(Michael Farber) Invitation To Topological Robotic

Zurich Lectures in Advanced Mathematics
Edited by
Erwin Bolthausen (Managing Editor), Freddy Delbaen, Thomas Kappeler (Managing Editor), Christoph Schwab,
Michael Struwe, Gisbert Wstholz
Mathematics in Zurich has a long and distinguished tradition, in which the writing of lecture notes volumes
and research monographs play a prominent part. The Zurich Lectures in Advanced Mathematics series
aims to make some of these publications better known to a wider audience. The series has three main constituents: lecture notes on advanced topics given by internationally renowned experts, graduate text books
designed for the joint graduate program in Mathematics of the ETH and the University of Zurich, as well
as contributions from researchers in residence at the mathematics research institute, FIM-ETH. Moderately
priced, concise and lively in style, the volumes of this series will appeal to researchers and students alike,
who seek an informed introduction to important areas of current research.
Previously published in this series:
Yakov B. Pesin, Lectures on partial hyperbolicity and stable ergodicity
Sun-Yung Alice Chang, Non-linear Elliptic Equations in Conformal Geometry
Sergei B. Kuksin, Randomly forced nonlinear PDEs and statistical hydrodynamics in 2 space dimensions
Pavel Etingof, Calogero-Moser systems and representation theory
Guus Balkema and Paul Embrechts, High Risk Scenarios and Extremes A geometric approach
Demetrios Christodoulou, Mathematical Problems of General Relativity I
Camillo De Lellis, Rectifiable Sets, Densities and Tangent Measures
Published with the support of the Huber-Kudlich-Stiftung, Zrich
Michael Farber
Invitation to
Topological Robotics
Author:
Prof. Michael Farber
Department of Mathematical Sciences
Durham University
DH1 3LE Durham
United Kingdom
E-mail: [email protected]
2000 Mathematics Subject Classification (primary; secondary): 58E05, 57R70; 57R30

Key words: Betti numbers, configuration space, mechanical linkage, navigational complexity, polyhedra,
robotics
ISBN 978-3-03719-054-8
The Swiss National Library lists this publication in The Swiss Book, the Swiss national bibliography,
and the detailed bibliographic data are available on the Internet at http://www.helveticat.ch.
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in other ways, and storage in data banks. For any kind of use permission
of the copyright owner must be obtained.
2008 Michael Farber

Contact address:
European Mathematical Society Publishing House
Seminar for Applied Mathematics
ETH-Zentrum FLI C4
CH-8092 Zrich
Switzerland
Phone: +41 (0)44 632 34 36
Email: [email protected]
Homepage: www.ems-ph.org
Printed on acid-free paper produced from chlorine-free pulp. TCF
Printed in Germany
987654321
To Madeleine
Preface
Topological robotics is a new mathematical discipline studying topological problems inspired by robotics and engineering as well as problems of practical robotics requiring topological tools. It is a part of a
broader newly created research area called computational topology.
The latter studies topological problems appearing in computer science
and algorithmic problems in topology.
This book is based on a one-semester lecture course Topics of
Topological Robotics which I gave at the ETH Z
urich during April
June 2006. I describe here four selected mathematical stories which
have interesting connections to other sciences.
Chapter 1 studies conguration spaces of mechanical linkages, a
remarkable class of manifolds which appear in several elds of mathematics as well as in molecular biology and in statistical shape theory.
Methods of Morse theory, enriched with new techniques based on properties of involutions, allow eective computation of their Betti numbers.
We describe here a recent solution of the conjecture raised by Kevin
Walker in 1985. This conjecture asserts that the relative sizes of bars of
a linkage are determined, up to certain equivalence, by the cohomology
algebra of the linkage conguration space.
In Chapter 1 we also discuss topology of random linkages, a probabilistic approach to topological spaces depending on a large number of
random parameters.
In Chapter 2 we describe a beautiful theorem of Swiatoslaw R. Gal
[38] which gives a general formula for Euler characteristics of conguration spaces F (X, n) of of n distinct particles moving in a polyhedron X,
for all n. The Euler Gal power series is a rational function encoding
all numbers (F (X, n)) and Gals theorem gives an explicit expression
for it in terms of local topological properties of the space.
Chapter 3 deals with the knot theory of the robot arm, a variation of
the traditional knot theory question motivated by robotics. The main
result (which in my view is one of the remarkable jewels of modern
mathematics) is an unknotting theorem for planar robot arms, proven
recently by R. Connelly, E.Demaine and G. Rote [10].
viii
PREFACE
Chapter 4 discusses the notion of topological complexity of the robot motion planning problem TC(X) and mentions several new results
and techniques. The number TC(X) measures the complexity of the
problem of navigation in a topological space X viewed as the conguration space of a system. In this chapter we explain how one may use
stable cohomology operations to improve lower bounds on the topological complexity based on products of zero-divisors. These results were
obtained jointly with Mark Grant.
I would like to thank Jean-Claude Hausmann, Thomas Kappeler
and Dirk Schuetz for useful discussions of various parts of the book.
My warmest thanks go also to Mark Grant who helped me in many ways
to make this text readable and in particular for his advice concerning
the material of Chapter 4.
Problems of topological robotics can roughly be split into two main
categories: (A) studying special topological spaces, conguration spaces
of important mechanical systems; (B) studying new topological invariants of general topological spaces, invariants which are motivated and
inspired by applications in robotics and engineering. Class (A) includes
describing the topology of varieties of linkages, conguration spaces of
graphs, knot theory of the robot arm topics which are partly covered
below. The story about TC(X) represents a theory of class (B).
The book is intended as an appetizer and will introduce the reader
to many fascinating topological problems motivated by engineering.
Michael Farber
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Chapter
1.1.
1.2.
1.3.
1.4.
1.5.
1.6.
1.7.
1.8.
1.9.
1.10.
1.11.
1. Linkages and Polygon Spaces . . . . . . . . .

Conguration space of a linkage . . . . . . . . . .
The robot arm workspace map . . . . . . . . . . .
Varieties of quadrangles: n = 4 . . . . . . . . . .
Short, long and median subsets . . . . . . . . . .
The robot arm distance map . . . . . . . . . . . .
Poincare polynomials of planar polygon spaces
Morse theory on manifolds with involutions . .
Proof of Theorem 1.7 . . . . . . . . . . . . . . . . .
Maximum of the total Betti number of MW . . .
On the conjecture of Kevin Walker . . . . . . . .
Topology of random linkages . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
2
5
6
9
11
16
19
24
30
31
36
Chapter
2.1.
2.2.
2.3.
2.4.
2.5.
2.6.
2.7.
2.8.
2. Euler Characteristics of Conguration Spaces .

The Euler Gal power series . . . . . . . . . . . . . .
Conguration spaces of manifolds . . . . . . . . . . .
Conguration spaces of graphs . . . . . . . . . . . . .
Recurrent formula for euX (t) . . . . . . . . . . . . . .
Cut and paste surgery . . . . . . . . . . . . . . . . . .
Cut and paste Grothendieck ring . . . . . . . . . . .
Cones and cylinders . . . . . . . . . . . . . . . . . . . .
Proof of Theorem 2.3 . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
41
41
45
46
49
54
55
57
59
Chapter
3.1.
3.2.
3.3.
3.4.
3.5.
3.6.
3.7.
3. Knot Theory of the Robot Arm .

Can a robot arm be knotted? . . . . .
Expansive motions . . . . . . . . . . .
Innitesimal motions . . . . . . . . . .
Struts and equilibrium stresses . . . .
Maxwell Cremona Theorem . . . .
The main argument . . . . . . . . . . .
Global motion . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
61
61
64
67
68
72
75
77
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Chapter
4.1.
4.2.
4.3.
4.4.
4.5.
4.6.
4.7.
4.8.
CONTENTS
4. Navigational Complexity of Conguration Spaces

Motion planning algorithms . . . . . . . . . . . . . . . . .
The concept TC(X) . . . . . . . . . . . . . . . . . . . . . .
The notion of relative complexity . . . . . . . . . . . . .
Navigation functions . . . . . . . . . . . . . . . . . . . . .
TC(X), cohomology, and cohomology operations . . .
Simultaneous control of multiple objects . . . . . . . .
Collision-free motion planning . . . . . . . . . . . . . . .
Motion planning and the immersion problem . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
87
87
89
99
102
106
114
117
121
Recommendations for further reading . . . . . . . . . . . . . . . . . . 125

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
CHAPTER 1
Linkages and Polygon Spaces

In this chapter we study conguration spaces of mechanical linkages,
also known as polygon spaces. These spaces are quite important in
various engineering applications: in molecular biology they describe
varieties of molecular shapes, in robotics they appear as spaces of all
possible congurations of some mechanisms, and they play a central
role in statistical shape theory, see [63]. Mathematically, these spaces
are also very interesting: generically they are smooth closed manifolds,
however for some special values of parameters they have singularities.
Mathematical study of linkages and more general mechanisms has
a long history going back to the Middle Ages. Engineering discoveries
concerning linkages played an important role in the industrial revolution. Among the most famous are the pantograph, Watts linkage and
Peaucelliers inverser, see [51], [62], [16].
Topological theory of linkages was initiated by W. Thurston and
his students and collaborators, see for instance [95]. We want to mention also the thesis of S. H. Niemann [79] written in Oxford in 1978.
Kevin Walker [100] in his 1985 Princeton undergraduate thesis gives
an amazingly deep picture of conguration spaces of linkages. Some
results mentioned in [100] were not rigorously proven there, but subsequent work of other authors conrmed most of Walkers statements.
Further signicant progress in topology of linkages was made by J.Cl. Hausmann [47] and M. Kapovich and J. Millson [59]. Non-generic
polygon spaces were studied by A. Wenger [101] and the Japanese
school (see, e.g. [56]). An explicit expression for the Betti numbers of
conguration spaces of linkages in R3 were given by A. A. Klyachko
[65] who used methods of algebraic geometry. Later J.-Cl. Hausmann
and A. Knutson [48] applied methods of symplectic topology (toric
varieties) to compute the multiplicative structure of the cohomology
in the case of linkages in R3 . Historic comments concerning other
developments can be found in corresponding places in the text.
In this chapter we cover several specic topics of topology of linkages
without attempting to represent everything known. Our choice is based
purely on the authors personal preferences. Our rst goal is to explain
1. LINKAGES AND POLYGON SPACES
(following [28]) how one may compute Betti numbers of planar polygon
spaces. We also discuss classication results for these spaces in terms
of combinatorics of chambers and strata of their length vectors (the
Walker conjecture); here we briey describe results of [29]. Finally
we present a probabilistic approach to polygon spaces, which is very
eective in situations when one does not know the bar lengths and the
number of links n is large, n ; these results are described in [31]
in more detail.
The style of the exposition is not uniform and varies while we advance into the chapter. In the beginning it is very elementary, with
many simple examples, pictures and explanations. Sections 1.51.8
are written in the style of a research article containing theorems and
full proofs. In several concluding sections we adopt the survey style,
simply stating major results and referring to original articles for proofs
and further details.
A special role is played by section 1.7 on Morse theory of manifolds
with involution which can be read independently of the other material
of the chapter and may be applied in various contexts.
1.1. Conguration space of a linkage
Consider a simple planar mechanism consisting of n bars of xed lengths
l1 , . . . , ln connected by revolving joints forming a closed polygonal chain,
see Figure 1.1. The positions of two adjacent vertices are xed but the
other vertices are free to move so that angles between the bars change
but the lengths of the bars remain xed and the links are not disconnected from each other. Our task is to understand the topology of
l2
l1
l4
l3
l5
Figure 1.1. Linkage.

the conguration space of this mechanism. Recall that in general the
conguration space of a system S is dened as the space of all possible
states of S. The conguration space of the linkage will be denoted by
MW where
(1.1)
Z = (l1 , l2 , . . . , ln ) Rn+ ,
l1 > 0, . . . , ln > 0,
1.1. CONFIGURATION SPACE OF A LINKAGE
is the collection of the bar lengths, called the length vector of the linkage. Clearly, a conguration of the linkage is fully determined by angles
the bars make with the horizontal direction. Hence, the conguration
space MW can be identied with
1
(1.2) MW = {(u1 , . . . , un ) S S ;
n
=
li ui = 0, un = e1 }.
i=1
Here ui is +
the unit vector in the direction of the bar number i and the
condition ni=1 li ui = 0 expresses the property of the polygonal chain
to be closed. The equation un = e1 means that the last bar always
points in the direction opposite to the x-axis.
The topological space MW can also be understood as the moduli
space of planar n-gons with sides of length l1 , . . . , ln , viewed up to the
action of orientation-preserving isometries of the plane. It is important
to emphasize that our n-gons have cyclically oriented sides which are
labeled by integers 1, 2, . . . , n as shown on Figure 1.2. For any such planar n-gon there is a unique rotation of the plane bringing the polygon
in the position with a given side pointing in a xed direction. Hence
there is a one-to-one correspondence between the conguration space
of the mechanism shown on Figure 1.1 and the variety of all dierent
polygonal shapes of planar n-gons with sides of lengths l1 , . . . , ln . This
l5
l2
l3
l4
l1
Figure 1.2. Closed planar polygon.

explains why the space MW is also called the polygon space it parameterizes shapes of all planar n-gons with sides l1 , . . . , ln . Therefore
one can equivalently write
1
(1.3) MW = {(u1 , . . . , un ) S S ;
n
=
li ui = 0}/SO(2).
i=1
In (1.3) the group of rotations SO(2) acts on the vectors u1 , . . . , un

diagonally, i.e., a rotation R SO(2) acting on the n-tuple of unit
vectors (u1 , . . . , un ) produces (Ru1 , Ru2 , . . . , Run ).
Spaces MW appear also in the statistical shape theory, see [63], dealing with shapes of nite clouds of points viewed up to orientationpreserving Euclidean isometries. A nite set z1 , . . . , zn R2 has the
center c = (z1 + + zn )/n and is partly characterized by the distances
li = |zi c|. Let us assume that all li > 0, i.e., zi == c. Then
ui = (zi c)/li S 1 , i = 1, . . . , n
+
is a collection of unit vectors satisfying ni=1 li ui = 0. If z1( , . . . , zn( R2
is obtained from z1 , . . . , zn R2 by applying an orientation-preserving
planar isometry, then the corresponding set u(1 , . . . , u(n S 1 is obtained
from u1 , . . . , un S 1 by a global orientation-preserving rotation. This
explains why MW parameterizes all planar shapes with given distances
li from the central point.
Any permutation : {1, . . . , n} {1, . . . , n} denes a dieomorphism
: S 1 S 1 S 1 S 1 ,
given by
(u1 , . . . , un ) = (u(1) , . . . , u(n) ).
Clearly, maps the polygon space MW S 1 S 1 dieomorphically
onto the polygon space MW! S 1 S 1 where Z( = (l(1) , . . . , l(n) ).
Hence, we conclude that the order in which the numbers l1 , l2 , . . . , ln
appear in the length vector (1.1) is irrelevant for the dieomorphism
type of MW .
The case n = 3 is trivial. By the remark
above without loss of generality we may assume that l1 l2 l3 . If the triangle inequalP3
P2
ity l1 < l2 + l3 is satised, then MW consists of
two points two triangles with sides l1 , l2 , l3
P1
which are congruent to each other via a reection of the plane. If l1 = l2 + l3 , then MW
consists of a single point the degenerate triangle.
In the remaining case l1 > l2 + l3 the
P2
P3
variety MW is empty, MW = .
The following lemma describes all cases
when the variety MW is empty, n 3.
Lemma 1.1. For any n 3, the variety MW is empty if and only if one
of the links li is longer than the sum of all other links, i.e.,
(1.4)
li > l1 + + li1 + li+1 + + ln .
1.2. THE ROBOT ARM WORKSPACE MAP
Proof. It is obvious that (1.4) implies that MW = .

Below we assume that li l1 + + li1 + li+1 + + ln for all i
and prove that MW == . In the case n = 3 the statement is obvious.
We will argue by induction for n 4.
We claim that for n 4 there always exists a pair of adjacent1
edges li , li+1 such that
(1.5)
li + li+1 l1 + + li1 + li+2 + + ln .
Indeed, assuming the contrary, we obtain

2(li + li+1 ) > L = l1 + + ln ,
i = 1, . . . , n
and adding all these inequalities gives 4L > nL, a contradiction.

Now, if (1.5) holds, by the induction hypothesis there exists a closed
(n 1)-gon with sides l1 , . . . , li1 , li + li+1 , li+2 , . . . , ln ; it can be viewed
as an n-gon with collinear sides li , li+1 .
!
1.2. The robot arm workspace map
In this section we analyze the simplest robot arm having two links (see
Figure 1.3) with lengths l1 l2 , its conguration space C, work space
W and the associated workspace map : C E. A conguration
B
l2
l1
Figure 1.3. Robot arm with two bars.

of the arm is determined by two angles which make the bars with
the x-axis. Since these angles are independent, we obtain that the
conguration space C is the two-dimensional torus C = T 2 = S 1 S 1 .
The workspace is the variety of positions on the end point of the arm
G (the location of the gripper). It is easy to see that the work space
is a planar annulus having an external circle of radius R = l1 + l2 and
1The
notion adjacent is understood cyclically, i.e., l1 and ln are also adjacent.
C=T
Figure 1.4. Robot arm map : C = T 2 W .

an internal circle of radius r = l1 l2 . In the special case l1 = l2 the
internal circle degenerates into a point. The robot arm workspace map
(1.6)
:CW
associates with each conguration of the arm the corresponding position of the gripper, see Figure 1.4. Analytically is given by
(1.7)
(u1 , u2 ) = l1 u1 + l2 u2 ,
u1 , u2 S 1 .
Lemma 1.2. (a) The preimage 1 (w) of any internal point w of the
annulus W consists of exactly two congurations which are symmetric
to each other with respect to the line passing through the origin and w.
(b) If |w| = l1 + l2 , the preimage 1 (w) contains a single conguration with u1 = u2 .
(c) If |w| = l1 l2 > 0, the preimage 1 (w) contains a single
conguration with u1 = u2 .
(d) In the case l1 = l2 , the preimage 1 (0) equals the anti-diagonal
= {(u, u); u S 1 }.
(e) Let T 2 denote the diagonal = {(u, u); u S 1 }. The complement T 2 ( ) is a union of two connected components (which
we denote C+ and C ) and maps each of them dieomorphically onto
the domain l1 l2 < |w| < l1 + l2 .
Proof. To prove (e) we note that T 2 ( ) is the set of linearly
independent pairs of unit vectors u1 , u2 S 1 and C+ is dened as the
set of pairs u1 , u2 with det(u1 , u2 ) > 0 (the positive frames). The other
set C is dened as the set of negative frames.
Remaining statements of the lemma are obvious.
!
1.3. Varieties of quadrangles: n = 4
Next we use the method of J. Milgram and J. Trinkle [75] to understand
the variety MW in the case n = 4. This elementary discussion will give
1.3. VARIETIES OF QUADRANGLES: n = 4
us some experimental material which will be helpful in handling the

general case. We assume that
(1.8)
l1 l2 l3 l4 .
Our purpose is to study the variety of shapes of all quadrangles with

sides of lengths l1 , l2 , l3 , l4 . The side AD will stay xed and parallel to the x-axis. To understand the variety MW in this situation we
examine possible positions of the point C. Clearly,
C must lie in the work space W of the robot arm B
P2
with links l1 and l2 (see 1.2); on the other hand C
C
must be on the circle of radius l3 centered at D.
P1
Let us rst assume that l1 > l2 so that W is
P3
a planar ring with exterior radius R = l1 + l2 and
D
A
P4
interior radius r = l1 l2 > 0. Depending on the
values of l3 and l4 the circle of radius l4 with center
D (it is drawn in bold on Figure 1.5) may intersect the ring W in ve
dierent ways, see Figure 1.5.
A
(5,2,1,1)
B
(4,2,1,1)
C
(3,2,1,1)
D
(3,2,2,1)
E
(4,3,3,1)
Figure 1.5. Mutual positions of the annulus W and the

circle; each case is illustrated by a sample length vector.
Case A: Here we assume l3 +l4 < r. Then the bold circle lies entirely
inside the central part of the ring. The variety MW = is empty.
Case B: If l3 + l4 = r, the bold circle is tangent to the ring. The
variety MW consists in this case of a single point.
Case C: If l3 + l4 > r but l3 l4 < r, the bold circle intersects the
ring in an arc. By Lemma 1.2 each internal point of the ring represents
two dierent congurations of the linkage, each boundary point of the
arc represents a single conguration. Hence, the conguration space
MW is obtained from two copies of the arc by identifying their end point.
We see that MW = S 1 is homeomorphic to a circle.
Case D: The situation shown on Figure 1.5, D happens when l3 +
l4 > r and l3 l4 = r (the rst inequality is obviously superuous).
Now the bold circle is tangent to the ring. Each point of the bold circle
represents two dierent congurations except the tangency point where

these two congurations become identical. Hence the moduli space MW
is topologically S 1 S 1 , the wedge of two circles.
Case E: Finally, if l3 l4 > r, the conguration space MW is a disjoint
union of two circles S 1 5 S 1 .
Hence we see that, under the assumption l1 > l2 , the moduli space
MW can be
,
single point,
S 1,
S 1 S 1,
S1 5 S1
depending on the values of l1 , . . . , l4 . The cases B and D should be

viewed as special as they happen for certain resonance values of the
parameters and are characterized by equations: l1 = l2 +l3 +l4 (case B)
and l1 + l4 = l2 + l3 (case D). The remaining cases A, C, E are generic:
small perturbations of the length vector Z do not result in topological
changes for MW . We see that for a generic Z the variety MW is a closed
manifold of dimension 1. In cases B and D the variety has singular
points (i.e., points where it is not locally homeomorphic to R) and
these points correspond to collinear congurations of the linkage.
In the special case l1 = l2 the analysis is slightly dierent. The
ring W degenerates into a disc |w| R such that the center 0 W
represents a circle of dierent congurations 1 (0) = {(u, u); u
S 1 } of the robot arm l1 , l2 (see Lemma 1.2). The mutual position of
the workspace W and the circle with center D of radius l4 could be as
shown on Figure 1.6.
F
(3,3,2,1)
G
(2,2,1,1)
H
(1,1,1,1)
Figure 1.6. The workspace and the bold circle for l1 = l2 .

Case F: Here we assume that l1 = l2 > l3 > l4 . Clearly MW = S 1 5S 1 .
Case G: l1 = l2 > l3 = l4 . The image of the canonical map : MW
W is shown on Figure 1.6G it is a circle passing through the center.
The moduli space MW is shown on Figure 1.7, case G. MW contains the
circle 1 (0) = , see Figure 1.7G. The points a and b represent two
collinear congurations a = (1, 1, 1, 1), b = (1, 1, 1, 1). Each of
the arcs and connecting a and b is mapped onto the bold circle
shown on Figure 1.6G.
1.4. SHORT, LONG AND MEDIAN SUBSETS
Case G,
Case H,
Figure 1.7. The conguration space MW in cases F and G

Case H: l1 = l2 = l3 = l4 . The image of : MW W is shown
on Figure 1.6H, it is the circle passing through the center 0 W and
tangent to boundary W . In this case the conguration space MW is
the union of three circles (two preimages of the bold circle and the
preimage of the center) such that any two have a common point (this
statement requires a justication which is left as an exercise for the
reader). Compared with case G, we see that the middle points of the
arcs and (see Figure 1.7G) become identied, see point c shown on
Figure 1.7H.
1.4. Short, long and median subsets
A length vector Z = (l1 , l2 , . . . , ln ), li > 0 is called generic if
(1.9)
n
=
li ii == 0 for ii = 1.
i=1
Geometrically this can be expressed as follows. A length vector Z is

generic if the moduli space MW contains no collinear congurations; in
other words it is not possible to make a closed collinear n-gon with given
side lengths. For instance, the length vectors (2, 2, 1, 1) and (1, 1, 1, 1)
which were mentioned in the previous section are not generic. On the
other hand, length vector (3, 3, 2, 1) is generic. The set of generic length
vectors Z coincides with the complement in the positive quadrant li > 0
of the union of nitely many hyperplanes (1.9).
Theorem 1.3. If the length vector Z = (l1 , . . . , ln ) is generic, then the
moduli space of planar linkages MW is a closed orientable manifold of
dimension n 3.
A proof will be given later. If Z is not generic, then MW has singularities which will be described in the sequel.
10
We will call a subset J {1, . . . , n} short with respect to the length

vector Z if
=
=
li <
(1.10)
li .
iJ
iJ
/
A subset of a short set is short. As we shall see later, the set of all
short subsets with respect to Z fully determines the dieomorphism
type of MW . Hence one expects to be able to express all topological
characteristics of MW in terms of the family of all short subsets with
respect to Z.
A subset J {1, . . . , n} is called long with respect to Z if
=
=
li >
(1.11)
li ,
iJ
iJ
/
i.e., if its complement is short. By Lemma 1.4 the moduli space of

planar linkages MW is empty if and only if there exists a one-element
long subset J = {i} {1, . . . , n}. There exist no disjoint long subsets.
J {1, 2, . . . , n} is a median subset with respect to Z if
=
=
(1.12)
li =
li .
iJ
iJ
/
The complement of a median subset is also a median subset. Clearly,

median subsets exist only when the vector Z is not generic. A subset
J {1, . . . , n} is median if and only if neither J nor its complement
J are short. Hence, the family of short subsets determines the families
of median and long subsets.
As an illustration we give the following table describing families of
short subsets for n = 4 (varieties of quadrangles) in all cases AH (see
the previous section) under the assumption (1.8).
Case
Family of short subsets
MW
A
J {2, 3, 4}
B
J " {2, 3, 4}
single point
C
J = {1} or J " {2, 3, 4}
S1
D
|J| 1 or J = {2, 4} or J = {3, 4}
S1 S1
E & F |J| 2 and J == {1, 2}, {1, 3}, {2, 3}
S1 5 S1
G
|J| 1 or J = {3, 4}
Figure 1.7 left
H
|J| 1
Figure 1.7 right
To explain our notations, let us mention that in Case B the short
subsets are all subsets of {2, 3, 4} except J = {2, 3, 4}.
1.5. THE ROBOT ARM DISTANCE MAP
11
Generic cases are A, C, E & F and in these cases the conguration

space MW is a one-dimensional manifold conrming Theorem 1.3.
The reader is invited to compare these results with another approach to classifying the quadrangle spaces described in [49], Table 2,
page 39.
1.5. The robot arm distance map
A robot arm is a simple mechanism consisting of n bars (links) of xed
length (l1 , . . . , ln ) connected by revolving joints as shown on Figure 1.8.
The initial point of the robot arm is xed on the plane. The moduli
l3
l2
l4
l1
Figure 1.8. Robot arm with n links.

space of the robot arm (dened as the space of possible shapes) is
(1.13)
W = {(u1 , . . . , un ) S 1 S 1 }/SO(2).
Clearly, W is dieomorphic to torus T n1 of dimension n 1. A dieomorphism can be specied, for example, by assigning to a conguration
1
1
n1
(u1 , . . . , un ) the point (1, u2 u1
(measur1 , u3 u1 , . . . , un1 u1 ) T
ing angles between the directions of the rst and the other links).
The moduli space of polygons MW (where Z = (l1 , . . . , ln )) is naturally embedded into W .
MP
0
fP
Figure 1.9. Function fW : W R.
12
Dene a real-valued function on W as follows:

(1.14)
fW : W R,
$2
$
n
$
$=
$
$
l i ui $ .
fW (u1 , . . . , un ) = $
$
$
i=1
Geometrically the value of fW equals the negative of the squared distance

between the initial point of the robot arm to the end of the arm shown
by the dotted line on Figure 1.8. Note that the maximum of fW is
achieved on the moduli space of planar linkages MW W .
An important role is played by the collinear congurations of the
robot arm, i.e., such that ui = uj for all i, j, see Figure 1.10. We will
label such congurations by long and median subsets J {1, . . . , n},
assigning to a long or median subset J the conguration pJ W given
by pJ = (u1 , . . . , un ) where ui = 1 for i J and ui = 1 for i
/ J. Note
i J
/J
i
Figure 1.10. A collinear conguration pJ of the robot arm.

that the congurations pJ , pJ W are identical where J {1, . . . , n}
is a median subset and J denotes its complement. Note also that pJ
lies in MW if and only if J is median.
Lemma 1.4. The critical points of fW : W R lying in W MW are
exactly the collinear congurations pJ corresponding to long subsets J
{1, . . . , n}. Each pJ , viewed as a critical point of fW , is nondegenerate
in the sense of Morse and its Morse index equals n |J|.
Proof. View fW as a function on the torus T n with the angular coordinates (u1 , . . . , un ), where ui S 1 . The
moduli space W is the factorn
space W = T /SO(2). Write ui = exp( 1 i ), where 0 i < 2.
Then
fW (u1 , . . . , un ) = (
n
=
li cos i )2 (
i=1
n
=
i=1
li2 2
=
i<j
n
=
li sin i )2
i=1
li lj cos(i j ).
13
We nd
=
fW
lj sin(j i ).
= 2li
i
j=1
n
(1.15)
The equation
(1.16)
f,
i
= 0 gives
sin i
n
=
lj cos j = cos i
n
=
lj sin j .
j=1
j=1
We conclude that either

n
=
lj cos j = 0 =
j=1
n
=
lj sin j ,
j=1
or tan i is independent of i. The rst possibility happens if and only

if the tuple (u1 , . . . , un ) T n represents a closed polygon, i.e., a conguration lying in MW . If the second possibility happens, then for any
pair of indices i and j either i = j or i = j . This shows that
any critical point of function fW lying in W MW is represented by a
collinear conguration. The inverse statement is obvious from (1.15).
Let J {1, 2, . . . , n} be a long subset. Consider the corresponding
critical point pJ = (u1 , . . . , un ) where ui = 1 for i J and ui = 1 for
i
/ J. Then
(1.17)
LJ =
n
=
li ui > 0.
i=1
We will calculate the Hessian of fW at point pJ . Note that fW (pJ ) =

(LJ )2 . Using (1.15) we nd
+
li k1=i lk cos(i k ), if i = j,
1 2 fW
=
(1.18)
2 i j
li lj cos(j i ),
if i =
= j.
Since cos(i j ) = ui uj , we may rewrite (1.18) in the form
li2 (di 1) if i = j,
2
1 fW
(1.19)
(pJ ) =
l l u u if i == j,
2 i j
i j i j
where
di =
ui L J
.
li
14
This shows that the Hessian of fW at pJ is congruent to the matrix DE,

where E is an n n-matrix with all entries 1 and D is a diagonal with
diagonal entries di . Note that di > 0 for i J and di < 0 for i
/ J.
We are going to calculate the signature of the Hessian D E using
the Sylvester criterion and Lemma 1.5. Let us reorder the set {1, . . . , n}
such that the indices of J follow the indices which are not in J; in other
words we assume that J = {k, k + 1, . . . , n}.
By Lemma 1.5 (see below) the determinant of the principal minor
of size r of D E equals
r =
(1.20)
r
9
dj (1
j=1
r
=
d1
).
=1
Let us show that the number in the brackets in formula (1.20) is always
non-negative and it vanishes only for r = n. Indeed, if r < k then
dj < 0 for 1 j r, and the statement is obvious. For r k we have
(1.21)
r
=
d1
L1
J (LJ
r
=
l i ui ) =
n
=
li ui 0,
i=r+1
i=1
=1
L1
J
as ui = +1 for i J; the vanishing in (1.21) happens only for r = n,

as is clear from (1.21).
This shows that the sign of the minor r is given by
sign(r ) =
()r
for 1 r < k,
()k1 for k r < n
and n = 0. We see that the number of negative eigenvalues of D E

equals n |J|, the number of positive eigenvalues is |J| 1, and there
is a unique zero eigenvalue.
The function fW viewed as a function on the torus T n is SO(2)invariant. Each pJ is represented by a circle in T n and the Hessian
vanishes in the direction tangent to the circle. It follows, that fW :
W R is nondegenerate at pJ and its Morse index equals n |J|.
!
Lemma 1.5. Let A be an n n-matrix of the form
(1.22)
A = D E,
15
where D = diag(d1 , d2 , . . . , dn ) is an (n n)-diagonal matrix, and E is

an (n n)-matrix with all entries 1. Then
&
8
n
n
=
9
1
di 1
det A =
.
d
i=1 i
i=1
Proof. In general, if
+X and Y are two square matrices, then det(X +
det(R ), where the matrix R is obtained by
Y ) equals the sum
replacing some columns of X by the corresponding columns of Y ; the
sum contains 2n terms. Since all columns of E are equal, replacing
more than one column of D by the columns of E gives a matrix with
zero determinant. This implies our statement.
!
Now we are able to prove Theorem 1.3:
Proof of Theorem 1.3. Consider the robot arm with n 1 links
having the length vector Z( = (l1 , . . . , ln1 ). Its moduli space W 9
T n2 is dieomorphic to the torus T n2 , see 1.5. The function fW! :
W R (given by (1.14)) has only Morse critical points in W MW! =
fW1
! ((, 0)). By Lemma 1.4 the critical values of fW! are of the form
ln2 where ln > 0 is such that the vector
Z = (l1 , . . . , ln1 , ln )
is not generic. One has
(1.23)
2
MW = fW1
! ((ln ) ).
This implies that MW is a closed manifold of dimension n 3.
The following theorem describes singularities of MW .

Theorem 1.6. Assume that the length vector Z = (l1 , . . . , ln ) is not
generic. Then MW is a compact (n 3)-dimensional manifold having
nitely many singular points which are in one-to-one correspondence
with collinear congurations pJ where J {1, . . . , n} is a median subset2. Near each conguration pJ the moduli space MW is homeomorphic
to the cone over the product of two spheres
S |J|2 S n|J|2 .
2Recall
that pJ = pJ M% for a median subset J {1, . . . , n}.
16
Proof. The proof is similar to that of Theorem 1.3. Consider the

moduli space W = T n2 of the robot arm with links l1 , . . . , ln1 and
the function fW! : MW! R where Z( = (l1 , . . . , ln1 ). Let J {1, . . . , n}
be a median subset such that n
/ J. Then J can be viewed as a long
subset of {1, . . . , n 1}. The corresponding collinear conguration
qJ W is a critical point of fW! having the Morse index n 1 |J| (by
Lemma 1.4). The value fW! (qJ ) equals ln2 . We obtain that locally the
n2|J|
2
S |J|2
preimage fW1
! (ln ) = MW is a cone over the product S
as claimed.
!
1.6. Poincar
e polynomials of planar polygon spaces
The next theorem gives a general formula for the Poincare polynomials
of moduli spaces MW . It describes the Betti numbers of all varieties
MW including the cases when MW has singularities (i.e., when the length
vector Z is not generic).
Theorem 1.7. For a given length vector Z = (l1 , . . . , ln ), x a link of
the maximal length li , i.e., such that li lj for any j = 1, 2, . . . , n.
For k = 0, 1, . . . , n 3, denote by ak the number of short subsets of
{1, . . . , n} of cardinality k + 1 which contain i, and by bk the number
of median subsets of {1, . . . , n} of cardinality k + 1 containing i. The
homology group Hk (MW ; Z) is free abelian of rank
(1.24)
ak + an3k + bk ,
for any k = 0, 1, . . . , n 3.
By Theorem 1.7 the Poincare polynomial
p(t) =
n3
=
dim Hk (MW ; Q) tk
k=0
of MW can be written in the form

(1.25)
q(t) + tn3 q(t1 ) + r(t)
where
(1.26)
q(t) =
n3
=
k=0
ak t ,
r(t) =
n3
=
bk t k ;
k=0
the numbers ak and bk are described in the statement of Theorem 1.7.

Clearly bk = 0 for all k, assuming that the length vector Z = (l1 , . . . , ln )
is generic.
POLYNOMIALS OF PLANAR POLYGON SPACES

1.6. POINCARE
17
A proof of Theorem 1.7 is given below in 1.8. It is based on the

technique of Morse theory in the presence of involution developed in
section 1.7.
Next we illustrate the statement of Theorem 1.7 by several examples.
Example 1.8. Suppose that Z = (3, 2, 2, 1). Then a0 = 1 and b1 = 1
and all other numbers ai and bi vanish. We obtain that the Poincare
polynomial of MW is 1 + 2t. This is consistent with the fact that MW =
S 1 S 1 , as established earlier.
Example 1.9. Assume that Z = (3, 2, 1, 1). Then a0 = 1 and all other
numbers ai and bi vanish. We obtain that the Poincare polynomial of
MW is 1 + t. It is consistent with our earlier result MW = S 1 (see 1.4,
Case C).
Example 1.10. Suppose Z = (4, 3, 3, 1). Then a0 = 1, a1 = 1 and all
other numbers ai and bi vanish. We obtain that the Poincare polynomial of MW is 2(1 + t). We know that MW = S 1 5 S 1 , see table in 1.4,
Case E.
Example 1.11. Suppose that n = 5 and l1 = 3, l2 = 2, l3 = 2, l4 = 1,
l5 = 1. Then l1 = 3 is the longest link and short subsets of {1, . . . , 5}
containing 1 are {1}, {1, 4} and {1, 5}. Hence a0 = 1, a1 = 2 and by
Theorem 1.7 the Poincare polynomial of MW equals 1 + 4t + t2 . We
conclude that MW is a closed orientable surface of genus 2.
Example 1.12. Consider the zero-dimensional Betti number a0 +an3 +
b0 of MW as given by Theorem 1.7. Lets assume again that
l1 l2 ln .
If the one-element set {1} is long, then ak = 0 and bk = 0 for all k and
MW = . If {1} is median, then ak = 0 for all k and b0 = 1 while bk = 0
for k > 0; the moduli space MW consists of a single point.
We assume below that {1} is short; then all one-element subsets of
{1, . . . , n} are short and MW == (see Lemma 1.1). We obtain in this
case that a0 = 1. Let us show that the number an3 equals 0 or 1.
By the denition, an3 coincides with the number of long two-element
subsets {r, s} {1, . . . , n} not containing 1. There may exist at most
one such pair: if {r( , s( } is another long pair with r == r( , r == s( ,
18
then {1, r} and {r( , s( } would be two disjoint long subsets which is
impossible.
Example 1.13. As another example consider the equilateral case Z =
(1, 1, . . . , 1) with n = 2r + 1 odd. It is clearly generic and hence bk = 0.
We may x the rst link l1 as being the longest. The short subsets in
this case are subsets of {1, . . . , n} of cardinality r. Hence we nd
that the number ak equals
.n1(
for k r 1,
k
(1.27)
ak =
0,
for k r.
By Theorem 1.7 the Betti numbers of MW are given by
.n1(
for k < r 1,
. (
for k = r 1,
2 n1
bk (MW ) =
(1.28)
r1
.n1(
for k > r 1.
k+2
This coincides with the result of Theorem 1.1 from [58], see also Theorem C in [57].
Example 1.14. Consider now the equilateral case Z = (1, 1, . . . , 1) with
n = 2r + 2 even. We x l1 = 1 as the longest link. The short subsets
are all subsets of cardinality r and the median subsets are all subsets
of cardinality r + 1. Hence. we (nd that ak is given by formula (1.27),
bk = 0 for k =
= r and br = 2r+1
. Applying Theorem 1.7 we obtain
r
.n1(
for k r 1,
. (
n
for k = r,
bk (MW ) =
(1.29)
r
.n1(
for r + 1 k n 3.
k+2
This is consistent with the result of Theorem 1.1 from [58].
We will use later the observation (following from the last two examples) that the total Betti number
n3
=
i=0
bi (M(1,1,...,1) )
1.7. MORSE THEORY ON MANIFOLDS WITH INVOLUTIONS
19
for the equilateral linkage with n links equals

<
;
n1
n1
(1.30)
2
r
where r = [ (n1)
].
2
1.7. Morse theory on manifolds with involutions
Our main tool in computing the Betti numbers of the moduli space of
planar polygons MW is Morse theory of manifolds with involution.
Theorem 1.15. Let M be a smooth compact manifold with boundary.
Assume that M is equipped with a Morse function f : M [0, 1]
and with a smooth involution : M M satisfying the following
properties:
(1) f is -invariant, i.e., f ( x) = f (x) for any x M ;
(2) The critical points of f coincide with the xed points of the
involution;
(3) f 1 (1) = M and 1 [0, 1] is a regular value of f .
Then each homology group Hi (M ; Z) is free abelian of rank equal to
the number of critical points of f having Morse index i. Moreover, the
induced map
: Hi (M ; Z) Hi (M ; Z)
coincides with multiplication by (1)i for any i.
As an illustration for Theorem 1.15 consider a surface in R3 (see Figure 1.11) which is symmetric with respect to the z-axis. The function f
is the orthogonal projection onto the z-axis, the involution : M M
is given by (x, y, z) = (x, y, z). The critical points of f are exactly
z
Figure 1.11. Surface in R3 .

the intersection points of M with the z-axis.
20
Proof of Theorem 1.15. Choose a Riemannian metric on M which

is invariant with respect to .
Let p M be a critical point of f . By our assumption, p must be
a xed point of , i.e., (p) = p. We claim that the dierential of at
p is multiplication by 1, i.e.,
(1.31)
dp (v) = v,
for any v Tp M.
Firstly, since is an involution, dp must have eigenvalues 1. Assume

that there exists a vector v Tp M with dp (v) = v. Then the geodesic
curve starting from p in the direction of v is invariant with respect
to , implying that p is not isolated in the xed point set of . This
contradicts our assumption and hence dp must have eigenvalue 1
only. Note that dp is diagonalizable as dp preserves the Hessian of f
at p,
(1.32)
H(f )p : Tp (M ) Tp (M ) R
which is a nondegenerate quadratic form. This proves (1.31).

Consider the gradient vector eld v of f with respect to the Riemannian metric. We will assume that v satises the transversality
condition, i.e., all stable and unstable manifolds of the critical points
intersect transversally. v is -invariant which means that
(1.33)
v (x) = dx (vx ),
x M.
The Morse Smale chain complex (C (f ), ) of f has the critical

points of f as its basis and the dierential is given by
=
(1.34)
(p) =
[p : q] q
q
where in the summation is taken over the critical points q with Morse
index ind(q) = ind(p) 1. The incidence numbers [p : q] Z are
dened as follows
=
(1.35)
[p : q] =
i(), i() = 1,
where : (, ) M are trajectories of the negative gradient ow

( (t) = v(t) satisfying the boundary conditions (t) p as t
and (t) q as t +.
Observe that if is a trajectory as above, then is another such
trajectory. Indeed, using (1.33) we nd ( )( = d ( ( ) = d (v(t) ) =
v ((t)) .
21
Figure 1.12. Two symmetric trajectories of the negative gradient ow.

Theorem 1.15 would follow once we show that
(1.36)
i() + i( ) = 0,
i.e., the total contribution to (1.35) of a pair of symmetric trajectories

is zero. Hence all incidence coecients vanish [p : q] = 0 and the
dierentials of the Morse Smale complex are trivial.
To prove (1.36) we rst recall the denition of the sign i()
{1, 1}, see [77]. For a critical point p of f we denote by W u (p) and
W s (p) the unstable and stable manifolds of p. Recall that W u (p) is the
union of the trajectories : (, ) M satisfying the dierential
equation ( (t) = v(t) and the boundary condition (t) p as t
. The stable manifold W s (p) is dened similarly but the boundary
condition in this case becomes (t) p as t +.
Fix an orientation of the stable manifold W s (p) for every critical
point p M . Since W s (p) and W u (p) are of complementary dimension
and intersect transversally at p, the orientation of W s (p) determines a
coorientation of the unstable manifold W u (p), for every p.
If ind(p) ind(q) = 1, then W u (p) and W s (q) intersect transversally
along nitely many connecting orbits (t) and the structure near each
of the connecting orbits looks as shown on Figure 1.13. Note that the
normal bundle to W u (p) along coincides with the normal bundle to
in W s (q). Hence, the coorientation of W u (p) together with the natural
orientation of the curve (t) determine an orientation of W s (q) along
. We set i() = 1 i this orientation coincides with the prescribed
orientation of W s (q); otherwise we set i() = 1.
To compare i() with i( ) we rst observe that the involution
preserves the stable and unstable manifolds W s (p) and W u (p) and
for every critical point p the degrees of the restriction of on these
submanifolds equal
(1.37) deg( |W u (p) ) = (1)ind(p) ,
deg( |W s (p) ) = (1)nind(p) ,
22
W (p)
(t)
W (q)
Figure 1.13. The stable and unstable manifolds along (t).

as follows from (1.31). Hence, applying the involution to the picture
shown on Figure 1.13, we have to multiply the coorientation of W u (p)
by (1)nind(p) and multiply the orientation of W s (q) by (1)nind(q) .
As a result the total sign will be multiplied by
(1)nind(p) (1)nind(q) = (1)ind(p)ind(q) = 1.
This proves (1.36) and completes the proof of the rst statement of the
theorem. The second statement of the theorem follows from the rst
one combined with (1.37).
!
Theorem 1.16. Let M be a smooth compact connected manifold with
boundary. Suppose that M is equipped with a Morse function f : M
[0, 1] and with a smooth involution : M M satisfying the properties
of Theorem 1.15. Assume that for any critical point p M of the
function f we are given a smooth closed connected submanifold
Xp M
with the following properties:
(1) Xp is -invariant, i.e., (Xp ) = Xp ;
(2) p Xp and for any x Xp {p}, one has f (x) < f (p);
(3) the function f |Xp is Morse and the critical points of the restriction f |Xp coincide with the xed points of lying in Xp .
(4) For any xed point q Xp of the Morse indexes of f and of
f |Xp at q coincide. In particular,
dim Xp = ind(p).
Then each submanifold Xp is orientable and the set of homology classes
realized by {Xp }pFix( ) forms a free basis of the integral homology group
23
H (M ; Z). In other words, we claim that the inclusion induces an

isomorphism
7
(1.38)
Hi (Xp ; Z) Hi (M ; Z)
dim Xp =i
for any i.
Proof of Theorem 1.16. First we note that each submanifold Xp
is orientable. Indeed, Theorem 1.15 applied to the restriction f |Xp
implies that f |Xp has a unique maximum and unique minimum3 and
the top homology group Hi (Xp ; Z) = Z is innite cyclic where i =
dim Xp = ind(p).
For a regular value a R of f we denote by M a M the preimage
f 1 (, a]. It is a compact manifold with boundary. It follows from
Theorem 1.15 that f has a unique local minimum and therefore M a
is either empty or connected. For a slightly above the minimum value
f (p0 ) = min f (M ) the manifold M a is a disc and the homology of M a
is obviously realized by the submanifold Xp0 = {p0 } M a .
We proceed by induction on a. Our inductive statement is that
the homology of M a is freely generated by the homology classes of the
submanifolds Xp where p runs over all critical points of f satisfying
f (p) a.
Suppose that the statement is true for a and the interval [a, b] contains a single critical value c. Let p1 , . . . , pr be the critical points of f
lying in f 1 (c). Set
r
/
X=
Xp i
i=1
(the disjoint union). Then f induces a Morse function g : X R and

we set
X a = g 1 (, a].
Consider the Morse Smale complexes C (M a ), C (M b ), C (X) and
C (X a ); the rst two are constructed using the function f and the
latter two are constructed using the function g. We have the following
Mayer Vietoris-type short exact sequence of chain complexes
(1.39)
3since
0 C (X a ) C (X) C (M a ) C (M b ) 0
we assume that M and Xp are connected.
24
which (by the arguments indicated in the proof of Theorem 1.15) have
trivial dierentials and hence the sequence
(1.40)
0 Hi (X a ) Hi (X) Hi (M a ) Hi (M b ) 0
is exact (all homology groups have coecients Z). It follows from

Lemma 1.17 below and the construction of the Morse Smale complex
(compare [77], 7) that the homomorphism (which appears in (1.39)
and (1.40)) coincides with the sum of the homomorphisms induced by
the inclusions X M b and M a M b .
For i < dim Xpk we have Hi (Xpak ) Hi (Xpk ) is an isomorphism
(by Theorem 1.15). For i dim Xpk we have Hi (Xpak ) = 0. Hence we
obtain an isomorphism induced by the inclusions
7
(1.41)
Hi (Xpk ) Hi (M a ) Hi (M b ).
indpk =i
This shows that Hi (M b ) is freely generated by the homology classes

of the submanifolds Xp satisfying f (p) < b and dim Xp = i. This
completes the step of induction.
!
Here is a minor variation of the Morse lemma which has been used
in the proof.
Lemma 1.17. Let f : Rn R be a smooth function having 0 Rn
as a nondegenerate critical point and suppose that for some k n the
restriction f |Rk {0} : Rk {0} R also has a nondegenerate critical
point at 0 Rk . Then there exists a neighborhood U Rn of 0 and
a local coordinate system x : U Rn such that x(Rk {0} U )
Rk {0} and
(1.42)
f (x1 , . . . , xn ) = x21 + + x2n + f (0).
Proof. One simply checks that the coordinate changes in the standard
proof of the Morse lemma (compare [76], 2) can be chosen so that the
subspace Rk {0} is mapped to itself.
!
1.8. Proof of Theorem 1.7
Consider the moduli space W of the robot arm (dened by (1.13)) with
the function fW : W R (dened by (1.14)). There is an involution
(1.43)
:W W
1.8. PROOF OF THEOREM 1.7
25
given by
(1.44)
(u1 , . . . , un ) = (
u1 , . . . , un ).
Here the bar denotes complex conjugation, i.e., the reection with respect to the real axis. It is obvious that formula (1.44) maps SO(2)orbits into SO(2)-orbits and hence denes an involution on W . The
xed points of are the collinear congurations of the robot arm, i.e.,
the critical points of fW in W MW , see Lemma 1.4. Our plan is to
apply Theorems 1.15 and 1.16 to the sublevel sets
(1.45)
W a = fW1 (, a]
of fW . Recall that the values of fW are non-positive and the maximum

is achieved on the submanifold MW W . From Lemma 1.4 we know
that the critical points of fW are the collinear congurations pJ . The
latter are labelled by long subsets J {1, . . . , n} and pJ = (u1 , . . . , un )
where ui = 1 for i J and ui = 1 for i
/ J. One has
(1.46)
fW (pJ ) = (LJ )2 .
(1.47)
(LJ )2 < a < 0
+
Here LJ = ni=1 li ui with pJ = (u1 , . . . , un ).
The number a which appears in (1.45) will be chosen so that
for any long subset J such that the manifold W a contains all the critical
points pJ . The situation is shown schematically on Figure 1.14.
MP
a
fP
Wa
Figure 1.14. Function fW : W R and the manifold W a .

For each subset J {1, . . . , n} we denote by ZJ the length vector
obtained from Z = (l1 , . . . , ln ) by integrating all links li with i J into
one link. For example, if J = {1, 2} then ZJ = (l1 + l2 , l3 , . . . , ln ). We
denote by WJ the moduli space of the robot arm with the length vector
ZJ . It is obvious that WJ is dieomorphic to a torus T n|J| . We view
WJ as being naturally embedded into W . Note that the submanifold
26
WJ W is disjoint from MW (in other words, WJ contains no closed

congurations) if and only if the subset J {1, . . . , n} is long.
Lemma 1.18. Let J {1, . . . , } be a long subset. The submanifold
WJ W has the following properties:
(1) WJ is invariant with respect to the involution : W W ;
(2) the restriction of fW onto WJ is a Morse function having as
its critical points the collinear congurations pI where I runs
over all subsets I {1, . . . , n} containing J;
(3) for any such pI the Morse indexes of fW and of fW |WJ at pI
coincide;
(4) in particular, f |WJ achieves its maximum at pJ WJ .
Proof. (1) is obvious. Statements (2) and (3) follow from Lemma 1.4
applied to the restriction of fW onto WJ . Here we use the assumption
that J is long. Under this assumption the long subset for the integrated
length vector ZJ are in one-to-one correspondence with the long subsets
I {1, . . . , n} containing J. Statement (4) follows from (3) as the
Morse index of fW |WJ at point pJ equals n |J| = dim WJ .
!
Applying Theorems 1.15 and 1.16 and taking into account Lemma
1.18 we obtain:
Corollary 1.19. One has:
(1) If a satises (1.47) then the manifold W a (see (1.45)) contains
all submanifolds WJ where J {1, . . . , n} is an arbitrary long
subset.
(2) The homology classes of the submanifolds WJ form a free basis
of the integral homology group H (W a ; Z).
Next we examine the homomorphism
(1.48)
: Hi (W a ; Z) Hi (W ; Z)
induced by the inclusion : W a W .

Below we will assume that l1 lj for all j {1, . . . , n}, i.e., l1 is
the longest link. This may always be achieved by relabelling.
We describe a specic basis of the homology H (W ; Z). For any
subset J {1, 2, . . . , n} we denote by WJ the moduli space of congurations of the robot arm with length vector ZJ where all links li
with i J are integrated into a single link. Note that WJ is naturally
embedded into W and
W J MW =
27
if and only if the set J is long. Since W is homeomorphic to the torus

T n1 , it is easy to see that a basis of the homology group H (W ; Z)
is formed by the homology classes of the submanifolds WJ where J
{1, . . . , n} runs over all subsets containing 1. We will denote the homology class of WJ by
(1.49)
[WJ ] Hn|J| (W ; Z).
Assuming that J, J ( {1, . . . , n} are two subsets with |J| + |J ( | =

n + 1, the classes [WJ ] and [WJ ! ] have complementary dimensions in
W and their intersection number is given by
(1.50)
[WJ ] [WJ ! ] =
1, if |J J ( | = 1,
0,
if |J J ( | > 1.
Indeed, if J J ( = {i0 } then WJ WJ ! consists of a single point {p},

the moduli space of a robot arm with all links integrated into one
link. Let us show that the intersection WJ WJ ! is transversal. A
tangent vector to W at p = (u1 , . . . , un ) can be labelled by a vector
w = (1 , . . . , n ) Rn (an element of the Lie algebra of the torus T n )
viewed up to adding vectors of the form (, , . . . , ). Such a tangent
vector w is tangent to the submanifold WJ i i = j for all i, j J.
Given w as above it can be written as
w = w( + w(( + (i0 , . . . , i0 )
where w( has coordinates 0 on places i J and coordinates i i0 on
places i
= J; coordinates of w(( vanish on places i
/ J and are i i0
on places i J. Hence every tangent vector to W is a sum of a tangent
vector to WJ and a tangent vector to WJ ! .
Now suppose that |J J ( | > 1. We will show that then the submanifold WJ ! can be continuously deformed inside W to a submanifold
WJ( ! such that WJ WJ( ! = . This would prove the second claim in
(1.50). Let us assume that {1, 2} J J ( . Dene gt : WJ ! W by
gt (u1 , . . . , un ) = (eit u1 , u2 , . . . , un ),
t [0, 1].
Here satises 0 < < . Then WJ( ! = g1 (WJ ! ) is clearly disjoint from
WJ ; indeed, the links l1 and l2 are parallel in WJ and make an angle
in WJ( ! .
28
It follows that the intersection form in the basis [WJ ], [WJ ! ]

H (W ; Z), where J 2 1, J ( 2 1, has a very simple form:
[WJ ] [WJ ! ] =
(1.51)
1, if J J ( = {1},
0,
if |J J ( | > 1.
In particular, given [WJ ] with 1 J, its dual homology class lying

in H (W ; Z) equals [WK ] where K = CJ {1}; here CJ denotes the
complement of J in {1, . . . , n}.
Denote by A H (W a ; Z) (correspondingly, B H (W a ; Z)) the
subgroup generated by the homology classes [WJ ] where J {1, . . . , n}
is long and contains 1 (correspondingly, J is long and 1
/ J). Then
Hi (W a ; Z) = Ai Bi .
(1.52)
Similarly, one has
(1.53)
Hi (W ; Z) = Ai Ci Di ,
where:
A is as above;
C H (W ; Z) is the subgroup generated by the homology
classes [WJ ] with J {1, . . . , n} short and 1 J;
D is the subgroup generated by the classes [WJ ] H (W ; Z)
where J is median and contains 1.
It is clear that (see (1.48)) is identical when restricted to Ai ,
compare (1.52) and (1.53). We claim that the image (Bi ) is contained
in Ai . This would follow once we show that
(1.54)
[WJ ] [WK ] = 0
assuming that [WJ ] Bi and [WK ] is the dual of a class [WJ ! ] Ci or

[WJ ! ] Di , see (1.51). We have
(1)
(2)
(3)
(4)
J is long and 1
/ J,
J ( is short or median and 1 J ( ,
|J| = |J ( |,
K = CJ ( {1}.
Here CJ ( denotes the complement of J ( in {1, . . . , n}. By (1.50), to

prove (1.54) we have to show that under the above conditions one has
29
|J K| > 1. Indeed, suppose that |J K| = 1, i.e., J K = {j},

a single element subset. Then J ( is obtained from J by removing the
index j and adding the index 1 which leads to a contradiction: indeed,
J is long, lj l1 and J ( is either short or median.
Corollary 1.20. The kernel of the homomorphism
i : Hi (W a ; Z) Hi (W ; Z)
has rank equal4 to rk Bi and the cokernel has rank rk Ci + rk Di .
Below we omit the coecient group Z from the notation.
One has
(1.55) Hj (W, W a ) 9 Hj (N, N ) 9 H n1j (N ) 9 H n1j (MW ).
Here N denotes the preimage fW1 ([a, 0]). Note that MW is a strong deformation retract of N : the gradient ow of fW gives such a deformation
retraction.
Hence we obtain the short exact sequence
(1.56)
0 coker(n1j ) H j (MW ) ker n2j 0
which splits since the kernel of n2j is isomorphic to Bn2j (see

above) and hence it is free abelian.
This proves that the cohomology H (MW ) has no torsion and therefore the homology H (MW ) is free as well (by the Universal Coecient
Theorem). The cokernel of n1j is isomorphic to Cn1j Dn1j
as we established earlier. We nd that the rank of cokern1j equals
the number of subsets J {1, . . . , n} which are short or median and
have cardinality |J| = j + 1. In other words,
(1.57)
rk(cokern1j ) = aj + bj ,
where we use the notation introduced in the statement of Theorem 1.7.

The rank of the kernel of n2j equals the rank of Bn2j , i.e.,
the number of long subsets J {2, . . . , n} of cardinality |J| = j + 2.
Passing to the complements, we nd
(1.58)
rk(ker n2j ) = an3j
i.e., the number of short subsets containing 1 with |J| = n 2 j.

4Note
that the kernel of i (viewed as a subgroup) is distinct from Bi in general.
30
Combining (1.57), (1.58) with the exact sequence (1.56) we nally

obtain
rk Hj (MW ) = rk H j (MW ) = aj + bj + an3j .
This completes the proof, compare (1.24).
1.9. Maximum of the total Betti number of M"
It is well known that moduli space of pentagons MW with a generic
length vector Z = (l1 , . . . , l5 ) is a compact orientable surface of genus
not exceeding 4, see [75]. In the equilateral case, i.e., if Z = (1, 1, 1, 1, 1),
MW is indeed an orientable surface of genus 4 and hence the above upper
bound for pentagons is sharp.
In this section we state theorems generalizing this result for arbitrary n and give sharp upper bounds on the total Betti number
(1.59)
n3
=
bi (MW ).
i=0
Theorem 1.21. Let Z = (l1 , . . . , ln ) be a length vector, li > 0. Denote

by r the number [ n1
]. Then the total Betti number of the moduli space
2
MW does not exceed
;
<
n1
n1
Bn = 2
(1.60)
.
r
This estimate is sharp: Bn equals the total Betti number of the moduli
space of planar equilateral n-gons, see Examples 1.13 and 1.14.
Note that for n even the equilateral linkage with n sides is not
generic and hence Theorem 1.21 does not answer the question about
the maximum of the total Betti number on the set of all generic length
vectors with n even.
Theorem 1.22. Assume that n is even and Z = (l1 , . . . , ln ) is a generic
length vector. Then the total Betti number of MW does not exceed
(1.61)
Bn( = 2 Bn1 ,
where Bk is dened by (1.60). This upper bound is achieved on the

length vector Z = (1, 1, . . . , 1, i) where 0 < i < 1 and the number of
ones is n 1.
1.10. ON THE CONJECTURE OF KEVIN WALKER
31
Note that M(1,...,1,U) is dieomorphic to the product M(1,...,1) S 1

(the number of ones in both cases equals 2r + 1), see5 Prop. 6.1 of
[49]. Hence the total Betti number of M(1,...,1,U) is twice the total Betti
number of M(1,...,1) .
Proofs of Theorems 1.21 and 1.22 can be found in [28].
The asymptotic behavior of the number Bn (dened by (1.60)) can
be recovered using available information about Catalan numbers
;
<
1
22r
2r
Cr =
,
r
r+1
r3/2
see [98]. One obtains the asymptotic formula
2
# 5
2
(1.62)
.
Bn 2n1 1
n
1.10. On the conjecture of Kevin Walker
In 1985 Kevin Walker in his study of topology of polygon spaces [100]
raised an interesting conjecture in the spirit of the well-known question
Can you hear the shape of a drum? of Marc Kac. Roughly, Walkers
conjecture asserts that one can recover relative lengths of the bars of
a linkage from intrinsic algebraic properties of the cohomology algebra
of its conguration space. In this section we survey results of a recent
paper [29] answering the conjecture for polygon spaces in R3 . In [29]
it is also proven that for planar polygon spaces the conjecture holds
in several modied forms: (a) if one takes into account the action of a
natural involution on cohomology, (b) if the cohomology algebra of the
involutions orbit space is known, or (c) if the length vector is normal.
Some results mentioned below allow the length vector to be non-generic
(the corresponding polygon space has singularities).
Walkers conjecture [100] states that for a generic length vector
Z the cohomology ring of MW determines the length vector Z up to a
natural equivalence (described below). To state the Walker conjecture
in full detail we recall the dependence of the conguration space MW on
the length vector
(1.63)
Z = (l1 , . . . , ln ) Rn+ .
Here Rn+ denotes the set of vectors in Rn having non-negative coordinates. Clearly, MW = MtW for any t > 0. Also, MW is dieomorphic to
MW! if Z( is obtained from Z by permuting coordinates.
5This
can also be easily deduced from Lemma 1.4, statement (4).
32
unit simplex, i.e.,

Denote by A = An1 Rn+ the interior of the+
the set given by the inequalities l1 > 0, . . . , ln > 0,
li = 1. One can
view A as the quotient space of Rn+ with respect to R+ -action. For any
subset J {1, . . . , n} we denote by HJ Rn the hyperplane dened
by the equation
=
=
li =
(1.64)
li .
iJ
iJ
/
One considers the stratication

(1.65)
A(0) A(1) A(n1) = A.
Here the symbol A(i) denotes the set of points Z A lying in at least
n 1 i linearly independent hyperplanes HJ for various subsets J. A
stratum of dimension k is a connected component of the complement
A(k) A(k1) . By Theorem 1.1 of [49], manifolds with singularities MW
and MW! are dieomorphic if the vectors Z and Z( belong to the same
stratum.
Lemma 1.23. Two length vectors Z, Z( An1 lie in the same stratum
of An1 if and only if the family of all subsets J {1, 2, . . . , n} which
are short with respect to Z coincides with the family of all subsets J
{1, 2, . . . , n} which are short with respect to Z( .
Proof. J is long i the complement J is short and J is median i
neither J nor J is short. Hence, vectors Z, Z( satisfying conditions of
the lemma have identical families of short, long and median subsets.
This clearly implies that Z and Z( lie in the same stratum of A.
!
Strata of dimension n 1 are called chambers. Vectors Z lying in
chambers are generic. Non-generic length vectors lie in walls separating
chambers and hence satisfy linear equations (1.64) for some J.
Walkers conjecture: Let Z, Z( A be two generic length vectors; if the corresponding polygon spaces MW and MW! have isomorphic graded integral cohomology rings, then for some permutation :
{1, . . . , n} {1, . . . , n} the length vectors Z and (Z( ) lie in the same
chamber of A.
It is important to recall that polygon spaces (1.3) come with a
natural involution
(1.66)
: MW MW ,
(u1 , . . . , un ) = (
u1 , . . . , un )
33
induced by complex conjugation; a similar involution played an important role earlier in section 1.8. Geometrically, this involution asso!2
!3
!1
!4
!4
!1
!3
!2
ciates to a polygonal shape the shape of the reected polygon. The

xed points of are the collinear congurations, i.e., degenerate polygons. In particular we see that : MW MW has no xed points i
the length vector Z is generic. Clearly, induces an involution on the
cohomology of MW with integral coecients
(1.67)
: H (MW ; Z) H (MW ; Z).
Theorem 1.24. [29] Suppose that two length vectors Z, Z( An1 are
ordered, i.e., Z = (l1 , l2 , . . . , ln ) with l1 l2 ln > 0 and
similarly for Z( . If there exists a graded ring isomorphism of the integral
cohomology algebras
f : H (MW ; Z) H (MW! ; Z)
commuting with the action of the involution (1.67), then Z and Z( lie in
the same stratum of A. In particular, under the above assumptions the
moduli spaces MW and MW! are -equivariantly dieomorphic.
Let Z, Z( An1 be length vectors (possibly non-generic) lying in
the same stratum. Then there exists a dieomorphism : MW! MW
which is equivariant with respect to the involution , see page 36 and
Remark 3.3 in [49].
W denote the factor-space of MW with respect to the involution
Let M
W is given by
(1.66). An alternative denition of M
(1.68)
W = {(u1 , . . . , un ) S 1 S 1 ;
M
n
=
i=1
li ui = 0}/O(2).
34
Theorem 1.25. [29] Suppose that two generic ordered length vectors
Z, Z( An1 are such that there exists a graded algebra isomorphism
W ; Z2 ) H ( M
W! ; Z2 )
f : H (M
of cohomology algebras with Z2 coecients. If n == 4, then Z and Z( lie
in the same chamber of A.
Theorem 1.25 is false for n = 4. Indeed, for the length vectors
W and M
W! are circles.
Z = (2, 1, 1, 1) and Z( = (2, 2, 2, 1) the manifolds M
However MW and MW! are not dieomorphic (the rst is S 1 and the
second is S 1 5 S 1 ) and thus Z and Z( do not lie in the same chamber.
In [29] we also prove a result in the spirit of Walkers conjecture
for the spatial polygon spaces. These spaces are dened by
n
=
2
2
(1.69) NW = {(u1 , . . . , un ) S S ;
li ui = 0}/SO(3).
i=1
Points of NW parameterize the shapes of n-gons in R3 having sides of

length Z = (l1 , . . . , ln ). If the length vector Z is generic then NW is a
closed smooth manifold of dimension 2(n 3).
Theorem 1.26. [29] Suppose that two generic ordered length vectors
Z, Z( An1 are such that there exists a graded algebra isomorphism
f : H (NW ; Z2 ) H (NW! ; Z2 )
of cohomology algebras with Z2 coecients. If n == 4, then Z and Z( lie
in the same chamber of A. This theorem remains true if the cohomology
algebras are taken with integral coecients.
Theorem 1.26 is false for n = 4: for length vectors Z = (2, 1, 1, 1)
and Z( = (2, 2, 2, 1) lying in dierent chambers (see above) the manifolds
NW and NW! are both dieomorphic to S 2 .
Definition 1.27. A length vector Z = (l1 , . . . , ln ) is called normal if
J == where J runs over all subsets J {1, . . . , n} with |J| = 3
which are either median or long with respect to Z.
A stratum of An1 is called normal if it contains a normal vector.
Clearly, any vector lying in a normal stratum is normal. A length
vector Z with the property that all subsets J of cardinality 3 are short is
normal since then the intersection J where |J| = 3 equals {1, . . . , n}
as the intersection of the empty family.
If Z = (l1 , . . . , ln ) where 0 < l1 l2 ln , then Z is normal
if and only if the set {n 3, n 2, n 1} is short with respect to Z.
35
Indeed, if this set is not short, then neither of the sets {n 3, n 2, n},
{n 3, n 1, n}, {n 2, n 1, n} is short and the intersection of
these four sets of cardinality three is empty. On the other hand, if the
set {n 3, n 2, n 1} is short, then any median or long subset of
cardinality three J {1, . . . , n} contains n and therefore J, where
|J| = 3, also contains n.
Examples of non-normal length vectors are (1, 1, 1, 1, 1) (for n = 5)
and (3, 2, 2, 2, 1, 1) for n = 6. Only 7 chambers out 21 are normal for
n = 6. However, for large n it is very likely that a randomly selected
length vector is normal. For n = 9, where there are 175428 chambers
up to permutation, 86% of them are normal. It is shown in [31] that
the (n 1)-dimensional volume of the union Nn An1 of all normal
strata satises
vol(An1 Nn )
n6
<
,
vol(An1 )
2n
i.e., for large n the relative volume of the union of non-normal strata
is exponentially small.
Theorem 1.28. [29] Suppose that Z, Z( An1 are two ordered length
vectors such that there exists a graded algebra isomorphism between the
integral cohomology algebras H (MW ; Z) H (MW! ; Z). Assume that
one of the vectors Z, Z( is normal. Then the other vector is normal as
well and Z and Z( lie in the same stratum of the simplex A.
Consider the action of the symmetric group n on the simplex An1
induced by permutations of vertices. This action denes an action of
n on the set of strata and we denote by cn and by cn the number
of distinct n -orbits of chambers (or chambers consisting of normal
length vectors, respectively).
Theorems 1.241.28 imply:
Theorem 1.29. [29] (a) For n == 4 the number of distinct dieomorphism types of manifolds NW , where Z runs over all generic vectors of
An1 , equals cn ;
(b) for n == 4 the number of distinct dieomorphism types of mani W , where Z runs over all generic vectors of An1 , equals cn ;
folds M
(c) the number xn of distinct dieomorphism types of manifolds MW ,
where Z runs over all generic vectors of An1 , satises cn xn cn ;
(d) the number of distinct dieomorphism types of manifolds with
singularities MW , where Z varies in An1 , is bounded above by the number of distinct n -orbits of strata of An1 and is bounded below by the
number of distinct n -orbits of normal strata of An1 .
36
Statements (a), (b), (c), (d) remain true if one replaces the words
dieomorphism types by homeomorphism types or by homotopy
types.
It is an interesting combinatorial problem to nd explicit formulae
for the numbers cn and cn and to understand their behavior for large
n . For n 9, the numbers cn have been determined in [49], by giving
an explicit list of the chambers. The following table gives the values cn
and cn for n 9:
n 3 4 5
cn 2 3 7 21 135 2470 175428

cn 1 1 2
65
1700
151317
We refer the reader to the original paper [29] for proofs of theorems
mentioned in this section and more details.
1.11. Topology of random linkages
In this section we study polygon spaces NW and MW assuming that the
number of links n is large, n . This approach is motivated by
applications in topological robotics, statistical shape theory and molecular biology. We view the lengths of the bars of the linkage as random
variables and study asymptotic values of the average Betti numbers
when n tends to innity. We describe a surprising fact (established in
[30], [31]) that for a reasonably ample class of sequences of probability measures the asymptotic values of the average Betti numbers are
independent of the choice of the measure.
The following picture summarizes our description of the eld of
topological spaces Z 7 NW viewed as a single object. The open simplex
n1 is divided into a huge number of tiny chambers, each representing
a dieomorphism type of manifolds NW . The symmetric group n acts
on the simplex n1 mapping chambers to chambers and manifolds NW
and NW! are dieomorphic if and only if the vectors Z, Z( lie in chambers
belonging the same n -orbit.
The main idea of this section is to use methods of probability theory and statistics in dealing with the variety of dieomorphism types
of conguration spaces NW and MW for n large. In applications dierent
manifold types appear with dierent probabilities and our intention is
to study the most frequently emerging manifolds NW and the mathematical expectations of their topological invariants. Formally, we view
the length vector Z n1 as a random variable whose statistical
1.11. TOPOLOGY OF RANDOM LINKAGES
37
behavior is characterized by a probability measure n . Topological

invariants of NW become random functions and their mathematical expectations might be very useful for applications. Thus, one is led to
study the average Betti numbers6
*
*
b2p (NW )dn and
(1.70)
bp (MW )dn
n1
n1
where the integration is understood with respect to Z. Here NW denotes

the polygon space in R3 (see (1.69)) and MW denotes the planar polygon
space (1.3). The main result of this section states that for p xed and
n large these average Betti numbers can be calculated explicitly up to
an exponentially small error. More precisely,
<
*
p ;
=
n1
b2p (NW )dn
i
n1
i=0
and
<
n1
.
bp (MW )dn
p
n1
It might appear surprising that the asymptotic values of average Betti
numbers do not depend on the sequence of probability measures n
which are allowed to vary in an ample class of admissible probability
measures described below. First we have to dene what is meant by
an admissible sequence of measures.
For a vector Z = (l1 , . . . , ln ) we denote by
;
|Z| = max{|l1 |, . . . , |ln |}

the maximum of absolute values of coordinates. The symbol n1
denotes the open unit simplex, the set of all vectors Z = (l1 , l2 , . . . , ln )
Rn such that li > 0 and l1 + + ln = 1. Let n denote the Lebesgue
measure on n1 normalized so that n (n1 ) = 1. In other words, for
a Lebesgue measurable subset A n1 one has
n (A) =
vol(A)
vol(n1 )
where the symbol vol denotes the (n 1)-dimensional volume. For an

integer p 1 we write
(1.71)
p = pn1 = {Z n1 ; |Z| (2p)1 }.
Clearly, p q for p q.
6It
is well known that all odd-dimensional Betti numbers of N% vanish, see [65].
38
Definition 1.30. [31] Consider a sequence of probability measures n

on n1 where n = 1, 2, . . . . It is called admissible if n = fn n where
f)n : n1 R is a sequence of functions satisfying: (i) fn 0, (ii)
f dn = 1, and (iii) for any p 1 there exist constants A > 0
n1 n
and 0 < b < 2 such that
fn (Z) A bn
(1.72)
for any n and any Z pn1 n1 .

Note that property (iii) imposes restrictions on the behavior of the
sequence n only in domains pn1 .
Example 1.31. Consider the unit cube !n Rn+ given by the inequalities 0 li 1 for i = 1, . . . , n. Let n be the probability measure
on Rn+ supported on !n Rn+ such that the restriction n |!n is the
Lebesgue measure, n (!n ) = 1. As in [30] consider the sequence of
induced measures n = q (n ) on simplices n1 where q : Rn+ n1
is the normalization map q(Z) = tZ where t = (l1 + + ln )1 . It is easy
to see that n = fn n where fn : n1 R is a function given by
(1.73)
fn (Z) = kn |Z|n ,
Z n1 .
Here kn is a constant which can be found (using (ii) of Denition 1.30)

from the equation
*
1
(1.74)
|Z|n dn .
kn =
n1
as the
If Z pn1 then fn (Z) kn (2p)n . We can represent n1
p
union A1 An where
Ai = {(l1 , . . . , ln ) n1 ; li (2p)1 },
!
>n1
2p1
Clearly, n (Ai ) = 2p
and hence
i = 1, . . . , n.
<n1
2p 1
.
n (
p ) 1 n
2p
;
!
>n1 <
2p1
1
n
Using (1.74) we nd that kn (2p) 1 n 2p
. This shows
;
n1
that the sequence (2p)n kn remains bounded as n , implying (iii)

of Denition 1.30. Hence, the sequence of measures {n } is admissible.
1.11. TOPOLOGY OF RANDOM LINKAGES
39
The following two theorems are the main results of this section.
Theorem 1.32. [31] Fix an admissible sequence of probability measures
n and an integer p 0, and consider the average 2p-dimensional Betti
number (1.70) of polygon spaces NW in R3 for large n . Then there
exist constants C > 0 and 0 < a < 1 (depending on the sequence of
measures n and on the number p but independent of n) such that the
average Betti numbers (1.70) satisfy
$
$
$ *
<
p ;
= n 1 $$
$
$ < C an
$
(1.75)
b2p (NW )dn
$
$
i
$
$ n1
i=0
for all n.
Theorem 1.33. [31] Fix an admissible sequence of probability measures
n and an integer p 0, and consider the average p-dimensional Betti
number (1.70) of planar polygon spaces for large n . Then there
exist constants C > 0 and 0 < a < 1 (depending on the sequence of
measures n and on the number p but independent of n) such that
$
$
$ *
;
<$
$
n 1 $$
n
$
bp (MW )dn
(1.76)
$
$<C a
p
$ n1
$
for all n.
Proofs of Theorems 1.32 and 1.33 can be found in [31].
Matthew Hunt [52] applied Monte Carlo simulation to compute
numerically the average Betti numbers of planar polygon spaces MW for
various n 11. His numerical results conrm Theorem 1.33.
CHAPTER 2
Euler Characteristics of Conguration Spaces

In this chapter we describe a beautiful result of S. Gal [38] which
expresses explicitly the Euler characteristics of various conguration
spaces associated with polyhedra.
2.1. The Euler Gal power series
For a nite simplicial polyhedron X we denote by F (X, n) the space of
all congurations of n distinct particles moving in X. In other words,
F (X, n) is dened as the subspace of the Cartesian product
F (X, n) X n = X X
of n copies of X consisting of all n-tuples (x1 , . . . , xn ) satisfying xi == xj
for i =
= j. Conguration spaces of this kind appear in robotics in
problems of simultaneous control of multiple objects (robots) avoiding
collisions.
The symmetric group n acts freely on F (X, n) by permuting the
particles. The factor
B(X, n) = F (X, n)/n
is the space of all subsets of cardinality n in X. The notation B intends
to bring association with braids; the fundamental group 1 (B(X, n))
is the braid group of X.
Our aim is to compute the Euler characteristics (F (X, n)) and
(B(X, n)) of conguration spaces F (X, n) and B(X, n) for a xed
polyhedron X and various values of n. These numbers are related by
(2.1)
(B(X, n)) =
(F (X, n))
n!
where n = 1, 2, . . . . One formally denes F (X, 0) and B(X, 0) as singletons (i.e., spaces consisting of a single point) so that
(B(X, 0)) = (F (X, 0)) = 1.
42
2. EULER CHARACTERISTICS OF CONFIGURATION SPACES
With each nite polyhedron X one associates a sequence of integers

(2.1) which may be organized into a formal power series with integer
coecients
=
=
tn
n
(B(X, n)) t =
(F (X, n)) .
(2.2) euX (t) =
n!
n=0
n=0
The latter is called the Euler Gal power series of X. The constant
term of euX (t) is 1. We shall see that euX (t) has a fairly simple expression while the individual numbers (2.1) are much more involved.
Theorem 2.1. For any nite polyhedron X the Euler Gal power
series euX (t) represents a rational function
(2.3)
euX (t) =
p(t)
,
q(t)
where p(t) and q(t) are polynomials with integral coecients satisfying
p(0) = 1 = q(0),
deg(p) deg(q) = (X).
Theorem 2.1 will be made more precise later in Theorem 2.3.

Theorem 2.1 implies that the numbers n = (B(X, n)) satisfy a
linear recurrence relation:
Corollary 2.2. Given a nite simplicial polyhedron X, there exist
integers a1 , . . . , ar Z (for some r depending on X) such that for any
n r one has
(2.4)
n = a1 n1 + a2 n2 + + ar nr .
Theorem 2.3 stated below describes explicitly the polynomials p(t)

and q(t) appearing in formula (2.3) in terms of local topological properties of X.
Recall the notion of link of a simplex in a simplicial complex. Let
be a simplex of X. The link of (denoted L ) is the union of all
simplices X such that = and and lie in a common
simplex of X. Clearly L is a subcomplex of X.
Figure 2.1 illustrates this notion. We see a one-dimensional simplex
which is incident to three two-dimensional simplices. The link L
consists of three points and the cone C(L ) over L is shown on Figure
2.1 on the right. This example tells us that the cone C(L ) describes
the geometry of X near in the direction perpendicular to .
It will be convenient for us to deal with polyhedral cell complexes.
Let us briey recall the relevant denitions, see [85] for more detail.
2.1. THE EULER GAL POWER SERIES
43
C(L)
Figure 2.1. Link of a simplex.

We consider RN with coordinates x = (x1 , . . . , xn ) and with the metric
d(x, y) = sup |xi yi | so that any ball {x; d(x, y) i} is a cube centered at y. A cell RN is the convex hull of a nite set of points
v1 , . . . , vm RN . We say that the set v1 , . . . , vm is the set of vertices
of if the convex hull of any proper subset of v1 , . . . , vm is a proper
subset of . We refer to [85], pages 13, 14 for the denition of a face
of a cell; notation ( < .
A polyhedral cell complex K is a nite collection of cells lying in
some Euclidean space RN such that with each cell it contains all its
faces and such that the intersection of any pair of cells , K
is a face of both and . The underlying polyhedron X = |K| =
has the following important property: any point x X has a cone
neighbourhood C(L) where L is compact. If x lies in the interior of a
cell , then a metric ball of small radius with center x is topologically
the product of a Euclidean disk of dimension dim and a cone C(L )
where L is compact. This L is called the link of the cell .
If X is an m-dimensional piecewise-linear
manifold with boundary (we refer to the book
[85] for basic denitions) then for any cell of
dimension d lying in the interior of X one has
L 9 S md1 . If belongs to the boundary X

L
then L is topologically the disk Dmd1 .
It will be convenient for us to introduce the
invariant
(2.5)
(X)
= 1 (X) = (C(X), X),
the reduced Euler characteristic. Here C(X) denotes the cone over X.
The reduced Euler characteristic behaves well with respect to the join
operation:
(2.6)
(X
Y ) = (X)
(Y
).
44
This follows from the observation

(C(X Y ), X Y ) = (C(X), X) (C(Y ), Y )
using the multiplicative property of the Euler characteristic. Note also
the useful formula
(2.7)
(S
k ) = (1)k+1 .
The statement given below is not used in the sequel and therefore it
is left as an exercise. It is a special case of Theorem 2.3 the equation
for the rst order terms in t.
Exercise: Show that for a nite polyhedral cell complex X one has
=
(2.8)
(L
),
(X) =
where runs over all cells of X.

Next we state an important addition to Theorem 2.1:
Theorem 2.3 (see Theorem 2 in [38]). Let X be a nite polyhedral cell
complex. Then the polynomials p(t) and q(t) which appear in formula
(2.3) can be chosen as follows:
9
(2.9)
p(t) =
[1 + t(L
)]
dim =even
and
(2.10)
q(t) =
[1 t(L
)] .
dim =odd
In (2.9) and in (2.10) runs over all cells of X having even or odd
dimension, correspondingly.
Corollary 2.4. The zeros of the rational function euX (t) are of the
form
(2.11)
t = (L
)1 ,
where is an even-dimensional cell with (L

) == 0. Poles of euX (t)
are of the form
(2.12)
t = (L
)1 ,
where is an odd-dimensional cell with (L

) =
= 0.
2.2. CONFIGURATION SPACES OF MANIFOLDS
45
The proofs of Theorems 2.1 and 2.3 will be completed in section

2.8. In the following sections we examine the statements of these
theorems in several special cases.
2.2. Conguration spaces of manifolds
Here we apply Theorems 2.1 and 2.3 in the case of manifolds.
Theorem 2.5. Let X be a piecewise-linear compact manifold, possibly
with boundary. Then
(1 + t)(X) , if dim X is even,

(2.13)
euX (t) =
(1 t)(X) , if dim X is odd.

Theorem 2.5 was mentioned in [35] in the special case when dim X
is even and X = . Passing to binomial expansions Theorem 2.5 may
be restated as follows:
( 1) . . . ( k + 1) if dim X is even,
(F (X, k)) =
( + 1) . . . ( + k 1) if dim X is odd.
Here X is a compact manifold, possibly with boundary, and = (X).
Theorem 2.5 may also be obtained by examining the towers of
Fadell Neuwirth brations [21]: if X is a manifold without boundary,
then projecting onto the rst coordinate gives a locally trivial bration
F (X, n) X. Its bre above a point p X equals F (X {p}, n 1),
the conguration space of n 1 distinct points in X {p}. Using the
multiplicative property of the Euler characteristic1 we nd
(2.14)
(F (X, n)) = (F (X {p}, n 1)) (X).
Iterating we obtain
(F (X, n)) = (X) (X1 ) (Xn1 ),
where each Xi is obtained from X by removing i distinct points. This
gives the formulae mentioned above since (Xi ) = (X) (1)dim X i.
Next we give a proof of Theorem 2.5 based on Theorem 2.3. For
every cell lying in the interior of X one has L = S nd 1 and
1It
states that for a bration E B with bre F one has (E) = (B)(F ).
46
(L
) = (1)nd where n = dim X and d = dim . If is a cell
lying in the boundary, then (L
) = 0. Hence Theorem 2.3 gives
(2.15)
.
((X)(X)
.
euX (t) = 1 + (1)dim X t
This implies (2.13) since for dim X even one has (X) = 0 and for
dim X odd, (X) = 2(X).
2.3. Conguration spaces of graphs
Next we examine the special case of Theorem 2.3 when X = is a nite
graph, i.e., a one-dimensional nite simplicial complex. For any vertex
v the link Lv is the discrete set of vertices which are connected to
v by an edge in . Hence
(L
v ) = 1 (v)
where (v) denotes the valence of v. For any edge e the link Le is
empty and therefore
(L
e ) = 1.
Applying Theorem 2.3 we nd
Theorem 2.6. The Euler Gal power series of a graph is given by
the formula
(2.16)
eu (t) = (1 t)E
0
v
[1 + t(1 (v))]
? 0
. (2
4
. (
t + [1 + t(1 (v))] .
= 1 + E1 t + E+1
2
v
Here E denotes the total number of edges in and v runs over all
vertices of .
Observe that in the product appearing in (2.16) the univalent vertices (v) = 1 give no contribution. As another observation note that
subdividing an edge by introducing a new vertex of valence 2 makes no
change to the Euler Gal series (2.16) as two new terms cancel each
other.
As an illustration we use formula (2.16) to compute explicitly the
Euler characteristic (F (, 2)), which equals twice the coecient of t2
in the above series.
2.3. CONFIGURATION SPACES OF GRAPHS
47
Corollary 2.7. For any nite graph one has

=
(2.17)
(F (, 2)) = ()2 + ()
((v) 1)((v) 2).
v
Proof. The coecient of t2 in (2.16) equals

<
;
=
E+1
1=
(2.18)
(1 (v))(1 (w)).
+E
(1 (v)) +
2
2
v
v1=w
In the last sum the summation
is over all ordered pairs of distinct
+
vertices (v, w). Since 2E = v (v) the second term in (2.18) is
VE
=
1=
(v)
(w),
2 v
w
where V is the number of vertices. Similarly, the third term in (2.18)

equals
1 2
1=
(v)(w).
(V V ) (V 1) 2E +
2
2 v1=w
Hence by Theorem 2.3 we nd that the Euler characteristic (F (, 2))
equals twice the coecient of t2 in (2.18), i.e.,
E(E + 1) + 2V E F + (V 2 V ) (V 1)4E,
+
where F denotes v (v)2 . Formula (2.19) can be rewritten as
(2.19)
(V 2 2V E + E 2 ) + (V E) 2V + 6E F
which is equivalent to (2.17).
As an example consider graph shown on Figure 2.2. It consists
Figure 2.2. Graph .
48
of edges incident to a vertex. The the Euler Gal series is

eu (t) =
1 + t(1 )
.
(1 t)
Hence,
(F ( , n)) =
( + n 2)!
[(n 1) 2n + 1] .
( 1)!
This result was obtained in [39] by a dierent method.

Next we analyze the behavior of (F (X, n)) assuming that the number of particles n tends to innity.
Proposition 2.8. Assume that is a connected graph with () < 0.
Then for large n one has the asymptotic formula
(2.20)
(B(, n)) c nE 1 .
Here E ( = E V + V ( with V ( denoting the number of vertexes v of

satisfying (v) == 2 and the constant c is given by
0
(2 (v))
c =
(v)1=2
(E ( 1)!
In the product, v runs over all vertexes with (v) == 2.

Proof. One may write
9
!
!
(1 + t
(v)) = a0 (1 t)V + a1 (1 t)V 1 + + aV ! ,
(v)1=2
where
(v) = 1 (v) and
9
aV ! =
(2 (v)) == 0.
(v)1=2
Using the binomial expansion

(1 t)
<
;
=
k+n1
n=0
tn
2.4. RECURRENT FORMULA FOR euX (t)
49
and Theorem 2.6 we obtain that (F (, n)) equals

(E ( + n 1)!
(E ( + n 2)!
(n () 1)!
! 1
+
a
+ + a0
V
(
(
(E 1)!
(E 2)!
(() 1)!
C
3
(E ( + n 1)!
E( 1
! + aV ! 1
+ .
=
a
V
(E ( 1)!
E( 1 + n
aV !
For n large the rst term in square brackets dominates and we may
ignore the other terms. We obtain that asymptotically one has
(2.21)
(F (, n))
9
(v)1=2
(2 (v))
(n + E ( 1)!
.
(E ( 1)!
This implies the statement of Lemma 2.8.
2.4. Recurrent formula for euX (t)

Having examined various special cases of Theorem 2.3 of S. Gal, we
start preparing auxiliary results which will be used in the proof of
Theorem 2.3. Our exposition essentially follows the original paper [38].
A central role in the proof of Theorem 2.3 is played by the following
statement:
Theorem 2.9. Let X be a nite polyhedral cell complex. For any
cell X denote by +4 X a suciently small open contractible
neighborhood of an internal point of . Then
=
(2.22) (F (X, n)) =
(F (X +4), n 1) (L
),
where runs over all simplices of X. Formula (2.22) can be equivalently rewritten as
=
d
(2.23)
euX (t) =
euX&3 (t) (L
).
dt
Figure 2.3 shows such neighborhoods +4 for edges and vertices of

a graph.
In general the set +4 is homeomorphic to the product
int intC(L )
where C(L ) denotes the cone over the link L of . The interior of the
cone C(L) is dened as intC(L) = C(L) L, i.e., as the cone without
50
!"
!"
Figure 2.3. Neighborhood +4 in the case when X is a graph.

the base. Recall the standard convention that the cone over an empty
set is a one-point set (the vertex).
In the case n = 1, Theorem 2.9 coincides with the statement of
Exercise 1.
Proof of Theorem 2.9. Observe that both sides of (2.22) are invariant under subdivision. This is obvious with regards to the left-hand
side of (2.22) as it is independent of the cell structure. Suppose that
under a subdivision the interior of a cell of dimension d is divided
into cells 1 , . . . , k where the dimension of i is di . Clearly
k
=
(1)di = (1)d
(2.24)
i=1
as follows from invariance of the Euler characteristic. Under the subdivision the term
(2.25)
(F (X +4, n 1) (L
)
of (2.22) is being replaced by the sum

(2.26)
k
=
(F (X +i 4, n 1) (L
i ).
i=1
Note that
(2.27)
F (X +i 4, n 1) = F (X +4, n 1)
for all i = 1, . . . , k. Besides, one has

Li = S ddi 1 L
and using (2.6) and (2.7) we nd
(2.28)
(L
i ) = (1)ddi (L
).
Combining (2.28), (2.27) with (2.24) one obtains equality between

(2.25) and (2.26).
51
Hence we may assume that X is a simplicial complex. Fix a metric

d on X compatible with the simplicial structure such that each cell of
X is a regular simplex with side of length 8. The symbols B (x) and
B (A) denote open metric balls of radius around x X and A X
respectively.
We will study the projection
: F (X, n) X
onto the rst factor.
We denote by pk the projection onto the k-dimensional skeleton
(k)
X of X; it is a map x 7 pk (x) dened only for points x X which
are suciently close to X (k) and suciently far from X (k1) . Here
pk (x) X (k) denotes the closest point to x lying in X (k) .
Set i = 1/8 and dene inductively the following closed subspaces
of F (X, n):
(2.29)
Ak Bk Ck Ak+1 .
Here Ak F (X, n) denotes the set of all congurations (x1 , . . . , xn ),

xi X, xi == xj , such that the rst particle x1 lies far from the
skeleton X (k1) in the following sense:
(2.30)
d(x1 , X (l) ) il ,
for l < k.
Note that A0 = F (X, n) and Ak+1 = for k dim X.

The set Bk Ak contains all congurations (x1 , . . . , xn ) Ak such
that either d(x1 , X (k) ) ik , or d(x1 , X (k) ) < ik and x1 is the only
particle amongst x1 , x2 , . . . , xn lying in the ball BUk (pk (x1 )).
Finally, Ck Bk consists of all congurations of Bk satisfying
d(x1 , X (k) ) ik /2.
We prove below that for k = 0, 1, . . . one has:
(i) Bk is a deformation
+ retract of Ak and hence (Ak , Bk ) = 0;
(F (X +4, n 1) (L
);
(ii) (Bk , Ck ) =
dim =k
(iii) Ak+1 is a deformation retract of Ck and hence (Ck , Ak+1 ) = 0.
52
Statements (i) (iii) imply Theorem 2.9. Indeed,

(F (X, n)) = (A0 ) =
(Ak , Ak+1 )
k=0
[(Ak , Bk ) + (Bk , Ck ) + (Ck , Ak+1 )]
k=0
(F (X +4, n 1) (L
).
Let us start proving (ii). Consider a conguration (x1 , . . . , xn ) lying

in Bk Ck . Then d(x1 , X (k) ) < ik /2 and d(x1 , X (l) ) il for all l < k.
This means that x1 lies in the union of disjoint neighborhoods
N = {x X; d(x, ) < ik /2} (Ak ),
one for each k-dimensional cell , see Figure 2.4. Clearly, N is home-
N2
N3
N 1
Figure 2.4. Neighborhoods N .
omorphic to the product int(C(L )) and the boundary of N in
(Bk ) is homeomorphic to L . Over N the map : F (X, n) X
is a locally trivial bration with bre F (X +4, n 1). Hence we
obtain
+
(Bk , Ck ) =
(F (X +4, n 1) () (CL , L )
dim =k
+
dim =k
(F (X +4, n 1) (L
).
53
Next we prove (i). We want to construct a deformation retraction

of Ak onto Bk . A conguration (x1 , . . . , xn ) satisfying (2.30) belongs to
Ak Bk if d(x1 , X (k) ) < ik and the ball BUk (pk (x1 )) contains besides x1
some other particle xi where i > 1. Our plan is to apply a stretching
isotopy moving the particles away from pk (x1 ) until either x1 leaves
the neighborhood BUk (X (k) ) or all other particles are outside the ball
BUk (x1 ).
We use the following notation. Let Y be a topological space and
CY = [0, ] Y /{0 Y } be a cone over Y . We shall say that a point
z = (t, y) CY has modulus |z| = t and argument arg(z) = y. Note
that the argument function arg : CY 0 Y is not dened on the
vertex of the cone 0 CY . The functions | | and arg form polar
coordinates on CY .
Let = (x1 , . . . , xn ) be a conguration in Ak Bk such that
d(x1 , X (k) ) < ik . The closure of the ball BUk (pk (x1 )) is a cone over
its boundary and hence may be viewed with its polar coordinates. The
numbers
= () = min{d(xi , pk (x1 ); i > 1}
and
r = r() = max(, d(x1 , X (k) ))
are continuous functions of a conguration Ak such that their

rst point x1 lies near X (k) . Note that r > 0. Let the function
: [0, 2ik ] (0, 2ik ] [0, 2ik ] be given for 0 < R ik by
(t, R) =
R ik
2i2k R2
t
+
t2
k
k
R(2i R)
R(2i R)
and (t, R) = t for R [ik , 2ik ]. It has the following properties:

(a) for any R (0, 2ik ] the function t 7 (t, R) is a homeomorphism [0, 2ik ] [0, 2ik ] preserving the endpoints;
(b) (R, R) = ik for R (0, ik ];
(c) (t, R) t for R [ik , 2ik ].
Using the function we dene a homeomorphism F, : X X
(where [0, 1]) as the identity map outside the ball B2Uk (pk (x1 ) and
given by the formula
F, (y) = (|y|, (1 )r + ik ) arg(y),
y = (|y|, arg(y))
inside this ball. Clearly, F, is a homeomorphism of X which depends

continuously on Ak and [0, 1]. Because of (c), the map F,1 is
the identity map X X. Figure 2.5 shows that under F, every point
moves away from pk (x1 ) along the line connecting these points. For
54
F, (y)
x1
pk(x1)
Figure 2.5. Homeomorphism F, : X X.

= 0 the homeomorphism F,0 : X X has the following property:
the inequality
d(F,0 (xi ), pk (x1 )) ik
holds either for i = 1 or for all i > 1. In other words, the conguration
(F,0 (x1 ), . . . , F,0 (xn )) lies in Bk . Also, if the initial conguration =
(x1 , . . . , xn ) lies in Bk , then r ik and therefore F, (y) = y for any
y X. The homotopy H : Ak Ak ,
H () = (F, (x1 ), . . . , F, (xn )),
= (x1 , . . . , xn ),
[0, 1]
is a deformation retraction of Ak onto Bk . This proves (i).

Statement (iii) follows using arguments similar to those used in the
proof of (ii). If a conguration = (x1 , . . . , xn ) lies in Ck Ak+1 , then
x1 satises
ik /2 d(x1 , X (k) ) < ik
and one applies a stretching isotopy of X moving x1 away from the
skeleton X (k) . For full details we refer to [38], page 64.
!
2.5. Cut and paste surgery
Let X be a nite polyhedron. A subpolyhedron S X is collared if
it has a neighborhood U X such that (U, S) is homeomorphic to
(S (1, 1), S {0}). We say that a polyhedron Y is obtained from
X by C&P-surgery along S if Y is the result of cutting X along S and
then gluing back by a piecewise linear automorphism of S. Here C&P
stands for cut and paste.
The following statement is a crucial observation:
Theorem 2.10. If Y is obtained from X by C&P -surgery along S,
then
(2.31)
euX (t) euY (t),
2.6. CUT AND PASTE GROTHENDIECK RING
55
Figure 2.6. C&P-surgery.

or, equivalently,
(2.32)
(F (X, n)) = (F (Y, n))
for any n.
Proof. We prove (2.32) by induction on n. It is trivial for n = 0 and
is obvious for n = 1. Suppose that it is true for n 1. By Theorem 2.9
we have
=
(F (X +4), n 1) (L
X
(2.33) (F (X, n)) =
),
and
(2.34)
(F (Y, n)) =
(F (Y +4), n 1) (L
Y ).
There is a natural one-to-one correspondence between the cells of X

and Y and we claim that the contribution of each simplex to formulae
(2.33) and (2.34) are equal.
If does not lie in UU = S (i, i) then its links in X and in Y
Y
are equal, LX
= L , and obviously the spaces X +4 and Y +4 are
related by C&P surgery along S. Hence, by induction hypothesis, one
has (F (X +4), n 1) = (F (Y +4, n 1)).
If intersects UU , then the above argument applies with S replaced
Y
by a slightly shifted parallel copy S ( of S. We conclude that LX
= L
and the spaces X +4 and Y +4 are related by a C&P-surgery
(along S ( ) and hence (F (X +4), n 1) = (F (Y +4, n 1)) by
induction.
!
2.6. Cut and paste Grothendieck ring
The operation of C&P-surgery and the relation of piecewise linear
homeomorphism generate an equivalence relation on the set of all nite polyhedra. The equivalence class of X is denoted [X]. The set of
56
all equivalence classes is a semi-ring with addition and multiplication

given by
[X] + [Y ] = [X 5 Y ],
[X] [Y ] = [X Y ].
The usual construction gives a Grothendieck ring C&P; elements of

C&P are represented by formal dierences [X] [Y ].
Proposition 2.11. The Euler Gal series euX (t) determines a homomorphism
euX (t)
Z[[t]]
[X] [Y ]
7
euY (t)
from the additive group of the Grothendieck ring C&P to the multiplicative group of formal power series having integral coecients and
constant term 1.
Proof. Consider
euX/Y (t) =
(F (X 5 Y, n))
n=0
tn
.
n!
It is easy to see that

(2.35)
n
=
(F (X 5 Y, n))
(F (X, k)) (F (X, n k))
=
,
n!
k!
(n
k)!
k=0
which can also be expressed by the equation

euX/Y = euX euY .
Proposition 2.11 now follows from Theorem 2.10.
Consider the following example. Suppose that X = A B where
A
S
Figure 2.7. C& P-surgery.

A and B are closed subpolyhedra and the intersection A B = S
has a neighborhood U = (S [1, 1]) X such that (U A, S) 9
2.7. CONES AND CYLINDERS
57
(S [0, 1], S {0}) and (U B, S) 9 (S [1, 0], S {0}). Then in

the ring C&P one has
(2.36)
[X] = [A] + [B] [S [1, 1]].
To see this start with the disjoint union X 5 (S [1, 1]), cut along
S 5 S (one copy S X and another S = S 0 S [1, 1]) and
glue one of these copies to another one and vice versa. The result is
homeomorphic to A 5 B which proves (2.36). Combining this result
with Proposition 2.11 we obtain:
Corollary 2.12. Under the conditions described above one has
(2.37)
euX (t) =
euA (t) euB (t)

.
euSI (t)
2.7. Cones and cylinders

In this section we compare the Euler Gal series of a cone and a
cylinder having the same base.
Proposition 2.13. Let X be a nite polyhedron. Then one has
(2.38)
euCX (t)
= 1 + t(X).
euXI (t)
Proof. Suppose that X is given a simplicial structure. Assume that

CX has the simplicial structure of the cone. Simplices of CX are of
three types: (i) simplices of X, (ii) cones over the simplices of X and
(iii) the vertex v of the cone. If is a simplex of X, then the link of
viewed as a simplex of CX equals the cone over its link in X, i.e.,
LCX
= C(LX
CX
). In particular, we see that (L

) = 0 for X. If
is a simplex of X, then the link of the simplex ( = C of CX equals
LX
. The link of the vertex v equals Lv = X.
We view the cylinder X I as the cone CX with a small open ball
around the vertex v removed. We shall use notations introduced in
Theorem 2.9. Let be a simplex of X and let ( = C CX be its
cone. Applying (2.36) to the neighborhood + ( 4 twice (for CX and for
X I) one nds
[CX] = [CX + ( 4] + [+ ( 4] [+ ( 4 I]
and
[X I] = [X I + ( 4] + [+ ( 4] [+ ( 4 I].
58
Subtracting, one obtains

[CX + ( 4] = [CX] + [X I + ( 4] [X I] C&P.
By Proposition 2.11,
(2.39)
euCX&! 3 (t) =
euCX (t)
euXI&! 3 (t).
euXI (t)
If ( = v is the vertex of the cone CX, then

CX + ( 4 = X I
and L! = X.
Theorem 2.9 gives
(2.40)
eu(CX (t) =
euCX&3 (L
CX
)
where runs over all simplices of CX. As explained above simplices of

type (i) do not contribute to (2.40). Using (2.39) we obtain that (2.40)
equals
(2.41)
euCX (t) =
euXI&3 (L
XI
) + euXI (t) (X).
euXI (t)
In the rst sum in (2.41) the symbol runs over all cells of X I of
the form I where is a simplex of X. However, links of the cells
of X I which intersect the top and the bottom faces of the cylinder
X I are contractible and therefore for such cells (L
) = 0. Hence,
we may apply Theorem 2.9 again to obtain:
(2.42)
eu(CX =
euCX
eu(XI + euXI (X),
euXI
which is equivalent to the dierential equation

;
(2.43)
euCX
euXI
<(
= (X).
This proves Proposition 2.13 since the constant term of Euler Gal
power series is always 1.
!
59
2.8. Proof of Theorem 2.3

Let X be a nite polyhedron and a cell of X. In the Grothendieck
ring C&P we have the equation
[X +4] = [X] + [+4 I] [+4].
(2.44)
Clearly, +4 is homeomorphic to the cone CY over the space

Y = +4 9 S dim 1 L ,
(Y
) = (1)dim (L
).
Hence applying Proposition 2.13 we obtain

euY
euX&3 = euX
euCY
= euX
= euX
1
1 + t(Y
)
1
1+
(1)dim
By Theorem 2.9 we have

=
=
euX&3 (L
) = euX
eu(X =
(L
) t
(L
)
.
dim
1 + (1) (L
) t
The last equation can be rewritten in the form

2 %
B5
9
d
d
dim
(log euX ) =
log
[1 + (1)dim (L
)t](1)
dt
dt
or, equivalently,
(2.45)
euX = C
p(t)
q(t)
where C is a constant. Comparing the free terms we nd that C = 1.

This completes the proof of Theorem 2.3.
CHAPTER 3
Knot Theory of the Robot Arm

The classical knot theory studies subsets K Rn , called knots,
viewed up to the equivalence relation of ambient isotopy. The precise
meaning of the word knot depends on the context; most common
knots are formed by embeddings of spheres (e.g. circles) or disks,
subject to requirements of being smooth, piecewise linear, or locally
at. An isotopy is a one-parameter family of homeomorphisms ht :
Rn Rn which depends continuously on a real parameter t [0, 1]
such that h0 is the identity map Rn Rn . Two subsets K1 , K2 Rn
are ambient isotopic if there exists an isotopy ht : Rn Rn such that
h1 (K1 ) = K2 . One may think of ht as being a continuous deformation of
the whole space Rn ultimately bringing K1 onto K2 . In some situations
knot theory provides invariants of knots and classies their types.
Unknotting theorems of knot theory state that under specic assumptions various knots are all equivalent to each other. One of
the historically rst unknotting theorems was proven by L. Antoine in
1921. It states that there are no nontrivial knots formed by planar
circles or arcs:
Theorem 3.1. For any simple closed curve C R2 there exists an
isotopy ht : R2 R2 taking C onto the standard circle S 1 = D2
R2 . Similarly, any simple arc1 L R2 is ambient isotopic to the
straight line segment.
Surprisingly this result is false for arcs in R3 ; examples of wild
arcs in R3 can be found in [36].
3.1. Can a robot arm be knotted?
In this chapter we consider the following robotical variation of the
knotting problem. A robot arm is a mechanism with hinges at its
vertices and rigid bars at its edges. The hinges can be folded but the
bars must maintain their length and cannot cross. Mechanisms of this
1An
arc is a subset homeomorphic to the interval [0, 1].
62
3. KNOT THEORY OF THE ROBOT ARM
kind appear widely in robotics. Related mathematical problems were

studied in discrete and computational geometry [4], in knot theory [6],
in molecular biology and polymer physics [37].
Let Z = (l1 , . . . , ln1 ) be a xed length vector, where li > 0. Consider
the planar robot arm with length vector Z having no self-intersections.
An admissible conguration p of the arm is given by a sequence of
distinct points p1 , p2 , . . . , pn R2 (positions of the elbows of the arm)
such that
|pi pi+1 | = li , i = 1, . . . , n 1,
the closed line segments [pi , pi+1 ] and [pj , pj+1 ] are disjoint assuming
that |i j| > 1 and besides [pi , pi+1 ] [pi+1 , pi+2 ] = {pi+1 }, see Figure
3.1.
pi
li
pi+1
Figure 3.1. Embedded planar robot arm.

Alternatively, a conguration is uniquely determined by the position of the rst point p1 R2 and by the unit vectors
ui =
pi pi1
,
|pi pi1 |
i = 2, . . . , n
xing the directions of the bars.

We denote by XW the space of all admissible congurations of the
arm. Clearly, XW is an open subset of the product R2 T n1 ; we
emphasize that only sequences p = (p1 , u2 , . . . , un ) R2 T n1 determining non-self-intersecting congurations of the arm belong to the
conguration space XW .
The central problem we study in this chapter is whether the space
XW is path-connected. In other words, we ask if there exists a continuous motion of an arbitrary conguration of the robot arm bringing it
into a straight line segment such that, in the process of the motion, no
self-intersections are created and the lengths of the bars pi pi+1 remain
constant. This question, known also as the carpenters rule problem,
3.1. CAN A ROBOT ARM BE KNOTTED?
63
has a long history. It was mentioned in Kirbys well-known list of

Problems in Low-Dimensional Topology (see [64], Problem 5.18) but it
remained unresolved for years despite the eorts of many mathematicians. The problem is relevant to various topological applications in
molecular biology and in robotics. The answer was found in 2003 by
R. Connelly, E. Demaine and G. Rote [10]:
Theorem 3.2. The conguration space XW is path-connected. Moreover, the factor space XW /G (where G = SE(R2 ) denotes the group of
orientation, preserving isometries of R2 ) is contractible.
An alternative approach was developed by I. Streinu [91]. A recent
survey covering some related problems and generalizations can be found
in [11]. The recent book [16] oers a wealth of additional information.
One may view Theorem 3.2 as an unknotting result for planar robot
arms analogous to the theorem of Antoine mentioned above. Theorem
3.2 becomes false for knots of robotic arms in space. An example of
a knotted arm in R3 is shown in Figure 3.2; see [6], [73] for more
information.
Figure 3.2. Knotted robot arm in R3 .

Similar questions can be asked regarding conguration spaces of
embeddings into R2 of more complicated metric graphs (linkages), see
Figure 3.3. It was recently discovered that trees may knot, i.e., the
space of embeddings of a metric tree into R2 can be disconnected (see
for example [5]). Here we consider metric embeddings, i.e., such that
each bar of the tree is mapped onto a planar segment of length equal
to the length of the bar.
In this chapter we describe a proof of Theorem 3.2 following essentially the original paper [10]; we only suggest an improvement of the
last argument of [10] leading to the construction of a global motion,
see 3.7.
64
Figure 3.3. Two embeddings of a metric tree into R2

which cannot be continuously deformed one into another.
3.2. Expansive motions
A deformation of a planar robot arm is a family of continuous functions
p1 (t), . . . , pn (t) R2 of a real parameter t [0, 1] such that |pi (t)
pi+1 (t)| = li for any t and no self-intersections happen during the motion. This last condition is the main source of diculty of the problem.
To deal with it we introduce a class of motions p1 (t), . . . , pn (t) R2
such that the distances |pi (t) pj (t)| between any pair of vertices
is nondecreasing. Such motions are called expansive. In this section we show that self-intersections do not occur in expansive motions p1 (t), . . . , pn (t) under the condition that the initial conguration
p1 (0), . . . , pn (0) is self-intersection free.
Consider a smooth deformation p1 (t), . . . , pn (t) R2 and velocities
vi = p i (0) of the vertices of the arm. Since the distance between pi
and pi+1 is constant, we obtain
(3.1)
+vi+1 vi , pi+1 pi 4 = 0.
This is a consequence of the fact that the derivative of |pi+1 (t) pi (t)|2
equals twice the scalar product 2 +p i+1 (t) p i (t), pi+1 (t) pi (t)4. The
velocity vector w of a point c lying on the bar pi pi+1 is given by
(3.2)
w=
|c pi |
|c pi+1 |
vi +
vi+1 .
|pi pi+1 |
|pi pi+1 |
Lemma 3.3. Let p1 , p2 , p3 R2 be three pairwise distinct points,

and let v1 , v2 , v3 R2 be vectors (describing velocities of the points
3.2. EXPANSIVE MOTIONS
65
p1 , p2 , p3 ) satisfying
(3.3)
(3.4)
(3.5)
+v2 v1 , p2 p1 4 = 0,
+v3 v2 , p3 p2 4 0,
+v3 v1 , p3 p1 4 0.
For a point c of the open interval (p1 p2 ) set

w=
|c p2 |
|c p1 |
v1 +
v2 .
|p1 p2 |
|p2 p1 |
(This is the velocity at point c if we view the interval [p1 , p2 ] as a rigid

body whose end points move with velocities v1 and v2 correspondingly,
see (3.2)). Then
(3.6)
+v3 w, p3 c4 0.
Inequality (3.6) is strict if at least one of the inequalities (3.4) or (3.5)

is strict.
v3
p3
p1
v1
c
v2
p2
Figure 3.4. Velocities of points of a triangle.

The statement of Lemma 3.3 is illustrated by Figure 3.4. It can be
also expressed as follows: if the vertices of a triangle move such that
the distance |p2 p1 | remains constant and the distances |p3 p1 |,
|p3 p2 | are nondecreasing, then the distance |p3 c| is nondecreasing
as well.
Proof. Without loss of generality we may assume that
v 1 = 0 = v2 ,
i.e., the bar [p1 , p2 ] is motionless (and hence w = 0) and only the point
p3 moves.
66
First we assume that p3 p1 and p3 p2 are linearly independent.

Let (p3 p2 ) denote a vector which is perpendicular to p3 p2 and
such that
+(p3 p2 ) , p3 p1 4 > 0.
Similarly, denote by (p3 p1 ) a vector which is perpendicular to
p3 p1 and such that +(p3 p1 ) , p3 p2 4 > 0. It follows from (3.4)
and (3.5) that v3 lies inside the cone spanned by the vectors (p3 p1 )
and (p3 p2 ) and hence one can write
v3 = (p3 p1 ) + (p3 p2 ) ,
0, 0.
On the other hand one has

p3 c =
|p1 c|
|p2 c|
(p3 p2 ) +
(p3 p1 ).
|p1 p2 |
|p2 p1 |
Therefore the scalar product +v3 , p3 c4 equals

|p1 c|
+(p3 p1 ) , p3 p2 4
|p1 p2 |
|p2 c|
+
+(p3 p2 ) , p3 p1 4 0.
|p2 p1 |
Note in the last formula the coecients of both and are positive.
Finally, if one of the inequalities (3.4) or (3.5) is strict, then either
> 0 or > 0 and hence +v3 , p3 c4 > 0.
Now consider the remaining case when p3 p1 and p3 p2 are linearly dependent, i.e., the points p1 , p2 , p3 lie on a straight line. There
are two subcases: (a) p3 lies between p1 and p2 and (b) p3
/ [p1 , p2 ].
Assuming (a), we see that (3.6) is obvious, however none of (3.4), (3.5),
(3.6) can be satised with a strong inequality.
In the case (b) the statement of Lemma 3.3 is obvious.
!
Corollary 3.4. For any expansive motion, the distance between any
pair of points of the robot arm is nondecreasing. In particular, no selfintersection may occur in an expansive motion p1 (t), . . . , pn (t) starting
with a conguration p1 (0), . . . , pn (0) without self-intersections.
Proof. Consider a point c lying on the bar p1 p2 and its distance to
another vertex p3 (see Figure 3.5, left). We know that p3 moves such
that the distances from p3 to p1 and p2 are nondecreasing. By Lemma
3.3, the distance from p3 to c is nondecreasing as well.
3.3. INFINITESIMAL MOTIONS
p3
p1
p4
p2 p1
67
p5
p2
Figure 3.5. Mutual distances under an expansive motion.

Next, consider two internal points c and d of bars p1 p2 and p4 p5 ,
see Figure 3.5, right. We observe that the distances p4 c and p5 c are
nondecreasing (by the above statement), hence applying Lemma 3.3 to
triangle p4 p5 c, we obtain that the distance cd is nondecreasing.
!
3.3. Innitesimal motions
Given an expansive motion p1 (t), . . . , pn (t), consider the velocity vectors vi = p(i (0) describing the velocities of motion of the vertices at
time t = 0. The set of velocity vectors v1 , . . . , vn is called an innitesimal motion. Which innitesimal motions correspond to expansive
motions? Since the length |pi+1 (t) pi (t)| must be constant, we have
(3.7)
+vi+1 vi , pi+1 pi 4 = 0,
where pj = pj (0). On the other hand, the distance |pi (t) pj (t)| must
be nondecreasing in t for |i j| > 1 and hence
(3.8)
+vi vj , pi pj 4 0,
for |i j| > 1.
In general, it is unrealistic to require that a strong inequality holds

in all inequalities (3.8) since a part of the robot arm between points
pi and pj may be a straight line segment. For example, assume that
pi+1 pi and pi+2 pi+1 are parallel. Then (vi+1 vi ) (pi+1 pi )
and (vi+2 vi+1 ) (pi+2 pi+1 ) imply (vi+2 vi ) (pi+2 pi ).
Definition 3.5. Let p1 , . . . , pn R2 be a conguration of the robot arm. An innitesimal motion v1 , . . . , vn is called expansive if it
satises (3.7) and for any pair of indices i < j such that the points
pi , pi+1 , . . . , pj are not on a straight line, one has
(3.9)
+vj vi , pj pi 4 > 0.
One of the key steps in the proof of Theorem 3.2 is the following
statement [10].
68
Theorem 3.6. For any admissible conguration p1 , . . . , pn R2 of the

robot arm there exists an innitesimal expansive motion v1 , . . . , vn .
Note that an innitesimal expansive motion v1 , . . . , vn is automatically not identically zero if p1 , . . . , pn is not a straight line conguration: one of the strict inequalities (3.9) will be satised.
3.4. Struts and equilibrium stresses
Now we reformulate the problem by introducing struts segments of a
new type compared to bars. In contrast to bars, which must have constant length during the motion, struts are permitted to increase their
length or to stay the same length but are not allowed to shorten. In an
admissible conguration p1 , . . . , pn R2 where each pi is connected
to pi+1 by a bar, we connect by a strut any pair of vertices pi and pj
such that the part of the arm between pi and pj is not a straight line
segment. We obtain a planar framework of the type shown on Figure
3.6 (right), where the struts are indicated by the dotted lines.
Figure 3.6. The initial conguration (left) and the planar framework obtained by adding struts (right).
Let us denote by B the set of bars and by S the set of struts. We
will write [i, j] S or [i, j] B to indicate that the segment connecting
pi and pj is a strut or a bar, correspondingly. Here [i, j] denotes an
unordered pair of indices. The conditions describing expansive innitesimal motions are:
+vi vj , pi pj 4 = 0,
if [i, j] B,
(3.10)
+vi vj , pi pj 4 > 0,
if [i, j] S.
3.4. STRUTS AND EQUILIBRIUM STRESSES
69
This is a system of linear equations and inequalities and Theorem 3.6

essentially states that it always has a solution.
At this point we will invoke the following well-known criterion for
existence of solutions of systems of linear equations and inequalities.
The following statement is a special case of Theorem 22.2 from [83].
Theorem 3.7. Let ai RN for i = 1, . . . , m and let k be an integer,
1 k m. Then one and only one of the following two alternatives
holds:
(a) There exists a vector x RN such that
+ai , x4 > 0
for
i = 1, . . . , k,
+ai , x4 = 0
for i = k + 1, . . . , m.
(b) There exist real numbers 1 , . . . , m such that

m
=
i ai = 0,
i=1
where each of the numbers 1 , . . . , k is non-negative and at least one

of them is positive.
We intend to apply this theorem to decide whether the system of
equations and inequalities (3.10) admits a solution. As the space RN
we take the direct sum RN = R2 R2 of n copies of the plane
R2 . For any pair i, j = 1, . . . , n with i < j we denote by aij RN the
vector which has i-th component pi pj and j-th component pj pi
and all other components are zero. The vector
x = v 1 v n RN
must then satisfy
(3.11)
(3.12)
+aij , x4 > 0,
+aij , x4 = 0,
if [i, j] S,
if [i, j] B.
By Theorem 3.7, this system has a solution if and only if there does
not exist a function assigning a real number
ij = ji ,
i == j
to any bar or strut such that:

(a) numbers associated to struts [i, j] S are non-negative, ij 0,
(b) at least one of the numbers ij associated to those struts is
positive,
70
(c) for any i = 1, . . . , n the sum

=
(3.13)
ij (pj pi ) = 0 R2
j; [i,j]BS
vanishes. In the last sum the summation is over all indices j such that
the pair [i, j] belongs to B S.
There is a useful mechanical interpretation of the weights ij . One
can view ij as the stresses induced in the bar or strut [i, j] B S.
A negative stress means that the edge is pushing on its endpoints by
an equal amount, a positive stress means that that the edge is pulling
on its endpoints by an equal amount, and zero means that the edge
induces no force. The major condition (3.13) can now be interpreted
by saying that the total force applied to any vertex equates to zero.
For that reason a stress
(3.14)
[i, j] 7 ij ,
[i, j] B S
satisfying (3.13) is called an equilibrium stress.

A stress (3.14) is called proper if ij 0 for any strut [i, j] S.
Summarizing the above discussion we obtain that Theorem 3.6 is
equivalent to the following statement:
Theorem 3.8. Let p1 , . . . , pn R2 be an admissible planar conguration of the robot arm. Then the planar network obtained from
the conguration by adding struts admits no proper equilibrium stress
[i, j]
7 ij such that ij > 0 is positive for at least one strut [i, j] S.
--1
--1
1
--1
--1
Figure 3.7. Equilibrium stress on a bar and strut graph

(left). No proper nonzero equilibrium stress exists on the
bar and strut graph shown on the right.
We illustrate Theorem 3.8 by Figure 3.7. On the left picture we see
an equilibrium stress on a bar and strut graph. On the right picture
no proper equilibrium stress exists, as can be easily understood by
examining forces near the end points of the robot arm.
3.4. STRUTS AND EQUILIBRIUM STRESSES
71
The next step is called planarization, its goal is to reduce the

problem of existence of equilibrium stress on a bar and strut framework
to a similar problem for a planar graph, i.e., such that any two edges
meet only at a vertex. This operation is illustrated by Figure 3.8. We
simply add new vertices at all crossing points.
Figure 3.8. The planar framework obtained from the

bar and strut graph of Figure 3.6 (right).
Suppose that the initial graph admits an equilibrium stress ij .
We claim that the obtained planar framework admits an equilibrium
stress as well. Indeed, suppose that an edge [pi , pj ] of the initial graph
having the stress ij is subdivided into a number of smaller edges, then
each of the new edges [p(k , p(l ] gets stress
(3.15)
(
kl
= ij
|pi pj |
.
|p(k p(l |
The equilibrium condition (3.13) remains preserved at all vertices of

(old and new) since the actual force is obtained by multiplying the
stress to the length of the edge, and the forces corresponding to two
subintervals cancel each other at newly created vertices.
It may happen that in the obtained planar framework several edges
overlap. In this case we add new vertices as explained above and merge
the multiple edges. The stress ( corresponding to multiple edges equals
the sum of the initial stresses of the merged edges. Clearly, this operation also preserves the equilibrium condition (3.13).
When doing planarization we make the following convention. Splitting a bar into several edges produces bars, and splitting a strut produces struts. When we merge several edges, the resulting edge will be
a strut if all merged edges were struts. The result is a bar if at least
one of the merged edges is a bar.
72
If the initial stress was proper, i.e., the struts had non-negative
stresses, then the obtained stress ( is proper as well.
Suppose that the initial graph had a strut with a positive stress.
After performing the operations of subdivision and merging edges, a
part of this edge remains to be a strut with a positive stress.
3.5. Maxwell Cremona Theorem
As the result of planarization described in the previous section we obtain a convex polygonal domain Y R2 subdivided into nitely many
polyhedral cells, see Figure 3.8. The exterior of Y in R2 will be denoted
by F0 . We consider continuous functions f : R2 R which are ane
when restricted onto every cell F of Y or onto the exterior F0 . In other
words,
(3.16)
f (x) = +aF , x4 + bF ,
xF
where F denotes either F0 or a cell of Y . If F and F ( are adjacent cells

and e is the edge separating them, then
+aF , x4 + bF = +aF ! , x4 + bF !
for x e. Dierentiating both sides of the above equation in the direction of e we obtain that the vector aF aF ! must be perpendicular to
e. Hence we may write
(3.17)
aF ! aF = e e ,
e R,
where e denotes the vector perpendicular to e, having the same length

as e and pointing from F towards F ( .
e
F
e
Figure 3.9. Stress determined by a piecewise ane

function f : R2 R.
We obtain a stress function
e
7 e
3.5. MAXWELL CREMONA THEOREM
73
assigning real valued weights2 to the edges of Y . Note that the order of
F and F ( is irrelevant: if the roles of F and F ( are interchanged, then
the vector e reverses and the value e remains unchanged.
Clearly, e is the jump of the slope of f when crossing the edge e
along the line perpendicular to the edge e. The graph of f is a piecewise
linear surface with edges lying above the edges of . One may write
e = tan(e )
where e is the dihedral external angle between the faces of the graph
of f lying above F and F ( , see Figure 3.10. Edges of having positive
stresses are called valleys and edges having negative stresses are called
mountains.
e >0
F
Valley
e <0
Mountain
Figure 3.10. Geometric interpretation of the stress determined by a piecewise ane function.
The following statement is known as the Maxwell Cremona theorem; it was known to James Clerk Maxwell and Luigi Cremona in the
nineteenth century.
Theorem 3.9. (i) The stress function e
7 e dened by the equation
(3.17) is an equilibrium stress on the planar framework . (ii) Any
equilibrium stress on can be realized by a continuous piecewise ane
function f : R2 R which is ane when restricted to any cell of Y
and on the exterior of Y . Such f is unique up to addition of an ane
function.
Proof. Consider a vertex p and the cells F1 , . . . , Fk incident to
it; one of the Fi s can be the exterior F0 of Y . We will assume that
F1 , . . . , Fk appear in cyclic order in the anticlockwise direction. Let ei
2Our
sign conventions here dier from the ones used in [10].
74
denote the edge separating Fi and Fi+1 . Assume that ei connects p to

a vertex qi where i = 1, . . . , k, i.e., ei = qi p. Then
aFi+1 aFi = ei e
i ,
(3.18)
where e
i is the vector perpendicular to ei satisfying |ei | = |ei | and
pointing from Fi to Fi+1 . In other words, ei = (qi p) is obtained

by rotating ei 90 degrees in the anticlockwise direction. Adding all
equations (3.18) we obtain
2 k
5
k
=
=
ei (qi p) =
ei (qi p)
= 0
i=1
i=1
which is equivalent to the equilibrium condition (3.13). This completes

the proof of (i).
To prove (ii) consider an equilibrium stress e 7 e on . We want
to assign a vector aF R2 to each cell of Y and to the exterior F0 of Y
such that (3.17) holds. Fix aF0 R2 arbitrarily. Any sequence of cells
F1 , . . . , Fk such that each pair of neighbors Fi and Fi+1 have a common
edge allows us to dene (using (3.17)) the vector aFk if the initial vector
aF1 is given. Choosing a dierent path from F1 to Fk gives the same
value aFk as follows from the equilibrium condition (3.13). Indeed,
the equilibrium condition says that the total increment obtained while
surrounding a vertex equals zero. Starting from F0 we may nd a
sequence of cells as above ending at any given cell F , which allows us
to determine the vectors aF for all cells F in a unique way such that
the compatibility condition (3.18) is satised.
Next we want to choose constants bF such that the ane functions
dened on dierent cells by +aF , x4 + bF form a continuous function
R2 R. In other words we require that the functions corresponding
to dierent cells coincide on their intersections.
If we have two adjacent cells F and F ( and we have already xed the
constant bF , then there is a unique value of bF ! such that the resulting
function is continuous on the union F F ( . If the constant terms bF and
bF ! are chosen arbitrarily, then the dierence +aF , x4+bF +aF ! , x4bF !
is a constant function on F F ( .
Suppose that we have a sequence of cells F1 , F2 , . . . , Fk (listed in
cyclic order) which are incident to a vertex p, see Figure 3.11. Given
the value of bF1 , one may use this sequence to determine subsequently
the constants bFi , where i = 1, . . . , k. Potentially there may appear a
discontinuity along the last edge ek separating Fk and F1 . We claim
that this discontinuity does not happen. Indeed, the dierence between
the two functions on the edge ek (the restrictions from F1 and from Fk )
3.6. THE MAIN ARGUMENT
75
e1
F2
F1
p
ek
Fk
Figure 3.11. Adjacent cells surrounding a vertex.

is (a) a constant function; (b) it vanishes at the vertex p. Hence the
dierence is identically zero. This fully proves statement (ii) since we
may x the constant bF0 in an arbitrary way.
!
3.6. The main argument
Our next goal is to prove the following theorem:
Theorem 3.10. Consider a planar convex polygon Y subdivided into
nitely many polyhedral cells and let R2 denote the 1-skeleton of
the subdivision. Let be a polyhedral arc (such as the one shown
on gure below in bold). The edges lying in are called bars, and
the edges lying in the complement of are called struts. Then there
does not exist a stress function assigning weights e R to edges e
of satisfying the following three properties:
(a) the stress of any strut e is non-negative e 0;
(b) for at least one strut e > 0;
(c) for any vertex p one has
=
(3.19)
e (q p) = 0 R2 .
q
In (3.19) q runs over all vertices of incident to p and e denotes the

edge e = [p, q].
The proof of Theorem 3.10 will be based on

the following lemmas.
Suppose that such a stress function e 7 e
exists. By the Maxwell Cremona theorem
we may construct a continuous piecewise ane
function f : R2 R which is ane on the cells
of Y which is identically zero outside Y and such
that the stress determined by f equals .
76
Lemma 3.11. Any mountain in the graph of f projects onto a bar e

of the framework .
Proof. Indeed, a strut has a non-negative stress and therefore over
any strut a at or a valley of the graph lies. Hence all mountains must
project onto bars.
!
Denote by M the set of points where f achieves its maximum. Note
that M may have several connected components.
Lemma 3.12. Let v be a vertex on the boundary of M , and let b1 , . . . , bk
be the bars incident to v, in cyclic order. Consider a small disk D
around v.
(1) If there is an angle of at least at v between two consecutive
bars, say bi and bi+1 , then the pie wedge P of D bounded by bi
and bi+1 belongs to M (see Figure 3.12).
(2) If there are no bars or only one bar incident to v, i.e., k 1,
then the entire disk D belongs to M .
b1
c
b2
b3
b2
b1
v
b3
Figure 3.12. Illustration of Lemma 3.12. Solid lines

are bars, dotted lines are struts, the shaded pie wedge P
is contained in M .
Proof. (1) Edges lying in the boundary M must be mountains, and
hence they must project onto bars, by Lemma 3.11. Since there are no
bars in the pie wedge P , we obtain that P must be either disjoint from
M or must be completely contained in M . Assuming that P is disjoint
from M , consider the level set f = m i (where m is the maximal
value of f and i > 0 is small) in P . It is a star-shaped polygonal arc
around v (see Figure 3.12, right) starting at a point of bi and ending
at a point of bi+1 . Convex vertices of this arc (denoted by c on Figure
3.7. GLOBAL MOTION
77
3.12, right) lie on mountains emanating from v and reex vertices of

the arc (denoted r on Figure 3.12) lie on valleys emanating from v.
Since the angle of the pie wedge is at least , the arc must contain at
least one convex vertex in P . By Lemma 3.11 there must be a bar in
P , a contradiction.
Statement (2) is a special case of (1) and follows from the same
argument.
!
Now we may complete the proof of Theorem 3.10. If the equilibrium
stress with properties described in the statement of Theorem 3.10 exists, then the boundary
of any connected component of the maximum
region M R2 for the corresponding piecewise
ane function f : R2 R has the following
property: every convex point v M has at
least three incident bars, as follows from Lemma 3.12. However we
know that the set of bars is an arc and hence each vertex must
have at most two incident bars. This excludes many possibilities for
f , for example we see that f may not have an isolated maximum as
shown on the picture.
We obtain that M must have no convex
points (by Lemma 3.12). Therefore M must
coincide with the exterior of a closed polygonal
M
curve. But then the boundary M is a closed cycle consisting entirely of bars which contradicts
our assumption concerning .
These arguments prove Theorem 3.8. They also prove Theorem 3.6
which is equivalent to Theorem 3.8 as was mentioned earlier.
3.7. Global motion

In this section we complete the proof of Theorem 3.2. First we recall
our notation. Given a length vector Z = (l1 , . . . , ln1 ), we consider the
space XW R2 T n1 of all congurations of the planar robot arm
having n 1 bars of length li without self-intersections (see Figure
W the factor-space XW /G where G is the group of
3.1). We denote by X
W represent dierent
orientation-preserving isometries of R2 . Points of X
W is
shapes of the arm. Our main goal is to show that the space X
contractible, as Theorem 3.2 states.
W as an open subset of the torus T n2 . An emWe will view X
n2
bedding XW T
is given by assigning to a conguration the set
78
of angles which the bars of the arm make with the rst bar. Alternatively, we may think that the rst bar [p1 p2 ] is pinned at points
W parameterizes the variety of all
p1 = (0, 0) and p2 = (l1 , 0); then X
possible positions of p3 , . . . , pn which are determined by all possible
angles 3 , 4 , . . . , n with the x-axis, see Figure 3.13. Given a congp3
pn
n
p2
p1
pn-1
Figure 3.13. Planar robot arm.
uration p XW represented by the sequence p1 = (0, 0), p2 = (l1 , 0),

W at p is a sequence of velocity
p3 , . . . , pn R2 , a tangent vector V to X
vectors v1 , v2 , . . . , vn satisfying v1 = 0 = v2 and equations (3.7). It
makes sense to speak of expansive tangent vectors
W ),
V = (v3 , . . . , vn ) Tp(X
vi Tpi (R2 ),
having in mind those which satisfy the conditions of Denition 3.5.

Our goal is to prove the following theorem:
Theorem 3.13. There exists a continuous tangent vector eld V on
the torus T n2 with the following properties:
W T n2 the tangent vector Vp is expansive;
(a) for any p X
(b) the vector Vp Tp(T n2 ) vanishes if and only if p T n2 is
either the straight line conguration (see Figure 3.14) or is a conguration having self-intersections;
W is C -smooth.
(c) The restriction V |X
l1
p1
l2
p2
ln-1
pn-1
pn
Figure 3.14. Straight line conguration.

Proof. First we show that such an expansive vector eld v exists in
W . Theorem 3.6
a small open neighborhood of any conguration p X
guarantees that an expansive tangent vector Vp exists for any admissible
W . Now we want to nd a continuous family of such
conguration p X
W.
vectors Vq where q varies in a neighborhood of p X
3.7. GLOBAL MOTION
79
One may parameterize expansive tangent vectors as follows. Let

p = (p1 , p2 , . . . , pn ) be an admissible conguration of the robot arm
and let V = (v1 , v2 , . . . , vn ) be an expansive tangent vector V
W ). Equation (3.7) is equivalent to
Tp(X
(3.20)
vi+1 vi = i+1 (pi+1 pi ) ,
i+1 R.
Here for a planar vector v R2 we denote by v the vector obtained

from v by rotating it 90 degrees in the anticlockwise direction.
Geometrically, i has the meaning of the angular velocity of rotation
of pi about the previous point pi1 . One may also write
i = i
(3.21)
in term of the angles i shown on Figure 3.13. Since v1 = 0 = v2 we

have
(3.22)
vi =
i
=
r (pr pr1 ) ,
i = 3, . . . , n.
r=3
We formally set 1 = 2 = 0. The numbers 3 , . . . , n parameterize

tangent vectors V = (v1 , . . . , vn ) satisfying (3.7). Inequality (3.9) can
now be expressed in the form
(3.23)
j
=
r det(pj pi , pr pr1 ) > 0
r=i+1
under the assumption that not all numbers det(pj pi , pr pr1 )

vanish for r = i + 1, . . . , j. Here for two planar vectors p, q R2 we
denote
det(p, q) = +p, q 4,
the scalar product of p and q . Inequality (3.23) can be rewritten as
(3.24)
j
=
r det(ps ps1 , pr pr1 ) > 0.
r, s=i+1
If k denotes the angle between pk pk1 and the x-axis (as shown on
Figure 3.13) then one has
det(ps ps1 , pr pr1 ) = ls lr sin(s r )
and inequality (3.24) turns into
=
(3.25)
lr ls (r s ) sin(s r ) > 0.
i+1r<sj
80
Theorem 3.6 states that for any admissible conguration (3 , . . . , n )

there exist numbers 1 , . . . , n such that 1 = 2 = 0 and (3.25) holds
for any i + 1 < j and r == r+1 for some i + 1 r < j (i.e., such that
the points pi , pi+1 , . . . , pj do not lie on a line).
Now, suppose that p = (p1 , . . . , pn ) is a conguration without
straight pieces, i.e., r == r+1 . Assume that a vector (0, 0, 3 , . . . , n )
represents an innitesimal expansive motion at p. Then the same vector represents an innitesimal expansive motion at any conguration
q which is suciently close to p, as follows from (3.23). This clearly
gives a smooth expansive vector eld dened in a neighborhood of p.
This simple argument fails to work near congurations which have
some adjacent collinear edges. Indeed, if the points pi+1 , pi+2 , . . . , pj
lie on a straight line, then i+1 = i+2 = = j and the LHS of (3.25)
is identically zero. However, we may not guarantee that the LHS of
(3.25) will be positive at congurations close to p = (p1 , . . . , pn ) which
are not collinear between pi+1 and pj .
Consider now the general case when the conguration in question
p = (p1 , . . . , pn ) has collinear arcs as shown on Figure 3.15. For a
sequence of indices k0 = 1 < k1 < < km < km+1 = n the points
pk , pk +1 , . . . , pk+1 1 , pk+1
lie on a straight line for = 0, 1, . . . , m while points
pk 1 , pk , pk +1
are not collinear. The corresponding sequence of angles (3 , 4 , . . . , n )
pn
p k2
p1
p k1
p km
Figure 3.15. Robot arm with straight line segments.

satises
(3.26)
k +1 = k +2 = = k+1 .
One may view this conguration as a robot arm with fewer vertices
p1 , pk1 , . . . , pkm , pn and apply to it Theorem 3.6. We obtain that there
3.7. GLOBAL MOTION
81
W ) repreexists an expansive tangent vector V = (v1 , . . . , vn ) Tp(X

sented by a sequence of real numbers (1 , . . . , n ) having the following
additional property:
(3.27)
i = j
where k < i, j k+1 .
This is obvious from the interpretation of numbers i as angular velocities i (see above) since all subdivided parts of a segment move with
the same velocity when viewed as parts of the integrated arm.
W near
Consider now admissible congurations q = (q1 , . . . , qn ) X
p. Each such conguration can be also described by a sequence of angles
(3 , 4 , . . . , n ). A tangent vector eld V near p will be described by
i =
i (3 , . . . , n ), where i =
a sequence of C -smooth functions
3, . . . , n. We set
(3.28)
i = i i + i ,
i = 3, . . . , n.
Let us show that condition (3.25) is satised for q lying suciently

close to p, i.e., when i is close to i for all i. First we see that
i = i for q = p and hence (3.25) holds for i = i . Given two
indices i + 1 < j such that the part of p is not straight between i and
j, then inequality (3.25) is satised and hence it also holds for close
congurations q. Finally consider the case when the part of p between
i and j is collinear. Then using (3.27) and (3.28) we nd
=
r
s ) sin(s r )
lr ls (
i+1r<sj
lr ls (s r ) sin(s r ).
i+1r<sj
Clearly there exists i > 0 such that for all x == 0, |x| < i one has
x sin x > 0. Hence, each term in the above sum is non-negative and
the sum vanishes if and only if the segment of conguration q between
the vertices qi and qj is collinear. This proves that V is an expansive
tangent vector eld in a small open neighborhood of p.
Now, if locally the elds VU are constructed on an open covering
W T n2 , one takes a smooth partition of unity {U } subor{U } of X
dinate to this covering and denes
=
(3.29)
V =
U VU .
U
W T n2 .
This gives the required eld in X
82
The next step is to extend it to a vector eld on the whole torus

T
. We may assume that each of the local elds VU is bounded by
a common constant C (note that tangent vectors are elements of the
Lie algebra of T n2 , which is Rn2 ). This can be achieved by scaling
(multiplying by a small constant). Then the global eld (3.29) will be
also bounded by C. Find a smooth function : T n2 R such that
W.
0 and the set of zeros of is exactly the complement T n2 X
n2
Then the product w = V is a continuous vector eld on T
satisfying our requirements.
This completes the proof of Theorem 3.13.
!
n2
Proof of Theorem 3.2. Now we are able to complete the proof of

the main theorem of this chapter Theorem 3.2.
Consider the vector eld V on the torus T n2 given by Theorem
W T n2 .
3.13. We know that V vanishes on the complement of X
W is the straight line conguration. For
Besides, Vp0 = 0 where p0 X
any p XW with p == p0 the vector Vp is nonzero and is expansive. The

W
eld V is smooth everywhere except possibly on the boundary of X
where it is only continuous.
W , consider a solution x(p, t) T n2 of the initial value
Given p X
problem
dx
= Vx ,
dt
(3.30)
x(p, 0) = p,
t 0.
Here we use the well-known general theorem about existence of solutions of ordinary dierential equations with continuous right-hand side.
One may view x(p, t) as a continuous process of modication of the robot arm starting with the initial conguration p for t = 0. Our aim is
to show that x(p, t) converges to p0 for t .
Lemma 3.14. Let p1 , p2 , . . . be a sequence of admissible congurations,
W , and let tn > 0 be a sequence of real numbers. If the limits
pn X
lim x(pn , tn ) = q T n2 ,
lim pn = p T n2
W , then q is an admisexist and p is an admissible conguration, p X
sible conguration as well, i.e., q XW .

Proof. Any conguration p = (p1 , . . . , pn ) determines a continuous
map sp : [0, L] R2 dened as follows. Here L = l1 + + ln1 is the
3.7. GLOBAL MOTION
83
total length of the robot arm and for

r1
=
i=1
li a
r
=
li
i=1
the point sp(a) lies on the segment [pr , pr+1 ] and satises
|sp(a) pr | = a
r1
=
li .
i=1
Clearly, sp is analogous to the arc-length parametrization of smooth

curves.
W T n2 can be characterized
The set of admissible congurations X
as the set of congurations p for which sp is injective. For two distinct
numbers a, b [0, L] dene a function Fa,b : T n2 R given by
Fa,b (p) = |sp(a) sp(b)|.
W if and only if Fa,b (p) > 0 for all
A conguration p T n2 lies in X
a == b.
To prove Lemma 3.14 we assume the contrary, i.e., that q is not
admissible. Then Fa,b (q) = 0 for some a < b. We know that Fa,b (p) =
c > 0 is positive and hence Fa,b (pn ) > c/2 for all suciently large n.
Therefore, by Corollary 3.4,
Fa,b (x(pn , tn )) Fa,b (pn ) > c/2
for all large n. This implies, by passing to the limit, that
lim Fa,b (x(pn , tn )) = Fa,b (q) c/2,
a contradiction.
W the solution x(p, t) stays in the domain

Corollary 3.15. For p X
n2
W T
X
and, moreover, any limit point of the set {x(p, t), t 0}
belongs to XW .
W is smooth and therefore (invoking a wellThe vector eld V |X
known result from the theory of ordinary dierential equations) we
conclude now that the solution x(p, t) of the dierential equation (3.30)
is smooth as a function of p and t.
W one has
Let us now show that for any initial conguration p X
(3.31)
lim x(p, t) = p0 ,
84
W denotes the straight line conguration. In other words,

where p0 X
we claim that solution of the dierential equation (3.30) implements
the full straightening of the robot arm. Consider the smooth function
F : T n2 R given by
=
F (p1 , . . . , pn ) =
|pi pj |2 .
i+1<j
W has a unique critical point p = p0 (the maxiThe restriction F |X

W there exists an
mum). Indeed, for any admissible conguration p X
n2
expansive tangent vector Vp Tp(T
) and the derivative Vp(F ) > 0
is positive assuming that p == p0 since at least one of the inequalities
(3.9) will be satised.
Consider a limit point q of the trajectory {x(p, t); t 0}. We
W (by Lemma 3.14). Assume that q == p0 . Then
know that q X
Vq(F ) = i > 0.
We claim that x(p, t) converges to q as t . If this statement is
false, then there exists an open neighborhood U of q such that for any
t > 0 there is T > t such that x(p, T )
/ U . We may assume that U
(
is so small that for any q U one has Vq! (F ) > i/2. Find a smaller
neighborhood U0 U of q such that the distance between U0 and the
complement T n2 U is > 0 positive. Let C > 0 be such that
|Vp| C for all p T n2 (here we use the standard trivialization of
the tangent bundle of the torus T n2 ). There is a sequence tn
such that x(p, tn ) U0 (since q is a limit point) and a sequence Tn > tn
such that x(p, Tn )
/ U . We may assume without loss of generality that
Tn < tn+1 . Then one has
*
<
Tn
tn
*
|x|dt
Tn
tn
|Vx(p,t) |dt C (Tn tn ),
i.e., Tn tn > /C. Hence we obtain that

*
F (x(p, Tn )) F (x(p, tn )) =
Tn
tn
Vx(p,t) (F )dt > i/2 /C
and therefore
F (x(p, tn+1 )) F (x(p, Tn )) F (x(p, tn )) +
i
,
2C
a contradiction (since F is bounded above). This proves that x(p, t)

converges to q as t .
3.7. GLOBAL MOTION
85
Now suppose that t is such that x(p, ) U for all t. Using the
Mean Value Theorem we nd some satisfying t such that
F (x(p, )) F (x(p, t)) = Vx(p,) (F ) ( t) > i/2 ( t)
which again contradicts the fact that F is bounded. Hence we obtain
a contradiction to our initial assumption that q == p0 .
This argument proves that q = p0 is the only limit point of the
trajectory {x(p, t); t 0} and hence for any admissible conguration
W the trajectory x(p, t) converges to p0 as t .
pX
W is contractible. Dene a deformaOur nal goal is to show that X
tion
W [0, 1] X
W
H:X
by the formula
;
<
W , [0, 1)
H(p, ) = x p,
, for p X
1
and
H(p, 1) = p0 .
We claim that H is continuous, which is equivalent to the following
statement: For any neighborhood U of p0 there exists a neighborhood
V of p and a real number T such that for any t T and q V one
has x(q, t) U .
The negation of this statement is equivalent to the existence of
W converging to p and a
a neighborhood U of p0 , a sequence pn X
sequence tn such that x(pn , tn )
/ U . Set
S = {x(pn , t); n N, t 0}.
W . By our assumptions,
By Lemma 3.7 the closure S is contained in X
1
there is i > 0 such that S F ([0, L i]).

W the function x 7 Vx (F ) vanishes only for
We know that for x X
x = p0 . Hence we obtain that for all x S one has Vx (F ) > 0 and
therefore, using compactness, there exists c > 0 such that Vx (F ) c
for all x S.
We obtain that
F (x(pn , tn )) F (x(pn , 0)) = Vx(pn ,) (F ) tn c tn .
This implies that the sequence F (x(pn , tn )) is unbounded, a contradiction.
!
CHAPTER 4
Navigational Complexity of Conguration Spaces

In this chapter we discuss topological invariants which arise in connection with the problem of construction and design of robot motion
planning algorithms. Given a mechanical system, a motion planning
algorithm is a function which assigns to any ordered pair of states of the
system (A, B), where A is the initial state and B is the desired state,
a continuous motion of the system starting at the state A and ending
at the state B. Such an algorithm allows the system to function in an
autonomous regime. A survey of algorithmic motion planning can be
found in [87]; see also the textbook [67].
In this chapter we study the topological invariant TC(X) introduced
originally in [22], see also [23] and [27]. It is a numerical homotopy
invariant inspired by problems of robotics, quite similar in spirit to
the classical Lusternik Schnirelmann category cat(X). Intuitively,
TC(X) is a measure of the navigational complexity of X viewed as the
conguration space of a system. Knowing TC(X) allows us to predict
instabilities in the behavior of the system and is helpful in practical
situations while constructing motion planning algorithms for real life
machines.
Formally TC(X), as well as cat(X), are special cases of the notion
of genus of a bration introduced by A. Schwarz [93]. In 19871988 S.
Smale [88] and V.A. Vassiliev [99] applied the notion of Schwarz genus
to study complexity of algorithms for solving polynomial equations.
Our discussion of TC(X) brings into the realm of algebraic topology
a broad variety of important engineering applications and gives a new
motivation for studying the concept of Schwarz genus.
4.1. Motion planning algorithms
Let X denote the conguration space of a mechanical system. States
of the system are represented by points of X, and continuous motions
of the system are represented by continuous paths : [0, 1] X. Here
the point A = (0) represents the initial state and (1) = B represents
the nal state of the system. The space X is path-connected if and
only if the system can be brought to an arbitrary state from any given
88
4. NAVIGATIONAL COMPLEXITY
state by a continuous motion. We are now interested in algorithms

producing such motions.
Denote by P X = X I the space of all continuous paths
: I = [0, 1] X.
The space P X is supplied with the compact-open topology [89]. Let
(4.1)
: PX X X
be the map which assigns to a path the pair ((0), (1)) X X of

the initial nal congurations. It is easy to see that is a bration
in the sense of Serre, see [89], chapter 2, 8, Corollary 3.
Definition 4.1. A motion planning algorithm is a section of bration
(4.1).
In other words a motion planning algorithm is a map (not necessarily continuous)
(4.2)
s : X X PX
satisfying
(4.3)
s = 1XX .
Note that s(A, B)(t) X is a continuous function of t I for any

pair of points A, B X.
A motion planning algorithm s : X X P X is continuous if the
suggested route s(A, B) of going from A to B depends continuously on
the states A and B.
Do there always exist continuous motion planning algorithms?
Lemma 4.2. A continuous motion planning algorithm in X exists if
and only if the space X is contractible.
Proof. Assume that there exists a continuous motion planning algorithm s : X X P X. For A, B X the image s(A, B) is a
path starting at A and ending at B. Fix B = B0 X and dene
F (x, t) = s(x, B0 )(t). Here F : X [0, 1] X is a continuous deformation with F (x, 0) = x and F (x, 1) = B0 for any x X. Hence X is
contractible.
Conversely, assume that X is contractible. Then there exists a
continuous homotopy F : X [0, 1] X connecting the identity map
4.2. THE CONCEPT TC(X)
89
X X with the constant map X X onto a point B0 X. Using

F one may connect any two given points A and B by the path s(A, B)
which is the concatenation of the path F (A, t) and the inverse path
to F (B, t). This denes a continuous motion planning algorithm in
X.
!
Corollary 4.3. For a system with non-contractible conguration space
any motion planning algorithm must be discontinuous.
This explains why motion planning algorithms appearing in industry are quite often discontinuous.
4.2. The concept TC(X)
Conguration spaces of mechanical systems appearing in most industrial applications are semi-algebraic sets, i.e., they are nite unions of
sets of the form
{x Rn ; P (x) = 0, Q1 (x) > 0, . . . , Ql (x) > 0}
where P, Q1 , . . . , Ql R[X1 , . . . , Xn ] are polynomials with real coecients, see [13]. It is well known that any semi-algebraic set is homeomorphic to a polyhedron, see Theorem 3.12 of [13].
Recall that a polyhedron is dened as a subset X RN homeomorphic to the underlying space of a nite-dimensional simplicial complex,
compare [85]. It is well known that any smooth manifold is homeomorphic to a polyhedron.
Since our main goal is to study motion planning algorithms for real
robotics applications, we may assume that the conguration space X
is homeomorphic to a polyhedron.
Definition 4.4. Let X be a polyhedron. A motion planning algorithm
s : X X P X is called tame if X X can be split into nitely
many sets
(4.4)
X X = F1 F2 F3 Fk
such that:
1. The restriction s|Fi : Fi P X is continuous, i = 1, . . . , k;
2. Fi Fj = , where i == j;
90
3. Each Fi is a Euclidean Neighborhood Retract (ENR), see below.

For a xed pair of points (A, B) Fi , the path t 7 s(A, B)(t) X
produced by the algorithm s is continuous as a function of t, it starts
at point A X and ends at point B X. The curve s(A, B) depends
continuously on (A, B) assuming that the pair of points (A, B) varies
in the set Fi X.
Recall the denition of an ENR, see [17]:
Definition 4.5. A topological space X is a Euclidean Neighborhood
Retract (ENR) if it can be embedded into a Euclidean space X RN
such that for some open neighborhood X U RN there exists a
retraction r : U X, r|X = 1X .
A subset X RN is an ENR if and only if it is locally compact
and locally contractible, see [17], Chapter 4, 8.
In particular, the class of ENRs includes all nite-dimensional cell
complexes and all manifolds.
Motion planning algorithms which appear in industrial applications
are tame. As we mentioned earlier, the conguration space X is usually
a semi-algebraic set, and the algorithm is described by several dierent
rules for various scenarios Fj X X for the initial nal congurations. These sets Fi are also given by equations and inequalities
involving real algebraic functions; thus they are semi-algebraic as well.
The functions s|Fj : Fj P X are also often real algebraic, hence they
are continuous.
Definition 4.6. The topological complexity of a tame motion planning algorithm (4.2) is the minimal number of domains of continuity k
in any representation of type (4.4) for s.
Definition 4.7. The topological complexity TC(X) of a nitedimensional polyhedron X is the minimal topological complexity of
tame motion planning algorithms in X.
TC(X) = 1 if and only if the polyhedron X is a contractible.
Example 4.8. Let us show that TC(S n ) = 2 for n odd and besides
TC(S n ) 3 for n even.
Let F1 S n S n be the set of all pairs (A, B) such that A == B.
We may construct a continuous section s1 : F1 P S n by moving
91
A towards B along the shortest geodesic arc. Consider now the set
A
B
--A
F2 S n S n of all pairs of antipodal points (A, A). If n is odd we

may construct a continuous section s2 : F2 P S n as follows. Fix a
non-vanishing tangent vector eld v on S n ; such v exists for n odd.
Move A towards the antipodal point A along the semi-circle tangent
to vector v(A).
--A
In the case when n is even the above procedure has to be modied
since for n even any vector eld v tangent to S n has at least one zero.
However, we may nd a tangent vector eld v having a single zero
A0 S n . Write F2 = {(A, A); A == A0 } and dene s2 : F2 P S n
as in the previous paragraph. The set F3 = {(A0 , A0 )} consists of a
single pair; dene s3 : F3 P S n by choosing an arbitrary path from
A0 to A0 .
Next we explain the relation between the invariant TC(X) and the
notion of Schwarz genus of a bration. Recall that the Schwarz genus
of a bration p : E B is dened as the minimal number k such that
there exists an open cover of the base B = U1 U2 Uk with the
property that over each set Uj B there exists a continuous section
sj : Uj E of p : E B, see [93].
The concept of Schwarz genus generalizes the notion of Lusternik
Schnirelmann category cat(X); the latter is dened as the smallest
integer k such that X admits an open cover
X = U1 U2 Uk
92
with the property that each inclusion Ui X is null-homotopic, see

[12]. It is easy to see that cat(X) coincides with the genus of the Serre
bration P0 X X. Here P0 (X) is the space of all paths in X which
start at the base point x0 X.
Proposition 4.9 gives a similar interpretation of TC(X):
Proposition 4.9. Let X be a polyhedron. Then the number TC(X)
coincides with the Schwarz genus of the path bration : P X X X.
In the proof of Proposition 4.9 we use the following fact:
Lemma 4.10. Let Y, Z P be two disjoint closed subsets of a polyhedron P . Then there exists a closed sub-polyhedron F P which
contains Y and is disjoint from Z, i.e., Y F and F Z = .
Proof of Lemma 4.10. Let P P be a locally nite family of nite
subpolyhedra such that P = P ; here runs over an index set A. For
each A the intersections Y P and Z P are disjoint compact
sets, hence we may nd a sequence of real numbers i > 0 such that
dist(Y P , Z P ) > i
where dist is a xed metric on P compatible with its topology. Using
the local niteness of the family {P } we may subdivide P such that
every simplex lying in P has diameter < i . Let F denote the union
of all simplices of P whose closure intersects Y . Then clearly F is a
closed sub-polyhedron of P containing Y and disjoint from Z.
!
Proof of Proposition 4.9. Denote by g the Schwarz genus of
and by k the number TC(X). Consider a splitting X X = F1 Fk
as in Denition 4.4. We show that one may enlarge each Fi to an open
set Ui admitting a continuous section of . This implies the inequality
g k.
We will use the following property of ENRs (see [17], chapter 4,
Corollary 8.7): If F X and both spaces F and X are ENRs, then
there is an open neighborhood U X of F and a retraction r : U F
such that the inclusion j : U X is homotopic to ir, where i : F X
denotes the inclusion.
In our situation, for any i = 1, . . . , k the sets Fi and X X are
ENRs, hence the statement of the previous paragraph implies that
there exists an open neighborhood Ui X X of the set Fi and a
continuous homotopy hi : Ui X X, where [0, 1], such that
hi0 : Ui X X is the inclusion and hi1 is a retraction of Ui onto Fi .
93
We may describe now a continuous map s(i : Ui P X with the

property s(i = 1Ui . Given a pair (A, B) Ui , the path hi (A, B) in
X X is a pair of paths (, ), where is a path in X starting at
the point (0) = A and ending at a point (1), and is a path in X
starting at B = (0) and ending at (1). Note that the pair ((1), (1))
belongs to Fi ; therefore the section si = s|Fi : Fi P X denes a path
= si ((1), (1)) P X
connecting the points (1) and (1). Now we set s(i (A, B) to be the
concatenation of , , and 1 (the reverse path of ):
s(i (A, B) = 1 .
Next we want to show that k g. Suppose one has an open cover
U1 Ug = X X such that each of the sets Ui admits a continuous
section of . We show that for i = 1, . . . , g one may nd subsets Fi Ui
satisfying the properties of Denition 4.4.
The sets Y1 = X X (U2 Ug ) and Z1 = X X U1 are
closed and disjoint. Applying Lemma 4.10 we see that there exists a
closed sub-polyhedron F1 containing Y1 and disjoint from Z1 .
We proceed by induction. Suppose that for some 1 < i < g we
have constructed sets F1 , . . . , Fi1 X X satisfying the following
properties:
(a) each Fj is a polyhedron lying in Uj ,
= j (,
(b) Fj Fj ! = for j =
(c) The closure of each Fj is contained in the union F1 Fj ,
(d) F1 Fi1 Ui Ug = X X.
The set Pi = X X (F1 Fi1 ) is an open subset of a
polyhedron, hence it is a polyhedron on its own, see [85]. The sets
Yi = Pi (Ui+1 Ug ) and Zi = Pi Ui
are disjoint and closed in Pi . Applying Lemma 4.10 we nd a closed
polyhedral subset Fi Pi which contains Yi and is disjoint from
Zi . This completes the step of induction. Thus, we nd subsets
F1 , . . . , Fg1 and we nally dene Fg = X X (F1 Fg1 ).
We obtain a splitting of
X X = F1 Fg
satisfying all properties of Denition 4.4. Therefore g k.
Definition 4.11. Let X be a topological space. Its topological complexity TC(X) is dened as the Schwarz genus of bration (4.1).
94
Several equivalent characterizations of TC(X) are given by Proposition 4.12 in the case when X is an ENR.
Proposition 4.12. Let X be an ENR. Then TC(X) = k = Z = r
where the numbers k = k(X), Z = Z(X), r = r(X) are dened as
follows:
(a) k = k(X) is the minimal integer such that there exist a section
s : X X P X of bration (4.1) and an increasing sequence
of k open subsets
U1 U2 Uk = X X
with the property that for any i = 0, 1, . . . , k 1 the restriction
s|(Ui+1 Ui ) is continuous; here U0 = .
(b) l = l(X) is the minimal integer such that there exist a section
s : X X P X of (4.1) and an increasing sequence of k
closed subsets
F1 F2 Fk = X X
with the property that for any i = 0, 1, . . . , k 1 the restriction
s|(Fi+1 Fi ) is continuous; here F0 = .
(c) r = r(X) is the minimal integer such that there exist a section
s : X X P X of bration (4.1) and a splitting
G1 G2 Gr = X X,
Gi Gj = ,
i == j
where each Gi is a locally compact subset of X X and each

restriction s|Gi : Gi P X is continuous, for i = 1, . . . , r.
Remark 4.13. In the denition of r(X) in (c) one may drop the requirement Gi Gj = for i =
= j and the obtained number will be
unchanged. This follows from the arguments of the proof of Proposition 4.12 given below.
Proof. Suppose that TC(X) = s and W1 W2 Ws = X X is
an open cover such that each open set Wi admits a continuous section
si : Wi P X of bration (4.1). Set Ui = W1 W2 Wi where
i = 1, . . . , s. Then U1 U2 Us = X X and one denes a
section s : X X P X by the rule: s(x, y) = si (x, y) where x, y X
and i is the smallest index such that (x, y) Wi . Clearly,
Ui+1 Ui = Wi+1 (W1 Wi )
95
and the restriction s|(Ui+1 Ui ) is continuous. This shows that

TC(X) = s k = k(X).
Let s denote TC(X) and W1 , . . . , Ws be as above. The space X X
is normal (since X and hence X X are metrizable). Therefore one
may nd closed subsets Vi Wi such that V1 Vs = X X. Setting
Fi = V1 V2 Vi (where i = 1, . . . , s) and repeating the argument
of the previous paragraph we obtain the inequality TC(X) Z(X).
Suppose that U1 U2 UW = X X is an increasing chain
of open subsets satisfying conditions of item (a). Then the sets
Gi = Ui (U1 Ui1 )
are locally closed and form a splitting of X X satisfying the conditions
of item (c). Hence, k = k(X) r = r(X).
The inequality Z r follows similarly.
Finally we want to show that r = r(X) s = TC(X). Suppose
that G1 G2 Gr = X X is a decomposition of X X into
pairwise disjoint locally compact subsets Gi such that each Gi admits
a continuous section si : Gi P X of (4.1). Such sections are in
one-to-one correspondence with continuous homotopies hit : Gi X
where t [0, 1] and the end maps hi0 , hi1 : Gi X are projections of
Gi X X onto the rst and the second coordinates correspondingly.
Let Wi X X be an open subset such that Gi = Gi Wi ; such Wi
exists since Gi is a locally closed subset of X X, see [17], chapter 4,
Lemma 8.3. Using Exercise 2 at the end of chapter 4 in [17], we nd an
open set Gi Ui Wi and a homotopy Hti : Ui X connecting the
projections of Ui onto the rst and the second coordinates. The latter
homotopy can be interpreted as a continuous section Si : Ui P X of
(4.1). As the result we obtain an open cover U1 , U2 , . . . , Ur of X X
such that each set Ui admits a continuous section of (4.1). Therefore,
TC(X) r.
!
Recall that the multiplicity (V) of a cover V = {Vi }iI is dened
as the maximal cardinality of a subset J I such that the intersection
iJ Vi =
= is nonempty.
Corollary 4.14. Let X be an ENR. Then TC(X) equals the minimal
multiplicity (V) of open (or closed) covers V = {V1 , . . . , Vm } of X X
having the property that the bration (4.1) admits a continuous section
over each of the sets Vi where i = 1, . . . , m.
96
Proof. If k = TC(X) and U = {U1 , . . . , Uk } is an open (closed) cover

of X X having the above property, then clearly (U) k = TC(X).
Hence to prove Corollary 4.14 one only has to prove the inequality
TC(X) (V)
for any closed (or open) cover V = {V1 , . . . , Vm } with the property that
the bration (4.1) admits a continuous section over each set Vi .
For (x, y) X X denote by (x, y) the number of sets Vi containing the point (x, y). The multiplicity (V) of the cover V equals
(V) =
max
(x,y)XX
(x, y).
For i = 1, 2, . . . , (V) we write

Wi = {(x, y) X X; (x, y) (V) i + 1}.
Each Wi is open (or closed, respectively) and one has
W1 W2 W(V) = X X.
Moreover, every complement Wi+1 Wi is a disjoint union of a family
of subsets, each lying in one of the sets Vj . It follows that there exists a continuous section of bration (4.1) over each Wi+1 Wi . The
inequality TC(X) (V) is now a consequence of Proposition 4.12. !
Corollary 4.14 is closely related to characterization of the topological complexity TC(X) in terms of degrees of instabilities of robot
motion planning algorithms, see [23].
Corollary 4.15. Let X be a polyhedron. Then
(4.5)
TC(X) 2 dim X + 1.
Proof. This corollary is obviously true for any ENR as follows from
the previous corollary. We state and prove it for polyhedra since in
this case we may give an explicit construction of a motion planning
algorithm.
Let X (r) = X r X r1 denote the union of interiors of all simplices
of dimension r where r = 0, 1, . . . , n. Dene
,
Gi =
X (r) X (s) , i = 0, 1, . . . , 2n.
r+s=i
97
The result of Corollary 4.15 would follow from Proposition 4.12 once
we show that each set Gi admits a continuous section si : Gi P X.
One has
6
Gi =
(r) (s)
r+s=i
where (r) , (s) run over open simplices of X of dimensions r and s

correspondingly. Hence a continuous section si can be described on
each connected component (r) (s) of Gi separately.
Given (r) and (s) , consider a path leading from a point x0
(r) to a point y0 (s) . Then one can move from any point x (r) to
any point y (s) by rst going to x0 (along the straight line segment
in (r) ), then following , and nally going from y0 to y, along the
straight line path in (s) . This proves (4.5).
!
A more general upper bound was found in [23], Theorem 5.2:
Theorem 4.16. If X is an r-connected polyhedron then
(4.6)
TC(X) <
2 dim(X) + 1
+ 1.
r+1
Example 4.17. Let us show that for any suspension X = Y where

Y is an ENR one has
(4.7)
TC(X) 3.
Recall that suspension Y is dened as the factor space of the cylinder

p
0
Y
Y
q
Y [0, 1] where each of the subsets Y 0 and Y 1 is collapsed to a

single point, denoted by p and q correspondingly. Note that removing
98
either of p or q makes the suspension X contractible. We see that for

y X {p} there exists a path y in X {p} starting at q, ending at
y and depending continuously on y.
To prove (4.7) one can use property (b) of Proposition 4.12. Consider the closed subsets
F1 F2 F3 = X X
where F1 consists of a single pair (p, p) X X and
F2 = {p} X X {p}.
Dene a section s : X X P X as follows. We set s(p, p) to be
the constant path at p. The complement F2 F1 is the union of two
disjoint sets p (X {p}) and (X {p}) p. For y == p we dene
the path s(p, y) as the concatenation of a xed path 0 from p to q
and the path y , see above; besides, we dene s(y, p) as s(p, y), the
inverse of the path s(p, y). For (x, y) F3 F2 (i.e., when x =
= p and
x == p) we set s(x, y) = x y . We see that the restriction s|(Fi+1 Fi )
is continuous for i = 0, 1, 2 and Proposition 4.12 now gives TC(X) 3.
Example 4.18. The arguments of the previous example can be used to
prove the following more general statement. Let X be a path-connected
ENR and Y = X Z where Z is a nite set. Then
(4.8)
TC(X) TC(Y ) + cat(Y ) + 1.
We leave details of the proof as an exercise.

Proposition 4.19. For any topological space X one has
(4.9)
cat(X) TC(X) cat(X X).
Proof. This follows directly from Denition 4.11, see [22], Theorem
5.
!
Exercise: Let G be a connected Lie group. Then
(4.10)
TC(G) = cat(G).
As a corollary we obtain
(4.11)
TC(SO(3)) = cat(SO(3)) = cat(RP3 ) = 4.
4.3. THE NOTION OF RELATIVE COMPLEXITY
99
This example is important for robotics since SO(3) is the conguration

space of a rigid body in R3 xed at a point. Let us also mention
another consequence of (4.10),
(4.12)
TC(T n ) = cat(T n ) = n + 1.
Finally we mention that there exists yet another interpretation of

the notion TC(X) as a feature of random motion planning algorithms,
see [27].
4.3. The notion of relative complexity
In this section we introduce a relative version of the notion of topological complexity. We also prove that the topological complexity is
homotopy invariant.
Definition 4.20. Let X be a topological space and A X X be a
subspace. Then the number TCX (A) is dened as the Schwarz genus
of the bration : PA X A where PA X P X is the space of all
paths : [0, 1] X such that the pair of end points ((0), (1)) lies
in A. In other words, TCX (A) is the smallest integer k such that there
is an open cover U1 U2 Uk = A with the property that each
Ui A is open and the projections X Ui X on the rst and the
second factors are homotopic to each other, for each i = 1, . . . , k.
Clearly TC(X) = TCX (X X).
We state for future reference:
Lemma 4.21. For a subset A X X the following properties are
equivalent:
(i) TCX (A) = 1,
(ii) the projections X A X are homotopic,
(iii) the inclusion A X X is homotopic to a map A X X
with values in the diagonal X X X.
Let us mention a few obvious inequalities:
(4.13)
TCX (A) TC(X)
and
(4.14)
TCX (A) catXX (A).
If A B X X then
(4.15)
TCX (A) TCX (B).
100
Let Y X be a subspace. Then

(4.16)
TCX (Y Y ) TC(Y ).
Lemma 4.22. Suppose that the sets A B X X are such that B

can be deformed into A inside X X. Then
(4.17)
TCX (A) = TCX (B).
Proof. Indeed, suppose that ht : B X X is a homotopy where

h0 : B X X is the inclusion and h1 maps B into A. Set k =
TCX (A) and let U1 Uk = A be an open cover such that the
projections X Ui X are homotopic. Hence there is a homotopy
si : Ui I X with si (a, b, 0) = a and si (a, b, 1) = b for (a, b) Ui .
Set Wi = h1
1 (Ui ) where i = 1, . . . , k. Then W1 Wk = B is an
open cover. For (x, y) Wi the homotopy ht (x, y) can be viewed as a
pair of paths (, ) in X starting at (0) = x and (0) = y and ending
at a pair ((1), (1)) = (a, b) lying in Ui A. The concatenation
of , si (a, b, ) and 1 is a path connecting x to y which depends
continuously on (x, y) Wi . This shows that TCX (B) TCX (A) and
together with (4.15) proves (4.17).
!
Lemma 4.23. Let X be an ENR and let A X X be a locally compact
subset. Then there is an open neighbourhood A U X X such
that
(4.18)
TCX (A) = TCX (U ).
This statement follows by applying Denition 4.20 and the result

of Exercise 2 from [17], chapter 4, 8.
If A1 , . . . , Ak X X are open subsets covering X X, then
clearly
(4.19)
TC(X) TCX (A1 ) + + TCX (Ak ).
The next statement claims that inequality (4.19) is true in a slightly

more general situation.
Proposition 4.24. Let X be an ENR and let X X be covered by
locally compact sets A1 , . . . , Ak XX, i.e., XX = A1 A2 Ak .
Then TC(X) TCX (A1 ) + + TCX (Ak ).
Proof. It follows from Lemma 4.23.
Lemma 4.25. Let Y X be a retract. Then TC(Y ) TC(X).
4.3. THE NOTION OF RELATIVE COMPLEXITY
101
Proof. Suppose that U1 Uk = X X is an open cover with

a continuous section si : Ui P X for each i = 1, . . . , k where k =
TC(X). Let r : X Y be a retraction. For (x, y) Ui (Y Y )
the formula r(si (x, y)(t)) denes a path in Y from x to y depending
continuously on (x, y) Ui (Y Y ) and on t [0, 1]. Hence the
sets Vi = Ui (Y Y ) form an open cover of Y Y with the required
properties. Therefore, TC(Y ) k.
!
Corollary 4.26. If Y X is a retract and X can be deformed into
Y , then TC(X) = TC(Y ).
Proof. By inequality (4.16) and Lemma 4.22 we have
TC(Y ) TCX (Y Y ) = TCX (X X) = TC(X).
The opposite inequality is given by Lemma 4.25.
Corollary 4.27. The topological complexity is homotopy invariant.

In other words, if topological spaces X and Y are homotopy equivalent
then TC(X) = TC(Y ).
Proof. Any two homotopy equivalent spaces can be realized as deformation retracts of a single space and the result follows from the
previous corollary.
!
Attaching a cell to a space may increase its category by at most 1.
The similar result for the topological complexity reads as follows:
Proposition 4.28. Let Y be an ENR and let
X = Y f (en1 1 enr r )
be obtained from Y by attaching simultaneously several cells via a continuous map f : S n1 1 5 5 S nr 1 Y . Then
(4.20)
TC(X) TC(Y ) + cat(Y ) + 1.
Proof. Let Z X be the nite set Z = {p1 , . . . , pr } containing a

single point pi eni i lying in the interior of each of the attached cells.
Then X Z is homotopy equivalent to Y and the result follows from
Example 4.18.
!
102
There are simple examples when inequality (4.20) is sharp. An even

dimensional sphere S n is obtained from one point Y by adding a cell.
In this case TC(Y ) = cat(Y ) = 1 and TC(X en ) = 3.
Another example: let Y be a contractible graph (a tree) and let X
be obtained from Y attaching r > 1 one-dimensional cells. Then again
TC(Y ) = cat(Y ) = 1 and TC(X) = 3.
Finally we mention the following relation with the notion of relative
Lusternik Schnirelmann category:
Lemma 4.29. For a subset A X and a point x0 X one has
TCX (A x0 ) = catX (A) = TCX (x0 A).
The relative category catX (A) is dened as the smallest cardinality
of an open cover of A with the property that each of its elements is
null-homotopic in X.
Several variations of the notion of TC were studied in [34]. They
were motivated by the natural requirement to have motion planning
algorithms which are symmetric, i.e., such that the motion from A to
B is the reverse of the motion from B to A.
4.4. Navigation functions
The material of this section is inspired by discussions with Daniel Koditschek and Elon Rimon and by reading their papers [66],[82]. Rimon
and Koditschek studied mechanisms which navigate to a xed goal
using a gradient ow technique. We modify slightly their approach by
allowing variable targets; therefore our navigation functions depend on
two variables, the source and the target. Two parametric navigation
functions lead to universal motion planning algorithms for moving from
an arbitrary source to an arbitrary target, instead of having the target
destination xed, as it was in [66], [82].
In this section we assume that the conguration space of our system
is a closed smooth manifold M without boundary.
Definition 4.30. A smooth function F : M M R is called a
navigation function for M if
(a) F (x, y) 0 for all x, y M ,
(b) F (x, y) = 0 if and only if x = y,
(c) F is nondegenerate in the sense of Bott.
The last property means that the set of critical points of F has
several connected components S1 , S2 , . . . , Sk M M where each Si
4.4. NAVIGATION FUNCTIONS
103
is a smooth submanifold and the Hessian of F is nondegenerate on

the normal bundle to Si . As follows from properties (a), (b), one of
these critical submanifolds Si is the diagonal M M M where
M = {(x, x); x M }. We will always assume that S1 = M .
Example 4.31. Suppose that M Rn is a smooth submanifold embedded into Euclidean space. Then the function F : M M R
given by
(4.21)
F (x, y) = |x y|2
for x, y M satises conditions (a) and (b). Condition (c) is also

satised provided that the submanifold M is generic. A pair (x, y)
M M is a critical point of F if and only if the tangent spaces to M
at the points x and y (i.e., Tx M and Ty M ) are perpendicular to the
Euclidean segment [x, y] Rn connecting x and y.
The main idea of the method of navigation functions is the possibility of exploiting the gradient ow of a navigation function (with
respect to a xed Riemannian metric on M ) for constructing motion
planning algorithms. An explicit description of motion planning algorithms based on navigation functions is given below after Example 4.34;
it utilizes notations introduced in the proof of the following theorem.
Theorem 4.32. Let F : M M R be a navigation function for M .
Consider the connected components S1 , S2 , . . . , Sk M M of the set
of critical points of F and denote by ci R the corresponding critical
values, i.e., F (Si ) = {ci }. Then one has
=
(4.22)
TC(M )
Nr .
rCrit(F )
Here Crit(F ) R denotes the set of critical values of F and for r

Crit(F ) the symbol Nr denotes the maximum of the numbers TCM (Si )
where i runs over indices satisfying ci = r, i.e.,
Nr = max{TCM (Si )}.
ci =r
Proof. Consider the negative gradient ow of F with respect to a

Riemannian metric on M . An integral trajectory of this ow is a
pair of smooth curves x(t) M , y(t) M satisfying the dierential
equation
(4.23)
(x(t),
y(t))
= grad(F )(x(t), y(t)),
t [0, ),
104
and the initial conditions x(0) = A, y(0) = B. As t tends to innity

the trajectory (x(t), y(t)) approaches one of the critical submanifolds
Si . For every i = 1, 2, . . . , k we denote by Fi M M the set of
all pairs of initial conditions (A, B) such that the limit of (x(t), y(t))
belongs to Si . We have
(4.24)
M M = F1 Fk ,
and Fi Fj = , i == j. Each Fi is a submanifold (although not necessarily compact), hence an ENR. More precisely, Fi is homeomorphic
to the total space of a vector bundle over Si of rank
2 dim M ind Si dim Si
where ind Si denotes the Bott index of Si , viewed as a critical submanifold of F . The projection i : Fi Si is given by
i (A, B) = lim (x(t), y(t))
t
where (x(t), y(t)) is the trajectory (4.23) satisfying the initial conditions
x(0) = A, y(0) = B.
For a critical value r Crit(F ) denote by Cr and Br the unions
,
,
Cr =
Fi and Br =
Si .
ci =r
ci =r
Clearly Br is the set of all critical points of F lying on the level r

and Cr is the stable manifold of Br . The retraction qr : Cr Br is
continuous (where qr |Fi = i ) and we have
,
M M =
Cr .
rCrit(F )
By Proposition 4.12,
TC(M )
TCM (Cr )
rCrit(F )
and for r Crit(F ) R Corollary 4.26 implies

TCM (Cr ) = max{TC(Fi )} = max{TC(Si )} = Nr .
ci =r
ci =r
4.4. NAVIGATION FUNCTIONS
105
Remark 4.33. The proof above uses the observation that dierent
critical submanifolds Si , Sj M M lying on the same level set of
F can be aggregated, i.e., united into a single submanifold so that
the map qr (see above) remains continuous. This allows signicant
improvement of the upper bound replacing the summation operation
by the maximum, compare (4.22).
More generally, two critical submanifolds Si , Sj can be aggregated
for this purpose assuming that there are no orbits of the gradient ow
of F connecting them.
Note also that the proof of Theorem 4.32 does engage property (b)
from Denition 4.30.
Example 4.34. Let M = S 1 S 1 S 1 = T n be the n-dimensional
torus. Consider the navigation function F : M M R given by
F (u, z) =
n
=
|ui zi |2 .
i=1
Here u = (u1 , u2 , . . . , un ) M and z = (z1 , z2 , . . . , zn ) M where

ui , zi S 1 C denote complex numbers lying on the unit circle. The
critical submanifolds SJ M M of F are in one-to-one correspondence with subsets J {1, 2, . . . , n} where SJ is dened as the set of all
pairs of congurations (u, z) such that ui = zi for i J and uj = zj
for j
/ J. The critical value of submanifold SJ equals 4|J|, where |J|
denotes the cardinality of J. Note that each SJ is dieomorphic to the
torus T n and the relative topological complexity TCM (SJ ) equals 1, as
it is easy to see. Hence in this example we have n + 1 critical values
0, 4, 8, . . . , 4n and each of the numbers Nr appearing in (4.22) equals
1. Therefore, Theorem 4.32 gives the inequality
TC(T n ) n + 1
which, as we know from (4.12), is sharp.
Next we construct a specic motion planning algorithm which uses
navigation functions F : M M R. We will employ the notations
introduced in the proof of Theorem 4.32. For simplicity we will assume
below that all critical values ci are distinct.
Denote by ri the relative topological complexity TCM (Si ), see 4.3.
Find a decomposition
Gi1 Gi2 Giri = Si Si ,
Gij Gij ! = ,
j == j ( ,
106
into locally compact subsets (as in Proposition 4.12(c)) and continuous

sections
(4.25)
sij : Gij P M,
i = 1, . . . , k,
j = 1, . . . , ri
of the path space bration P M M M . Consider the subsets

Vji = i1 (Gij ),
i = 1, . . . , k,
j = 1, . . . , ri .
If (A, B) Vji and i (A, B) = (a, b) Gij Si we may move from A

to B by rst following the trajectory x( ) of the ow (4.23) arriving at
a, then following the path sij (a, b) which starts at a and ends at b and
nally following the inverse path y( ), the solution of (4.23).
It is convenient to introduce a new time parameter = (t) given
by
*
(4.26)
t
0
|grad(F)(x(), y())|2 d.
Then the trajectory (x( ), y( )) (see (4.23)) reaches the critical submanifold Si in nite time
(4.27)
= F (A, B) Fi
where Fi = F (Si ) denotes the constant value which the function F

attains on Si , as follows from the equation dF /d = 1.
Equations (4.23) in the special case of navigation function (4.21)
have the form
(4.28)
x = 2Px (y x),
y = 2Py (x y),
where Px : Tx Rn Tx M and Py : Ty Rn Tx M are orthogonal

projections.
4.5. TC(X), cohomology, and cohomology operations
In this section X denotes a path-connected topological space.
Definition 4.35. Let u H (X X; R) be a cohomology class, where
R is a coecient system on X X. We say that u has weight k 0
if k is the largest integer with the property that for any open subset
A X X with TCX (A) k one has u|A = 0.
4.5. COHOMOLOGY OPERATIONS
107
This denition is inspired by the work of E. Fadell and S. Husseini

[20] and Y. Rudyak [86]; it is similar (but not identical) to the notion
of weight introduced in [34], [32].
We will denote the weight of u by wgt(u). The weight of the zero
cohomology class equals . Knowing weights of cohomology classes
may be helpful in nding lower bounds for the topological complexity
TC(X). The following statement is an immediate consequence of the
denition.
Proposition 4.36. If there exists a nonzero cohomology class u
H (X X; R) with wgt(u) k, then TC(X) > k.
Hence, our goal is to nd nonzero cohomology classes of highest
possible weight.
Lemma 4.37. For u H (X X; R) one has wgt(u) 1 if and only
if
u|X = 0 H (X; R( ).
(4.29)
Here X X X denotes the diagonal and R( = R|X .

Proof. The statement follows from Lemma 4.21.
Cohomology classes satisfying (4.29) are called zero-divisors; this

term was suggested in [22].
Example 4.38. Any cohomology class u H j (X; R), where R is an
abelian group, determines a zero-divisor
u = 1 u u 1 H j (X X; R).
(4.30)
Zero-divisors are easy to describe in the case of cohomology with

coecients in a eld. By the K
unneth theorem any cohomology class
n
u H (X X; k) (where k is a eld) can be represented in the form
u=
m
=
ai b i ,
ai H i (X; k),
bi H ni (X; k).
i=1
The class u is a zero-divisor if and only if

m
=
ai bi = 0,
i=1
i.e., when the result of replacing the cross-product by the cup-product

is trivial. Hence many examples of zero-divisors are available.
108
Cohomology classes having higher weight can be obtained as cupproducts of zero-divisors as concluded from the following statement:
Lemma 4.39. Let u H n (X X; R) and v H m (X X; R( ) be
two cohomology classes. Then the weight of their cup product uv
H n+m (X X; R R( ) satises
(4.31)
wgt(uv) wgt(u) + wgt(v).
Proof. Denote r = wgt(u), s = wgt(v). Any open subset A X X

with TCX (A) r + s can be represented as the union A = B C where
B, C X X are open subsets with TCX (B) r and TCX (C) s.
Then u|B = 0 and hence there exists a class u( H n (X X, B; R)
with u( |(X X) = u. In a similar manner, there exists a renement
v ( H m (X X, C; R( ) with v ( |(X X) = v. Then the cup-product
u( v ( H n+m (X X, A; R R( ) satises (u( v ( )|A = 0 = (uv)|A.
!
Corollary 4.40. If the cup-product of k zero-divisors
ui H (X X; Ri ),
i = 1, . . . , k
is nonzero, then TC(X) > k.

Proof. It follows from Proposition 4.36 and Lemmas 4.37 and 4.39.
!
Corollary 4.40 gives a very eective tool for dealing with TC. As
an illustration we will compute the topological complexity of spheres,
graphs and orientable surfaces.
Proposition 4.41. One has TC(S n ) = 2 for n odd and TC(S n ) = 3
for n even.
Proof. Let u H n (S n ; Q) denote the fundamental class. Then
u = 1 u u 1 H n (S n S n ; Q)
is a nonzero zero-divisor and
u2 = [1 + (1)n ] u u.
We see that u2 is nonzero for n even and hence, by Corollary 4.40,
TC(S n ) 3 for n even. Similarly, Corollary 4.40 implies that TC(S n )
2 for n odd. The inverse inequalities were obtained in Example 4.8. !
Now we calculate topological complexity of graphs.
109
Proposition 4.42. If X is a connected nite graph then
1, if b1 (X) = 0,
2, if b1 (X) = 1,
TC(X) =
3, if b (X) > 1.
1
Proof. If b1 (X) = 0, then X is contractible and hence TC(X) = 1. If
b1 (X) = 1, then X is homotopy equivalent to the circle and therefore
TC(X) = TC(S 1 ) = 2, see above.
Assume now that b1 (X) > 1. Then there exist two linearly independent classes u1 , u2 H 1 (X). Thus
ui = 1 ui ui 1,
i = 1, 2
are zero-divisors and their product u2 u1 u1 u2 == 0 is nonzero.

This implies TC(X) 3 by Corollary 4.40. On the other hand, we
know that TC(X) 3 by Corollary 4.15. Therefore, TC(X) = 3.
!
Proposition 4.43. Let g denote a closed orientable surface of genus
g. Then
1
3, for g = 0, or g = 1,
TC(g ) =
5, for g 2.
Proof. The case g = 0 is covered by Proposition 4.41. In the case
g = 1 the surface 1 is a Lie group and hence TC(1 ) = cat(1 ) =
3 as follows from the exercise at the end of 4.2. Let us show that
TC(g ) 5 for g 2. Indeed, for g 2 one may nd cohomology
classes u1 , v1 , u2 , v2 H 1 (g ; Q) such that ui uj = ui vj = vi vj = 0 for
i=
= j and u2i = 0, vi2 = 0, and besides, u1 v1 = u2 v2 = A H 2 (g ; Q)
is the fundamental class. Then, using the notation (4.30), we obtain
u1 u2 v1 v2 = 2A A == 0.
Hence, the product of four zero-divisors is nonzero and TC(g ) 5
follows from Corollary 4.40. The inverse inequality TC(g ) 5 is a
special case of Corollary 4.15.
!
Next we describe a result from [32] which allows us to nd cohomology classes of weight 2 which are not necessarily cup-products.
Let R and S be abelian groups. A stable cohomology operation of
degree i,
(4.32)
: H (; R) H +i (; S),
110
is a family of natural transformations : H n (; R) H n+i (; S), one

for each n Z, which commute with the suspension isomorphisms, see
[74]. It follows that commutes with all Mayer Vietoris connecting
homomorphisms, and each homomorphism (4.32) is additive, i.e., is a
group homomorphism.
Definition 4.44. The excess of a stable cohomology operation , denoted e(), is dened to be the largest integer n such that (u) = 0 for
all cohomology classes u H m (X; R) with m < n.
Consider a few examples. For an extension 0 R( R R(( 0
of abelian groups the Bockstein homomorphism
: H n (; R(( ) H n+1 (; R( )
has excess 1. The excess of the Steenrod square
Sq i : H (; Z2 ) H +i (; Z2 )
equals i and for any odd prime p the excess of the Steenrod power
operation
P i : H n (; Zp ) H n+2i(p1) (; Zp )
equals 2i, see [46], pages 489490. More generally, the excess of a
composition of Steenrod squares = Sq I = Sq i1 Sq i2 . . . Sq in satises
e() max {ik ik+1 in }.
1kn
It is easy to see that for an admissible sequence I = i1 i2 . . . in (i.e. such

that ik 2 ik+1 for all k) the excess equals
=
e() =
(ik 2ik+1 ),
k
which coincides with the standard notion of excess, see [74], page 27.
As noted above, any cohomology class u H j (X; R) determines a
class
u = 1 u u 1 H j (X X; R)
where denotes the cohomology cross product. Note that u is a zerodivisor and hence wgt(u) 1. Observe that
(u) = (p2 (u) p1 (u)) = p2 ((u)) p1 ((u)) = (u),
by the naturality and additivity of (here p1 , p2 : X X X are the
projections onto each factor).
111
Theorem 4.45. (Compare [32], Theorem 6) Let : H (; R)

H +i (; S) be a stable cohomology operation of degree i and excess
e() n. Then for any cohomology class u H n (X; R) of dimension n the class
(u) = (u) = 1 (u) (u) 1 H n+i (X X; S)
has weight at least 2. In symbols,
wgt((u)) 2.
(4.33)
Proof. Let A X X be an open subset with TCX (A) 2. Then

A = B C where B and C are open and such that their projections
B
C
C
pB
1 : B X, p2 : B X, p1 : C X, p2 : C X are pairwise
mutually homotopic, i.e.,
B
pB
1 9 p2 ,
(4.34)
C
pC
1 9 p2 .
Consider the element

(4.35)
A
u|A = (pA
2 ) (u) (p1 ) (u) H (A; R).
The homomorphism F of the Mayer Vietoris sequence for A,
H n1 (B C; R) H n (A; R) H n (B; R) H n (C; R)

takes the class u|A to zero, F (u|A) = 0, as follows from (4.34) and
(4.35). Hence, u|A = (w) for some w H n1 (B C; R). Therefore,
(u)|A = (u|A) = (w) = (w) = 0,
since is a stable operation of excess n and w has degree n 1. !
In [32] we applied Theorem 4.45 to compute the topological complexity of some lens spaces
L2n+1
= S 2n+1 /Zm .
m
Here the cyclic group Zm = {1, , . . . , m1 } C of m-th roots of
unity acts freely on the unit sphere S 2n+1 Cn+1 by pointwise multiplication. In the literature L2n+1
is known as Lm (1, 1, . . . , 1), see page
m
144 of [46].
112
To illustrate our method we state here a few results from [32].

Theorem 4.46. (See [32], Theorem 13) Suppose that p is an odd prime
and n is such that its p-adic expansion,
n = n0 + n1 p + + nk p k ,
where
ni {0, 1, . . . , m 1},
involves only digits ni satisfying ni (p 1)/2. Then the topological

equals
complexity of the lens space L2n+1
p
(4.36)
) = 4n + 2.
) = 2 dim(L2n+1
TC(L2n+1
p
p
Sketch of the proof. The cohomology H i (L2n+1

; Zp ) is Zp for 0
p
i 2n + 1 and vanishes for i > 2n + 1, see [46]. As generators one
can choose x H 1 (L2n+1
; Zp ) and y = (x) H 2 (L2n+1
; Zp ), where
p
p
1
2
: H (; Zp ) H (; Zp ) is the mod p Bockstein homomorphism. As
; Zp ) coincides with the factor-ring Zp [x, y]/I
a graded algebra H (L2n+1
p
where I is the ideal generated by y n+1 and x2 (see [46], Example 3E.2).
Since is a stable cohomology operation of excess 1, we have by
Theorem 4.45,
(4.37)
wgt((x)) = wgt((x))) = wgt(y) 2,
2n+1
where x = 1 x x 1, y = 1 y y 1 H (Lp2n+1 Lm
; Zp ).
The cohomology class
; <
2n
2n
n
x y = (1)
L2n+1
; Zp )
x (y n y n ) H 4n+1 (L2n+1
p
m
n
. (
is nonzero provided that the binomial coecient 2n
is not divisible
n
by p. The latter condition is equivalent to the requirement that all
digits ni in the p-adic decomposition of n are small, i.e., satisfy
ni (p 1)/2, see [32], Lemma 19. Using (4.37) and Corollary 4.40
we obtain that under the assumptions of the theorem one has
TC(L2n+1
) > wgt(x) + 2n wgt(y) 4n + 1.
p
This proves the lower bound TC(L2n+1
) 4n + 2.
p
2n+1
The reverse upper bound TC(Lp ) 2(2n + 1) = 4n + 2 holds for
all lens spaces; we refer to [32] for more detail.
!
Next we state without proof two results from [32].
Let (n) denote the number of 1s in the dyadic expansion of n.
113
Theorem 4.47. (See [32], Theorem 14) Assume that m = 2r . Then

one has
(4.38)
) = 4n + 2
) = 2 dim(L2n+1
TC(L2n+1
m
m
for lens spaces Lm of dimension 2n+1 for all n satisfying (n) r 1.

The following result gives the answer for all 3-dimensional lens
spaces:
Theorem 4.48. (See [32], Corollary 15) The topological complexity of
the 3-dimensional lens space L3m is given by
1
TC(L3m )
4, for m = 2,
6, for m 3.
The topological complexity of rational formal simply-connected topological spaces coincides with their zero-divisors cup-length plus 1, as
shown by L. Lechuga and A. Murillo [69]. In other words, Corollary
4.40 is sharp in this case.
In a recent preprint [8] Daniel Cohen and Goderzi Pruidze computed explicitly the topological complexity of right-angled Artin groups.
Let = (V , E ) be a nite graph with no loops or multiple edges. Here
V is the set of vertices of the graph and E is the set of edges. The
right-angled Artin group associated to is the group G = G with generators corresponding to vertices v V of and relations vw = wv
corresponding to edges {v, w} E . R. Charney and M. Davis [7]
and J. Meier and L. Van Wyk [72] describe explicitly the Eilenberg
MacLane space X of type K(G , 1). The main result of Cohen and
Pruidze [8] states that
(4.39)
TC(X ) = z() + 1,
where
z() = max |K1 K2 |
K1 ,K2
is the largest number of vertices of covered by two cliques K1 and

K2 in .
Recall that a clique K in a graph is a subset of vertices such that
any two vertices v, w K are connected by an edge in .
114
4.6. Simultaneous control of multiple objects

In this section we consider motion planning algorithms to control simultaneously several systems, assuming that the systems do not interact.
The case of interacting systems will be discussed in the following section.
We will start with the case of two systems S1 and S2 . Let X and Y
be the corresponding conguration spaces. The conguration space of
the system S1 S2 consisting of S1 and S2 , viewed as a single system,
is the product X Y ; indeed, a state of S1 S2 is a pair consisting of
a state of S1 and a state of S2 .
Thus, we have to understand the topological complexity of products
X Y and explicit constructions of motion planning algorithms in the
products X Y .
Theorem 4.49. Let X and Y be ENRs. Then one has
(4.40)
TC(X Y ) TC(X) + TC(Y ) 1.
Proof. Let s : X X P X be a section such that for some tower

of open subsets
= U0 U1 U2 Uk = X X
the restriction s|(Ui+1 Ui ) is continuous for i = 0, 1, . . . , k 1. Here
k = TC(X), see Proposition 4.12. Similarly, let s( : Y Y P Y be a
section such that for some tower of open subsets
= U0( U1( U2( Uk( ! = Y Y
(
the restriction s( |(Ui+1
Ui( ) is continuous for i = 0, 1, . . . , k ( 1, where
(
k denotes TC(Y ).
Dene the product section s s( : (X Y ) (X Y ) P (X Y )
as follows: if x1 , x2 X and y1 , y2 Y , then
(s s( )((x1 , y1 ), (x2 , y2 ))(t) = (s(x1 , x2 )(t), s( (y1 , y2 )(t)).

In other words, we apply s with respect to the x-coordinate and s( with
respect to the y-coordinate. Dene
,
Wn =
Ui Uj( (X X) (Y Y ) (X Y ) (X Y )
i+j=n+1
where n = 0, . . . , k + k ( 1. Then
= W0 W1 Wk+k! 1 = (X X) (Y Y )
4.6. SIMULTANEOUS CONTROL OF MULTIPLE OBJECTS
and
Wn Wn1 =
115
(
)
(Ui Ui1 ) (Uj( Uj1
i+j=n
(
))
is a disjoint union and the section (s s( )|((Ui Ui1 ) (Uj( Uj1
is continuous. The result now follows from Proposition 4.12.
!
Another proof of Theorem 4.49 can be found in [22].

Arguments similar to those used in the proof of Theorem 4.49 prove
the following inequality:
Theorem 4.50. Let A, B X be locally compact subsets of an ENR
X. Then
TCX (A B) catX (A) + catX (B) 1.
Theorem 4.49 suggests the notation
:
TC(X)
= TC(X) 1,
the reduced topological complexity.
Corollary 4.51. For ENRs X1 , . . . , Xk one has
(4.41)
: 1 X2 Xk )
TC(X
k
=
: i ).
TC(X
i=1
Hence, if one controls simultaneously k systems having conguration spaces X1 , . . . , Xk , the total conguration space is the Cartesian
product X1 X2 Xk and its topological complexity is bounded
above by the sum of topological complexities of individual systems according to inequality (4.41).
To obtain a lower bound we introduce the following notation. For
a topological space X we denote by zcl(X) the largest integer k such
that there exist k zero-divisors u1 , . . . , uk H (X X; Q) having a
nontrivial cup-product u1 u2 . . . uk == 0 H (X X; Q). Corollary 4.40
gives the inequality
:
TC(X)
zcl(X).
(4.42)
Lemma 4.52. One has
zcl(X Y ) zcl(X) + zcl(Y ).
116
Proof. Set us k = zcl(X) and l = zcl(Y ). Further, let u1 , . . . , uk

H (X X; Q) and v1 , . . . , vs H (Y Y ; Q) be zero-divisors having
nontrivial products u1 u2 . . . uk == 0 and v1 v2 . . . vl == 0. Then
ui = ui 1 1 H (X X Y Y ; Q) 9 H ((X Y ) (X Y ); Q)
is a zero divisor and
vi = 1 1 vj H (X X Y Y ; Q) 9 H ((X Y ) (X Y ); Q)
is a zero-divisor and the product
u1 u2 . . . uk v1 v2 . . . vl = (u1 u2 . . . uk ) (v1 v2 . . . vl )
is nonzero. Hence, zcl(X Y ) k + l.
: 1 X2 Xk ) +k zcl(Xi ).
Corollary 4.53. TC(X
i=1
Corollary 4.54. Suppose that one controls simultaneously k systems
having path-connected conguration spaces X1 , . . . , Xk . Assume that
: i ) M and (2) each Xi has a
(1) for some constant M one has TC(X
(Xi ; Q) =
nontrivial reduced rational cohomology H
= 0. Then
: 1 X2 Xk ) kM
k TC(X
: 1 Xk ) grows linearly with k.
i.e., the number TC(X
Example 4.55. Imagine that we control k identical systems, each having conguration space X = R2 B, where B R2 is a ball (repre:
: n ) = n.
senting an obstacle). Then TC(X)
= 1, zcl(X) = 1 and TC(X
n
Here X denotes the n-fold Cartesian product X X.
In the case when X = R3 B, where B R3 is a ball, the answers
:
: n ) = 2n.
are slightly dierent: TC(X)
= 2, zcl(X) = 2 and TC(X
Comparing these two cases we conclude that the complexity of the
motion planning problem in R3 is twice the complexity of the planar
motion planning problem.
So far we have discussed the problem of centralized control which
is characterized by centralized decision making. Now let us consider
algorithms of distributed control, i.e., when the controllable objects
have their own motion planning algorithms and behave independently
of the behavior of the others. For simplicity we will assume that the
objects we control are identical, each object has conguration space X
and all motion planning algorithms are tame, see Denition 4.4.
4.7. COLLISION-FREE MOTION PLANNING
117
The motion planning algorithm of the i-th object is given by a

splitting F1i F2i Fsii = X X and by dening a continuous
section sij : Fji P X for j = 1, . . . , si . Here clearly si TC(X). The
domains of continuity for the system of k objects are of the form
Fr11 Fr22 Frkk
where 1 ri ki . We see that any distributed motion planning
algorithm has at least
s1 s2 . . . sk TC(X)k
domains of continuity.
This result has an important implication in control theory:
Theorem 4.56. The topological complexity of centralized control of k
identical objects is linear in k. The distributed control has an exponential in k topological complexity TC(X)k .
Hence, the centralized simultaneous control of a large number of
objects k is more ecient than the distributed one.
4.7. Collision-free motion planning
In this section we briey discuss the problem of nding the topological
complexity TC(F (Rm , n)) of the conguration space F (Rm , n) of n
distinct points in the Euclidean space Rm as well as the problem of
nding TC(F (, n)) where is a graph. These two problems can be
viewed as instances of the problem of simultaneous control of multiple
objects avoiding collisions with each other.
A motion planning algorithm in F (Rm , n) takes as an input two
congurations of n distinct points in Rm and produces n continuous
curves A1 (t), . . . , An (t) Rm , where t [0, 1], such that Ai (t) == Aj (t)
for all t [0, 1], i == j and (A1 (0), . . . , An (0)) and (A1 (1), . . . , An (1)) are
the rst and the second given congurations. In other words, a motion
planning algorithm in F (Rm , n) moves one of the given congurations
into another, avoiding collisions.
The following theorem was obtained jointly with S. Yuzvinsky [26].
Theorem 4.57. One has
m
TC(F (R , n)) =
2n 1 for any odd m,

2n 2 for m = 2.
118
It is an interesting combinatorial problem to nd explicit motion

planning algorithms in F (Rm , n) with 2n local rules. Such algorithms
might have some interesting industrial applications. In [27] we suggested an algorithm having n2 n + 1 local rules.
The proof of Theorem 4.57 uses the theory of subspace arrangements. Consider the set
Hij = {(y1 , . . . , yn ); yi Rm , yi = yj } Rnm .
Here i, j {1, 2, . . . , n}, i < j. The set Hij is a linear subspace of Rnm
of codimension m. The system of subspaces {Hij }i<j is an arrangement
of linear subspaces of codimension m. Our approach to the problem is
to view the union
,
H=
Hij
i<j
as the set of obstacles:

F (Rm , n) = Rnm H.
We will only comment here on the easy case when m 3 is odd.
Then F (Rm , n) is (m 2)-connected and in particular it is simply
connected. Its cohomology algebra is generated by the cohomology
classes
eij H m1 (F (Rm , n)), i == j
which arise as follows. Consider the map
ij : F (Rm , n) S m1 ,
Then
(y1 , y2 , . . . , yn ) 7
yi yj
S m1 .
|yi yj |
eij = ij [S m1 ]
where [S m1 ] is the fundamental class of the sphere S m1 .

The cohomology classes eij satisfy the following relations:
(4.43)
e2ij = 0,
and eij ejk + ejk eki + eki eij = 0
for any triple i, j, k. It follows that a product ei1 j1 ei2 j2 . . . eik jk is nonzero
if and only if the subgraph of the full graph on vertices {1, 2, . . . , n}
having the edges (ir , jr ) contains no cycles.
Hence for m 3 the conguration space F (Rm , n) has the homotopy type of a polyhedron of dimension (n 1)(m 1). Since it is
(m 2)-connected we may use inequality (4.6) to nd
TC(F (Rm , n)) <
2(n 1)(m 1) + 1
1
+ 1 = 2n 1 +
.
m1
m1
4.7. COLLISION-FREE MOTION PLANNING
119
We obtain the inequality TC(F (Rm , n)) 2n 1. To show that

it is an equality we shall use the cohomological lower bound given by
Corollary 4.40. Set eij = 1 eij eij 1. It is a zero-divisor and
(
eij )2 = 2 eij eij == 0. Here we use the assumption that m is odd.
The product
=
n
9
(
e1i )2
i=2
0
equals = (2)n1 m m where m = ni=2 e1i . The monomial m == 0
is nonzero and hence the product is nonzero.
The opposite inequality TC(F (Rm , n)) 2n 1 follows now from
Corollary 4.40. This completes the proof of Theorem 4.57 in the case
m 3 odd. Details of the proof in the case m = 2 can be found in
[26].
In a forthcoming joint paper with Mark Grant we show that
TC(F (Rm , n)) = 2n 2
for all even m.
In paper [33] we studied a more general problem of controlling
multiple particles such that there are no collisions between the particles
and a set of obstacles, which can be movable (however the trajectory
of the obstacles is known in advance).
Next we discuss the conguration spaces F (, n) where is a connected graph. These spaces were studied by R. Ghrist, D. Koditschek
and A. Abrams [39], [40], [2] ; see also [38], [90]. To illustrate the
importance of these conguration spaces for robotics one may mention
the control problems where a number of automated guided vehicles
(AGV) have to move along a network of oor wires. The motion of
the vehicles must be safe: it should be organized so that collisions do
not occur. If n is the number of AGV, then the natural conguration
space of this problem is F (, n) where is a graph.
The rst question to ask is whether the conguration space
F (, n) is connected. Clearly F (, n) is disconnected if = [0, 1]
is a closed interval (and n 2) or if = S 1 is the circle and n 3.
These are the only examples of this kind as the following simple lemma
claims:
Lemma 4.58. Let be a connected nite graph having at least one
essential vertex. Then the conguration space F (, n) is connected.
120
An essential vertex is a vertex which is incident to three or more

edges.
Theorem 4.59. Let be a connected graph having an essential vertex.
Then the topological complexity of F (, n) satises
(4.44)
TC(F (, n)) 2m() + 1,
where m() denotes the number of essential vertices in .

A proof can be found in [24].
Theorem 4.60. Let be a tree having an essential vertex. Let n be
an integer satisfying n 2m() where m() denotes the number of
essential vertices of . In the case n = 2 we will additionally assume
that the tree is not homeomorphic to the letter Y viewed as a subset
of the plane R2 . Then the upper bound (4.16) is exact, i.e.,
(4.45)
TC(F (, n)) = 2m() + 1.
Paper [24] contains a sketch of the proof and also an explicit description of a motion planning algorithm in F (, n) (assuming that
is a tree) having precisely 2m() + 1 domains of continuity.
If is homeomorphic to the letter Y , then m() = 1 and F (, 2) is
homotopy equivalent to the circle S 1 . Hence in this case TC(F (, 2)) =
2. The equality (4.45) fails in this case.
For any tree one has TC(F (, 2)) = 3 assuming that is not
homeomorphic to the letter Y . This example shows that the assumption n 2m() of Theorem 4.60 cannot be removed: if is a tree
with m() 2, then the inequality above would give TC(F (, 2)) =
2m() + 1 5.
Here are more examples. For the graphs K5 and K3,3
K5
K3,3
one has
(4.46)
TC(F (K5 , 2)) = TC(F (K3,3 , 2)) = 5.
In these examples the equality (4.45) is violated.
4.8. MOTION PLANNING AND THE IMMERSION PROBLEM
121
4.8. Motion planning and the immersion problem

In this last section we briey consider the problem of computing the
topological complexity of the real projective spaces. The main result
[25] states that the problem of computing the number TC(RPn ) is
equivalent to a classical problem of manifold topology which asks what
is the minimal dimension N of Euclidean space such that there exists an
immersion RPn RN . The immersion problem for the real projective
spaces was studied by many people and a variety of important results
were obtained; a relatively recent survey can be found in [14]. However
at the moment the immersion dimension of RPn as a function of n is
not known.
The problem of nding motion planning algorithms in the projective
space RPn can be viewed as an elementary problem of topological
robotics. Indeed, points of RPn represent lines through the origin in
the Euclidean space Rn+1 and hence a motion planning algorithm in
RPn describes how a given line A in Rn+1 should be moved to another
prescribed position B.
A
B
O
Lines through the origin in R3 may represent metallic bars xed at the
xed point by a revolving joint; this situation is common in practical
robotics.
If the angle between the lines A and B is acute, then one may rotate
A towards B in the two-dimensional plane spanned by A and B such
that A sweeps the acute angle. Hence the problem reduces immediately
to the special case when the lines A and B are orthogonal. In this case,
if the intention is to use simple rotations, one needs a continuous choice
of the direction of rotation in the plane spanned by A and B.
Note that the Lusternik Schnirelmann category of the real projective spaces is well known and easy to compute: cat(RPn ) = n + 1,
see for example Proposition 4.5 of [53]. Using the general properties
of the topological complexity mentioned above we may write
n + 1 TC(RPn ) 2n + 1.
More precisely, one can show that TC(RPn ) 2n for all n and the
equality holds i n is a power of 2.
122
The following is the main result of [25]:

Theorem 4.61. For any n == 1, 3, 7 the number TC(RPn ) equals the
smallest k such that the projective space RPn admits an immersion
into Rk1 .
For the special values n = 1, 3, 7 one has TC(RPn ) = n + 1, as it is
easy to see [25].
Below is the table of the values TC(RPn ) for n 23, see [25]. It
is obtained by combining Theorem 4.61 with the information on the
immersion problem available in the literature.
n
1 2 3 4 5 6 7 8 9 10 11 12
TC(RPn ) 2 4 4 8 8 8 8 16 16 17 17 19
n
13 14 15 16 17 18 19 20 21 22 23 24
TC(RPn ) 23 23 23 32 32 33 33 35 39 39 39 ?
Explicit motion planning algorithms in RPn with n 7 could be
constructed using multiplication of the complex numbers, the quaternions, and the Cayley numbers, see [25].
The following theorem gives a direct construction of a motion planning algorithm in RPn starting from an immersion RPn Rk .
Theorem 4.62. (See [25]) Suppose that the projective space RPn can
be immersed in Rk . Then TC(RPn ) k + 1.
Proof. Imagine RPn being immersed in Rk . Fix a frame in Rk
and extend it, by parallel translation, to a continuous eld of frames.
Projecting orthogonally onto RPn , we nd k continuous tangent vector elds v1 , v2 , . . . , vk on RPn such that the vectors vi (p) (where
i = 1, 2, . . . , k) span the tangent space Tp (RPn ) for any p RPn .
A nonzero tangent vector v to the projective space RPn at a point
A (which we understand as a line in Rn+1 ) determines a line vA in Rn+1 ,
which is orthogonal to A, i.e., vA A. The vector v also determines an
orientation of the two-dimensional plane spanned by the lines A and vA,
see the gure.
A
4.8. MOTION PLANNING AND THE IMMERSION PROBLEM
123
For i = 1, 2, . . . , k let Ui RPn RPn denote the open set of

all pairs of lines (A, B) in Rn+1 such that the vector vi (A) is nonzero
and the line B makes an acute angle with the line v!
i (A). Let U0
n
n
RP RP denote the set of pairs of lines (A, B) in Rn+1 making an
acute angle.
The sets U0 , U1 , . . . , Uk cover RPn RPn . Indeed, given a pair
(A, B), there exist indices 1 i1 < < in k such that the vectors
vir (A), where r = 1, . . . , n, span the tangent space TA (RPn ). Then the
n+1
#
lines A, v#
and therefore
i1 (A), . . . , vin (A) span the Euclidean space R
the line B makes an acute angle with one of these lines. Hence, (A, B)
belongs to one of the sets U0 , Ui1 , . . . , Uik .
We may describe a continuous motion planning strategy over each
set Ui , where i = 0, 1, . . . , k. First dene it over U0 . Given a pair
(A, B) U0 , rotate A towards B with constant velocity in the twodimensional plane spanned by A and B so that A sweeps the acute
angle. This denes a continuous motion planning section s0 : U0
P (RPn ). The continuous motion planning strategy si : Ui P (RPn ),
where i = 1, 2, . . . , k, is a composition of two motions: rst we rotate
line A toward the line v!
i (A) in the 2-dimensional plane spanned by A
and v!
i (A) in the direction determined by the orientation of this plane
(see above). On the second step rotate the line v!
i (A) towards B along
the acute angle similarly to the action of s0 .
!
Jes
us Gonzalez [42] studied relations between the topological complexity and the immersion dimension for lens spaces. See also the paper
of J. Gonzalez, L. Zarate [43].
Finally, we want to warn the reader about a mistake in paper [81]
which attempts to compute the topological complexity of real Grassmannians. The authors claim in [81] that TC(X) = cat(X X) for
any space (see Theorem 1.8 of [81]). This is incorrect in general; for
example TC(S 1 ) = 2 and cat(S 1 S 1 ) = 3. This error compromises all
statements made in [81].
Recommendations for further reading

To my regret that many exciting mathematical stories of topological robotics have not been mentioned in my lectures. To redress this shortfall
I point out some of them below and give the reader additional bibliographic references.
A relation between topology and robotics was pioneered by D. Gottlieb [44] who observed that the inverse kinematic problem of robotics
reduces to the task of nding a section of a specic smooth map and
methods of dierential topology can be used to decide existence or
nonexistence of a continuous section. A related mathematical problem of snake charming was studied by J.-Cl. Hausmann [50] and E.
Rodriguez [84].
Methods of symplectic topology play a key role in studying polygon
spaces in R3 . This approach was initiated in a fascinating paper of A.
A. Klyachko [65] and developed further by M. Kapovich and J. Millson [60] and J.-Cl. Hausmann and A. Knutson [48]. A.A. Klyachko
[65] used the toolbox of symplectic topology and algebraic geometry to
nd the Poincare polynomials of spatial polygon spaces. J.-Cl. Hausmann and A. Knutson [48] went one step further and determined the
structure of cohomology algebras of these spaces.
One of the key results of topology of conguration spaces of graphs
(which were mentioned briey in Chapter 2) is the theorem stating that
they are non-positively curved. Using a more technical language, these
conguration spaces are CAT(0)-spaces, and therefore aspherical. This
theorem was discovered independently by A. Abrams and R. Ghrist
[39], [1], [2], and J. Swiatkowski [90] who in fact attributes this result
to M. Davis and T. Januszkiewicz (unpublished). Many interesting
consequences of this theorem for robotics and engineering were found
by Robert Ghrist and his collaborators, see for example [41].
The work of V. de Silva and R. Ghrist [15] uses topology for solving
an important engineering problem of coverage in sensor networks.
My recommendations for further reading of topics of computational
topology include [18], [55], [80] and [103]. I also suggest that the
reader browse through the volume Topology and Robotics published
126
RECOMMENDATIONS FOR FURTHER READING
in 2007 in the AMS series Contemporary Mathematics. This volume is a collection of papers written by participants of the conference
Topology and Robotics held in ETH Zurich in 2006.
The book [16], which appeared when these notes had been completed, is an amazing additional source of beautiful mathematics which
uses geometry and topology in service of algorithms in computer science and engineering.
Bibliography
[1] A. Abrams, Conguration spaces of colored graphs, Geom. Dedicata 92
(2002), 185194.
[2] A. Abrams, R. Ghrist, Finding topology in a factory: conguration
spaces. Amer. Math. Monthly 109 (2002), no. 2, 140150.
[3] L. Antoine, Sur lhomeomorphie de gures et de leur voisinages, J. Math.
Pure Appl. 86(1921), 221325.
[4] T. Biedl, E. Demaine, M. Demaine, S. Lazard, A. Lubiw, J. ORourke,
S. Robbins, I. Streinu, G. Toussaint, S. Whitesides, Locked and unlocked polygonal chains in three dimensions, Discrete and Computational
Geometry 26(2001), 283287.
[5] T. Biedl, E. Demaine, M. Demaine, S. Lazard, A. Lubiw, J. ORourke, S.
Robbins, I. Streinu, G. Toussaint, S. Whitesides, A note on reconguring
tree linkages: trees can lock, Discrete Applied Mathematics, 117(2002),
293297.
[6] J. Cantarella, H. Johnston, Nontrivial embeddings of polygonal intervals
and unknots in 3-space, Journal of Knot theory and its Ramications,
7(1998), 10271039.
[7] R. Charney, M.W. Davis, Finite K(, 1)s for Artin groups. Prospects
in topology (Princeton, NJ, 1994), 110124, Ann. of Math. Stud., 138,
Princeton Univ. Press, Princeton, NJ, 1995.
[8] D. C. Cohen, G. Pruidze, Motion planning in tori, Bulletin of the London
Mathematical Society, 40(2008), 249262.
[9] D. C. Cohen, G. Pruidze, Topological complexity of basis-conjugating
automorphism groups, Preprint arXiv:0804.1825.
[10] R. Connelly, E. Demaine, G. Rote, Straightening polygonal arcs and
convexifying polygonal cycles, Discrete and Computational Geometry,
30(2003), 205239.
[11] R. Connelly, Expansive motions, Preprint 2006.
[12] O. Cornea, G. Lupton, J. Oprea, D. Tanre, Lusternik Schnirelmann
category, AMS, 2003.
[13] M. Coste, An Introduction to Semialgebraic Geometry. Dip. Mat. Univ.
Pisa, Dottorato di Ricerca in Matematica, Istituti Editoriali e Poligraci
Internazionali, Pisa 2000.
[14] D. Davis, Immersions of projective spaces: a historical survey, Contemporary Mathematics, AMS, 146(1993), 3137.
[15] V. de Silva, R. Ghrist, Coverage in sensor networks via persistent homology, Algebr. Geom. Topol. 7 (2007), 339358.
[16] E. D. Demaine, J. ORourke, Geometric folding algorithms; Linkages,
origami, polyhedra, Cambridge University Press, 2007.
128
BIBLIOGRAPHY
[17] A. Dold, Lectures on Algebraic Topology, Springer Verlag, 1972.

[18] H. Edelsbrunner, J. Harer, Afra Zomorodian, Hierarchical Morse
Smale complexes for piecewise linear 2-manifolds. ACM Symposium
on Computational Geometry (Medford, MA, 2001). Discrete Comput.
Geom. 30 (2003), no. 1, 87107.
[19] P. Erdos, Problem number 3763, American Mathematical Monthly,
46(1935), 627.
[20] E. Fadell and S. Husseini, Category weight and Steenrod operations,
Bol. Soc. Mat. Mexicana (2) bf 37 (1992), no. 12, 151161.
[21] E. Fadell, L. Neuwirth, Conguration spaces, Math. Scand., 10(1962),
111118.
[22] M. Farber, Topological Complexity of Motion Planning. Discrete and
Computational Geometry 29 (2003), 211221.
[23] M. Farber, Instabilities of Robot Motion. Topology and its Applications
140(2004), 245266.
[24] M. Farber, Collision free motion planning on graphs, in: Algorithmic
Foundations of Robotics IV, M. Erdmann, D. Hsu, M. Overmars, A.
Frank van der Stappen editors, Springer, 2005, 123138.
[25] M. Farber, S. Tabachnikov, S. Yuzvinsky, Topological Robotics: Motion
Planning in Projective Spaces. International Mathematical Research
Notices, 34(2003), 18531870.
[26] M. Farber, S. Yuzvinsky, Topological Robotics: Subspace Arrangements
and Collision Free Motion Planning. Transl. of AMS, 212(2004), 145
156.
[27] M. Farber, Topology of robot motion planning, In Morse Theoretic
Methods in Nonlinear Analysis and in Symplectic Topology, Paul Biran, Octav Cornea, Francois Lalonde editors, pages 185230, Springer
2006.
[28] M. Farber and D. Schuetz, Homology of planar polygon spaces, Geometria Dedicata, 125(2007), pp. 7592.
[29] M. Farber, J.-Cl. Hausmann, D. Schuetz, On the conjecture of Kevin
Walker, Preprint math.AT/0708.2995.
[30] M. Farber, T. Kappeler, Betti numbers of random manifolds, Homology, Homotopy and Applications, 10 (2008), No. 1, 205222.
[31] M. Farber, Topology of random linkages, Algebraic and Geometric
Topology, 8(2008), pp. 155171.
[32] M. Farber, M. Grant, Robot motion planning, weights of cohomology classes, and cohomology operations, Proceedings of the AMS
136(2008), 33393349.
[33] M. Farber, M. Grant, S. Yuzvinsky, Topological complexity of collision
free motion planning algorithms in the presence of multiple moving obstacles, Topology and Robotics (M. Farber, R. Ghrist et al editors),
Contemporary Mathematics AMS, vol. 438, 2007, 7583.
[34] M. Farber and M. Grant, Symmetric motion planning, Topology and
Robotics, (M. Farber, R. Ghrist et al editors), Contemporary Mathematics AMS, vol. 438, 2007, 85104.
[35] Y. Felix, J.-Cl. Thomas, Rational Betti numbers of conguration spaces,
Topology and its applications, 102(2000), 139149.
BIBLIOGRAPHY
129
[36] R. H. Fox, E. Artin, Some wild cells and spheres in three-dimensional

space, Annals of Mathematics, 49(1948), 979990.
[37] M.D. Frank Kamenetskii, Unravelling DNA, Addison Wesley, 1997.
[38] S.R. Gal, Euler characteristic of conguration spaces of a complex, Colloq. Math., 89(2001), 6167.
[39] R. Ghrist, Conguration spaces and braid groups on graphs in robotics,
in: Knots, braids, and mapping class groups , dedicated to Joan Birman, pages 2940, AMS/IP Stud. Adv. Math., 24, AMS, Providence,
2001.
[40] R.W. Ghrist, D.E. Koditschek, Safe cooperative robot dynamics on
graphs, SIAM J. Control Optim. 40 (2002), no. 5, 15561575.
[41] R. Ghrist, V. Peterson, The geometry and topology of reconguration,
Advances in applied mathematics, 38(2007), 302323.
[42] J. Gonzalez, Topological robotics in lens spaces. Math. Proc. Cambridge
Philos. Soc. 139 (2005), no. 3, 469485.
[43] J. Gonzalez, L. Z
arate, BP-theoretic instabilities to the motion planning
problem in 4-torsion lens spaces. Osaka J. Math. 43 (2006), 581596.
[44] D. Gottlieb, Topology and the Robot Arm, Acta Applicandae Mathematicae 11(1988), 117121.
[45] M. Grant, Topological complexity of motion planning and Massey products, Proceedings of the M. M. Postnikov Memorial Conference, 2007,
to appear.
[46] A. Hatcher, Algebraic Topology, Cambridge University Press, 2002.
[47] J.-Cl. Hausmann, Sur la topologie des bras articules, In Algebraic
Topology, Poznan, Springer Lecture Notes, 1474(1989), 146159.
[48] J.-Cl. Hausmann, A. Knutson, Cohomology rings of polygon spaces, Ann.
Inst. Fourier (Grenoble), 48(1998), 281321.
[49] J.-Cl. Hausmann and E. Rodriguez, The space of clouds in an Euclidean
space, Experimental Mathematics, 13 (2004), 3147.
[50] J.-Cl. Hausmann, Contr
ole des bras articules et transformations de
M
obius, Enseignement Math., 51(2005), 87115.
[51] D. Hilbert, S. Cohn-Vossen, Geometry and the imagination, Chelsea
Publishing Company, 1952.
[52] M. Hunt, Linkages and their conguration spaces, 4-th year undergraduate project, Durham University, 2007.
[53] I.M. James, On the category in the sense of Lusternik Schnirelamnn,
Topology 17(1978), 331348.
[54] D. Jordan and M. Steiner, Conguration spaces of mechanical linkages,
Discrete and Computational Geometry, 22(1999), 297315.
[55] T. Kaczynski, K. Mischaikow and M. Mrozek, Computational homology.
Applied Mathematical Sciences, 157. Springer-Verlag, New York, 2004.
[56] Y. Kamiyama, The homology of singular polygon spaces, Canad. J.
Math. 50 (1998), 581594.
[57] Y. Kamiyama, M. Tezuka, T. Toma, Homology of the conguration
spaces of quasi-equilateral polygon linkages, Trans. AMS, 350(1998),
48694896.
[58] Y. Kamiyama, M. Tezuka, Topology and geometry of equilateral polygon
linkages in the Euclidean plane, Quart. J. Math., 50(1999), 463470.
130
BIBLIOGRAPHY
[59] M. Kapovich, J.L. Millson, On the moduli space of polygons in the

Euclidean plane, J. Di. Geometry 42(1995), 133164.
[60] M. Kapovich, J.L. Millson, The symplectic geometry of polygons in
Euclidean space, J. Di. Geometry 44(1996), 479513.
[61] M. Kapovich, J.L. Millson, Universality theorems for conguration
spaces of planar linkages, Topology 41(2002), 10511107.
[62] A.B. Kempe, How to draw a straight line: a lecture on linkages, Macmillian, London, 1877.
[63] D.G. Kendall, D. Barden, T.K. Carne and H. Le, Shape and Shape Theory, John Wiley & Sons, Chichester 1999.
[64] R. Kirby, Problems of low-dimensional topology, In William H. Kazez,
editor, Geometric topology: Proceedings of the 1993 Georgia International Topology Conference, Athens, Georgia, volume 2.2 of AMS/IP
Studies in Advanced Mathematics, pages 35473. American Mathematical Society and International Press, 1997.
[65] A.A. Klyachko, Spatial polygons and stable congurations of points in the
projective line, Algebraic geometry and its applications, Aspects Math.,
E25, Vieweg, Braunschweig, 1994, 6784.
[66] D. E. Koditschek, E. Rimon, Robot navigation functions on manifolds
with boundary, Adv. in Appl. Math. 11 (1990), no. 4, 412442.
[67] J.-C. Latombe, Robot Motion Planning. Kluwer Academic Publishers,
1991.
[68] H. Lebesgue, Lecons sur les constructions geometriques, Gauthier Villars, Paris, 1950.
[69] L. Lechuga, A. Murillo, Topological complexity of formal spaces, Topology and Robotics, M. Farber, R. Ghrist et al editors, Contemporary
Mathematics, 438(2007), 105114.
[70] James Clerk Maxwell, On reciprocal gures, frameworks and diagrams of
forces, Transactions of the Royal Society of Edinburgh, 26(1870), 140.
[71] James Clerk Maxwell, On Bows method of drawing diagrams in graphical statics, with illustrations from Peaucelliers linkage, Cambridge
Philosophical Society Proceedings 2(1876), 407414.
[72] J. Meier, L. VanWyk, The Bieri Neumann Strebel invariants for
graph groups, Proc. London Math. Soc. (3) 71 (1995), no. 2, 263280.
[73] K. Millett, Knotting of regular polygons in 3-space, Journal of Knot
theory and its ramications, 3(1994), 263278.
[74] R.E. Mosher, M.E. Tangora, Cohomology operations and applications in
homotopy theory, Harper Row Publishers, 1968.
[75] R. J. Milgram, J.C. Trinkle, The geometry of conguration spaces for
closed chains in two and three dimensions, Homology, Homotopy and
Applications 6(2004), 237267.
[76] J. Milnor, Morse theory, Princeton Univ. Press, 1969.
[77] J. Milnor, Lectures on the h-cobordism theorem, Princeton Univ. Press,
1966.
[78] E.E. Moise, Geometric topology in dimensions 2 and 3, Springer Verlag, 1977.
[79] S.H. Niemann, Geometry and Mechanics, PhD thesis, St. Catherines
College, Oxford, 1978.
BIBLIOGRAPHY
131
[80] P. Niyogi , S. Smale and S. Weinberger, Finding the Homology of Submanifolds with High Condence from Random Samples, Discrete Comput. Geom. 39(2008), no. 13, 419441.
[81] K.J. Pearson and T. Zang, Topological complexity and motion planning
in certain real Grassmannians, Applied Mathematics Letters, 17(2004),
499502.
[82] E. Rimon, D.E. Koditschek, The construction of analytic dieomorphisms for exact robot navigation on star worlds, Trans. Amer. Math.
Soc. 327 (1991), no. 1, 71116.
[83] R.T. Rockafellar, Convex analysis, Princeton University Press, 1970.
[84] E. Rodriguez, The snake charmer algorithm, Topology and Robotics,
(M. Farber, R. Ghrist et al editors), Contemporary Mathematics AMS,
vol. 438, 2007, 1932.
[85] C. P. Rourke, B.J. Sanderson, Introduction to Piecewise-linear topology,
Springer-Verlag, 1972.
[86] Y. B. Rudyak, On category weight and its applications, Topology, 38
(1999), no. 1, 3755.
[87] M. Sharir, Algorithmic motion planning. Handbook of Discrete and
Computational Geometry. J.E. Goodman, J. ORourke eds. CRC Press,
Boca Raton, FL, 1997, 733754.
[88] S. Smale, On the topology of algorithms, I, J. of Complexity, 3(1987),
8189.
[89] E. Spanier, Algebraic Topology, Mebman Hell, 1966.
[90] J. Swiatkowski, Estimates for the homological dimension of conguration
spaces of graphs, Colloq. Math., 89(2001), 6979.
[91] I. Streinu, Pseudo-Triangulations, Rigidity and Motion Planning, Discrete and Computational Geometry, 34(2005), 587635.
[92] N. Shvalb, M. Shoham, D. Blanc, The conguration space of arachnoid
mechanisms, Forum Mathematicum, 17(2005), 10331042.
[93] A.S. Schwarz, The genus of a ber space. Amer. Math. Sci. Transl.
55(1966), 49140.
[94] W. Thurston, Shapes of polyhedra, Preprint 1987.
[95] W. Thurston, J. Weeks, The mathematics of three-dimensional manifolds, Scientic American, July 1986, 94106.
[96] J.C. Trinkle, R.J. Milgram, Complete path planning for closed kinematic
chains, The International Journal of Robotics Research, 21(2002), 773
789.
[97] J.C. Trinkle, R.J. Milgram, The geometry of conguration spaces for
closed chains in two and three dimensions, Homology, homotopy and
applications, 6(2004), 237267.
[98] I. Vardi, Computational recreations in mathematics, Redwood City, CA,
Addison-Wesley, 1991.
[99] V.A. Vassiliev, Cohomology of braid groups and complexity of algorithms,
Functional Analysis and its Appl., 22(1988), 1524.
[100] K. Walker, Conguration spaces of linkages, Undergraduate thesis,
Princeton Univ., 1985.
[101] A.Wenger, Etudes des espaces de congurations de certains syst`emes
articules, Travail de dipl
ome, University of Geneva, 1988.
132
BIBLIOGRAPHY
[102] S. Yuzvinsky, Topological complexity of generic hyperplane complements,

Topology and Robotics, M. Farber, R. Ghrist et al editors, Contemporary Mathematics, 438(2007), 115119.
[103] A. Zomorodian, G. Carlsson, Computing persistent homology. Discrete
Comput. Geom. 33 (2005), no. 2, 249274.
Index
Admissible sequence of measures, 38

Average Betti numbers, 37
Bockstein homomorphism, 110
Centralized control, 116
Chamber, 32
Cohomology operation, 109
Collared subpolyhedron, 54
Collinear congurations, 12
Conguration space, 2
Conguration spaces of graphs, 46
Conguration spaces of manifolds, 45
C&P-Grothendieck ring, 55
C&P-surgery, 54
Distributed control, 116
ENR, 90
Equilateral linkage, 18
Euclidean neighbourhood retract, 90
Euler Gal power series, 42
Expansive innitesimal motion, 67
Generic length vector, 9
Innitesimal motion, 67
Isotopy, 61
Median subset, 10
Morse Smale complex, 20
Motion planning algorithm, 88
Navigation function, 102
Normal length vector, 34
Planar linkage, 2
Poincare polynomial, 16
Polygon space, 3
Polyhedral cell complex, 43
Random linkage, 36
Reduced Euler characteristic, 43
Reduced topological complexity, 115
Relative topological complexity, 99
Robot arm, 5
Robot arm distance map, 11
Schwarz genus, 91
Short subset, 10
Singularities of M% , 15
Spatial polygon spaces, 34
Steenrod square, 110
Stratum, 32
Stress, 70
Strut, 68
Subspace arrangement, 118
Length vector, 3
Lens space, 111
Link of a simplex, 42
Linkage, 2
Long subset, 10
Lusternik Schnirelmann category,
91
Variety of quadrangles, 6
Manifold with involution, 19

Maxwell Cremona theorem, 73
Zero-divisor, 107
Tame motion planning algorithm, 89

Topological complexity TC(X), 90
Walkers conjecture, 31
Weight of cohomology class, 106
Workspace, 5

(Michael Farber) Invitation To Topological Robotic

Uploaded by

Copyright:

Available Formats

(Michael Farber) Invitation To Topological Robotic

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Michael Farber) Invitation To Topological Robotic

Uploaded by

Copyright:

Available Formats

Zurich Lectures in Advanced Mathematics

2000 Mathematics Subject Classification (primary; secondary): 58E05, 57R70; 57R30

2008 Michael Farber

1. Linkages and Polygon Spaces . . . . . . . . .

2. Euler Characteristics of Conguration Spaces .

3. Knot Theory of the Robot Arm .

4. Navigational Complexity of Conguration Spaces

Recommendations for further reading . . . . . . . . . . . . . . . . . . 125

Linkages and Polygon Spaces

1. LINKAGES AND POLYGON SPACES

Figure 1.1. Linkage.

1.1. CONFIGURATION SPACE OF A LINKAGE

Figure 1.2. Closed planar polygon.

In (1.3) the group of rotations SO(2) acts on the vectors u1 , . . . , un

1. LINKAGES AND POLYGON SPACES

li > l1 + + li1 + li+1 + + ln .

1.2. THE ROBOT ARM WORKSPACE MAP

Proof. It is obvious that (1.4) implies that MW = .

li + li+1 l1 + + li1 + li+2 + + ln .

Indeed, assuming the contrary, we obtain

and adding all these inequalities gives 4L > nL, a contradiction.

Figure 1.3. Robot arm with two bars.

notion adjacent is understood cyclically, i.e., l1 and ln are also adjacent.

1. LINKAGES AND POLYGON SPACES

Figure 1.4. Robot arm map : C = T 2 W .

1.3. VARIETIES OF QUADRANGLES: n = 4

us some experimental material which will be helpful in handling the

Our purpose is to study the variety of shapes of all quadrangles with

Figure 1.5. Mutual positions of the annulus W and the

1. LINKAGES AND POLYGON SPACES

represents two dierent congurations except the tangency point where

depending on the values of l1 , . . . , l4 . The cases B and D should be

Figure 1.6. The workspace and the bold circle for l1 = l2 .

1.4. SHORT, LONG AND MEDIAN SUBSETS

Figure 1.7. The conguration space MW in cases F and G

Geometrically this can be expressed as follows. A length vector Z is

1. LINKAGES AND POLYGON SPACES

We will call a subset J {1, . . . , n} short with respect to the length

i.e., if its complement is short. By Lemma 1.4 the moduli space of

The complement of a median subset is also a median subset. Clearly,

1.5. THE ROBOT ARM DISTANCE MAP

Generic cases are A, C, E & F and in these cases the conguration

Figure 1.8. Robot arm with n links.

Figure 1.9. Function fW : W R.

1. LINKAGES AND POLYGON SPACES

Dene a real-valued function on W as follows:

Geometrically the value of fW equals the negative of the squared distance

Figure 1.10. A collinear conguration pJ of the robot arm.

1.5. THE ROBOT ARM DISTANCE MAP

We conclude that either

or tan i is independent of i. The rst possibility happens if and only

We will calculate the Hessian of fW at point pJ . Note that fW (pJ ) =

1. LINKAGES AND POLYGON SPACES

This shows that the Hessian of fW at pJ is congruent to the matrix DE,

as ui = +1 for i J; the vanishing in (1.21) happens only for r = n,

()k1 for k r < n

and n = 0. We see that the number of negative eigenvalues of D E

1.5. THE ROBOT ARM DISTANCE MAP

where D = diag(d1 , d2 , . . . , dn ) is an (n n)-diagonal matrix, and E is