Eat

Elementary Applied Topology
Robert Ghrist University of Pennsylvania
EE
Contents
Preface 1 Manifolds
1.1 Manifolds . . . . . . . . . . . . . 1.2 Conguration spaces of linkages . 1.3 Derivatives . . . . . . . . . . . . 1.4 Vector elds . . . . . . . . . . . 1.5 Braids and robot motion planning 1.6 Transversality . . . . . . . . . . 1.7 Signals of opportunity . . . . . . 1.8 Stratied spaces . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vii
2 3 5 6 7 9 12 14 16 20 22 23 23 25 25 26 27 29 30 32 34
Complexes
2.1 Simplicial and cell complexes . . . . . . . . . . . 2.2 Vietoris-Rips complexes and point clouds . . . . . 2.3 Flag complexes and networks . . . . . . . . . . . 2.4 Phylogenetic trees and links . . . . . . . . . . . . 2.5 Witness complexes and the sampling problem . . 2.6 ech and random samplings . . . . . . . . . . . . 2.7 Nerves and neurons . . . . . . . . . . . . . . . . 2.8 Strategy complexes and planning with uncertainty 2.9 Decision tasks and consensus . . . . . . . . . . . 2.10 Discretized graph conguration spaces . . . . . . 2.11 State complexes and reconguration . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 3.2 3.3 3.4 3.5 Counting . . . . . . . . . . . Curvature . . . . . . . . . . . Obstructions to vector elds Index theory for vector elds Tame topology . . . . . . . . iii . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
Euler Characteristic
38 39 40 41 42
37
iv
Contents
3.6 3.7 3.8 3.9 3.10 3.11
Euler calculus
. .
. . .
. . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
44 45 46 48 50 51 54
Target enumeration
The Fubini Theorem .
Euler integral transforms . Intrinsic volumes . . . . . . . .
Gaussian random elds . . . . . . . . . . .
Exercises
Homology
4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 4.11 4.12 Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Homology examples Coecients . . . . . .
55
56 57 59 60 61 62 62 63 63 64 66 67 69
Singular homology
Reduced homology . ech homology . . . . .
Relative homology Local homology .
Homology of a relation Functoriality . . . . . . .
Inverse kinematics
Winding number and degree . . . . . . . . . . . . . . . .
Exercises
Sequences
5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10 5.11 5.12 5.13 5.14 Homotopy invariance . Exact sequences Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
72 73 75 75 76 78 79 80 80 81 82 84 86 88 90
Mayer-Vietoris
Degree and local computation . Borsuk-Ulam theorems . Euler characteristic . Lefschetz index . . . . . . . . . . . . . . . . . . .
Equivalence of homology theories Coverage in sensor networks . Homological coverage Persistent homology . . . . . . . . . . . . . . . . . . . . .
The space of natural images . Zigzag persistence . . . . . . . . . . . . . . . . . . . . .
Exercises
Cohomology
6.1 6.2 6.3 Dual spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Cochain complexes . Poincar duality . .
93
94 94 98
Contents 6.4 Alexander duality . . . . . . . . . . . . 6.5 Euler integration over ad hoc networks 6.6 Convex geometry . . . . . . . . . . . 6.7 Forms and de Rham cohomology . . . 6.8 Currents . . . . . . . . . . . . . . . . 6.9 Laplacians . . . . . . . . . . . . . . . 6.10 Circular coordinates in data sets . . . . Exercises . . . . . . . . . . . . . . . . . . . .
7 Morse Theory
v . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 99 101 101 105 106 108 109

111
7.1 Classical Morse theory . . . . . . . . . . 7.2 4-valued Euler integration . . . . . . . . 7.3 Qualitative signal processing . . . . . . . 7.4 Stratied Morse theory . . . . . . . . . 7.5 Conley Index . . . . . . . . . . . . . . . 7.6 Combinatorial Morse theory . . . . . . . 7.7 L-S category . . . . . . . . . . . . . . . 7.8 Unimodal decomposition in statistics . . 7.9 Bundles and sectional category . . . . . 7.10 Topological complexity of path planning Exercises . . . . . . . . . . . . . . . . . . . . . 8.1 Presheaves . . . . . . . 8.2 Examples of sheaves . . 8.3 Operations on sheaves . 8.4 Sheaf cohomology . . . 8.5 Pursuit and evasion . . 8.6 Network coding . . . . 8.7 Categories . . . . . . . 8.8 Diagrams and nerves . . 8.9 Functors . . . . . . . . 8.10 Clustering functors . . . 8.11 Natural transformations 8.12 Sheaves and categories Exercises . . . . . . . . . . . . A.1 A.2 A.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
112 114 115 117 117 120 122 123 125 126 129 132 134 136 137 139 140 141 143 143 144 146 146 147
Sheaves & Other Functors
131
Background
On point-set topology . . . . . . . . . . . . . . . . . . . . . . . 149 On linear and abstract algebra . . . . . . . . . . . . . . . . . . . 150 On homotopy theory . . . . . . . . . . . . . . . . . . . . . . . . 152
157
149
Bibliography
vi
Contents
2HAB=?A
viii
What topology is
Preface
f 'f g X!Y
opology is the mathematics of spaces and maps between them. Implicit are the notions of proximity and continuity. Proximity in a space is controlled by its topology: the collection of open neighborhoods of all points in . Nested systems of neighborhoods communicate proximity without necessarily requiring a metric distance. This thin notion of closeness suces to dene all the necessary notions of continuity, convergence, and connectivity familiar to students more at home with calculus in Euclidean spaces. The term map always means a continuous function between spaces. Topology explores deformations of maps and spaces, and the basic equivalence relations of topology emphasize qualitative features. Certainly, a homeomorphism (a bijective map with continuous inverse) counts as a topological equivalence. A more general deformation arises at the level of maps. A homotopy of maps : is a continuous 1-parameter family of maps J : interpolating and . We having maps write for homotopic maps. Homotopic spaces are those Id; . : and : with Id: and Topology does its work by means of distinguishing spaces and maps to various degrees of resolution. A topological invariant of spaces [maps] is an assignment of some (usually algebraic) datum to spaces [maps] which respects the equivalence relation of homotopy: homotopic spaces [maps] are sent to the same invariant. Counting the number of connected components of a space is a simple topological invariant. Topology is, from a certain modern perspective, all about arrows: maps, invariants, sequences, diagrams, functors. The simple gures adorning chapters are drawn with this dictum in mind.
h Y
!X
hg '
f X!Y
gh '
f ;f X ! Y f f X'Y
What topology can do

Topology was built to distinguish qualitative features of spaces and mappings. It is good for, inter alia: 1. Characterization: Topological properties extract qualitative features. For example, the genus of a surface, or the number of connected components of an object, give global characteristics important to classication. This is important in extracting high-level features. 2. Continuation: Topological properties are usually robust. The number of components is not something that changes with a small error in measurement. This is vital to applications in scientic disciplines, where data is never not noisy. 3. Integration: Topology is the premiere tool for converting local data into global properties. As such, it is rife with principles and tools (Mayer-Vietoris, Excision, spectral sequences, sheaves) for integrating local data. 4. Obstruction: Topology often provides tools for answering feasibility of certain problems, even when the answers to the problems themselves are hard to compute. These characteristics, classes, degrees, or obstructions take the form of algebraic-topological quantities.
Preface
What topology cannot do
ix
Topology is ckle. There is no recourse to tweaking epsilons should desiderata fail to be found. If the reader is a scientist or applied mathematician hoping that topological tools are a quick x, take this text with caution. The reward of reading this book with care may be limited to the realization of new questions as opposed to new answers. It is not uncommon that a new mathematical tool contributes to applications not by answering a desired question but by revealing a dierent (and perhaps more signicant) underlying principle.
What this text is
This text is a quick tour of examples to motivate the subject of applied topology. The interested reader will have to look elsewhere for more details. The intent is to span a broad array of ideas, tools, perspectives, and applications; this precludes depth, both in the mathematics and in the applications. The author would have called this text Cartoons in Applied Topology were it not for the resulting confusion. The subject is in its infancy, and it seems certain that even a more detailed treatment of the examples given here would likely appear quaint in less than a decade. The best approach, perhaps, is to make the text intentionally simplistic, in the hopes that it will lure the unsuspecting reader to greater depths and prepare for the eld as it will be. The chapters are organized according to mathematical sub-topic, rather than according to application domain. This raises an interesting philosophical question about the nature of applied mathematics: is it how dierent branches of mathematics embed in the physical world, or is how dierent applications implicate and are aided by mathematics? The organization of the text reects the author's view that applied mathematics concerns the incarnation of mathematical objects and structures. The text begins with the necessary background in spaces, emphasizing examples and avoiding the set-theoretic denitions which, though necessary, may overly discourage the interested scientist. The goal is to skip through the formalisms in order to get to applications as quickly as possible. This reects the author's learning of the subject of topology: despite the best eorts of brilliant topologists (including Profs. Dranishnikov, Hatcher, Krsti, and Vogtmann), the author never learned much of anything in the subject without rst nding some physical manifestation of the principle, no matter how cartoonish. This book represents a partial collection of such cartoons. This is not a mathematics text: proofs are seldom given, for the sake of making the exposition quick and painless. The reader should not conclude that the subject is quickly or painlessly learned. This text, properly used, is the impetus for future work: hard, slow, and fruitful. This is not a text in computational topology: the reader may look to several excellent sources [?, ?] for the problem of algorithmic complexity of the topological objects explored in this text. The author believes that the questions of What is it good for? and How do I compute that? are neither independent nor inseparable. Experts will likely be exasperated with this text, for many reasons. The text is
Preface
meant not for experts but for students, to point them in the direction of better texts to come. The author is also exasperated with a text that was written with suboptimal constraints on space and time.
Acknowledgements
A debt of thanks is due to the organizers and attendees of the 2009 CMBS meeting at Cleveland State University, Peter Bubenik and John Oprea above all The young eld of Applied Topology has numerous practitioners and supporters, and their consistent support for the author are worth more than words can repay. Of particular note is Benjamin Mann, the primum mobile of the eld. The vision and guidance of (chronological order) Philip Holmes, Bob Williams, Konstantin Mischaikow, Dan Koditschek, and Gunnar Carlsson have been invaluable. The author has had the pleasure of collaborating with and learning from Yuliy Baryshnikov, Vin de Silva, John Etnyre, Yasu Hiraoka, Michael Robinson, and Rob Vandervorst during this last decade of development in the eld. Thank you for your patience. The work behind and writing of this text was made possible through the generous support of DARPA, the NSF, the ONR, and the AFOSR. )TEX. All gures were drawn by the author using Adobe Illustrator; typesetting by L
Chapter 1
Spaces: Manifolds
Chapter 1. Manifolds
he most intuitive and initially useful topological constructs are smooth manifolds and smooth functions between them. Manifolds are the extension of domains familiar from calculus curves and surfaces to higher-dimensional settings.
A topological n-manifold is a space M locally homeomorphic to Rn . That is, there is a cover U = fU g of M by open sets along with maps : U ! Rn that are continuous bijections onto their images with continuous inverses. In order to do dierential calculus, one needs a smoothing of a manifold. This consists of insisting that the maps
1.1 Manifolds
are smooth (innitely dierentiable, or C 1 ) whenever U \ U 6= ;. The pairs (U ; ) are called charts; they generate a maximal atlas of charts which specify a smooth structure on M . Charts and atlases are rarely explicitly constructed, and, if so, are often immediately ignored. The standard tools of multivariable calculus the Inverse and Implicit Function Theorems lift to manifolds and allow for a simple means of producing interesting examples. Smooth curves are 1-manifolds, easily classied. Any connected curve is dieomorphic (smoothly homeomorphic with smooth inverse) to either R or to the circle S1 ; thus, compactness suces to distinguish the two. The story for 2-manifolds surfaces introduces two more parameters. Compact surfaces can be orientable or non-orientable, and the existence of holes or handles is captured in the invariant called genus. The sphere S2 is the orientable surface of genus zero; the torus T2 the orientable surface of genus one; their nonorientable counterparts are the projective plane P2 and the Klein bottle K 2 respectively. The Classication Theorem for Surfaces states that any compact surface is dieomorphic to the orientable or non-orientable surface of some xed genus g 0. The spatial universe is, seemingly, a 3manifold . The classication of 3-manifolds is a delightfully convoluted story [?], with recent, spectacular progress [?] that perches this dimension between the simple (2) and the wholly bizarre (4).
manifolds in this text should be assumed Hausdor and paracompact. The reader for whom these terms are unfamiliar is encouraged to ignore them for the time being. This is a simplication, ignoring whatever complexities black holes, strings, and other exotica produce.
All
1 : (U \ U ) ! (U \ U )
1.2. Conguration spaces of linkages distance to the origin. The n-sphere is an n-dimensional manifold. The 0-dimensional sphere S0 is disconnected it is the disjoint union of two points. For n > 0, Sn is a connected n as manifold dieomorphic to the compactication of R n follows. Consider the quotient space obtained from R ?, where ? is an abstract point whose neighborhoods consist of ? union the points in Rn suciently far from the origin. This abstract space is dieomorphic to Sn via a dieomorphism that sends the origin and ? to the south and north poles of the sphere Sn respectively.
Example 1.1 (Spheres) The n-sphere, Sn , is the set of points in Euclidean Rn+1 unit
Pn , is dened as the n +1 space of all 1-dimensional linear subspaces of R , with the topology that says neighborhoods of a point in Pn are generated by small open cones about the associated line. That Pn is an n-manifold for all n is easily shown (but should be contemplated until
Example 1.2 (Projective spaces) The real projective space,
it appears obvious). Projective 1-space, P1 , is dieomorphic to S1 . The projective plane, P2 , is a non-orientable surface dieomorphic to the following quotient spaces: 1. Identify opposite sides of a square with edge orientations reversed; 2. Identify antipodal points on the boundary of the closed unit ball B R2 ; 3. Identify antipodal points on the 2-sphere S2 . For any n, Pn is dieomorphic to the quotient Sn =a, where a : Sn ! Sn is the antipodal map a(x ) = x . The space P3 is dieomorphic to the space of rotation matrices, SO3 , the group of real 3-by-3 matrices with determinant 1. Among the many possible extensions of projective spaces, the Grassmannian spaces arise in numerous contexts. The Grasmannian Gn k is dened as the space of all k -dimensional subspaces of Rn , with topology induced in like manner to Pn . The Grassmannian is a manifold that specializes to +1 Pn = Gn . 1
1.2 Conguration spaces of linkages

Applications of manifolds and dierential topology are ubiquitous in rational mechanics, Hamiltonian dynamics, and mathematical physics and are well-covered in standard texts [?, ?]. A simple application of (topological) manifolds to robotics falls under a dierent aegis. Consider a planar mechanical linkage consisting of several at, rigid rods joined at their ends by pins that permit free rotation in the plane. One can use out-of-plane height (or mathematical license) to assert that interior intersections of rods are ignorable. The conguration space of the linkage is a topological space that assigns a point to each conguration of the linkage a relative positioning of the rods up to equivalence generated by rotations and translations in the plane and which
assigns neighborhoods in the obvious manner. A neighborhood of a conguration is all congurations obtainable via a small perturbation of the mechanical linkage. The conguration space of a planar linkage is almost always a manifold, the dimension of which conveys the number of mechanical degrees of freedom of the device. Four rods of lengths fLi g4 1 are linked in a cyclic chain. When one rod is anchored, the system is seen to have one mechanical degree of freedom. The conguration space is thus one-dimensional and almost always a manifold. If, as illustrated, one has a single short rod, then this rod can be rotated completely about its anchor, causing the opposing rod to rock back-and-forth. This linkage is used to transform spinning motion (from, say, a motor) into rocking motion (as in a windshield wiper). The conguration space of the illustrated linkage is S1 S1 , the coproduct or disjoint union of two circles. The second circle comes from taking the mirror image of the linkage along the axis of its xed rod in the plane and repeating the circular motion there: this forms an entirely separate circle's worth of conguration states. Many other familiar manifolds are realized as conguration spaces of planar linkages (with judicious use of the third dimension to mitigate bar crossings). The undergraduate thesis of Walker [?] has many examples of orientable 2-manifolds as conguration spaces of planar linkages. The linkage illustrated in the margin has conguration space a closed, orientable surface of genus g ranging between 0 and 4, depending on the length between the xed ends of the linkage [?]. The reader is encouraged to try building a linkage whose conguration space yields an interesting 3-manifold. The realization question this exercise prompts has a denitive answer (albeit with a convoluted attribution and history [?, ?, ?, ?]):
Example 1.3 (Crank-rocker) The canonical example of a simple useful linkage is the Grashof 4-bar, or crank-rocker linkage, used extensively in mechanical components.
Theorem 1.4.
Any smooth compact manifold is dieomorphic to the conguration
space of some planar linkage.
This remarkable result provides great consolation to students whose ability to conceptualize geometric dimensions greater than three is limited: one can sense all the complexities of manifolds by hand via kinematics. The reader is encouraged to build a few congurable linkages and to determine the dimensions of the resulting conguration spaces.
Example 1.5 (Robot arms) A robot arm is a special kind of mechanical linkage in
which joints are sequentially attached by rigid rods. One end of the arm is xed
1.3. Derivatives
(mounted to the oor) and the other is free (usually ending in a manipulator for manufacturing, grasping, pick-n-place, etc.). Among the most commonly available joints are pin joints (cf. an elbow) and rotor joints (cf. rotation of a forearm), each with conguration space S1 . Ignoring the (nontrivial!) potential for collision of an arm in a convoluted conguration, the conguration space of this arm in R3 has the 1 n topology of the n-torus, T = n 1 S , the cross product of n circles, where n is the number of rotational or pin joints. There are natural maps associated with this conguration space, including the map to R3 which records the location of the end of the arm, or the map to SO3 that records the orientation (but not the position) of an asymmetric part grasped by the end manipulator.
1.3 Derivatives
Derivatives, vector elds, gradients, and more are familiar constructs of calculus that extend to arbitrary manifolds by means of localization. Dierentiability is a prototypical example. A map between manifolds f : M ! N is dierentiable if pushing it down via charts yields a dierentiable map. Specically, whenever f takes p 2 U M to f (p ) 2 V N , one has f 1 a smooth map from a subset of Rm to a subset of Rn . The derivative of f at p is therefore dened as the derivative of f 1 , and one must check that the choice of chart does not aect the result. It suces to use charts and coordinates to understand derivatives, but it is not satisfying. A deeper inquiry leads to a signicant construct in dierential topology. The tangent space to a manifold M at a point p 2 M , Tp M , is a vector space of tangent directions to M at p , where the origin of Tp M is abstractly identied with p itself. This notion is the rst point of departure from the calculus mindset in elementary calculus classes there is a general confusion between tangent vectors (e.g., from vector elds) and points in the space itself. It is tempting to illustrate the tangent space as a vector space of dimension dim M that is tangent to the manifold, but this pictorial representation is dangerously ill-dened in what larger space does this tangent space reside? Do dierent tangent spaces intersect? There are several ways to correctly dene a tangent space. The most intuitive uses smooth curves. Dene Tp M to be the vector space of equivalence classes of dierentiable curves : R ! M where (0) = p . Two such curves, and ~ are equivalent if and only if 0 (0) = ~ 0 (0) in some (and hence any) chart. An element of Tp M is of the form v = [ 0 (0)], where [] denotes the
equivalence relation. The vector space structure is inherited from that of the chart in Rn . A tangent vector coincides with the intuition from calculus in the case of M = Rn . Invariance with respect to charts implies that the derivative of f : M ! N at p 2 M is realizable as a linear transformation Dfp : Tp M ! Tf (p) N . In any particular chart, a basis of tangent vectors may be chosen to realize Dfp as the Jacobian matrix of partial derivatives at p . The next step is crucial: one forms the disjoint union of all tangent spaces Tp M , p 2 M , into a self-contained space T M called the tangent bundle of M . An element of T M is of the form (p; v ), where v 2 Tp M . The natural topology on T M is one for which a neighborhood of (p; v ) is a cross product of a neighborhood of v in Tp M with a neighborhood of p in M . In this topology, T M is a smooth manifold of dimension equal to 2 dim M . For example, the tangent bundle of a circle is dieomorphic to S1 R1 . However, it is not the case that T S2 = S2 R2 . That this is so is not so obvious.
1.4 Vector elds

The formalism of tangent spaces and tangent bundles simplies the transition of calculus-based ideas to arbitrary manifolds; a ready and compelling exemplar (in light of Example 1.9) is the topology and dynamics of vector elds. A vector eld on M is a choice of tangent vectors (p ) 2 Tp (M ) which is continuous in p . Specically, : M ! T M is a (continuous) map satisfying = IdM for : T M ! M the projection map taking a tangent vector at p to p itself. Such a map is called a section of T M . As in the case of dierential equations on Rn , vector elds can be integrated to yield a ow. Given a vector eld on M , the ow associated to is the family of dieomorphisms t : M ! M satisfying: 1. 0 (x ) = x for all x 2 M ; 2. s +t (x ) = t (s (x )) for all d (x ) = (x ). 3. dt t
x 2 M and s; t 2 R;
One thinks of t (x ) as determining the location of a particle at x following tangent to for t units of time. For M non-compact or insuciently smooth, one must worry about existence and uniqueness of solutions: such questions are not considered in this text. Smooth vector elds on compact manifolds always yield smooth ows. The dynamics of vector elds and ows is the fundamental overlap between topology and dierential equations.
Example 1.6 (Equilibria) The primal objects of inquiry in dynamics are the equilibrium solutions: a vector eld is said to have a xed point at p if (p ) = 0. An isolated xed point may have several qualitatively distinct features based on stability.
1.5. Braids and robot motion planning The
stable manifold of a xed point p is the set
W s (p) = fx 2 M : tlim (x ) = pg: !1 t

For typical xed points of a typical vector eld, W s (p ) is in fact a manifold, as is the unstable manifold, W u (p ), dened by taking the limit t ! 1 in (1.1) above. A sink is a xed point p whose stable manifold contains an open neighborhood of p ; such equilibria are fundamentally stable solutions. A source is a p whose unstable manifold contains an open neighborhood of p ; such equilibria are fundamentally unstable. A saddle equilibrium satises dim W s (p ) > 0 and dim W u (p ) > 0; such solutions are, like the Nash equilibria they encompass, minimax solutions.
(1.1)
Example 1.7 (Periodic orbits) If a continuous-time dynamical system a ow
is imagined to be supported by the skeleton of its equilibria, then the analogous circulatory system would be comprised of the periodic orbits. A periodic orbit of a ow is an orbit ft (x )gt 24 satisfying t +T (x ) = (x ) for some xed T > 0 and all t 2 R. The minimal such T > 0 is the period of the orbit. Periodic orbits are submanifolds dieomorphic to S1 . One may classify periodic orbits as being stable, unstable, saddle-type, or degenerate. The existence of periodic orbits in contrast to that of equilibria is a computationally devilish problem. On S3 , it is possible to nd smooth, xed-point-free vector elds whose set of periodic orbits is, respectively: (1) all of S3 (e.g., the Hopf eld) [?]; (2) empty [?]; or (3) innite, expressing all types of knotting of loops possible [?]. The dynamics of vector elds goes well beyond equilibria and periodic orbits (see, e.g., [?]); however, for typical systems, the blood-and-bones of periodic orbits and equilibria, together with the musculature of their stable and unstable manifolds, gives the basic structure for reasoning about the body of behavior.
1.5 Braids and robot motion planning

A dierent class of conguration spaces is inspired by applications in multi-agent robotics. Consider an automated factory equipped with mobile robots. A common goal is to place several such robots in motion simultaneously, controlled by an algorithm that either guides the robots from initial positions to goal positions (in a warehousing application), or executes a cyclic pattern (in manufacturing applications). These robots are costly and cannot tolerate collisions. As a rst step at modeling such
a system, assume the location of each robot is a mathematical point in a space X (typically a domain in R2 or R3 ). The conguration space of n distinct labeled points on a topological space X , denoted Cn (X ), is the space ( n ) n C (X ) = X ; = f(xi )n (1.2) 1 : xi = xj for some i 6= j g:
1
The set , the pairwise diagonal, represents those congurations of n points in X which experience a collision this is the set of illegal congurations for the robots. Of course, in practice, robots are not point-like, and near-collisions are unacceptable. From the point of view of topology, however, removing a suciently small neighborhood of gives a space equivalent to Cn (X ), and the conguration space Cn (R2 ) forms an acceptable model for robot motion planning on an unobstructed oor. There are applications for which labeling the points is important, with warehousing, in which robots move specic packages, being a prime example. However, in some settings, such as mobile security cameras in a building, anonymity is not detrimental any camera will do. The unlabeled conguration space, denoted UCn (X ), is dened to be the quotient of Cn (X ) by the natural action of the symmetric group Sn which permutes the ordered points in X . UCn (X ) = Cn (X )=Sn : This space is given the quotient topology: a subset in UCn (X ) is open if and only if the union of preimages in C(X ) is open. Conguration spaces of points on a manifold M are all (non-compact) manifolds of dimension dim Cn (M ) = n dim M . The space Cn (S1 ) is homeomorphic to (n 1)! disjoint copies of S1 Rn 1 , while UCn (S1 ) is a connected space.
is very relevant to mobile robot motion planning (e.g., on a factory oor); it is also among the most topologically interesting conguration spaces. Consider the case of n robots which begin and end at xed congurations, tracing out non-colliding routes in between. This complex motion corresponds to a path in Cn (R2 ), or, perhaps, UCn (R2 ), if the robots are not labeled. (If the path has the same beginning and ending conguration, it is a loop, an image of S1 in the conguration space.) How many dierent ways are there for the robots to wind about one another en route from their starting to ending locations? The space-time graph of a path in conguration space is a braid, a weaving of strands tracing robot paths. A deformation in the motion plan equals a homotopy of the path (xing the endpoints), which itself corresponds to moving the braid strands in such a way that they cannot intersect. From this, and a few sketches, the reader will reason correctly that there are innitely many inequivalent motion plans between starting and ending congurations.
2
Example 1.8 (Braids) The conguration space of points on
1.6. Transversality
both in modeling complex mechanisms or automated guided vehicles. One of many useful techniques for controlling robotic systems to execute behaviors and avoid collisions is to place a vector eld on the conguration space and use the resulting ow to guide systems towards a goal conguration. The task: given a conguration space X and a specied goal point or loop in X , construct an explicit vector eld on X having the goal as a global attractor, so that (almost) all initial conditions will ow to and remain on the goal. Programming such a vector eld is a challenge, requiring explicit coordinate systems. Additional features are also desirable: near the collision set (e.g., from Equation (1.2)), the vector eld should be pointed away from the collision set. For example, consider the problem of mobilizing a pair of robots on a circular track so as to patrol the domain while remaining as far apart as possible. This can be done by means of a vector eld on C2 (S1 ). Coordinatize C2 (S1 ) T2 as f(1 ; 2 ) ; 1 6= 2 g. The reader may check that the following vector eld has f1 = 2 g as an attracting periodic solution and f1 = 2 g as a repelling set:
Example 1.9 (Navigation elds) The use of conguration spaces in robotics is widespread,
_1 = 1 + sin(1 2 ) ; _2 = 1 + sin(2 1 ):
All initial conditions evolve to this desired state of circulation on S1 . This is an excellent model not only for motion on a circular track, but also for an alternating gait in legged locomotion, where each leg position/momentum is represented by S1 . One reason for specifying a vector eld on the entire conguration space (instead of simply dictating a path from initial to goal states) is so that if the robot experiences an unexpected failure in executing a motion, the vector eld automatically corrects for the failure. Enlarging the problem from path-planning to eld-planning returns stability and robustness. This technique has been successfully applied in visual servoing [?], in robots that juggle [?], in automated guided vehicles [?], and in hopping- [?], walking[?], and insectoid- [?] robot locomotion. This approach to robot motion planning makes extensive use of dierential equations and dynamics on manifolds of conguration spaces.
1.6 Transversality
Genericity is often invoked in applications, but seldom explained in detail. Intuition is an acceptable starting point: consider the following examples of generic features of smooth manifolds and mappings: 1. 2. 3. 4. 5. Two intersecting curves in R2 generically intersect in a discrete set of points. Three curves in R2 generically do not have a point of mutual intersection. Two curves in Rn generically do not intersect for n > 2. Two intersecting surfaces in R3 generically intersect along curves. A real square matrix A is generically invertible.
10 6. 7. 8. 9. 10.
Chapter 1. Manifolds Critical points of a R-valued function on a manifold are generically discrete. The roots of a polynomial are generically distinct. The xed points of a vector eld are generically discrete. The conguration space of a planar linkage is generically a manifold. A generic map of a surface into R5 is injective.
Some of these seem obviously true; others less so. All are provably true with precise meaning using the theory of transversality. Two submanifolds V; W in M are transverse, written V J W , if and only if,
Tp V Tp W = Tp M 8 p 2 V \ W:
(1.3)
Otherwise said, the tangent spaces to V and W span that of M . For example, two planes in R3 are transverse if and only if they are not identical. Note that the absence of intersection is automatically transverse. A central theme of topology is the lifting of concepts from spaces to maps between spaces. The notion of transverse maps is the rst of many examples of this principle. Two smooth maps f : are transverse, written f J g , if and only if:
V !M
and
g:W !M
(1.4)
Dfv (Tv V ) Dgw (Tw W ) = Tp M 8f (v ) = g (w ) = p:
This means that the degrees of freedom in the intersection of images of f and g span the full degrees of freedom in M . Note that submanifolds V; W M are transverse if and only if the inclusion maps V : V ! M and W : W ! M are transverse as maps. Likewise, a map f is transverse to a submanifold W M if and only if f J W . This mapcentric denition does not constrain the images of maps to be submanifolds. This is one hint that dierential tools are ecacious in the management of singular behavior. A point q 2 N is a regular value of f : M ! N if f J fq g. This is equivalent to the statement that, for each p 2 f 1 (q ), the derivative is a surjection the matrix of Dfp is of rank at least n = dim N . One benet of transversality is that localized linear-algebraic results can be pulled back to global results.
Theorem 1.10 (Preimage Theorem). Consider a dierentiable between smooth manifolds. If f J W for W N a submanifold,
submanifold of
map then
M of dimension
dim
f :M!N f 1 (W ) is a
(1.5)
f 1 (W ) = dim M + dim W
dim
N:
1.6. Transversality
11
The proof of the Preimage Theorem relies on linearization and a simple fact from linear algebra: for a linear transformation T : U ! V ,
dim ker
T = dim U + dim coker T
dim
V:
(1.6)
This provides an eective means of constructing manifolds without the need for explicit charts: it is often used in the context of a regular value of a map. 1. The sphere Sn is the inverse image of 1 under f : Rn+1 ! R given by f (x ) = kx k.
It is a manifold of dimension n = (n + 1) 1 + 0. Tnnis the inverse image f 1 (1; : : : ; 1) of the map f : Cn ! Rn given 2. The torus by f (z ) = k =1 kzk k. Its dimension is n = 2n n + 0. 3. The matrix group On of rigid rotations of Rn (both orientation preserving and reversing) can be realized as the inverse image f 1 (I ) of the map from n-by-n real matrices to symmetric n-by-n real matrices given by f (A) = AAT . The 1 1 dimension of On is n2 2 n(n + 1) + 0 = 2 n(n 1). 4. The determinant map restricted to On is in fact a smooth map to the 0-manifold S0 = f1g. As such, SOn , as the inverse image of +1 under this restricted det, is a manifold of the same dimension as On , which is therefore seen to be a disjoint union of two compact manifolds.
In the above examples, the transversality condition is checked by showing that the mapping f has a derivative of full rank at the appropriate (regular) value. Such regularity seems to fail rarely, for special singular values. This intuition is the driving force behind the Transversality Theorem. A subset of a topological space is said to be residual if it contains a countable intersection of open, dense subsets. A property dependent upon a parameter 2 is said to be a generic property if it holds for in a residual subset of even when that subspace is not explicitly given. For reasonable (e.g., Baire) spaces, residual sets are dense, and hence form a decent notion of topological typicality.
For M and N smooth manifolds and W N a submanifold, the set of smooth maps f : M ! N with f J W is residual in C 1 (M; N ), the space of all smooth (innitely dierentiable) maps from M to N . If W is closed, then this set of transverse maps is both open and dense.
Theorem 1.11 (Transversality Theorem).
The proof of this theorem relies heavily on Sard's Lemma: the regular values of a suciently smooth map between manifolds are generic in the codomain.
Example 1.12 (Fixed points of a vector eld) The xed point set of a dieren-
tiable vector eld on a compact manifold M is generically nite, thanks to transversality. Recall from 1.4 that a smooth vector eld is a section, or smooth map : M ! T M with ((p )) = p for all p 2 M . The zero-section T M is the set f(p; 0)g of all zero vectors. The xed point set of is therefore the preimage 1 ( ). For a generic perturbation of , this set is a submanifold of dimension
dim Fix() = dim
M + dim M
dim T M = 0:
12
A 0-dimensional compact submanifold is a nite point set. With more careful analysis of the meaning of transversality, it can be shown that the type of xed point is also constrained: on 2-d surface, only source, sinks, and saddles are generic xed points (cf. 3.3).
Example 1.13 (Beacon alignment) Consider three people walking along generic smooth
paths in the plane
R . How often are their positions collinear? One considers the map f : C (R ) ! R which computes
2 3 2
the (signed) area of the triangle spanned by the three locations at an instant of time:
f (v1 ; v2 ; v3 ) = det [v2 v1 ; v3 v1 ] :

This map has zero as a regular value, and Theorem 1.10 implies that the set of collinear congurations is a submanifold W = f 1 (0) of C3 (R2 ). A set of three paths is a generic map from R (time) into C3 (R2 ). It follows from Theorems 1.11 and 1.10 that, generically, one expects collinearity at a discrete set of times, since
dim
+ dim W
dim C3 (R2 ) = 1 + 5
6 = 0:
This, moreover, implies a stability in the phenomenon of collinearity: at such an alignment, a generic perturbation of the paths perturbs where the alignment occurs, but does not remove it.
1.7 Signals of opportunity

Applications of transversality are alike: (1) set up the correct maps/spaces; (2) invoke transversality; (3) count dimensions. This has some simple consequences, as in computing the generic intersection of curves and surfaces in R3 . Other consequences are not so obvious. The following theorem states that any continuous map of a source manifold into a target manifold has, after generic perturbation, a submanifold as its image when the dimension of the target is large enough. How large? The critical bound comes, as it must, from the proper transversality criterion and a dimension count:
Theorem 1.14 (Whitney Embedding Theorem).
Any continuous function
f :M! M is
between smooth manifolds is generically perturbed to a smooth injection when
dim
N > 2 dim M .
Proof. The conguration space C2 M = M M

a manifold of dimension 2 dim
C2 f : C2 M
2
M.
The graph of
M of two distinct points on induces a map
!C M N N
(x; y ) 7! (x; y; f (x ); f (y )):
1.7. Signals of opportunity
13
The set of points on which f is non-injective is (C2 f ) 1 (C2 M N ), where N N N is the diagonal. According to the appropriate! transversality theorem, generic perturbations of f induce generic perturbations of C2 f . From Theorems 1.10 and 1.11, the generic dimension of the non-injective set is:
dim C2 M + dim(C2 M N ) dim(C2 M N N ) = 2 dim
dim
N < 0:
Thus, there are no self-intersections, and the result is a smooth submanifold. A simple application of Whitney's theorem, due to Michael Robinson [?] informs a problem of localization via signals. Consider a scenario in which one wants to determine location in an unknown environment. Certainly, a GPS device would suce. Such a device works by receiving signals from multiple satellite transmitters and utilizing known data about the transmitters to determine location, within the tolerances of the signal reception. This sophisticated system requires many independent components to operate, including geosynchronous satellites, synchronized clocks, and more, all with nontrivial power constraints. Though wonderfully useful, GPS devices are not universally available: they do not operate underwater or indoors; they are unreliable in urban canyons; the need for satellites and synchronized clocks can limit availability. There is no reason, however, why other signals could not be used. In fact, passive signals arising naturally from TV transmissions, radio, even ionospheric waves induced by lightning strikes [] provide easily measured pulses with which to attempt to reconstruct location. There is a small but fascinating literature on the use of such signals of opportunity to solve localization and mapping problems. The following is a simple mathematical model for localization via signals of opportunity. Consider a connected open domain D Rk which is a k -manifold. Assume there exist N transmitters of xed location which asynchronously emit pulses whose times of arrival can be measured by a receiver at any location in D. Given a receiver located at an unknown point of D, do the received TOA (time of arrival) signals uniquely localize the receiver? Consider the signal prole mapping T : D ! RN which records the TOA of the (received, identied, and ordered) transmitter pulses. If we assume that the generic placement of transmitters associated with this system provide a generic perturbation to T within C 1 (D; RN ), then the resulting perturbation embeds the domain D smoothly for N > 2k . Thus, the mapping T is generically injective, implying unique channel response and the feasibility of localization in D via TOA. For example, this implies that a receiver can be localized to a unique position in a planar domain D using only a sequence of ve or more transmission signals, globally readable from generically-placed transmitters. Note: it is not assumed that the signals are perfectly spherical, nor does the receiver compute any distances-to-transmitters. In practice, it may be most feasible to realize a localization scheme a posteriori, performing a reconstruction of the receiver location after-the-fact. ! This follows from the more general multi-jet transversality theorem [?, Thm. 4.13]
14
This result is greatly generalizable [?]. First, one may modify the codomain to record dierent signal inputs. For example, using TDOA (time dierence of arrival) merely reduces the dimension of the signal codomain by one and preserves injectivity for suciently many pulses. Second, one may quotient out the signal codomain by the action of the symmetric group SN to model inability to identify target sources. This does not change the dimension of the signals codomain: the system has the same number of degrees of freedom. The only change is that the codomain has certain well-mannered singularities inherited from the action of SN . It follows from transversality that any reasonable signal space of sucient dimension preserves the ability to localize based on knowledge of the image of D under T . Transversality and dimension-counting provide a critical bound on the number of signals needed to disambiguate position, independent of the types of signals used.
1.8 Stratied spaces

The application of Whitney's Theorem to signal localization in the previous subsection is questionable in practice. Signals do not propagate unendingly, and the physical realities of signal reection/echo, multi-bounce, and diraction conspire to make manifold theory less than applicable in this setting. The addition of signal noise further frustrates a dierential-topological approach. Finally, the assumption that D is a manifold is a poor one. In realistic settings, the domain has a boundary: signals are bouncing o of walls, building exteriors, and other structures that, at best, are piecewise-manifolds. One approach to this last diculty is to enlarge the class of manifolds. A k -manifold with boundary is a space locally homeomorphic to either Rk or Rk 1 [0; 1), with the usual compatibility conditions required for a smoothing. The boundary of D, @ D, is therefore a manifold of dimension k 1. Many of the tools and theorems of this chapter (e.g., tangent spaces and transversality) apply with minor modications to manifolds with boundary. Yet this is not enough in practice: further generalization is needed. A k -manifold with corners is a space, each point of which has a neighborhood locally homeomorphic to { } x 2 Rk : xi 0; i = 1; : : : ; m ; for some 0 m k , where m may vary from point to point. A true manifold has m = 0 everywhere; a manifold with boundary has m 1. The analogues of smoothings, derivatives, tangent spaces, and other constructs are not dicult to generate. The boundary of a manifold with corners no longer has the structure of a smooth manifold, as, e.g., is clear in the case of a cube I n = [0; 1]n . Note however, that such a boundary is assembled from manifolds of various dimensions, suitably glued together. Such piecewise-manifolds are common in applications. Consider the solution to a polynomial equation p (x ) = 0, for x 2 Rn . An application of transversality theory shows that the
1.8. Stratied spaces
15
solution set is, for generic choices of coecients of p , a manifold of dimension n 1; however, nature does not always deal out such conveniences. Innumerable applications call for the solution to a specic polynomial equation. The null set of a polynomial, even when not a true manifold, can nevertheless be decomposed into manifolds of various dimensions, glued together in a particular manner. There is a hierarchy of such stratied spaces which deviate from the smooth regularity of a manifold. An intuitive denition of a stratied space is a space X , along with a nite partition X = [i Xi , such that each Xi is a manifold of dimension dim Xi = ni . Some control over how these manifolds are pieced together is needed but not integral to this discussion: a stratied space is a piecewise-manifold. Typical examples of stratied spaces include singular solutions to polynomial or real-analytic equations. More physical examples are readily generated. Recall the setting of planar linkage conguration spaces. The 4bar mechanism gives a 1-d manifold, except when the lengths satisfy L1 = L3 and L2 = L4 . In this case, the conguration space is a pair of circles with singularities either two or three depending on whether or not all the lengths are the same. These singularities have physical signicance: they correspond to congurations which are collinear. Upon building such a linkage, one can feel the dierence as it passes through a singular point.
Notes
1. Homeomorphic manifolds are not always dieomorphic. In dimensions three and below, they are. Each Rn has a unique smooth structure except n = 4, which has an uncountable number of exotic smoothings. Spheres S7 have exactly 27 distinct smooth structures [?]. It is unknown if S4 has non-standard smoothings. 2. A manifold is orientable if it has an atlas such that all transition maps 1 between charts preserve orientation (have determinants with positive derivative). The projective plane P2 and Klein bottle K 2 are well-known non-orientable surfaces. 3. It is instructive and highly recommended to build a complex mechanical linkage and investigate its topology la main. For planar linkages, at cardboard, wood, metal or plastic with pin joints works well. For 3-d linkages, the author uses wooden dowels
16
with latex tubing for rotational joints. Unfortunately, no higher spatial dimensions are available for resolving intersections between edges in spatial linkages. With practice, the user can tell the dimension of the conguration space by feel, without trying to compute (past dimension 10, discernment seems to fail). Other tools of advanced calculus follow in patterns similar to those of derivatives, including dierential forms, integration of forms, Stokes' Theorem, partial dierential equations, and more. The reader unfamiliar with calculus on manifolds should consult one of the many excellent introductions [?, ?], some tuned to applications in mechanics [?, ?]. Transversality is a topological approach to genericity. Probability theory oers complementary and, often, incommensurate approaches. There are some phenomena which are topologically generic, but have small (even zero?) measure in a desired probability measure. Caveat lictor. One must be careful in proving genericity results, as the specication of a topology on function spaces is required. The Transversality Theorem works with the C r topology, for r 2 N or r = 1. Tweaking the topology of the function spaces allows for relative versions, which allow perturbations to one domain while holding the function xed elsewhere. Higher derivative data associated to a map f : M ! N is encoded in the r -jet j r f taking values in a jet bundle J r (M; N ). This j r f records, for each p 2 M , the Taylor polynomial of f at p up to order r . As always, the computations are done at the chart level and shown to be independent of coordinates. The Jet Transversality Theorem says that, for any submanifold W J r (M; N ) of the jet bundle, the set of f : M ! N whose r -jets are transverse to W is a residual subset of C 1 (M; N ). The study (and even the denition) of stratied spaces is much more involved than here indicated. Various forms of the Transversality Theorem apply to stratied spaces [?]; they are sucient to derive a form of Whitney's Theorem on embeddings and allow for unique channel response in a transmitter-receiver system outtted with corners, walls, and reections.
4.
5.
6.
7.
8.
Exercises
1.1. Compute the dimension dim Gn k of the Grassmannian. 1.2. What is the conguration space of a 4-bar linkage in R3 ? Remember to factor out the action of the Euclidean group (translations and rotations). What happens in the degenerate cases having collinearities? 1.3. What is the conguration space of a unicycle riding on the plane R2 ? What is the conguration space of an airplane in R3 ? [Do not include velocities, only consider positions and orientations, constrained by allowing motion only to turn or to move forward.] 1.4. Can a planar mechanical linkage admit a conguration space that is non-orientable? 1.5. Compute the homeomorphism types of C2 (R3 ) and C3 (R3 ). 1.6. Show that UC2 (S1 ) is homeomorphic to a Mbius band (without boundary). This example is the basis for various musings about perceptions of musical chords [?]. One assumes a topology on the space of frequencies of notes yielding S1 (octaves are identied). A chord is considered to be a point in a conguration space of points on
Exercises
17
a circle. For musical chords, these congurations are unlabeled. 1.7. Consider the conguration space of a ray in R3 based at the origin: it is homeomorphic to S2 , naturally. Remove K disjoint compact balls from R3 (none containing the origin) and consider the resulting conguration space of rays which may not intersect any of the balls. Find examples for which the resulting conguration space is not connected. 1.8. The codimension of a submanifold V M is codimM W = dim M dim W . Restate Theorem 1.10 more concisely in terms of codimension. 1.9. Use codimension and transversality (and some ridiculous physical assumptions) to argue which is the more dangerous situation in which to nd oneself: a mineeld, a tornado storm, or a hurricane? 1.10. Use transversality to prove that polynomials with real coecients generically do not have repeated real roots. 1.11. Modify the proof of Lemma 1.12 to show that a self-map of a compact manifold f : M ! M has, generically, a nite number of xed points. 1.12. Prove that the boundary of a manifold with boundary is a manifold of codimension one. 1.13. Instead of TOA or TDOA signals, assume DOA (direction of arrival), in which the receiver in Rk detects the direction (point in Sk 1 ) of each transmitter. How many transmitters must be heard (as a function of k ) to ensure disambiguation? What goes wrong in the case k = 1? [?]
Chapter 2
Spaces: Complexes
20
Chapter 2. Complexes change in focus from spaces that resemble the smooth surfaces of advanced calculus to those assembled from simple pieces motivates the eventual transition to algebraic means of assembling spaces and invariants.
2.1 Simplicial and cell complexes

Let S be a discrete set and P(S ) its power set the set of all subsets. An abstract simplicial complex is a collection of nite subsets of S , X P(S ), closed from below by restriction: for each 2 X , all subsets of are also in X . Each element 2 X is called a simplex, or rather a k -simplex, where j j = k + 1. Given a k -simplex , its faces are the simplices corresponding to all subsets of S . For example, if S = fsi gi , the subcollection fs2 ; s5 ; s7 g is a 2-simplex with three 1-simplex faces, three 0-simplex faces, and, of course, the empty-set face. The most familiar simplicial complexes are abstract graphs, simplicial complexes without simplices of dimension higher than one. An abstract graph is often presented as a pair (V; E ), where V is the set of vertices (or 0-simplices) and E is a collection of distinct unordered pairs in V . These edges are the 1-simplices of the graph. Abstract graphs are ideal for expressing pairwise relations between objects. Relations of a higher order point to simplicial models. One class of elementary but illuminating examples occurs naturally in the form of independence of objects. Given a nite collection of objects O = fx g, an independence system is a collection of unordered subsets of O declared independent. Any such system must be closed with respect to restriction (any subset of an independent set is also independent) and thus denes a simplicial complex: the independence complex IO is the abstract simplicial complex on vertex set O whose k -simplices are collections of k + 1 independent objects.
Example 2.1 (Statistical independence) Independence of objects occurs in multiple contexts, including linear independence of a collection of vectors or linear independence of solutions to linear dierential equations. More subtle examples include statistical independence of random variables: recall that the random variables X = fXi gn 1 are statistically independent if their probability densities fXi are jointly multiplicative, i.e., the probability density fX of the combined random variable (X1 ; : : : ; Xn ) satises fX = i fXi . The independence complex of a collection of random variables compactly encodes statistical dependencies in a topological manner.
As with the setting of manifolds, one should rapidly metabolize the formal definition and progress to drawing pictures. One topologizes a simplicial complex X as
2.1. Simplicial and cell complexes a quotient space built from topological simplices. The space:
21
standard k -simplex is the
xi = 1 : i =0 The faces of k are copies of j for j < k via restriction of Rk +1 to coordinate subspaces; via this restriction, one sees how these faces are glued along the boundary of k . To topologize an abstract simplicial complex X , one takes a copy of k for each k -simplex of X , and glues them together respecting faces. This ill-named geometric realization of X is performed inductively as follows. Dene the k -skeleton of X , k 2 N, to be the quotient space: ( )/ (k ) (k 1) k X = X ;
:dim =k
k =
2 [0; 1)k +1 :
(2.1)
where is the equivalence relation that identies faces of k to the corresponding faces of in X (j ) for j < k . Thus, X (0) is a discrete set, equivalent to S . The 1-skeleton X (1) is a space homeomorphic to a collection of vertices and edges connecting vertices: a topological graph. By construction, X (k ) X (k 1) and one therefore identies X with 1 X = X (k ) :
k =0 The abuse of notation is intentional; the topological space X will be used interchangeably with the abstract simplicial complex X . The space X is given the weak topology: a subset U X is open if and only if U \ X (k ) is open for all k . A simplicial complex is one of many possible data structures for building a space. For example, one may take the basic k -dimensional building block to be a k -cube I k , where I = [0; 1]. The faces of the cube are obviously smaller-dimensional cubes; attaching maps between faces identify cubes and are likewise clear. The resulting class of cubical complexes are quite natural in many applications. Digital cameras store data in pixels on a 2-dimensional cubical complex. Many applications in numerical and nite element analysis work well on 3-dimensional cubes (or voxels) arranged in a lattice. More general still is the class of spaces referred to in this text as cell complexes and in the topology literature as CW complexes. These have as their basis cells closed
22 balls B k , and for attaching maps, @Bk to X (k 1) .

arbitrary
Chapter 2. Complexes continuous functions from the boundary
The reader may nd it simpler to work with cell complexes whose boundary maps are homeomorphisms of @B k to their images, but the more general cell complexes are certainly omnipresent. This text falls back to the simplicial and cubical setting frequently, but the reader should note the ubiquity and utility of CW/cell complexes in the literature. For other specialized types of complexes (e.g., prodsimplicial and -) see [?, ?]. The remainder of this chapter is a menagerie of interesting simplicial and cellular complexes that have been found useful in applications.
Consider a discrete subset Q Rn of Euclidean space. If one receives Q as a collection of data points a point cloud sampled from a submanifold, it may be desirable to reconstruct the underlying submanifold from the point cloud. A graph based on a point cloud Q can be ecacious in discerning shape, as any child doing a dot-to-dot can attest, but a simplicial complex may improve matters. Choose a constant > 0. The Vietoris-Rips complex of scale on Q, VR (Q), is the simplicial complex whose simplices are all those nite collections of points in Q of pairwise-distance . Note the diculty in choosing the correct to best approximate the underlying structure. Although this complex is based on concrete points in Rn , there may be simplices of dimension much larger than n (when suciently many points are crowded together). The degree to which a Vietoris-Rips complex VR of xed accurately captures the topology of the point cloud is mixed. Consider a point cloud Q in Rn and the projection map : VR ! Rn taking the vertices VR(0) to Q and taking a k -simplex of VR to the convex hull of the associated vertices in Q. The image of in Rn is called the shadow, S, of the Vietoris-Rips complex. There is no hope of VR and S being homeomorphic, as the domain of the projection is likely to have higher dimension than the codomain. Even homotopy equivalence is too much to ask, as it is possible to arrange data points in Q so as to have a simplicial sphere Sk for arbitrary k > 1 which projects in S to a convex set: such bubbles are artifacts of the Vietoris-Rips complex and not intrinsic to the point cloud. Nevertheless, for large-scale topological holes, the Vietoris-Rips complex seems to capture the topology correctly.
2.2 Vietoris-Rips complexes and point clouds
2.3. Witness complexes and the sampling problem
23
2.3 Witness complexes and the sampling problem

The Vietoris-Rips complex of a large data set is too large to store and manipulate, even with the simplication to the input problem implicit in the ag complex structure. The witness complex constructions of Carlsson and de Silva [?] greatly reduce size at the expense of topological precision. This is by now a multifarious subject, with a variety of related denitions built to suit dierent data sets. The following is a elementary version, suitable as an introduction. Given a (large) point cloud of nodes Q in a Euclidean space Rn , choose a (small) set of landmarks L Q. The weak witness complex, WL , is an abstract simplicial complex on the vertex set L dened as follows. A subcollection of points S L is a simplex in the witness complex if and only if for each subset T S there is a weak witness xT 2 Q with every point in T closer to xT than to any point in L T . Simple examples indicate that for a reasonable choice of landmarks, the witness complex provides a decent topological approximation to the underlying structure of the point cloud. More recent generalizations to strong and parameterized witness complexes span the gap between computability and rigorous recovery of the correct topological type [?].
2.4 Flag complexes and networks

The Vietoris-Rips complex is an example of a more general class of simplicial complexes that ll a frame. One of the advantages of this construction is its parsimonious representation with regards to input and storage of the data. To encode a typical simplicial complex into a form that computational software can manage, it is necessary to list all the simplices. In contrast, note that the Vietoris-Rips complex is entirely specied by knowing the vertices Q and the edges (depending on ). The remaining, higher-dimensional, simplices are all accounted for: the graph determines the complex. Given a graph, the ag complex" is the maximal simplicial complex having the graph as its 1-skeleton. Otherwise said, whenever there appears to be the outline of a simplex, one lls it in. The Vietoris-Rips complex is the ag complex associated with the distance- graph of a point cloud Q. Flag complexes also arise as higher-dimensional models of communications networks. Consider a collection of points Q R2 that represent the locations of nodes in a wireless communications " This is sometimes called a clique complex, especially in topological combinatorics.
24
Chapter 2. Complexes
network. Ignoring the details of the communications protocols, consider the following simple model of communication: nodes broadcast their unique identities, received by nodes which are within range; these neighbors then establish communication links and in so doing assemble an ad hoc network. Due to irregularities in transmission characteristics, ambient characteristics of the domain, and signal bounce, the communication links formed may not be solely a function of geometric distance between nodes. Nevertheless, forming the ag complex associated with these edges may recover some of the rough structure implicit in the global network. In Chapter 4, a method for addressing coverage-type problems is considered using these ag complexes.
2.5 ech and random samplings

The diculty of building topologically accurate simplicial models is classical. In the context of Vietoris-Rips complexes, it is frustrating to have higher-dimensional features (as, e.g., in the cross-polytopes above) appearing. One way to circumvent this is to sort through the higher-dimensional simplices of the Vietoris-Rips complex more carefully. Given a point cloud Q in Rn and a length parameter to be the simplicial > 0, dene the ech complex C is a colcomplex built on Q as follows. A k -simplex of C lection of k + 1 distinct elements xi of Q such that the net intersection of diameter balls at the xi 's is nonempty. The ech complex is a subcomplex of the Vietoris-Rips complex VR ; it is sometimes a proper subcomplex. The disadvantages of a ech complex are its storage requirements (one cannot simply store the 1-skeleton and ll the rest in) and its construction (one must check many intersections to build the full complex). These are compensated is always homotopic to the union for by a topological accuracy: the ech complex C of diameter- balls about Q. Niyogi, Smale, and Weinberger [?] have used this property of ech complexes to prove a result about randomly sampled point clouds. Consider a smooth submanifold M Rn . Let Q be a collection of points sampled at random from M . It is intuitively obvious that for a suciently dense sampling, the union of suciently sized balls about the points is homotopic to M : this union of balls can be captured by the ech complex of Q. The ability of the ech complex to approximate M is conditional upon the density of sampling and the manner in which M is geometrically embedded in Rn . Let denote the injectivity radius of M : the largest number such that all normal rays to M of length are mutually
2.6. Nerves and neurons nonintersecting.
25
Theorem 2.2 (Niyogi-Smale-Weinberger [?]).

a smooth compact submanifold density of
2 (Q) is homotopic to M . =2 for < 3=5, then the ech complex C
Additional results, such as when Q is a sampling of points from are also available [?] and relevant to the statistics of point clouds.
Q is such
M RN ,
Assume a collection
where
of points on
has injectivity radius
that the minimal distance from any point of
to
.
If the
is less than
with noise,
2.6 Nerves and neurons

As the Vietoris-Rips complex is a metric-ball version of the more general ag complex construction, so also is the ech complex the metric-ball version of a more general object. Consider X an arbitrary topological space. Given a nite collection U = fU g of subsets of X , one builds the nerve of U, N(U), as follows. The k -simplices of N(U) correspond to nonempty intersections of k + 1 distinct elements of U. For example, vertices of the nerve correspond to elements of U; edges correspond to pairs in U which intersect nontrivially. This denition respects faces: the faces of a k -simplex are obtained by removing corresponding elements of U, leaving the resulting intersection still nonempty. Examples of nerves based on, e.g., convex subsets of Rn , suce to suggest the following classical generalization of the result for ech complexes:.
Theorem 2.3 (Nerve Lemma).

the union of elements of
If
is a nite collection of subsets of
non-empty intersections of subcollections of
U.
contractible, then
NU
with all
is homotopic to
One of the more recent and interesting applications of nerves is in neuroscience. The work of Curto and Itskov [?] considers how neural activity can represent external environments. In particular, the authors consider the impacts of external stimuli on certain place cells in the dorsal hippocampus of rats. These cells experience dynamic electro-chemical activity which is known to strongly correlate to the rat's location in physical space. Each such cell group is assumed to determine a specic location in space, and the collection of such location patches,
26
or place elds, forms a cover U satisfying the hypotheses of the Nerve Lemma. It suces to assume that place elds exist and are stable, omni-directional, and have ring elds that are convex. Curto and Itskov argue that these assumptions are generally satised for place elds of dorsal hippocampal place cells recorded from a freely foraging rat in a familiar open eld environment. The following experiment is proposed: place cells are monitored, with the corresponding activity recorded as a time-series. These time series show occasional bursts of activity, such spike trains being common data forms in neuroscience. Cell groups are identied by correlating spike-train rings that occur simultaneously (in practice, within a window of time depending on some experimenter-chosen parameter). The correlation of spiking activity provides the intersection data of the place elds. This, in turn, allows for the computation of the nerve of the place elds based on spike train correlation. These investigations suggest that rats may build a structured representation of their external environment within an abstract space of stimuli. Of note is the lack of metric data the construction of the physical environment is purely topological and can be eected without reference to coordinates. This is an increasingly signicant constraint in biological systems, where metrics can be rare or needlessly articial.
2.7 Phylogenetic trees and links

Mathematical biology is a prime target for topological methods, given the noisy nature of Nature. In contrast to the neuroscience example above is an analogue of conguration spaces associated to genetic data in the form of phylogenetic trees data structures for organizing and comparing taxonomic and genetic data. The relevant objects of study are rooted, end-labeled, metric tress. A tree is a graph without cycles. The leaves of a tree are vertices of degree 1 (only one edge is attached), and a root is a dedicated leaf. A phylogenetic tree is a rooted tree with (1) leaves labeled by data (genes, phenotypes, or other taxonomic data), and (2) all interior edges (those not connected to a lead) labeled with positive real numbers (and represented as the length of the edge). These trees can be readily illustrated as planar graphs, but the embedding of the tree into the plane is not part of the data. The following construction comes from the seminal paper [?]. Let Tn denote the following tree space, the space of rooted metric trees with n labeled (non-root) leaves. For simplicity, assume that the label on the root is 0 and labels on the leaves are 1; 2; 3; : : : ; n. This space has a natural cube complex structure as follows. All non-leaf vertices of the graphs are assumed to be essential in that the number of edges attached is greater than two.
2.7. Phylogenetic trees and links
27
Two such trees are equivalent if there is a mapping between them that preserves the graph structure (takes vertices to vertices and edges to edges), the leaf labels, and the interior edge lengths. Two trees will be considered isomorphic if there is a mapping between them that preserves the graph structure and the labels. A xed isomorphism class of phylogenetic trees is parameterized by the interior edge lengths: this yields an open cube of dimension equal to the number of interior edges. To compactify this to a closed cube, reparameterize all interior edge lengths to the open interval (0; 1) and add the boundary faces. Those faces with one or more 0 factors may be thought of as having the corresponding interior edges collapsed (zero length edges). These edge-collapse faces are identiable as trees of a dierent isomorphism class. By gluing together all such cubes according to isomorphism class, a nite cube complex Tn is obtained. As topological spaces, Tn may seem too simple: it is contractible to an origin, the radial graph with no interior edges, by shrinking each interior edge continuously to length zero. Despite this simplicity, the manner in which the various cubes are assembled is topologically (and geometrically) of interest. The best way to analyze this assembly is through a simplical model called a link. Let X denote a cell complex built from simplices, cubes, or other simple polyhedral basis cells, and let v 2 X (0) be a 0-cell. The link of v in X , `kX (v ), is the simplicial complex whose k simplices consists of (k + 1)-dimensional cells attached to v , with faces inherited from faces of cells in X .
Example 2.4 (Tree spaces) The tree space T3 is a letter Y. The space T4 consists
of 15 2-dimensional squares (corresponding to the 15 possible binary rooted trees with 4 labeled leaves) glued together so as to have a common corner (corresponding to the radial rooted tree with 4 leaves) and a link equal to the petersen graph.
Tn
Theorem 2.5 (Billera-Holmes-Vogtmann [?], Vogtmann [?]).

origin in is a ag complex that has the homotopy type of
(n
1)!
The link of the 3 copies of
Sn
glued together at a single point.
28
There are numerous reasons why scientists and engineers might want to work with spaces of objects rather than merely the objects themselves. The tree spaces Tn (though arising in other contexts in topology and algebraic geometry as certain moduli spaces) were generated in response to challenges in statistics associated to genetic data [?]. Experimental data is used to generate phylogenetic trees, and this data does not spring forth fulled formed from the mind of the researcher: the data are often noisy and not perfectly repeatable. A collection of numerical or vector-valued data can be averaged, but what does it mean to average a sequence of phylogenetic trees in a scientically meaningful manner? This requires a geometry (and thus a topology) on the space of all possible phylogenetic trees.
2.8 Strategy complexes and planning with uncertainty

Many of the complexes of this chapter approximate manifolds or provide models of data; however, complexes also have broad applicability in engineering systems as spaces that collate states. Consider a transition graph, a directed graph G whose vertices V represent states of an abstract system, and whose edges E represent deterministic actions. For concreteness, consider the example [?] of a transportation network, where vertices are locations and edges represent ights, freeways, subways, trains, or other discrete enter-exit modes of transport. Other examples include motion-planning in robotics with discrete states (robots move from landmark to landmark). Planning on the transition graph is straightforward: nd a directed path from initial to goal, then execute the appropriate actions sequentially. However, stepping onto a ight bound for Newark does not guarantee arrival at Newark. Worse still are the kinds of uncertainties precipitated by an adversary. Ill weather and air travel does not exactly t that scheme, but other situations do exhibit adversarial forms of uncertainty for which, if something may go amiss, it must be mitigated. Erdmann [?] models adversarial uncertainty in the transition graph via branched edges. Actions may be deterministically chosen, but once chosen, the outcome is one of several possibilities, out of the control of the chooser. These possibilities are not necessarily chosen probabilistically; an adversary may determine any outcome among the terminal edges. Note that the state chosen by the adversary is known to the game-player, but only after the nondeterministic action has been chosen and executed. In such a setting, the notion of a guaranteed strategy for reaching a goal state despite adversarial meddling seems dicult if not impossible. The question of path-planning within a nondeterministic transition graph is one amenable to simplicial data structures. Given a nondeterministic transition graph G, dene the strategy complex, G , as the abstract simplicial complex dened on the set of actions E (G), in which k -simplices consist of collections of k +1 disjoint actions (including all branches of any chosen nondeterministic action) whose unions contain
2.9. Decision tasks and consensus

no directed cycles
29
. This relation is closed under restriction (a smaller collection of such edges clearly also contains no directed cycle), and is sensible, as a directed cycle means that a skilled adversary can cycle the player through states ad innitum. The intuition is that G consists of trap-free strategies: each simplex of G is a strategy for progressing to one or more states. For g 2 V (G) a goal state, say that G has a complete strategy for attaining g if for N large there exists a mapping from V (G) to E (G) which associates to v an action based at v , so that one attains g with the execution of at most N actions. Note: the goal is certain, but the path is not. The obstruction to attainability lies in the strategy loopback complex: denote by G g the strategy complex of G augmented by deterministic arrows from g to each v 2 V (G) g .
Theorem 2.6 (Erdmann [?]).

dimension
A nondeterministic transition graph
complete strategy for attaining a goal
jV (G)j
2; else it is contractible.
g 2 V (G) i G
possesses a
is homotopic to a sphere of
The proofs of this and related theorems in [?] use nerves extensively. Additional results include realizability of simplicial complexes as strategy complexes, incorporation of stochastic nondeterminism, and decompositions of strategy complexes by means of deterministic subsystems.
2.9 Decision tasks and consensus

A related use of simplicial structures assists in multi-agent decision-making. Consider the problems of consensus or distribution among collaborative agents, a common scenario in biology (animal ocking), sociology (voting, beliefs), engineering systems (network clocks), and robotics (cooperative action). The problem of collaborative consensus has a large literature: mathematical approaches depend sensitively on the modeling assumptions. Recent interest in synchronization phenomena [] focus on graph-theoretical methods. Hidden beneath these 1-dimensional models is a rich higher-dimensional topological structure underlying inuence, consensus, and division.
Example 2.7 Consider the problem of N transmitters attempting to broadcast without interference over a frequency spectrum with K N distinct, non-interfering channels. Build a simplicial complex on the vertex set corresponding to transmitter i choosing channel j , where simplices (up to and including dimension N 1) correspond to several transmitters having made distinct choices. Each n-simplex encodes the setting in which n + 1 of the transmitters have cooperated to share specic chan-
30
nels, while the other N n 1 transmitters are viewed as having failed/declined to choose. These simplices form a complex using the natural boundary map derived from transmitters `forgetting' a choice. For example, three transmitters choosing from four channels gives rise to a simplicial 2-torus T2 . Work of Herlihy and Shavit [?] and their collaborators provides a topological approach to consensus/decision problems in the context of distributed, faulttolerant computation. In this setting, it is presumed that there are N independent processors Pi who share a common read/write memory element for communication purposes. For simplicity and concreteness, assume that there are M > 1 possible labels or states and that each processor has an initial preference. The problem is whether there exists an algorithm for the processors, via read/write communication to the common memory, to come to deterministic nite-time consensus. The interesting catch is that all computations are asynchronous (each processor works at its own speed) and faulty (sometimes, a processor crashes and can engage no further in communication). The problem in reasoning about such systems is that processors which take a very long time to reach decisions are not readily distinguished from processors which have failed and will never respond. A fault tolerant distributed algorithm is one which will terminate in nite time even in the event that some (but not all) processors fail irrevocably. The approached used in [?] begins with a simplicial model of the problem. Given N processors with M possible initial states, dene the input complex I to be the abstract simplicial complex whose k -simplices are labelings of the states of k + 1 distinct processors. The output complex O is the abstract simplicial complex whose k -simplices are labelings of the states of k + 1 distinct processors which are consistent with a desired output. The intuition behind a k -simplex is that k + 1 of the processors exhibit a legal system state and the N (k + 1) remaining processors have failed to report. Because of this, the topology of the complexes (encoded by the simplex faces) correlates with failure of processors. A decision task consists of the input and output complexes, along with a set of constraints. For example, if all N processors begin in consensus with the same label, then the putative consensus algorithm must of necessity output that precise label as the consensus state. Because of the relationship between the simplices, faces, and failures, it is possible to reduce questions of existence of distributed fault-tolerant consensus algorithms in terms of the existence of a map from the input complex to the output complex.
2.10. Discretized graph conguration spaces
31
Example 2.8 In the case of N processors trying to come to binary consensus (i.e.,
two labels, zero and one), the input complex I ' SN 1 is a simplicial sphere and N 1 the output complex O = N 1 is a disjoint pair of simplices: all zeros and all ones. The decision task is to come to consensus, with the proviso that initial consensus terminates the program immediately. Thus, any wait-free fault-tolerant protocol would have to induce a surjective continuous map from I to O. This is impossible, since I is connected and O is not. There is no distributed asynchronous fault-tolerant solution to the problem of binary consensus. This can be shown by other, perhaps more direct means; however, the Herlihy-Shavit theorem has resolved more complex consensus problems for which other solutions were unknown. These deeper examples rely not on common-sense notions of connectivity but rather on holes of higher order the homology theory of Chapter 4. One of the key theorems [?] gives necessary and sucient conditions for consensus stated in terms of existence of maps (with details concerning simplicial subdivisions and colorings). The salient feature of the result is that the topology of the decision task can determine whether an algorithm exists. This type of implicit inference is a hallmark of topological methods.
2.10 Discretized graph conguration spaces

Consider a nite graph G, now thought of not as an abstract system of states, but rather as a physical, geometric highway of paths. Keeping track of vehicles or robots or tokens on G leads, as in 1.5, to conguration spaces. The conguration space of n labeled points on G, Cn (G) is important for motion planning in robotics when the automated vehicles are constrained by tracks, guidewires, or optical paths [?]; however, Cn (G) is in general neither a manifold nor a simplicial/cell complex. The following constructions [?, ?] are approximations to these spaces by cubical complexes. Dene the discretized conguration space of G as:
~; Dn ( G ) = ( G G )
(2.2)
~ denotes the set of all product cells in Gn = G where G whose closures intersect the topological diagonal . Equivalently, Dn (G) is the set of congurations for which, given any two tokens on G and any path in G connecting them, the path contains at least one entire (open) edge. Thus, instead of restricting tokens to be at least some intrinsic distance apart (i.e., removing an neighborhood of ), one now restricts tokens on G to be at least one full edge apart. Note that Dn (G) is a subcomplex of the cubical complex Gn and a subset of Cn (G) (it does not contain partial cells that arise when cutting along the diagonal), and is, in fact, the largest subcomplex of Gn that does not intersect . With this natural cell structure, one can think of the vertices (0-cells) of Dn (G) as discretized congurations arrangements of labeled tokens at the vertices of the
32
graph. The edges of Dn (G), or 1-cells, tell us which discrete congurations can be connected by moving one token along an edge of G. Each 2-cell in Dn (G) represents two physically independent edges: one can move a pair of tokens independently along disjoint edges. A k -cell in Dn (G) likewise represents the ability to move k tokens along k closure-disjoint edges in G in an independent manner.
Example 2.9 (Digital microuidics) An excellent physical instantiation of the discretized conguration space is implied by work on digital microuidics (see, e.g., the work of Fair in [?, ?]). In digital microuidics, small (1mm diameter) droplets of
uid can be quickly and accurately manipulated in an inert oil suspension between two plates. The plates are embedded with a grid of wires, providing discrete localization of droplets via electrowetting a process that exploits current-induced dynamic surface tension eects to propel and position a droplet. Applying an appropriate current drives the droplet a discrete distance along the wire grid. The goal of this and related microuidics research is to create an ecient lab on a chip in which droplets of various chemicals or liquid suspensions can be positioned, mixed, and then directed to the appropriate outputs. Using the grid, one can manipulate many droplets in parallel. The discretized conguration space of droplets on the graph of the electrical grid captures the topology of the multi-droplet coordination problem. Note, however, that collisions are not always to be avoided chemical reactions depend on controlled collisions.
conguration spaces whose structures are illuminated by discretization. The discretized conguration space D2 (K5 ) of two labeled points on a complete connected graph of ve vertices, K5 , is a 2-dimensional cube complex. Fix a single vertex v 2 D2 (K5 ): this corresponds to a pair of distinct labeled tokens on vertices of K5 . From this vertex v 2 D2 (K5 ) emanate six edges: each token can move to any of the three non-occupied vertices of K5 . For each edge e incident to v , the token not implicated by e can move to two of the three available vertices without interfering with e ; thus, e is incident to exactly two 2-cells of D2 (K5 ). There are a total of six such 2-cells incident to v , arranged cyclically as shown. Thanks to the symmetry of K5 , the complex looks the same at each vertex, and D2 (K5 ) is thus a topological 2-manifold: a compact surface. With the tools of the next chapter at our disposal, it will be shown that this surface is of genus six. The discretized conguration space is an accurate cubical model of the topological conguration space.
Example 2.10 (A nonplanar graph) There are several interesting examples of graph
2.11. State complexes and reconguration
33
Theorem 2.11 (Abrams [?]).

Cn (G).
so as to split each edge of
Let
connecting a vertex to itself ) and let
G into n
G be a nite topological graph (with no edges G0 be the subdivision of G that inserts vertices 1 edges. Then Dn (G0 ) is homotopic to Cn (G0 ) =
2.11 State complexes and reconguration

Discretized conguration spaces of graphs are generalizable to a very broad class of systems ranging from metamorphic robots to protein chains which are recongurable based on local rules. The resulting conguration spaces, called state complexes, are cube complexes with interesting topological and geometric properties. Fix a graph G and an alphabet of labels A, used to label the vertices V (G), each such labeling V (G) ! A dening a state u. A recongurable system is a collection of states fu g, closed under the actions of a xed set of local recongurations, or generators. Each generator consists of a support subgraph supp G and an unordered pair of local states, labelings of the vertex set of supp by elements of A. A generator is admissible at a given state u if one of the generator's local states matches the restriction of u to supp . The trace of a generator is the (nonempty) subset of the support trace supp on which the local states dier. A simple familiar example is that of discretized motion-planning on a graph from 2.10. An alphabet A = f0; 1; : : : ; N g labels the vertices of G as empty (0) or occupied with a (numbered) token. Local recongurations are supported on the closure of each edge in G; each generator has supp = trace equal to an edge with local state a labeling of one vertex with 0 and the other 0 < i N for some i , the other local state reversing these labelings. This system has a well-dened notion of a conguration space, DN (G), that generalizes to any recongurable system as follows. A collection of generators f g admissible at a given state u is said to commute if their supports and traces are pairwise non-intersecting:
trace
\ supp = ?
8 6= :
This means that the changes implicated by a generator (its trace) do not impact the applicability of the other generators (their supports): commutativity encodes physical independence. Dene the state complex of a recongurable system to be the cube complex S with an abstract k -cube for each collection of k admissible commuting generators. That is, the 0-skeleton consists of the states in the recongurable system; the 1-skeleton is the graph whose edges pair two states which dier by the local states of a generator; 2-cells correspond to commutative pairs of generators acting on a state; etc. For recongurable systems that admit only nitely many admissible generators at any given state, the state complex S is a locally compact cubical complex.
Example 2.12 (Articulated planar robot arm) This recongurable system models the position of an articulated robotic arm with one end xed at the origin and which can (1) rotate at the end and (2) ip a 2-segment corner from up-right to right-up and vice versa. This arm is positive in the sense that it may extend up and to the
34
right only. A grammatical model is simplest: states are length n words in two letters. Specically, let G be a linear graph on n vertices with an alphabet A = fU; Rg encoding up and right respectively. Generators are of two types: (1) exchange the subwords UR and RU along an edge; (2) change the terminal letter in the word. Supports (and traces) are (1) a closed edge, and (2) the terminal vertex, respectively. Generators commute when the associated subwords are disjoint. The state complex in the case n = 5 has cubes of dimension at most three. In this case also, although the transition graph (the 1-skeleton of S) for this system is complicated, the state complex itself is contractible.
coordination problems. Consider a collection of n planar graphs (Gi )n 1 , each embedded in the plane of a common workspace (with intersections permitted). On each Gi , a robot Ri with some particular xed size/shape is free to translate along edges of Gi . The coordination space of this system is dened to be the space of all congurations in i Gi for which there are no collisions the geometric robots Ri have no overlaps in the workspace. To model this with a recongurable system, let the underlying graph be the disjoint union G = i Gi . The generators for this system are as follows. For each edge 2 E (Gi ), there is exactly one generator . The trace is the (closure of the) edge itself, trace( ) = , and its generator corresponds to sliding the robot Ri from one end of the edge to the other. The support, supp( ) consists of this trace union any other edges in Gj (j 6= i ) for which the robot Rj sliding along the edge can collide with Ri as it slides along . The alphabet is A = f0; 1g and the local states for have zeros at all vertices of all edges in the support, except for a single 1 at the boundary vertices of . Any state of this recongurable system has all vertices labeled with zeros except for one vertex per Gi with a label 1. The resulting state complex is a cubical complex that approximates the coordination space. In the case where Gi = G is the same for all i and the robots Ri are suciently small, this recongurable system has S = Dn (G). It appears to be dicult to characterize which cubical complexes can arise as
Example 2.13 (Robot path coordination) The following example is inspired by robot
2.11. State complexes and reconguration
35
state complexes for some recongurable system: the problem is closely entwined with deep ideas in geometric group theory [?]. A few results are known: e.g., it is impossible to obtain a sphere Sn for n > 1, or, indeed, anything resembling a sphere.
Theorem 2.14 ([?, ?]).
n > 1 into S is homotopic to a constant map
State complexes are
aspherical:
Sn
!?
any map of a sphere
Sn
for
Notes
1. The Nerve Lemma is usually attributed to Leray [ ], whose version, concerning sheaves, is much deeper than that stated, which echoes antecedents in the work of ech [ ] and Alexandrov [ ]. Alexandrov investigated spaces like ech complexes in [ ], prior to the work of ech. 2. Work on
graphical methods#
is an attempt to encode statistical independence that is
reminiscent of Example 2.1. The literature would seem to benet from the language of simplices; however, the directedness associated to conditional independence may explain the convoluted terminology there employed. 3. One virtue of topological methods is their coordinate-free nature: it is not necessary to know exact locations. If, however, one does possess strong geometric data in the form of coordinates, there are numerous ways to improve on the Vietoris-Rips construction. One popular approach is the
alpha complex of Edelsbrunner [?] which restricts a ech
complex by a Delaunay triangulation to produce a topologically accurate simplicial model with computationally small footprint. 4. Theorem 2.2 is remarkable for its explicit computation of density bounds it was previously known [ , the submanifold.
? ?]
that a suciently dense sampling recovers the topology of
The bounds in Theorem 2.2 have been applied to computing the
topological type of data sets in reaction-diusion systems [ ]. 5. The Nerve Lemma is ubiquitous in topology and its applications. that It is often stated
be comprised of convex sets, as these and all nonempty intersections thereof
are contractible. It will sometimes be mentioned that the Nerve Lemma breaks down if the sets involved are not contractible. This statement is true but should propel the reader to more sophisticated methods rather than to despair. 6. A recent survey article [ ] gives a broad treatment of applications of geometric and topological combinatorics to a broad class of problems in phylogenetics. This is a very active branch of applied topology and geometry. 7. Many results about the topology of state complexes are highly dependent on their geometric features [ ]. Theorem
??
follows from some simple observations combined
with the powerful geometric tools of Gromov [ ]. 8. State complexes are undirected versions of the
[ ]. They may also be viewed as weak versions of the very interesting class of combinatorial cell complexes known as
high-dimensional automata
of Pratt
hom complexes.
The monograph of Koslov [ ] has
a wealth of information on this latter class of spaces.
# The
name apparently is meant to connote graph-based as opposed to pictorial.
36
Exercises
2.1. Consider the graph whose vertices are triangulations of a regular graph in the case
n-gon, and whose

Draw this
edges are given by ipping the diagonal edge of two incident triangles.
n = 6 as the 1-skeleton of a polytope.
2.2. Show that every cubical complex can be subdivided to give a simplicial complex, and that every simplicial complex can be subdivided to give a cubical complex. 2.3. Assume a legged robot may lift each leg independently. To remain in a stable standing position, certain combinations of legs must remain on the ground. For example, for the hexapod insectoid RHex [ ], at least three legs with at least one per side must remain on the ground. Dene an abstract simplicial complex for a legged robot based on legs o-the-ground and compute what this complex is in the case of RHex. 2.4. Consider four points in shadow. Show that if in their
R2
and let
VR
be the Vietoris-Rips complex and
image,
VR contains two disjoint closed edges then VR contains at least one 2-simplex.
R
its
whose interiors intersect
2.5. Consider four compact convex sets in
R2
having the property that any three sets Does
intersect nontrivially. What are the possible nerves of this collection of sets?
the Nerve Theorem help you deduce whether or not all four sets intersect nontrivially? 2.6. Show that the ech and Vietoris-Rips complexes t together: for any point set
Rn ,
(Q) VR2 (Q) C 2 (Q): C
What improvement to the constants can be made if
n = 2?
(see, e.g., [ ])
2.7. Is it possible to have a nondeterministic graph whose every action is nondeterministic but which nevertheless has a complete strategy for attaining a goal state? 2.8. What are the input and output complexes associated to
3 processors and 3 labels,
with the desired outcome being neither consensus (all the same label) nor diversity (all three labels represented among choices) but in between (exactly two labels are chosen by the processors). 2.9. Let
denote a Y-graph (a graph with three edges emanating from a central vertex).
Show that
C 2 (Y )
embeds in
R3 and draw a picture of it.

D2 (Y)?
How does it compare to the
discretized conguration space 2.10. Show that 2.11.
D3 (Y) cannot be homotopic to C3 (Y). 2 Show that D (K3;3 ), the discretized conguration space of complete bipartite graph K3;3 , is a topological 2-manifold.
two labeled points on the
2.12. Show that the state complex of Example
?? is always contractible, for any length robot

G
arm. In the limit as length goes to innity, this starts to look like a conguration space of a smooth curve in the plane with nonnegative partial derivatives. 2.13. Consider the following recongurable system: ve edges (a pentagon);
A;
A = f1; 2; 3; 4; 5g,
is a cyclic graph with ve nodes and
and each state is a bijection of nodes to
generators have support and trace equal to an edge of the graph, with local states
that exchange labels on vertices; there is one such generator per edge and per vertex labeling. Show that the state complex of this system is an orientable surface.
Chapter 3
Euler Characteristic
38
Chapter 3. Euler Characteristic
he simplest and most elegant non-obvious topological invariant is the Euler characteristic, an integer-valued invariant of spaces. This chapter is comprised of observations and applications of this invariant, meant to serve as motivation for the algebraic tools to follow that justify its remarkable properties.
3.1 Counting
The Euler characteristic is a generalization of counting. Given a nite, discrete set X , the Euler characteristic is its cardinality (X ) = jX j. Connect two points together by means of an edge (in a cellular/simplicial structure); as the resulting space has one fewer component, the Euler characteristic is decremented by one. Continuing inductively, the Euler characteristic counts vertices with weight +1 and edges with weight 1. This intuition of counting connected components works at rst; however, for certain examples, the addition of an edge does not change the count of connected components. Note that this occurs precisely when a cycle is formed. To ll in such a cycle in the gure with a 2-cell would return to the setting of counting connected components again, suggesting that 2-cells be weighted with +1. This intuition of counting with weights inspires the combinatorial denition of Euler characteristic. Given a space X and a decompo sition of X into a nite number of open cells X = , where each k -cell is homeomorphic to Rk , the Euler characteristic$ of X is dened as (X ) = ( 1)dim : (3.1) This quantity is well-dened for nite cellular objects and independent of the decomposition of X into cells: is a homeomorphism invariant. It is not a homotopy invariant for non-compact cell complexes, as, e.g., it distinguishes ((0; 1)) = 1 from ([0; 1]) = 1. Among compact nite cell complexes, it is a homotopy invariant, as will be shown in Chapter 5. Euler characteristic can determine the homotopy type of a compact connected graph; e.g., such is a tree (a contractible graph) if and only if = 1. Euler characteristic is also sharp invariant among connected compact orientable 2-manifolds: = 2 2g , where g equals the genus. Any compact convex subset of Rn has = 1. Removing k disjoint convex open sets from such a convex superset results in a compact space with Euler characteristic 1 k ( 1)n .
Example 3.1 (Conguration spaces of graphs) Recall from Example 2.10 that the
discrete conguration space D2 (K5 ) of two points on a complete connected graph of ve vertices is a cubical 2-manifold. Being built from nitely many cells, it is clearly $ This is sometimes called the combinatorial or geometric or even motivic Euler characteristic.
3.2. Curvature
39
compact. It is also orientable and connected, as the reader may check. To determine the genus of this conguration space, one computes the Euler characteristic. Vertices of D2 (K5 ) correspond to ordered pairs of distinct vertices of K5 , of which there are (5)(4) = 20. Edges correspond to pairs of one closed edge and one disjoint vertex: there are (2)(10)(5 2) = 60 such. Faces correspond to ordered pairs of closure-disjoint edges in K5 , of which there are (10)(3) = 30. Thus, (D2 (K5 )) = 20 60 + 30 = 10 and this surface has genus g = 1 1 = 6. 2 The remainder of this chapter complements the interpretation of with more geometric, dynamical, and analytic perspectives.
as counting
3.2 Curvature
A blend of Euler characteristics and integration is prevalent in geometry. The classical result that initiated the subject is the Gauss-Bonnet Theorem. Let M be a smooth surface embedded in R3 . The Gauss map is the map : M ! S2 that associates to each point of M the direction of the unit vector normal to M in R3 . The Gauss curvature = detD is the determinant of the derivative of the Gauss map. Note that the curvature is a geometric quantity: rigid translations and rotations leave it invariant, but stretching and deformation of M change . This change is local, not global.
Theorem 3.2 (Gauss-Bonnet).

oriented surface in respect to area on
For
M a compact smooth
(3.2)
M equals
M
, the integral of Gauss curvature with
dA = 2(M ):
That this is the beginning of a much larger story is evidenced by the following mild generalization. Let M be an oriented surface in R3 , piecewise smooth over a cell structure, perhaps with boundary a piecewise-smooth curve(s). Again, 2(M ) equals the integral of a curvature function over M , but the curvature measure, like M , is stratied. 1. On 2-cells of M , d means dA, Gauss curvature times the area element; 2. On 1-cells of M , d means kg ds , geodesic curvature times the length element; 3. On 0-cells of M , d means the angle defect. With this interpretation, the Gauss-Bonnet formula can be d = d + d + d = 2(M ); (3.3) M M (0) M (1) M (2) i.e., the integral over the 2-cells of Gauss curvature plus the integral over 1-cells of geodesic curvature, plus the integral over 0-cells of angle defect equals 2(M ). There are several corollaries of this result relevant to discrete and dierential geometry: written as
40
Chapter 3. Euler Characteristic 1. A smooth, closed surface has total (integrated) Gauss curvature constant, no matter how the surface is deformed. 2. For a geodesic triangle, d vanishes along the geodesic edges and the sum of the angles of the triangle equals plus the integral of Gauss curvature over the triangle face. This recovers classical notions of angle-sums for triangles on spheres and other curved surfaces. 3. For M the boundary of a convex polyhedron in R3 , all faces and edges are at, (M ) = (S2 ) = 2, and the sum of the vertex angle defects (in this case, 2 minus the sum of the face angles) equals 4 .
The reader may rightly suspect whether an embedding in R3 is at all needful. Recall, again, from Example 2.10, the cubical 2-manifolds built from arranging six squares around each of 20 vertices. Placing a at Euclidean metric on each square yields a space with an intrinsic geometry, locally at but for the vertices. Since each vertex has angle defect 2 6 = , Gauss-Bonnet implies that the 2 surface has Euler characteristic = 20( )=2 = 10.
3.3 Obstructions to vector elds

Euler characteristic has many uses, one of which is as an obstruction to the existence of certain vector elds. Recall that a vector eld is a continuous (or smooth) assignment of a tangent vector to each point in a manifold.
Theorem 3.3.
if
A connected compact manifold
without
boundary possesses a nonvanishing vector eld if and only
(M ) = 0.
This is not terribly useful in dimensions 1 and 3, since all compact manifolds in these dimensions have Euler characteristic zero. However, since (S2 ) = 2, this does imply that there are no xed-point-free vector elds on a 2sphere, a fact manifestly clear to meteorologists.%
Proof. (sketch) The interesting direction (only if) reveals the obstructive nature of . The easiest proof without invoking advanced tools involves imposing a dense cellular mesh and working cell-by-cell, assuming smoothness. Assume that M is a compact closed n-manifold that admits a smooth nonvanishing vector eld . Place a smooth cell structure on M with all cells dieomorphic to convex polytopes of suciently small diameter. Invoking transversality and rening as needed, perturb the cell structure so that the vector eld is transverse to all cells. As the cells are suciently small and the vector eld is nonvanishing, is approximately linear in a neighborhood of each cell. On each n-dimensional cell , consider the set of faces % This also is the source of the moniker Hairy Ball Theorem, to wit, you can't comb the hair on
a ball at.
3.4. Index theory
41 By transversality, I ( ) is nonempty for each top-dimensional cell and contains both the interior cell of (homeomorphic to an open disc Dn ) along with some boundary faces, the union of which (by linearity of ) is homeomorphic to the open disc Dn 1 . Thus, (I ()) = ( 1)n +( 1)n 1 = 0 and (M ) = (I ( )) = 0, since the sets I ( ) partition all cells of M .
I () on which points to the interior of .
This invocation of transversality is admittedly glib. With better tools (from Chapters 4-5), it will be possible to drop the assumptions about smoothness and manifold structure: see Theorem 5.12.
3.4 Index theory
Consider, for simplicity, a vector eld V on an orientable 2-manifold . Let p 2 Fix(V ) be an isolated xed point of V . Let Bp denote a suciently small ball about p with boundary = @Bp . The index of V at p , IV (p ) is dened to be the following line integral:
IV (p ) =
where (V ) is the angle made by V in local coordinates. More specically, if in (x; y ) coordinates based at p , the vector eld is of the form V = (x; _ y _ )T , then the integrand d is so that IV (p ) represents the (signed integer) number of turns the vector makes in a small curve about p . This index is well-dened and independent of (a suciently small) Bp , thanks to Green's Theorem. Among the nondegenerate xed points, sources and sinks have index +1; saddle points have index 1. It is clear that Equation (3.4) extends to dene an index IV ( ) to any closed curve which avoids Fix(V ). One argues (in a manner not unlike those used in Green's Theorem or in contour integration) that index is additive over Fix(V ). For R2 Fix(V ) a simple closed curve, let D be a disc whose boundary is @D = . Then, IV ( ) = IV (p ): (3.5) p2D\Fix(V ) of competing species x (t ), y (t ):
1 d(V ); 2
(3.4)
x _ dy y _ dx d = x ; 2 _ +y _2
Example 3.4 (Population dynamics) Consider the following dierential equation model
x _ = 3x x 2 2xy ; y _ = 2y xy y 2 :
(3.6)
42
The equilibrium solutions consist of (0; 0) [mutual death], (3; 0) [species x dominates], (0; 2) [species y dominates], and (1; 1) [coexistence]. Linearization reveals that the coexistence solution is a saddle point; thus, I(1; 1) = 1. It is clear (from the equations and from the interpretation) that the x and y axes are invariant sets. Is there a periodic orbit? No. The index of a periodic orbit must by Equation (3.4) equal IV ( ) = +1, since is everywhere tangent to the vector eld. However, by Equation (3.5), must surround a collection of xed points whose indices sum to +1. If (t ) is a nontrivial periodic solution, then it cannot intersect the xed point set or the x or y axes, and the only remaining enclosable xed point has negative index. The additivity present in Equation (3.5) is reminiscent of the Euler characteristic. The rationale for this equation (and the proper denition of the index of a vector eld in all dimensions) will become clearer in Chapters 4-5. The following classical theorem is a hint of these deeper connections:
Theorem 3.5 (Poincar-Hopf).

eld
For a continuous vector
with isolated xed points on a compact manifold
Fix(V )
M,
IV (p ) = (M ):
(3.7)
The proofs of Theorems 3.3 and 3.5 compute locally (on cells and xed points respectively), then add up this local data to return the global . This integrative technique motivates an extension of Euler characteristic from sets to certain functions over sets: an integration theory. To complete this integration theory requires an excursion to consider exactly which sets have a well-dened Euler characteristic.
3.5 Tame topology

Euler characteristic is well-dened for spaces having a decomposition into a nite number of cells. Though an explicit cell structure is often present in, say, the simplicial setting, some (sub)spaces do not come with an externally imposed cell decomposition. Consider, for example, a conguration space of points in a domain, or the level set of a smooth function f : Rn ! R. Besides the lack of an explicit cell structure, worse things can occur when presented with a space ab initio, as with the graph of sin(1=x ) for 0 < x < 1. Though this is homeomorphic to an interval, it is a wild type of equivalence. The closure of this set in R2 is not homeomorphic to the closed
3.5. Tame topology
43
interval [0; 1]. The intersection of this set with the x -axis is an innite set of points. In applications, one wants to rule out such oddities and focus on spaces (and mappings) that are for all intents and purposes tame. Dierent mathematical communities focus on, e.g.: piecewise linear (PL) spaces, describable in terms of ane sets and matrix inequalities; semialgebraic sets, expressible in terms of a nite set of polynomial inequalities; or subanalytic sets, dened in terms of images of analytic mappings [?]. Logicians have created an axiomatic reduction of such classes of sets in the form of an o-minimal structure.& An o-minimal structure O = fOn g is a sequence of Boolean algebras An of subsets of Rn (families of sets closed under the operations of intersection and complement) which satises certain axioms: 1. O is closed under cross products; 2. O is closed under axis-aligned projections Rn ! Rn 1 ; 3. O1 consists of all nite unions of points and open intervals. Elements of O are called tame or, more properly, denable sets. Canonical examples of o-minimal systems include semialgebraic sets and subanalytic sets. Some deeper results of o-minimal geometry require having at least all algebraic sets [?], but for our purposes, piecewise-linear sets form an acceptable o-minimal structure. Note the last axiom: this niteness is the crucial piece that drives the theory. Given a xed o-minimal structure, one can work with tame sets with relative ease. Tame mappings are likewise easily dened: a (not necessarily continuous) function between tame spaces is tame (or denable) if its graph (in the product of domain and range) is a tame set. A denable homeomorphism is a tame bijection between tame sets. To repeat: denable homeomorphisms are not necessarily continuous. Such a convention makes the following theorem concise:
Theorem 3.6 (Triangulation Theorem).

disjoint union of open standard simplices.
Any
tame set is denably homeomorphic to a nite The
intersection of the closures of any two simplices in this denable triangulation is either empty, or the closure of another open simplex in the triangulation.
This result implies that tame sets always have a well-dened Euler characteristic, as well as a well-dened dimension (the max of the dimensions of the simplices in a triangulation). The surprise is that these two quantities are not only topological invariants with respect to denable homeomorphism; they are complete invariants.
& The term derives from order minimal, in turn coming from model theory. The text of Van den Dries [?] is a beautifully clear reference.
44

Two tame sets in an o-minimal structure are denably homeo-
Theorem 3.7 ([?]).
morphic if and only if they have the same dimension and Euler characteristic.
This result reinforces the idea of a denable homeomorphism as a scissors equivalence. One is permitted to cut and rearrange a space with abandon. Recalling the utility of such scissors work in computing areas of planar sets, the reader will not be surprised to learn of a deep relationship between tame sets, the Euler characteristic, and integration.
3.6 Euler calculus

It is possible to build a topological calculus based on Euler characteristic. The integral in this calculus depends on the following additivity principle:
Lemma 3.8.
For
A and B denable sets,

(3.8)
(A [ B) = (A) + (B) (A \ B):
for A X denable. The measurable functions in this integration theory are integervalued and constructible, meaning that for h : X ! Z, all level sets h 1 (n) X are tame. Denote by CF(X ) the set of bounded compactly supported constructible functions on X . The Euler integral is dened to be the homomorphism X : CF(X ) ! Z given by: 1 h d = s(fh = s g): (3.10) X s= 1 Alternately, using the denition of tame sets, one may write h 2 CF(X ) as h = c , where c 2 Z and f g is a decomposition of X into a disjoint union of open cells, then (3.11) h d = c ( ) = c ( 1)dim X
' The proper term is a valuation, not a measure; the abuse of terminology is to prompt the reader to think explicitly in terms of integration theory.
This additivity has been foreshadowed in the imagery of as counting in 3.1, in the Gauss-Bonnet Theorem of 3.2, in the proof of Theorem 3.3, and in the additivity of the xed point index of 3.4. The reader may prove this via triangulation and induction, until the homological tools Chapters 4-5 are available. The similarity between Equation (3.8) and the denition of a measure is not a coincidence. Following ideas that date back to Blaschke (and likely earlier), one denes a measure' d over denable sets via: A d = (A); (3.9) X
3.7. Target enumeration
45
That this sum is invariant under the decomposition into denable cells is a consequence of Lemma 3.8 and Theorem 1.
3.7 Target enumeration

A simple application of Euler integration to data aggregation demonstrates the utility of this calculus. Consider a nite collection of targets, represented as discrete points in a space W . Assume a eld of sensors, each of which observes some subset of W and counts the number of targets therein. The sensors will be assumed to be distributed over a region so densely as to be approximated by a topological space X . In a particular system of sensors in X and targets in W , let the sensing relation be the relation S W X where (w; x ) 2 S i a sensor at x 2 X detects a target at w 2 W . The horizontal and vertical bers (inverse images of the projections of S to X and W respectively) have simple interpretations. The vertical bers target supports are those sets of sensors which detect a given target in W . The horizontal bers sensor supports are those targets observable to a given sensor in X . Assume that the sensors are additive but anonymizing: each sensor at x 2 X counts the number of targets in W detectable and returns a local count h(x ), but the identities of the sensed targets are unknown. This counting function h : X ! Z is, under the usual tameness assumptions, constructible. A natural problem in this context is to aggregate the redundant anonymous target counts: given h and some minimal information about the sensing relation S, determine the total number of targets in W . This is not as easy as it sounds, and it is even less easy when X is not a continuum space, but rather a discretization thereof. It is therefore remarkable that a purely topological solution exists, independent of knowing a decomposition of the counting function h.
Theorem 3.9 ([?]).

function for target supports 1 , then .
# = N X h d
(
If h : X ! N where h is a counting U of uniform Euler characteristic (U ) = N 6= 0 for all
Proof.
h d =
U
d =
U d =
(U ) = N #:
(3.12)
For contractible supports (such as in the setting of beacons visible on star-convex
46
domains), the target count is, simply, the Euler integral of the function. This solves a problem in the aggregation of redundant data, since many nearby sensors with the same reading are detecting the same targets; in the absence of target identication (an expensive signal processing task, in practice), it is nontrivial to aggregate the redundancy. Notice that the restriction N = 0 is nontrivial. If h 2 CF(R2 ) is a nite sum of characteristic functions over annuli, it is not merely inconvenient that 42 h d = 0, it is a fundamental obstruction to disambiguating sets. Given this, it is all the more remarkable that sets with 6= 0 can be enumerated easily. Since the solution is in terms of an integral, local and distributed computations may be used in practice.
3.8 The Fubini Theorem

Euler characteristic is like a volume in another aspect: it, like volume, is multiplicative under cross products, cf. area is base-times-height.
Lemma 3.10.
For
X and Y
denable,
(X Y ) = (X )(Y ).
Proof. The product X Y has a denable cell structure using products of cells from
dim( ) = dim
and
Y.
For cells
X and Y , the lemma holds via the exponent rule since + dim . The rest follows from additivity.
The assertion that d is an honest measure is supported by this fact and its corollary: the Euler integration theory theorem. Calculus students know admits a Fubini that f (x; y )dx dy = f (x; y )dy dx . Real-analysis students learn to pay attention to ner assumptions on measurability (cf. tameness assumptions). This familiar result is the image of a deeper truth about integrations and projections: Given F : X ! Y , one can integrate over the bers, or level sets, of F rst, then integrate the resulting function over the projected base Y .
Theorem 3.11.
all
h 2 CF (X ),
Let
F :X!Y
be a tame mapping between tame spaces. Then for
h d =
1(
y)
h(x ) d(x ) d(y ):
(3.13)
Proof. If X is homeomorphic to U Y and F is projection to the second factor, the result follows from Lemma 3.10. The o-minimal Hardt Theorem [?] says that Y has a denable partition into tame sets Y such that F 1 (Y ) is denably homeomorphic
3.9. Euler integral transforms to U Y for U denable, and that of the integral completes the proof.
47
F : U Y ! Y acts via projection.
Additivity
plication of the Fubini Theorem to time-dependent targets. Consider a collection of vehicles, each of which moves along a smooth curve i : [0; T ] ! R2 in a plane lled with sensors that count the passage of a vehicle and increment an internal counter. Specically, assume that each vehicle possesses a footprint a support Ui (t ) R2 which is a compact contractible neighborhood of i (t ) for each i , varying tamely in t . At the moment when a sensor x 2 R2 detects when a vehicle comes within proximity range when x crosses into Ui (t ) that sensor increments its internal counter. Over the time interval [0; T ], the sensor eld records a counting function h 2 CF(R2 ), where h(x ) is the number of times x has entered a support. As before, the sensors do not identify vehicles; nor, in this case, do they record times, directions of approach, or any ancillary data.
Example 3.12 (Enumerating vehicles) The following example demonstrates an ap-
Proposition 3.13 ([?]).

to
42
h d.
The number of vehicles is equal
given by the union of slices (Ui (t ); t ) for t 2 [0; T ]. Each such tube has = 1. The integral over R2 [0; T ] of the sum of the characteristic functions over all N tubes is, by Theorem 3.9, N , the number of targets. Consider the projection map p : R2 [0; T ] ! R2 . Since p 1 (x ) is fx g [0; T ], the integral over p 1 (x ) records the number of (necessarily compact) connected intervals in the intersection of p 1 (x ) with the tubes in R2 [0; T ]. This number is precisely the sensor count h(x ) (the number of times a sensor detects a vehicle coming into range). By the Fubini Theorem, 42 h d must equal the integral over the full R2 [0; T ], which equals N .
Proof. Each target traces out a compact tube in R2 [0; T ]
3.9 Euler integral transforms

Euler integration admits a variety of operations which mimic analytic structures of particular relevance to signal processing, imaging, and inverse problems.
Example 3.14 (Convolution and Minkowski sum) On a real vector space V , a convolution operation with respect to Euler characteristic is straightforward. Given
f ; g 2 CF(V ), one denes
(f
g )(x ) =
f (t )g (x t ) d(t ):
(3.14)
48
There is a close relationship between convolution and the Minkowski sum: for A and convex, A B = A+B , where A + B is the set of all vectors expressible as a sum of a vector in A and a vector in B [?, ?]. This, in turn, is useful in applications ranging from computer graphics [?] to motion-planning for robots around obstacles [?]. For non-convex shapes, convolution of indicator functions can take on nonzero values besides 1.
One of the most general integral transforms is the Radon transform of Schapira [?, ?, ?]. Consider a locally closed denable relation S W X (that may or may not come from sensing), and let W and X denote the projection maps of W X to their factors. The Radon transform with kernel S is the map RS : CF(W ) ! CF(X ) given by (RS h)(x ) = h W d(w ): (3.15) X 1 (x )
Example 3.15 (Target enumeration) Consider the sensor relation S W X , and a nite set of targets T W as dening an atomic function T 2 CF(W ). Observe that the counting function which the sensor eld on X returns is precisely the Radon transform RS T . In this language, Theorem 3.9 is equivalent to the following: assume that S W X has vertical bers W1 (w ) \ S with constant EulercharacteristicN . Then, RS : CF(W ) ! CF(X ) scales integration by a factor of N : X RS = N W .
A similar regularity in the Euler characteristics of bers allows a general inversion formula for the Radon transform [?]. One must choose an `inverse' relation S0 X W.
X W have bers Sw and S0w satisfying (1) (Sw \ S0w 0 ) = 6= for all w 0 6= w 2 W . Then for all
(RS0 RS )h = (
Theorem 3.16 (Schapira inversion formula).
W X (Sw \ S0w ) = w 2W h 2 CF(W )

Assume that
and
S0
for all ,
; and (2)
)h +
h W :
(3.16)
Recall that the sensor counting eld h : X ! Z is equal to RS T , where T W is the set of targets. If the conditions of Theorem 3.16 are met, then the inverse Radon transform RS0 h = RS0 RS T is equal to a multiple of T plus a multiple of W . Thus, one can localize the targets by performing the inverse transform. It is remarkable that counting data alone can yield not only target counts but target positions as well. By changing T from a discrete set to a collection of (say, contractible) compact sets, one notes that Radon inversion has the potential to not merely localize but recover shape. This topological tomography is the motivation for Schapira's incisive paper [?].
Example 3.17 (Medical imaging) Schapira's paper outlines an application in imag-
ing most natural in the setting of MRI. Assume that W = R3 and one scans a compact subset T W by slicing R3 along all at hyperplanes, recording simply the Euler
3.9. Euler integral transforms
49
characteristics of the slices of T . Since a compact subset of a plane has Euler characteristic the number of connected components minus the number of holes (which, in turn, equals the number of bounded connected components of the complement), it is feasible to compute an accurate Euler characteristic, even in the context of noisy readings. This yields a constructible function on the sensor space 3 X = AG 3 2 (the ane Grassmannian of 2-planes in R ) equal to the Radon transform of T . Using the same sensor relation to dene the inverse transform is eective. Since Sw = P2 and Sw \ Sw 0 = P1 , one has = (P2 ) = 1, 1 = (P ) = 0, and the inverse Radon transform, by (3.16), yields T exactly: one can recover the shape of T based solely on connectivity data of black and white regions of slices.
Example 3.18 (Sensor beams) Let the targets T be a nite disjoint collection of points in W = Dn , the open unit disc in Rn . Assume that the boundary @W is lined with sensors, each of which sweeps a ray over W and counts, as a function of bearing, the number of targets intersected by the beam. The bearing of a ray at a point p 2 @W lies in the open hemisphere of the unit tangent bundle to W at p . This open hemisphere projects to the open unit disc in Tp @W . Thus, the sensor space X is homeomorphic to T (@W ) = T S n 1 , the tangent bundle of @W . The sensing relation will be seen to be self-dual: S = S0 . Any point in W is seen by any sensor in @W along a unique bearing angle. Thus, the sensor relation S has vertical bers (target supports) which are sections of T Sn 1 and hence spheres of Euler measure = (Sn ) = 1 + ( 1)n . Any two distinct vertical bers project to X and intersect along the subset of rays from @W that pass through both points in W : this is a discrete set of cardinality (hence ) = 2. For n even, one has = 2, = 0, and Equation (3.16) implies that the inverse Radon transform gives 2T +2(#T )W . Thus, on even-dimensional spaces, targets may be localized with boundary beam sensors. One may likewise assume that the target set T is not a collection of points but rather a disjoint collection of compact sets in W ; target shapes will then be recovered, assuming the ability of beams to measure Euler characteristic of slices exactly. Example 3.19 (Fourier transforms and curvature) One relationship between Euler
and Lebesgue measures is encoded in the Gauss-Bonnet theorem. This, too, can be lifted from manifolds to denable sets and then to CF(Rn ). The mechanism is the
50
2 Sn
microlocal Fourier transform. Given h 2 CF(Rn ), a point x 2 Rn , and a unit vector

1
, the microlocal Fourier transform is dened via: h d(y ); (M F h)x () = lim+ !0 B (x ) (y x )0
(3.17)
where B denotes the open ball of radius . Like the classical Fourier transform, M F takes a frequency vector and integrates over isospectral sets dened by dot product. For Y a compact tame set, M FY () is the constructible function on Y that, when averaged over , yields a curvature measure dY on Y implicated in the Gauss-Bonnet Theorem (as shown by Brcker and Kuppe [?]): for any U Rn open,
dY = Vol Sn
5n
M FY
d d:
The Euler-calculus interpretation of the Gauss-Bonnet Theorem says, 1 M F d d = Y = (Y ); Y Vol Sn 1 5n 1 Y Y (3.18) and this extends via linearity of M F to all of CF(Y ).
3.10 Intrinsic volumes

The humble combinatorial denition of has evolved in this chapter to the more analytic construct of a measure. This has a grand history in the subject of integral geometry. There is a classical family of measures on Euclidean Rn that mediate Euler and Lebesgue measure and entwine topological with geometric data. The k th intrinsic volume k is characterized uniquely by the following: for all A and B tame in Rn , 1. 2. 3. 4.
Additivity: k (A [ B ) = k (A) + k (B ) k (A \ B ); Euclidean invariance: k is invariant under rigid motions of Rn ; Homogeneity: k ( A) = k (A) for all 0; and Normalization: k of a closed unit ball in Rn equals 1.
These measures generalize Euclidean n-dimensional volume (n = dvoln ) and Euler characteristic (0 = d). There are several equivalent denitions, all revolving about the notion of an average Euler characteristic. One way to dene the intrinsic volume k (A) is in terms of the Euler characteristic of all slices of A along ane codimension-k planes: k (A) = n k (A \ P ) d(P ) = n k A d d(P ); (3.19) Pln Pln P
Intrinsic volumes are also known as Hadwiger measures, quermassintegrale, Lipschitz-Killing curvatures, Minkowski functionals, and, likely, a few more names unknown to the author.
3.10. Intrinsic volumes
51
where is an appropriate measure on Pln n k , the space of ane (n k )-planes in Rn . The details of this construction are left to the literature: the point here stressed is that all the intrinsic volumes are certain Lebesgue-averaged Euler integrals. These intrinsic volumes are more than isolated examples. Consider the vector space of Euclidean-invariant valuations additive functionals on tame sets. A valuation is compact-continuous if it is continuous with respect to the Hausdor metric on compact convex sets. A classical theorem of Hadwiger [?] characterizes such valuations:
Theorem 3.20 (Hadwiger Theorem).

Euclidean-invariant valuations on
Rn
has
fk gn
The vector space of compact-continuous

0 as a basis.
These amalgamations of Euler and Lebesgue measure are the key to several interesting applications.
Example 3.21 (Microstructure coarsening) Anyone who has observed foam in a pilsner glass knows that the cell walls of foam evolve over time so that some cells grow, while others shrink unto disappearance (neglecting pops). The same processes are ubiquitous in the microstructure coarsening of froth, metals, and some ceramics [?]. It has long been known how idealized cells evolve and coarsen in the 2-dimensional setting. The von Neumann - Mullins formula states that the area A(t ) of a cell in an ideal dynamic 2-d cell structure changes as: ( ) dA j C0 j = 2M 1 ; (3.20)
dt
where M is a mobility constant, is a surface tension constant, and jC0 j equals the number of corners (vertices) of the cell. Thus, cells with fewer than six corners shrink: those with more grow. The proper extension of this formula to three-dimensional cell
structures was recently discovered by MacPherson and Srolovitz [?]. The key to the extension was to note that the 1 in Equation (3.20) is in fact the Euler characteristic
This measure is derived from the Haar measure on the Grassmannian /n n k and Lebesgue measure on the orthogonal 4k . See [?] for a complete discussion.
52
of the (contractible, compact) cell. To lift from 2-d to 3-d requires lifting the relevant interior volumes from 0 (= ) to 1 . The MacPherson-Srolovitz formula for the volume V (t ) of a cell a in 3-d structure is: ( dV jC1 j ) ; = 2M 1 (3.21)
dt
where 1 is the intrinsic 1-volume of the cell and jC1 j is the total length of all the edges (1-dimensional boundary curves) of the cell. When contemplated in the light of intrinsic volumes, it is clear that Equation 3.21 is in fact an elegant relationship between dierent (intrinsic) volumes, the left hand side implicating 3 and the right implicating 1 .
3.11 Gaussian random elds

Most readers will be comfortable with the utility (if not the details) of stochastic processes. A real-valued stochastic process is, in brief, a collection fh(x ) : x 2 X g of random variables h(x ) parameterized by a space X . A random eld is a stochastic process whose arguments vary continuously over the parameter space X . For example, a random eld may be used to model the height of water on a noisy sea or the magnitude of noise in a sea of cell-phone towers, both examples having as parameter space a geometric domain of dimension two (or three, with the addition of time). As the simplest random variables are Gaussian (having well-dened nite mean and variance with the usual normal distribution), so are the simplest random elds. A Gaussian random eld with parameter space X Rn is a random eld h over X k such that, for each k 2 N and each k -tuple fxi gk 0 X , the collection fh (xi )g0 of random variables has a multivariate Gaussian distribution. For simplicity, attention will be restricted to Gaussian distributions with the following features: 1. 2. 3.
jx y j.
Centered: all h(x ) have zero mean, Efh(x )g = 0; Common variance: all h(x ) have xed variance 2 > 0; Stationary-isotropic: the covariance Efh(x )h(y )g is a function of distance
respect to x . Adler [?], Taylor [?], Worsley [?], and others have led in the exploration of Gaussian random elds from a geometric perspective. There is implicit in this work no small amount of topology, with Euler characteristics featuring prominently. Let h(x ), x 2 T Rn , be a stationary random Gaussian eld. Knowing as much as possible about the tomography of the eld h is important in medical imaging, astronomy, and a host of other applications [?]. For example, one might wish to know the expected number of peaks in a eld, or some other qualitative properties associated to the excursion sets fh s g. One of the principal results in this area is that, although the
These last two assumptions imply a well-dened constant , the second spectral moment, which, roughly, measures the variance of directional derivatives of h with
3.11. Gaussian random elds
53
ne details of the expected eld excursion sets cannot be computed, the expected Euler characteristics are not merely computable, but computed and commensurate with experimental data [?]. The following classical result of Rice from the 1940s [?] helped initiate the subject: for h(x ), x 2 [0; T ], a stationary, zero mean, almostsurely continuous Gaussian process with nite variance 2 and second spectral moment , the expected number of up-crossings at h = s locations where h increases past s is given by
E#fh % s g = e
The observation that the number of upcrossings #fh % s g is in fact the Euler characteristic fh s g of the upper excursion set presages the deeper work of [?] for elds over higher-dimensional domains. It appears very difcult to understand the expected shape or even topological type of these excursion sets; however, the simplication that introduces and its commensurability with integral techniques yields to computational eort. The following result is a highly illustrative example:
s2 2 2
T: 2
(3.22)
h(x ), x 2 Rn is a centered, stationary, isotropic 2 Gaussian random eld with variance , second spectral moment , and sucient regularity. The expected Euler characteristic of the upper excursion set fh s g n over a compact denable subset X R is
Theorem 3.22 ([?, ?]).
Assume that
E(fh s g) = e
where
s2 2 2
dim X
k =0
(2 )
k +1
2
( p )k

Hek
(s )
k (X );
(3.23)
Hek
is the degree
intrinsic volume of
T.
Hermite polynomial in one variable and
k (X ) is the k th
Additional work [?] allows one to relax the assumptions of the eld being stationary, isotropic, and Gaussian. In addition, expectations for intrinsic volumes k of excursion sets are likewise derivable.
This consists of nondegeneracy assumptions on the joint distributions of the rst and second derivatives of f and regularity assumptions on the covariance functions of the second derivatives. See [?, ?] for (many) details.
54
Notes
1. The scissors equivalence implicit in Theorem is the hint of deeper structures. characteristic and integration over
CF
is an elementary version of
motivic integration,
Euler
of great current interest in algebraic geometry. 2. Since the Euler measure is the universal motivic measure over topological spaces, any other scissors-invariant group-valued measure on denable topological spaces must factor through
.
One way around this depressing result is to extend
valued measure of sequences of spaces.
[ ] on Euler characteristics associated to conguration spaces, where one encodes the Euler characteristics of
The monograph [ ] details a theory of Gal
to a polynomialt.
X tame, the resulting series is the Taylor series of a rational function.

functions from algebraic geometry [ , integration with respect to
UCn (X )
as coecients of a series in a formal variable
For
Similar algebraic
manipulations for Euler characteristics of sequences of spaces appears in work on
d[t ] over ltered spaces awaits development.

?
? ?].
-
It is to be suspected that a full theory of
3. Euler integration is here, as in [ ], presented as a combinatorial theory. Higher perspectives are more illuminating. Chapter 8 and [ ] give an interpretation in terms of sheaf theory. The literature on
? normal cycles [?, ?] allows, thanks to results of Kashiwara [?], an approach in terms of conormal cycles and related geometric measure theory.
to
4. The issue of numerical Euler integration how to approximate an integral with respect
d
based on sampled data is extremely interesting, technical, and relevant It appears to have received little treatment from the Mathematics
to applications. community.
5. It is possible and protable to perform Radon transforms (and inversion thereof ) with weighted kernels: see [ ]. A few other integral transforms will appear in 7.3. 6. Intrinsic volumes are well-dened on all of
sets thanks to the close relationship with
CF (Rn )
d.
not merely compact denable Continuity of these measures is
possible, but requires a more delicate function space topology, relying on currents (see Chapter 6). 7. Research in geometric random elds has the reputation for being very dicult to understand. This is perhaps due more to the typical mutual ignorance of researchers in probability and in topology. It may help matters to employ (1) the language of
Euler integration, as it removes much of the mystery as to the natural appearance of
in this work, and (2) the o-minimal framework, as it would likely subsume many of
denable sets in the statement of Theorem
the nondegeneracy assumptions. As a rst step, I have taken the liberty of invoking
??:
this is not what Adler proves, nor
have I re-tooled Adler's proof (and all of probability theory) to the o-minimal setting. Caveate.
Exercises
3.1. Use Gauss-Bonnet to show that any sphere in
R3
built from regular hexagons and This result has
pentagons must have exactly 12 pentagons: no more and no less.
implications constraining the possible arrangements of carbon into buckyballs. 3.2. A
periodic orbit
of a vector eld is a owline that is homeomorphic to
S1 .
Use the
Poincar-Hopf theorem to show that any periodic orbit in the plane encloses at least one xed point, and the sums of indices of the enclosed (assumed isolated) xed
Exercises
points equals one. 3.3. Prove directly that the image of a tame set again tame.
55
under a denable map
f : X ! Rn
is
3.4. Computing Euler integrals via Equation (3.10) can be tricky, since the level sets are usually not compact. The following formula, a manifestation of the Fundamental
Theorem of Integral Calculus, is often more stable:
X
Prove this formula.
h d =
1
s =0
f h > s g f h < s g :
(3.24)
3.5. Why does the Fubini theorem for Euler integration not have the customary caveats about checking that the level sets of 3.6. Consider a collection
n + 1 video cameras surveying a region. There are some nite collection of distinct targets in [i Ui , but the cameras cannot identify them, but can determine which targets are seen in common by multiple cameras. Given the function h : N(U) ! N which assigns to each (open) simplex of the nerve N(U) the number of targets in the corresponding intersection of elements of U, show that h d is the total number N
F are measurable sets? k = fUi gn of convex sets in R , thought of as regions seen by 0
of targets visible. 3.7. Show that convolution in Euler integration is multiplicative:
f g d =
f d
g d.
3.8. Compute the (Euler) convolution of the characteristic functions of two round circles of radii
r1 and r2 in the plane.
How do these change as a function of radii?
3.9. In the Radon inversion of Example 3.18, explain from physical reasoning why in the
3.10.
3.11.
n = 1 there is no hope of inverting the Radon transform. Compute the microlocal Fourier transform of P for P a polygon in the plane. that the resulting -average yields the angle defect concentrated at vertices. n n The following integral transform D : CF(R ) ! CF(R ) is from [?]:
case
Show
Dh ( x ) =
where
lim !0
B (x )
h d;
3.12. 3.13.
B (x ) is a small open ball about x . Compute DA for A a compact convex Rn . Show that DD = Id: this is why D is called the dual transform. 2 Compute 1 of a submanifold of R with boundary. What familiar quantity do you
subset of obtain?
3.14. Show that for
A open and denable in Rn , k (A) = ( 1)n k k (A)
Chapter 4
Homology
58
Chapter 4. Homology
ecomposing a space into cells, counting, and canceling according to parity of dimension, is ecacious for dening one topological invariant. More rened invariants arise upon enriching this data with linear-algebraic sensibilities. This leads to the notion of homology.
4.1 Chains
The crucial construction of this chapter is that which converts a decomposition of a space in terms of simple pieces into a collection of vector spaces (or modules) and linear transformations (or homomorphisms): an algebraic version of a simplicial complex. The initial discussion proceeds using the language of linear algebra and passes to more general algebraic constructs. The reader who is weak in either linear or abstract algebra is encouraged to consult the Appendix (at least) for a review before proceeding.
For simplicity, mod-2 arithmetic is used initially. Recall that F2 = f0; 1g is the eld with two elements. Counting in this eld is analogous to ipping a light switch. To build intuition, consider a nite cell complex X , each cell of which is outtted with a metaphorical light switch. The building blocks of a rudimentary homology for X are as follows. 1. Dene k -chains Ck as the vector space over the eld F2 = f0; 1g with basis the k -cells of X . 2. Consider the boundary maps the linear transformations @k : Ck ! Ck 1 which send a basis k -cell to its boundary (as an abstract sum of basis (k 1)cells, each with coecient 1). The collection of chains and boundary maps is assembled into a
chain complex:
@0
/0:
/ Ck
@ @k / Ck 1 k
@2 / C1
@1 / C0
(4.1)
The chain complex is graded, in this case by the dimension of the simplices. It is benecial to denote the chain complex as a single object C = (C ; @ ) and to write @ for the boundary operator acting on any chain of unspecied grading. It seems at rst foolish to algebraicize the problem in this manner why bother with vector spaces which simply record whether a cell is present (1) or absent (0)? Why express the boundary of a cell in terms of linear transformations, when the geometric meaning of a boundary is clear? By the end of this chapter, probably, and the next, certainly, this objection will have been forgotten.
n+1 . The top-dimensional boundary map @ is, in the obvious basis, a dimension k n +1
a single n-simplex n . The resulting chain complex has Ck of Example 4.1 ( Consider )
4.2. Homology examples
59
1-by-n matrix with all entries 1. In contrast, for I n = [0; 1]n , the n-dimensional cube with cell structure inherited from the (n) interval I = [0; 1] with two endpoints, the chain complex has Ck of dimension 2n k k . Note that in all cases, the boundary map @0 is
the zero map. The following lemma is deep.
Lemma 4.2.
@k = 0 for all k .
The boundary of a boundary is null:
@k 1
from the simplex's list, if present; else, do nothing. The graded boundary operator @ : C ! C is thus a formal sum of face operators @ = 2V D . It suces to show that @ 2 = 0 on each basis simplex . Computing the composition in terms of face operators, one obtains: (4.2) @ 2 = D D : 6= Each (k 2)-face of the k -simplex is represented exactly twice in the image of D D over all 6= . Thanks to F2 coecients, the sum over this pair is zero. The homology of C, H (C), is a sequence of F2 -vector spaces built from the following subspaces of chains.
Proof. For simplicity, consider the case of an abstract simplicial complex on a vertex set V = fv g. The face operator D acts on a simplex by removing the vertex v
Corollary 4.3.
For all
k , im @k +1 is a subspace of ker @k .
A cycle of C is a chain with empty boundary, i.e., an element of ker @ . Homology is an equivalence relation on cycles of C. Two cycles in Zk = ker @k are said to be homologous if they dier by an element of Bk = im @k +1 . The homology of X is the sequence of quotient vector spaces Hk (X ) over F2 , for k 2 N, given by:
Hk (X ) = Zk =Bk = ker @k / im @k +1 = cycles/ boundaries:
(4.3)
To repeat: Zk = ker @k is the subspace of cycles in Ck , and Bk = im @k +1 is the subspace of boundaries. Homology inherits the grading of C and will be therefore denoted H (X ) when a dimension is not specied. Elements of H (X ) are homology classes.
4.2 Homology examples

Example 4.4 (Graphs) Consider G a topological graph a compact 1-d cell complex. As @0 = 0, all 0-chains are 0-cycles: Z0 = C0 . The boundary subspace B0 Z0
60
Chapter 4. Homology
consists of nite unions of vertices, with an even number in each connected component of the graph. If one chooses the vertex set of G as a basis for Z0 , it follows that any two basis elements are homologous if and only if they are the endpoints of an end-to-end sequence of edges in G that is, if and only if they lie in the same connected path-component of G. The dimension of H0 (G) is therefore the number of such path components. Any element of Z1 (G) consists of a nite union of cyclic end-to-end sequences of edges. As a graph is a 1-dimensional complex, there are no higher dimensional chains with which to form homology classes: for graphs, 1-cycles are homologous if and only if they are identical. Correspondingly, H1 (G) collates the number of linearly independent cyclic chains of edges, and Hk (G) = 0 for k > 1. Note that homology gives the collection of 1-cycles in G the structure of a vector space, complete with the notions of sums, spans, and bases. compact orientable surface of genus g , the homology is concentrated in dimensions less than three, since all cells are of dimension less than three. The relation to genus is as follows:
Example 4.5 (Surfaces) For Sg a simplicial complex homeomorphic to a connected
2g dim Hk (Sg ) = 1 0
It is worth stressing that H1 (X ) does not measure whether a cycle in X is contractible whether it can be shrunk to a point in X continuously. For example, on an orientable surface, any loop (1-cycle) which divides the surface into two connected components is nullhomologous, though it may not be contractible. A few facts (presented without proof) may help the beginning reader build up an intuition for what it is that homology measures:
: : : :
k =0 k =1 : k =2 k >2
(4.4)
1. dim H0 (X ) is equal to the number of path components of X . 2. For X contractible, Hk (X ) = 0 for all k > 0. 2 a disc in the plane with n disjoint discs removed, dim H0 (D2 ) = 1, 3. For Dn n 2 ) = n , and dim Hk (D2 ) = 0 for k > 1. dim H1 (Dn n 4. The sphere Sn has Hk = 0 for all k except H0 and Hn , both of dimension one. 5. For a compact manifold M of dimension n, dim Hk (M ) = dim Hn k (M ). In particular, Hn (M ) has dimension one if and only if it is connected.! ! Let the reader beware: this result requires the use of . coecients. 2
4.3. Coecients 6. For a disjoint union of spaces
61
X = A B, Hk (X ) = Hk (A) Hk (B). Hk
(T n ) =
( )
Example 4.6 (Cross products) A torus Tn = (S1 )n has homology satisfying

dim
n k :
The reader will rightly suspect a relationship with polynomial algebra, via the the coecients of (1 + t )n . Consider the Poincar polynomial of X , Pt (X ) = k ck t k , with coecients ck := dim Hk (X ). The Poincar polynomial of S1 is 1 + t . The simple version of the Knneth Theorem states that Pt (X Y ) = Pt (X )Pt (Y ), from which the computation for Tn follows.
4.3 Coecients
The homology introduced in Equation (4.3) should be denoted the cellular homology of X with F2 coecients, cell (X ; F ). These additional adjectives are necessary or H 2 due to the plethora of homology theories in existence. Any graded sequence C = (C ; @ ) of vector spaces and linear transformations with the property @ 2 = 0 is a chain complex suitable for building a homology theory. For example, instead of the Ck being vector spaces over a eld F2 , other elds may be used. For any eld F, one may construct chains Ck for a cell complex X as F-vector spaces. Such coecients permit expressiveness: R is eective in describing simplices' intensities, in contrast to F2 's involutive switch.
Example 4.7 (Kirchho's current rule) Consider an electric circuit as a 1-dimensional

cell complex, with circuit elements (resistors, capacitors, etc.) located at certain vertices. Kirchho's current rule states that the current owing through the edges of the circuit satises a conservation principle: at each node, the sum of the incoming currents equals the sum of the outgoing currents. In the language of this chapter: current is a 1-cycle. Note the unambiguous need for R coecients, since current is measured as a real quantity.
There is an important development when going beyond F2 coecients: 1 6= 1. The eect of this algebraic unfolding is the need for an orientation associated to simplices. For 1-chains, this is naturally encoded as directedness. In calculus class, one uses a twist of the hand for orientations on curves and surfaces: a more formal approach is needed for the present setting. Recall that an abstract simplicial complex consists of simplices on a formal vertex set V . Each k -simplex of X is specied as k = fvo ; v1 ; : : : ; vk g for some choice of vi 2 V . An oriented simplex is a simplex with an ordering of vertices: [vo ; v1 ; : : : ; vk ]. An orientation on is an equivalence class of orderings up to even transpositions; thus, [vo ; v1 ; v2 ] = [vo ; v2 ; v1 ] = [v2 ; v0 ; v1 ]. This yields a well-dened multiplicative action of f 1; +1g on oriented chains: multiplication by 1 is the chain analogue of reversing orientation. In the algebra of chains,
62
Chapter 4. Homology
as with contour integration, two chains of opposite orientation cancel. The boundary operator incorporates orientations as follows (using the face operators from the proof of Lemma 4.2):
@ =
k i =0
( 1)i Di :
(4.5)
The reader should check that the use of ( 1)i 's suces to keep Lemma 4.2 in eect. The resulting homology is cell (X ; F). Field denoted H (X ; F), or, more formally, H coecients and linear algebra are, still, not the full story. The generalization to Z coecients in which each Ck is a Z-module is particularly relevant. One visualizes chains C (X ; Z) over a cell complex X with integer coecients as recording a nite collection of simplices with orientation and multiplicity. Ultimately, one works with a chain complex over an R-module (for R a ring) with the boundary maps module homomorphisms. ent homology groups. For F2 2 F2 Hk (K 2 ; F2 ) = F F 2
Example 4.8 (Coecients) Using dierent coecients can lead to genuinely dierexample, the Klein bottle
K 2 satises:
: : : :
k =0 : k =0 k = 1 : ; H (K 2 ; Z ) = Z Z Z 2 : k =1 : k k =2 0 : k >1 k >2
In this example, Z coecients yield dierent ranks of homology groups at grading one and two. The Klein bottle has no 2-cycle (a boundaryless nonzero chain) in Z coecients, but it does in F2 : thus, K 2 is a non-orientable surface. Note also the presence of torsion in H1 (K 2 ; Z) = Z Z2 , where Z2 = Z=2Z is the quotient of Z by the subgroup of evens. This indicates that the meridian and longitudinal curves on K 2 , unlike those of the torus T2 , are qualitatively dierent: sliding a meridional curve along a longitude reverses orientation. In general, the presence of a torsional element in H (X ; Z) is indicative of some type of twisting in X which, in surface examples, manifests itself in non-orientability. ple, and useful example of torsional phenomena where coecients matter. Consider SO3 , the group of real 3-by-3 matrices with determinant 1. These are precisely the orientation-preserving rotations of Euclidean 3-space. As one can demonstrate using a variety of physical devices (plates, belts, fermions, etc.), there is a torsional
Example 4.9 (Rotations and projections) Rotations in R3 oer a fascinating, sim-
4.4. Singular homology
63
core writhing within: H1 (SO3 ; Z) = Z2 . Since SO3 = P3 , this is an example of the n homology of projective spaces. In general, dim Hk (P ; F2 ) = 1 for all 0 k n. The story for integer coecients is quite dierent:

Hk (Pn ; Z) = Z2 :
0 :
k = 0 or k = n; odd 0 < k odd < n :

else
4.4 Singular homology

A denition with sucient strength has emerged: a chain complex C = (Ck ; @k ) is any sequence of R-modules Ck with homomorphisms @k satisfying @k @k +1 = 0. Though the generators of C may be simplices or cells of a space X , there are many dierent objects worthy of being counted and compared. To be noted are the commonalities (e.g., notions of dimension associated to the objects counted) and the distinctions (e.g., auxiliary structures imposed). The next several sections delineate a few major homology theories (with others to follow in later chapters). The rst such example is singular homology the most common and generally useful theory. The reader will have observed that cellular homology is, like Euler characteristic, independent of how the space is decomposed into cells. The best method of proof comes not from parsimonious combinatorial renements of cell structures, but rather by limits to uncomputable abundance. One of the rst to take such an approach was Vietoris, who introduced his homology theory for metric spaces [?]. In this theory, one xes an > 0 and considers the chain complex is generated by k -tuples of distinct points in X of pairwise distance . These simplices do not t together to form a triangulation of X : there are far too many. However, for reasonable spaces (e.g., metric nite cell complexes), the homology of the Vietoris chain complex stabilizes as ! 0+ , and this agrees with the cellular homology of a cell complex. The reason that Vietoris' theory is mostly forgotten lies in the ecacy of singular homology, which requires neither a metric space nor the explicit limiting process. The singular chain complex of a topological space X is the complex Csing = (Ck ; @k ), where the generators of Ck are continuous maps k : k ! X from the standard simplex k to X . The boundary of a singular simplex is the restriction of the map to @ k . Notice that there is a decoupling between the grading and the dimension of the image of the singular simplex. Indeed, the image of a singular simplex may be very convoluted, with no sensible notion of dimension other than in its preimage. The
64
Chapter 4. Homology
sing ; @ ) is enormous certainly of uncountably innite resulting chain complex (C dimension except in the most trivial cases. However, in this lies its exibility. For two spaces that are homeomorphic, there is an equivalence between their singular chain complexes that guarantees equivalent homologies. The yet more useful and general homotopy invariance of singular homology is the true reward for the unwieldy bulk of the singular complex, as will be shown in the next chapter.
4.5 Reduced homology

It is sometimes convenient to modify a homology theory in grading zero. Given a ~ chain complex C over an R-module, dene the reduced chain complex C
/ Cn
@n / Cn 1 @n
@2 / C1
@1 / C0

/
where : C0 ! R sends each basis 0-chain to the sum of the coecients." It is ~k for the chain groups, where k 2 Z and: advantageous to write C Ck : k 0 ~k = R :k= 1 ; C
/0;
(4.6)
:k
< 1
with all boundary maps as above.
Proposition 4.10.
homology
For a nonempty chain complex C with coecients in G , the reduced ~ (C) satises Hk (C) ~k (C) for k > 0 and H0 (C) ~0 (C) R. H =H =H
The reader may think of the codomain of as C 1 , having a single basis element at grading 1. Reduced homology simplies certain results: it is convenient to have ~ = 0, as such an acyclic space plays the role a nonempty contractible space have H of a zero in homological algebra.
4.6 ech homology

Any homology theory counts objects with grading and cancelation. While the objects and gradings are ultimately indicative of cells and dimensions, the correspondence may be disguised. Consider the case of the ech homology of a cover. Let X be a topological space and U = fU g a collection of open sets whose union is X . (U) = (C k ; @ k ). The 0-chains C 0 have as Consider the following chain complex C basis the elements of U, and the k -chains Ck have as basis non-empty intersections of k + 1 (distinct, unordered) elements of U. The resulting chain complex is outtted with the obvious boundary operator: k k (U0 \ \ Uk ) = (4.7) @ ( 1)i Uj : i =0 j 6=i
" For
.-vector spaces, : C0 ! ..
4.7. Relative homology
65
@ = 0 and the ech In F2 coecients, the signs are superuous; one checks that @ (U)), is well-dened. The reader will be reminded U := H (C homology of the cover, H (U) is the algebraic shadow of the of the nerve complex of Chapter 2; the complex C geometric nerve. A homological version of the Nerve Lemma (Theorem 2.3) follows:
Theorem 4.11.
of
~ 0), then the ech homology of U H (U) agrees with the singular homology of X : H = H (X ).
U
are acyclic (
If all nonempty intersections of elements
The ech homology of a space X is well-dened, independent of the cover, by taking a limit of ner and ner covers. The details of this limiting procedure require a few tools from Chapter 8 and are not here detailed: see, e.g., [?]. This is particularly useful for characterizing holes in non-tame spaces that the more well-behaved singular theory cannot abide.
4.7 Relative homology
Homology relative to a subcomplex is an important variant. Let A X be a subset (or subcomplex, as appropriate) and C(X ) a chain complex (singular, cellular, etc.) on X . There are two natural chain complexes implicating A. The rst, C(A), consists of subgroups Cn (A) < Cn (X ) with the obvious restriction of @ . This chain complex yields the homology of A. The more subtle construct is the chain complex of the quotient: let C(X; A) denote the quotients Cn (X; A) = Cn (X )=Cn (A) of chains on X modulo chains on A, with the induced boundary maps @ on quotients. These relative chains do not vanish on A; rather, they are equivalence classes of chains in X which are identical o of A. The resulting relative homology H (X; A) collates homology classes of relative cycles (chains in X whose boundaries lie in A).
Example 4.12 (Reduced homology) The reduced homology of a space X is isomorphic to the homology of X relative to a basepoint p 2 X . The only nontrivial chain in X whose boundary is a nonzero multiple of p is, precisely, a multiple of p and a zero-cycle. Thus, by denition, Hk (X; fp g) = Hk (X ) for k > 0 and H0 (X; fpg) has rank one less than H0 (X ), since the subspace corresponding to the homology class [fpg] has been quotiented out.
One of the foundational theorems about homology concerns relative homology.
Theorem 4.13 (Excision).

interior of
A.
Then
Let U A X with the closure of U H (X U; A U ) = H (X; A).
contained in the
66
Chapter 4. Homology Excision implies that for the pair (X; A), what happens inside of
A is irrelevant.
Corollary 4.14.
~ (X=A). H
For
complex of a cell complex
A X a closed subX , H (X; A) =
The need for a subcomplex is superuous. In the singular setting, it suces to have A closed and possessing an open neighborhood which deforms continuously onto A (unlike, say, a closed subset of Q).
Example 4.15 (Relative homology) The homology of a closed disc Dn relative to its
boundary @Dn has as a nontrivial generator Dn itself, as a relative n-cycle. From Corollary 4.14 and the fact that Dn =@Dn is homeomorphic to Sn , one sees H (Dn ; @Dn ) = ~ (Sn ). H
The local homology of p 2 X is the (singular) homology H (X; X p ). This is a little hard to visualize, given that Corollary 4.14 is inapplicable, but a moment's thought yields the intuition that it measures something of the features of X in an arbitrarily small neighborhood of p . The correctness of this intuition follows from excision: local homology is, truly, local.
4.8 Local homology
Corollary 4.16.
Let
open neighborhood of
p 2 X be a (closed) point and V an p. Then H (X; X p) = H (V; V p):
that is, local homology can be computed from an arbitrarily small neighborhood of
p.
Proof. Apply Excision, with U = X V and A = X p. Example 4.17 (Local orientation) It follows from Example 4.15 that the local ho~ (Sn ) and thus nontrivial of rank one in mology of an n-manifold is isomorphic to H dimension n. This leads to a painless denition of orientation, a concept that usually requires recourse to a dextrous sleight of hand. A local orientation at p 2 M is a choice of generator for the n-dimensional local homology at p (in Z coecients). A global orientation on a compact n-manifold M is a choice of generator for Hn (M ; Z). A global orientation, in keeping with one's experience from calculus class, is a consistent choice of local orientations.
4.9. Homology of a relation
67
A relation between two sets X and Y is a subset R X Y . A point x 2 X is related to a point y 2 Y if and only if (x; y ) 2 R. For example, the sensing modality used for target enumeration in Chapter 3 is a relation between sensors and targets collating which sensors detect which targets. It is possible and protable to build homologies based on relations. For R X Y a relation, dene a chain complex C(XR ) as follows: generators of Ck (XR ) are unordered (k + 1)-tuples of points in X related to some xed y 2 Y . Dually, one constructs generators of C(Y R ) from vertical bers of R unordered tuples of points of Y related to a xed x 2 X . The boundary operators for C(XR ) and C(Y R ) forget points in the tuples with sign, mimicking the boundary operator of a simplicial complex (the property of being related to a common element is closed under subsets). The homology of the relation, H (R), is dened as follows:
4.9 Homology of a relation
Theorem 4.18 (Dowker [?]). H (R) := H (XR ) = H (Y R ).

This result was originally used to prove an equivalence between ech and Vietoris homology theories, in the case where X is a metric space, Y is a dense net of points in X , and R records which points in X are within > 0 of points in Y . The nerve of the cover of X by metric balls about Y has homology H (Y R ), and the Vietoris homology of X is captured by H (XR ). Other applications are more salient to this text.
Example 4.19 (Ad hoc communications networks) An ad hoc communications network among a nite set X of devices is generated when devices broadcast their identities (assumed unique) and any other devices within range hear the signal and establish a link. The communications network can be encoded in a relation R X X , where (xi ; xj ) 2 R if and only if messages sent from device xi are heard by device xj . By convention, the diagonal = f(xi ; xi )g R. One may wish or assume or impose that R is symmetric (devices are heard by those they hear); this case permits topologizing the network as a graph whose edges are symmetric communication links. In a nonsymmetric case, there are two possible directed graphs: one encoding hears and one encoding heard by. Higher-dimensional simplicial models are also available. Following the constructions of Theorem 4.18, let X R be the simplicial complex on X determined by vertical bers of R any collection of devices which hear device xi 2 X forms a simplex of X R . Dual to this is XR , the complex on X determined by horizontal bers of R any collection of devices heard by device xi forms a simplex of XR . These two simplicial
68 structures on
Chapter 4. Homology
X have, by Theorem 4.18, isomorphic homology: H (X R ) = H (R) = H (XR ):
One imagines this could be useful in controlling certain dynamical systems over networks where nodes carry information that evolves under the inuence of neighboring nodes (neighborhood connoting a heard or heard by relation). One suspects that issues of consensus and convergence under evolution are more delicate for systems with H (R) nontrivial.
4.10 Functoriality
Homology gives a natural measure of the qualitative features not only of spaces but also of maps between spaces. This property, at the heart of algebraic topology, is functoriality, and it is from this principle the power of the subject emerges. Given a (cellular or continuous) map f : X ! Y , there is a graded induced homomorphism f : C (X ) ! C (Y ) at the level of (cellular or singular) chains given by f () = f (where is a basis k -chain). For a continuous map, the induced chain map respects the boundary operation, f @ = @f . Passing to homology yields the induced homomorphism H(f ): H X ! H Y . This map is as it must be: for a cycle, dene H(f )[] = [f ] = [f ]. Functoriality means that induced homomorphisms on homology are an algebraic reection of the properties of continuous maps between spaces. 1. Given f : X ! Y , H (f ): H X ! H Y is a (graded) homomorphism. 2. The identity map Id : X ! X induces the isomorphism Id : H X ! H X . 3. Given f : X ! Y and g : Y ! Z , H (g f ) = H (g ) H (f ). There is hardly a more important feature of homology than this. One implication in the sciences is inference. It is sometimes the case that what is desired is knowledge of the homology of an important but unobserved space X ; the observed data comprises the homology of a pair of spaces Y1 , Y2 , which are related by a map f : Y1 ! Y2 that factors through a map to X , so that f = f2 f1 with f1 : Y1 ! X and f2 : X ! Y2 . If the induced homomorphism H (f ) is known, then, although H (X ) is hidden from view, inferences can be made.
Example 4.20 (Experimental imaging data) The problem of measuring topological features of experimental data by means of imaging is particularly sensitive to threshold eects. Consider, e.g., an open tank of uid whose surface properties are experimentally measured and imaged. Perhaps the region of interest is the portion of the uid surface above the ambient height h = 0; the topology of the set A = fh 0g must be discerned, but can only be approximated by imprecise pixellated images of fh 0g. Similar situations arise in MRI data, where the structure of a tissue of interest can be imaged as an approximation. Given a reasonably close image, is an observed topological feature (say, a hole, or lack of connectivity) true?
4.11. Inverse kinematics
69
If the goal is to accurately capture the topological features of an exact but unmeasurable set, functoriality may assist. Let, e.g., A = fh 0g be the desired but unseen set. If one can measure approximants A A A+ from below and above and then match common features of these images, then one has a simple diagram of the form: H ( ) * / H A / H A+ ; H A where the map H () is the induced map on the inclusion : A ! A+ that itself factors through inclusion to the invisible desideratum A. Any nonzero element in the image of H () must factor through a nonzero homology class in A: one can discern the presence of a true hole with two imprecise observations and a map between them. This simple observation is greatly extendable to the concept of persistence, as will be seen in 5.12-5.14.
4.11 Inverse kinematics

Other applications of homology are obstructive. Consider the idealized robot arm of Example 1.5 consisting of N rotational joints and rigid rods, grasping a part at the end of the arm. In many applications in manufacturing, one must perform part placement manoeuver the arm so as to locate the part in the correct location and in the correct orientation. The available motions are controlled by the N rotational joints. One topologizes the problem as per [?]: consider the kinematic map : TN ! SO3 taking the ordered sequence of rotations to the net orientation of the part at the end of the arm. As is common when mathematicians study robot arms, the mechanisms are considered insubstantial, and no thought is wasted on the problem of self-intersection. The critical issue is that of understanding and inverting the kinematic map. Given a part orientation, is there a sequence of rotations to realize it? Certainly, is onto for a choice of axes which span R3 . It is the inverse kinematic map which is problematic. Given a xed part orientation, are all nearby part orientations realizable via small changes in the rotations required? This local problem seems to be a reasonable assumption; a global version is not.
70
Chapter 4. Homology
There is no continuous section to
Proposition 4.21.
s : SO3 ! TN
satisfying
s = Id.
in
.
That is, there is no map
Proof. Consider a putative section s : SO3

Examples 4.6 and 4.9,
s =
Id. Then on
H1
Z coecients one has, from

(/
! TN
with
Id
Z2
H (s )
/ ZN
H()
Z2 ;
H (s ) = 0
and
H() H(s ) = 0.
Id, an identity map, thanks to functoriality. This yields a contradiction: since ZN has no nonzero elements of nite order,
the composition of which is
H () H (s ) = H ( s ) =
This result means that a continuous assignment of robot arm rotation angles as a function of part orientation is impossible. This decidedly non-intuitive result is a direct consequence of the algebraic topology of conguration spaces, as revealed by H1 . Note, however, that, in general, one is dependant on luck. Consider the case where the part grasped by the robot arm is rotationally symmetric about the last axis, such as a bolt or peg to be inserted. The desideratum is no longer an orientation in SO3 but rather in S2 , the direction in which the last axis points. Repeating the argument of Proposition 4.21 with the map 0 : TN ! S2 leads to frustration, as H1 (S2 ) = 0. There is no analogous contradiction in assuming that the map H (0 ) H (s 0 ) is the identity, since it is the identity homomorphism on H1 (S2 ) = 0. One does not conclude that the inverse kinematic map s 0 : S2 ! TN necessarily exists. In fact, s 0 does not exist. At the level of H2 , there is a diagram:
Id
H2 (S2 )
H (s 0 )
H2 (TN ) K
+ 2 / 0 ) H2 (O S ) H ( KKK KKK H() KKK% H2 (SO3 )
(4.8)
Note that 0 factors through as 0 : Tn ! SO3 ! S2 by enacting and ignoring all but the last axis in the frame. The top-row composition H (0 ) H (s 0 ) cannot be the identity map on H2 (S2 ), since H2 (SO3 ) = 0 while H2 (S2 ) 6= 0. Again, functoriality reveals what individual homology groups do not. There is no continuous assignment of rotation angles to a directional axis for a robot arm manipulating a part.
4.12 Winding number and degree

It is helpful to think of an induced map on homology as akin to a winding or linking: to what extent are the holes of X wound about the holes of Y via f : X ! Y ? The
4.12. Winding number and degree
71
full complexity of homomorphisms between abelian groups gives an algebraic picture of the wrapping and winding that maps can execute. Indeed, induced homomorphisms are the right way to express the classical notion of winding numbers. A continuous simple closed curve : S1 ! R2 in the plane separates the plane into two connected regions: dim H0 (R2 (S1 )) = 2. The mod-2 winding number of about a point p 2 R2 not in the image of is, intuitively, the number which represents whether p is inside (1) or outside (0) the image of . Of the many denitions the reader may have seen (either involving integrals or the clever counting of intersections), the best is via homology. For : S1 ! R2 fp g, consider the induced homomorphism:
H( ): H1 (S1 ) ! H1 (R2 fpg):

Both the domain and codomain of H ( ) are of rank one, and the map H ( ) is therefore multiplication by a constant: that constant deg( ) is the winding number (perhaps in F2 or Z depending on the coecients used). There is no need to restrict to non-selfintersecting curves: any : S1 ! R2 fp g determines a homology class. If, instead of F2 coecients, the integers Z are used, then the class in H1 (R2 fp g; Z) is the winding number of about p. The winding number measures the algebraic number of times the image of wraps about p , with a choice of orientations (both of and R2 fpg) determining the sign. From the fact that Hn Sn = Z, any self-map of Sn induces a homomorphism on Hn which is multiplication by an integer: this is the degree of the map. The resulting degree theory is an important classical topic, some of the important points of which are as follows: 1. 2. 3. 4. 5.
deg is a homotopy invariant. deg(f g ) = (deg f )(deg g ). deg Id = 1. deg a = ( 1)n for a : Sn ! Sn the antipodal map a(x ) = deg f = deg g if and only if f ' g [the Hopf Theorem].
x.
Example 4.22 (Index theory and vector elds) The index theory for vector elds introduced in 3.3 is better understood in terms of degree. Let V be a vector eld on a manifold M with isolated xed point p , and let Bp be a suciently small ball about p. The index of V at p, IV (p), is dened to be the degree of the map V j@Bp : @Bp = Sn ! Rn f0g ' Sn , where V is represented in local coordinates on Bp . This index is well-dened, since p is an isolated xed point. By the properties of degree, the index is independent of the neighborhood Bp chosen, so long as its intersects Fix(V ) in p alone. Example 4.23 (Linking number) Degree is well-dened for any map between oriented compact n-manifolds, since Hn = Z for such. One example of this form of degree is in classical knot theory, the study of embeddings S1 ,! S3 . One intuits
72
Chapter 4. Homology
that there are many inequivalent ways to tie a simple closed curve in R3 . Degree can be used to characterize how two disjoint oriented knots link together. To dene the linking number of two disjoint oriented knots, K1 and K2 , parameterize each knot as 1 ; 2 : S1 ! S3 and dene `k (K1 ; K2 ) to be the degree of the map : T2 ! S2 given by:
The reader may observe that this quantity is invariant of the parametrization chosen, so long as orientation is respected. Linking number has been found useful in a variety of contexts, from owlines in uid- and magnetohydrodynamics [?] to closed DNA strands [].
( ) ( ) : 1 ; 2 7! k 2 (2 ) 1 (1 )k : 2 2 1 1
(4.9)
which interpolate between ordered and disordered structure. Besides being of great commercial interest, liquid crystals oer fascinating observable defects whose classication is inherently topological. For concreteness, consider the nematic liquid crystals, composed of axisymmetric rod-like molecules whose alignment (or director eld), under a continuum assumption based on average behavior, is a continuous function of position apart from certain defects or disclinations [?]. A topological classication of such defects is intricate [], but an initial step uses degree as follows. Motivated by Example 4.22, one sets up a map whose induced action on homology yields an appropriate degree.
Example 4.24 (Nematic liquid crystals) Liquid crystals are a mesophase of matter
In a nematic liquid crystal that is taken so thin as to be approximated by a plane region to which the molecules align (a so-called Schlieren texture), the director eld assigns to points in R2 a planar direction in P1 , well-dened and continuous o the set of singular point-defects. To assign an index to such a defect at x 2 R2 , choose an oriented small loop encircling x and consider the restriction : ! P1 = S1 . The resulting index I (x ) 2 Z is an undirected version of that used for vector elds. In the literature, one often divides this integer quantity by 2 so as to obtain agreement with the vector eld index when the line eld is orientable [?]. In the case of a nematic in R3 , the direction eld takes values in P2 . Following the planar case, choose a suciently small ball B around an isolated point-defect x
4.12. Winding number and degree
73
and consider the restriction of : @B = S2 ! P2 and its induced map H (): H2 (S2 ) ! 2 2 H2 (P ). From Example 4.9, H2 (P ; Z) = 0, and the (integer-valued) degree is trivial; however, passing to F2 coecients for H2 yields a well-dened index I (x ) 2 F2 , since H2 (P2 ; F2 ) = F2 . For singularities occurring along a disclination curve, index is measured by a loop = S1 locally linking the disclination. The resulting map on 1 homology H (): H1 (S ) ! H1 (P2 ) is again going to return a degree I 2 F2 , which, like that of a point-singularity, is sensitive to orientability of the direction eld about the disclination.
Notes
1. The reader will note, perhaps with displeasure, that computational issued have not been discussed: it is too large and uid a topic. See [ ,
? ?, ?]
for an introduction to
the many techniques available for fast computation of homology, some of which are distributable (a near-necessity for realistic scientic applications). The short version
seems to be that cellular homology is computable in time near-linear in the number of cells. The curse of dimensionality (high-dimensional complexes have many, many cells) remains a challenge, however. 2. The Knneth theorem of example 4.6 holds for singular or cellular homology in eld coecients, for spaces with nite dimensional homology groups. There is a more general version which accommodates more general coecients. In the case of
Z coecients
?
and spaces with torsion, it is exact but algebraically delicate: see, e.g., [ ]. 3. 4.5 is the rst hint that taking advantage of negative dimensional chains and homology may be permissible. There are homology theories in which
H
takes on signicant
meaning for negative gradings. 4. Theorem 4.18 on the homology of relations is not the strongest possible result: as with the Nerve Lemma, a homotopy equivalence between the two complexes exists. 5. This treatment of winding numbers begins with an invocation of the classic Jordan Curve Theorem, which, in the language of this text, says: Let homeomorphic to
Sk .
Then
~p (Sn H
A) is Z for p = n k
be a subset of
and
Sn
else. Note that
there is no assumption of tameness in this statement. 6. Linking numbers are just the beginning of the algebraic topology of knots and links a vast subject. Knot theory has found some applications, notably in physics, biology, uids dynamics, and dierential equations. 7. The treatment of singularities in nematic liquid crystals in Example 4.24 is woefully incomplete, given the tools here used. Very interesting phenomena occur as the eld is continuously changed and disclination lines are enticed to collide or entangle, or as point-defects move so as to encircle a disclination line. The homological index is incapable of handling the resulting (non-abelian!) phenomena, and tools from homotopy theory (see A.3) are required. Beyond the simple nematics are other types of
liquid crystals, with bi-axial nematic, cholesteric, and smectic phases being the most common. 8. There are recent applications of homology and linking numbers to problems of in robotics [ ,
D R2 that prevents the object from being able to be moved 2 `to innity' by means of Euclidean motions (translations and rotations) in R D. The
the smallest discrete set
? ?],
where, given a xed geometric object in
R2 , one wishes to choose
caging
related problem of performing motion-planning in robotics about obstacles by means
74
Chapter 4. Homology
of specifying homology classes leads to computable optimization methods [ ].
Exercises
4.2. 4.3.
4.1. Verify, for every compact space
H (X ; F) with eld coecients Hk (X ; F). Give a proof of Lemma 4.2 for the case of cube complexes with F2 and Z coecients. Triangulate a Mbius strip M and compute the homologies H (M ; F2 ) and H (M ; Z).
has been computed or stated, that
(X ) =
whose homology
k k ( 1) dim
Can you see any hint of non-orientability from your computations? 4.4. The
4.5. 4.6.
4.7.
p via a X ) in terms of the individual homologies H (X ). n n 1 ), where Sn 1 is an equator of Sn . Compute H (S ; S Let X be a simplicial or cubical complex and v a 0-cell of X . What is the relationship between local homology at v and the link of v (as dened in 2.7)? Let X be the set of all web sites and R X X the links-to relation whose vertical
point from each factor and identifying them all to a single point quotient map. Compute
wedge sum of spaces
p 2 X
H (
is the quotient space obtained by choosing a single
ber over a xed web page are other web pages to which the base page links. Speculate on the topology of with
4.8. Show that an abstract simplicial complex must have 4.9.
X with k n + 1 simplices of dimension n H n (X ) = 0. 2 2 2 2 2 Consider the torus T as the quotient R =Z . Let f : T ! T be given by the linear
transformation
H (R).
Do your guesses change if
is the set of Twitter users
the follows relation?
for
a; b; c , and d
a b c d
constants. What is the induced map
H(f ) on homology (say, in R
coecients)? 4.10. What happens to the results of 4.11 if self-intersections of the robot arm are forbidden? What would you need to know about the resulting conguration space of the robot arm (now a proper subset of 4.11. Consider the kinematic map
R2 .
: Tn
! SO
Tn ) to still have a lack of section s ?
2 of a planar robot arm holding a at part in
Is there a homological obstruction to the section
Can you construct
s?
s : SO2 ! Tn
with
s = Id?
4.12. Prove that the degree of a map
f : Sn ! Sn is zero if f
R
3
is not surjective. Construct a
counterexample to the converse. 4.13. Give an explicit construction of a map 4.14. Give an example of a pair of knots in
f : Sn ! Sn with degree k , for each k 2 Z.

whose linking number is zero but which cannot
be separated (let the reader consider what the correct denition of separated should mean). 4.15. What happens to linking number when the orientation of one or both curves is reversed? 4.16. How would the classication of singularities in Example 4.24 change if the molecules were disc-like, as opposed to rod-like? What if the molecules were bi-axial (having the shape of an ellipsoid with no rotational symmetries)?
Chapter 5
Sequences
76
Chapter 5. Sequences
omology takes as its input a chain complex a hierarchical assembly line of parts and returns the global features. The elemental tools for analyzing homology are a similar linelike devices: chain complexes, chain maps, and, especially, exact sequences. When combined with functoriality, they assemble into commutative diagrams in homology. This chapter unpacks these tools and puts them to use.
5.1 Homotopy invariance

Homology begins by replacing topological spaces with complexes of algebraic objects (vector spaces, abelian groups, or modules, depending on one's preferences). Further topological notions continuous functions, homeomorphisms, homotopies, etc. also have analogues at the level of chain complexes. A chain map is a map ' : C ! C0 between chain complexes that is a homomorphism on chain groups respecting the grading and commuting with the boundary maps. This is best expressed in the form of a commutative diagram:

' @ = @ 0 ' .
/ Cn+1
/ Cn
@ / Cn
@0 / 0 Cn
@
/

/
(5.1)
' @0 0 / Cn +1
0 / Cn
'
'
1
@0
Commutativity means that homomorphisms are path-independent in the diagram; e.g., Chain maps are the analogues of continuous maps, since, via respect for the boundary operators, neighbors are sent to neighbors. The appropriate generalization of a homeomorphism to chain complexes is therefore an invertible chain 0 . Clearly, a homeomorphism map one which is an isomorphism for all Ck ! Ck sing sing f : X ! Y induces a chain map f : CX ! CY which is an isomorphism. As sing (X ) sing (Y ). The extension to homotopy is more subtle. Recall that such, H = H f ; g : X ! Y are homotopic if there is a map F : X [0; 1] ! Y which restricts to f on X f0g and to g on X f1g. A chain homotopy between chain maps ' ; : C ! C0 is a homomorphism F : C ! C0 sending k -chains homomorphically to (k + 1)-chains so that @ 0 F F @ = ' :
@ / / Cn+1 @ / Cn / Cn 1 @ { { { { { F {{{ ' F {{{ ' F {{{ ' {{{ {{ {{ {{ {{ { }{{ }{{ F }{{ }{{ 0 0 0 / Cn / / Cn / Cn +1 1 @0 @0 @0
(5.2)
One calls F a map of degree +1, indicating the upshift in the grading. Note the morphological resemblance to homotopy of maps: a chain homotopy maps each k chain to a k + 1-chain, the algebraic analogue of a 1-parameter family. The dierence between the ends of the homotopy, @ 0 F F @ , gives the dierence between the chain maps.
5.2. Exact sequences
77
Lemma 5.1.
Chain homotopic maps induce the same homomorphisms on homology.
2 H (C). Assuming chain homotopic maps from C to C0 ,

Proof. Consider []
'
and
are
H(' )[] H( )[] = [(' )] = [(@ 0 F + F @ )] = [@ 0 F ] = 0;

since
is a cycle and @ 0 (F ) is a boundary.
The following theorem is proved by constructing an explicit chain homotopy:
Theorem 5.2.
Csing X
! Csing Y
Homotopic maps
f;g : X ! Y
induce chain homotopic maps
f ; g :
The idea behind the proof is simple. For each singular k -simplex , one considers the F -image of [0; 1] as a family of singular k -simplices parameterized by the homotopy. This prism is then triangulated into singular (k + 1)-simplices that encode the homotopy. It is shown that this chain map P (called a prism operator [?]) is a chain homotopy.
Corollary 5.3.
Singular homology is a homotopy invariant.
5.2 Exact sequences

n.
A complex C = (C ; @ ) is exact when its homology vanishes: ker @n = im @n+1 for all Exact sequences are as ubiquitous as the nullhomologous spaces they mirror.
Example 5.4 (Exact sequence) The following simple examples of exact sequences
help build intuition: 1. Two abelian groups are isomorphic, G = H , if and only if the following sequence is exact: /H /0 /G 0
78
Chapter 5. Sequences 2. The rst isomorphism theorem for groups says that for a homomorphism G , the following sequence is exact:
' of
/ ker
'
/
' / im
'
/0
Such a 5-term sequence framed by zeroes is called a short exact sequence. In any such short exact sequence, the second map is injective and the penultimate map is surjective. 3. More generally, the kernel and cokernel of a homomorphism ' : G ! H t into an exact sequence:
/ ker
'
/
' / H
/ coker
'
/0
4. Consider dierentiable functions C = C 1 (R3 ) and X, the vector space of vector elds on R3 . These t together into an exact sequence,
C1
R
/
C r
/ X r / X
r / C
/0;
(5.3)
where r is the gradient dierential operator from vector calculus, and the initial R term in the sequence represents the constant functions on R3 . This sequence encodes the fact that curl-of-grad and div-of-curl vanish, as well as the fact that, on R3 , all curl-free elds are gradients and all div-free elds are curls. This one exact sequence compactly encodes many of the relations of rst-year undergraduate vector calculus.
The most important examples of exact sequences are those relating homologies of various spaces and subspaces. The critical technical tool for the generation of such weaves an exact thread through a loom of chain complexes.
Theorem 5.5 (Snake Lemma).

0
induces the
Any short exact sequence of chain complexes
/ A
i / B
j
/ C
/0;
long exact sequence:

/
Hn (A )
H(i ) / H (j ) / Hn (B ) Hn (C )
/ Hn
(A )
H (i ) /
(5.4)
Moreover, the long exact sequence is functorial: exact sequences and chain maps
a commutative diagram of short
0 0
/ A
/ B
/ C
/0 /0
f /A ~
g /B ~
h /C ~
5.3. Pairs
induces a commutative diagram of long exact sequences
79
Hn (A )
/ /
Hn (B )
/ /
Hn (C )
/

/
Hn 1 (A )
/ /
(5.5)
/
~ ) Hn (A
H(f )

H (g )

H (h )

H (f )
~ ) Hn (B
~ ) Hn (C
~ ) Hn 1 (A
An exact sequence of chain complexes means that there is a short exact sequence in each grading, and these short exact sequences t into a commutative diagram with respect to the boundary operators. The induced connecting homomorphism : Hn (C) ! Hn 1 (A) comes from the boundary map in C as follows: 1. 2. 3. 4. 5. Fix [ ] 2 Hn (C). By exactness, = j ( ) for some 2 Bn . By commutativity, j (@ ) = @ (j ) = @ = 0. By exactness, @ = i for some 2 An 1 . Set [ ] := [] 2 Hn 1 (A).
Any topologist, no matter how wedded to geometric intuition, must possess a thorough understanding of the connecting homomorphism. This is perhaps best grasped via animation; a static illustration is a poor substitute, but nevertheless conveys the critical shift in grading that exactness and weaving enacts. The reader should demonstrate that is well-dened and that the resulting long exact sequence is indeed exact. Doing so solidies the invaluable technique of diagrammatic argument.
5.3 Pairs
Given A X (a subset in the singular category, or a subcomplex in the cellular), the following short sequence is exact:
/ C (A)
i / C (X )
j / C (X; A)
/0;
where i : A ,! X is an inclusion and j : (X; ?) ,! (X; A) is an inclusion of pairs. This yields the long exact sequence of the pair (X; A):
/
Hn (A)
H (i ) / H (j ) / Hn (X ) Hn (X; A)

/
Hn 1 (A)
/
(5.6)
The connecting map takes a relative homology class [] homology class [@] 2 Hn 1 (A).
2 Hn (X; A) to the
80 cation of the long exact sequence of the pair (Dk ; @Dk ):

/
Example 5.6 (Spheres) Computing the homology of the sphere Sk is a simple appli-
Hn (Dk )
H(j ) / Hn (Dk ; @Dk )

/
Hn 1 (@Dk )
H (i ) / Hn
(D k )
/
By denition, @Dk = Sk 1 . By excision, H (Dk ; @Dk ) = ~ (Dk =@Dk ) ~ (Sk ). For all n > 0, Hn (Dk ) = 0. As the H =H rst and last terms in the sequence above vanish, exactness yields the recursion relation
~n (Sk ) H = Hn 1 (Sk 1 );
for all n > 1. Beginning with the explicit and trivial compu~ (S0 ), one inducts to show that H ~n (Sk ) tation of H = Z for n = k , and = 0 else.
5.4 Mayer-Vietoris
Another important sequence is derived from a decomposition of X into subsets (or subcomplexes) A and B . In the singular category, one requires X = int(A) [ int(B ); in the cellular case, subcomplexes suce. Consider the short exact sequence
/ C (A \ B )
/ C (A) C (B )
/ C (A + B )
/ 0;
with chain maps : c 7! (c; c ), and : (a; b) 7! a + b. The term on the right, C (A + B ), consists of those chains which can be expressed as a sum of chains on A and chains on B. In cellular homology with A; B subcomplexes, C (A + B) = C (X ); in the singular category, one shows (via the techniques of ech homology) that H (A + B ) = H (X ). In both settings, the resulting long exact sequence yields the Mayer-Vietoris sequence:
/
Hn (A \ B)
H() / H( Hn (A) Hn (B )
Hn (X )

/
Hn 1 (A \ B)
/
The connecting map decomposes a homology class in X into portions in A and B , then takes the boundary of one of these portions in A \ B . This sequence captures the integrative nature of homology one builds from local to global. Mayer-Vietoris as follows. Let A and B be hemispheres intersecting in an equato~ (A) ~ (B ), one obtains by exactness that : Hn (Sk ) rial Sk 1 . As H = 0 = H = k 1 Hn 1 (S ) for all n > 1 and all k .
Example 5.7 (Spheres, redux) The computation of H (Sk ) can be carried out via
5.5. Degree and local computation
81
5.5 Degree and local computation

Computing degree (4.12) is often possible by means of local computations. For example, given a point p 2 R2 and a closed curve : S1 ! R2 fp g, the winding number of about p is easily computed as follows. Draw a ray from p and perturb it (via transversality, if desired) to intersect the image of in a nite set of points fqi gK 1 . At each intersection point qi , the curve kisses or crosses the ray; either the curve traverses (left-to-right or right-to-left, since both are oriented) or it osculates, touching the ray and immediately turning back. Each action contributes a local degree: 1 if crossing and 0 if kissing. The net winding number of about p is the sum of these local contributions to degree. This simple example of a local computation rewards rumination. What happens if, instead of a curve in the plane, one needs to know whether a cycle in an ad hoc nonlocalized network surrounds a node (as in, e.g., a network of security cameras [?])? To know whether a curve surrounds a point in the plane, it suces to know the local behavior of the curve at a (small) nite number of points. What the curve does elsewhere is irrelevant. This is the pattern and stamp of topology. This intuitively simple procedure has a rigorous footing in transversality [?]; better still is the use of local homology. Assume f : Sn ! Sn and q 2 f 1 (p ) is an isolated point in the inverse image. Dene the local degree of f at q to be
deg(f ; q ) = deg
H(f ) : Hn (V; V q ) ! Hn (U; U p);
(5.7)
for U and V suciently small neighborhoods of q and p satisfying f (U ) V . The local degree is an integer since Hn (Sn ; Sn ?) = Z for any point ? 2 Sn . The validity of local computation can be shown using basic tools:
Theorem 5.8.
Q = fqi gK 1 .
Assume that for
f : Sn ! Sn , p has discrete inverse image f 1 (p) = f=

K i =1
deg(f ; qi ):
Then
deg
(5.8)
(Sn ; Sn p) form a commutative diagram:
Proof. From Theorem 5.5, the long exact sequences of the pairs (Sn ; Sn Q) and
Hn (Sn Q)
/
Hn (Sn )

H (j ) / Hn (Sn ; Sn
Q)
/
Hn 1 (Sn Q) :
/
Hn (Sn p)
H (f )
/ Hn (Sn ) = / Hn (Sn ; Sn p ) H (j )
H (f )
H (f )
Hn 1 (Sn p)
H (f )
82
The rst terms in both rows and the last term in the bottom row vanish. By exactness, the lower map H (j ) is an isomorphism. Thus, by commutativity ( ) H(j ) H(f ) n n n n n deg f = deg Hn (S ) ! Hn (S ; S Q) ! Hn (S ; S p ) :
Choose a small neighborhood V of p so that f 1 (V ) = i Ui is a disjoint collection of neighborhoods of the qi . It follows that ( ) n n (1) Hn (S ; S Q) = Hn Ui ; Q Hn (Ui ; qi ); =(2) i i
where the isomorphisms come from (1) excision, and (2) behavior of homology under disjoint unions. The induced map H (f ) : Hn (Ui ; qi ) ! Hn (V; p ) is by denition multiplication by deg(f ; qi ). Thus, ( ) n deg f = deg Hn (S ) ! Hn (Ui ; qi ) ! Hn (V; p) = deg(f ; qi ): i i
5.6 Borsuk-Ulam theorems

There are a number of theorems which fall under the name Borsuk-Ulam; all concern spheres and the antipodal map a : Sn ! Sn that sends x 7! x . The following is the key step in these results, written in the language of degree theory.
Theorem 5.9 (Borsuk-Ulam).

degree; even maps of
Sn
Odd maps of
Sn
have odd
have even degree.
key to which is a long exact sequence to track antipodes. Assume that f is odd, so that f a = a f . Recall from Example 1.2, the quotient space Sn =a consisting of equivalence classes of antipodal points is one denition of the real projective space Pn . Let : Sn ! Pn denote the quotient projection map; this map is a 2-to-1 local homeomorphism. Let f : P2 ! P2 be the induced map. There is a commutative diagram of short exact sequences of chain complexes in
Proof. The proof in the (harder) odd case is sketched, the
coecients,
/0; /0
0 0
/ C (P n ) / C
/ C (Sn ) / C
/ C (Pn ) f / C (Pn )
(P n )
f

f
(Sn )
5.6. Borsuk-Ulam theorems
83
where is the transfer map, lifting a chain in Pn to a pair of chains in Sn . The F2 coecients ensures that the sequence is exact. The corresponding long exact sequences yields:
0 0
/ Hn (Pn ) H( ) /
Hn (Sn )
H ( ) / Hn (Pn )
/ Hn / Hn
(P n )
H (f )
/0; /0
H (f ) H (f ) H ( ) H ( ) / Hn (Pn ) / Hn (Sn ) / Hn (Pn ) H (f )

1
(P n )
By exactness and knowledge of H (Pn ), one argues that in the above diagram, and H ( ) are isomorphisms, while H ( ) = 0. By inducting on the dimension n and using the commutative square penultimate to the right, one shows that H (f ) is an isomorphism. By commutativity, this implies that H (f ) : Hn (Sn ; F2 ) ! Hn (Sn ; F2 ) is an isomorphism. This, in turn, is the mod 2 reduction of H (f ) : Hn (Sn ; Z) ! Hn (Sn ; Z); i.e., deg(f ) mod 2 = 1.
There are a number of famous corollaries of Theorem 5.9 whose proofs may be found in standard texts [?, ?]. 1. 2.
faces (simplices of @ n+1 ) whose images in Rn intersect. 3. Stone-Tukey: Given a collection of n Lebesgue-measurable bodies in Rn , there is a hyperplane which bisects evenly the volume of each body. n 4. Lusternik-Schnirelmann: Any cover U = fUi gn 0 of S by n + 1 open sets must have at least one element Uj containing an antipodal pair of points.
Borsuk-Ulam: Any map Sn ! Rn must identify some pair of antipodal points. Radon: Any map n+1 ! Rn of an n + 1-simplex has a pair of disjoint closed
Several of these theorems have physical interpretations. For example, it is common to express the rst corollary above as saying that (assuming meterological continuity) some antipodal pair of points on the earth have the same temperature and barometric pressure. Applications more relevant to this text lie in economics and fair division problems: see, e.g., [?] for applications of Stone-Tukey to fair division. The recent book of Matouek [?] contains a wealth of applications to combinatorics.
84
5.7 Euler characteristic

Sequences give an easy justication of the topological invariance of the Euler characteristic. This comes from lifting the notion of Euler characteristic from a cell complex to an arbitrary (nite) chain complex C = (C ; @ ): (C) = ( 1)k dim Ck : (5.9) k The reason for the alternating sum is to take advantage of cancelations that permit the following.
Lemma 5.10.
The Euler characteristic of a chain complex and its homology are
identical, when both are dened.
and @ 2 = 0, one has dim Zk = dim from which it follows that
Proof. The proof for eld coecients uses simple linear algebra. Since Hk = Zk =Bk
Hk + dim Bk
and dim Ck = dim Zk + dim
Bk
dim Ck = dim
Hk + dim Bk + dim Bk 1 : k
telescopes.
Multiply this equation by ( 1)k ; the sum over
Corollary 5.11.
For a nite compact cell complex
(X ) =
X with subcomplexes A and B,

(5.10) (5.11) (5.12)
(A [ B) = (A) + (B) (A \ B) (X A) = (X ) (A)

It follows that
( 1)k dim Hk (X )
is a homotopy invariant among this class of spaces.
These results follow from applications of Lemma 5.10 to (1) the chain complex for cellular homology; (2) the Mayer-Vietoris sequence; and (3) the long exact sequence of the pair (X; A) respectively, the last requiring a little excisive eort to relate C (X; A) to X A.
5.8 Lefschetz index

There is a generalization of Euler characteristic from spaces to self-maps. For any chain map ' : C ! C on a nite-dimensional chain complex C, dene the Lefschetz index as the alternating sum of the traces on the induced homomorphism, (' ) = k ( 1)k trace (H(') : Hk (C) ! Hk (C)). This index, like the Euler characteristic which it mimics, is intimately connected to the question of xed points, not
5.9. Equivalence of homology theories
85
of vector elds, but of self-maps in general. For a self-map f : X ! X of a nite cell complex X , dene its Lefschetz index as that of f , or, equivalently, H (f ): (f ) = (f ) = ( 1)k trace (H(f ) : Hk X ! Hk X ) : (5.13) k
Theorem 5.12.
point if
(f ) 6= 0.
For
a nite cell complex, any map
f :X!X
must have a xed
Proof. The technical portion of the proof (omitted) is to show that f may be approximated by a map (also labeled f ) that respects a cell structure on X (maps the n-skeleton X (n) to itself for all n) without modifying the xed point set. Assuming this, if f is xed-point-free, then the cellular approximation takes no cell to itself. Thus, the trace of the chain map f : C ! C vanishes. The analogue of Lemma 5.10 holds: the alternating sum of traces of H (f ) equals the alternating sum of traces of f via the same telescoping sum argument. Thus, (f ) = 0.
points that is a small perturbation of the identity map. The Lefschetz index of this map is (Id) = (M ), since the identity map on H has trace equal to the dimension of H . Indeed, it follows that for any nite cell complex X with 6= 0, any map f : X ! X homotopic to the identity has a xed point.
M with 6= 0 has at least one xed point. If a nonvanishing vector eld existed, the time- map of the ow (for suciently small) would be a map of M without xed
This provides a simple proof of Theorem 3.3, that any vector eld on a manifold
5.9 Equivalence of homology theories

The equivalence of the various homology theories discussed can be shown in a number of ways. One clean approach uses a diagrammatic lemma from homological algebra.
Lemma 5.13 (The Five Lemma).

of the form
Given a commutative diagram of abelian groups
=

=
/

/ /

/ /

=
/

(5.14)
=
vertical maps are iso-
whose top and bottom rows are exact, and whose four
outer
morphisms; the middle vertical map is an isomorphism as well.
Theorem 5.14.
On cell complexes, singular and cellular homology are isomorphic.
Proof. The standard proof inducts on k , the dimension of the skeleta of the cell complex X and uses the long exact sequence of the pair (X (k ) ; X (k 1) ) of skeleta in S ) and cellular (H C ) homologies. For a restricted class of regular cell comsingular (H plexes (attaching maps are homeomorphisms), the proof is slightly simpler. Choose
86
a cell-by-cell assembly sequence for X and induct on this sequence order. For X0 a point, the proof is trivial. Assume that e is a k -cell which when added to Xi yields Xi +1 . The cell e is glued to Xi along its boundary @e , homeomorphic to the sphere Sk 1 . The Mayer-Vietoris sequences for (Xi ; e ) in cellular and singular homology t together in a commutative diagram,
C @e Hn
S @e Hn / / H C Xi n
Ce Hn
/
CX Hn i +1
/ /
C @e Hn 1
/
C X HC e Hn n 1 1 i
/
=
=
S X HS e Hn i n
SX Hn i +1
/
S @e Hn 1
=

=
S X HS e Hn n 1 1 i
with vertical arrows induced from the map converting a cellular chain to a singular chain. By induction and previous computations of the homologies of balls and spheres, four of the ve vertical maps are isomorphisms. The 5-Lemma completes the proof.
5.10 Coverage in sensor networks

Sensors devices which return data tied to a location are ubiquitous. Although many sensors commonly observed are stand-alone or global devices, there is an increasing push to network multiple local sensors, thanks to progress in miniaturization and wireless communications. The problem of collating distributed pieces of sensor data over a communications network is a substantial engineering challenge for which the tools of topology and homological algebra seem strangely tting. One simple-to-state problem is that of coverage. Fix a domain D R2 and consider a nite collection Q of sensors in a the plane R2 . The nodes have two functions: they (1) sense a neighborhood of their locale in D; and (2) communicate with other sensors. Both of these actions are assumed to be local in the sense that individual nodes cannot extract sensing data from or communicate data over all of D. The problem of coverage, or more precisely, blanket coverage, is the question of whether there are holes in the sensor network are there any regions in D which are not sensed? The reader is likely to have encountered failure of blanket coverage generated by a network of cell phone towers: Coverage lost implies a hole. Other important coverage problems include barrier coverage, in which one wants to determine whether a sensor network separates D or surrounds a critical region, and sweeping coverage, the time-dependent problem familiar to users of robotic vacuum sweepers. The reader of this text will no doubt be led to thinking of coverage as a homological problem, with the ech theory as a particularly good source for tools. However, as individual sensors have specic ranges and geometrically-dened coverage domains, most coverage problems fall under the aegis of computation geometry [?]. As opti-
5.11. Homological coverage
87
mization problems where should one place the sensors for maximal coverage these are known as art gallery problems [?]; in the context of randomly-placed sensors, geometric probability becomes important [?]. Where a topological approach is protable is in the setting where location coordinates and distributions are unknown. The following simple application of homology to blanket coverage makes use of several of the constructs of this chapter. Specic geometric assumptions are kept to a minimum, for clarity and ease of proofs: 1. Sensors are modeled as a collection of nodes Q R2 . 2. Each sensor is assumed to have a unique identication which it broadcasts; certain neighbors detect the transmission and establish a communication link. 3. Communication is symmetric and generates a communications graph G on Q. 4. Sensor coverage regions are correlated to communications: the convex hull of any subset of sensors S Q which pairwise communicate is contained in the union of coverage regions of S . 5. One xes a cycle C G whose image in R2 is a simple closed curve bounding a domain D R2 .
D is contained in the coverage region of the network. The
Under these assumptions, one wants to know whether
critical assumption is the fourth, connecting the communications and sensing of the network. It is satised by systems with radially-symmetric communications networks (or unit disc graphs) and radially symmetric sensing regions with the proper ratio between sensing and communication [?, ?], but asymmetric systems are entirely permissible in this framework. These assumptions are chosen to be weak enough to be applicable in realistic systems; however, some important considerations e.g., time-variability, node failure, false echoes, communications errors are not modeled.
5.11 Homological coverage

Looking at a picture of a sensor network communications graph, one is tempted (as in 2.4) to search for coverage holes by looking for a long 1-cycle in the graph. Without the use of coordinates, it is problematic how to know whether such a 1-cycle bounds a hole or not. The assumption that k -tuples of nodes in pairwise communication have a convex hull in R2 which is sensor-covered is meant to capture the correspondence between local sensing and local communication. As stated, it is explicitly simplicial, and motivates the use of the ag complex F of the communications graph G to model the topology of the sensed region. The following theorem gives a criterion for coverage based on homology.
Theorem 5.15 ([?]).
Given
Q R2 , G, F, C, and D as above, then all of D is covered
by the sensors if, equivalently:
88
1.
2. There exists
[C] = 0 2 H1 (F). [] 2 H2 (F; C) with [] = @ = C.
Proof. Equivalence of the two conditions comes from the long exact sequence of the
pair (F; C) induced by the inclusion : C ,! F,

/
H2 (F; C)

/
H1 (C)
H (i ) / H1 (F)
/
;
since ker H (i ) = im by exactness. Consider the simplicial projection map : F ! R2 which sends vertices of the abstract complex F to the corresponding node points of Q D and which sends a k -simplex of F to the (potentially singular) k -simplex given by the convex hull of the vertices implicated. By construction, takes the pair (F; C) to (R2 ; @ D), acting as a homeomorphism on the second terms of the pairs. This map induces the following diagram on long exact sequences of the pairs:

/
/ H2 (F; C)

/
H1 (C)
/
:
(5.15)
H2 (R2 ; @ D)
H()
/
H ( )
/
H1 (@ D)
Functoriality implies that the diagram is commutative: H ( ) = H ( ) . By assumption, @ = C 6= 0; hence, H ( ) [] = H ( )[@] 6= 0. Commutativity implies that H ( )[] 6= 0, and thus H ( )[] 6= 0. If the sensors do not cover some point p 2 D, then p does not lie in the image of ; thus, the map : F ! R2 is a 2 2 2 2 composition of maps F ! R2 p ,! R , where Rp = R p R . Diagram (5.15) thus becomes: H2 (F; C) / H1 (C) : (5.16) oo o o o H() H ( ) ooo wooo / H2 (R2 ; @ D) / H1 (@ D) H2 (R2 p ; @ D) However,
2 H2 (R2 p ; @ D) = 0, as the long exact sequence of the pair (Rp ; @ D) reveals:
H2 (R2 p)
/
H2 (R2 p ; @ D)

/
H1 (@ D)
= / H1 (R2 p)
/
(5.17)
1 The rst term in (5.17) is zero since R2 p is homotopic to S , which has vanishing H2 . The next terms are H1 (@ D) = H1 (R2 p ) = Z, and, moreover, the image of 2 H() : H1 (@ D) ! H1 (Rp ) = Z is the winding number (4.12) of @ D about p 2 R2 . Since p lies in the interior of D, the winding number is 1, and H () is an isomorphism. Exactness of (5.17) implies that H2 (R2 p ; @ D) = 0. Commutativity of diagram (5.16) completes the proof.
5.12. Persistent homology
89
1. The assumption on sensor coverage is conservative it species that certain regions are guaranteed to be covered while passing no information about lack of coverage elsewhere. As such, the homological coverage criterion cannot be if-and-only-if. It, like the assumptions on which it is built, is a conservative criterion. 2. It is common to use the assumption that nodes communicate if and only if they are within a certain radius. This leads to replacing the ag complex F of the communications graph with the Vietoris-Rips complex of Q. 3. When the homological coverage criterion fails, choosing a basis for H1 (F) which is sparse (in the sense of implicating few nodes and edges) gives information about where the coverage holes may reside. 4. A relative cycle 2 Z2 (F; C) with @ = C suces to cover D: only those nodes implicated in are required to be actively sensing/communicating. This allows one to conserve power or establish a sleep-wake cycle by homological means.
5.12 Persistent homology

The capstone example of this chapter is a short survey of the exciting work being done in data analysis using sequences and homologies. The motivation is that, for a parameterized family of spaces (e.g., Vietoris-Rips complexes) modeling a pointcloud data set, those homology classes which persist over a larger parameter range have greater statistical signicance. This branch of applied topology is advancing very rapidly; see [?] for initial works, [?] for a recent survey, and [?] for a comprehensive text on the subject. The work described here spans contributions of Carlsson, de Silva, Edelsbrunner, Zomorodian, and others. Consider a ltration of a space X by subspaces Xi a sequence of inclusions
X0 , ! X1 , ! , ! XN = X:
(5.18)
This is motivated by, e.g., a sequence of Vietoris-Rips or ech complexes of a set of data points with an increasing sequence of parameters (i )i . This topological picture is converted to a sequence of chain complexes,
C (X0 ) ! C (X1 ) ! ! C (XN ) = C (X );
90
then, after passing to homology in eld coecients, to a sequence of (graded) vector spaces with linear transformations between them:
H (X0 ) ! H (X1 ) ! ! H (XN ) = H (X ):

There is, of course, no need to restrict to the case of metric-based simplicial complexes. For the general case, one begins with a persistence complex: a sequence of chain complexes P = (Ci ), i 2 Z, together with chain maps x : Ci ! Ci +1 . (For notational simplicity, one does not index the chain maps x .) Note that each Ci = (C;i ; @ ) is a complex.
i !j (P) is i < j , the (i; j )-persistent homology of P, denoted H j i dened to be the image of the homomorphism H (x ) : H (Ci ) ! H (Cj ) induced by xj i.
Denition 5.16.
For
There is a good deal more algebraic structure in the interleaving of persistent homology groups, as explained in [?]. Fix a eld of coecients F and place a graded F[x ]-module structure on P with x , the shift map, as multiplication. That is, a unit monomial x n 2 F[x ] sends Ci to Ci +n via n applications of x . One assumes a nite-type condition that each Ci is nitely generated as a F[x ]-module and that the sequence stabilizes in i in both directions (in the case of a bi-innite sequence of chain complexes). As the ltering of P is via chain maps x , P is free as an F[x ]-module. The resulting homology H (P) retains the structure of an F[x ]-module, but, unlike the chain module, is not necessarily free, but rather a direct sum of free and cyclic submodules. The classication of F[x ]-modules follows from the Structure Theorem for Principal Ideal Domains, since the only graded ideals of F[x ] are of the form x n F[x ].
Theorem 5.17 (Structure Theorem [?]).

with eld
For a nite-type persistence module
coecients,
H (P) =
x t i F [x ]
x rj (F[x ]=(x sj F[x ])) ;

.
(5.19)
for some collection of integer constants
frj ; sj ; ti gi;j
The Structure Theorem has a natural interpretation. The free portions of Equation (5.19) are in bijective correspondence with those persistent homology generators which come into existence at parameter ti and which persist for all future parameter values. The cyclic elements correspond to those homology generators which appear at parameter rj and disappear at parameter rj + sj . At the chain level, the Structure Theorem provides a birth-death pairing of generators of P (excepting those that are eternal or co-eternal). The parameter intervals arising from the basis for H (P) in Equation (5.19) inspire a visual snapshot of H (P) in the form of a barcode. A barcode is a graphical representation of H (P) as a collection of horizontal line segments
5.13. The space of natural images
91
in a plane whose horizontal axis corresponds to the parameter and whose vertical axis represents an (arbitrary) ordering of homology generators.
Theorem 5.18 (Barcode Theorem [?]).

intervals in the barcode of
H (P) containing the parameter interval [i; j ]. dim H (Ci ) is equal to the number of intervals containing i .
A barcode is an infographic of a parameterized dim
dim
i !j (P) H
is equal to the number of In particular,
H .
5.13 The space of natural images

One recent example of discovering topological structure in a high-dimensional data set comes from natural images. A collection of 4167 digital photographs of random outdoor scenes was assembled in the late 1990s by van Hateren and van der Schaaf [?, ?]. Mumford, Lee, and Pederson [?] sampled this data by choosing at random 5000 three-pixel by three-pixel squares within each digital image and retaining the top 20% of these with respect to contrast. The full data set consists of roughly 8,000,000 vectors in R9 whose components represent grey-scale intensities. By normalizing with respect to mean intensity and high-contrast images (those away from the origin), and by utilizing a certain norm for contrast [?], the data set M lies on a topological seven-sphere S7 R8 . A cursory visualization reveals points distributed seemingly densely over the entire S7 , prompting judicious use of density ltrations. A codensity function is used in [?] as follows. Fix a positive integer k > 0. For any point x in the data set, dene k (x ) as the distance in Rn from x to k th nearest neighbor of x in the data set. For a xed value of k , k is a positive distribution over the point cloud which measures the radius of the ball needed to enclose k neighbors. Values of k are thus inversely related to the point cloud density. The larger a value of k used, the more averaging occurs among neighbors, blurring ner variations. Denote by M[k; T ] the subset of M in the upper T -percent of density as measured by k . This is a two-parameter subset of the point cloud which, for reasonable values of
92
and T , represents an appropriate core. The rst interesting persistent homology computation on this data set occurs at the level of H1 . Taking a density threshold of T = 25 at neighbor parameter k = 300, with 5000 points sampled at random from M[k; T ], computing the barcode for the rst homology H1 reveals a unique persistent generator. This indicates that the data set is diused about a primary circle in the 7-sphere. The structure of the barcode is robust with respect to the random sampling of the points in M[k; T ]. In practice, witness complexes (as in 2.3) provide small enough spaces for computations to be done quickly. A closer examination of the data corresponding to this primary circle reveals a pattern of 3-by-3 patches with one light region and one dark region separated by a linear transition. This curve between light and dark is linear and appears in a circular family parameterized by the angle of the transition line. As seen from the barcode, this generator is dominant at the threshold and codensity parameters chosen. An examination of the barcodes for the rst homology group H1 of the data set ltered by codensity parameter k = 15 and threshold T = 25 reveals a dierent persistent H1 . The reduction in k leads to less averaging and more localized density sensitivity. The barcode reveals that the persistent H1 of samples from M[k; T ] has dimension ve. This does not connote the presence of ve disjoint circles in the data set. Rather, by focusing on the generators and computing the barcode for H0 , it is observed that, besides the primary circle from the high-k H1 computation, there are two secondary circles which come into view at the lower density parameter. A close examination of
these three circles reveals that each intersects the primary circle twice, yet the two secondary circles are disjoint. As noted in [?], each secondary circle regulates images with three contrasting regions and interpolates between these states and the primary circle. The dierence between the two secondary circles lies in their bias for horizontal and vertical stratication respectively. The barcodes for the second persistent homology H2 are more volatile with respect to changes in density and thresholding. This is not surprising: the lowest order terms in any series expansion are always most easily perceived. However, there is indication of a persistent H2 generator (in F2 coecients) at certain settings of k and T . Combined with the basis of H1 generators, one obtains predictive insight to the structure of the space of high-contrast patches. At certain density thresholds,
5.14. Zigzag persistence
93
the H2 barcode suggests a two-dimensional completion of the low-k persistent H1 basis into a Klein bottle K 2 . The primary and secondary circles appear with the
appropriate intersection properties. A comparison of homology computations in F2 and F3 coecients resolves the ambiguity that H2 (K 2 ; F2 ) = H2 (T2 ; F2 ) and veries 2 2 that the persistent surface found is K and not T .
5.14 Zigzag persistence
The icon of persistence is the inclusion sequence, ! ! ! ! , where arrows connote inclusion of spaces or chains or the induced homomorphisms on homology. However, other non-monotone sequences are possible and relevant to data management [?]. For example, sequences of the form ! ! , arise in consistency tests for sampling point clouds as follows. Given a large set of nodes Q, as in 5.13, it may be infeasible to construct the full Vietoris-Rips complexes and compute persistent homology. In the literature, a small sampling of points is taken; perhaps random or according to a witness complex construction as in 2.3. To check for accuracy of the sampling, one could compare the resulting homologies; however, it is not the dimensions of the homology that matter, but the correspondence. Assume that several samples of the data all detect the presence of a single hole. Is it the same hole? Or are there many holes in the true data set, each sampling detecting a dierent one? Given an ordered sequence of small subsamples, build simplicial complexes Xi based on them. One means of performing comparisons is to build the sequence of
94 spaces and inclusions:
X1
X;1 [ X 2 cH HH vv HH v v HH v HH vv v v
X2
X3
X4
This sequence, like the sequence of inclusions in Equation (5.18), is linear. Were the analogue of Theorem 5.17 to exist, it would classify homology classes which persist over the entire sequence. Such homology classes would be consistent over samples. The recent results of Carlsson and de Silva [?, ?] use representation theory to classify indecomposable submodules of the homology of linear sequences of the form
either to the left or to the right. In classical persistence, all arrows go to the right. In consistency tests, the arrows alternate. In general, the resulting zigzag persistence and associated barcodes are extremely powerful. The moral of these latter sections and of this chapter is that the homology of a sequence is worth more than a sequence of homologies.
! ! ! ! ; where each is a vector space over F and each ! is a linear transformation going
Notes
1. The Snake Lemma and 5-Lemma are two of many wonderfully useful general results in homological algebra: for others, see [ ,
? ?, ?].
does not exactly make for colorful reading, but [ ], combined with the mantra that complexes are algebraic representations of spaces, is the best place to start. 2. This chapter on exact sequences is just the beginning of diagrammatic homological algebra. The next step is to build a
The subject of homological algebra
C;
with horizontal
and vertical
techniques lead quickly to a
spectral sequence,
@0
double complex,
a bi-graded 2-d array of chains
chain maps satisfying
@@ 0 + @ 0 @
= 0.
Such
a structure reminiscent of a book
whose pages are double complexes, outtted with homomorphisms that turn the pages. Such structures, though notationally intricate, are quite powerful. 3. The Radon Theorem is usually stated in terms of convex hulls of points in Euclidean space, and its proof is often by means of convex geometry tricks. The deeper meaning of the Radon Theorem is, like that of the Helly Theorem (see Chatper 6), topological in nature. The Radon Theorem (and all the Borsuk-Ulam type results) are greatly
generalizable, with the Colored Tverberg Theorem being one of the most general [ ]. It is to be suspected that such generalizations of the Borsuk-Ulam Theorem are useful perhaps to economics most readily. 4. One should not underestimate the utility of local degree computations as in 5.5; the ability to infer global features from a small number of local measurements is greatly desirable in scientic and technological settings. 5. The Cellular Approximation Theorem casually alluded to in the proof of Theorem 5.12 is deep and signicant: any map between CW complexes can be homotoped to a
cellular map relative to a subcomplex on which the map is already cellular [ ].
5.14. Zigzag persistence
95
6. The extension of the coverage criterion to higher-dimensional networks is possible, but less clean than the 2-d case presented here. boundary of the domain in question. The diculty resides in specifying the
For regions in the plane, it is not dicult to
image choosing a cycle or even establishing a fence of sensors. In a triangulated
Rn , one must specify
cycle: this seems awkward to the author, but some applications
may permit this condition, in which case a simple modication of the existing proofs suce. A persistent homology approach is given in [ ]. 7. Why was the ech complex of the sensor cover not used to obtain a coverage criterion? Determining the depths of overlaps of coverage sets is work-intensive; cf. the use of Vietoris-Rips complexes versus ech in point cloud data. 8. As stated, the criterion for computing homological coverage in sensor networks is centralized, in the sense that nodes must upload connectivity data to a central computer. More desirable is a decentralized or distributed computation, performable by
nodes communicating with neighbors.
homology have just recently emerged [ ,
? ?, ?].
Algorithms for decentralized computation of
9. The available perspectives on persistent homology for an author to choose from are daunting, and the treatment in this chapter is necessarily elementary. See the books by Edelsbrunner-Harer and Zomorodian and the book-in-progress by Blumberg, Carlsson, and Vejdemo-Johansson for proper details from several perspectives. 10. The computation of persistent homology is a well-motivated problem which is, fortunately, solved [ ].
However nearly any homology computation of a realistically large
and high-dimensional data set is nontrivial. To perform computations, it can be ecacious to pass to a witness complex. Random sampling, multiple trials, and comparisons justify this reduction [ ]. 11. Carlsson and de Silva use the representation theory of
quivers
prominently in classi-
cation of zig-zag persistence. From a classication theorem of Gabriel [ ], it is shown that not only linear sequences, but any Dynkin diagram has a nice zigzag persistence classication theorem. 12. Multi-dimensional persistence is a pertinent and dicult generalization. The paper of Carlsson and Zomorodian [ ] gives an excellent introduction to the diculties inherent in the subject.
Chapter 6
Cohomology
98
6.1 Duals
The rst form of duality one encounters in mathematics is combinatorial.# The number of ways to choose k items from n k is exactly the same as the number of ways to (not) choose n k items. This symmetry in counting manifests itself in numerous numerical miracles, from the geometric symmetry of Pascal's triangle, to the fact that (S2n+1 ) = 0. The manner in which duality presents itself in Topology is best discovered through the familiar constructs of linear algebra and calculus. The dual space of a real vector space V is V _ , the vector space of all linear functionals V ! R. The dual space satises dim(V _ ) = dim(V ), and (V _ )_ = V for V nite-dimensional. There is a corresponding notion of duality for linear transformations. If f : V ! W is linear, the dual map or adjoint of f is f _ : W _ ! V _ given by (f _ ( ))(v ) = (f (v )). Note how the dual transformation reverses the direction of the map: it is the archetypical construct of this chapter.
Chapter 6. Cohomology ohomology mirrors homology with all the arrows reversed. The duality implicit in this theory is a deep and meaningful construct whose utility is subtle and easily underestimated.
Example 6.1 (Gradients) Dual vector spaces play an important role in calculus on manifolds. The cotangent space to a manifold M at p 2 M is the vector space dual
The cotangent spaces, like their tangent space duals, t together to form a bundle of vector space over M , the cotangent bundle T M . The analogue of a vector eld is a 1-form: a choice of Tp M continuous in p. For example, given a real-valued function f : M ! R, the gradient of f is the 1-form df which, in local coordinates fxi gn 1 evaluates to
Tp M = (Tp M )_ to the tangent space.
ordinates used to express it. It is a common mistake to conate the gradient of a function with a vector eld rf : this is permissible in Euclidean space, but the pairing is not well-dened in general. It is best to imagine a gradient df not as a vector eld, but as a ruler eld a eld of rulers with direction, orientation, and scale along which tangent vectors are measured.
# That
@f @x dxi ; i =1 i where dxi is the dual to the xi unit tangent vector. The Chain Rule implies that df is independent of the local codf =
grammatical dualities cannot fail to precede even these is not a false statement.
6.2. Cochain complexes
99
6.2 Cochain complexes
A cochain complex is a sequence C = (C ; d ) of R-modules C k and module homomorphisms d k : C k ! C k +1 with the property that d k +1 d k = 0 for all k . The cohomology of a cochain complex is,
Hk (C) = ker d k
im d k 1 :
Cohomology classes are equivalence classes of cocycles in ker d . Two cocylces are cohomologous if they dier by a coboundary in im d . The simplest means of constructing cochain complexes is to dualize a chain complex (C ; @ ). _ , the module of Given such a complex (with coecients in, say, R), dene C k = Ck homomorphisms Ck ! R. The coboundary d is the dual of the boundary @ , so that
d d = @ _ @ _ = (@ @ )_ = 0_ = 0:
The coboundary operator d can be presented explicitly: (df )( ) = f (@ ). For k -cell, d implicates the cofaces those (k + 1)-cells having as a face.
Example 6.2 (Simplicial cochains) Examples in the simplicial category are illustrative. Consider a triangulated disc with a 1-cocycle on edges using F2 coecients. This is the coboundary of a 0cochain which labels vertices on the left with zero and on the right with one. This cocyle is thus the zero cohomology class. On the other hand, if one considers a surface with some nontrivial H1 say, an annulus then one can construct a similar 1-cocycle which is nonvanishing in H 1 . The astute reader will notice the implicit relationship between such cocycles and gradients of a local potential over the vertices, with cohomology class in H 1 dierentiating between those which are or are not globally expressible as a gradient of a potential.
C with R-coecients, freely generated by (oriented) simplices f g in a simplicial complex X . The dual basis cochains can be thought of as characteristic functions f g. The natural pairing between C and its dual C_in these bases permits an integral interpretation: for basis elements and , dene to be the evaluation taking the value 1 if and only if = and 0 else. This notation is illustrative. By denition of the coboundary d as the dual of the boundary @ , one has (d )( ) = (@ ). Using linearity and the integral notation,
Example 6.3 (Integration) Consider a chain complex
100 one has for all cochains
Chapter 6. Cohomology
2 C p and chains c 2 Cp+1 ,
d =
@c
(6.1)
This reformulation should motivate scientists and engineers who know from early education the utility of Stokes' Theorem.
Example 6.4 (Kirchho's voltage rule) Consider an electric circuit as a 1-dimensional

cell complex, with circuit elements (resistors, capacitors, etc.) located in the interiors of edges. Kirchho's voltage rule states that the sum of the voltage potential differences across any loop in the circuit is zero. In the language of this chapter: voltage is a 1-coboundary.
Example 6.5 (Cuts and ows) Consider a directed graph G: a 1-d cell complex
with each edge oriented. A classical set of problems in combinatorial optimization concerns ows on G. Choose two nodes of G to be the sender (s) and receiver (r) nodes, and assume that the graph is connected from s ! r (respecting direction). A
for each node except s and r, the sum of the in-pointing edge ow values equals the sum of the out-pointing edge ow values: this net sum is constant over G. Motivated by problems in transportation and railway shipping, the classical max ow problem seeks to maximize this sum for a choice of ow. The problem is constrained in that each edge is assigned a capacity that dominates the possible ow value on that edge. The min-cut-max-ow theorem states that the maximum possible ow from s to r equals the minimal cut capacity, where, equivalently: 1. A cut is a partition of nodes of G into two sets, one containing s and the other r. The cut capacity is the sum of edge capacities over all edges which point from the s node-partition to the r node-partition. 2. A cut is a subset of edges of G whose removal disconnects s and r (there are no longer directed paths from sender to receiver). The cut capacity is the sum of the edge capacities of the removed edges. 3. A cut is an open subset of G containing r but not s. The cut capacity is the sum over the boundary of this set of all incoming edge capacities. In the language of this chapter, given G, s, and r, cuts and ows correspond to various chains and cochains. A ow is a choice of relative 1-cocycle in Z 1 (G; s [ r; Z), where
ow on G is an assignment of coecients (in, say, N or R+ ) to edges of G so that,
6.3. Cohomology
101
the sign of the Z-coecients is consistent with the edge orientations. Since the graph is nite, the ow may also be viewed as a well-dened relative 1-cycle in H1 (G; s [ r; Z). A cut is a 1-coboundary in B 1 (G) via (1) above, choosing a 0-cochain constant on the sender-receiver partition: the coboundary of this 0-cochain has support on all edges between partition elements. Alternately, a cut can be viewed as a nontrivial relative cocycle in H 1 (G; s [ r; Z). The astute reader rightly suspects duality lurking in the min-cut-max-ow theorem, especially when viewing a ow in relative H1 and a cut in relative H 1 .
6.3 Cohomology
The denition of cohomology in terms of dualizing chain complexes, though less than intuitive, is ecient. By dualizing results of the previous two chapters, one immediately obtains the following standard results: 1. Cohomology theories: cellular, singular, local, ech, relative, reduced; all with arbitrary coecients. 2. Functoriality and the homotopy invariance of cohomology. 3. Cohomological long exact sequences of pairs; the Mayer-Vietoris sequence; excision. It must be stressed that to correctly interpret and use these results arrows must be reversed. For example, the induced homomorphisms on cohomology twist composition: H(f g ) = H(g ) H(f ), in keeping with the way that duals of linear transformations behave. This manifests itself in, e.g., the long exact sequence of a pair (X; A) as follows. The short exact sequence of cochain complexes,
C (X; A)
j
/ C (X )
i / C (A)
/0;
becomes, via the Snake Lemma (Theorem 5.5), the long exact sequence,
/ H n 1 (A)
H (j ) / n H(i ) / n / n H (X; A) H (X ) H (A)

/
where, in this text, the induced and connecting homomorphisms have the same notation in homology and cohomology and are distinguishable via context and direction. The beginner may be deated at learning that cohomology seems to reveal no new data, at least as far as linear algebra can count.
Hn (X ; F)_ .
Theorem 6.6 (Universal Coecient).
For
a space and
a eld,
H n (X ; F ) =
The situation is more delicate in Z-coecients, but it is nevertheless true that cohomology is determined by knowing the homology and certain algebraic properties of the coecient ring. Students may wonder why they should bother with cohomology, as it alike to homology in every respect except intuition. In the beginning, one should repeat the mantra that duality often simplies algebraic diculties.
102
connectivity data. The proof in the case of cohomology is simpler than for homology, since, by denition, H 0 = ker d 0 . There is nothing to quotient out, due to arrow reversal. Elements of H 0 (X ) are locally-constant functions on X ; a suitable basis consists of characteristic functions of connected components of X . Note the dierences. In homology, H0 determines the number of path-connected components (homologous 0-cycles are connected by 1-chain paths) while cohomology H 0 measures connected components (as seen by functionals). These subtleties mirror the distinction between homology and cohomology. local and global coboundaries is the popular optical illusion of the impossible tribar. When one looks at the tribar, the drawn perspective is locally realizable one can construct a local depth function. However, a global depth function cannot be dened. The impossible tribar is a cartoon of a non-zero class in H 1 (properly speaking, H 1 (S1 ; R+ ), using projective geometry for coordinates [?]). An even more cartoonish example that evokes nontrivial 1-cocycles is the popular game of Rock, Paper, Scissors, for which there are local but not global ranking functions. A local gradient of rock-beats-scissors does not extend to a global gradient. Perhaps this is why customers are asked to conduct rankings (e.g., Netix movie rankings or Amazon book rankings) as a 0-cochain, and not as a 1-cochain (i.e., asking which-of-these-two-is1 better?): nontrivial H is, in this setting, undesirable. The Condorcet paradox that individually consistent comparative rankings can lead to global inconsistencies is a favorite topic in voting theory. Its best explanation cohomology is less popular.
Example 6.7 (Connectivity) Note that the dimension of H 0 , like that of H0 , yields
Example 6.8 (1-Cocycles) One cartoon for understanding this distinction between
6.4 Poincar duality

Homology and cohomology of manifolds have a dimensional symmetry as an expression of duality. Based on data from spheres, compact orientable surfaces, tori, and the Knneth theorem in 4.2, one might guess that for M a compact orientable n-manifold and coecients in a eld, dim Hk M = dim Hn k M , and likewise with cohomology. This is true and is a version of duality due to Poincar. At the level of cellular homology, this duality has a geometric interpretation. Consider a compact surface with a polyhedral cell structure, and let C be the cellular chain complex with F2 coecients. There is a dual polyhedral cell structure, yielding a chain complex C, where the dual cell structure places a vertex in the center of each original 2-cell, has 1-cells transverse to each original 1-cell, and, necessarily, has as its 2-cells neighborhoods of the original vertices. Each dual 2-cell is a polyhedral n-gon, where n is the degree of the original 0-cell dual. Note that these cell decompositions are truly dual and have the eect of reversing the dimensions of cells: k -cells generating Ck
6.4. Poincar duality
103
are in bijective correspondence with (2 k )-cells (on a surface) generating a modied k _ _ cellular chain group C 2 k . The dual complex C consisting of C := C k and d := @ entwines with C in a diagram:
0 0
/
/ C2
/ C1 /
/ C0 /
/0 /0
C0
=
d

=
d

=
C1
C2
The reader should check that the vertical maps are isomorphisms and, crucially, that the diagram is commutative. The equivalence of singular and cellular (co)homology, along with Theorem 6.7, implies that, for a compact surface with F2 coecients, Hk = _ H . Though this style of proof generalizes to higher-dimensional H2 k H = 2 k = 2 k manifolds, a better explanation for the symmetries present in (co)homology is more algebraic, and proceeds in a manner that adapts to non-compact manifolds as well. To unwrap this (and state a general symmetry result), a modied cohomology theory is helpful. Given a (singular, simplicial, cellular) cochain complex C on a space X , consider the subcomplex C c of cochains which are compactly supported: each cochain is zero outside some compact subset of X . (In the simplicial or cellular setting, this is equivalent to building cochains from a nite number of basis cochains.) The coboundary map restricts k ! C k +1 and d 2 = 0, yielding a well-dened to d : Cc c (X ). This cohocohomology with compact supports, Hc mology satises the following: 1. 2. 3. 4. omorphism) invariant. k (Rn ) = 0 for all k except k = n, in which cases it is of rank one. Hc (X ) Hc = H (X ) for X compact. Hc (X ) = H (X; X K ) for K a suciently large$ compact set.
is not a homotopy invariant, but is a proper-homotopy (and hence a homeHc
Compactly supported cohomology compactly expresses the duality implicit in manifolds:
Theorem 6.9 (Poincar duality).

phism
n k (M ; F ). PD : Hk (M ; F2 ) ! Hc 2
For
an
n-manifold, there is a natural isomor-
For M a compact n-manifold, Hk (M ; F2 ) = H n k (M ; F2 ). The coecients may be modied at the expense of worrying about orientability of the manifold and the torsional elements: see, e.g., [?, ?, ?] for details. With a slight change of perspective (in 6.9), a more precise form of PD will be given.
$ This is subtle: in brief, one orders compact sets by inclusion and uses induced maps to take a limit. See, e.g., [?, 3.3].
104
6.5 Alexander duality

Poincar duality can be adapted to several related settings involving manifolds. Among the most useful is Alexander duality.
Theorem 6.10 (Alexander duality).

is an isomorphism
For
A Sn A ).
a com-
pact locally contractible (nonempty proper) subset, there
~k (A) ! H ~n k 1 (Sn AD : H
One proof of this result wields many of the tools of homology and cohomology. Choose a family of small open neighborhoods of A that continuously deform onto A, and denote by U a suciently small representative of that family. For k > 0,
n k (Sn A) Hk (Sn A) =(1) Hc =(2) H n k (Sn A; (Sn A) (Sn U )) = H n k (Sn A; U A) =(3) H n k (Sn ; U ) ~ n k 1 (U ) =(4) H ~ n k 1 (A ); =(5) H
where the nontrivial isomorphisms come from (1) Poincar duality, (2) properties of compactly supported cohomology, (3) excision, (4) the long exact sequence of the pair, and (5) denition of U . Slight modications are required for k = 0. The condition on A is a form of tameness and is crucial.
6.6 Numerical Euler integration

Alexander duality is a key step in developing a highly eective numerical method for performing integration with respect to Euler characteristic over a non-localized planar network. Recall from 3.7 that certain problems in data aggregation over a network are expressible as an integral with respect to d over a tame space X Rn . This assumption of integrating sensors over a continuous domain is highly unrealistic; however, expressing the true answer as an integral over a continuum provides a clue as to how to approximate the result over a discretely sampled domain. Given an integrand h 2 CF(Rn ) sampled over a discrete set, computational formulae such as Equation (7.4) suggest that the estimation of the Euler characteristics of the upper excursion sets is an eective approach. However, if the sampling occurs over a network with communication links, then it is potentially dicult to approximate those Euler characteristics. Taking the ag complex of the network can lead to the existence of
6.7. Convex geometry

fake holes
105
higher-dimensional spheres (cf. 2.2) that ruin an Euler characteristic approximation.
Theorem 6.11.
For
h : R2 ! N constructible and upper semi-continuous,

42
h d =
1
s =0
(0 fh > s g
0 fh s g + 1) ;
(6.2)
where
0 = dim H0 , the number of connected components of the set.
Proof. Let A be a compact nonempty subset of R2 . From the homological denition

of the Euler characteristic and compactness of A,
(A) =
1
s =0
( 1)s dim Hs (A);
where H denotes singular homology in R coecients. Since A R2 , Hs (A) = 0 for all s > 1, and it suces to compute (A) = dim H0 (A) dim H1 (A). By Alexander duality, ~0 (R2 A) = dim H0 (R2 A) 1; dim H1 (A) = dim H 0 (R2 A; A) = dim H Since h is upper semi-continuous, the set R2 A = fh s g, one has:
A = fh > s g
is compact. Noting that
h d =
1
s =0
fh > s g =
1
s =0
dim
H0 (fh > s g) (dim H0 (fh s g) 1) :
Corollary 6.12.
that
The degree of sampling required to ensure
exact approximation of
h d over a planar network G is h.
correctly samples the connectivity of all the upper
and lower excursion sets of
This formulation is extremely important to numerical implementation of this integration theory to planar sensor networks, which, in practice, are not dense enough to ll regions without any holes. Thanks to Alexander duality, these can be mitigated, so long as the connectivity of the excursion sets is properly sampled.
6.7 Convex geometry

The following theorem is a classic result in convex geometry and geometric combinatorics. Using (co)homology yields a transparent proof.
106
Theorem 6.13 (Helly's Theorem).

compact convex subsets of
Rn
Let
U =
such that every
(n + 1)-tuple
U
fU g
be a collection of
M > n+1
of distinct elements of
has a point in common. Then all elements of
have a point in common.
Proof. Induct on M , beginning at M = n + 2. Consider the nerve
N of U. It is a subcomplex of the (n + 1)-simplex which, by hypothesis, contains all faces. If the common intersection is empty, then N = @ n+1 ' Sn . As the cover U is by convex sets, it is a good cover (all nonempty intersections are acyclic) and, thanks to Theorem 2.3,
In other words, N ' Sn has the homology of a subset of Rn . However, it is impossible for a subset A Rn to have the homology type of Sn via Alexander duality. Note that the hypotheses for Theorem 6.11 are satised, and thus,
(U) H = H (N) = H ([ U ) :
~n Hn (A) =H
~ 1 (X ) = 0 for all X nonempty. Thus, N 6' Sn , and the cover U must where, note, H have a common intersection point. The induction step is a simple modication.
The reader may have seen proofs of Helly's Theorem based on convex geometry or functional analysis. One benet of the topological approach is that extensions to the non-compact and non-convex world are natural and easily discerned. The critical ingredient is that the resulting cover U is acyclic, in the sense that all sets and their ~ = 0. nonempty intersections have H
n 1 (Rn
A) = 0;
6.8 Forms and Calculus

In the setting of dierential manifolds, cohomology ows naturally from multivariable calculus [?, ?]. One begins with multilinear algebra. Given a R-vector space V , let (V ) denote the algebra of forms on V alternating multilinear maps from products of V to R. Given a basis fxi gn 1 for V , explicit generators for (V ) are given by the _ dual 1-forms dxi 2 V , where dxi : V ! R returns the xith coordinate of a vector in V . These 1-forms combine to yield p-forms which acts on p-tuples of elements of V . The product in the algebra (V ) is called the wedge product and denoted ^. It is alternating, meaning, in particular, that dxi ^ dxj = dxj ^ dxi for all i and j . Scalar multiplication in (V ) is over R and the sum is induced by that on V _ . The algebra (V ) is graded:
p=0 where p (V ) is the vector space of p -forms, with basis dxi1 ^ ^ dxip for 1 i1 < < ip n = dim V . A p-form takes as its argument an ordered p-tuple of
(V ) =
p (V );
6.8. Forms and Calculus
107
vectors in V and returns a real number in a manner that is multilinear and alternating, n cf. determinants. The uniqueness of the determinant implies that dim = 1; the p alternating property of ^ implies that = 0 for all p > n = dim V . These constructions pass to manifolds. Recall from 1.3 that for an n-manifold M , the tangent bundle T M is a collection of n-dimensional vector spaces, parameterized by points in M . A (smooth) vector eld V is a choice of elements of Tx M varying (smoothly) in x , or, more precisely, a section taking x 7! V (x ) 2 Tx M . In like manner, the vector spaces of p -forms, p (Tx M ), form a bundle a family parameterized by points of M of vector spaces, of which the cotangent bundle T M is the case p = 1. As with the gradient 1-forms of Example 6.1, a p -form eld (shortened to p -form in practice) is a section : M ! p (T M ) giving x 2 p (Tx M ) varying smoothly in x . The space of all such sections the p-forms on M is denoted
p =
p (M ). On all manifolds,
0 (M ) = C 1 (M ; R), since dim 0 = 1. Likewise,
p (M ) = 0 for all k > dim M . Algebraic operations on pass to operations on
operating pointwise. For example, the wedge product extends to ^ :
p
q !
p+q . In the passage from to
, form elds change from point-to-point: such changes are measured by a derivative. The appropriate dierential operator for forms is d :
p !
p+1 , dened implicitly via: 1. d is linear with respect to addition of forms and scalar multiplication; 2. d satises a Leibniz rule: d ( ^ ) = d ^ + d ^ ; 3. On 0-forms, d is the dierential d : f 7! df .
Example 6.14 (Vector calculus on R3 ) On Euclidean R3 , the 1-forms

1 are repre-
sented as = fx dx + fy dy + fz dz , where the coecient functions are smooth. Some of these forms can be written as dierentials = dg for some g 2
0 . The 2-forms
2 are represented as = gx dy ^ dz + gy dz ^ dx + gz dx ^ dy . Every 3-form on R3 is of the form h dx ^ dy ^ dz for some h. The dierential d is familiar to students of vector calculus. Recall from Equation (5.3) how grad, curl, and div tie together functions C = C 1 (R3 ) and vector elds X. There is a commutative diagram,
=
C

r / r / X X = =
d
/
1
/
2
r / : C =
d
/
3
(6.3)
On
0 (R3 ), d is the gradient operator, taking a function f not to its gradient vector eld, but to the more natural gradient 1-form df . On
2 (R3 ), d is the curl operator; on
2 (R3 ), d acts as the divergence operator. Here, the vertical arrows identify 0- and 3-forms with functions, and identify vector elds with 1- and 2-forms in the obvious
108 ways:
~ = Fx^ ^ ! Fx dx + Fy dy + Fz dz = F F i + Fy ^ j + Fz k ~ ~ = Fx^ ^ ! Fx dy ^ dz + Fy dz ^ dx + Fz dx ^ dy = F F i + Fy ^ j + Fz k ~
The reader should return to simplicial examples of cochains and coboundaries and be convinced that what is measured on the algebraic level is indeed a discrete analogue of gradients, curls, and divergences. tous in mathematical physics. Maxwell's equations admit a particularly simple interpretation. The calculus version of Maxwell's equations (on Euclidean R3 , in a vacuum) are as follows:
Example 6.15 (Maxwell's equations) The language of dierential forms is ubiqui-
~ ~ =0 rB c @t = r E ~ 1 @E ~ = 4; ~ 4 ~ rE c @t = r B c J ~, B ~, J ~, , and c are the electric eld, magnetic eld, current, charge density, where E ~ and speed of light, respectively. Using the Euclidean structure to convert a eld F into a 1-form F or a 2-form as in Example 6.15, one obtains: ~ ~ F d (cE ~ ^ dt + B ~) = 0 d (cB ~ ^ dt E ~ ) = 4J ~ ^ dt dx ^ dy ^ dz;
These equations can be made more compact still and extended to arbitrary geometric manifolds using some of the constructions in 6.11. The language of dierential forms is designed for integration: a p -form is perhaps best thought of as an object that can be integrated over a p -dimensional domain. Specically, given an oriented p -dimensional submanifold with corners, S , there is an integral operator S :
p ! R dened by evaluating the p -form pointwise on oriented p tuples of tangent vectors to S and integrating on coordinate charts using the standard Lebesgue integral. As is the case in the more familiar setting of line integrals in vector calculus, the Chain Rule implies an invariance of the integral with respect to local orientation-preserving coordinate representations. In Example 6.18, the integral of d over an oriented piecewise-smooth closed curve is, precisely, the denition of the winding number of about 0: cf. Equation (3.4). If one ignores the dextrous handwaiving details about induced orientations on boundaries, the fundamental theorem of vector calculus is clear:
~ 1 @B
6.9. De Rham cohomology
109
For
Theorem 6.16 (Stokes' Theorem).

dimensional submanifold with corners in
M,
p-form
and
an oriented
p + 1-
d =
@S
6.9 De Rham cohomology

The antisymmetry property of forms is reminiscent of the cancelations via judicious 0 choice of ( 1) s at the heart of all homology and cohomology. Thanks to this antisymmetry and the commutativity of mixed partial derivatives, d 2 = 0. This prompts the interpretation of
as a complex: the de Rham complex of a manifold M is the cochain complex dR C = (
; d ) of forms with coboundary the exterior derivative d . The de Rham cohomology of M , dR H (M ) is the cohomology of
. In this theory, it is traditional to denote the cocycles as closed forms and the coboundaries as exact forms. A de Rham cohomology class is an equivalence class of closed forms modulo
exact forms.
Example 6.17 (de Rham cohomology) The reduced de Rham cohomology of Rn is

trivial for all n, thanks to the Fundamental Theorem of Integral Calculus. On the punctured plane P = R2 f0g, the closed 1-form d = (x dy y dx )=(x 2 + y 2 ) is not exact. Despite being denoted d, there is no single-valued 0-form whose gradient 1-form is as above; thus, d denes a non-trivial cohomology class [d] 2 dR H 1 (P ).
scends to a product on cohomology. By dening [] ^ [ ] := [ ^ ], one notes that since and are closed, d ( ^ ) = d ^ + d ^ = 0; furthermore, if or is exact, then so is ^ . Thus, ^ : dR H q dR H p ! dR H p+q turns dR H into a ring. Since, locally, a basis Euclidean k -form measures oriented projected k -dimensional volumes, the wedge product inherits a volumetric interpretation. On the torus Tn with angular coordinates i , the 1-forms di , i = 1 : : : n generate the cohomology ring dR H (Tn ). For example, the generator for dR H n (Tn ) = R is the volume form d1 ^ ^ dn . It is no coincidence that the de Rham cohomology of Tn has the same dimension as in the singular theory.
Example 6.18 (Wedge) In de Rham cohomology, the wedge product for forms de-
Theorem 6.19 (de Rham).
For
M a manifold, dR H (M ) = H (M ; R).
A little more eort yields a calculus-based version of Poincar duality. Let

c (M ) denote the complex of compactly supported forms on M , complete with dierential d . This, of course, yields a well-dened cohomology dR H c with compact supports, . which, as an extension of Theorem 6.20, is isomorphic to Hc
110
For M an oriented manifold n, wedge and integration of forms yields isomorphisms
Theorem 6.20 (Poincar Duality, de Rham version).

of dimension
= p _ ^ : dR Hp (M ) ! dR Hn c (M )
= : Hp (M ; R) ! dR Hp (M )_
It is an instructive exercise to show that integration descends to homology and cohomology. For [] 2 dR H p , 2
p 1 , S a boundaryless (p 1)-dimensional submanifold and T a p -dimensional submanifold with boundary, then
S+@T
+ d =
@S
d +
@S
d =
via liberal use of Stokes' Theorem. Thus, only [S ] and [] matter. The isomorphisms of Theorem 6.21 eect the Poincar Duality isomorphism PD : Hp (M ; R) ! n p _ dR H c (M ) of 6.4.
6.10 Currents
On smooth manifolds, calculus provides the convenient language of forms for cohomology. Duals of forms provide an extremely exible interpolation between smooth and discrete homological structures on manifolds that allow one to talk about, inter alia, the homology class of a vector eld. This section touches on analytic tools based on geometric measure theory [?, ?, ?, ?]. To avoid the numerous technicalities involving regularity and rectiability [?], let the reader assume (via restriction to the o-minimal structure of globally subanalytic sets) sucient (piecewise) smoothness where needed. _ Fix M an oriented manifold of dimension n. Let
p (M ) = (
p c (M )) be the space of p -currents real-valued functionals on compactly supported p -forms. Currents have a homological nature. Given any p -current T 2
p , the boundary of T , @T 2
p 1 , is dened as the adjoint to the exterior derivative: @T () = T (d). A cycle is a current T with @T = 0. Clearly, @ 2 = 0, and there is a resulting complex (
(M ); @ ) with resulting de Rham homology dR H (M ). The analogue of Theorem 6.20 holds: dR H (M ) = H (M ; R). The chief advantage in using currents is their visualizability. An oriented p dimensional submanifold in M is a p -current, since one can integrate a p -form over it: its de Rham homology class coincides with its singular homology class. For example, any (piecewise-smooth) oriented knot or link is a 1-current, since one can integrate a 1-form over oriented curves. Upon xing a volume form on M , a piecewise-smooth vector eld V is likewise a 1-current, since any 1-form pairs with V pointwise as (V ), which may then be integrated over M (to a nite value, thanks to compact support of ). The 2-currents on a manifold with volume form can range in shape from oriented surfaces to pairs of vector elds to a pair of tangent vectors at a single point.
6.10. Currents
111
Example 6.21 (Volume preserving links) One beauty of the language of currents is
that it allows one to compare both knots and vector elds on a manifold. For example, it follows from a result of D. Sullivan [?, Prop. II.25] that any volume-preserving vector eld on an oriented manifold can be realized as a limit of 1-currents supported on a compact 1-dimensional submanifold: an oriented link. This implies that any volume-preserving ow on S3 is the limit (in the sense of 1-currents) of a sequence of ever-lengthening, ever-coiling links. This suggests a reformulation of knot/link theory in terms of volume-preserving vector elds on S3 . What is the appropriate vector eld analogue of the unlink?
Example 6.22 (Helicity) It has been known for a long time what is the appropriate asymptotic analogue of linking number for volume-preserving vector elds on S3 [?, ?, ?]. The construction is as follows: given any two points x; y 2 S3 , evolve them forward under the ow of the vector eld for times s and t respectively, until the owlines come close to their starting points (that this happens for almost-every x and y innitely often follows from the Poincar Recurrence Theorem for volume-preserving ows [?]). Close these curves with short paths and compute the linking number (welldened for almost all x and y ). The limit of this linking number, normalized by st , converges as s; t ! +1 to a function `k (x; y ), which, when integrated over S3 S3 with the conserved volume form, yields the asymptotic linking number of the ow,
`k (V ) =
53 53
`k (x; y )dx dy:
The techniques of forms and currents makes the computation of this seeminglyintractable quantity elementary. A volume-preserving vector eld V on S3 is closed and nullhomologous as a 1-current; this implies that the vector eld contracted into the volume form yields an exact 2-form (V; ; ) = d for some . The helicity of V is the integral of the wedge product of with its dierential: H(V ) := ^ d; (6.4) 53 One shows well-denedness with respect to choice of via Stokes' Theorem. Arnol'd [?] (following Moat [?] ( following Cal ugarean u [?])) showed that the helicity is the asymptotic linking number:
Theorem 6.23 (Helicity Theorem [?]).
H(V ) = `k (V )
As a corollary, the helicity is an invariant of V under the action of volumepreserving dieomorphisms of S3 , since linking numbers are unchanged by such. This is of great signicance in uid dynamics, since the velocity eld of an ideal uid evolves in time according to the Euler equations (Example ??), and the energy of the uid
112
(the integral of the norm of the velocity eld) is bounded below by helicity H is a topological measure of a uid's inability to relax [?, ?]. The proof of the Helicity Theorem in [?] uses currents on S3 S3 to capture linking behavior. concerning Euler characteristic and intrinsic volumes have a representation in terms of currents. The normal cycle of a tame set A Rn is a special (n 1) Rn current NA on the unit cotangent bundle T1 = Sn 1 n n R . For A R of positive codimension, the normal cycle is best visualized as having support on the set of points a `unit' distance from A. The conormal cycle of a tame set A Rn is a particular n-current CA 2
n (T Rn ) on the cotangent bundle. It is, for lack of a better explanation in this text, the cone over the normal cycle. Each of the intrinsic volumes k of 3.10, including Euler charac1 n teristic = 0 , can be dened as the integral of a canonical form (k 2
n c (T1 R ) n n or !k 2
c (T R )) over the appropriate cycle:
Example 6.24 (Normal and conormal cycles) Many of the constructs of Chapter 3
k (A) =
NA
k =
CA
!k :
(6.5)
The valuation property of the intrinsic volumes is expressed in terms of addition of currents: e.g., NA[B = NA + NB NA\B and likewise for C. This means, e.g., that one can see the dierence between the Dn Euler characteristic of a compact disc (Dn ) = 1 and of its interior ( ) = ( 1)n = n @Dn ) as being a reection. When subtracting the conormal cycle C@Dn from (D n D C , the support in Rn is the same, but the orientation in each axis is reversed. For n odd, this results in an orientation-reversal, reected in the sign change.
6.11 Laplacians
With the addition of a geometric structure, there is another manifestation of Poincar duality in cohomology for manifolds via partial dierential equations. Recall that for a vector space V of dimension n, the algebra of forms, (V ), displays a combinatorial (n ) duality: dim p (V ) = p = dim n p (V ). Fix a geometry on V in the form of an inner product h; i and choose an orthonormal basis fxi gn 1 . Fix also an orientation on V in the form of an equivalence class of orderings [xi ]n 1 of basis elements up to even permutations. Dene the Hodge star ? : p (V ) ! n p (V ) on basis elements as follows: ?dxi1 ^ dxi2 ^ ^ dxip := dxip+1 ^ ^ dxin 1 ^ dxin ;
6.11. Laplacians
113
where the ordering [xij ] respects orientation. Extend to all of (V ) via linearity. The Hodge star depends only on the inner product and the orientation, not on the basis itself. It satises a signed duality ?? = ( 1)p(n p) Id. If an oriented manifold M is Riemannian, meaning that there is a smoothlyvarying inner product g (; ) on tangent spaces of M , then the Hodge star extends to ? :
p !
n p . For example, in Euclidean R3 with the standard basis as per Example 6.15, ?F ~ = F ~ . Every Riemannian manifold has a well-dened volume form 2
n which, in local orthonormal coordinates, is dx1 ^ ^ dxn and which is given by = ?M . The Hodge star yields an inner product on each
p via integration:
h; i :=
^ ?:
With this geometry in place, one may dene a codierential :

p !
p 1 given by the adjoint: h; d i = h; i; more explicitly, := ( 1)p(n p) ?d?. The Laplacian is the operator :
!
given by:
:= (d + )2 = d + d:
Note that the Laplacian is degree zero, and for p = 0 is the familiar second-order dierential operator. The Laplacian carefully entwines analytic, geometric, and topological features.The harmonic forms are dened as H (M ) := ker , the kernel of the Laplacian.
p (M ) has an orthogonal decomposition:
Theorem 6.25 (Hodge Theorem).
For
M a compact oriented Riemannian manifold,
p = d
p 1 H p
p+1 :
Corollary 6.26.
For
M a compact oriented Riemannian manifold, H (M ) = dR H (M ).
? is an incarnation of Poincar duality. Let 2 Hp be a harmonic form. Theorem 6.26 implies that d = = 0. This implies that ? is also harmonic, since
(?) = ( 1)p(n p) (d?d? + ?d?d )? = (d?d + ?d ) = 0:
The trouble of working with geometry has the following payo: the Hodge star
One may therefore realize the Poincar dual as ? : H k (M ) ! H n k (M ); cf. Theorem 6.21. One of the benets of using dierential-topological constructs is the ability to import and export ideas between smooth and discrete frameworks. There is a simple simplicial analogue of the Hodge theorem which has the advantage of requiring no
114
forms, dierentiability, or manifold structures, but merely an implicit geometry. Consider a cell complex X with cellular cochain complex C = (C ; d ) in R coecients. Choose an inner product h; i so that indicator functions over the cells of X are orthogonal. The implicit choice is in the weight of each simplex h ; i = w . With this inner product structure, use the adjoint of d to dene the combinatorial Laplacian: = d + d . As in the smooth theory, the harmonic cochains are H := ker . The following combinatorial Hodge theorem is a simple exercise.
Example 6.27 (Distributed homology) The Laplacian is a local operator and, as such, is well-suited to distributed computation. The work of Tahbaz-Salehi and Jadbabaie [?] details the use of the simplicial Laplacian to distributed computation of the homological coverage criterion of 5.11. By Theorems 5.15 and ??, veried coverage in a sensor network in R2 follows from showing that ker = 0 on combinatorial 1forms of the ag complex F of the network. It is easy to see that the heat equation, d dt = , has 0 as an asymptotically stable solution if and only if ker = 0. Thus, by solving a heat equation with random initial conditions, one can safely (to the degree one trusts in random initial conditions) conclude coverage if the solution converges to zero. That this equation can be solved locally and in a distributed manner should come as no surprise to the reader who has spent time with the heat equation. The paper [?] details the use of gossip-style algorithms for computing ker in a distributed manner.
6.12 Circular coordinates in data sets

In 5.13 the problem of determining the topology of a point cloud was addressed by means of persistent homology. A cohomological approach becomes the appropriate tool for addressing a related problem of coordinatizing a point cloud. Assume for the present that Q Rn is a pointcloud whose topology is known or suspected to be sufciently circular so as to merit outtting circular coordinates : Q ! S1 in a manner that respects the underlying topology of the space X Rn (homotopic to S1 ) that Q is presumed to sample. Many of the existing algorithms for assigning circular coordinates to a point cloud [?] have some convexity assumptions: X must be either convex [?] or the isometric embedding of a convex set [?]. The solution of de Silva, Morozov, and VejdemoJohansson [?] is a slick application of algebraic topology that highlights the particular benets of cohomology and the role of coecients. The outline of their work [?] is as follows. 1. One begins with the following result from homotopy theory: for any space X , the group of homotopy classes [X; S1 ] of maps X ! S1 is isomorphic to H 1 (X ; Z). The coordinatization function : X ! S1 therefore is naturally approachable via cohomology.
6.12. Circular coordinates in data sets
115
2. To nd a cohomology class for X based on a sampling of nodes Q, compute the persistent cohomology of a sequence of Vietoris-Rips complexes, as in 5.125.13. By the Structure Theorem (Theorem 5.17), this requires eld coecients; for numerical reasons (to avoid roundo errors), coecients in a nite eld Fq are preferred. 3. A (persistent) class [p ] 2 H 1 (X ; Fp ) can be converted to a integral class [] 2 H1 (X ; Z) by means of the following process. The short exact sequence of p coecients 0 ! Z ! Z ! Fp ! 0 yields a short exact sequence of chain complexes on X and, via Theorem 5.5, a long exact sequence on cohomology: / H 1 (X ; Fp ) / H 2 (X ; Z) p / H 2 (X ; Z) / / H 1 (X ; Z) The kernel of p : H 2 (X ; Z) ! H 2 (X ; Z) consists of p -torsional cohomology classes: for p > 2 these would seem to be rare occurrences in organic spaces X living behind data sets. By exactness, ker (p ) = coker ; when this is zero, it implies that H 1 (X ; Z) ! H 1 (X ; Fp ) is surjective, and (persistent) classes in Fp coecients therefore lift to integral classes. 4. The resulting integer cocycle is perhaps a poor S1 -coordinatization all the circular motion may be concentrated over a small subset of X . To relax to a smooth circular coordinate system, lift to R coecients and nd a cohomologous harmonic cocycle 2 H 1 (X ). Thanks to the local-averaging properties of the Laplacian, this 1-cocycle integrates to a well-regulated coordinate function : X ! S1 .
For details on computational aspects and implementation, see [?]. This work illustrates well the utility of cohomology, while highlighting the delicate interplay between real, integral, and cyclic coecients.
Notes
1. This chapter, above all in this text, is woefully incomplete: a short, motivational text such as this cannot do justice to cohomology theory. move on with all haste to, e.g., [ ], to learn the theory properly. 2. The conation of objects with duals is ubiquitous and insidious. Examples include
The interested reader should
confusing gradient 1-forms with vector elds and dening simplicial chains as functions from simplices to coecients. Let the reader resolve not to do so, even in the (barely permissible) nite-dimensional case. (The author hopes that sermons such as this will help expiate his own sins in this regard.)
3. The idea of the impossible tribar as a 1-cocycle was suggested by Penrose [ ] (and was noted to the author by Vin de Silva). 4. One of the many cohomology theories not covered in this text is related to conguration
X a topological space and 4 a ring, let C k be the set of functions (not k +1 necessarily maps!) f : X . This gives a complex C with dierential d taking !4 f (x0 ; : : : ; xk ) to the sum i ( 1)i f (x0 ; : : : ; xi 1 ; xi +1 ; : : : ; xk ). The set of all functions which vanish in a neighborhood of the grand diagonal (x; : : : ; x ) forms a subcomplex C0 . The Alexander-Spanier cohomology of X is H (C=C0 ). It is, for reasonable spaces, isomorphic to the singular H (X ).
spaces. Given
116
5. Helly's Theorem is important in a wide array of combinatorics and optimization problems: see [ ]. That is has a purely topological proof is a testament to the power of topological methods. The homological proof was known to Helly [ ] and many topological generalizations exist [ ]. It is remarkable how many experts nevertheless consign Helly's Theorem to convex geometry. 6. There is, as one might suspect, a deeper form of duality, of which Poincar, Alexander, and Lefshetz are emanations.
Verdier duality
+
for sheaves is perhaps the best
encapsulation of the scope and power of duality theorems in (co)homology. 7. The wedge product algebraic
cup product ^: H p H q
^ in de Rham cohomology is the smooth counterpart to the more ! Hp q , making singular cohomology H a ring
for any space, not just smooth manifolds. 8. Hodge theory is merely a hint at how partial dierential equations (in this case, Laplace's equation) on geometric manifolds can lead to topological invariants. Many subtler and deeper invariants come from other PDEs using auxiliary structures (gauge theory, Seiberg-Witten equations) and have implications in string theory and algebraic geometry. 9. A very clean proof of the de Rham Theorem (6.20) uses a double complex (see the notes to Chapter 5).
Exercises
6.1. Compute the simplicial cohomology of 6.2. Consider
: thus, c (X ) := c , the Euler characteristic dened using Hc
T2 , K 2 , and P2 in Z coecients by hand.

1
Show that, as in the denition of the o-minimal Euler characteristic of Chapter 3,
k k =0 dim Hc (X ; R).
c (Rn ) = ( 1)n . A more challenging exercise is to show that for any locally compact tame set X , c (X ) = (X ).
characteristic. 6.4. What is the analogue of Alexander duality for subsets of relationship between
6.3. Show that any compact boundaryless odd-dimensional manifold has vanishing Euler
H (A) and
H (Rn
A)?.
Rn ? Rn
Specically, what is the
6.5. Prove the Jordan Curve Theorem: any (tame) subset of divides
Rn
homeomorphic to
Sn
into exactly two connected components one bounded, one unbounded.
(It holds in non-tame cases, but requires more delicate hypotheses and analysis.) 6.6. Give a proof of boundary
@M , Hk (M ) = Hn
Lefshetz duality:
k (M; @M )
6.7. Consider a point charge at rest in
Rt , the time axis.

generator of
H k (M ) = Hn 3 R and let X be the

and
2 dR H (X ).
for
a compact manifold of dimension
k (M; @M ).
with
exterior of the point times
Assume that the magnetic eld vanishes and that the current and
electric elds are radial. Use Maxwell's equations to show that the Faraday form is exact but the Maxwell form generates
Thus, a point charge in
R3
is a
of its exterior.
6.8. Show that helicity is independent of the choice of 1-form 6.9. Show that the vector eld vector eld on
S3 :
=(
it is called the
Hopf eld.
x2 ; x1 ; x4 ; x3 ) is a well-dened volume-preserving
Compute its asymptotic linking number
and helicity. How much can you say about the orbits of this eld?

Eat

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Eat

Enviado por

Direitos autorais:

Formatos disponíveis

Elementary Applied Topology

Robert Ghrist University of Pennsylvania

3.6 3.7 3.8 3.9 3.10 3.11

The Fubini Theorem .

Euler integral transforms . Intrinsic volumes . . . . . . . .

Gaussian random elds . . . . . . . . . . .

Reduced homology . ech homology . . . . .

Relative homology Local homology .

Homology of a relation Functoriality . . . . . . .

Winding number and degree . . . . . . . . . . . . . . . .

The space of natural images . Zigzag persistence . . . . . . . . . . . . . . . . . . . . .

Sheaves & Other Functors

What topology can do

What topology cannot do

What this text is

1.2 Conguration spaces of linkages

Any smooth compact manifold is dieomorphic to the conguration

space of some planar linkage.

1.4 Vector elds

1.5. Braids and robot motion planning The

stable manifold of a xed point p is the set

W s (p) = fx 2 M : tlim  (x ) = pg: !1 t

Example 1.7 (Periodic orbits) If a continuous-time dynamical system  a ow 

1.5 Braids and robot motion planning

Example 1.8 (Braids) The conguration space of points on

Dfv (Tv V )  Dgw (Tw W ) = Tp M 8f (v ) = g (w ) = p:

T = dim U + dim coker T

Theorem 1.11 (Transversality Theorem).

f (v1 ; v2 ; v3 ) = det [v2 v1 ; v3 v1 ] :

1.7 Signals of opportunity

Theorem 1.14 (Whitney Embedding Theorem).

Any continuous function

between smooth manifolds is generically perturbed to a smooth injection when

Proof. The conguration space C2 M = M M

M of two distinct points on induces a map

(x; y ) 7! (x; y; f (x ); f (y )):

1.7. Signals of opportunity

1.8 Stratied spaces

1.8. Stratied spaces

2.1 Simplicial and cell complexes

standard k -simplex is the

22 balls B k , and for attaching maps, @Bk to X (k 1) .

Chapter 2. Complexes continuous functions from the boundary

2.2 Vietoris-Rips complexes and point clouds

2.3. Witness complexes and the sampling problem

2.3 Witness complexes and the sampling problem

2.4 Flag complexes and networks

2.5 ech and random samplings

2.6. Nerves and neurons nonintersecting.

Theorem 2.2 (Niyogi-Smale-Weinberger [?]).

has injectivity radius

that the minimal distance from any point of

2.6 Nerves and neurons

Theorem 2.3 (Nerve Lemma).

is a nite collection of subsets of

non-empty intersections of subcollections of

2.7 Phylogenetic trees and links

2.7. Phylogenetic trees and links

Theorem 2.5 (Billera-Holmes-Vogtmann [?], Vogtmann [?]).

The link of the 3 copies of

glued together at a single point.

2.8 Strategy complexes and planning with uncertainty

2.9. Decision tasks and consensus

Theorem 2.6 (Erdmann [?]).

Gaussian random elds . . . . . . . . . . .

1.2 Conguration spaces of linkages

Any smooth compact manifold is dieomorphic to the conguration

1.4 Vector elds

stable manifold of a xed point p is the set

W s (p) = fx 2 M : tlim (x ) = pg: !1 t

Example 1.7 (Periodic orbits) If a continuous-time dynamical system a ow

Example 1.8 (Braids) The conguration space of points on

Dfv (Tv V ) Dgw (Tw W ) = Tp M 8f (v ) = g (w ) = p:

Proof. The conguration space C2 M = M M

1.8 Stratied spaces

1.8. Stratied spaces

is a nite collection of subsets of

2.10. Discretized graph conguration spaces

2.10 Discretized graph conguration spaces

2.11. State complexes and reconguration

2.11 State complexes and reconguration

2.11. State complexes and reconguration

that a suciently dense sampling recovers the topology of

(Q) VR2 (Q) C 2 (Q): C

discretized conguration space 2.10. Show that 2.11.

is a cyclic graph with ve nodes and

3.3 Obstructions to vector elds

boundary possesses a nonvanishing vector eld if and only

I () on which points to the interior of .

with isolated xed points on a compact manifold

tame set is denably homeomorphic to a nite The

A and B denable sets,

(A [ B) = (A) + (B) (A \ B):

If h : X ! N where h is a counting U of uniform Euler characteristic (U ) = N 6= 0 for all