Zsuzsanna Dancso - On A Universal Finite Type Invariant of Knotted Trivalent Graphs

On a universal finite type invariant of knotted trivalent graphs
by
Zsuzsanna Dancso
A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Graduate Department of Mathematics University of Toronto
Copyright c 2011 by Zsuzsanna Dancso
Abstract
On a universal nite type invariant of knotted trivalent graphs Zsuzsanna Dancso Doctor of Philosophy Graduate Department of Mathematics University of Toronto 2011 Knot theory is not generally considered an algebraic subject, due to the fact that knots dont have much algebraic structure: there are a few operations dened on them (such as connected sum and cabling), but these dont nearly make the space of knots nitely generated. In this thesis, following an idea of Dror Bar-Natans, we develop an algebraic setting for knot theory by considering the larger, richer space of knotted trivalent graphs (KTGs), which includes knots and links. KTGs along with standard operations dened on them form a nitely generated algebraic structure, in which many topological knot properties are denable using simple formulas. Thus, a homomorphic invariant of KTGs provides an algebraic way to study knots. We present a construction for such an invariant. The starting point is extending the Kontsevich integral of knots to KTGs. This was rst done in a series of papers by Le, Murakami, Murakami and Ohtsuki in the late 90s using the theory of associators. We present an elementary construction building on Kontsevichs original denition, and discuss the homomorphicity properties of the resulting invariant, which turns out to be homomorphic with respect to almost all of the KTG operations except for one, called edge unzip. Unfortunately, edge unzip is crucial for nite generation, and we prove that in fact no universal nite type invariant of KTGs can intertwine all the standard operations at once. To x this, we present an alternative construction of the space of KTGs on which a homomorphic universal nite type invariant exists. This space retains ii
all the good properties of the original KTGs: it is nitely generated, includes knots, and is closely related to Drinfeld associators. The thesis is based on two articles, one published [Da] and one preprint [BD1], the second one joint with Dror Bar-Natan.
iii
Acknowledgements
I would like to express my deepest gratitude to my advisor, Dror Bar-Natan, from whom I learned how to do and enjoy mathematics, who spent large amounts of his time working with me, and seemed to have believed in me throughout. Like most graduate students probably do, I have hit a few walls in the past ve years: got stuck on problems, got discouraged, have lost focus or enthusiasm, and each time Dror tried idea after idea to get me mathematically, psychologically, technologically and generally un-stuck. He spent countless hours discussing mathematics with me, encouraged me to attack diculties I had given up on, suggested new directions when projects hit dead ends, told me that he thought I could make it in academia when I did not think I could, waited paitently when I was preoccupied with life outside mathematics, introduced me to other mathematicians, pushed me to give talks and taught me a lot about how to do it better, gave me many computer tutorials, funded my travels, and has overall been a spectacular advisor. As much as it still seems like a bit of a miracle, he may have succeeded, and I will never be able to thank him enough. I believe that choosing to work with Dror was one of the best choices I have ever made. Dror- thank you for everything. Without doubt I would never have made it here without the love and support of my husband Balazs, also a mathematician, and also one of my best choices ever made. I met Balazs shortly after starting my undergraduate studies in math, and he has been a limitless source of inspiration, encouragement and warmth ever since, even at times when I did not deserve it. Among many other gifts, he has given me the best piece of graduate school advice I have ever recieved: in my rst year in Toronto I was complaining to him about my doubts regarding academia and what an academic career entails, and whether I am good enough to even try. He told me that was nonsense, that graduate school is a fabulous time of ones life, a rare combination of relative freedom nanncially and in terms of responsibilities, an opportunity to make good friends and spend time and share experiences with them, a time to do interesting research and learn new things and iv
meanwhile have fun. I took this advice to heart, and this made for a (possibly unusually) smooth and happy graduate school experience, during which Balazs and I made many happy memories. Bali- thank you for being you and loving me. I am lucky and grateful to have the most wonderful family one could wish for: my parents, who have always encouraged me to strive for achievement; my grandmother, who will undoubtedly be the biggest fan of this thesis, albeit not the most expert one; and my brother and sister, who can always make me laugh. Anyu, Apu, Nyonyo, Marcsi, Tomi- thank you for the childhood where I am always happy to return. Last but not least, I would like to thank all of my fantastic friends, old and new, for always being there. Thank you for all you have done with me, for me, or because of meyou know who you are.
Contents
1 Introduction 2 Preliminaries 2.1 2.2 Finite type invariants- the classical view . . . . . . . . . . . . . . . . . . Finite type invariants- the algebraic view . . . . . . . . . . . . . . . . . . 2.2.1 2.2.2 2.3 Algebraic structures and expansions . . . . . . . . . . . . . . . . . Knotted trivalent graphs . . . . . . . . . . . . . . . . . . . . . . .
1 5 5 10 10 11 22 24 25 32 34 34 35 37 37 37 41 48 48
The denition of the Kontsevich Integral . . . . . . . . . . . . . . . . . . 2.3.1 2.3.2 2.3.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Universality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3 The Kontsevich integral of knotted trivalent graphs 3.1 The naive extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.1 3.1.2 3.2 The good properties . . . . . . . . . . . . . . . . . . . . . . . . . The problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Eliminating the divergence . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 3.2.2 The re-normalized integral Z1 . . . . . . . . . . . . . . . . . . . . The good properties . . . . . . . . . . . . . . . . . . . . . . . . .
3.3
Corrections - constructing a KTG invariant . . . . . . . . . . . . . . . . . 3.3.1 Missing moves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
3.3.2 3.3.3 3.3.4 3.4
Syzygies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Translating to equations . . . . . . . . . . . . . . . . . . . . . . . The resulting invariants . . . . . . . . . . . . . . . . . . . . . . .
49 50 56 61 65 65 68 71 75 75 89 90
Non-homomorphicity: unzip . . . . . . . . . . . . . . . . . . . . . . . . .
4 Fixing unzip: dotted KTGs 4.1 4.2 4.3 The space of dotted KTGs . . . . . . . . . . . . . . . . . . . . . . . . . . The associated graded space and homomorphic expansion . . . . . . . . . An equivalent construction . . . . . . . . . . . . . . . . . . . . . . . . . .
5 Applications 5.1 5.2 The relationship with Drinfeld associators . . . . . . . . . . . . . . . . . A note on the Kirby band-slide move and the LMO invariant . . . . . . .
Bibliography
vii
Chapter 1 Introduction
The goal of this thesis is to construct an algebraic setting for studying knot theory. In itself, knot theory is not very algebraic: the space of knots has few operations dened on it (most notably connected sum and cabling), and these are not nearly enough to make knots nitely generated. However, we can embed knots in a larger, richer space of knotted trivalent graphs (KTGs), which, along with standard operations dened on them, form a nitely generated algebraic structure. Our objective is to construct a homomorphic expansion (or universal nite type invariant) of KTGs: an invariant with a strong universality property which is very well-behaved under all KTG operations in the appropriate sense. This idea is due to Dror Bar-Natan, and was raised in [BN2]. One reason this approach is of interest is that several knot properties (such as knot genus, unknotting number and ribbon property, for example) are denable by short formulas involving knotted trivalent graphs and the aforementioned operations. Therefore, such an operation-respecting invariant yields algebraic necessary conditions for these properties, i.e. equations in the target space of the invariant. The extension of Z presented in Chapters 3 and 4 is the rst example for such an invariant. Unfortunately, the target space of Z is too complicated for it to be useful in a computational sense. However, we hope that by nding sucient quotients of the target space more computable 1
Chapter 1. Introduction invariants can be born.
The rst step of our construction is to extend the Kontsevich integral Z of knots to knotted trivalent graphs. The extension is a universal nite type invariant of knotted trivalent graphs, which is almost homomorphic: intertwines almost all standard operations dened on the space of graphs with their counterparts induced on the target space. These are changing the orientation of an edge; deleting an edge; unzipping an edge (an analogue of cabling of knots); and connected sum. The Kontsevich invariant Z was rst extended to knotted trivalent graphs by Murakami and Ohtsuki in [MO], and later by Cheptea and Le in [CL]. Both papers extend the combinatorial denition of Z, using q-tangles (a.k.a. parenthesized tangles) and building on a signicant body of knowledge about Drinfelds associators (even associators in [CL]) to prove that the extension is a well-dened invariant. In Chapter3 we try to extend Z naively, replacing knots by knotted trivalent graphs in the original denition. Unfortunately, the result is neither convergent nor an isotopy invariant. Thus, one needs to apply re-normalizations to make it convergent, and corrections to make it invariant. Using q-tangles, [MO] and [CL] do not have to deal with the convergence issue, while similar invariance issues arise in both approaches. The main purpose of this approach to extending the Kontsevich integral is to eliminate the black box quality of the previous constructions, which results partly from the depth of the ingredients that go into them, and partly from the fact that the proofs are spread over several papers ([MO, LMMO, LM] or [CL, MO]). Our construction diers from previous ones in that we build the corrected extension step by step on the naive one. After re-normalizations to make the extension convergent (a baby version of re-normalization used in quantum eld theory), non-invariance errors arise. We x some of these by introducing counter terms (corrections) that are precisely the inverses of the errors, and we show that thanks to some syzygies, i.e. dependencies between the errors, all the other errors get corrected automatically. This is where the
Chapter 1. Introduction
bulk of the diculty lies. The proofs involve mainly elementary Kontsevich integral methods and combinatorial considerations. The problem with the extended invariant, which we have alluded to before, is that it is only almost a homomorphism: it fails to commute with the unzip operation, which plays a crucial role in the nite generation of KTGs. Although its behavior with respect to unzip is well-understood, it is not homomorphic, and we show that this is not a shortcoming of this sepcic construction: any expansion (universal nite type invariant) of KTGs will display a similar anomaly, i.e., can not commute with all four operations at the same time. The goal of Chapter 4 is to x this problem by proposing a dierent denition of KTGs, which we will call dotted knotted trivalent graphs, or dKTGs. These form a nitely generated algebraic structure on which a truly homomorphic expansion exists. We present two (equivalent) constructions of this space. In one we replace the unzip, delete and connected sum operations by a more general set of operations called tree connected sums. In the other, we restrict the set of edges which we allow to be unzipped. We show that the invariant of Chapter 3 can easily be modied to produce a homomorphic expansion of dKTGs. In particular, this provides an algebraic description of the Kontsevich integral (of knots and graphs), due to the fact that dKTGs are nitely generated, i.e. theres a nite set (of size 4) of dKTGs such that any dKTG can be obtained from these using the above operations. This is described for KTGs in more detail in [Th], and is easily modied to dKTGs in Chapter 4. Sine our extension of the Kontsevich integral commutes with all operations, it is enough to compute it for the graphs in the generating set. As knots are special cases of knotted trivalent graphs, this also yields an algebraic description of the Kontsevich integral of knots. Finally, in Chapter 5 we describe two applications. The rst is the connection to Drinfeld associators, in particular, the construction of an associator as the value of the
Chapter 1. Introduction
invariant on a tetrahedron graph. The second is a simple (free of associators and local considerations) proof of the (recently falsely disputed) theorem that the LMO invariant is well behaved with respect to the Kirby II (band-slide) move.
Chapter 2 Preliminaries
2.1 Finite type invariants- the classical view
The main reference for this section is Chmutov and Duzhins expository paper on the Kontsevich integral, [CD]. The theory of nite type (or Vassiliev) invariants grew out of the idea of V. Vassiliev to extend knot invariants to the class of singular knots. By a singular knot we mean a knot with a nite number of simple (transverse) double points. The extension of a C-valued knot invariant f follows the rule f () = f () f (). A nite type (Vassiliev) invariant is a knot invariant whose extension vanishes on all knots with more than n double points, for some n N. The smallest such n is called the order, or type of the invariant. The set of all Vassiliev invariants forms a vector space V, which is ltered by the subspaces Vn , the Vassiliev invariants of order at most n:
V0 V1 V2 ... Vn ... V 5
Chapter 2. Preliminaries This ltration allows us to study the simpler associated graded space:
grV = V0 V1 /V0 V2 /V1 ... Vn /Vn1 ... The components Vn /Vn1 are best understood in terms of chord diagrams. A chord diagram of order n is an oriented circle with a set of n chords all of whose endpoints are distinct points of the circle (see the gure on the right). The actual shape of the chords and the exact position of endpoints are irrelevant, we are only interested in the pairing they dene on the 2n cyclically ordered points, up to orientation preserving dieomorphism. The chord diagram of a singular knot S 1 S 3 is the oriented circle S 1 with the pre-images of each double point connected by a chord. Let Cn be the vector space spanned by all chord diagrams of order n, and let Fn be the vector space of all C-valued linear functions on Cn . A Vassiliev invariant f Vn determines a function [f ] Fn dened by [f ](C) = f (K), where K is some singular knot whose chord diagram is C. The fact that [f ] is well dened (does not depend on the choice of K) can be seen as follows: If K1 and K2 are two singular knots with the same chord diagram, then they can be projected on the plane in such a way that their knot diagrams coincide except possibly at a nite number of crossings, where K1 may have an over-crossing and K2 an under-crossing or vice versa. But since f is an invariant of order n and K has n double points, a crossing ip does not change the value of f (since the dierence would equal to the value of f on an (n + 1)-singular knot, i.e. 0). The kernel of the map Vn Fn is, by denition, Vn1 . Thus, what we have dened is an inclusion in : Vn /Vn1 Fn . The image of this inclusion (i.e. the set of linear maps on chord diagrams that come from Vassiliev invariants) is described by two relations, the four-term, or 4T , and the framing independence, or F I relation, also known as the one-term relation.
Chapter 2. Preliminaries
The dotted arcs in the pictures that follow mean that there might be further chords attached to the circle, the positions of which are xed throughout the relation. The four-term relation asserts that
f(
) f(
) + f(
) f(
) = 0,
for an arbitrary xed position of (n 2) chords (not drawn here) and the two additional chords as shown. This follows from the following fact of singular knots:
1 1 0 11 11 00 1 0 f (00 ) + f ( 0 ) + f ( 00) + f ( 0 ) = 0, 1 0 1 11 00 11 00 11 1 0
which is easy to show using the isotopy below:
The framing independence relation arises from the Reidemeister 1 move of knot diagrams:
f(
) = 0,
To explain the name of this relation and because we will need this later, let us take a short detour to talk about framed knots. A framing on a curve is a smooth choice of a normal vector at each point of the curve, up to isotopy. This is equivalent to thickening the curve into a band, where the band is always orthogonal to the chosen normal vector. Note that due to the smooth choice of normal vectors, the band has an even number of twists on it, and thus forms an oriented surface. Observe that framed knots are a Z-extension of knots, as the number (with sign) of double twists determines the framing up to isotopy.
Chapter 2. Preliminaries A knot projection (knot diagram) denes a framing (blackboard framing), where we always choose the normal vector that is normal to (and
points up from) the plane that we project to. Every framed knot (up to isotopy) can be represented as knot projection with blackboard framing, as adding a double twist amouns to adding a kink to the diagram, as shown in the gure on the right. Of the Reidemeister moves, R2 and R3 do not change the blackboard framing of a knot diagram. R1, however, adds or eliminates a kink, and hence changes the framing. However, by dropping R1, we get something slightly larger than framed knots: R2 and R3 also leave the rotation number of the knot projection xed, while R1 does not. In fact, there are four types of R1 moves, depending on whether the kink is to the left or to the right of the strand and whether the crossing is positive or negative, as shown below. or R+ R or L+ or L
Left/right aects the rotation number and +/ aets the framing. Indeed, one can verify easily that a succession of L+ and R or L and R+ can be trivialized using only R2 and R3. Hence, knot projections modulo R2 and R3 form a Z2 - extension of knots, where framing accounts for one factor of Z and rotation number accounts for the other. To get rid of rotation number and obtain a characterization of framed knots, one needs R2, R3 and an additional relation which we will call R1 , which lets us eliminate an L+ followed by an L, as shown on the right. (This is eqvivalent, via R2 and R3, to eliminating an R+ followed by R.) The same can be said for links, with the dierence that both framing and rotation number adds a factor of Z for each link component. The framing independence relation we work with here arises from R1, meaning that the knot invariants we work with are independent of any framing on the knot. This will change later as we turn to (framed) knotted trivalent graphs. R1
We dene the algebra A, as a direct sum of the vector spaces An generated by chord diagrams of order n considered modulo the FI and 4T relations. By an abuse of notation, from now on we consider 4T and F I to be relations in Cn , i.e. dene the appropriate (sums of) chord diagrams to be zero. The multiplication on A is dened by the connected sum of chord diagrams, which is well dened thanks to the 4T relations (for details, see for example [BN1]). C-valued linear functions on An are called weight systems of order n. The above construction shows that every Vassiliev invariant denes a weight system of the same order. The famous theorem of Kontsevich, also known as the Fundamental Theorem of Finite Type Invariants, asserts that every weight system arises as the weight system of a nite type invariant. The proof relies on the construction of a universal nite type invariant, by which we mean a knot invariant which takes its values in the graded completion of A with the property that when evaluated on a singular knot, the lowest order term of the result is the corresponding chord diagram. Note that a universal nite type invariant is not a nite type invariant, but rather encompasses all the information contained in all nite type invariants. The (essentially unique) universal nite type invariant of knots is the Kontsevich Integral Z, the main ingredient of this thesis. To prove Kontsevichs Theorem, given any weight system, one gets the appropriate nite type invariant (up to invariants of lower type) by pre-composing it with Z. Our overall goal is to develop an algebraic context for the subject by embedding knots in the nitely generated space of (dotted) knotted trivalent graphs and extending the Kontsevich integral to obtain a universal nite invariant of this space, which also respects the algebraic structure (is homomorphic). Let us begin by introducing a general notion of algebraic structures and expansions, and re-introducing nite type invariants from this point of view.
10
2.2
Finite type invariants- the algebraic view
The goal of this section is to introduce the theory of nite type invariants by putting it in the general algebraic context of expansions. We rst dene general algebraic structures, projectivizations and expansions, followed by a short introduction to nite type invariants of knotted trivalent graphs (including the special case of knots and links) as an example: knotted trivalent graphs form a nitely presented algebraic structure which includes knots and links, the projectivization of this structure is the corresponding space of chord diagrams (a graded space), and the word expansion translates to universal nite type invariant. The presentation follows [BD1] and [BN5].
2.2.1
Algebraic structures and expansions
An algebraic structure O is some collection (O ) of sets of objects of dierent kinds, where the subscript denotes the kind of the objects in O , along with some collection of operations , where each is an arbitrary map with domain some product O1 Ok of sets of objects, and range a single set O0 (so operations may be unary or binary or multi-nary, but they always return a value of some xed kind). We also allow some named constants within some O s (or equivalently, allow some 0-nary operations). The operations may or may not be subject to axioms an axiom is an identity asserting that some composition of operations is equal to some other composition of operations. Any algebraic structure O has a projectivization. First extend O to allow formal linear combinations of objects of the same kind (over some eld, here we will work over Q), extending the operations in a linear or multi-linear manner. Then let I, the augmentation ideal, be the sub-structure made out of all such combinations in which the sum of coecients is 0. Let I m be the set of all outputs of algebraic expressions (that is, arbitrary compositions of the operations in O) that have at least m inputs in I (and
Chapter 2. Preliminaries possibly, further inputs in O), and nally, set proj O :=
m0
11
I m /I m+1 .
Clearly, with the operations inherited from O, the projectivization proj O is again algebraic structure with the same multi-graph of spaces and operations, but with new objects and with new operations that may or may not satisfy the axioms satised by the operations of O. The main new feature in proj O is that it is a graded structure, the degree m piece being I m /I m+1 . Given an algebraic structure O let l O denote the ltered structure of linear combinations of objects in O (respecting kinds), ltered by the powers (I m ) of the augmentation ideal I. Recall also that any graded space G =
nm m
Gm is automatically ltered, by
Gn
m=0
An expansion Z for O is a map Z : O proj O that preserves the kinds of objects and whose linear extension (also called Z) to l O respects the ltration of both sides, and for which (gr Z) : (gr O = proj O) (gr proj O = proj O) is the identity map of proj O. In practical terms, this is equivalent to saying that Z is a map O proj O whose restriction to I m vanishes in degrees less than m (in proj O) and whose degree m piece is the projection I m I m /I m+1 . A homomorphic expansion is an expansion which also commutes with all the algebraic operations dened on the algebraic structure O.
2.2.2
Knotted trivalent graphs
A trivalent graph is a graph which has three edges meeting at each vertex. We require that all edges be oriented and that vertices be equipped with a cyclic orientation, i.e. a cyclic ordering of the three edges meeting at the vertex. We allow multiple edges; loops (i.e., edges that begin and end at the same vertex); and circles (i.e. edges without a
Chapter 2. Preliminaries vertex). Given a trivalent graph , its thickening is obtained

3 3
12
from it by thickening verices as shown on the right, and gluing the resulting thick Ys in an orientation preserving
1 2 1 2
manner, where the cyclic orientation at the vertex denes the orientation of the surface. Hence, is a two-dimensional oriented1 surface with boundary.
A Knotted Trivalent Graph (KTG) is an isotopy class of embeddings of a thickened trivalent graph in R3 , as shown. This is equivalent to saying
that the edges of the graph are framed and the framings (normal vectors) agree at vertices: the embedding of the thickened graph has a dark side and a white side due to being oriented, and by denition, the framing is the normal vector to the surface along the middle of the ribbon on the dark side, so at vertices the three normal vectors agree. Conversely, such a framed embedding obviously determines an embedding of the thickening, up to isotopy. As a special case, framed knots and links are knotted trivalent graphs. The skeleton of a KTG is the combinatorial object (trivalent graph ) behind it. Isotopy classes of KTGs also have a characterisation in terms of graph projections (graph diagrams) and Reidemeister moves. Graph diagrams are projections onto a plane with only transverse double points preserving the over- and under-strand information at the crossings. Unframed KTGs correspond to graph diagrams modulo the Reidemeister moves R1 R2, R3 and R4 (see for example [MO]). R1, R2 and R3 are the same as in the knot/link case. R4 involves moving a strand in front of or behind a vertex:
R4a :
R4b :
Since we are considering framed KTGs, we need a charcterisation of these, which is

1
There is a slightly dierent non-oriented version which we do not describe in detail here.
Chapter 2. Preliminaries similar to that of framed knots:
13
Proposition 2.2.1. KTGs (according to our denition, framed) are in one-to-one correspondence with graph diagrams modulo Reidemeister moves R1 , R2, R3 and R4, where R1 eliminates adjacent positive and negative left kinks, as in the case of framed knots. Proof sketch. We understand a graph diagram to come with the blackboard framing. First we show that any isotopy class of KTGs can be represented by a blackboard framed graph diagram. Choose a plane to project to, and deform small neighborhoods of the vertices to at Ys parallel to the plane. Now each edge has a (not necessarily even) integer number of twists on it. The goal is to have even numbers on each edge, since a double twist can be represented by a kink in the blackboard framed projection. To achieve this, x a spanning tree of the graph and iron it: align it with the blackboard framing by moving all twists o the tree, starting from the root and progressing
leaves root
towards the leaves, using the isotopy pictured on the right.
pushing twists off the tree
The goal is for the tree to be at ribbon and for all the twists to lie on the curves connecting the leaves of the ironed spanning tree. Since the thickened graph is an oriented surface, each of these curves contain an even number of twists. These can now be represented on the projection as kinks. Next, we need to prove that isotopies of KTGs can be realised by Reidemeister moves, the other direction being obvious. Suppose two KTGs, represented by blackboard framed graph diagrams, are isotopic. Fix a spanning tree of the common skeleton. We leave it to the reader to convince him- or herself that one can use R2, R3 and R4 to bring the spanning trees into identical positions, move shrink kinks to a small size and move them o the trees. The moving of kinks, in particular, can be done by the move shown on the right, which is a combination of R2, R3 and R4 moves.
leaves moving kinks off the tree root
R2, R3, R4
14
Now one can again use R2, R3 and R4 to bring the arcs connecting the ends of the spanning trees into the same position, so the only dierence is the number of kinks on the arcs. Kinks move freely on the strands and commute with each other using R2, R3 and R4, and if the graphs are isotopic, then the framing, determined by the signed count of kinks on each arc, is the same. So one can use R1 to cancel all the kinks that can be cancelled, and the remaining essential ones can be brought into the same position. As an algebraic structure, KTGs have a dierent kind of objects for each skeleton. The sets of objects are the sets of knottings K() for each skeleton graph . There are four kinds of operations dened on KTGs: Given a trivalent graph , or a knotting K(), and an edge e of , we can switch the orientation of e. We denote the resulting graph by Se (). In other words, we have dened unary operations Se : K() K(Se ()). We can also delete the edge e, which means the two vertices at the ends of e also cease to exist to preserve the trivalence. To do this, it is required that the orientations of the two edges connecting to e at either end match. This operation is denoted by de : K() K(de ()). Unzipping the edge e (denoted by ue : K() K(ue ()), see gure below) means replacing it by two edges that are very close to each other. The two vertices at the ends of e will disappear. This can be imagined as cutting the band of e in half lengthwise. In the case of a trivalent graph , we consider its thickening and similarly cut the edge
e in half lengthwise. Again, the orientations have to match, i.e. the edges at the vertex where e begins have to both be incoming, while the edges at the vertex where e ends must both be outgoing.
e
ue ()
Given two graphs with selected edges (, e) and ( , f ), the connected sum of these graphs along the two chosen edges, denoted #e,f , is obtained by joining e and f by
15
a new edge. For this to be well-dened, we also need to specify the direction of the new edge, the cyclic orientations at each new vertex, and in the case of KTGs, the framing on the new edge. To compress notation, let us declare that the new edge be oriented from towards , have no twists, and, using the blackboard framing, be attached to the right side of e and f , as shown:
#e,f
As a short detour, we show how some topological knot properties are denable by KTG formulas. Theorem 2.2.2. 1. {Knots bounding a surface of genus k}= { : K(
... 1 2 k
)},
where , the boundary operator, is a certain xed composition of KTG operations. 2. {Knots of unknotting number k}= {xn () : K(
... 1 2 k
), dk () = O}, where
x, the crossing change operation is a given combination of KTG operations, dk refers to deleting the k middle edges, and O denotes the unknot. 3. {Ribbon knots}={uk1 () : K(
... 1 2 3 k
), dk1 () = OO...O for some
k}, where uk1 denotes unzipping the (k 1) connecting edges of the k-dumbbell graph, while dk1 refers to deleting these same edges, and OO...O denotes a trivial link of k components. Before proving the theorem, let us review the denition of ribbon knots. A knot is ribbon, if it bounds a singular disk in R3 such that
not ribbon ribbon
all the singualrities are transverse of ribbon type, as shown in the gure on the right. This is not the most standard denition (which is phrased in terms
16
of the preimages of the singularities, which form a 1-manifold in D2 ), but is eqvivalent to it. Proof (sketch). 1. Imagining the camel graph
... 1 2 k
as a band graph, i.e. an oriented sur...
face, the boundary operator turns it into its boundary:
The reader can check that the genus of the k-humped camel surface is k, which implies the statement. All that is left to show is that is a composition of KTG operations. Morally, is unzipping all edges, but of this cant really be done: there arent enough vertices for it. The solution is to plant a triangle at each vertex of the camel by taking connected sum with a tetrahedron, followed by two unzips, as shown:
u2
After this trick, it is possible to unzip all the old edges to get exactly the boundary we want. 2. The crossing change operation x corresponding to a given edge acts the following
x
way:
. We leave it as an exercise for the reader to verify that this
can be written as a composition of KTG operations and provides a description of the unknotting number. 3. The proof has two directions. We rst show that any uk () for as stated is ribbon. Deleting the connecting edges of the dumbbell graph produces a trivial link of k components, therefore each of the k circles bound disjoint embedded disks. Adding back the connecting edges, these now pass through the interiors of these disjoint disks transversely (after possibly a small perturbation). Unzipping these edges will
17
unite all the disks into one disk, with ribbon type singularities where the edges passed through the original disks. For the other direction, cut up the disk at each ribbon singularity:
If there are (k 1) ribbon singularities, we obtain k disjoint embedded disks. Connecting the two sides of each cut with edges gives the dumbbell graph we are looking for. As an algebraic structure, KT G is nitely generated2 (see [Th]), by two elements, the trivially embedded tetrahedron and the twisted tetrahedron, shown on the right (note that these only dier in framing). As described in the general context, we allow formal Q-linear combinations of KTGs and extend the operations linearly. The augmentation ideal I is generated by dierences of knotted trivalent graphs of the same skeleton. KT G is then ltered by powers of I, and the projectivization A := proj KT G also has a dierent kind of object for each skeleton , denoted A(). The classical way to lter the space of KTGs, which leads to the theory of nite type invariants, is by resolutions of singularities, as described above in the case of knots and links. An n-singular KTG is a trivalent graph immersed in R3 with n transverse double points. A resolution of such a singular KTG is obtained by replacing each double point by the dierence of an over-crossing and an under-crossing, which produces a linear combination of 2k KTGs. Resolutions of n-singular KTGs generate the n-th piece of the ltration.
2
In the appropriate sense it is also nitely presented, however we do not pursue this point here.
18
Theorem 2.2.3. [BD1] The ltration by powers of the augmentation ideal I coincides with the classical nite type ltration. Proof. Let us denote the n-th piece of the classical nite type ltration by Fn , and the augmentation ideal by I. First we prove that I = F1 . I is linearly generated by dierences, i.e., I = 1 2 , where 1 and 2 are KTGs of the same skeleton. F1 is linearly generated by resolutions of 1-singular KTGs, i.e. F1 = , where and dier in one crossing change. Thus, it is obvious that F1 I. The other direction, I F1 is true due to the fact that one can get to any knotting of a given trivalent graph (skeleton) from any other through a series of crossing changes. To prove that I n Fn , we use that I = F1 . (F1 )n is generated by formulas containing n 1-singular KTGs, possibly some further non-singular KTGs, joined by connected sums (the only binary operation), and possibly with some other operations (unzips, deletes, orientation switches) applied. The connected sum of a k-singular and an l-singular KTG is a (k + l)-singular KTG. It remains to check that orientation switch, delete and unzip do not decrease the number of double points. Switching the orientation of an edge with a double point only introduces a negative sign. Unzipping an edge with a double point on it produces a sum of two graphs with the same number of double points. Deleting an edge with a double point on it produces zero. Thus, an element in (F1 )n is n-singular, therefore contained in Fn . The last step is to show that Fn I n , i.e., that one can write any n-singular KTG as n 1-singular, and possibly some further non-singular KTGs with a series of operations applied to them. The proof is in the same vein as proving that KTGs are nitely generated [Th], as illustrated here on the example of a 2-singular knotted theta-graph, shown on the right. In the gures, a trivalent vertex denotes a vertex, while a 4-valent one is a double point. As shown in the gure below, we start by taking a singular twisted tetrahedron for each double point,
19
#4
slide
u3
u and slide
a (non-singular) twisted tetrahedron for each crossing, and a standard tetrahedron for each vertex, as shown in the gure below. We then apply a vertex connected sum (the composition of a connected sum and two unzips, as dened at the beginning of the proof of Theorem 5.1.1) along any tree connecting the tetrahedra, followed by sliding and unzipping edges, as shown below. The result is the desired KTG with an extra loop around it. Deleting the superuous loop concludes the proof. As in the classical theory of nite type invariants, A() is best understood in terms of chord diagrams. A chord diagram of order n on a skeleton graph is a combinatorial object consisting of a pairing of 2n points on the edges of , up to orientation preserving homeomorphisms of the edges. Such a structure is illustrated by drawing n chords between the paired points, as seen in the gure on the right. From the nite type point of view, a chord represents the dierence of an over-crossing and an under-crossing (i.e. a double point). Chord diagrams are factored out by two classes of relations, the 4T relations: + = 0,
and the Vertex Invariance relations (V I), (a.k.a. branching relation in [MO]):
(1)
+ (1)
+ (1)
= 0.
20
In both pictures, there may be other chords in the parts of the graph not shown, but they have to be the same throughout. In V I, the sign (1) is 1 if the orientation of the edge the chord is ending on is outgoing, and +1 if it is incoming. The 4T relation is proven the same way as we have seen in the case of knots, and the V I relation arises from a similar lassoe isotopy around a vertex. Although it is easy to see that these relations are present, showing that there are no more is dicult, and is best achieved by constructing an expansion (in nite type language, a universal nite type invariant) QKT G A, rst done in [MO] by extending the Kontsevich integral Z of knots, and later in [Da], which constitutes most of Chapter 3 of this thesis. The resulting invariant will be denoted by Z2 (as it is built through a two-step construction). The nite type theory of knots and links is included in the above as a special case. On knots, there is no rich enough algebraic structure for the nite type ltration to coincide with powers of the augmentation ideal with respect to some operations. However, knots and links form a subset of KTGs, and the restriction of I n to that subset reproduces the usual theory of nite type invariants of knots and links, and Z2 restricts to the Kontsevich integral. Now we turn to the question of whether Z2 is homomorphic with respect to the algebraic structure of KT G. To study this we rst have to know the operations induced on A by Se , de , ue and #e,f . Given a graph and an edge e, the induced orientation switch operation is a linear map Se : A() A(se ()) which multiplies a chord diagram D by (1)k where k is the number of chords in D ending on e. This is due to the fact that switching the orientation of an edge turns an under-crossing into an over-crossing and vice versa. Note that this generalizes the antipode map on Jacobi diagrams, which corresponds to the orientation reversal of knots (see [Oh], p.136). The induced edge delete is a linear map de : A() A(de ()), dened as follows:
21
when the edge e is deleted, all diagrams with a chord ending on e are mapped to zero (an over- and an under-crossing become the same when one of the participating edges is deleted), while those with no chords ending on e are left unchanged, except the edge e is removed. Edge delete is the generalization of the co-unit map of [Oh] (p.136), and [BN1]. The induced unzip is a linear map ue : A() A(ue ()). When e is unzipped, each chord that ends on it is replaced by a sum of two chords, one ending on each new edge (i.e., if k chords end on e, then ue sends this chord diagram to a sum of 2k chord diagrams). There is an operation on A(O) corresponding to the cabling of knots: references include [BN1] (splitting map) and [Oh] (co-multiplication). The graph unzip operation is the graph analogy of cabling, so the corresponding map is analogous as well. For graphs and , with edges e and e , the induced connected sum #e,e : A() A( ) A(#e,e ) acts in the obvious way, by performing the connected sum operation on the skeletons and not changing the chords in any way. This is well dened due to the 4T and V I relations. (What needs to be proven is that we can move a chord ending over the attaching point of the new edge; this is done in the same spirit as the proof of Lemma 3.1 in [BN1], using hooks; see also [MO], gure 4.) As it turns out (see [MO, Da]), Z2 is almost homomorphic: it intertwines the orientation switch, edge delete, and connected sum operations. However, Z2 does not commute with edge unzip. The behavior with respect to unzip is wellunderstood (showed in [Da] using a result of [MO]), and is described by the formula shown in the gure on the right. Z2 () Z2 (u()) 1/2 1/2 u( 1/2 )
Here, denotes the Kontsevich integral of the un-knot. A formula for was conjectured in [BGRT1] and proven in [BLT]. The new chord combinations appearing on the right commute with all the old chord endings by 4T. A dierent way to phrase this formula
22
is that Z2 intertwines the unzip operation ue : K() K(ue ()) with a renormalized chord diagram operation ue : A() A(ue ()), ue = i2 1/2 ue i 1/2 , where i 1/2 denotes the operation of placing a factor of 1/2 on e, ue is the chord-diagram unzip operation induced by the topological unzip, and i2 1/2 places factors of 1/2 on each daughter edge. So we have Z2 (ue ()) = ue Z2 (). This is an anomaly: if Z2 was honestly homomorphic, there should be no new chords appearing, i.e., Z2 should intertwine unzip and its induced chord diagram operation. Furthermore, as we prove in Chapter 3, the problem is not a shortcoming of only the Z2 constructed here: homomorphic expansions on KTG dont exist. Our main goal in Chapter 4 is to x this by changing the space of KTGs slightly, and constructing a homomorphic expansion on this new space, which we call dotted knotted trivalent graphs, or dKTGs. We also prove that the changes dont eect the good properties we like KTGs for.
2.3
The denition of the Kontsevich Integral
In this section we present the classical construction of the Kontsevich integral of knots, with proofs or at least proof sketches. The extension in Chapter 3 builds on this to a large extent, and uses the techniques explained here. The main reference we follow is [CD], further references on the subject include [BN1] and [Ko]. Let us represent R3 as a direct product of a complex plane C with coordinate z and a real line with coordinate t. We choose a Morse embedding for the oriented knot K: an embedding into C R such that the coordinate t is a Morse function on K (see the gure below). The Kontsevich integral of K is an element in the graded completion of A (throughout this section, A denotes the space of chord diagrams on an oriented circle), and is dened by the following formula:
t t4 2 1 t3 3 t2 4
23
DP
t1
z1
z1
Z(K) =
m=0 t
min <t1 <...<tm <tmax
P =(zi ,zi )
(1)P DP (2i)m
dzi dzi zi zi i=1
ti noncritical
In the formula, tmin and tmax are the minimum and maximum of the function t on K. The integration domain is the m-dimensional simplex tmin < t1 < ... < tm < tmax divided by the critical values into a number of connected components. The number of summands in the integral is constant in each of the connected components. In each plane {t = tj }, choose an unordered pair of distinct points (zj , tj ) and (zj , tj ) on K, so that zj (tj ) and zj (tj ) are continuous functions. P denotes the set of such pairs for each j. The integrand is the sum over all choices of P . For a pairing P , P denotes the number of points (zj , tj ) or (zj , tj ) in P where t decreases along the orientation of K. DP denotes the chord diagram obtained by joining each pair of points in P by a chord, as the gure above shows. Over each connected component, zj and zj are smooth functions. By mean the pullback of this form to the simplex. The term of the Kontsevich integral corresponding to m = 0 is, by convention, the only chord diagram with no chords and with coecient one, i.e, the unit of the algebra A. From now on, we will refer to this element as 1 A.
m dzi dzi i=1 zi zi
we
24
2.3.1
Convergence
We review the proof of the fact that each integral in the above formula is convergent. By looking at the denition, one observes that the only way the integral may not be nite is the (zi zi ) in the denominator getting arbitrarily small near the critical points (the boundaries of the connected components of the integration domain). This only happens near a local minimum or maximum in the knot - otherwise the minimum distance between strands is a lower bound for the denominator. Roughly speaking, if a chord ck is separated from the critical value by another long chord ck+1 ending closer to the critical value, as shown below, then the smallness in the denominator corresponding to chord ck will be canceled by the smallness of the integration domain for ck+1 , hence the integral converges:
zcrit
zk+1 zk
ck+1 ck
zk
zk+1
More precisely, the integral for the long chord can be estimated as follows (using the gures notation):
tcrit tk
dzk+1 dzk+1 C zk+1 zk+1
tcrit tk
d(zk+1 zk+1 ) =
C|(zcrit zk ) (zk+1 (tcrit ) zk+1 (tk ))| C |zk zk | For some constants C and C. So the integral for the long chord is as small as the denominator for the short chord, therefore the integral converges. Thus, the only way a divergence can occur is the case of an isolated chord, i.e. a chord near a critical point that is not separated from it by any other chord ending. But, by the one term relation, chord diagrams containing an isolated chord are declared to be zero, which makes the divergence of the corresponding integral a non-issue.
25
2.3.2
Invariance
Since horizontal planes cut the knot into tangles, we will use tangles and their properties to prove the invariance of the Kontsevich integral in the class of Morse knots. By a tangle we mean a 1-manifold embedded in [0, 1]3 , whose boundary is the union of k points on the bottom face of the cube, positioned at ( 1 1 2 1 k 1 , , 0), ( , , 0), ..., ( , , 0); k+1 2 k+1 2 k+1 2
and l points on the top face, positioned at ( 1 1 2 1 l 1 , , 1), ( , , 0), ..., ( , , 1). l+1 2 l+1 2 l+1 2
See example on the right. Two tangles are considered equal if there is an isotopy of the cubes that xes their boundary and takes one tangle to the other. Tangles can be multiplied by stacking one cube on top of another and rescaling, if the number of endpoints match. A tangle chord diagram is a tangle supplied with a set of horizontal chords considered up to a dieomorphism of the tangle that preserves the horizontal bration. Multiplication of tangles induces a multiplication of tangle chord diagrams in the obvious way. If T is a tangle, the space AT is a vector space generated by all chord diagrams on T , modulo the set of tangle one- and four-term relations: The tangle one-term (or framing independence) relation is the same as the framing independence relation for knots, i.e., asserts that a tangle chord diagram with an isolated chord is equal to zero in AT . For dening the tangle 4T relation, consider a tangle consisting of n parallel vertical strands. Denote by tij the chord diagram with a single horizontal chord connecting the i-th and j-th strands, multiplied by (1) , where stands for the number of endpoints of the chord lying on downward-oriented strands.
... ... ...
26
tij = (1)
The tangle 4T relation can be expressed as a commutator in terms of the tij s: [tij + tik , tjk ] = 0, and is illustrated in the gure below:
i j k i j k i j k
+ i j k
=0
One can check that by closing the vertical strands into a circle respecting their orientations, the tangle 4T relation carries over into the ordinary 4T relation of knots. We take the opportunity here to mention a useful lemma, which is a direct consequence of the tangle 4T relations. A slightly dierent version of this appears in [BN1], and a special case is stated in [MO]. Lemma 2.3.1. Locality. Let T be the trivial tangle consisting of n parallel vertical strands, and D be any tangle chord diagram on T such that no chords end on the j-th string. Let S be the sum
i=j
tij in AT . Then S commutes with D in AT .
The Kontsevich integral is dened for tangles the same way it is dened for knots (by placing a tangle in the picture instead of a knot). The advantage is that thangles can be multiplied, and by Fubinis theorem, Z is multiplicative: Z(T1 )Z(T2 ) = Z(T1 T2 ), whenever the product T1 T2 is dened. This implies the important fact that the Kontsevich integral of the vertical connected sum of knots is the connected sum of the Kontsevich integrals of the summands. By vertical connected sum we mean that the two knots are placed directly above one
27
another and connected by close parallel vertical lines. By the invariance results that will follow this generalizes to any connected sum, and is sometimes referred to as the factorization property, or multiplicativity, of Z. Proposition 2.3.2. The Kontsevich integral is invariant under horizontal deformations (deformations preserving the t coordinate, as shown in the gure below) of the knot, which leave the levels of the critical points xed.
Proof. Decompose the knot into a product of tangles without critical points, and other (thin) tangles containing one unique critical point. The following lemma, which could be considered the heart of the invariance of Z, addresses the case of tangles without critical points. The proposition then follows from the lemma by taking a limit. (See [CD] for more details.) Lemma 2.3.3. Let T0 be a tangle without critical points and T a horizontal deformation of T0 into T1 , such that T xes the top and the bottom of the tangle. Then Z(T0 ) = Z(T1 ). Proof. Let denote the dierential form in the m-th term of the Kontsevich Integral: =
P =(zi ,zi )
(1)#P DP (2i)m
dzi dzi zi zi i=1
Since there are no critical points, the integration domain for any is the entire msimplex = {tmin < t1 < ... < tm < tmax }. Consider the product of this simplex with the unit interval: = 0 I , and apply Stokes theorem: =

d. {f aces}.
The form is exact: d = 0. The boundary of the domain is = 0 1 + To prove the lemma it is enough to show that restricted to each face is zero.
28
This is the case for the faces dened by t1 = tmin or tm = Tmax , since this implies dz1 = dz1 = 0, (or dzm = dzm = 0), because z1 and z1 (or zm and zm ) do not depend on . For the faces where tk = tk+1 for some k, the endpoints of the k-th and (k + 1)-th chords might coincide, meaning we may not get a chord diagram at all. So, to dene the prolongation of and DP to such a face, we agree to place the k-th chord a little lower than the (k + 1)-th chord, in the case where some of their endpoints belong to the same string. The summands of belong to three subcases: 1) The k-th and (k + 1)-th chords connect the same two strings: we have zk = zk+1 and zk = zk+1 or vice versa, so d(zk zk ) d(zk+1 zk+1 ) = 0 and so the restriction of to the face is zero. 2) The endpoints of the k-th and (k + 1)-th chords belong to four dierent strings: it is easy to check that all choices of chords in this part of appear in mutually canceling pairs. 3) There are three dierent strings containing the endpoints of the k-th and (k + 1)-th chord: a slightly more involved computation shows that this part of is indeed zero.
We will need a modication of the proof of the next lemma for the KTG case, so we present this proof in full detail. Proposition 2.3.4. Moving critical points. Let T0 and T1 be two tangles that dier only in a thin needle (possibly twisted), as in the gure, such that each level {t = c} intersects the needle in at most two points and the distance between these is , a small positive number (it is enough that other parts of the knot do not wind through the needle). Then Z(T0 ) = Z(T1 ).
T0
T1
29
Proof. Z(T0 ) and Z(T1 ) can only dier in terms in which some chords end on the needle. If the chord closest to the end of the needle connects the two sides of the needle (isolated chord), then the corresponding diagram is zero by the F I (1T ) relation. So, we can assume that the one closest to the needles end is a long chord, suppose the endpoint belonging to the needle is (zk , tk ). Then, there is another choice for the k-th chord which touches the needle at the opposite point (zk , tk ), as the gure shows, and DP will be the same for these two choices.
zk
zk
zk
The corresponding two terms appear in Z(T1 ) with opposite signs due to (1)P , and the dierence of the integrals can be estimated as follows:
tc tk1
d(ln(zk zk )) = ln 1 +
tc tk1
d(ln(zk zk )) = ln
zk (tk1 ) zk (tk1 ) zk (tk1 ) zk (tk1 )
zk (tk1 ) zk (tk1 ) zk (tk1 ) zk (tk1 )
C|zk (tk1 ) zk (tk1 )| C,
where tc is the value of t at the tip of the needle, and C is a constant depending on the minimal distance of the needle to the rest of the knot. If the next, (k 1)-th chord is long, then the double integral corresponding to the k-th and (k 1)-th chords is at most:
tc tk2 tc tk1 tc tk2
d(ln(zk zk ))
tc tk1
d(ln(zk zk )) d(ln(zk1 zk1 )) zk1 (tc ) zk1 (tc ) zk1 (tk2 ) zk1 (tk2 )
d(ln(zk1 zk1 )) = C ln CC ,
where C is another constant depending on the ratio of the biggest and smallest horizontal distance from the needle to the rest of the knot.
30
If the (k 1)-th chord is short, i.e. it connects zk1 and zk1 that are both on the needle, then we can estimate the double integral corresponding to the k-th and (k 1)-th chords:
tc tk2 tc tk1 tc tk1 tc tk2 tc
d(ln(zk zk ))
d(ln(zk zk ))
dzk1 dzk1 zk1 zk1
(zk (tk1 ) zk (tk1 ))
dzk1 dzk1 = |zk1 zk1 |
=C
tk2
d(zk1 zk1 ) = C|zk1 (tk2 ) zk1 (tk2 )| C.
Continuing to go down the needle, we see that the dierence between Z(T0 ) and Z(T1 ) in degree n is proportional to (C )n , for a constant C = max{C, C }, and by horizontal deformations we can make tend to zero, therefore the dierence tends to zero, concluding the proof. So far we have proved the invariance of the Kontsevich integral in the class of Morse knots: to move critical points, one can form a sharp needle using horizontal deformations only, then shorten or lengthen the needles arbitrarily, then deform the knot as desired by horizontal deformations. However, Z is not invariant under straightening humps, i. e., deformations that change the number of critical points, as shown below. (We note that straightening the mirror image of the hump shown is equivalent to this one, see Section 3.3 for the details.)
To x this problem, we apply a correction factor, using the proposition below, which is a consequence of the lemma that follows it. We sketch the proofs briey here, see [CD] for more details.
31
Proposition 2.3.5. Let K and K be two knots diering only in a small hump in K that is straightened in K (as in the gure). Then Z(K ) = Z(K)Z( ).
Lemma 2.3.6. Faraway strands dont interact. Let K be a Morse knot with a distinguished tangle T , with tbot and ttop being the minimal and maximal values of t on T . Then, in the formula of the Kontsevich integral, for those components whose projection on the tj axis is contained in [tbot , ttop ], it is enough to consider pairings where either both points (zj , tj ) and (zj , tj ) belong to T , or neither do.
Proof. (Sketch) We can shrink the tangle T into a narrow box of width , and do the same for the rest of the knot between heights tbot and ttop . It is not hard to show that the value of the integral corresponding to long chords (connecting the tangle to the rest of the knot) then tends to zero. Proof of Proposition 2.3.5. Using the notation of Lemma 2.3.6, choose T to include just the hump, so by the assertion of the Lemma, there will be no long chords connecting the hump to the rest of the knot in Z(K ) or in Z( ). Also, in Z( ), there are no
chords above or below the hump, since the highest (resp. lowest) of those would be an isolated chord. Since the constant term of Z( ) is 1, it has a reciprocal in the graded completion of
A (i.e. formal innite series of chord diagrams). Using this we can now dene an honest
Chapter 2. Preliminaries knot invariant Z by setting Z (K) = Z(K) Z( )c/2 ,
32
where c is the number of critical points in the Morse embedding of K that we use to compute Z.
2.3.3
Universality
Here we state Kontsevichs theorem, and the main idea of the proof, which will apply in the case of the extension to graphs wihout any changes. A complete, detailed proof can be found in [CD] or [BN1], and Kontsevichs original paper [Ko]. Theorem 2.3.7. Let w be a weight system of order n. Then the a Vassiliev invariant of order n given by the following formula has weight system w: K w(Z (K)). Proof. (Sketch.) Let D be a chord diagram of order n, and KD a singular knot with chord diagram D. The theorem follows from the fact that Z (KD ) = D + {higher order terms}. Since the denominator of Z always begins with 1 (the unit of A), it is enough to prove that Z(KD ) = D + {higher order terms}. Because of the factorization property and the fact that faraway strands dont interact (Lemma 2.3.6), we can think locally. Around a single double point, we need to compute the dierence of Z on an over-crossing and an under-crossing. These can be deformed as follows: Z() Z() = Z Z .
33
Since the crossings on the bottom are now identical, by the factorization property, its enough to consider Z Z Z .
equals 1 (the unit of A(2 ), where A(2 ) stands for chord diagrams on two
upward oriented vertical strands), as both zi (t) and zi (t) are constant. In Z , the rst term is 1, as always, so this will cancel out in the dierence.
The next term is the chord diagram with one single chord, and this has coecient
1 2i tmax dzdz tmin zz
= 1, by Cauchys theorem.
So the lowest degree term of the dierence is a single chord with coecient one. Now putting KD together, the lowest degree term in Z(KD ) will be a chord diagram that has a single chord for each double point, which is exactly D.
Note that the crucial property of Z that Z (KD ) = D + {higher order terms} is the universality of Z , i.e., the fact that Z is an expansion. In later chapters we will refer to the nal (isotopy invariant) version of the Kontsevich integral as Z.
Chapter 3 The Kontsevich integral of knotted trivalent graphs

In this chapter we build an extension of the Kontsevich integral of knots to knotted trivalent graphs. With the exception of Theorem 3.4.1, which is from the preprint [BD1], everything in this chapter appeared in the paper [Da].
3.1
The naive extension
We start by trying to extend the denition of the Kontsevich integral to knotted trivalent graphs (and trivalent tangles) the natural way: consider a Morse embedding of the graph (or tangle) in R3 , as shown below. Although the graph is not planar, we understand the framing to be the blackboard framing, in a slightly generalised sense: the normal vector dening the framing is parallel to the plane iR R and its inner product with (i, 0) is positive1 . Dene the integral by the same formula as before, requiring that t1 , ..., tn are non-critical and also not the heights of vertices. (We do not do any correction or renormalization yet.) We denote this naive extension by Z0 .
This requires that the curve is never parallel to (i, 0), but this can be achieved by a small preturbation for both a specic embedding and isotopies.
1
34
Chapter 3. The Kontsevich integral of knotted trivalent graphs

t t4 2 1 t3 3 t2 4
35
DP
t1
z1
z1
Z() =
m=0 t
min <t1 <...<tm <tmax
P =(zi ,zi )
(1)P DP (2i)m
dzi dzi zi zi i=1
ti noncrit,nonvtx
3.1.1
The good properties
The ordinary Kontsevich integral has several nice properties which we hope would hold true for the extension and corresponding graph operations. We list these properties here with intuitive justications, but without proofs for the time being. Factorization property. The extension of the integral seems to preserve the factorization property of the Kontsevich integral, meaning that it would be multiplicative with respect to stacking trivalent tangles and the (vertical) connected sum of graphs (i.e. it would commute with the connected sum operation). Nice behaviour under orientation switches. Z should commute with the orientation switch operation due to the signs (1) in the formula that denes Z. (In other words, when we switch the orientation of an edge, the coecients of each chord diagram in the result of the integral will be multiplied by (1)k , where k is the number of chords ending on the edge we switched the orientation of).
Chapter 3. The Kontsevich integral of knotted trivalent graphs Nice behavior under the edge delete operation.
36
Assuming (mistakenly) that our extension is a convergent knotted graph invariant, consider an embedding of the graph in which the edge e to delete is a straight vertical line, with the top vertex forming a Y , the bottom vertex resembling a shape. (Of course such an embedding exists within each isotopy class.) Now, if we delete the edge e, then in the result of the integral every chord diagram in which a chord ended on e will disappear (go to zero), and the coecient of any other chord diagram stays unchanged (as the integral used to compute it is unchanged). In other words, the extended Kontsevich integral commutes with the edge delete operation.
Nice behavior under edge unzip. Let the embedding of the graph be as above. When we unzip the vertical edge e, we do it so that the two new edges are parallel and very close to each other. In the result of the integral, the chord diagrams that contained k chords ending on e will be replaced by a sum of 2k chord diagrams, as each chord is replaced by the sum of two chords, one of them ending on the rst new edge, the other ending on the second. (Since for each choice of zi on e we now have two choices.) The coecient for the sum of these new diagrams will be the same as the coecient of their parent, (since the two new edges are arbitrarily close to each other). If we were to choose a chord to have both ends on the two new parallel edges, the resulting integral will be zero, as zi zi will be a constant function. Again, the coecients of the diagrams that dont involve chords ending on e are unchanged. Therefore, the extended Kontsevich integral, assuming it exists and is an invariant, commutes with the unzip operation.
37
3.1.2
The problem
The problem with the extension is that the integral as dened above is divergent. Causing this are possible short chords near a vertex (i.e. those not separated from a vertex by another chord ending). These are just like the isolated chords in the knot case but, contrary to the knot case, we have no reason to factor out by all the chord diagrams containing such chords. Also, if we want to drop the F I (a.k.a. 1T ) relation for the sake of working with framed graphs, we have to x the divergence coming from the isolated chords near critical points as well.
3.2
Eliminating the divergence
The Kontsevich Integral has been previously extended to framed links (and framed tangles) by Le and Murakami in [LM], [LM2], and by Goryunov in [Go]. We use essentially the same method as Le and Murakami, a simple version of a renormalization technique from quantum eld theory, which extends to trivalent vertices easily. In short, we know the exact type of divergence, and thus we divide by it to get a convergent integral. Goryunovs approach was dierent, using an -shift of the knot along a general framing (not necessarily the blackboard framing).
3.2.1
The re-normalized integral Z1
We rst restrict our attention to a vertex of a shape, x a scale and chose a small . We change the integral at the vertex by opening up the two lower strands at a distance from the vertex, to a width at the height of the vertex.
38
The old strands (solid lines on the picture) up to distance from the vertex, and above the vertex, will be globally active, meaning that we allow any chords (long or short) to end on them. The opening strands (dashed lines on the picture) are locally active, meaning that we allow chords between them, but chords from outside are not allowed to end on them. We dene the value of Z1 as the limit of this new integral as tends to zero. We will do the same to a vertex of a Y -shape, however, we will have to restrict our attention to these two types of vertices. (I.e. we do not allow vertices to be local minima or maxima.) Of course, any isotopy class graph embeddings into R3 contains a representative such that all vertices are of one of these two types, but this will cause a problem with the invariance of Z1 , which will need to be xed later. To get an invariant of framed graphs, we use the same method to re-normalize at the critical points and thereby make isolated chords cause no divergence, this is why we can drop the F I (or 1T ) relation. Proposition 3.2.1. The re-normalized integral Z1 is convergent. Proof. It suces to consider the case of a -shaped vertex, the other cases are similar. Let us x a -shaped vertex v. We will be computing the integral for the tangle (slice of ) between a xed level and tv , which does not contain any critical or vertex levels other than tv , and we assume that this part of the graph all lies in one plane and all strands except for the legs of the are vertical. As the value will turn out to be invariant under horizontal deformations, we can do this without loss of generality, and
39
since the factorization property also remains true, proving the convergence for this tangle is enough. If the highest chord is a long chord, the convergence is obvious. Suppose the highest k chords, call them ci , ci1 ,..., cik+1 , are short, as ahown in the gure below, and rst compute the integral corresponding to these. t
tv tv ci ci1 ci2
w(t)
The globally active part corresponding to the highest short chord ci can be computed as below.
tv ti1
...
cik+1 cik
tv ti1
dzi dzi = zi zi
d ln(zi zi ) = ln
zi (tv ) zi (tv ) . zi (ti1 ) zi (ti1 )
The locally active part on the other hand:

tv tv
d(ln(zi zi )) = ln
. zi (tv ) zi (tv )
The integral for this highest chord is the sum of these, and is therefore equal to: ln = ln ln(zi (ti1 ) zi (ti1 )). zi (ti1 ) zi (ti1 )
Let us simplify notation by calling w(t) (for width) the distance between the legs of the at the level t. In this notation the result for the top chord is ln ln w(ti1 ). For the next short chord underneath, the double integral corresponding to the two chords is:
tv ti2
ln ln(w(ti1 )) d ln(w(ti1 )) =
2 tv ti2
1 ln ln(w(ti1 )) 2
1 ln ln w(ti2 ) 2
40
Continuing in this fashion for the k short chords, we see that the value of the integral between the two critical levels is depend on . Now suppose the next chord, cik is long. Note that this chord can only end on the globally active part, and hence the integral corresponding to the top (k + 1) chords is
tv tik1 1 k!
ln ln w(tik )
, which is nite and does not
1 k d(zik zik ) . ln ln w(tik ) k! zik zik

dw(tik ) 2
Note that since the other end of cik ends on a vertical chord, d(zik zik ) = and zik zik = C + vertex. So we have
tv tik1 w(tik ) , 2
where C is the distance of the vertical strand from the
1 k d(w(tik )) ln ln w(tik ) . k! w(tik ) + 2C
We need to show that this integral is nite and converges as 0. Since the integrand does not change sign near tv , the value (assuming it is nite) changes monotonically as approaches 0, therefore it is enough to show that the integral is bounded by a number that does not depend on . We have the following obvious bounds: 1 2C + w( )
tv tik1 tv tik1
1 k ln ln w(tik ) d(w(tik )) k!
1 k d(w(tik )) ln ln w(tik ) k! w(tik ) + 2C 1 k ln ln w(tik ) d(w(tik )). k!

tv 1 tik1 k!
1 2C
tv tik1
Thus, it is enough to show boundedness for
ln ln w(tik ) d(w(tik )). Inte-
grating by parts, we nd that the value of this integral is

k
j=1
1 w(tik ) ln ln w(tik )j j!
tv tik1
+ [w(tik )]tv = tik1
41
Assuming that w(tv ) = (which we can assume without loss of generality), we obtain that the above formula equals 1 k (ln ln )j j! j=1
k
j=1
1 w(tik1 )(ln ln w(tik1 ))j + w(tik1 ). j!
The terms of the rst sum and the lone both tend to zero as approaches zero, and therefore the integral is bound between two dierent constant multiples of
k
j=1
1 w(tik1 )(ln ln w(tik1 ))j w(tik1 ), j!
and hence it converges as 0. Showing the convergence for any further chords (short or long) is easy and is left to the reader.
3.2.2
The good properties
Let us call the deletion (respectively, unzip) of an edge that is embedded as a vertical line segment vertical edge delete (respectively, vertical unzip). By vertical connected sum, we mean placing one KTG above the another and connecting them by an edge that is a vertical line segment. Theorem 3.2.2. Z1 is invariant under horizontal deformations that leave the critical points and vertices xed, and rigid motions of the critical points and vertices, explained below. Z1 has the factorization property, and commutes with orientation switch, vertical edge delete, edge unzip and connected sum. Moreover, it has good behavior under changing the renormalization scale .
Chapter 3. The Kontsevich integral of knotted trivalent graphs By rigid motions of critical points we mean shrinking or extending a sharp needle, like in the case of the standard Kontsevich integral (Lemma 2.3.4), with the dierence that we do not allow twists on the needle, but require the two sides of the needle to be parallel straight lines. This dif-
42
ference is due to dropping the framing independence relation, as adding or eliminating twists would change the framing. We are going to study the role of framing in more detail after proving Theorem 3.2.2. For vertices, a rigid motion is moving the vertex down two very close edges without twists, as shown in the gure on the right. To prove that the integral commutes with the vertical edge unzip operation and to investigate the behavior under changing the scale , we will use the following lemma: Lemma 3.2.3. Let w1 , w2 be distinct complex numbers and let be another complex number. Let B be the 2-strand rescaling braid dened by the map [, T ] [, T ] C2 t (t, et w1 , et w2 ). Then Z1 (B) = exp t12 (T ) A(2 ), 2i
where t12 is the chord diagram with one chord between the two vertical strands. Proof. The m-th term of the sum in the dening formula of Z is 1 tm (2i)m 12
T T T
...
t1 tm1 T
dln(etm w1 etm w2 )...dln(et1 w1 et1 w2 ) =

T T
1 tm m (2i)m 12
...
t1 tm1
dtm dtm1 ...dt1 =
(t12 (T ))m , (2i)m m!
which proves the claim.
43
We note that Lemma 3.2.3 is easily extended to the case of the n-strand rescaling braid, dened the same way, where in the statement t12 would be replaced by though we dont use this generalization here. We also state the following reformulation, which follows from Lemma 3.2.3 by elementary algebra: Lemma 3.2.4. For the two strand rescaling braid B where the bottom distance between
L
tij ,
the strands is l, and the top distance is L, shown here: Z1 (B) = exp independently of T and . Now we proceed to prove Theorem 3.2.2: ln(L/l)t12 , 2i
Proof. Factorization property. The factorization property for tangles is untouched by the renormalization, as the height at which tangles are glued together must be non-critical and not contain any vertices. For the vertical connected sum of knotted graphs 1 and 2 , denoted 1 #2 , if we connect the maximum point of the 1 with the minimum of 2 , the minimum and maximum renormalizations will become vertex renormalizations when computing the Kontsevich integral of 1 #2 . Invariance. To prove invariance under horizontal deformations that leave the critical points and vertices xed, we use the same proof as in the case of the standard integral (Proposition 2.3.2), i.e. cut the graph into tangles with no critical points or vertices, and thin tangles containing the vertices and critical points, apply Lemma 2.3.3 to the former kind, then take a limit. For invariance under rigid motions of critical points, it is enough to consider the case of a maximum, the case of a minimum being strictly similar. Since we have proven the invariance under horizontal deformations and the needle is not twisted, we can assume
44
that the sides of the needle are two parallel lines and is the horizontal distance between them. By the factorization property, the value of Z1 for the needle extended (see gure) can be written as a product of the values for the part under the needle, the two parallel
Extended Retracted
strands, and the renormalization for the critical point that is the tip of the needle, as shown on the right. The value of Z1 for the needle retracted is the product of the value for the part under the needle and the renormalization part. What we have to show therefore is that in the rst case (needle extended) the coecients for any diagram that contains any chords on the parallel strands tends to zero as the width of the needle tends to zero. This is indeed the case: the integral is 0 for any diagram on two parallel strands that contains any short chord, since d(zk zk ) = 0. For long chords, the highest long chord can be paired up with the one ending on the other strand, as in the proof of 2.3.4. The reason for this is that their dierence commutes with any short chords that occur in the renormalization part, by the Locality Lemma 2.3.1. Now we can use the same estimates as in Proposition 2.3.4 to nish the proof. To prove invariance under rigid motions of vertices, let us assume that all edges are outgoing. All other cases are proven the same way after inserting the appropriate sign changes. Similarly to the case of critical points, we can assume that the part we shrink consists of two parallel strands at horizontal distance . We need to prove that the dierence of the values of Z1 for the two pictures shown below tends to zero as tends to zero. For the value corresponding to the left picture, just like in the needle case, we can assume that there are no short chords connecting the two parallel strands. The long chords ending on the parallel strands come in pairs, with the same sign, and their coecients are the same
45
in the limit. These pairs commute with any short chords in the renormalization part by Lemma 2.3.1. Also, by the vertex invariance relation, each sum of a pair of such chords equals one chord ending on the top vertical edge, which, from the right side integral, will have the same coecient as the former sum, as 0. This concludes the proof. Good behavior under orientation switch The renormalization doesnt change anything about the signs that correspond to the orientations of the edges, so Z1 still commutes with orientation switches. Good behavior under vertical edge delete. When deleting a vertical edge, the renormalization that was originally inserted for the two vertices at either end of the edge becomes exactly the renormalization we need for the two critical points that replace the vertices, as on the gure. Good behavior under vertical edge unzip. It is slightly harder to see that Z1 commutes with unzipping an edge. If we unzip a vertical edge e and then compute the value of Z1 , the vertices on either end of e disappear, so no renormalization will occur. However, if we rst compute the integral and then perform unzip on the result in A(), then the coecients for each resulting chord diagram will be as if we had computed them using re-normalizations as in the picture on the right. What we need to show is that the contribution from the upper renormalization will cancel the contribution from the lower. Let T denote the trivalent tangle from the lower renormalization to the upper renormalization (including any faraway part that is not on the picture). Let us divide T into three tangles: let T1 denote the lower renormalization area (short chords only), T2 the unzipped edge and any faraway part of the graph at this height, and T3 the upper renormalization area. By the factorization property, Z(T ) = Z(T1 )Z(T2 )Z(T3 ) In the integrals result, the short chords occurring at either renormalization part can slide up and down the unzipped edge by Lemma 2.3.1, as they commute with the pairs
Chapter 3. The Kontsevich integral of knotted trivalent graphs of incoming long chords.
46
In other words, Z(T1 ) commutes with Z(T2 ), and therefore Z(T ) = Z(T2 )Z(T1 )Z(T3 ). Z(T1 )Z(T3 ) = exp
ln(/)t12 2i
exp
ln(/)t12 2i
= 1 A(T ) by the reformulated rescal-
ing Lemma 3.2.4, so the re-normalizations cancel each other.
Good behavior under changing the scale . Chang
ing the scale to some other amounts to adding a small, two-strand rescaling braid at each vertex and critical point, as shown. By Lemma 3.2.4, this means that when changing the scale from to , the element exp point, and exp minimum. It remains to study the behavior of Z1 under framing changes. To do this, we rst compute the Z1 -value of a crossing. Proposition 3.2.5. Z1 () = e
t12 2
ln( /)t12 2i ln(/ )t12 2i
A(2 ) is placed at each -shaped vertex and maximum = exp

ln( /)t12 2i
A(2 ) is placed at each Y -vertex and
and Z1 () = e
t12 2
, where t12 is a chord between the
two strands. We understand the distance between the top and bottom ends of strands to be the same. Proof. First let us compute Z1 of the square of the crossing, which, by horizontal deformations, can be deformed to one straight strand and one wrapping around it in a circle, as shown on the right. We parametrise the straight strand by z(t) = 0, and the wrapped around one by z(t) = e2it , both for 0 t 1. The m-th term of Z1 in the dening formula is computed as follows (in this case there is no dierence between Z0 and Z1 : 1 tm (2i)m 12
1 0 0 1 tm1 t2
...
0 tm1
d ln(e2it1 )d ln(e2it2 )...d ln(e2itm ) =

t2
= tm 12
0 0
...
0
dt1 dt2 ...dtm =
tm 12 . m!
47
this (due to the factorization property), namely Z1 () = e
Thus, the value of the square of the crossing is et12 , and Z1 () is the square root of
t12 2
. The computation for the
undercrossing can be done the same way, or one can use that the over-and undercrossing are tangle inverses. It is obvious from the factorization property, the invariance under horizontal deformations, and Lemma 3.2.4 that as long as the distance between the two strands is the same on the bottom and on the top, making the crossing tangle shorter, taller, thinner or wider does not change the result. This lets us compute how framing changes aect the value of Z1 . For now, we restrict ourselves to adding or removing a kink on the top of the graph, but after dealing with invariance issues in the next section, this will generalise to framing changes anywhere.
Proposition 3.2.6. Z1
= et/2 Z1
, where t denotes a short chord on the strand.
Proof. The statement follows from Proposition 3.2.5 and the factorisation property. The only point that needs a word of explanation is t being a short chord: the only thing that makes the chords coming from the crossing not isolated short chords in the rst place are the ones coming from the maximum, which are themselves isolated short chords, and hence commute with everything.
Note that framing changes amount to adding local chords to a strand, which are central (commute with all other chords). Positive and negative kinks add inverse chord diagrams, and this implies that after xing the invariance issues in the next chapter, the resulting Z2 is invariant under the R1 move, hence is a KTG invariant.
48
3.3
3.3.1
Corrections - constructing a KTG invariant

Missing moves
As in the case of knots, the Kontsevich integral is not invariant under certain deformations that do not change the isotopy class of the framed graph. In the case of knots, the only missing move to make Z an isotopy invariant was straightening a hump, and we xed this by multiplying with Z( )c/2 .
The situation here is more complex than in the case of knots. Z1 is invariant under deformations that do not change the number of critical points or the shape of vertices. To transform it into a knotted framed graph invariant Z2 , we need to make corrections to create invariance under the following eight moves:
1 2 3 4
Moves 1 and 2 will x the problem of straightening humps, just like in the case of knots. Moves 3, 4, 5 and 6 guarantee that we can switch a vertex freely from a shape to a Y shape and vice versa. Moves 7 and 8 are needed because we excluded vertices that are critical points at the same time. Any such vertex can be perturbed into a or a Y shape by an arbitrarily small deformation, but we need the value of the invariant to be the same whether we push the middle edge over to the left or to the right. Remark 1. Corrections will not eect the behavior of Z1 with respect to framing changes,
49
so everything that was said for Z1 stands for Z2 , with the added benet that due to the the invariance of Z2 (which is to be achieved in this section), framing changes will not need to happen at the top of the KTG anymore. This also implies that Z2 will be invariant under the R1 move, so it is an honest (framed) KT G invariant. To implement the corrections, we have four correctors available: we can prescribe local chord diagrams to put on minima, maxima, Y -vertices and -vertices. The ones on cups and caps are elements of A(1 ), the ones on vertices we can think of as elements of A(2 ), as one of the edges can be swept free of chords using the vertex invariance relation. The eight moves above dene eight equations (1 and 2 are equations in A1 , the rest of them in A2 , the unknowns being the four correctors). The question is whether this system of equations can be solved.
3.3.2
Syzygies
To reduce the number of equations that the correctors need to satisfy, we use the syzygies below:
1.
1 2 1
2.
3.
1
4.
5.
The rst syzygy implies that once Z2 is invariant under move 1, it is automatically invariant under move 2 (we used this fact already in the knot case). The second says that invariance under moves 1 and 3 implies invariance under move 5. By the third, invariance
50
under moves 2 and 4 implies invariance under 6. The fourth tells us that xing move 1, 3 and 4 xes 7. Finally, the fth syzygy shows that xing moves 2, 5 and 6 xes move 8 as well. Therefore, its enough to make Z invariant under moves 1, 3 and 4, which together imply invariance under everything else.
3.3.3
Translating to equations
We know already how to make Z1 invariant under move 1, as this was done in the knot case. We need to place a correction u A1 at each minimum, and a correction n A1 at each maximum, the only equation u and n need to satisfy being un = Z1 ( )1 A(1 ). This will settle move 1 the same way it did for knots. (The correction used in the knot case is just an instance of the general one described here. The tool used for knots was the lemma of a distinguished tangle, Lemma 2.3.6, which said that faraway parts of a knot dont interact. The proof of this applies in the graph context word by word.) Z1 ( ) is often denoted in the literature by 1 , where denotes the Kontsevich invariant of the un-knot. Now let us translate moves 3 and 4 into equations for correctors Y and . For simplicity, let us assume that in the picture for moves 3 and 4, the width of the opening at the top is 1. (This can be achieved by horizontal deformations when applying these moves on any graph.) Let us rst also assume that the xed scale is one. (We can correct this later as we know how changing eects the value of Z1 ). Finally, we choose one convenient set of orientations for the edges - all other cases will follow from this one by performing the orientation change operation on the appropriate edges. To compute the values of Z1 on the left side of move 3 and the right side of move 4, we will make use of the following lemma:
51
Lemma 3.3.1. Assuming that the width at the opening at the top of each of the pictures below is 1, and = 1, Z1 = , and Z1 = ,
i.e., the value of Z1 on these trivalent tangles is trivial (no chords at all). Proof. There are two types of chords appearing in Z1 : horizontal chords connect-
ing the left vertical strand to the diagonal edge, and horizontal chords connecting the diagonal edge to the right vertical strand. Note that there are no chords on the bottom part: the opening of the strands is of width 1, and so is the renormalization scale for the minimum, so when computing Z1 , we get exp
ln(1)t12 2i
as we showed in Lemma 3.2.4, which equals 1 A2 (since ln(1) = 0),
meaning no chords. Also, there are no chords connecting the left and right vertical strands, as these are parallel, so in the denition of Z, d(zi zi ) = 0. The key observation we will use is that the two types of existing chords commute. This is an elementary computation making repeated use of the Locality Lemma 2.3.1 and the VI relation: we are going to prove is that any chord diagram with chords as above is equivalent to one where all the horizontal chords on the right are on the top, followed by all the horizontal chords on the left at the bottom: =
To prove this, let us rst establish a few basic equalities. For the bottom chord on the right, the V I relation gives the following: = + (1)
We will use the following shorthand notation:
Chapter 3. The Kontsevich integral of knotted trivalent graphs + = ,
52
so the sum above will be denoted as as the root.
. We will refer to the xed end of the chords
Next, since a chord on the left commutes with both chords in this sum (obviously with the horizontal one, and by the Locality Lemma 2.3.1 with the short one), it commutes with the sum:
(2)
We can also pull only the box part of a sum over a left chord, this, when expanded, is a 4T relation:
(3)
And lastly we claim that we can separate nested sums, as we can slide the lower box over the sum in the middle, since it commutes with both parts by the equation above and the Locality Lemma 2.3.1:
(4),
Therefore, we can pull all the chords on the right to the top like we wanted to, by the following algorithm. First, we pull all the right ends of the chords over the right vertex, using (1), creating a number of boxes. Then we pull the box end of what was the lowest chord (which has the highest box now) up to its root, over left horizontal chords if needed, which we can do using (3). Continue by pulling the next box up to its root, over horizontal left
53
chords and the entire lowest sum, both of which are legal steps from above ((3) and (4)). Continue doing this until all the boxes are united with their roots. Then we can slide all the sums to the top as they commute with the horizontal left chords, as shown in (2). Now we reproduce the above procedure backwards, to turn the boxes back into right chords, except that, now, they are sitting above the left chords. This concludes the proof of the observation. Now its easy to show that all these chords indeed cancel out in Z1 : since
the left ones all commute with the right ones, the value doesnt change if we compute Z1 separately on the left and on the right. As the opening of strands is 1, as well as the renormalization scale, we have exp
ln(1)t12 2i
= 1 on both sides, by Lemma 3.2.4. This
concludes the proof, and the proof for the mirror image is the same.
An easy corollary is the following: Lemma 3.3.2. Again assuming that the width at the opening at the top of each of the pictures below is 1, and = 1, the following is true:
Z1 ) A(2 ).
=
a
= Z1
where a = Z1 (
Let us say a word about our slight abuse of notation in stating this lemma: In the statement, we mean that the edges of the answer are oriented according to the orientations of the edges of the picture were computing Z1 of, so the leftmost and rightmost sides are not really equal. For the denition of a A2 , due to the symmetry of the picture, it doesnt matter which way the two parallel edges go, as long as they are oriented the same way. In other words, S1 S2 (a) = a, where S1 and S2 are the orientation reversing maps. Note also that by denition of Z1 , a is an invertible element of A2 .
Chapter 3. The Kontsevich integral of knotted trivalent graphs Proof. First note that Z1 ( ) = Z1 ( )Z1 ( ),
54
since = 1 and by Lemma 2.3.6. The rst factor on the right side is 1, since = 1. Now the proof is essentially identical to the previous one. For the rst step of the key observation, we need to slide the right end of the lowest right chord over the bottom part of the picture. This is done by using one more VI relation. The resulting sum of two chords commutes with a by the Locality Lemma 2.3.1. Thus we can pull it over a and use the VI relation once more to get a single chord again. A few edge orientations dier from the ones in Lemma 3.3.1. (We chose orientations there to avoid negative signs, here were switching to the ones we will use later.) We can repeat the proof of Lemma 3.3.1 with the same edge orientations, then switch the ones we need to switch at the end. This causes no problems, as Z1 commutes with the orientation switch operation. Since no chords end on the edges were reorienting, the result doesnt change sign. A similar lemma, which is a crucial ingredient in the Murakami-Ohtsuki construction, [MO], is proved by Le, Murakami, Murakami and Ohtsuki in [LMMO]; the proof involves the computation of the Z-value of an associator in terms of values of the multiple Zeta function. Note that the element we call a is called b in [LMMO], and vice versa. We can now use Lemma 3.3.2 to compute the value of Z1 on the trivalent tangles that appear in moves 3 and 4. This is done in the following two corollaries:
Corollary 3.3.3. Still assuming that = 1, Z1 Z1 = a , and Z1 = a .
, Z1
a ,
Note that we are abusing notation again: the way we dened a, it was an element of A(2 ) where the strands were U-shaped. Due to the symmetry of the picture though, we can slide a either up the curve on the left, or up the curve on the right to a vertical position, we will get the same result. This is what we call a here.
55
Proof. Unzipping the vertical edge on the right in
, we get
Since Z1 commutes with vertical unzip, and in Z1 were unzipping, we can deduce that Z1 By the factorization property, Z1 = Z1 Z1 = a .
no chords end on the edge
However, the second factor is 1 (i.e. has no chords) by Lemma 3.2.4, since the opening of the strands at the top is 1, and so is the width of the renormalization at the bottom. Therefore, we deduce that Z1 = a .
Also, we can forget about the vertical strand on the right, as we know from Lemma 2.3.6 (which generalizes to KTGs word for word) that faraway strands dont interact. This proves the rst equality. The proof or the second equality is the same, with all the pictures mirrored. By the factorization property, Z1 = Z1 Z1 .
Again, the second factor is 1 (no chords at all), by Lemma 3.2.4, if we assume that the width of the opening where we cut is 1 (which we are free to do, by horizontal deformations), and the width we use for the vertex renormalization is also 1. The third equality then follows, as does the fourth by mirroring the pictures.
We are now ready to form the equations corresponding to moves 3 and 4 (still assuming that = 1).
56
The value of the corrected integral Z2 is constructed from the value of Z1 by placing correctors on each vertex and extremum. Let us call these correctors n, u, , and Y , for maxima, minima, -shaped vertices, and Y -shaped vertices respectively. The correctors n and u are elements of A(), and Y are elements of A(2 ), and they are placed on the graph as shown in the gure (edge orientation issues will be dealt with later): n u
Y
Using this terminology, we have proven the following proposition: Proposition 3.3.4. With corrections A2 , Y A2 , and u A1 , such that
=
u
=
u
a ,
and n = u1 , Z2 is invariant under moves 1-8, and therefore it is an isotopy invariant of knotted trivalent graphs. As before, = Z1 ( )1 . Note that we have used above that Z1 of a Y vertex with an opening of width 1 at the top equals 1, since = 1.
3.3.4
The resulting invariants
There is an obvious set of corrections satisfying the above equations: = a1 , Y = 1 and u = 1. This forces n = . Note that this and Y value works only with the orientations of the edges chosen as in the pictures above. Correct notation would be to say that = a1 , where in the subscript the rst arrow shows the orientation of the top strand of the vertex, the second stands for the lower left, and the last stands for the lower right strand. Its easy to now come up with a complete set of corrections for all orientations: if we change the orientation of one of the lower strands, we apply the corresponding orientation switch operation to the correction: = S1 (a1 ), = S2 (a1 ), =
57
S1 S2 (a1 ) = a1 . These satisfy the equations of Proposition 3.3.4, since Z1 commutes with the orientation switch operation. If we switch the orientation of all the edges, the proof remains unchanged (due to the fact that S1 S2 (a) = a), so we have = a1 , and then by the above reasoning, the rest follows: = S1 (a1 ), = S2 (a1 ), and = S1 S2 (a1 ) = a1 . It is worth noting that S1 (a) = S1 (S1 S2 (a)) = S2 (a). Since Y is trivial, orientation switches dont aect Y . For n and u the orientation of the strand doesnt matter, since these are elements in A1 , so each chord has two endings on the one strand, thus the orientation change operation acts trivially. Proposition 3.3.5. With the complete set of corrections described above, Z2 commutes with the orientation switch, edge delete and connected sum operation. Proof. It is obvious that Z2 commutes with the orientation switch, as Z1 has the same property and we designed the corrections ( in particular) to keep it true. We only need to deal with vertical edge delete: since we now have an isotopy invariant, we can rst deform the edge to be deleted into a straight vertical line with a Y vertex on top and vertex on bottom. Thus, its enough to check that the following equalities are true: = ,
and also for the opposite orientations of the strands. Each time appears above, we need to check the equations for all appropriate choices of depending on the orientations of the strands, for example and in the second equation. Indeed, = S2 (a1 ) = S2 (S1 S2 (a1 ) = S1 (a1 ) = , and
S1 (a)
= Z1 (
) = Z1 (
) = Z1 (
) = Z1 ( ),
58
where all equalities are understood in A(1 ), using the isomorphism A(1 ) A(S 1 ). The = rst equality is true by denition of a and the fact that Z1 commutes with orientation switches, using that = 1 and Lemma 2.3.6. The second one is again = 1 and Lemma 2.3.6, while the third is horizontal deformations and moving critical points, and the fourth is due to Lemma 2.3.6. Now commutativity with edge deletion follows by taking inverses on both sides. There is nothing to check for Y and u, as theyre both trivial. Connected sum is the reverse of edge delete, the fact that Z2 commutes with it is proved by backtracking the above proof, using that the connected sum is well-dened. Note that have said nothing about wether the invariant commutes with the unzip operation, and in fact it doesnt. Before we deal with this issue, let us produce a 1parameter family of suitable corrections, one of which will have relatively good behavior under unzip. We will use the following lemma, the statement of which appears in [MO]. The proof there is phrased in terms of q-tangles, but is based on the same trick.
x Lemma 3.3.6. a
=
x
bx
, where we dene b A2 to be b = Z1 (
), an
upside down a, and x R any real number. Proof. Throughout the proof we assume that all strand openings are of width 1 as well as = 1, however, the statement itself is just an equality in A(Y ), and thus independent of the choice of . Then, the same way we had done in Lemma 3.3.1 to Corollary 3.3.3, we can compute Z1 = b .
By multiplicativity, the fact that faraway strands dont interact (Lemma 2.3.6), and our previous computations,
b
Z1
=
a
Chapter 3. The Kontsevich integral of knotted trivalent graphs By an unzip (Z1 commutes with unzip), it follows that
59
Z1
= a
We know that adding a hump amounts to multiplying by a factor of 1 = Z1 ( ), therefore Z1 = Z1 1 ,
where multiplication by 1 is on the right strand. But Z1 = 1, due to = 1, as seen before. So we have
b a
= 1.
Using that commutes with everything due to the Locality Lemma 2.3.1, and multiplying by b1 , we get
a
b1
Since a and commute, the lemma follows for all x N (we can pull each b1 through the vertex one by one, then group as and s together). Then by Taylor expansions, the statement follows for all x.
Proposition 3.3.7. For each x R, the corrections = ax1 , u = x , Y = bx , and n = 1x make Z2 a universal nite type invariant of knotted trivalent graphs that commutes with the orientation change, edge delete and connected sum operation. Proof. We need to check that the equations translated from moves 1, 3 and 4 are satised to prove the invariance part: The equation coming from move 1 says nu = , which holds for the n and u above.
Chapter 3. The Kontsevich integral of knotted trivalent graphs Moves 3 and 4 were translated in Proposition 3.3.4 to Y
60
=
u
=
u
a .
This is satised since
bx
= ax
u
x
and the mirror image done the same way. The complete set of corrections follows from these by inserting the appropriate orientation switches. This will ensure that Z2 commutes with the orientation switch operation. The fact that Z2 is a universal nite type invariant is true by the exact same proof that applies to knots. For edge delete and connected sum, the proof is the same as for our rst set of corrections. Throughout the above calculations and proofs, we had always assumed that the renormalization scale was chosen to be 1. However, when we change the scale from to , all that happens is that factors of exp
ln( /)t12 2i
will be placed at -vertices and
maxima, while the reciprocal is placed at Y -vertices and minima. Therefore, if the scale is chosen dierently, we can account for this by multiplying the correction terms with the inverses of these factors. This means that after implementing the corrections, Z2 has no dependence on anymore. Remark 2. When restricted to links, the invariants corresponding to dierent values of x all restrict to the Kontsevich integral. For KTGs they are dierent invariants (this is demonstrated by the fact that they have dierent behavior with respect to the unzip operation). With the use of the symmetric corrections corresponding to x = 1/2, our construction coincides with that of [MO] and [CL]. In fact, as we shall see in the next section, while none of these invariants commute with the unzip operation, the behavior
61
of the one corresponding to x = 1/2 is well controlled and much nicer than that of the others. As a result, this is the only invariant in the sequence that can serve as a basis for xing the unzip issue in Chapter 4, and the only one that we can use to construct an associator, which we do in Chapter 5. Since this issue will be the center of our attention from now on, whenever we refer to Z2 in the future, we mean the invariant corresponding to x = 1/2.
3.4
Non-homomorphicity: unzip
There is one shortcoming of Z2 still: it fails to commute with the unzip operation. (As before, it is enough to restrict our attention to vertical unzip.) Commuting with unzip would require that Y = 1, in other words, after we unzip a vertical edge, we would want the top and the bottom corrections to cancel each other out, as the two renormalizations did before. Note that the chords ending on the unzipped edge come in sums of pairs, and and Y commute with these by the Locality Lemma 2.3.1, so they could indeed cancel each other out, if they were inverses of each other. However, any set of corrections that makes Z2 a KTG invariant can either let it commute with the edge delete operation or with the unzip, but not both. This is easy to show: As said above, for Z2 to commute with unzip, we need
Y
= 1.
Since Y has an inverse in A2 (otherwise it couldnt be a correction), this inverse has to be . Therefore, we have 1=
Y n
=
u
62
a contradiction. Here the second equality is due to the assumption that Z2 commutes with the edge delete operation: as discussed before, this property is equivalent to Y = ,
which implies the second equality. In the argument above, were ignoring edge orientation issues for simplicity. The edge orientations compatible with unzip are not the same as the ones compatible with edge deletion, so, strictly speaking, were not talking about the same and Y . However, they only dier by orientation switches, and orientation switches commute with taking inverses, so this doesnt interfere with the proof. We have shown that no set of corrections can produce a homomorphic expansion of KTGs, but one can still ask if such an invariant exists, and may be constructed in a dierent way. The answer to this question no: Theorem 3.4.1. There is no homomorphic expansion QKT G A, i.e. an expansion cannot intertwine all four operations at once. Since the proof of this theorem uses the notion of a Drinfeld associator, we defer it to Chapter 5, the end of Section 5.1, where we have the necessary denitions and techniques introduced. Now that the reader is convinced (or can at least take it on faith) that the nonhomomorphicity issue is unavoidable, let us understand the bad behavior of unzip better. Take our rst set of corrections as an example. For (vertical) unzip to be dened, either all edges need to be oriented upwards, or all downwards. In both cases = = a1 , and Y = 1. This means that in u(Z2 ()) an extra factor of a1 appears on the unzipped pair of edges, coming from Y . This factor is not present in Z2 (u()), which means that Z2 is not homomorphic with respect to unzip. In other words, Z2 (u()) = a u(Z2 ()), where multiplication by a is understood to occur on the pair of unzipped daughter edges. A similar problem occurs with any choice of corrections, however, in the
63
case of x = 1/2, which leads to the [MO] invariant, the extra factor can be simplied in a nice way. To do this we need the following fact, proved by Le and Murakami in [LM]. We sketch a simple proof of this using only the properties of Z1 and Z2 . Proposition 3.4.2.
b a
=
1 1
u()
Proof. By multiplying the tangles used to dene a and b, computing Z1 and unzipping a vertical edge with no chords, we get:
b a
= Z1 .
We can produce this graph by an unzip:
=u
, but this not being a vertical
unzip, it does not commute with Z1 , so let us use Z2 with our rst set of corrections:
a1 u()
Z2
= a uZ2
=au
For the second equality above, we use the fact that Z1 of this graph is trivial (everything cancels out), so in Z2 all we have is the corrections. Now we can get Z1 back from Z2 by undoing the corrections:
b a
which completes the proof. = Z1 =
1 1
u()
Chapter 3. The Kontsevich integral of knotted trivalent graphs Since the element on the right side of the equality is central (by the Locality Lemma 2.3.1), this implies that a and b commute, and hence ax bx = (ab)x . So in the case of the corrections corresponding to x = 1/2, Y = a1/2 b1/2 = (ab)1/2 . As a result, the behavior of unzip can be described in terms of factors, as shown in the gure on the right, or can be written as u(Z2 ()) = (ab)1/2 Z2 (u()) = 1/2 1/2 u( 1/2 ) Z2 (u()). Z2 () Z2 (u())
64
1/2 1/2 u( 1/2 )
Another way to say this is that Z intertwines the KTG unzip with a renormalized chord diagram unzip operation, u = i2 1/2 u i 1/2 , where i 1/2 is the injection of a factor of 1/2 on the edge, similarly i 1/2 plants a copy of 1/2 . Using this notation, we have that Z2 (u()) = uZ2 (). However, u is not the operation induced by unzip on the associated graded. In the next chapter, we use this choice of corrections and our understanding of the behavior of unzip to construct a modied space of KTGs, which we call dotted KTGs, on which a homomorphic expansion exists.
Chapter 4 Fixing unzip: dotted KTGs

This chapter, as well as the next one, is based on a submitted paper [BD1], joint with Dror Bar-Natan.
4.1
The space of dotted KTGs
In this section we dene the algebraic structure of dotted knotted trivalent graphs or dKTGs. It will later become apparent how the dots and crosses introduced make the exist A dotted trivalent graph is a graph which may have trivalent vertices, and two kinds of bivalent vertices, called dots, and anti-dots, the latter are denoted by crosses. Dots and crosses are part of the skeleton information. An example is shown on the right. Trivalent vertices are equipped with cyclical orientations and edges are oriented, as before. Like before, a dKT G has a well-dened thickening . The algebraic structure dKT G has a dierent kind of objects for each dotted trivalent graph skeleton. The objects K() corresponding to skeleton are embeddings of into
R3 , modulo ambient isotopy, or equivalently, framed embeddings of where the framing agrees with the cyclical orientations at trivalent verices. Obviously, dKT Gs are also represented by dKT G diagrams, with added Reidemeister 65
Chapter 4. Fixing unzip: dotted KTGs
66
moves to allow the moving of bivalent vertices and anti-vertices over or under an edge. We dene three kinds of operations on dKT G. Orientation reversal reverses the orientation of an edge, as before. Given two dKT Gs 1 and 2 of skeletons 1 and 2 , and with identical distinguished trees1 T1 and T2 , the tree connected sum #T1 ,T2 : K(1 ) K(2 ) 1 #T1 ,T2 2 is obtained by deleting the two trees, and joining corresponding ends by bivalent vertices, as shown. The orientations of the new edges are inherited from the leaves of the trees. We leave it to the reader to check that this operation is well-dened.
#T1 ,T2
We allow the distinguished trees to have dots and anti-dots on them, with the restriction that for each dot (resp. anti-dot) on T1 , T2 is required to have an anti-dot (resp. dot) in the same position. The cancel operation cd,a : K() K(cv,a ) is dened when a dKT G has a dot d and an adjacent anti-dot a, in which case cancel deletes both (as in the gure below). This requires the orientations of the three fused edges to agree. cd,a d a
Lemma 4.1.1. dKTGs with the above operations form a nitely generated algebraic structure. Proof. We will show that orientation switch, unzip, delete and edge connected sum (that is, a connected sum operation followed by unzipping the connecting edge) are
By identical distinguished trees we mean that the trees should be isomorphic as graphs. Furthermore, the KTGs 1 and 2 each embedded inside unit cubes so that the distinguished trees are the only parts positionsed outside of the cube, to the right and to the left respectively, forming mirror images of each other. This can clearly be done by suitable isotopies of 1 and 2 .
1
67
compositions of the new operations. In the proof that KT G is nitely generated (see [Th]), a connected sum is always followed by an unzip, so edge connected sum is sucient for nite generation. Furthermore, we need to show that it is possible to add dots and anti-dots using tree connected sum, which then also allows one to delete any dots or anti-dots using the cancel operation. Orientation switch is an operation of dKT Gs, so we have nothing to prove. Unzip can be written as a tree connected sum the following way:
e T1 T2
#T1 ,T2
The graph on the right is almost ue (), except for the dots which result from the tree connected sum. So to show that unzip can be written as a composition of the new operations, it is enough to show that it is possible to get rid of dots. This is achieved by taking a tree connected sum with a circle with three crosses and then canceling:
T1
T2
#T1 ,T2
c3
Edge delete and edge connected sum are done similarly, as illustrated by the gure below:
e T1 T2
#T1 ,T2
#T1 ,T2
T1 T2
68
We have shown above that to add one anti-dot we need to take tree connected sum with a circle with three anti-dots and a trivial tree (since the tree connected sum produces two dots which then have to be canceled). Similarly, to add one dot, we apply tree connected sum with a circle which has one anti-dot on it. Using that KT G is nitely generated by the two tetrahedrons, we have now shown that dKT G is nitely generated by the following four elements:
The reader might object that tree connected sum is innitely many operations under one name, so it is not fair to claim that the structure is nitely generated. However, only two of these (the tree needed for unzip and delete, and the trivial tree used for edge connected sum and vertex addition) are needed for nite generation. Later we will show a slightly dierent construction in which we only use the operations that are essential, however, we felt that tree connected sums are more natural and thus chose this version to be the starting point.
4.2
The associated graded space and homomorphic expansion
As before, the associated graded space has a kind of objects for each skeleton dotted trivalent graph , denoted A(), generated by chord diagrams on the skeleton , and factored out by the usual 4T and V I relations, the latter of which now applies to dots and anti-dots as well, shown here for dots:
69
(1)
+ (1)
= 0.
Orientation reversal acts the same way as it does for KTGs: if there are k chord endings on the edge that is being reversed, the diagram gets multiplied by (1)k . The tree connected sum operation acts on AdKT G the following way: if any chords end on the distinguished trees, we rst use the VI relation to push them o the trees. Once the trees are free of chord endings, we join the skeletons as above, creating bivalent vertices. Again, this operation is well-dened. The cancel operation deletes a bivalent vertex and an anti-vertex on the same edge, without any change to chord endings. Theorem 4.2.1. There exists a homomorphic expansion Z on the space of dKTGs, obtained from the Z2 (with x = 1/2) of Chapter 3 by placing a 1/2 near each dot and 1/2 near each anti-dot. We remark that since the invariant restricts to the Kontsevich integral Z of knots (without dots), it is not ambiguous to call it Z. Proof. First, note that Z is well-dened: it does not matter which side of a dot (resp. anti-dot) we place 1/2 (resp. 1/2 ) on: if one edge is incoming, the other outgoing, then these are equal by the V I relation, otherwise they are equal by the V I relation and the fact that S() = , where S denotes the orientation switch operation. Since Z2 is an expansion of KTGs, it follows that Z is an expansion. For homomorphicity, we must show that Z commutes with the orientation switch, cancel and tree connected sum operations. Orientation switch and cancel are easy. If an edge ends in two trivalent vertices, then on that edge Z coincides with Z2 and hence commutes with switching the orientation. If one or both ends of the edge are bivalent, then Z might dier from Z2 by a factor of (or two), but still commutes with S by the fact that S() = . (This is because the trivially
70
framed oriented un-knot is isotopic to itself with the opposite orientation.) Z commutes with cancel because the values of the dot and anti-dot are inverses of each other (and local, therefore commute with all other chord endings and cancel each other out). In terms of the old KTG operations (disregarding the dots for a moment, and ignoring edge orientation issues), a tree connected sum can be realized by one ordinary connected sum followed by a series of unzips:
T1 T2 # u u
We want to prove that Z(1 #T1 ,T2 2 ) = Z(1 )#T1 ,T2 Z(2 ). To compute the left side, we trace Z2 through the operations above. We assume that the trees have been cleared of chord endings in the beginning using the V I relation (and of course chords that end outside the trees remain unchanged throughout). Z2 commutes with connected sum, so in the rst step, no chords appear on the trees. In the second step, we unzip the bridge2 connecting the two graphs. As mentioned at the end of Chapter 3, Z2 intertwines unzip with u = i2 1/2 u i 1/2 . It is a simple fact of chord diagrams on KTGs that any chord diagram with a chord ending on a bridge is 0, thus the rst operation i 1/2 is the identity in this case. After the bridge is unzipped, i2 1/2 places 1/2 on the two resulting edges. The next time we apply u, i 1/2 cancels this out, the edge is unzipped, and then again 1/2 is placed on the daughter edges, and so on, until there are no more edges to unzip.
We use the word bridge in the graph theoretic sense: an edge of the graph such that deleting this edge would make its connected component break into two connected components. In this context, we mean the new edge resulting from the connected sum.
2
71
The operation i 1/2 will always cancel a 1/2 from a previous step. Therefore, at the end, the result is one factor of 1/2 on each of the connecting edges. We get 1 #T1 ,T2 2 by placing a dot on each of the connecting edges in the result of the above sequence of operations. Z adds a factor of 1/2 at each dot, which cancels out each 1/2 that came from the unzips. Thus, Z(1 #T1 ,T2 2 ) has no chords on the connecting edges, which is exactly what we needed to prove. Let us note that edge orientations can indeed be ignored: for the unzips used above to be legitimate, a number of orientation switch operations are needed, but since S() = , the action of these on any chord diagram that appears in the calculation above is trivial. If the trees had dots and anti-dots to begin with, provided that for every dot (resp. anti-dot) on T1 , T2 had an anti-dot (resp. dot) in the same position, these will cancel each other out, and we have already seen that Z commutes with the cancel operation.
4.3
An equivalent construction
In this section we describe an eqvivalent construction which may be less natural but simpler than the previous approach. Let the space dKTG have the same objects as dKTG, but we dene the operations dierently. We keep orientation reversal and cancel the same. Instead of tree connected sums, we introduce the following three operations: Edge delete is the same as in the space KT G, i.e., if orientations match, we can delete an edge connecting two trivalent verices, and as a result, those vertices disappear. Dotted unzip allows unzips of an edge connecting two trivalent vertices with one dot on it (technically, two edges), provided that orientations match at the trivalent vertices, as shown:
e f
ue,f ()
72
Dotted edge connected sum is the same as edge connected sum, except dots appear where edges are fused (and there are no conditions on edge orientations):
1 e1
2 e2
#e1 ,e2
Alternatively, one can allow dotted connected sums (a connected sum where a dot appears on the connecting edge). In this case, this construction is slightly stronger than the previous one, as dotted connected sum cannot be written in terms of the operations of the rst construction. Dotted edge connected sum is the composition of a dotted connected sum with a dotted unzip. The associated graded space is as in the case of dKTG, and the induced operations on it are the same for orientation reversal and cancel; are as in the case of KTGs for delete and dotted unzip; and are as one would expect for dotted edge connected sum (no new chords appear). Proposition 4.3.1. The two constructions are equivalent in the sense that every dKTG operation can be written as a composition of dKTG operations and every dKTG operation is a composition of dKTG operations, involving some constants (namely, a dumbbell garph and a trivially embedded tetha graph, as shown in the proof ). Proof. For the rst direction we only need to show that a tree connected sum can be written as a composition of dKTG operations. We have essentially done this before, in the proof of Theorem 4.2.1. The composition of operations required is one dotted edge connected sum, followed by a succession of dotted unzips, and orientation switches which we are ignoring for simplicity (as noted before, they dont cause any trouble):

T1 T2 # u
73
For the second direction, we need to write edge delete, dotted unzip and dotted edge connected sum as a composition of dKTG operations. For dotted edge connected sum, this was done in the proof of Theorem 4.2.1. Dotted unzip and edge delete are tree connected sums with given graphs and given trees, similar to the proof of Theorem 4.2.1, shown below:
e T1 T2
#T1 ,T2
c4
e T1 T2
#T1 ,T2
c2
Proposition 4.3.2. Z is a homomorphic expansion of dKTG. Proof. It is obvious from the homomorphicity of Z on dKTG that Z commutes with orientation switch, cancel, and dotted edge connected sum. Since dotted unzip and edge delete are unary operations, to show that Z commutes with them, we need to verify that the values of Z on given graphs we used to produce
Chapter 4. Fixing unzip: dotted KTGs these operations from tree connected sums are trivial, shown here for edge delete: Z(d()) = Z(c4 (#0 )) = c4 (Z()#Z(0 )).
74
Here, 0 denotes the dumbbell graph with four anti-dots, shown above. If Z(0 ) = 1, and provided that the edge to be deleted was cleared of chords previously, using the V I relation, then the right side of the equation equals exactly d(Z()). Since Z2 of the trivially embedded dumbbell (with no anti-dots) has a factor of on each circle, Z(0 ) is indeed 1, since two anti-dots add a factor of 1 on each circle. Unzip is done in an identical argument, where we use that the Z value of a trivially embedded theta-graph with one cross on each strand is 1. This follows from the homomorphicity of Z: a tree connected sum of two of these graphs is again the graph itself, as shown below. If the Z value of the graph is , we obtain that 2 = , and since values of Z are invertible elements of A, this implies that = 1.
Chapter 5 Applications
5.1 The relationship with Drinfeld associators
Associators are useful and intricate gadgets that were rst introduced and studied by Drinfeld [Dr1, Dr2]. The theory was later put in the context of parenthesized (a.k.a. non-associative) braids by [LM], [BN3] and [BN4]. Here we present a construction of an associator as the value of Z on a dotted KTG.1 We rst remind the reader of the denition. An associator is an element A(3 ) (where A(3 ) denotes chord diagrams on three upward-oriented vertical strands, subject to tangle 4T ), which satises three major equations, called the pentagon and the two hexagon equations, as well as two minor conditions. Let us remark here that according to a recent and surprising result due to H. Furusho [F] (see also [W]), it turns out that under some moderate restrictions (in the class of horizontal chord associators) the pentagon equations implies the hexagon equations. In [BD2], joint with Dror Bar-Natan, we present a simpler proof of this theorem. The pentagon is an equation in A(4 ). Before we write it, let us dene some necessary
This, along with the fact [MO] that Z2 can be constructed from associators, is the content of the often stated assertion that associators are equivalent to homomorphic expansions of knotted trivalent graphs.
1
75
Chapter 5. Applications
76
maps i : A(n ) A(n+1 ). i , for i = 1, 2, .., n, is the doubling (unzip) of the i-th strand, which acts on chord diagrams the same way unzip does. 0 adds an empty strand on the left, leaving chord diagrams unchanged. Similarly, n+1 adds a strand on the right. Multiplication in A(n ) is dened by stacking chord diagrams on top of each other. In this notation, we can write the pentagon equation as follows (products are read left to right and bottom to top): 4 () 2 () 0 () = 1 () 3 (). The hexagons are two equations in A(3 ), involving A(3 ) and R A(2 ). The permutation group Sn acts on A(n ) by permuting the strands. We denote this action by superscripts, as illustrated by the example on the right. The two hexagon equations are as follows: 2 (R) 231 = 3 (R) 213 (0 (R))213 2 (R1 ) 231 = 3 (R1 ) 213 (0 (R1 ))213 In addition to the pentagon and hexagon equations, associators are required to satisfy two minor conditions, non-degeneracy and horizontal unitarity (or mirror skew-symmetry), as below: 1. is non-degenerate: di () = 1 for i = 1, 2, 3, where di denotes the deletion of the i th strand, which acts the same way as it does for KTGs, and maps A(3 ) to A(2 ). 2. The mirror image of is 1 : 321 = 1 . There are several other useful symmetry properties that the associator constructed here satises. Some of these are required in constructions which involve associators as ingredients, for example, the construction of a universal nite type invariant of KTGs in 231 =
77
[MO] uses the rst two properties mentioned below. Stating some of these is notationally awkward, so let us introduce some useful alternative notation before dening the properties. For any map between free groups : Fm Fn , there exists a pullback : An Am , dened the following way. denes a map of a bouquet of m circles to a bouquet of n circles, up to homotopy, which induces on the fundamental groups. Removing the joining point of the bouquets, one obtains a partial covering map from m oriented lines to n oriented lines. Given a chord diagram c An , (c) is the sum of all possible lifts to the covering m lines, with negative signs wherever the orientations are opposite, and setting the image to be 0 when a chord ends on a strand not covered. Let us illustrate this on an example. Consider the map = (x1 , x1 x1 , x1 ) : F3 2 2 F2 , where x1 and x2 are the free generators of F2 , and the notation means that sends the rst free generator of F3 to x1 , the second to x1 x1 , and the third to x1 . In the covering sense, 2 2 this can be illustrated by the the left side of the gure on the right, while the right side shows the image (c) of a chord diagram c A2 in familiar, but awkward notation. Properties (1) and (2) are easily rephrased in terms of the new notation:
1. d1 () = 1 can be written as 1 () = 1 where 1 = (x2 , x3 ) : F2 F3 . Similarly
2 1 1
1 (2 (S2 (c)))
for d2 and d3 .
2. 321 = 1 means 2 () = 1 , where 2 = (x3 , x2 , x1 ) : F3 F3 .
Now let us state the further symmetry properties promised above.
Chapter 5. Applications 3. An associator is called horizontal if it lives in the sub-algebra Ahor (3 ), consisting of chord diagrams made only of only horizontal chords. The associator we construct here is not a horizontal chord associator, however, it possesses the following crucial property of horizontal chord associators:
3 ()
78
1 (S1 ())
=1
1,
where
3 = (x2 x1 , x3 x1 ) : F2 F3 . This, in fact, is the only consequence of hori1 1 zontality needed in [MO]. The gure on the right shows this property in the other notation. 4. is even or vertically unitary, meaning upside down is 1 : S1 S2 S3 () = 1 ,
equivalently 4 () = 1 , where 4 = (x1 , x1 , x1 ) : F3 F3 . 1 2 3
5. We will prove that the associator we construct has an additional symmetry, as shown in the gure below, which we
1 (3 (S2 (S3 )))35412
= .
call rotational symmetry: 5 () = , where 5 = (x1 , x1 x1 , x1 x1 ). 3 3 2
This arises from the order 3 rotational symmetry of the tetrahedron, and the reader
3 can check that indeed 5 = 1. This property is rather unsightly when written in
the other notation, nevertheless it is shown on the right. Note that for any chord diagram on a dKTG skeleton, one can pick any spanning tree and use the V I relation to sweep it free of chords. (In a
1 2 3 1 2 3
slight abuse of notation, by a vertex we shall mean a trivalent vertex, and by an edge, and edge connecting two trivalent vertices, which may have dots and crosses on it, so it may really be a path.) This sweeping trick induces a (well-dened) isomorphism from chord diagrams on a dKTG with a specied spanning tree to some A(n ). For example, there is an isomorphism from chord diagrams on a trivially embedded tetrahedron to A(3 ), as shown in the gure on the right. We will now prove that we have constructed an associator:
79
Theorem 5.1.1. The following and R satisfy the pentagon and hexagon equations, as well as properties (1) (5), and therefore is a (nice) associator:
=Z
1 2
R=Z
Note that the tetrahedron shown above is isotopic to the picture on the right, which might remind the reader more of the usual representation of an associator.
Proof. We rst prove that satises the pentagon equation. In the proof, we will use the vertex connected sum operation shown in the gure on the right. This can be thought of either
as a tree connected sum in the rst model, or the composition of a dotted connected sum composed with dotted unzips in the second model. Now let us consider the sequence of operations shown in the gure below. In each step, we are thinking of Z of the pictured graph, which commutes with all the operations. To save space, we will not write out the Zs.
80
1b 2b 1a 2a 3a
3b c5 #2
1b 2b 1a 3c 2c
3b u
1c
2a 3a
1c 2c
3c
3b u 1 1a 2a 3a 1b = 4 1c 3c
1c 2c 3c 1b 2b 2b 3b
2b 2 3 2b
1a 2a 3a 2c
Since Z is homomorphic, the result of this sequence of operations is 4 ()2 ()0 (), the left side of the pentagon equation. For the right side, we perform a vertex connected sum of two tetrahedra:
2e 3e c3 # 1e VI 3d 1d 2d 3d 1e 1e 2e 3e 1d 2d 1d 1e 1e 1 2 2d 2e 3e 2e 3 1e 3d 4 3d 3e 1d 2d 3d 3d =
The result can be written as 1 () 3 (), the right side of the pentagon equation. Since the two resulting dKTGs are isotopic (trivially embedded triangular prisms with crosses in the same positions), the two results have to be equal, and therefore satises the pentagon equation.
81
Morally, the hexagon equation amounts to adding a twist to one of the tetrahedra in the right side of the pentagon, on the middle crossed edge, which produces a triangular prism with a twist on the middle vertical edge. Unzipping this edge then gives a new twisted tetrahedron. More precisely, we carry this out in a similar fashion to the proof of the pentagon. For the left side of the rst hexagon equation we take vertex connected sum of two tetrahedra with a twisted theta graph:
1c 2b 3c 1b 1c 2c 3c 1 3c 1c 2c 2 = 1b 2b 2b 1a 2a 3a 2c
c5 #2
1a 2a 3a
1b
2b 3c 2c
1c u
2a 3a
1a
1a u 2a
2b 1b 3 2b 3a
The result is 2 (R) 231 , the left side of the rst hexagon equation. For the right side, we need to add an extra strand to the twisted theta graphs. This can be done without changing the value of Z: we have mentioned before that the Zvalue of a trivially embedded tetha graph with one cross on each strand is 1. Adding an extra strand can be realised by taking an edge connected sum with such a theta graph:
2d 1d c2 # 2d 1d
On to the right side of the hexagon equation, we now connect a tetrahedron with two twisted theta graphs with an added strand, and unzip. The reader can check that moving the extra strand from one side to the other can be done by an isotopy.
82
2d 1d 3e
2e
1e
2d 2f
c5 #2
1e 2e 3e 2f 1f
1d
1f
1 u2 3
2d 2 1d
1e 2e 3e 1f 2f =
1f 1d 2d
2f
2e 1e 3e
The result reads 3 (R) 213 (0 (R))213 , the right hand side of the hexagon. Is is apparent that the resulting twisted tetrahedron graphs of the left and right sides are isotopic, proving that and R satisfy the rst hexagon equation. For the second hexagon, we rst show that
(R21 )1 =
by taking a vertex connected sum with R:

1
2b 1b 1b
2 1 2
2b
c3 #
1a 2a 2a
1a
Note that the resulting graph is isotopic to a trivially embedded theta-graph with one cross on each edge. We have seen at the end of the proof of Proposition 4.3.2 that Z evaluated on this graph is trivial, which proves the claim. For a proof of the second hexagon equation, one substitutes the picture for (R21 )1 in place of R everywhere.
83
Remark. It is easy to check from the denition of Z that R21 = R. If one requires this additional property of R, then the strand swap of the second hexagon equation may be ommitted. In fact R, when thought of as a chord diagram on two strands, has only horizontal chords. More specically, R = e
t12 2
, where t12 is a chord connecting the
two strands, as seen from the step-by-step construction of Z: in Proposition 3.2.5 we computed that the Z1 -value of the crossing is e
t12 2
. The renormalizations of the top and
the bottom vertex cancel each other out, and the vertex corrections of Z2 are cancelled by the -factors of the three crosses. As an aside, we note that R21 = R could also be deduced from the properties of a homomorphic expansion, if dened slightly more generally. In this paper with work with KTGs that are oriented surfaces embedded in R3 . However, one could allow nonoriented surfaces, with a slight modication of the methods presented here. In this case, the Moebius band replaces the twisted tetrahedron as a generator of the space of dKTGs. The notation needs to be modied as blackboard framed graph diagrams are not enough in this case: we need an additional feature, the half twist, which comes in left and right varieties, and the vertex orientations need to be specied at each vertex (see the example on the right). As for Reidemeister moves, the half twists can move freely behind and in front of crossings and across dots and crosses. There are a few additional relations: one can move a twist across a vertex, two twists equal a kink, and two opposite twists cancel, shown below for one direction:
The homomorphic expansion is constructed from Z old by rst ignoring the twists, then placing a factor of ec/4 for each right twist and ec/4 for each left twist, where c is a short (local) chord on the appropriate strand (as well as the appropriate factors on dots and crosses).
84
In this notation, the twisted theta graph representing R is isotopic to the graph shown on the right. The Z-value of the trivially embedded theta
1 2
with three crosses is 1, as shown before, so the value of this graph is ec /4 on all three strands, which is obviously symmetric when pushed onto the rst
two strands. WHY EQUALS et12 /2? Now on to proving the additional symmetry properties:
1. To show non-degeneracy, observe that by homomorphicity, di () = Z di
1 2
Deleting any of the numbered edges from the tetrahedron yields a theta-graph with one cross on each edge, and, as mentioned before, the value of Z on this graph is indeed 1. 2. Horizontal unitarity follows from the fact that satises the pentagon and that R21 = R, as proven in REFERENCE. Alternatively, observe that 321 is the horizontal mirror image of 1 , and Z is invariant under taking mirror images for dKTGs with no crossings. This is easy to see from the sep-by-step denition of Z: since there are no crossings, the graph is embedded in a planar way. Place this planar picture in the plane R R C R, then taking a horizontal mirror image can be done by the map z z. The integral formula used to compute Z1 is unchanged under this map, and all steps after that are mirror symmetric. 3. The horizontal-like property has the following pictorial proof:
d 1 2 3
3 #
uS1 1
85
In the rst step, we are deleting the un-numbered edge which carries a cross. The second step doesnt change the Z-value of the graph (as an element of A(3 )), since we are taking a connected sum with a dKTG with a trivial Z-value. The rest follows from homomorphicity. The Z-value graph on the right is, on one hand, the same as required in the statement, as an element of A(2 ), on the other hand, it is the theta-graph with three crosses, which has trivial value. 4. Evenness follows from unitarity: the graphs representing S1 S2 S3 () and 321 are isotopic.
5. And nally, rotational symmetry is due to the rotational symmetry of the tetrahedron. The right hand side of the gure below is the same as the formula in the statement.
1 2
3 1
VI 3 3 1 2
2 1 3 1 3
1 3
1 2
The relationship to Drinfeld associators provides the tools we need to prove Theorem 3.4.1, which stated that there does not exist a homomorphic expansion (universal nite type invariant) of KTGs. Proof of Theorem 3.4.1. Let us assume that a homomorphic expansion of KTGs exists and call it Z. By denition, Z has to satisfy the following properties: If is a singular KTG (a KTG with nitely many transverse double points), and C is its chord diagram (chords connect the pre-images of the double points), then
Z() = C + higher order terms.
86
Z commutes with all KTG operations, in particular, with orientation switch, unzip, and connected sum: Z(Se ()) = Se (Z()), Z(ue ()) = ue (Z()), Z(#e,f ) = Z()#e,f Z(). Let us denote the degree k part of the values of Z by Zk . To prove the theorem, we deduce a sequence of lemmas from the above properties, until we get a contradiction. Lemma 5.1.2. Assume that Z is homomorphic, as above, and let Z(O) =: be the value of the trivially framed un-knot. Then, in A(O), 2 = . This implies that all positive degree components of are zero, i.e. = 1. Proof. Z(O) = Taking the connected sum of two un-knots implies: Z( )=
Unzipping the middle edge, we get an un-knot back, which proves that = 2 : = Z(O) =
= 2
To prove that the positive degree components of are zero, let us write out degree by degree. The universality of Z implies that the degree zero part is 1. Let us denote the degree k part by k , for 1 k.
= 1 + 1 + 2 + 3 + ... Now we compute 2 degree by degree:
87
2 = 1 + (21 ) + (1 + 22 ) + (1 2 + 2 1 + 23 ) + ... 2 Comparing term by term, we obtain that 1 = 21 , so 1 = 0, therefore 2 = 22 and so 2 = 0, and so on. By induction, the only term in the degree k component of 2 which does not involve a lower degree (hence zero) component of is 2k , so k = 2k , and hence k = 0. Note that the conclusion of Lemma 5.1.2 implies that = 1 as an element of A(), as A(O) A(). = Corollary 5.1.3. Z( ) = 1 A(2 ).
Proof. The Z-value of the dumbbell graph can be viewed as an element of A(2 ) by sweeping the middle edge free of chords, as we have done before. (In fact, this is not even necessary, as it is an easy property of chord diagrams that any chord diagram that any chord ending on a bridge makes a chord diagram zero.) The statement of the lemma is obviously true, as the dumbbell graph is the connected sum of two un-knots, and Z commutes with connected sum. Lemma 5.1.4. = Z
3 2
is an associator, with R = Z
Proof. The proof is identical to the proof of Theorem 5.1.1, omitting all dots and crosses.
Corollary 5.1.5. Z2
1 2
1 24
Proof. The universality of Z (using that switching the crossing in R gives R1 ) implies that the linear term of R is
1 2
It is a well-known fact (and can be checked by a short computation) that the only non-trivial solution up to degree 2 to the pentagon and hexagon equations with the above constraint on R is = 1 +
1 24
Chapter 5. Applications Corollary 5.1.6. Z2 ( ) = 0, in contradiction with Lemma 5.1.3.
88
Proof. Switching the orientation of edge 1 of the tetrahedron followed by unzipping the edge labeled e below results in a dumbbell graph:
S(1) e 2 3 ue S1 3 2 = S(1) 2 3
Therefore, in degree 2, we have
so Z2 ( zero.
) is some non-zero constant multiple of the right side above, which is non-
This contradiction concludes the proof of Theorem 3.4.1. Note that the associator constructed here is related to, but not the same as, the KZ associator [Dr1]. Our associator is the Z-value of the dKTG shown on the right (by an isotopy of the terahedron). Keep in mind that Z here is the end result of a step-by-step construction Z0 Z1 Z2 Z. Drinfelds KZ associator, translated to our language, is Z1
1 2 3
1 2 3
, where in the picture
the distance between the left and right vetical strands is 1, and = 1. It is easy to see that with = 1, Z1 of the gure on the right is the same as Z1
1 2 3
in the sense that chord diagrams on a tetrahedron can be regarded as chord diagrams on three strands, as explained before. To prove this, we can assume that the distance between strands 1 and 3 of the tetrahedron is 1 (horizontal deformations), and that due to Lemma 2.3.6 (faraway strands dont interact) there are no chords between the right strand and the rest of the picture. All the renormalizations except for the ones present in Z1
1 2 3
cancel.
89
However, in passing from Z1 to Z2 , we introduce 4 vertex corrections (as well as a minimum and a maximum, which together amount to a factor of ). The vertex corrections contain non-horizontal chords, so after this point our associator lives in a dierent space than Drinfelds KZ associator. Also, this explains why our associator has the evenness property while KZ does not. Going from Z2 to Z introduces some further factors, an additional though less signicant dierence.
5.2
A note on the Kirby band-slide move and the LMO invariant
In [LMMO] Le, Murakami, Murakami and Ohtsuki construct an invariant Z of links which induces an invariant of 3-manifolds, which was recently falsely disputed [Ga]. The key step is proving that Z is invariant under the Kirby band-slide move K2, shown below:
The problem is that this move is not a well-dened operation of links, so somewhat cumbersome local considerations (freezing local pictures or xing bracketings) need to be used. Z is dened to be a normalized version of the classical Kontsevich integral Z, where an extra factor of is placed on each link component.
Chapter 5. Applications In our language, let us consider the sub-structure of dKTG the objects of which are links with possibly one -graph component, where circles are required to have two dots on them, and the -graph component is required to have one dot on each edge. The only operation we allow is unzipping edges of the -graph. We think of the theta as the two link components on
90
e f uf
ue
K2
which we want to perform K2, fused together at the place where we perform the operation. Unzipping the middle edge of the theta gives back the original link (before K2), while unzipping a side edge produces the link after K2 is performed. The fact that Z commutes with K2 is then a direct consequence of the homomorphicity of Z with respect to dotted unzip, as summarized by the gure on the right. Note that in this case Z is indeed Z when restricted further to links, via replacing the dots by their values of 1/2 . In summary, we have proven the following: Theorem 5.2.1. There exists an expansion Z for links with possibly one knotted theta component, which is homomorphic with respect to unzipping any edge of the theta compo nent. When restricted to link (with no thetas), Z agrees with the invariant Z of [LMMO]. The reader may verify that the unzip property of Theorem 5.2.1 is exactly the equiv ariance property required for the use of Z in the construction of an invariant of rational homology spheres, see [LMO, BGRT2, BGRT3, BGRT4].
Bibliography
[BN1] D. Bar-Natan: On the Vassiliev knot invariants, Topology 34 (1995), 423472. [BN2] D. Bar-Natan: Algebraic knot theorya call for action,
http://www.math.toronto.edu/drorbn/papers/AKT-CFA.html [BN3] D. Bar-Natan: Non-associative tangles, Geometric Topology (proceedings of the Georgia international topology conference, W.H. Kazez, ed.), Amer. Math. Soc. and International Press, Providence, 1997, 139183 [BN4] D. Bar-Natan: On associators and the Grothendieck-Teichmuller Group I, Selecta Mathematica, New Series 4 (1998), 183212 [BN5] D. Bar-Natan: Finite Type Invariants of W-Knotted Objects: From Alexander to Kashiwara and Vergne, http://www.math.toronto.edu/drorbn/papers/WKO/ [BD1] D. Bar-Natan, Z. Dancso: Homomorphic expansions for knotted trivalent graphs, arxiv:1103.1896 [BD2] D. Bar-Natan, Z. Dancso: Pentagon and hexagon equations following Furusho, to appear in Proceedings of the AMS, arxiv:10101.0754 [BGRT1] D. Bar-Natan, S. Garoufalidis, L. Rozansky, D. P. Thurston: Wheels, wheeling, and the Kontsevich integral of the unknot, Israel Journal of Mathematics, 119 (2000), 217237 91
Bibliography
92
[BGRT2] D. Bar-Natan, S. Garoufalidis, L. Rozansky, D. P. Thurston: The rhus integral of rational homology 3-spheres I: A highly non trivial at connection on S3, Selecta Mathematica, New Series 8 (2002), 315339 [BGRT3] D. Bar-Natan, S. Garoufalidis, L. Rozansky, D. P. Thurston: The rhus integral of rational homology 3-spheres II: Invariance and Universality, Selecta Mathematica, New Series 8 (2002), 341371 [BGRT4] D. Bar-Natan, S. Garoufalidis, L. Rozansky, D. P. Thurston: The rhus integral of rational homology 3-spheres III: The Relation with the Le-Murakami-Ohtsuki Invariant, Selecta Mathematica, New Series 10 (2004), 305324 [BLT] D. Bar-Natan, T. Q. T. Le, D. P. Thurston: Two applications of elementary knot theory to Lie algebras and Vassiliev invariants, Geometry and Topology 71 (2003), 131 [CL] D. Cheptea, T. Q. T. Le: A TQFT associated to the LMO invariant of threedimensional manifolds Commun. Math. Physics 272 (2007), 601-634. [CD] S. V. Chmutov, D. Duzhin: The Kontsevich integral, Acta Applicandae Math. 66 2 (April 2001), 155190. [Da] Z. Dancso: On the Kontsevich integral for knotted trivalent graphs, Alg. and Geom. Topology 10 (2010), 13171365 [Dr1] V. G. Drinfeld: On quasi-Hopf algebras, Leningrad Math. J. 1 (1990), 14191457 [Dr2] V. G. Drinfeld: On quasitriangular quasi-Hopf algebras and a group closely connected with GalQ/Q, Leningrad Math. J. 2 (1990), 829860 [F] H. Furusho, Pentagon and hexagon equations Annals of Math., Vol. 171 (2010), No 1, 545556.
Bibliography [Ga] [Go] R. Gauthier: arXiv: 1010.2559, 1010.2422
93
V. Goryunov: Vassiliev invariants of knots in R3 and in a solid torus, Amer. Math. Soc. Transl. (2) 190 (1999), AMS, Providence, RI, 3759.
[Ko]
M. Kontsevich: Vassilievs knot invariants, Adv. in Soviet Math. 16 2 (1993) 137 150.
[LM] T.Q.T. Le, J. Murakami: Representation of the category of tangles by Kontsevichs iterated integral, Comm. Math. Physics 168 (1995), no. 3, 535562 [LM2] T. Q. T. Le, J. Murakami: The universal Vassiliev-Kontsevich invariant for framed oriented links, Compositio Math. 102 (1996), no. 1, 4164. [LMMO] T. Q. T. Le, H. Muarakami, J. Murakami, T. Ohtsuki: A three-manifold invariant via the Kontsevich integral, Osaka J. Math 36 (1999), 365395. [LMO] T.Q.T. Le, J. Murakami and T. Ohtsuki: On a universal perturbative quantum invariant of 3-manifolds, Topology 37 (1998), 539574 [MO] J. Murakami, T. Ohtsuki: Topological quantum eld theory for the universal quantum invariant, Communications in Mathematical Physics 188 3 (1997), 501 520. [Oh] T. Ohtsuki: Quantum invariants. A study of knots, 3-manifolds, and their sets, Series on knots and everything, 29, World Scientic, 2002. [Th] D. P. Thurston: The algebra of knotted trivalent graphs and Turaevs shadow world, Geom. Topol. Monographs 4 (2002), 337362. [W] T. Willwacher: M. Kontsevichs graph complex and the Grothendieck-Teichmueller Lie algebra, arXiv:1009.1654

Zsuzsanna Dancso - On A Universal Finite Type Invariant of Knotted Trivalent Graphs

Enviado por

Dados do documento

Descrição original:

Título original

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

Zsuzsanna Dancso - On A Universal Finite Type Invariant of Knotted Trivalent Graphs

Enviado por

Direitos autorais:

Formatos disponíveis

On a universal finite type invariant of knotted trivalent graphs

Copyright c 2011 by Zsuzsanna Dancso

Corrections - constructing a KTG invariant . . . . . . . . . . . . . . . . . 3.3.1 Missing moves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

3.3.2 3.3.3 3.3.4 3.4

Syzygies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Translating to equations . . . . . . . . . . . . . . . . . . . . . . . The resulting invariants . . . . . . . . . . . . . . . . . . . . . . .

Chapter 1. Introduction invariants can be born.

Finite type invariants- the algebraic view

Algebraic structures and expansions

Knotted trivalent graphs

Chapter 2. Preliminaries vertex). Given a trivalent graph , its thickening is obtained

Since we are considering framed KTGs, we need a charcterisation of these, which is

Chapter 2. Preliminaries similar to that of framed knots:

towards the leaves, using the isotopy pictured on the right.

pushing twists off the tree

), dk1 () = OO...O for some

as a band graph, i.e. an oriented sur...

face, the boundary operator turns it into its boundary:

. We leave it as an exercise for the reader to verify that this

The denition of the Kontsevich Integral

dzi dzi zi zi i=1

dzk+1 dzk+1 C zk+1 zk+1

tij in AT . Then S commutes with D in AT .

dzi dzi zi zi i=1

zk (tk1 ) zk (tk1 ) zk (tk1 ) zk (tk1 )

zk (tk1 ) zk (tk1 ) zk (tk1 ) zk (tk1 )

C|zk (tk1 ) zk (tk1 )| C,

dzk1 dzk1 zk1 zk1

(zk (tk1 ) zk (tk1 ))

dzk1 dzk1 = |zk1 zk1 |

d(zk1 zk1 ) = C|zk1 (tk2 ) zk1 (tk2 )| C.

Chapter 2. Preliminaries knot invariant Z by setting Z (K) = Z(K) Z( )c/2 ,

Chapter 3 The Kontsevich integral of knotted trivalent graphs

The naive extension

Chapter 3. The Kontsevich integral of knotted trivalent graphs

dzi dzi zi zi i=1

The good properties

Chapter 3. The Kontsevich integral of knotted trivalent graphs

Eliminating the divergence

The re-normalized integral Z1

Chapter 3. The Kontsevich integral of knotted trivalent graphs

Chapter 3. The Kontsevich integral of knotted trivalent graphs

zi (tv ) zi (tv ) . zi (ti1 ) zi (ti1 )

The locally active part on the other hand:

Chapter 3. The Kontsevich integral of knotted trivalent graphs

, which is nite and does not

1 k d(zik zik ) . ln ln w(tik ) k! zik zik

where C is the distance of the vertical strand from the

1 k d(w(tik )) ln ln w(tik ) . k! w(tik ) + 2C

1 k d(w(tik )) ln ln w(tik ) k! w(tik ) + 2C 1 k ln ln w(tik ) d(w(tik )). k!

Thus, it is enough to show boundedness for

ln ln w(tik ) d(w(tik )). Inte-

grating by parts, we nd that the value of this integral is

+ [w(tik )]tv = tik1

Chapter 3. The Kontsevich integral of knotted trivalent graphs

1 w(tik1 )(ln ln w(tik1 ))j + w(tik1 ). j!

1 w(tik1 )(ln ln w(tik1 ))j w(tik1 ), j!

The good properties

dln(etm w1 etm w2 )...dln(et1 w1 et1 w2 ) =

dtm dtm1 ...dt1 =

(t12 (T ))m , (2i)m m!

which proves the claim.

Chapter 3. The Kontsevich integral of knotted trivalent graphs

this (due to the factorization property), namely Z1 () = e