TREE DIAGRAMS FOR STRING LINKS

BLAKE MELLOR Mathematics Department Loyola Marymount University Los Angeles, CA 90045-2659 [email protected]

Abstract In previous work, the author dened the intersection graph of a chord diagram associated with a string link (as in the theory of nite type invariants). In this paper, we classify the trees which can be obtained as intersection graphs of string link diagrams.

Introduction

Intersection graphs of various kinds have an important role in graph theory [10]. Recently, intersection graph have also proved useful in topology, in the study of nite type knot invariants. The theory of nite type invariants allows us to interpret many important knot and link invariants in purely combinatorial terms, as functionals on spaces of chord diagrams. For knots, there is an obvious intersection graph associated with these diagrams, rst studied (in this context) by Chmutov, Duzhin and Lando [4]. In many cases, these graphs contain all the relevant information for the functionals coming from knot invariants [4, 7]. Some of the most important knot invariants, such as the Conway, Jones, Homy and Kauman polynomials, can be interpreted in terms of these intersection graphs [2, 8]. The machinery of nite type invariants and chord diagrams is easily extended to other topological objects, such as links, braids and string links. Recently, the author extended the notion of the intersection graph of the chord diagram to string links as well [9], presenting the possibility of reinterpreting invariants of string links in the same way we have reinterpreted knot invariants. An important part of this program is to understand which graphs can appear as intersection graphs in this context. For knots, this question was answered by (among others) Bouchet [3]; for string links, however, the class of graphs is much larger. As special cases, it contains the set of intersection graphs for knot diagrams (known to graph theorists as circle graphs [10]) and the set of permutation graphs, both of which have been studied extensively. This paper begins the process of classifying these intersection graphs with a very modest goal: to describe the (directed) trees which can arise as intersection graphs of chord diagrams for string links. We will not discuss the background of nite type invariants, instead looking at the relationship between the intersection graph and the chord diagram from a purely combinatorial viewpoint. For a discussion of how these diagrams arise in the theory of nite type invariants, see [1]. In Section 2 we will review the denitions of chord diagrams and intersection graphs for string links. In Section 3 we will classify the trees which arise as intersection graphs.

Acknowledgement: The author thanks Loyola Marymount University for supporting this work via a Summer Research Grant in 2003.

2
2.1

Preliminaries
Chord Diagrams

We begin by dening what we mean by a chord diagram. Since we are only considering chord diagrams which arise from string links, we will simply refer to these as chord diagrams, but the reader should be aware that these dier somewhat from the more usual chord diagrams for knots. Denition 1 A chord diagram of degree n with k components is a disjoint union of k oriented line segments (called the components of the diagram), together with n chords (unoriented line segments with endpoints on the components), such that all of the 2n endpoints of the chords are distinct. The diagram is determined by the orders of the endpoints on each component. In the theory of nite type knot and link invariants, these diagrams are usually organized into a graded vector space, considered modulo certain relations. For the purposes of this paper, however, this additional structure is irrelevant.

2.2

Intersection Graphs

The essential value of the intersection graph for knots (in which the chord diagram consists of chords in a bounding circle) is that it can detect when the order of two endpoints for dierent chords along the bounding circle is switched, since this changes the pair of chords from intersecting to non-intersecting or vice-versa. For chord diagrams for string links, the existence of a bottom and top for each component allows us to give a linear (rather than cyclic) ordering to the endpoints of the chords on each component, and so the notion of one endpoint being below another is well-dened. We want our intersection graphs to detect when this order is reversed. Denition 2 [9] Let D be a chord diagram with k components (oriented line segments, colored from 1 to k) and n chords. The intersection graph (D) is the labeled, directed multigraph such that: (D) has a vertex for each chord of D. Each vertex is labeled by an unordered pair {i, j}, where i and j are the labels of the components on which the endpoints of the chord lie. There is a directed edge from a vertex v1 to a vertex v2 for each pair (e1 , e2 ) where e1 is an endpoint of the chord associated to v1 , e2 is an endpoint of the chord associated to v2 , e1 and e2 lie on the same component of D, and the orientation of the component runs from e1 to e2 (so if the components are all oriented upwards, e1 is below e2 ). We count these edges mod 2, so if two vertices are connected by two directed edges with the same direction, the edges cancel each other. If two vertices are connected by a directed edge in each direction, we will simply connect them by an undirected edge. Examples of chord diagrams and their associated intersection graphs are given in Figure 1. Note that when the two chords have both endpoints on the same component i, our denition of intersection graph corresponds to the usual intersection graph for knots. Our denition also matches our intuition in the case of chord diagrams of two components, as shown in Figure 1. Note also that the total number of directed edges between a vertex v labeled {i, j} and a vertex w labeled {l, m} is given by the sum of the number of occurrences of i in {l, m} and the number of occurrences of j in {l, m}. In particular, if a vertex v has a label {i, i}, this number will be even (0, 2 or 4). Since we count directed edges modulo 2, this implies there is an (uncancelled) directed edge from v to another vertex w if and only if there is also an (uncancelled) directed edge from w to v. We will say that labeled directed multigraphs which have this property are semisymmetric.

{i,i}
{i,i}

{i,i}

{i,i}

{i,i}

{i,i}

{i,j} {i,j}

{i,j} {i,j}

=
{i,j} {i,j}
{i,j}

=
{i,j}

{i,j}
1 5 3 2 4 2 3 4 6

{i,i}
1

{j,k}
5 6

{k,k} {i,i} {i,j} {j,k}

{j,k}

Figure 1: Examples of intersection graphs for string links. In the last example the chords and vertices are numbered to clarify the bijection between them. Denition 3 A directed multigraph G, with each vertex labeled by a pair {i,j}, is semisymmetric if for every vertex v labeled {i,i}, and any other vertex w, there is a directed edge from v to w if and only if there is a directed edge from w to v.

Intersection Graphs and Trees

In this section we will give a characterization of the trees which arise as intersection graphs. For our purposes, a directed graph is a tree if it is an undirected tree - i.e. the graph is connected, and it has no cycles even if the orientations of the edges are ignored. We will rst consider chord diagrams with two components, and then the (easier) case of diagrams with more than two components.

3.1

2-component diagrams

In intersection graphs for diagrams with two components, labeled 1 and 2, there are three possible labels on the vertices: {1,1}, {2,2} and {1,2}. To begin with, we will consider marked, rather than labeled intersection

graphs. A vertex will be called marked if it is labeled {1,2}, and unmarked if it is labeled {1,1} or {2,2}. In other words, a vertex is marked if the corresponding chord has its endpoints on dierent components, and unmarked if the chord has both endpoints on the same component. The rst important observation is that not all marked trees are intersection graphs (otherwise the characterization would be rather trivial). Figure 2 gives an example of a marked tree which is not an intersection graph for any chord diagram on two components. The marked vertices are labeled by an asterisk (*).

*
Figure 2: A marked graph which is not an intersection graph To determine which marked trees are intersection graphs, we need to introduce the notion of light and heavy boughs of a tree. Denition 4 A bough of a vertex v in a marked tree T is a connected component of T \v (the graph which results from removing the vertex v and all edges incident to v). The bough is called light if it contains at most one marked vertex and this vertex (if present) is adjacent to v in T. Otherwise the bough is heavy. The main theorem of this section characterizes tree intersection graphs via boughs. Theorem 1 A marked tree T (i.e. T may or may not have marked vertices) is an intersection graph for a chord diagram on two components if and only if any vertex in T has at most 2 heavy boughs. Before we prove this theorem, we require two lemmas. Our rst lemma relates the notion of a light bough to the notion of a share of a chord diagram, introduced in [4]. Denition 5 A share of a chord diagram D is a subset S of the set of chords in D and two arcs A and B on the boundary components of D (A and B may be on the same or dierent components) such that every chord in S has both endpoints in A B and no other chord in D has an endpoint in A B. Lemma 1 Assume T is a marked tree which is the intersection graph for a chord diagram D, and v is a vertex in T. Further assume that T has at least one marked vertex. Then a bough of v is light if and only if the corresponding chords are a share in D. Proof: Recall that trees are connected graphs. We will rst assume that the chords corresponding to the vertices in a bough H of v form a share S in D, and show that H is light. We will show that any marked vertex in H must be adjacent to v. Working by contradiction, assume w is a marked vertex in H which is not adjacent to v. Since S is a share, all endpoints of the chords in S lie on two arcs A and B. Since the chord corresponding to w (which we will also call w) has an endpoint on each component of the chord diagram, we may assume that A and B lie on dierent components, and w has one endpoint in each. Since v is not in S, it has no endpoints in A or B. Since v and w are not adjacent, the endpoints of v must be either both above or both below the endpoints of w. In this case, the endpoints of v are both above or both below any endpoint in A B, so v is not adjacent to any vertex in H (since the directed edges in an intersection graph

are counted mod 2). But since T is connected, and H is a component of T \v, this is a contradiction, and we conclude that any marked vertex in H is adjacent to v. We still need to show that H has at most one marked vertex. Assume H has two marked vertices, w and u. By the argument above, these are both adjacent to v, so there is a path wvu in T . But, since H is a connected component of T \v, there is also a path from w to u in H which does not contain v. Joining these two paths gives us a cycle in T , which contradicts the assumption that T is a tree, so H can have at most one marked vertex. This proves H is a light bough of v. Now we assume that H is a light bough of v, and S is the set of chords corresponding to the vertices of H. Since H is connected, and T is a tree, there is a unique vertex w in H which is adjacent to v (otherwise, we will have a cycle as in the last paragraph). w may or may not be marked - in either case, no other vertices of H are marked. Our goal will be to show that we can nd minimal arcs A and B containing all the endpoints of chords in S (but not the endpoints of v), and these will be the arcs required for S to be a share. The rst step is to show that there exists some pair of arcs containing all the endpoints of chords in S, but not the endpoints of v. We rst consider the case when v is marked. Then v divides each of the two boundary components of the diagram into two pieces (giving a total of four segments), as shown in Figure 3. w connects two of these

w v

w v

or

Figure 3: v is marked four pieces, and . Since no other chords in S intersect v, and all other chords in S are unmarked, the other chords in S must have both endpoints in or both endpoints in . So and are arcs containing all the endpoints of S, but not the endpoints of v. If v is unmarked, it again divides the two boundary components into four pieces, three on one component and one on the other. Once again, w connects two of these pieces, and as shown in Figure 4. If w is

v w

w u

or
v

Figure 4: v is unmarked marked, then once again all chords of S must have both endpoints in or both in , and we can take these as the desired arcs. If w is also unmarked, there may be a chord u in S which connects to one of the other arcs, as shown in Figure 4. But this means there are no marked chords in the diagram: any marked chord would have to be connected to v (since the tree T is connected) by a path of chords which would at some point cross u, but this would mean S contains a marked chord, which is impossible. Since our lemma

assumes that T has at least one marked vertex, we can discard this case. If there is no such chord u, then we can take the arcs and as before. Now that we know there exists at least one pair of disjoint arcs on the boundary components of the chord diagram which contain all the endpoints of S but not the endpoints of v, let A and B be a minimal pair of such arcs - minimal in the sense of containing as few endpoints of chords in T as possible. Now say we have a chord u which is not in S. First assume that u has one endpoint in A B; without loss of generality, we may assume it lies in A. Since A B is minimal, there are endpoints of chords in S both above and below u in A (otherwise we could shrink A to exclude the endpoint of u, giving a smaller arc). Since H is connected, there is a path of chords connecting the endpoints above u to those below u. If the other endpoint of u is not in A, it must cross some chord in this path, but this means that u is in S, a contradiction. So the other endpoint of u must also be in A. By the same argument, any chord intersecting u must have at least one, and hence both, endpoints in A; in fact, all chords corresponding to vertices in the connected component of T containing u must have both endpoints in A. But since T is connected, this means that, in particular, v has both endpoints in A, a contradiction. We conclude that only chords in S have endpoints in A B, which shows that S is a share, and completes the proof of the lemma. 2 Note that if the tree T does not have any marked vertices, a vertex v may have a light bough which does not correspond to a share, as in Figure 5. However, in this case we are only looking at diagrams on a

Figure 5: A light bough which is not a share. single component. Bar-Natan [1] showed that this case is equivalent diagrams on knots, and if we change the bounding components to a Chmutov et al. [4] have already shown that all unmarked trees are restrict our attention to trees with at least one marked vertex. We now prove a lemma about the structure of trees where every We introduce the notion of the spine of a tree: (modulo the 4-term relation) to chord circle the boughs are shares [4]. Since intersection graphs, in this section we vertex has at most two heavy boughs.

Denition 6 Say that T is a marked tree. A spine of T is a path in T of maximal length such that the rst and last vertices in the path are marked. If T has a single marked vertex, that vertex is the spine; if it has no marked vertices the spine is empty. Lemma 2 If T is a marked tree where every vertex has at most two heavy boughs, and S is a spine for T , then every marked vertex in T which is not in S is adjacent to a vertex in the interior (i.e. not one of the endpoints) of S. Proof: Let v be a marked vertex in T which is not in S. Since T is connected, there is a path P connecting v to some vertex w in S (such that no other vertex on P is in S). Since S has maximal length, w must be an interior vertex of S, or the path P S would be longer than S. Also, P must be no longer than the path from w to either endpoint of S, or we could combine P with one of these paths to get a path between marked vertices longer than S. So if v is not adjacent to w, neither are either of the endpoints of S. But this means that w has three heavy boughs. So v must be adjacent to w, and we can conclude that every marked vertex of T which is not in S is adjacent to an interior vertex of S. 2 Proof of Theorem 1: We rst assume that T is an intersection graph, and show that any vertex v has at most two heavy boughs. We work by contradiction - assume v is a vertex with (at least) three heavy boughs; call these boughs C1 , C2 , C3 . So each Ci contains a marked vertex wi which is not adjacent to v.

Since w1 , w2 , w3 are non-adjacent marked vertices, the corresponding chords must be parallel, as shown in Figure 6. Without loss of generality, we order the chords w1 , w2 , w3 from top to bottom. Since T is a tree,

w1

w3

Figure 6: Parallel marked chords. there is a unique path from w1 to w3 , passing through v, and no chord in this path is adjacent to w2 (since the only chord in the path adjacent to any chord in C2 is v itself, and v is not adjacent to w2 ). But since w2 separates w1 and w3 , some chord in the path must cross w2 , which is a contradiction. So any chord v can have at most two heavy boughs. Now we will assume that any vertex v has at most two heavy boughs, and show that T is an intersection graph. Let S be a spine for T . S is just a path of marked and unmarked vertices, so S is the intersection graph for a chord diagram D, as shown in Figure 7. By Lemma 2, any marked vertex of T is adjacent to an
*

* *

Figure 7: Chord diagram for a path. interior vertex of S; let H be the union of S and the other marked vertices of T . Then we can add chords to D to produce a new diagram D with intersection graph H, as shown in Figure 8. All remaining vertices of T are unmarked. These vertices can be divided into connected components Ci , each of which is adjacent to one vertex of H. Let vi denote the vertex of Ci which is adjacent to H. By [4], Ci is the intersection graph for a barbell chord diagram consisting of the chord corresponding to vi and a share at either end, as shown in Figure 9. These diagrams can be attached to the desired chords of D (and any number can be attached to a single chord of D ) to obtain a chord diagram with intersection graph T , as in Figure 10. So T is an intersection graph, completing the proof. 2 We now turn our attention from marked trees to labeled trees, in which the marked vertices are labeled {1,2}, and the unmarked vertices are labeled either {1,1} or {2,2}, and ask when a labeled tree is an intersection graph. There are two obvious requirements: the underlying marked tree must be an intersection graph (so the tree must satisfy the condition of Theorem 1), and adjacent unmarked vertices must have the same label (since chords on dierent components of a chord diagram cannot intersect). However, these

* * * *

* * *

H
Figure 8: Expanding D to D
vi

D'

Figure 9: Chord diagram for Ci . conditions are not sucient, as shown by the example in Figure 11. We need to add one more condition, which is given by the following theorem: Theorem 2 A labeled tree T is an intersection graph for a chord diagram on two components if and only if it satises the following three conditions: The underlying marked tree is an intersection graph. Adjacent unmarked vertices have the same label. If S is a spine of T , and v and w are unmarked vertices in S, then the labels on v and w are the same if and only if there is an even number of marked vertices between them along S. Proof: We rst prove the suciency. If a labeled tree T satises the three conditions, then we can construct a chord diagram with intersection graph T as in Theorem 1. In particular, the third condition allows us to construct a chord diagram for the spine as in Figure 7. As for the necessity, it is clear that the rst two conditions are necessary for T to be an intersection graph. Assume T is an intersection graph for a chord diagram D. Let S be a spine of T , and v and w be two unmarked vertices along S. If there are no marked vertices between v and w, then they are connected by a path of unmarked vertices corresponding to chords in the chord diagram with both endpoints on the same component. So all of these chords must be on the same component, and the vertices must have the same label. Now assume there are marked vertices between v and w.

Figure 10: Adding barbells to a chord diagram

undesired intersection is forced {1, 2} {1, 1} {1, 2} {1, 1} {1, 2} 1

Figure 11: A labeled tree which is not an intersection graph. We will rst consider the case when all of the marked vertices between v and w along S are consecutive along S. Since the endpoints of the path S are marked vertices, there is a path m1 U1 vU2 M U3 wU4 m2 in S, where m1 and m2 are marked vertices, each Ui is a path of unmarked vertices, and M = x1 x2 ...xk is a path of k marked vertices. Since T is a tree, m1 and m2 are not adjacent to each other or to any of the vertices in M . If k = 1, we have the picture in Figure 12 (assuming, without loss of generality, that U1 vU2 is on component 1). Since the rst chord in U3 does not intersect any chord in U1 vU2 , it cannot have an endpoint on component 1 between m1 and x1 . However, since it is intersects x1 , it must have an endpoint on either side of x1 , and since it does not intersect m1 , one of these endpoints must lie between x1 and m1 . So this chord must have an endpoint on component 2. Since it is unmarked, it will have both endpoints on component 2, so the entire path U3 wU4 is on component 2, and v and w have dierent labels. If k > 1, the marked chords will cross each other as in Figure 13, which shows examples when k is even or odd. This arrangement is forced by the fact that each of the marked vertices is only adjacent to the ones immediately before and after it in the path M . The rst chord of U3 must intersect xk on the same side of xk1 as m2 . If k is odd, this means that the chords in U3 wU4 will be on component 2; if k is even, they will lie on component 1. So v and w will have the same label if and only if k is even. If the marked chords between v and w are not all consecutive, we can apply the argument above to each consecutive grouping. The labels of v and w will be the same if and only if the labels switch an even number of times, i.e. if there are an even number of stretches of marked vertices of odd length, which will happen exactly when the total number of marked chords in between v and w is even. This completes the proof. 2

v
x1

Figure 12: v and w separated by one marked chord.

m1
v v

m1

w m2 m2

1
k=4

1
k=5

Figure 13: v and w separated by odd or even number of marked chords.

3.2

Diagrams with more than 2 components.

Surprisingly, the directed trees which are intersection graphs for string link chord diagrams with more than 2 components is somewhat simpler than for diagrams with two components - we no longer need to consider heavy boughs. We will consider connected chord diagrams, meaning that we cannot separate the components into two nonempty sets so that all chords are between components in the same set. Our goal in this section is to prove the following theorem: Theorem 3 Let T be a connected, labeled directed tree with n colors (so each vertex has a label {i, j}, where 1 i, j n). Let mi,j denote the number of vertices with label {i, j}. Then T is an intersection graph for a connected chord diagram on n components if and only if the following conditions are met (possibly after relabeling the tree by a permutation of 1,..., n): 1. The labels of adjacent vertices must have at least one color in common. 2. T is semisymmetric (see Denition 3).

3. If v has label {i,j} and w has label {i,k}, where i, j and k are all distinct, then there is a directed edge between v and w. 4. mi,j = 0 if |i j| > 1. 5. mi,i+1 = 1 for 2 i n 2, m1,2 1 and mn1,n 1. 6. No two marked vertices are connected by a path of undirected edges. Remark: It is not intuitive that the theorem for many components should be simpler than that for two components - why could we not augment a 2-component chord diagram to a 3-component diagram by adding an additional component with a single chord connecting it to the previous diagram? The problem is that if the original diagram had a heavy bough (and so two marked chords connected by a path of unmarked chords), then the new diagram would not be a tree diagram (since the new chord would be adjacent to both of the marked chords in the heavy bough, completing a cycle). So heavy boughs are no longer permitted (this is one eect of Condition 6), and the diagrams are simpler. Proof: We will maintain the convention of the previous section, and refer to a vertex as marked if it is labeled by two separate colors and unmarked if it is labeled by only one color. We will rst show that these conditions are all necessary. Assume that a tree T is an intersection graph for a connected string link chord diagram D. Condition 1 is obvious, since two chords cannot intersect (in the sense of Denition 2) unless they have at least one endpoint on the same component. Conditions 2 and 3 are consequences of the discussion at the end of Section 2.2. To show the necessity of Condition 4, we will use the connection graph of the diagram D. Denition 7 Given a string link chord diagram D with n components (labeled 1, 2,..., n), the connection graph of D is the graph C(D) with n vertices (labeled 1, 2,..., n) where vertices i and j are adjacent in C(D) if and only if there is a chord connecting components i and j in D. We rst observe that if T = (D) is a tree, then so is C(D). If there is a cycle i1 i2 ...ik in C(D), then there are chords in D connecting components i1 and i2 , i2 and i3 ,..., ik and i1 . So T has vertices labeled {i1 , i2 }, {i2 , i3 },..., {ik1 , ik }, {ik , i1 }. Call these vertices v1 ,..., vk . Then, by Condition 3, each vi is adjacent to vi+1 (and vk is adjacent to v1 ), so there is a cycle v1 v2 ...vk in T . This is impossible, since T is a tree, so C(D) must also be a tree. Since D is a connected diagram, every vertex of C(D) has degree at least one. Moreover, no vertex of C(D) can have degree greater than two. Otherwise, there would be a vertex i in C(D) adjacent to three other vertices j, k, l. So there would be three chords in D with one endpoint on component i and the other on component j, k, l (respectively). This means there are corresponding vertices in T = (D) with labels {i, j}, {i, k} and {i, l}. By Condition 3, these three vertices are all adjacent to each other, meaning that T will have a 3-cycle. Since T is a tree, this is impossible. So C(D) is a connected tree where every vertex has degree at most two; this means that C(D) is a path. Starting at one end of the path, we renumber the vertices (and hence the components of D) from 1 to n. We will now have no chords between components i and j in D if |i j| > 1, which means mi,j = 0, giving us Condition 4. Moreover, we have that mi,i+1 1 for 1 i n 1. For the fth condition, assume that mi,i+1 > 1 for some i with 2 i n 2. So T will have at least two vertices v and w labeled {i, i + 1}. Since mi1,i 1 and mi+1,i+2 1, we also have vertices s and t labeled {i 1, i} and {i + 1, i + 2}, respectively. By Condition 3, s and t are both adjacent to both v and w, so we have a cycle svtw in T . Since T is a tree, this is impossible, so mi,i+1 = 1 for 2 i n 2, giving us Condition 5. For the nal condition, assume that two marked vertices v and w are connected by an undirected path vu1 u2 ...ur w (r 1 by Condition 3). Then all the unmarked vertices in the path have the same label, say {i, i}, which is one of the labels of v and w. If v and w share only one label, there is also a directed edge between them, giving a cycle in the graph T . Since T is a tree, this is impossible, so v and w must share both

their labels, {i, i 1}. By Condition 5, this label is either {1, 2} or {n 1, n}. Without loss of generality, assume the label is {1, 2}. Then there is a vertex x labeled {2, 3} which is adjacent to both v and w by directed edges, giving a cycle vu1 u2 ...ur wxv in T . Again, since T is a tree, this is impossible, which proves the nal condition. So any tree that is an intersection graph will satisfy the six conditions of the theorem. It remains to show that these six conditions are sucient. Let T be a connected, labeled directed tree with n colors which satises the conditions of the theorem. If necessary, relabel the vertices so Conditions 4 and 5 are satised. We rst consider the subgraph S of T induced by the marked vertices; S will still be connected by Conditions 3 and 5. We will rst construct a chord diagram with intersection graph S. Begin with an empty chord diagram with n components (all oriented upwards). Add m1,2 parallel chords between components 1 and 2. (Note that no two vertices labeled {1, 2} are adjacent, or they would form a 3-cycle with the unique, by Condition 5, vertex with lable {2, 3}.) If v is the unique vertex in S with label {2, 3}, there are k vertices labeled {1, 2} with an edge directed towards v, and m1,2 k such vertices with an edge directed away from v. Place a chord between components 2 and 3 so that the endpoint on component 2 is above k of the chords between components 1 and 2, and below the rest. We now place a chord between components 3 and 4, which is above or below the chord between components 2 and 3 depending on the direction of the edge in S. This corresponds to the unique (by Condition 5) vertex in S with label {3, 4}. Continue in this way through all the components, nally putting mn1,n parallel chords between components n 1 and n. The result is a chord diagram with intersection graph S. See Figure 14 for an example. Now we simply
{1,2} {1,2} {2,3} {3,4} {4,5} {5,6} {5,6} {5,6}

Figure 14: A chord diagram with intersection graph S need to add chords corresponding to the unmarked vertices in T . By Condition 6, no marked vertices are connected by paths of unmarked vertices. So the unmarked branches of the tree correspond to barbell chord diagrams, as in Figure 9, and these diagrams can be attached to the chord diagram for S as in Figure 10. The result is a chord diagram with intersection graph T , which completes the proof of Theorem 3. 2

References
[1] D. Bar-Natan, On the Vassiliev Knot Invariants, Topology 34, No. 2, pp. 423-472, 1995 [2] D. Bar-Natan and S. Garoufalidis, On the Melvin-Morton-Rozansky Conjecture, Invent. Math. 125, pp. 103-133, 1996 [3] A. Bouchet, Circle Graph Obstructions, J. Comb. Theory, Series B, 60, p. 107-144, 1994 [4] S. V. Chmutov, S.V. Duzhin, and S. K. Lando, Vassiliev Knot Invariants I, II, III, Advances in Soviet Mathematics 21, pp. 117-145, 1994 [5] M. Kontsevich, Vassilievs knot invariants, Advances in Soviet Mathematics 16, Part 2, pp. 137-150, 1993 [6] S. Lando, On a Hopf algebra in graph theory, J. Comb. Theory, Series B, 80, p. 104-121, 2000

[7] B. Mellor, The Intersection Graph Conjecture for Loop Diagrams, J. of Knot Theory and its Ramications 9, No. 2, 2000 [8] B. Mellor, A few weight systems arising from intersection graphs, Michigan Math. J., Vol. 51, No. 3, pp. 509-536, 2003. Also see arXiv:math.GT/0004080 [9] B. Mellor, Intersection Graphs for String Links, to appear in J. of Knot Theory and its Ramications, 2004, arXiv:math.GT/0312347 [10] T.A. McKee and F.R. McMorris, Topics in Intersection Graph Theory, SIAM, Philadelphia, 1999