft, and Summit.
3. A ripe Seed-vessel of its natural size.
4. The Seed-vessel cut transversely, to shew the situation and
number of the cells and valves.
5. A ripe Seed, natural size.
The White Day-Lily is from the same country, and of the same date in our
gardens, as the Blue; figured in the First Vol. Pl. VI. and was introduced
through the same medium. It is herbaceous, and generally flowers, if kept in
the hot-house, about August; having that true and constant character of the
genus, and from which it had its name, the producing but one solitary,
perfect flower, per diem, till all the blossoms on the spike are exhausted;
which, in this species, are much more abundant than in any of the others. It
is increased by the root or seed.
This plant, we presume, is the same as those specified under the different
titles of Lilium Longiflorum, and L. Japonicum, in Willdenow’s new edition
of the Species Plant.; the L. candidum, and L. Japonicum, of Thunberg’s
Japan; and the L. Longiflorum of the Linn. Trans. Vol. II. P. 333. The
Hemerocallis formerly figured by us, under the specific title of Cærulea, we
take to be, the Hemerocallis Japonica, and Lilium Cordifolium of
Willdenow; the H. Cordata of Thunberg’s Japan; and the Hemerocallis
Japonica, and Lilium Cordifolium of the Linn. Trans. Vol. II. p. 332.
Willdenow, who had never seen even dried specimens of the plants in
question, has hence been led to place the same plant, under different genera;
and to consider the Hemerocallis of the Botanists Repository, as only a
variety of H. Japonica, but still admitting it as an Hemerocallis. Now, as we
have no doubt, (nor do we think any one can, that will take the trouble to
examine the dissections given with each figure,) that if the one is an
Hemerocallis, the other must be admitted of the same family; so, have we
made no scruple in rejecting the generic name of Lilium, for the present
plant; although we would gladly have adopted the specific one of Longiflora,
had not that of Alba, already obtained so generally in our gardens; a rule, for
our direction, paramount to all others as to specific denomination.