Gordian split links in the Gehring ropelength problem Friedrich Bauermeister Abstract A thick link is a link in ℝ 3 \mathbb{R}^{3} such that each component of the link lies at distance at least 1 1 from every other component. Strengthening the notion of thickness, we define a thickly embedded link to be a thick link whose open radius- 1 2 \tfrac{1}{2} normal disk bundles of all components are embedded. The Gehring ropelength problem asks how large the sum of the lengths of the components of a thick (respectively thickly embedded) link must be, given the link homotopy (respectively isotopy) class of the link. A thick homotopy (isotopy) is a link homotopy (isotopy) of a thick (thickly embedded) link that preserves thickness throughout, and such that during the homotopy the total length of the link never exceeds the initial total length. These notions of thick homotopy and isotopy are more permissive than other notions of physical link isotopies in which the length of each individual component must remain constant (no “length trading”). We construct an explicit example of a thickly embedded 4-component link which is topologically split but cannot be split by a thick homotopy, and thick links in every homotopy class with 2 components that are non-global local minima for ropelength. This is the first time such local minima for ropelength have been explicitly constructed. In particular, we construct a thick 2-component link in the link homotopy class of the unlink which cannot be split through a thick homotopy. 1 Introduction In 2015, Coward and Hass [ 5 ] defined the notion of a physical link isotopy. A physical link isotopy is a link isotopy that keeps a link thickly embedded throughout, as well as keeping the length of each component constant. This made it possible to study a physical knot- and link theory that corresponds to knots and links tied from real pieces of rope. In the same paper, Coward and Hass construct a 2 -component Gordian split link which they define as a thickly embedded, topologically split link that cannot be split by a physical link isotopy. More generally, a Gordian pair of links is a set of two links in \mathbb{R}^{3} which are "topologically equivalent", but which are not equivalent under "physical deformations". Here "topologically equivalent" refers either to being in the same link isotopy or homotopy class, and "physical deformation" refers to either link homotopy or link isotopy, subject to additional geometric constraints on length and thickness. For length, there are two common constraints: either demanding that the length of each component individually does not increase throughout the homotopy/isotopy, or that their sum does not increase throughout the homotopy/isotopy. If only the sum of lengths is constrained, we say that "length trading" between components is permitted. The thickness constraint usually demands that links stay thick links (see Definition 3.3 ) or thickly embedded links (see Definition 3.4 ) throughout the homotopy or isotopy respectively. When considering link isotopy, sometimes additional curvature constraints are imposed. The original Gehring ropelength problem asked for the minimal length of a closed curve in \mathbb{R}^{3} , given that it is linked to another curve from which it keeps at least unit distance. The answer was quickly shown to be 2\pi , realized by two Hopf-linked unit circles. Given some link homotopy class, the general Gehring ropelength problem [ 3 ] asks for the minimal ropelength of a thick link in that homotopy class. A link realizing this minimal length is called an ideal link of the homotopy class. While the precise length of ideal links is known only in a few cases, there are various upper and lower bounds. Cantarella, Fu, Kusner, Sullivan, and Wrinkle [ 3 ] defined a notion of criticality for the general Gehring ropelength problem, which is a necessary but not sufficient criterion for being an ideal link. There are good reasons to study thick homotopies even though they are the least physical of the "physical deformations". First, once it is proven that a pair of "topologically equivalent" links does not have a thick homotopy between them, the pair of links will a fortiori be a Gordian pair in all other reasonable notions of physicality. Secondly, the lack of thick homotopies between two links in the same link homotopy class is related to local minima in the general Gehring ropelength problem: Two links in a given link homotopy class, one minimizing ropelength globally and one only locally, should not have a thick homotopy between them. This will be the strategy we use to construct Gordian pairs in every link homotopy class of two components. 2 Overview of existing results about Gordian pairs Isotopically Gordian Pairs Any Gordian pair Split link 2 -component split link 2 -component unlink Without length trading and with extra curvature constraints Coward and Hass [ 5 ] Coward and Hass [ 5 ] Coward and Hass [ 5 ] Ayala [ 1 ] Without length trading Coward and Hass [ 5 ] Coward and Hass [ 5 ] Coward and Hass [ 5 ] Ayala and Hass [ 2 ] , Kusner and Kusner [ 9 ] With length trading Kusner and Kusner [ 8 ] Section 5 With length trading and allowing homotopies instead of isotopies Kusner and Kusner [ 8 ] Section 5 Homotopically Gordian Pairs Any Gordian pair Split link 2 -component unlink All 2 -component links Without length trading Kusner and Kusner [ 8 ] Section 5 , Section 6 Section 6 Section 6 With length trading Kusner and Kusner [ 8 ] Section 5 , Section 6 Section 6 Section 6 Table 1: Summary of known examples for Gordian pairs in different settings. The first Gordian unlink was constructed by Coward and Hass in 2015 [ 5 ] . In 2023, Kusner and Kusner [ 8 ] constructed a Gordian pair of thickly embedded 7 -component links. The links in the pair are link isotopic to each other, but there is no thick homotopy between them. So even though the links are equivalent in the strongest topological sense, they are not equivalent even in the weakest "physical" sense. It was Kusner and Kusner who first coined the term Gordian pair . In [ 1 ] Ayala constructs a thickly embedded 2 -component link in the isotopy class of the unlink which cannot be split by thick isotopy under additional curvature constraints. This was the first example of a Gordian unlink. Recently, Ayala and Hass constructed a different example of a Gordian unlink without needing any additional curvature constraints. Independently, Kusner and Kusner announced the construction of a Gordian unlink [ 9 ] . Table 1 summarizes both the existing results on Gordian pairs and the author’s contributions in this paper. The table is split into two parts, considering links with respect to isotopy and homotopy respectively. The columns of each table denote what kind of example was constructed. The rows of each table denote which geometric constraints were considered, with the strength of constraints being relaxed from top to bottom. The paper referenced in each cell is the paper in which the first example of the corresponding Gordian pair was constructed. The tables give partial orders: the further down and to the right a cell in a given table is, the harder it is to construct an example for it. 3 Preliminaries 3.1 Definition . Let A,B\subseteq\mathbb{R}^{n} be non-empty sets, let x\in\mathbb{R}^{n} and let f,g be functions with image in \mathbb{R}^{n} . Then we denote the distance of A to B with \mathrm{dist}(A,B)\coloneqq\inf\{||a-b||\mid a\in A,b\in B\} ; and we define \mathrm{dist}(f,g) , \mathrm{dist}(f,A) , \mathrm{dist}(x,A) , \mathrm{dist}(x,f) by identifying x with \{x\} and f,g with their image. If f,g:M\to\mathbb{R}^{n} then ||f-g||\coloneqq\sup\{||f(x)-g(x)||\mid x\in M\} . 3.2 Definition . If L=\bigcup_{i=1}^{n}C_{i} is a link, then we define its ropelength as the sum of the lengths of its components. \mathrm{len}(L)=\sum_{i=1}^{n}\mathrm{len}(C_{i}). 3.3 Definition . Let L=\left(C_{1},\dots,C_{n}\right):S_{1}\sqcup\dots\sqcup S_{n}\to\mathbb{R}^{3} be a link. We say that L has thickness d if \mathrm{dist}(C_{i},C_{j})\geq d for every i,j\in\{1,\dots,n\},i\neq j . We define \tau(L) as the largest such d . If \tau(L)\geq 1 we call L a thick link . 3.4 Definition . We say that L is d -thickly embedded if L is \mathcal{C}^{1} and if the open radius- \tfrac{d}{2} normal disk bundle around L is embedded. We define \tau_{e}(L) as the largest such d . If \tau_{e}(L)\geq 1 we call L a thickly embedded link . 3.5 Definition . Let H:[0,1]\times\left(S_{1}\sqcup S_{2}\sqcup\dots\sqcup S_{n}\right)\to\mathbb{R}^{3} be a link homotopy (isotopy). We call H a d -thick homotopy (isotopy) if for all t\in[0,1] : 1. L_{t}=H(t,\cdot) has thickness d (is d -thickly embedded), 2. \mathrm{len}(L_{t})\leq\mathrm{len}(L_{0}) . If d=1 we call H a thick homotopy (isotopy). 3.6 Definition . For thick links L_{1} and L_{2} we define \displaystyle L_{1}\preceq L_{2} \displaystyle\iff\text{ There is a thick link homotopy from }L_{2}\text{ to }L_{1}. \displaystyle L_{1}\simeq L_{2} \displaystyle\iff L_{1}\preceq L_{2}\text{ and }L_{2}\preceq L_{1} For thickly embedded links L_{1} and L_{2} we define \displaystyle L_{1}\preceq_{e}L_{2} \displaystyle\iff\text{ There is a thick link isotopy from }L_{2}\text{ to }L_{1}. \displaystyle L_{1}\simeq_{e}L_{2} \displaystyle\iff L_{1}\preceq_{e}L_{2}\text{ and }L_{2}\preceq_{e}L_{1}. If L_{1}\preceq L_{2} or L_{1}\preceq_{e}L_{2} we say that L_{1} is reachable from L_{2} . 3.7 Proposition . If L_{1}\preceq L_{2} , then \mathrm{len}(L_{1})=\mathrm{len}(L_{2})\iff L_{2}\preceq L_{1}\iff L_{1}\simeq L_{2}, and the same for \preceq_{e} . Proof. Only the implication \mathrm{len}(L_{1})=\mathrm{len}(L_{2})\implies L_{2}\preceq L_{1} is interesting. Let H_{t} be the thick homotopy (isotopy) from L_{1} to L_{2} , then H_{1-t} is a thick homotopy (isotopy) from L_{1} to L_{2} . ∎ 3.8 Definition . A pair of thick (thickly embedded) links (L_{1},L_{2}) is called Gordian as thick links (as thickly embedded links) if L_{1} and L_{2} are in the same link homotopy (isotopy) class, but neither is reachable from the other. We will refer to a pair that is Gordian as thick links as homotopically Gordian , and to a pair that is Gordian as thickly embedded link as isotopically Gordian . When it is clear from context if we are working in the setting of thick links or thickly embedded links, we will simply call such a pair of links a Gordian pair . 3.9 Definition . A thick (thickly embedded) link L=\bigcup_{i\in I}C_{i} is called separated if there is some \emptyset\subsetneq J\subsetneq I such that \mathrm{conv}\left(\bigcup_{i\in J}C_{i}\right)\cap\mathrm{conv}\left(\bigcup_{i\in I\setminus J}C_{i}\right)=\emptyset. A thick (thickly embedded) link L is called split as a thick (thickly embedded) link if there is a thick (thickly embedded) link homotopy (isotopy) from L to some separated link L^{\prime} . 3.10 Definition . A thick (thickly embedded) link L is called a homotopically (isotopically) Gordian split link if L is topologically split through a link homotopy (isotopy), but not split as a thick (thickly embedded) link. It is easy to see that every Gordian split link L is part of a Gordian pair. If we can write L=L_{1}\cup L_{2} where L_{1} and L_{2} are not topologically linked with each other, define L^{\prime}=L_{1}\cup T(L_{2}) for some translation T sufficiently far away from L_{1} . Clearly L and L^{\prime} are in the same link homotopy (isotopy) class. But since L was a Gordian split link, we have L^{\prime}\npreceq L . Since \mathrm{len}(L)=\mathrm{len}(L^{\prime}) , we have L\npreceq L^{\prime} . Hence (L,L^{\prime}) is a Gordian pair. 3.11 Remark . For the link homotopy class with one component, we have L_{1}\preceq L_{2}\iff\mathrm{len}(L_{1})\leq\mathrm{len}(L_{2}). Therefore, no Gordian pair can exist in the 1 -component link homotopy class. 3.12 Remark . It is not known whether there is a Gordian pair of thickly embedded knots. There is some computational evidence for the existence of a Gordian unknot, i.e., a thickly embedded knot in the isotopy class of the unknot which has no thick isotopy to a circle [ 10 ] . 3.13 Definition . We call a thick link L a global minimum for ropelength if for every thick link L^{\prime} in the same link homotopy class as L , we have \mathrm{len}(L^{\prime})\geq\mathrm{len}(L). 3.14 Definition . We call a thick link L a local minimum for ropelength if there is some \varepsilon>0 such that for every thick link L^{\prime} ||L-L^{\prime}||<\varepsilon\implies\mathrm{len}(L^{\prime})\geq\mathrm{len}(L). Global and local minima of ropelength are interesting in and of themselves, and much is still unknown about them. For example, up until now, the only local minima proven to exist were also global minima. In [ 3 ] , Cantarella, Fu, Kusner, Sullivan, and Wrinkle use their criterion of ropelength criticality to construct suspected local and global minimizers in different link homotopy classes, for example for the Borromean rings. When studying thick homotopies, there is another interesting kind of minimization. 3.15 Definition (sink of ropelength) . If a thick link L is a minimum of the preorder \preceq , then we call L a sink of the ropelength functional. Sinks are exactly the configurations which cannot be shortened by a thick homotopy. We briefly define the notion of a sink more generally, and see how it relates to the notion of local minima. 3.16 Definition . Let X be a topological space, f:X\to\mathbb{R} a (not necessarily continuous) function. We call x\in X a sink of f if for all continuous \gamma:[0,1]\to X with \gamma(0)=x , we have (f\circ\gamma)(t)\leq f(x)\text{ for all t}\implies f\circ\gamma\text{ constant. } It is easy to see that every global minimum and every strict local minimum is a sink. However, even for "nice" spaces X and continuous functions f , there are local minima which are not sinks and sinks which are not local minima. 3.17 Example . Take X=\mathbb{R} and f(x)=\min(x,0) . Then x=1 is a local minimum, but not a sink. 3.18 Example . Take X=\mathbb{R} and f(x)=\begin{cases}x\sin\left(\frac{1}{x}\right),&x\neq 0\\ 0,&x=0\end{cases} Then x=0 is a sink but not a local minimum. 3.19 Definition . Let X be a topological space and f:X\to\mathbb{R} . We say that a subset S\subseteq X is a strictly locally minimizing set of f if f is constant on S and if for all x\in S , there is an open neighborhood U\subseteq X of S such that y\in U\implies f(y)\geq f(x) with equality only if y\in S . 3.20 Proposition . Let X be a topological space, f:X\to\mathbb{R} , and let S be a strictly locally minimizing set of f . If S\subseteq X is closed, then every x\in S is a sink of f . Proof. Let c=f(x) , let \gamma:[0,1]\to X with \gamma(0)=x and with f(\gamma(t))\leq c for all t\in[0,1] . Let J=\gamma^{-1}(S) . Clearly 0\in J , so J\neq\emptyset . Further, J is the preimage of a closed set under a continuous map, hence closed. Next we show that J is open. Let t\in J . Since \gamma(t)\in S we have f(\gamma(t))=c . Further there is a neighborhood U_{t} around t with t^{\prime}\in U_{t}\implies c=f(\gamma(t))\leq f(\gamma(t^{\prime})) and with equality if and only if \gamma(t^{\prime})\in S . By assumption f(\gamma(t^{\prime}))\leq c . Therefore f(\gamma(t^{\prime}))=c and hence \gamma(t^{\prime})\in S . This shows that J is open. Since J is a non-empty, closed and open subset of [0,1] we have J=[0,1] . So \gamma([0,1])\subseteq S\subseteq f^{-1}(c) , hence f\circ\gamma is constant. ∎ 4 Bounds on ropelength 4.21 Theorem (Original Gehring Problem) . Let C_{1} and C_{2} be two linked closed curves in \mathbb{R}^{3} with \mathrm{dist}(C_{1},C_{2})\geq 1 . Then \mathrm{len}(C_{1})\geq 2\pi . An elegant proof of this theorem was found by Marvin Ortel, and was first published in [ 3 ] . The statement can be strengthened and generalized in various ways. We will give a few of the most useful generalizations and some important consequences. 4.22 Definition . Let C:S^{1}\to\mathbb{R}^{3} be a closed curve, and let p\in\mathbb{R}^{3} . We define the p -cone of C as the disk K:\left(S^{1}\times[0,1]\right)/_{\sim}\to\mathbb{R}^{3} , K(x,y)=y\cdot C(x)+(1-y)\cdot p . Here (x,y)\sim(x^{\prime},y^{\prime})\iff y=0=y^{\prime} . If p is the center of mass of C , we call K the center-of-mass spanning disk of C . 4.23 Theorem . Let C be a closed curve, p a point in the convex hull of C , K the p -cone of C . Let r\geq 0 and let x_{1},x_{2},\dots,x_{n}\in K be the vertices of a convex polygon P\subseteq K , such that \mathrm{dist}(x_{i},C)\geq r for all i=1,\dots,n . Then \mathrm{len}(C)\geq 2r\pi+\mathrm{circumference}(P). In [ 5 ] , Coward and Hass prove a version of this theorem for the case where C is a \mathcal{C}^{1} curve, where n=2 , and where r=2 . Their proof easily generalizes to a proof for the statement with arbitrary n and r . By approximating arbitrary rectifiable curves by \mathcal{C}^{1} curves, we get our desired result. For n=2 , we get the following version of this theorem. 4.24 Proposition . Let C be a closed curve, p\in\mathrm{conv}(C) , K the p -cone of C . Let r\geq 0 and let x,y\in K be such that \mathrm{dist}(x,C),\mathrm{dist}(y,C)\geq r . Then \mathrm{len}(C)\geq 2\pi r+2||x-y||. We will use this result for r=1 or for r=1-\varepsilon for small \varepsilon . 4.25 Proposition . Let C be a closed curve, p a point in the convex hull of C , K the p -cone of C . Let A\subseteq K be such that \mathrm{dist}(A,C)\geq 1 . Then \mathrm{len}(C)\geq 2\pi+2\mathrm{diam}(A). We can now use these results to give some lower bounds for the lengths of curves in a thick link that are linked to a certain number of other curves. We already know from the original Gehring problem that if (C_{1},C_{2}) is a thick link with C_{1} and C_{2} topologically linked, then \mathrm{len}(C_{1})\geq 2\pi . It is natural to ask about the general case: If (C_{1},C_{2},\dots,C_{n+1}) is a thick link with C_{1} topologically linked to each of C_{2},\dots,C_{n+1} , what is the minimal length that C_{1} can have? This question was answered by Cantarella, Kusner, and Sullivan in [ 4 ] . Their formulation was originally only for the case of thickly embedded links, but it holds just as well for thick links. 4.26 Theorem . Let (C_{1},C_{2},\dots,C_{n+1}) be a thick link such that C_{1} is topologically linked to each of C_{2},\dots,C_{n+1} . Then \mathrm{len}(C_{1})\geq 2\pi+Q_{n} , where Q_{n} is the length of the shortest closed curve in \mathbb{R}^{2} surrounding n points of at least unit distance to each other. Here Q_{1}=0 , Q_{2}=2 , Q_{3}=3 , Q_{4}=4 . Asymptotically, Q_{n}\sim\sqrt{n} [ 4 ] . If L=(C_{1},\dots,C_{n}) is a thick link such that each curve C_{i} of L is topologically linked to k_{i} curves, then \mathrm{len}(L)\geq 2n\pi+\sum_{i=1}^{n}Q_{k_{i}} . If \mathrm{len}(L)=2n\pi+\sum_{i=1}^{n}Q_{k_{i}} , then L is a global minimum for ropelength in its link homotopy class. Until now, these are the only configurations proven to be global minimizers. Some examples and graphics of such global minimizers can be found in [ 3 ] . 4.27 Proposition . Let C be a closed curve, p a point in the convex hull of C , K the p -cone of C . Let 0\leq\varepsilon\leq 1 and let x,y\in\mathbb{R}^{3} be such that \mathrm{dist}(x,C),\mathrm{dist}(y,C)\geq 1 and \mathrm{dist}(x,K),\mathrm{dist}(y,K)\leq\varepsilon Then \mathrm{len}(C)\geq 2\pi+2||x-y||-(2\pi+4)\varepsilon. Proof. Let x^{\prime},y^{\prime}\in K with ||x^{\prime}-x||,||y^{\prime}-y||\leq\varepsilon . Then \mathrm{dist}(x^{\prime},C),\mathrm{dist}(y^{\prime},C)\geq 1-\varepsilon. Hence by Proposition 4.24 \displaystyle\mathrm{len}(C)\geq \displaystyle 2\pi(1-\varepsilon)+2||x^{\prime}-y^{\prime}|| \displaystyle\geq \displaystyle 2\pi(1-\varepsilon)+2||x-y||-4\varepsilon \displaystyle= \displaystyle 2\pi+2||x-y||-(2\pi+4)\varepsilon. ∎ 5 An isotopically Gordian split link exists The construction of our Gordian split link, as well as the basic structure of the proof, has a very similar flavor to that of Coward and Hass [ 5 ] . Both approaches derive a lower bound for the length needed to split a thickly embedded link by assuming that there is a thick homotopy/physical isotopy, tracking the intersections of a center-of-mass spanning disk and a curve throughout that homotopy/isotopy, and finally arguing that there needs to be some point in time where the intersections yield some lower bound through applications of propositions from section 4 . The main difference is in the "book-keeping" of how intersections are tracked. Coward and Hass need to resolve multiple kinds of singularities that might occur during the tracking of intersections. Our arguments use the homotopy invariance of the Brouwer degree and the mountain climbing theorem instead. We start out by stating the main proposition of this chapter, which establishes a lower-bound for the length necessary to split some thick links. 5.1 A lower bound on ropelength needed to split some thick links 5.28 Proposition . Consider a thick link L consisting of four curves, a main curve M linked to two auxiliary curves A^{1} and A^{2} , and a fourth curve C which we will call the center curve. Let D be the center-of-mass spanning disk of C . Assume the following: 1. D and M are \mathcal{C}^{1} and they intersect in exactly two points, i.e. there are exactly two pairs of elements (a,v),(b,w)\in S^{1}\times D^{1} such that M(a)=D(v),M(b)=D(w) , 2. these intersections are transverse, 3. and a\neq b . We get that a and b partition S^{1} into two parts, say I^{1} and I^{2} . Let M^{1} be the concatenation of M|_{I_{1}} with the straight line segment connecting M(a) and M(b) and let M^{2} be the concatenation of M|_{I_{2}} and the straight line segment connecting M(a) and M(b) . Further assume: 1. M^{1} is linked to A^{1} , 2. M^{2} is linked to A^{2} . Then, if H_{t} is a homotopy of L preserving thickness throughout, and such that \mathrm{conv}(C_{1}) and \mathrm{conv}(M_{1}) are disjoint, we have \sup_{t\in[0,1]}\left(\mathrm{len}(C_{t})\right)\geq 2\pi+4 . Therefore \displaystyle\sup_{t\in[0,1]}\mathrm{len}(H_{t})= \displaystyle\sup_{t\in[0,1]}\left(\mathrm{len}(M_{t})+\mathrm{len}(A^{1}_{t})+\mathrm{len}(A^{2}_{t})+\mathrm{len}(C_{t})\right) \displaystyle\geq \displaystyle\left(2\pi+2\right)+2\pi+2\pi+\sup_{t\in[0,1]}\mathrm{len}(C_{t}) \displaystyle\geq \displaystyle 8\pi+6, where we used that \mathrm{len}(M_{t})\geq 2\pi+2 and \mathrm{len}(A^{1}_{t}),\mathrm{len}(A^{2}_{t})\geq 2\pi by Theorem 4.21 and Theorem 4.26 . In particular, if \mathrm{len}(L)<8\pi+6 , there can be no thick homotopy splitting L . Figure 1: This plot shows the main curve M in black, the two auxiliary curves A^{1} and A^{2} in green and the center curve C in blue. The dotted line joins M(a) and M(b) . Assume we have some homotopy H_{t} , t=[0,1] , preserving thickness throughout, which splits the link in this figure. As the center curve C_{t} varies continuously, so does the center-of-mass spanning disk D_{t} of C_{t} . Imagine there are continuous functions t\mapsto a_{t},b_{t}\in S^{1} such that a_{0}=a , b_{0}=b , such that M_{t}(a_{t}),M_{t}(b_{t})\in D_{t} and such that for some T\in[0,1] we have a_{T}=b_{T} . Then a_{t} and b_{t} split S^{1} into intervals I^{1}_{t} and I^{2}_{t} . We can define continuously varying families of curves M^{1}_{t} and M^{2}_{t} as concatenations of M|_{I^{1}_{t}} and M|_{I^{2}_{t}} with the straight line segment from M_{t}(a_{t}) to M_{t}(b_{t}) respectively. Either M^{1}_{T} or M^{2}_{T} becomes a single point, while the other one becomes all of M . Without loss of generality assume M^{1}_{T} becomes a single point. Then (M^{1}_{T},A^{1}_{T}) becomes unlinked. But (M^{1},A^{1}) was linked, so at some point in time t\in[0,T] , M_{t}^{1} and A_{t}^{1} have to intersect. This intersection cannot be an intersection between M_{t} and A^{1}_{t} , so it has to be an intersection of the straight line segment with A_{t} . Since the endpoints of that straight line segment are on M_{t} , hence have distance at least 1 from A_{t} , we can conclude that the straight line segment at time t has length at least 2 . This means that M_{t}(a_{t}) and M_{t}(b_{t}) are at distance at least 2 from each other, and since they are both in the center-of-mass spanning disk of C_{t} , we get \mathrm{len}(C_{t})\geq 2\pi+4 by Proposition 4.25 . ∎ The main step omitted in this sketch is how, given a homotopy H preserving thickness, we can continuously keep track of the intersection points a_{t} and b_{t} , and why we can assume that they merge eventually. In this form the claim is wrong, but we will see that it can be replaced with a more technical, true claim in the same spirit. The first tool we need to use for our book-keeping will be the homotopy invariance of the Brouwer degree of a function around 0 . We will therefore briefly recall the following existence and uniqueness theorem defining the degree. This formulation is taken from Fixed Point Theory by Granas and Dugundji [ 6 ] . 5.29 Theorem (Brouwer degree around 0 ) . Let \mathcal{M}=\{(U,F)\mid U\subseteq\mathbb{R}^{n}\text{ open, bounded },F:\overline{U}\to\mathbb{R}^{n}\text{ continuous},0\notin F(\partial U)\}. Then there exists a unique function d:\mathcal{M}\to\mathbb{Z} such that 1. (Normalization) If 0\in U , then d(U,\mathrm{id}_{\overline{U}})=1 . 2. (Additivity) If F^{-1}(0)\subseteq U_{1}\cup U_{2}\subseteq U with U_{1},U_{2} open and disjoint, then d(U,F)=d(U_{1},F)+d(U_{2},F) . 3. (Homotopy Invariance) If F_{t} is a homotopy of functions such that for every t we have 0\notin F_{t}(\partial\Omega) , then the function t\mapsto d(U,F_{t}) is constant. We further need the following two facts about the Brouwer degree. If F:U\to\mathbb{R}^{n} is \mathcal{C}^{1} and if 0 is a regular value of F , then d(U,F)=\sum_{x\in F^{-1}(0)}\mathrm{sgn}(\det(DF(x))). If 0\notin F(\partial U) , then 0\notin F(U)\implies d(U,F)=0. 5.30 Lemma . Let \{L_{t}:[0,1]\to\mathbb{R}^{n+1}\}_{t\in[0,1]} and \{D_{t}:D^{n}\to\mathbb{R}^{n+1}\}_{t\in[0,1]} be homotopies of a curve and an n -dimensional disk respectively, such that 1. L_{0} and D_{0} are \mathcal{C}^{1} embeddings, 2. L_{0} intersects D_{0} exactly one time and such that this intersection is transverse, 3. L_{1} and D_{1} are disjoint, then there is some point in time t\in[0,1] such that L_{t} and D_{t}|_{\partial D^{n}} intersect or such that L_{t}|_{\partial[0,1]}=L_{t}|_{\{0,1\}} and D_{t} intersect. Proof. We define F_{t}:[0,1]\times D^{n}\to\mathbb{R}^{n+1},\quad F_{t}(x,y)=D_{t}(y)-L_{t}(x). Let D^{n+1}=[0,1]\times D^{n} . Since L_{0} and D_{0} intersect exactly once and since this intersection is transverse, we have d(D^{n+1},F_{0})=\pm 1 . Further, since L_{1} and D_{1} do not intersect, we have 0\notin F_{1}(D^{n+1}) , hence d(D^{n+1},F_{1})=0 . Since the Brouwer degree is homotopy invariant, this is only possible if there is some t\in[0,1] such that 0\in F_{t}(\partial D^{n+1}) . This proves our claim. ∎ The second tool we will need is the mountain climbing theorem. 5.31 Theorem (Mountain Climbing Theorem) . Let \gamma_{1},\gamma_{2}:[0,1]\to[0,T] be continuous such that \gamma_{1}(0)=0=\gamma_{2}(0) and \gamma_{1}(1)=T=\gamma_{2}(1) , and such that \gamma_{1} and \gamma_{2} are piecewise monotone, then there exist continuous functions \gamma^{\prime}_{1},\gamma^{\prime}_{2}:[0,1]\to[0,1] fixing endpoints with \gamma_{1}\circ\gamma^{\prime}_{1}=\gamma_{2}\circ\gamma^{\prime}_{2}. This result was first stated and proved by Whittaker in 1966 [ 11 ] . It has been strengthened since then, for example in 1969 by Huneke [ 7 ] . 5.32 Theorem (Disk separation lemma) . Let D^{2} be a disk and let K,L\subseteq D^{2} be closed such that D^{2}=K\cup L and such that K\cap S^{1}\times\{0\}=I_{A}\cup I_{B} and L\cap S^{1}\times\{0\}=I_{X}\cup I_{Y} are each the disjoint union of two compact intervals of positive length, with I_{A}\cap I_{X} , I_{B}\cap I_{X} , I_{A}\cap I_{Y} , and I_{B}\cap I_{Y} each containing exactly one element. Then either I_{A} and I_{B} are in the same connected component of K or I_{X} and I_{Y} are in the same connected component of L . Proof. Up to homeomorphism we can assume D^{2}=[0,1]^{2} and that W\coloneqq\{0\}\times[0,1]=I_{A} , E\coloneqq\{1\}\times[0,1]=I_{B} , S\coloneqq[0,1]\times\{0\}=I_{X} , N\coloneqq[0,1]\times\{1\}=I_{Y} . Let K_{1} be the connected component of W in K and let K_{2}=K\setminus K_{1} . Let L_{1} be the connected component of S in L and let L_{2}=L\setminus L_{1} . Then D^{2}=K_{1}\cup K_{2}\cup L_{1}\cup L_{2} . Toward a contradiction, assume that E\subseteq K_{2} and N\subseteq L_{2} . Since K and L were closed, so are K_{1} , K_{2} , L_{1} , and L_{2} . By Urysohn’s Lemma, there are functions g_{K}:D^{2}\to[0,1] and g_{L}:D^{2}\to[0,1] with g_{K}(z)=1 for all z\in K_{1} , g_{K}(z)=0 for all z\in K_{2} and g_{L}(z)=1 for all z\in L_{1} , g_{L}(z)=0 for all z\in L_{2} . We define the continuous map f:D^{2}\to D^{2},\quad f(z)=(g_{K}(z),g_{L}(z)) Note that z\in K_{1}\implies g_{A}(z)=1 , hence f maps K_{1} to E\subseteq K_{2} . Similarly f maps K_{2} to W\subseteq K_{1} and L_{1} to N\subseteq L_{2} and L_{2} to S\subseteq L_{1} . Any z\in D^{2} must be in at least one of the sets K_{1} , K_{2} , L_{1} , or L_{2} , and f(z) will be in a set disjoint to it. This means that f cannot have a fixed point, a contradiction to Brouwer’s fixed point theorem. We conclude that either E and W are in the same connected component of K , or S and N are in the same connected component of L . ∎ 5.33 Lemma . Consider two closed sets K,L\subseteq S^{1}\times[0,1] such that both K and L are locally path connected with K\cup L=S^{1}\times[0,1] , and such that K\cap(S^{1}\times\{0\})=I_{A}\cup I_{B} and L\cap(S^{1}\times\{0\})=I_{X}\cup I_{Y} are each the disjoint union of two compact intervals of positive length with I_{A}\cap I_{X} , I_{B}\cap I_{X} , I_{A}\cap I_{Y} , and I_{B}\cap I_{Y} each containing exactly one element. Further assume that both K\cap(S^{1}\times\{1\}) and L\cap(S^{1}\times\{1\}) are either empty or all of S^{1}\times\{1\} . Then either I_{A} and I_{B} are in the same path component of K or I_{X} and I_{Y} are in the same path component of L . Proof. Let q:S^{1}\times[0,1]\to(S^{1}\times[0,1])/\sim be the map defined by identifying all points in S^{1}\times\{1\} with each other. Then D^{2}\coloneqq(S^{1}\times[0,1])/\sim is a disk. We apply Theorem 5.32 to D^{2} and the sets q(K) and q(L) and conclude that either q(I_{A}) and q(I_{B}) are in the same connected component of q(K) , or q(I_{X}) and q(I_{Y}) are in the same connected component of q(L) . Without loss of generality assume the former. Since K is locally path connected and since q is an open map, q(K) is locally path connected and therefore q(I_{A}) and q(I_{B}) are in the same path component of q(K) . Take a path from q(I_{A}) to q(I_{B}) in q(K) . Since we assumed that K\cap(S^{1}\times\{1\}) is either empty or all of S^{1}\times\{1\} , we can always lift this path to a path from I_{A} to I_{B} in K . ∎ We are now able to prove the following result. 5.34 Lemma (Tracking Lemma) . Let M:S^{1}\to\mathbb{R}^{n+1} and D:D^{n}\to\mathbb{R}^{n+1} be \mathcal{C}^{1} , such that M and D intersect in exactly two points, and such that these intersections are transverse. Let M_{t} and D_{t} be homotopies of M and D respectively, so that M_{t} never intersects D_{t}|_{\partial D^{n}} and so that M_{1} and D_{1} are disjoint. Let \varepsilon>0 . Then there is some T\in[0,1] , some continuous \gamma:[0,1]\to[0,T] with \gamma(0)=0 and \gamma(1)=T , and two continuous maps x_{1},x_{2}:[0,1]\to S^{1} such that 1. x_{1}(0) and x_{2}(0) are the points at which M intersects D , 2. x_{1}(1)=x_{2}(1) , 3. \mathrm{dist}(M_{\gamma(t)}(x_{1}(t)),D_{\gamma(t)}),\mathrm{dist}(M_{\gamma(t)}(x_{2}(t)),D_{\gamma(t)})<\varepsilon for all t\in[0,1] . This lemma tells us that if we have a closed curve and an n -dimensional disk in \mathbb{R}^{n+1} that intersect nicely in exactly two spots, and if we have any homotopy splitting the curve and the disk without the curve passing through the boundary of the disk, then we can continuously track points in the parameter space of the curve such that the curve evaluated at these points stays arbitrarily close to the disk, and such that the parameter values eventually merge. This tracking does require us to "slide the homotopies back and forth through time", i.e. instead of considering M_{t} and D_{t} , we have to consider M_{\gamma(t)} and D_{\gamma(t)} for some curve \gamma . Proof. Assume without loss of generality that there is some c>0 such that for t\in[0,c] we have D_{t}=D and M_{t}=M . Let G:[0,1]\times S^{1}\to\mathbb{R},\quad G(t,x)=\mathrm{dist}(M_{t}(x),D_{t}), S=\{(t,x)\in[0,1]\times S^{1}\mid G(t,x)=0\}. Note that G is continuous and hence S is closed. Let \delta>0 be such that for all (t,x)\in[0,1]\times S^{1} with G(t,x)=0 we have (t^{\prime},x^{\prime})\in U_{\delta}(t,x)\implies G(t^{\prime},x^{\prime})<\varepsilon. Let Z\subseteq[0,1]\times S^{1} so that 1. Z contains an open neighborhood of S 2. Z is contained in the \delta -neighborhood of S 3. Z and the closure of its complement are locally path connected 4. Whenever there is a path in Z connecting two points in Z^{\circ} , there is also a path in Z^{\circ} connecting the same two points 5. Whenever there is a path in the closure of the complement of Z connecting two points in the complement of Z , there is a path in the complement of Z connecting the same two points. 6. \left(\{1\}\times S^{1}\right)\cap Z=\emptyset 7. \left(\{0\}\times S^{1}\right)\cap Z is a disjoint union of two compact intervals I_{A} and I_{B} of positive length, each containing one point in S . Such a set Z always exists, for example we could cover S with a finite number of closed radius- \tfrac{\delta}{2} balls in a suitable way. Let I_{X} and I_{Y} be the disjoint compact intervals so that I_{X}^{\circ}\cup I_{Y}^{\circ}=\left(\{0\}\times S^{1}\right)\setminus Z . Let A\subseteq Z be the connected component of I_{A} and let B\subseteq Z be the connected component of I_{B} in Z . Toward a contradiction, assume \mathbf{A\neq B} . By Lemma 5.33 , there is a path (\alpha,f):[0,1]\to[0,1]\times S^{1} from a point (0,x)\in I_{X}^{\circ} to a point (0,y)\in I_{Y}^{\circ} avoiding Z^{\circ} . By the properties of Z , we can assume that this path avoids Z . Since Z is closed, we can perturb the curve (\alpha,f) in the open set \left([0,1]\times S^{1}\right)\setminus Z so that the curve \alpha is piecewise monotone. Let T be the maximum value of \alpha . Without loss of generality assume that \alpha(\tfrac{1}{2})=T , else reparametrize (\alpha,f) . We now define (\alpha_{1},f_{1}),(\alpha_{2},f_{2}):[0,1]\to[0,1]\times S^{1} through (\alpha_{1},f_{1})(t)\coloneqq(\alpha,f)(\tfrac{t}{2}) and (\alpha_{2},f_{2})\coloneqq(\alpha,f)(1-\tfrac{t}{2}) . Note that \alpha_{1},\alpha_{2}:[0,1]\to[0,T] agree on endpoints and that both are piecewise monotone. We may assume that \gamma\coloneqq\alpha_{1}=\alpha_{2} , else we appeal to the mountain climbing theorem ( 5.31 ) for the existence of curves \alpha^{\prime}_{1},\alpha^{\prime}_{2}:[0,1]\to[0,1] , fixing endpoints, with \gamma\coloneqq\alpha_{1}\circ\alpha^{\prime}_{1}=\alpha_{2}\circ\alpha^{\prime}_{2} and then consider (\gamma,f_{1}\circ\alpha^{\prime}_{1}) and (\gamma,f_{2}\circ\alpha^{\prime}_{2}) . Let q:\mathbb{R}\to S^{1} be the standard covering. Lift f_{1} and f_{2} to paths \overline{f_{1}},\overline{f_{2}}:[0,1]\to\mathbb{R} , i.e. q\circ\overline{f_{1}}=f_{1} and q\circ\overline{f_{2}}=f_{2} . Pick the lifts so that \overline{f_{1}}(0)<\overline{f_{2}}(0)<\overline{f_{1}}(0)+1 . Then define h:[0,1]^{2}\to\mathbb{R} , h(t,x)=(1-x)\overline{f_{1}}(t)+x\overline{f_{2}}(t) . For a given t , the image of h(t,\cdot) corresponds to the interval between \overline{f}_{1}(t) and \overline{f}_{2}(t) . Looking at q(h(t,\cdot)):[0,1]\to S^{1} we get a continuously varying family of curves in S^{1} with endpoints f_{1}(t) and f_{2}(t) . Next we define L_{t}:[0,1]\to\mathbb{R}^{n+1},\quad L_{t}(x)=M_{\gamma(t)}(q(h(t,x))). The curve L_{0} is simply a reparametrization of a segment of M . We further define \tilde{D}_{t}:D^{n}\to\mathbb{R}^{n+1},\quad\tilde{D}_{t}(y)=D_{\gamma(t)}(y). Note that L_{0} intersects \tilde{D}_{0}=D_{0} exactly once, and that this intersection is transverse. Further note that \mathrm{im}(L_{1})\subseteq\mathrm{im}(M_{1}) and \tilde{D}_{1}=D_{1} . Since M_{1} and D_{1} don’t intersect, neither do L_{1} and \tilde{D}_{1} . By Lemma 5.30 , this is only possible if for some t\in[0,1] we have L_{t}(0)=f(t)\in\tilde{D}_{t} , or L_{t}(1)=g(t)\in\tilde{D}_{t} , or if L_{t} intersects \tilde{D}_{t}(\partial D^{n}) . By construction L_{t}(f_{1}(t)) and L_{t}(f_{2}(t)) avoid \tilde{D}_{t} , and we know that L_{t} and \tilde{D}_{t}(\partial D^{n}) are disjoint because we assumed that M_{t} and D_{t}|_{\partial D^{n}} are always disjoint. We have arrived at a contradiction. We conclude that our assumption that \mathbf{A\neq B} was wrong and therefore \mathbf{A=B} . By arguments exactly the same as in the proof-by-contradiction, we obtain two paths (\gamma,x_{1}),(\gamma,x_{2}):[0,1]\to Z\subset[0,1]\times S^{1} with x_{1}(0) and x_{2}(0) being the two intersections of M and D and with x_{1}(1)=x_{2}(1) . The fact that (\gamma,x_{1}) and (\gamma,x_{2}) have image in Z implies \mathrm{dist}(M_{\gamma(t)}(x_{1}(t)),D_{\gamma(t)}),\mathrm{dist}(M_{\gamma(t)}(x_{2}(t)),D_{\gamma(t)})<\varepsilon . ∎ We now proceed to prove Proposition 5.28 Proof of Proposition 5.28 . Let \varepsilon>0 and define \varepsilon^{\prime}=\tfrac{\varepsilon}{2\pi+4} . By Lemma 5.34 , there is some T\in[0,1] , some continuous \gamma:[0,1]\to[0,T] with \gamma(0)=0 and \gamma(1)=T , and two continuous maps x_{1},x_{2}:[0,1]\to S^{1} such that 1. x_{1}(0)=a and x_{2}(0)=b , 2. x_{1}(1)=x_{2}(1) , 3. \mathrm{dist}(M_{\gamma(t)}(x_{1}(t)),D_{\gamma(t)}),\mathrm{dist}(M_{\gamma(t)}(x_{2}(t)),D_{\gamma(t)})<\varepsilon for all t\in[0,1] . Assume without loss of generality that \min\left(\{t\in[0,1]\mid x_{1}(t)=x_{2}(t)\}\right)=1 . We define \left(\tilde{M}_{t},\tilde{A}^{1}_{t},\tilde{A}^{2}_{t},\tilde{C}_{t}\right)=\tilde{H}_{t}=H_{\gamma(t)} , and let \tilde{D}_{t} be the center of mass spanning disk of \tilde{C}_{t} . For t\in[0,1] , x_{1}(t) and x_{2}(t) partition S^{1} into two parts, say I_{t}^{1} and I_{t}^{2} . We define \tilde{M}_{t}^{1} as the concatenation of \tilde{M}_{t}|_{I^{1}_{t}} and the straight line segment between \tilde{M}_{t}(x_{1}(t)) and \tilde{M}_{t}(x_{2}(t)) . Analogously we define \tilde{M}_{t}^{2} as the concatenation of \tilde{M}_{t}|_{I^{2}_{t}} and the straight line segment between \tilde{M}_{t}(x_{1}(t)) and \tilde{M}_{t}(x_{2}(t)) . These are continuously varying families of curves. Either \tilde{M}^{1}_{1} or \tilde{M}^{2}_{1} becomes a single point. Without loss of generality assume that \tilde{M}^{1}_{1} becomes a single point. Since \tilde{M}^{1}_{0} was linked to \tilde{A}^{1}_{0} , and since \tilde{M}^{1}_{1} is a point and hence not linked to \tilde{A}^{1}_{1} , there must be some t\in[0,1] such that \tilde{M}^{1}_{t} and \tilde{A}^{1}_{t} intersect. This intersection must happen on the straight line segment between \tilde{M}_{t}(x_{1}(t)) and \tilde{M}_{t}(x_{2}(t)) , as \tilde{M}_{t} and \tilde{A}^{1}_{t} stay disjoint throughout. Let P_{1}:=\tilde{M}_{t}(x_{1}(t)) and P_{2}:=\tilde{M}_{t}(x_{2}(t)) , then there is some Q\in\tilde{A}^{1}_{t} on the line segment \overline{P_{1}P_{2}} with ||P_{1}-Q||,||P_{2}-Q||\geq 1 . This implies ||P_{1}-P_{2}||\geq 2 . Further, we have \mathrm{dist}(P_{1},\tilde{D}_{t})<\varepsilon^{\prime} and likewise \mathrm{dist}(P_{2},\tilde{D}_{t})<\varepsilon^{\prime} . By Proposition 4.27 , we get \mathrm{len}(C_{\gamma(t)})=\mathrm{len}(\tilde{C}_{t})\geq 2\pi+2||P_{1}-P_{2}||-(2\pi+4)\varepsilon^{\prime}\geq 2\pi+4-\varepsilon. Since \varepsilon>0 was arbitrary, we get \sup_{t\in[0,1]}\mathrm{len}(C_{t})\geq 2\pi+4 . ∎ 5.2 Constructing an isotopically Gordian split link Next we exhibit a thickly embedded split link L fulfilling the conditions of Proposition 5.28 , which has an initial length of less than 8\pi+6 . This means that L cannot be split through a thick homotopy. A fortiori, L cannot be split through a thick isotopy. So L is a Gordian split link both as a thick and as a thickly embedded link. We construct the link L=(M,A^{1},A^{2},C) such that the main curve M is in the XY -plane, the auxiliary curves A^{1} and A^{2} are in the XZ -plane, and the center curve is in the YZ -plane. The following plots, as well as numerical calculations of some of the curves’ lengths, were created and conducted with Mathematica. Figure 2: A plot of the thickly embedded Gordian split link L Figure 3: Plot of the radius- \tfrac{1}{2} tubes around the link L . To see how our Gordian split link is defined, we now look at the three different coordinate planes. In the plots of the coordinate planes, we will always plot the curves, as well as forbidden areas, i.e. the 1 -neighborhoods of the curves in other coordinate planes. We will also plot and reference selected points to make following the definition of the curves easier. The points are named A through X , so some points share a name with our curves M and C . It will always be obvious which is meant, so there should be no confusion. We define the curves in sequence. First we define the auxiliary curves A^{1} and A^{2} , then the main curve M and finally the center curve C . Figure 4: A plot of the XZ -plane. The auxiliary curves A^{1} and A^{2} are plotted in green, the forbidden areas belonging to main curve and center curve are plotted in gray and blue respectively. Figure 5: A plot of the XY -plane. The main curve M is plotted in black, the forbidden area belonging to the auxiliary curves A^{1} and A^{2} is plotted in green and the forbidden area belonging to center curve C is plotted in blue. Figure 6: A plot of the YZ -plane. The center curve C is plotted in blue, the forbidden area belonging to the auxiliary curves A^{1} and A^{2} is plotted in green and the forbidden area belonging to the main curve M is plotted in gray. The two auxiliary curves are unit circles in the XZ -planes with centers E:=\left(\tfrac{3}{2},0,0\right) and I:=\left(-\tfrac{3}{2},0,0\right) . The main curve M is defined as follows: In the XY -plane, take the unit circles with centers B:=\left(\tfrac{1}{2},0,0\right) , M:=\left(0,\tfrac{\sqrt{15}}{2},0\right) , C:=\left(-\tfrac{1}{2},0,0\right) , and S:=\left(0,-\tfrac{\sqrt{15}}{2},0\right) . The circles around M and B touch in exactly one point F . Similarly, we can define H , J , and L . Let G:=\left(0,\tfrac{\sqrt{15}}{2}-1,0\right) and J:=\left(0,-\left(\tfrac{\sqrt{15}}{2}-1\right),0\right) . The main curve M is defined as the concatenation of four unit circle segments (see Figure 5 ). Let \alpha=\angle FMG , then \alpha=\tfrac{\pi}{2}-\arctan(\sqrt{15})\approx 0.2527 . We get \mathrm{len}(M)=2\pi+8\alpha=6\pi-8\arctan(\sqrt{15})\approx 8.3046 . To define the center curve C , we parametrize the forbidden green region, which is the intersection of the solid, open 1 -torus around either of auxiliary curve with the YZ -plane. This region, which we will denote by \mathcal{G} , is given by \left(\left(\tfrac{3}{2}\right)^{2}+z^{2}+y^{2}\right)^{2}-4\left(\left(\tfrac{3}{2}\right)^{2}+z^{2}\right)<0, and its boundary \partial\mathcal{G} is a closed curve implicitly defined through \left(\left(\tfrac{3}{2}\right)^{2}+z^{2}+y^{2}\right)^{2}-4\left(\left(\tfrac{3}{2}\right)^{2}+z^{2}\right)=0. There is a unique straight line segment in the quadrant y,z\geq 0 connecting a point on the unit circle around G and a point on the boundary curve \partial\mathcal{G} while being tangent to both the unit circle and \mathcal{G} . We call the start of this line segment N and the end of it O . Let P=\left(0,0,\tfrac{\sqrt{7}}{2}\right) and M=\left(0,\tfrac{\sqrt{15}}{2},0\right) . We define the first quarter of our center curve C as the concatenation of the circle segment from M to N , the straight line segment from N to O , and the segment of the boundary curve \partial\mathcal{G} from O to P . The other three quarters of C are defined by reflecting the first segment from M to P along both the Y -axis and the Z -axis. Mathematica was not able to solve for the coordinates of N and O exactly, so they were calculated numerically instead. The angle \beta=\angle MGN was calculated as \beta\approx 1.2554 . The length of the straight line segment s from N to O was calculated as \mathrm{len}(s)\approx 1.0925 . The length of the segment b from O to P along \partial\mathcal{G} was numerically calculated as \mathrm{len}(b)\approx 0.2116 . The length of C is therefore \mathrm{len}(C)=4\left(\beta+\mathrm{len}(s)+\mathrm{len}(b)\right)\approx 10.2380 . Finally, we get \displaystyle\mathrm{len}(L)= \displaystyle 2\pi+2\pi+\left(6\pi-8\arctan(\sqrt{15})\right)+4\left(\beta+\mathrm{len}(s)+\mathrm{len}(b)\right) \displaystyle\approx \displaystyle 31.1090<31.1327\approx 8\pi+6. We have constructed L to be a thick link, but have not yet shown that L is thickly embedded. The open radius- \tfrac{1}{2} normal disk bundles around A^{1} , A^{2} and M are obviously embedded. To see that the open radius- \tfrac{1}{2} normal disk bundle around C is embedded, we need to check that the segment b is \mathcal{C}^{1,2} . This is true. In fact, all of \mathcal{G} is \mathcal{C}^{1,2} . We omit a proof as it is cumbersome and not very interesting. Numerical calculations show that the curvature of \partial\mathcal{G} peaks at the points P and V where it is around 1.51 . 5.35 Theorem (A 4 -component Gordian split link exists) . There exist 4 -component thickly embedded links that are Gordian split links as thick links (and hence a fortiori as thickly embedded links). The difference between the lower bound we derived in section 5.1 and the length of our constructed configuration L is less than 0.08\% of \mathrm{len}(L) . While it is unlikely that L can be shortened much, it is highly likely that the lower bound of 8\pi+6 from Proposition 5.28 can be significantly improved. A careful examination of the proof of Proposition 5.28 shows that we never needed to assume that the distances between A^{1} and C and between A^{2} and C stay at least 1 throughout the homotopy. In fact, we have not even used that A^{1} and A^{2} stay disjoint from C . We only needed that (A^{1},A^{2},M) and (M,C) stay thick links throughout the homotopy of L . 6 Homotopically Gordian pairs exist in every link homotopy class with 2 components 6.1 Circle decompositions - the idea We first sketch the idea for our construction, as well as the proof, informally. Consider a two-component thick link L made up of unit circle segments as in Figure 7 and parametrized as shown in Figure 8 Figure 7: A conjoined circle decomposition: The image of the blue curve is the union of three coplanar, pairwise tangent unit circles. The image of the red curve is the union of three pairwise tangent unit circles, each piercing the centers of 2 blue circles. Figure 8: A perturbation of a conjoined circle decomposition: This figure shows a thick link L^{\prime} close to L to help understand how L is parametrized. In the original configuration, there are two strands going through each circle’s center. After a small perturbation, for every circle, pick one point each on both of the strands through that circle minimizing their respective distance. This gives a set of two points for every circle. These sets are marked A,B,C,D,E,F in Figure 8 . We now split up the two link-components into 6 curves each, along the points making up the sets A,B,C,D,E,F , and pair them up as indicated by the color scheme. Let us calculate the sum S_{\text{blue}} of the lengths of the blue segments. We define the closed curve C_{\text{blue}} by connecting the two blue components by a straight line segment through the 2 points in B and by a straight line segment through the 2 points in F . Then the two strands piercing the blue circle will pierce the center-of-mass spanning disk of C_{\text{blue}} in (at least) one point each, say in the points p and q , where \{p,q\}=A . We have \mathrm{len}(C_{\text{blue}})\geq 2\pi+2||p-q||=2\pi+2\mathrm{diam}(A), S_{\text{blue}}=\mathrm{len}(C_{\text{blue}})-\mathrm{diam}(B)-\mathrm{diam}(F)\geq 2\pi+2\mathrm{diam}(A)-\mathrm{diam}(B)-\mathrm{diam}(F). Doing the same thing for all other colors and summing the lengths we get that all the terms involving diameters of sets A,B,C,D,E,F cancel, so that the only thing remaining is \mathrm{len}(L^{\prime})\geq 12\pi. This would then show that any thick link sufficiently close to the original configuration is at least as long, so the original configuration seems to be a local minimum. Besides the many missing details, there are two problems with this proof. First, even if our arguments are correct, we only get that the initial configuration is a local minimum, which is not yet sufficient to conclude that it is a sink. Secondly, we made a mistake. We do not know that the straight line segment keeps distance at least 1 to the points p and q intersecting the spanning disk of C_{\text{blue}} . But this was a condition implicitly used when we claimed C_{\text{blue}}\geq 2\pi+2||p-q|| . We will see that we can make adjustments to the argument to get \displaystyle\mathrm{len}(L^{\prime})\geq 12\pi \displaystyle+2\mathrm{diam}(A)-2\varphi(\mathrm{diam}(A)) \displaystyle+2\mathrm{diam}(B)-2\varphi(\mathrm{diam}(B)) \displaystyle+2\mathrm{diam}(C)-2\varphi(\mathrm{diam}(C)) \displaystyle+2\mathrm{diam}(D)-2\varphi(\mathrm{diam}(D)) \displaystyle+2\mathrm{diam}(E)-2\varphi(\mathrm{diam}(E)) \displaystyle+2\mathrm{diam}(F)-2\varphi(\mathrm{diam}(F)), where near 0 we have that \varphi(x)\sim x to first order, but with \varphi(x)>x for x>0 . This adjusted argument is no longer strong enough to show that the original configuration was a local minimum. In some ways, the reason why the argument fails is that there are too many "breaking points". We can fix this by augmenting L . Figure 9: A perturbation of an augmented conjoined circle decomposition. In the configuration shown in Figure 9 , we go through every one of the original 6 circles one additional time. This only introduces one new breaking point for each circle, which corresponds to one new term of -\varphi(\mathrm{diam}(\cdot)) for each of the sets A,B,C,D,E,F . At the same time, each of the circles introduces a term of 2\mathrm{diam}(\cdot) for the sets A,B,C,D,E,F . This gives the inequality \displaystyle\mathrm{len}(L^{\prime})\geq 24\pi \displaystyle+4\mathrm{diam}(A)-3\varphi(\mathrm{diam}(A)) \displaystyle+4\mathrm{diam}(B)-3\varphi(\mathrm{diam}(B)) \displaystyle+4\mathrm{diam}(C)-3\varphi(\mathrm{diam}(C)) \displaystyle+4\mathrm{diam}(D)-3\varphi(\mathrm{diam}(D)) \displaystyle+4\mathrm{diam}(E)-3\varphi(\mathrm{diam}(E)) \displaystyle+4\mathrm{diam}(F)-3\varphi(\mathrm{diam}(F)). Since \varphi(x)\sim x to first order, for small enough diameters this inequality yields \mathrm{len}(L^{\prime})\geq 24\pi with equality if and only if \mathrm{diam}(A)=\mathrm{diam}(B)=\dots=\mathrm{diam}(F)=0 . i.e. we have \mathrm{len}(L^{\prime})\geq 24\pi with equality if and only if L^{\prime} is also a circle decomposition. 6.2 Circle decompositions - the details We now formally define what we mean by a circle decomposition and a conjoined circle decomposition. Throughout the rest of this paper \frown will denote concatenation of curves. 6.36 Definition . Let C be a closed curve such that the definition domain of C decomposes into I_{1}\cup I_{2}\cup\dots\cup I_{n} , where each I_{i} is a compact interval of positive length with \max(I_{i})=\min(I_{i+1}) , for all i (here indices are understood to be cyclic and the definition domain is seen as an interval with endpoints identified). We define a_{i} via [a_{i},a_{i+1}]=I_{i} . Let \{A_{j}\}_{j=1}^{m} and \{\alpha_{l}\}_{l=1}^{k} each be a partition of \{1,2,\dots,n\} into non-empty sets. We say that \left(C,\{a_{i}\}_{i=1}^{n},\{A_{j}\}_{j=1}^{m},\{\alpha_{l}\}_{l=1}^{k}\right) is a circle decomposition of C if for all i_{1},i_{2}\in\{1,\dots,n\} and j\in\{1,\dots,m\} : 1. There is a bijection \{1,2,\dots,c\}\to A_{j} , k\mapsto i_{k} and some s_{1},\dots,s_{c}\in\{1,-1\} , such that C^{j}:=C_{i_{1}}^{s_{1}}\frown C_{i_{2}}^{s_{2}}\frown\dots\frown C_{i_{c}}^{s_{c}} is a curve of length 2\pi parametrizing a unit circle. Here C_{i_{k}}^{1}:=C|_{I_{k}} and C_{i_{k}}^{-1} is defined to be the curve obtained by inverting the orientation of C|_{I_{k}} . 2. C(a_{i_{1}})=C(a_{i_{2}})\iff\text{ there is some }l\text{ with }i_{1},i_{2}\in\alpha_{l} The length of a circle decomposition \left(C,\{a_{i}\}_{i=1}^{n},\{A_{j}\}_{j=1}^{m},\{\alpha_{l}\}_{l=1}^{k}\right) is equal to 2m\pi . 6.37 Remark . From here on we will write circle decompositions just as (C,\{a_{i}\},\{A_{j}\},\{\alpha_{l}\}) , without specifically referencing the index-sets. We want to be able to talk about multiple circle decompositions at once. For consistency and readability, we state some conventions that will be used throughout the rest of this section. 1. We implicitly use the definition K_{j}:=\bigcup_{i\in A_{j}}C(I_{i}) and the definition of z_{j}\in\mathbb{R}^{3} as the center of the circle K_{j} . 2. We use i to index points or intervals in parameter space, where I_{i}=[a_{i},a_{i+1}] . We call the points a_{i} breaking points. 3. We use j to index circles of the circle decomposition: Either for the collection of (indices representing) intervals A_{j} that map to a circle, or for the closed curve C^{j} parametrizing K_{j} . 4. We will use l to index sets \alpha_{l} of indices representing points a_{i} . We will call the sets \alpha_{l} point clusters, as they are made up of indices representing a subset of \{a_{i}\} clustered on C . Because of the way \alpha_{l} is defined, we can define C(\alpha_{l}):=C(a_{i}) for some (hence every) i\in\alpha_{l} . 6.38 Definition . Let (C,\{a_{i}\},\{A_{j}\},\{\alpha_{l}\}) be a circle decomposition. We say that two indices i_{1},i_{2}\in\alpha_{l} are distinguishable if there are small neighborhoods H_{i_{1}} around a_{i_{1}} and H_{i_{2}} around a_{i_{2}} such that C|_{H_{i_{1}}} and C|_{H_{i_{2}}} intersect only at C(\alpha_{l}) . 6.39 Definition . Let (C,\{a_{i}\},\{A_{j}\},\{\alpha_{l}\}) be a circle decomposition. If a_{i} is on the boundary of one of the intervals I_{i^{\prime}} with i^{\prime}\in A_{j} , then we say that a_{i} is along the circle C^{j} . We say that \alpha_{l} is along a circle C^{j} if there is some i\in\alpha_{l} with a_{i} along C^{j} . It will be useful to define the notion of a transverse intersection in the non-differentiable setting. 6.40 Definition (transverse intersection) . Let C:[0,1]\to\mathbb{R}^{n+1} be a curve and let D:[0,1]^{n}\to\mathbb{R}^{n+1} be a disk. We define D-C:[0,1]\times[0,1]^{n}\to\mathbb{R}^{n+1},\quad(D-C)(x,y)=D(x)-C(y). We say that C intersects D transversely at a point p=D(x)=C(y) if there is an open disk U\subseteq[0,1]^{n} around x and an open interval I\subseteq[0,1] around y such that D|_{U} and C|_{I} only intersect in p and such that the Brouwer degree of (D-C)|_{U\times I} around 0 is \pm 1 . Next we define conjoined circle decompositions. 6.41 Definition . Let (C_{1},\{a_{i}\},\{A_{j}\},\{\alpha_{l}\}) be a circle decomposition of C_{1} and let (C_{2},\{b_{i}\},\{B_{j}\},\{\beta_{l}\}) be a circle decomposition of C_{2} . We say that C_{1}\cup C_{2} have a conjoined circle decomposition if 1. \mathrm{dist}(C_{1},C_{2})=1, 2. K_{j} circle of C_{1} \implies there is some \beta_{l} such that z_{j}=C_{2}(\beta_{l}), and every intersection of C_{2} and D_{j} is transverse, 3. K_{j} circle of C_{2} \implies there is some \alpha_{l} such that z_{j}=C_{1}(\alpha_{l}), and every intersection of C_{1} and D_{j} is transverse. The last two conditions guarantee the existence of some functions h^{A} , h^{B} defined as follows: h^{B}(j)\text{, such that }C_{2}(\beta_{h^{B}(j)})=z_{j},\quad h^{A}(j)\text{, such that }C_{1}(\alpha_{h^{A}(j)})=z_{j}. So h^{A} takes an index j of some circle K_{j} of C_{2} to the index l of the point cluster \alpha_{l} which is mapped to z_{j} , and analogously for h^{B}(j) . The linking scheme of C_{1} and C_{2} is defined to be the data for which indices j belonging to the circle decomposition of C_{1} and j^{\prime} belonging to the circle decomposition of C_{2} the circles K_{j} and K_{j^{\prime}} are non-trivially linked. The data of the conjoined circle decomposition are the data of the individual circle decompositions along with the functions h^{A} and h^{B} , and the linking scheme. Every circle of one of the circle decompositions has its center pierced by the other curve, and this only ever happens transversely. The arc-length parametrization of both curves is piecewise \mathcal{C}^{1} as it consists of circle segments. Further, the transitions between circle segments happen only when a curve pierces a unit disk of a unit circle of the other curve. Because of the condition \mathrm{dist}(C_{1},C_{2})=1 , we know that each curve pierces these disks only at the center, and it does so perpendicularly. Therefore the arc-length parametrization of both curves is actually \mathcal{C}^{1} . 6.42 Remark . Let (C_{1},\{a_{i}\},\{A_{j}\},\{\alpha_{l}\}),(C_{2},\{b_{i}\},\{B_{j}\},\{\beta_{l}\}) be a conjoined circle decomposition. When we have a conjoined circle decomposition, we will refer to indices i , j , and l belonging to the circle decomposition of C_{1} as A -indices and to the indices i , j and l belonging to the circle decomposition of C_{2} as B -indices. This way of handling the many different (kinds of) indices implicitly, without defining and naming 6 different index sets to keep track of, will hopefully be easier to follow than the alternative. For technical reasons, we will need to consider a smaller class of conjoined circle decompositions. 6.43 Definition . Let (C_{1},\{a_{i}\},\{A_{j}\},\{\alpha_{l}\}),(C_{2},\{b_{i}\},\{B_{j}\},\{\beta_{l}\}) be a conjoined circle decomposition. We say that the conjoined circle decomposition is suitable , if 1. Every circle K_{j} of one of the curves is linked to exactly two circles K_{j_{1}},K_{j_{2}} of the other curve, 2. for every A -index l : \#\{j\text{ }A\text{-index }\mid l\text{ along }C_{1}^{j}\}<2\#(h^{A})^{-1}(l) , 3. for every B -index l : \#\{j\text{ }B\text{-index }\mid l\text{ along }C_{2}^{j}\}<2\#(h^{B})^{-1}(l) 4. for every \alpha_{l} , there are two distinguishable i_{1},i_{2}\in\alpha_{l} , 5. for every \beta_{l} , there are two distinguishable i_{1},i_{2}\in\beta_{l} . The first condition is not strictly necessary for our arguments, but it will make the book-keeping a lot easier. Note that we really do mean that every circle belonging to one curve is linked to two distinct circles of the other curve, not to two circles with distinct indices. This is important as we can have K_{j_{1}}=K_{j_{2}} even if j_{1}\neq j_{2} . The next two suitability conditions can be rephrased as "The number of circles broken up by any point cluster is less than two times the number of circles around that point cluster". This is the main condition needed to make our inequalities work out. The last two conditions are so that we can use Lemma 6.52 . We can now state the main theorem of this chapter. 6.44 Theorem . Let S be the set of 2 -component links in some link homotopy class, that all have conjoined circle decompositions with the same given index sets \{A_{j}\} , \{\alpha_{l}\} , \{B_{j}\} , \{\beta_{l}\} , the same index functions h^{A},h^{B} and the same linking scheme. Then S is a closed, strictly locally minimizing set. The main tool in proving Theorem 6.44 will be finding a lower bound for the length of "almost closed" curves avoiding some set of points. In the following lemmas, U_{r}(\cdot) will refer to the open r -neighborhood of a set or point and B_{r}(\cdot) will refer to the closed r -neighborhood of a set or point. 6.45 Lemma . Let D_{1} and D_{2} be two open disks in \mathbb{R}^{2} , with radii r_{1},r_{2} and centers z_{1},z_{2} respectively, with r_{1},r_{2}\in[\tfrac{1}{2},1] and ||z_{1}-z_{2}||<\tfrac{1}{2} . For all a,b\in\mathbb{R}^{2}\setminus\left\{D_{1}\cup D_{2}\right\} with ||a-b||<\tfrac{1}{2} and all m\in D_{1}\cap D_{2} not collinear with a and b , there exists a curve C from a to b , such that \mathrm{len}(C)\leq\arcsin(2||a-b||)-||a-b||, obtained from radial projection of the straight-line segment \overline{ab} from m . Proof. Take any a,b\in\mathbb{R}^{2}\setminus(D_{1}\cup D_{2}) and consider the straight line segment \overline{ab} . If \overline{ab} avoids D_{1}\cup D_{2} , then we take C=\overline{ab} . If \overline{ab} intersects D_{1}\cup D_{2} then we define C by radially projecting (from m ) the part of \overline{ab} that lies between the first and the final intersection of \overline{ab} and D_{1}\cup D_{2} onto \partial(D_{1}\cup D_{2}) and by leaving the rest of \overline{ab} the same. Then C=l_{1}\frown A\frown l_{2} where l_{1},l_{2} are straight line segments which are part of \overline{ab} and where A is the path on the boundary. There are three cases, either A is a circular arc, the concatenation of two circular arcs A_{1}\frown A_{2} , or the concatenation of three circular arcs A_{1}\frown A_{2}\frown A_{3} . Assume A is a circular arc, say from v_{1} to v_{2} and assume without loss of generality that the arc is on \partial D_{1} . Then \displaystyle\mathrm{len}(C) \displaystyle=\mathrm{len}(l_{1})+\mathrm{len}(A)+\mathrm{len}(l_{2}) \displaystyle\leq\mathrm{len}(l_{1})+\mathrm{len}(l_{2})+2r_{1}\arcsin\left(\frac{||v_{1}-v_{2}||}{2r_{1}}\right) \displaystyle\leq 2r_{1}\arcsin\left(\frac{\mathrm{len}(l_{1})+\mathrm{len}(l_{2})}{2r_{1}}\right)+2r_{1}\arcsin\left(\frac{||v_{1}-v_{2}||}{2r_{1}}\right) \displaystyle\leq 2r_{1}\arcsin\left(\frac{\mathrm{len}(l_{1})+||v_{1}-v_{2}||+\mathrm{len}(l_{2})}{2r_{1}}\right) \displaystyle\leq\arcsin\left(\mathrm{len}(l_{1})+||v_{1}-v_{2}||+\mathrm{len}(l_{2})\right) \displaystyle=\arcsin(||a-b||) \displaystyle\leq\arcsin(2||a-b||)-||a-b||, where we used that for x\geq 0 , the map r\mapsto 2r\arcsin(\tfrac{x}{2r}) is monotonically decreasing and that \tfrac{1}{2}\leq r_{1} . If A=A_{1}\frown A_{2} is a union of two circular arcs, then \partial D_{1} and \partial D_{2} intersect in exactly two points, say p and q . Assume without loss of generality that \{p\}=A_{1}\cap A_{2} and that \overline{ab} intersects D_{1} before D_{2} . We know that \overline{ab} must intersect \partial D_{1} twice, say at points w_{1},w_{3} , and it must intersect \partial D_{2} twice, say at points w_{2},w_{4} . These intersections happen in order w_{1},w_{2},w_{3},w_{4} . Then A_{1} a circular arc on D_{1} from w_{1} to p and A_{2} a circular arc on D_{2} from p to w_{4} . Let E_{1} be the circular arc from p to w_{3} on \partial D_{1} and let E_{2} be the circular arc from w_{2} to p on \partial D_{2} . Then A_{1}\frown E_{1} is the circular arc from w_{1} to w_{3} , hence \mathrm{len}(A_{1}\frown E_{1})=2r_{1}\arcsin\left(\tfrac{||w_{1}-w_{3}||}{2r_{1}}\right)\leq\arcsin(||w_{1}-w_{3}||) and similarly \mathrm{len}(E_{2}\frown A_{2})=2r_{2}\arcsin\left(\tfrac{||w_{2}-w_{4}||}{2r_{2}}\right)\leq\arcsin(||w_{2}-w_{4}||) . Note that E_{2}\frown E_{1} is a path from w_{2} to w_{3} , so ||w_{2}-w_{3}||\leq\mathrm{len}(E_{2})+\mathrm{len}(E_{1}) . This yields \displaystyle\mathrm{len}(A)\leq \displaystyle\arcsin(||w_{1}-w_{3}||)+\arcsin(||w_{2}-w_{4}||)-||w_{2}-w_{3}|| \displaystyle= \displaystyle\arcsin(||w_{1}-w_{2}||+2||w_{2}-w_{3}||+||w_{2}-w_{4}||)-||w_{2}-w_{3}|| \displaystyle\leq \displaystyle\arcsin(||w_{1}-w_{2}||+2||w_{2}-w_{3}||+||w_{3}-w_{4}||) \displaystyle+\arcsin(||w_{1}-w_{2}||+||w_{3}-w_{4}||) \displaystyle-||w_{1}-w_{2}||-||w_{2}-w_{3}||-||w_{3}-w_{4}|| \displaystyle\leq \displaystyle\arcsin(2||w_{1}-w_{4}||)-||w_{1}-w_{4}||. It follows that \mathrm{len}(C)\leq\arcsin(2||a-b||)-||a-b|| . Finally, assume that A=A_{1}\frown A_{2}\frown A_{3} is the union of three circular arcs. Without loss of generality assume \{p\}=A_{1}\cap A_{2} , \{q\}=A_{2}\cap A_{3} , that A_{1} and A_{3} are arcs on \partial D_{1} and that A_{2} is an arc on \partial D_{2} . Then A_{2} is the arc from p to q on \partial D_{2} . Let v_{1} denote the start of A_{1} and let v_{2} denote the end of A_{3} and let E_{2} be the arc from p to q on \partial D_{1} . Then A_{1}\frown E_{2}\frown A_{3} is an arc from v_{1} to v_{2} on \partial D_{1} . Therefore \mathrm{len}(A_{1}\frown E_{2}\frown A_{3})=2r_{1}\arcsin\left(\tfrac{||v_{1}-v_{2}||}{2r_{1}}\right)\leq\arcsin(||v_{1}-v_{2}||) . This yields \displaystyle\mathrm{len}(A)= \displaystyle\mathrm{len}(A_{1}\frown E_{2}\frown A_{3})+(\mathrm{len}(A_{2})-\mathrm{len}(E_{2})) \displaystyle\leq \displaystyle\arcsin(||v_{1}-v_{2}||)+\left(\mathrm{len}(A_{2})-\mathrm{len}(E_{2})\right) \displaystyle= \displaystyle\arcsin(||v_{1}-v_{2}||)+2r_{2}\arcsin\left(\frac{||p-q||}{2r_{2}}\right)-2r_{1}\arcsin\left(\frac{||p-q||}{2r_{1}}\right) \displaystyle\leq \displaystyle\arcsin(||v_{1}-v_{2}||)+\arcsin(||p-q||)-||p-q|| \displaystyle\leq \displaystyle\arcsin(||v_{1}-v_{2}||)+\arcsin(||v_{1}-v_{2}||)-||v_{1}-v_{2}|| \displaystyle\leq \displaystyle\arcsin(2||v_{1}-v_{2}||)-||v_{1}-v_{2}|| where we used that x\mapsto\arcsin(x)-x is monotonically increasing and that ||p-q||\leq||v_{1}-v_{2}|| . This then yields \mathrm{len}(C)\leq\arcsin(2||a-b||)-||a-b|| . ∎ 6.46 Lemma . Let x,y,m\in\mathbb{R}^{3} with ||x-y||,||x-m||,||y-m||<\tfrac{1}{2} Let a,b\in\mathbb{R}^{3}\setminus U_{1}(\{x,y\}) with ||a-b||<\tfrac{1}{2} . Let P be a plane through a , b and m . Then there is a curve C\subseteq P from a to b , avoiding U_{1}(\{x,y\}) with \mathrm{len}(C)\leq\arcsin(2||a-b||)-||a-b|| and such that C is obtained by radial projection of the straight line segment \overline{ab} away from m . Proof. Toward a contradiction, assume m\in\overline{ab} . Then ||a-x||\leq||a-m||+||m-x||\leq||a-b||+||m-x||<1 , which is a contradiction to ||a-b||,||m-x||<\tfrac{1}{2} . So m\notin\overline{ab} . This means that if a , b and m are collinear, then either a\in\overline{mb} or b\in\overline{am} . Without loss of generality assume that a\in\overline{mb} . Toward a contradiction, assume there is some c\in\overline{ab} with c\in U_{1}(x) . Since U_{1}(x) is convex and since a\in\overline{mc} , this implies a\in U_{1}(x) , which is not the case. We conclude that \overline{ab} and U_{1}(x) are disjoint. Analogously we can conclude that \overline{ab} and U_{1}(y) are disjoint. We define C=\overline{ab} . If a , b and m are not collinear, then by taking the intersection of U_{1}(x) and U_{1}(y) with P , the problem reduces to Lemma 6.45 . ∎ 6.47 Proposition . Let C:[0,1]\to\mathbb{R}^{3} be a curve. Let \bar{C}=C\frown S be the closed curve obtained by concatenating C with the straight line segment from a=C(1) to b=C(0) with ||a-b||<\tfrac{1}{2} . Let m be a point in the convex hull of \bar{C} , K the m -cone of \bar{C} and x,y\in K with 1\leq d(x,C),d(y,C) and ||x-y||,||x-m||,||y-m||<\tfrac{1}{2}. Let \varphi(t)=\arcsin(2t)-t . Then \mathrm{len}(C)\geq 2\pi+2||x-y||-\varphi(||a-b||) Proof. Let P be a plane containing m , a and b . Then, by Lemma 6.46 , there is a curve \tilde{C} in P connecting a to b obtained by projecting s radially away from m in P with \mathrm{len}(\tilde{C})\leq F(||a-b||) and with 1\leq d(x,\tilde{C}),d(x,\tilde{C}) . Let \tilde{K} be the m -cone of C\frown\tilde{C} . Note that m\in\mathrm{conv}(C\frown S)\subseteq\mathrm{conv}(C\frown\tilde{C}) and that x,y\in\tilde{K} . Since C\frown\tilde{C} avoids U_{1}(\{x,y\}) , we can apply Lemma 4.24 to get \mathrm{len}(C)+\mathrm{len}(\tilde{C})=\mathrm{len}(C\frown\tilde{C})\geq 2\pi+2||x-y||, hence \mathrm{len}(C)\geq 2\pi+2||x-y||-\mathrm{len}(\tilde{C})\geq 2\pi+2||x-y||-\varphi(||a-b||). ∎ This proposition can easily be generalized. 6.48 Proposition . Let C_{i}:[0,1]\to\mathbb{R}^{3} , i=1,\dots,k be curves. Let C be the curve obtained by interpolating between curves C_{i} and C_{i+1} with straight line segments s_{i,i+1} (here k+1 is interpreted as 1 ). Let m be in the convex hull of C , K the m -cone of C , and let x,y\in K with 1\leq d(x,C_{i}),d(y,C_{i}) for i=1,\dots,k and ||x-y||,||x-m||,||y-m||<\tfrac{1}{2} . Let \varphi(t)=\arcsin(2t)-t . Then \sum_{i=1}^{k}\mathrm{len}(C_{i})\geq 2\pi+2||x-y||-\sum_{i=1}^{k}\varphi\left(||C_{i}(1)-C_{i+1}(0)||\right). The following theorem is an "almost-closed" analogon of Proposition 4.25 and an easy consequence of the previous proposition. 6.49 Theorem . Let C_{i}:[0,1]\to\mathbb{R}^{3} , i=1,\dots,k be curves. Let C be the curve obtained by interpolating between curves C_{i} and C_{i+1} with straight line segments s_{i,i+1} (here k+1 is interpreted as 1 ). Let m be in the convex hull of c , K the m -cone of C , and let A\subseteq K with 1\leq\mathrm{dist}(x,C_{i}),\mathrm{dist}(y,C_{i}) for i=1,\dots,k and \sup(\{||x-m||\mid x\in A),\mathrm{diam}(A)<\tfrac{1}{2} . Let \varphi(t)=\arcsin(2t)-t . Then \sum_{i=1}^{k}\mathrm{len}(C_{i})\geq 2\pi+2\mathrm{diam}(A)-\sum_{i=1}^{k}\varphi\left(||C_{i}(1)-C_{i+1}(0)||\right),. Now that we have proved the main tool for our upcoming construction, we will state some technical lemmas. Since their proofs are standard, we have omitted them. 6.50 Lemma . Let D be a parametrized disk, and let C be a curve such that C and D intersect transversely. Then there is an \varepsilon>0 such that ||D-\tilde{D}||,||C-\tilde{C}||<\varepsilon\implies\tilde{C}\cap\tilde{D}\neq\emptyset. 6.51 Lemma . Let A\subseteq\mathbb{R}^{3} be compact and B\subseteq\mathbb{R}^{3} be closed with A\cap B=\{p\} . Then, for every \varepsilon>0 there exists a \delta>0 such that for all x\in A , d(x,B)<\delta\implies d(x,p)<\varepsilon . 6.52 Lemma . Let C_{1},C_{2},\dots,C_{n} be n curves, all defined on compact intervals such that p=C_{1}(x_{1})=C_{2}(x_{2})=\dots=C_{n}(x_{n}) is the only intersection point of C_{1} and C_{2} . Further assume that every C_{i} is an embedding. For every \varepsilon>0 there is a \delta>0 such that the following holds: Let ||C_{i}-\tilde{C}_{i}||<\delta for all i . Let y_{1},\dots,y_{n} be such that (\tilde{q}_{1},\tilde{q}_{2},\dots,\tilde{q}_{n})=(\tilde{C}_{1}(y_{1}),\tilde{C}_{1}(y_{2}),\dots,\tilde{C}_{n}(y_{n}))\in\tilde{C}_{1}\times\tilde{C}_{2}\times\dots\times\tilde{C}_{n} minimizes \mathrm{diam}\left(\{v_{1},v_{2},\dots,v_{n}\}\right)|_{v_{i}\in\tilde{C}_{i}} . Then for all i,j\in\{1,\dots,n\} , we have ||\tilde{q}_{i}-\tilde{q}_{j}||,||\tilde{q}_{i}-p||,|x_{i}-y_{i}|<\varepsilon. 6.53 Lemma . Let D be a parametrized disk and let C be a curve, both with compact definition domains, such that C intersects D in exactly one point p\in\mathbb{R}^{3} and such that this intersection is transverse. For every \varepsilon>0 there is a \delta>0 such that ||D-\tilde{D}||,||C-\tilde{C}||<\delta,\hskip 5.69046pt\tilde{q}\in\tilde{C}\cap\tilde{D}\implies||\tilde{q}-p||<\varepsilon. 6.54 Lemma . Let C:I\to\mathbb{R}^{3} be a uniformly continuous curve and \varepsilon>0 . Then there is a \delta>0 such that for all \tilde{C} with ||\tilde{C}-C||<\delta and all intervals a,b,a^{\prime},b^{\prime}\in I with ab\end{cases},\quad\bar{C}^{\prime}(t)=\begin{cases}C^{\prime}(t)\text{ if }t\in[a^{\prime},b^{\prime}]\\ C^{\prime}(a^{\prime})\text{ if }tb^{\prime}\end{cases}. 6.55 Lemma . Let C:I\to\mathbb{R}^{3} be a uniformly continuous curve and \varepsilon>0 . Then there exists a \delta>0 such that for every C^{\prime} with ||C^{\prime}-C||<\delta and every collection of n subcurves C_{i}:[a_{i},b_{i}]\to\mathbb{R}^{3} , i=1,\dots,n of C and of n subcurves C^{\prime}_{i}:[a^{\prime}_{i},b^{\prime}_{i}]\to\mathbb{R}^{3} of C^{\prime} with |a_{i}-a^{\prime}_{i}|,|b_{i}-b^{\prime}_{i}|<\varepsilon we have ||C-C^{\prime}||<\varepsilon, where C=\bar{C}_{1}\frown S_{1}\frown\bar{C}_{2}\frown S_{2}\frown\dots\frown\bar{C}_{n}\frown S_{n}, C^{\prime}=\bar{C}^{\prime}_{1}\frown S^{\prime}_{1}\frown\bar{C}^{\prime}_{2}\frown S^{\prime}_{2}\frown\dots\frown\bar{C}^{\prime}_{n}\frown S^{\prime}_{n}, are closed curves defined by concatenation of curves defined as follows: \bar{C}_{i} and \bar{C}^{\prime}_{i} are defined as in Lemma 6.54 , and S_{i},S^{\prime}_{i} are straight line segments interpolating between the curves \overline{C}_{i} , \overline{C}_{i+1} and the curves \overline{C}^{\prime}_{i} , \overline{C}^{\prime}_{i+1} respectively (here indices are understood to be cyclic). 6.56 Theorem . Let L=C_{1}\cup C_{2} be a two-component link, with a suitable conjoined circle decomposition. Then there is a neighborhood U around L , such that L^{\prime}\in U\implies\mathrm{len}(L^{\prime})\geq\mathrm{len}(L) with equality if and only if L^{\prime} has a suitable conjoined circle-decomposition with the same index sets, the same index functions h^{A} and h^{B} , and the same linking scheme. Proof. Let \tilde{L}=\tilde{C}_{1}\cup\tilde{C}_{2} be a thick link with ||L-\tilde{L}||=d . For each A -index l and B -index l , pick closed intervals \{H_{i}\}_{i\in\alpha_{l}} with a_{i}\in H_{i}^{\circ} , and pick closed intervals \{H_{i}\}_{i\in\beta_{l}} with b_{i}\in H_{i}^{\circ} , such that \{H_{i}\mid i\text{ }A\text{-index}\} are pairwise disjoint and such that \{H_{i}\mid i\text{ }B\text{-index}\} are pairwise disjoint. Now, define sets of points \{\tilde{a}_{i}\}_{i\in\alpha_{l}} such that \tilde{a}_{i}\in H_{i} and \mathrm{diam}(\tilde{C}_{1}(\{\tilde{a}_{i}\}_{i\in\alpha_{l}})) is minimized, and \{\tilde{b}_{i}\}_{i\in\beta_{l}} such that \tilde{b}_{i}\in H_{i} and \mathrm{diam}(\tilde{C}_{2}(\{\tilde{b}_{i}\}_{i\in\beta_{l}})) is minimized. Since the circle decomposition is suitable, we can assume that there are always two curves in \{C_{1}\mid_{H_{i}}\}_{i\in\alpha_{l}} only intersecting in C_{1}(\alpha_{l}) and two curves in \{C_{2}\mid_{H_{i}}\}_{i\in\beta_{l}} only intersecting in C_{2}(\beta_{l}) . This allows us to apply Lemma 6.52 to the families of curves \{C_{1}\mid_{H_{i}}\}_{i\in\alpha_{l}} and \{C_{2}\mid_{H_{i}}\}_{i\in\beta_{l}} . We get the following result. For any \mathbf{\delta_{1}>0} , there is \mathbf{\delta>0} such that \mathbf{d<\delta} implies |a_{i}-\tilde{a}_{i}|<\delta_{1},\quad|b_{i},\tilde{b}_{i}|<\delta_{1}, \mathrm{dist}(C_{1}(\alpha_{l}),\tilde{C}_{1}(\tilde{a_{i}}))<\delta_{1},\quad\mathrm{dist}(C_{2}(\beta_{l}),\tilde{C}_{2}(\tilde{b_{i}}))<\delta_{1}, \mathrm{diam}(\tilde{C}_{1}(\alpha_{l}))<\delta_{1},\quad\mathrm{diam}(\tilde{C}_{2}(\beta_{l}))<\delta_{1}, where \tilde{C}_{1}\left(\alpha_{l}\right):=\{\tilde{C}_{1}(\tilde{a}_{i})\mid i\in\alpha_{l}\} and \tilde{C}_{2}(\beta_{l}):=\{\tilde{C}_{2}(\tilde{b}_{i})\mid i\in\beta_{l}\} . For A -indices i , we define \tilde{I}_{i}:=[\tilde{a}_{i},\tilde{a}_{i+1}] and for B -indices i we define \tilde{I}_{i}:=[\tilde{b}_{i},\tilde{b}_{i+1}] . The decompositions \left(\tilde{C}_{1},\{\tilde{a}_{i}\},\{A_{j}\},\{\alpha_{l}\}\right) and \left(\tilde{C}_{2},\{\tilde{b}_{i}\},\{B_{j}\},\{\beta_{l}\}\right) should be thought of almost-closed versions of a conjoined circle decomposition of \tilde{L} . To every circle K_{j} of the link L we associate a spanning disk D_{j} and a center z_{j} . To investigate the length of \tilde{L} , we define perturbed versions of z_{j} and D_{j} for \tilde{C}_{1} , the case for \tilde{C}_{2} follows analogously. Lemma 6.55 gives us that, for every \mathbf{\delta_{2}>0} the straight line interpolation \tilde{C}_{1}^{j} between the curves \{\tilde{C}_{1}|_{\tilde{I}_{i}}\mid i\in A_{j}\} satisfies ||C_{1}^{j}-\tilde{C}_{1}^{j}||<\delta_{2} as long as \mathbf{\delta} and \mathbf{\delta_{1}} are sufficiently small. We define \tilde{z}_{j} as the center of mass of \tilde{C}_{1}^{j} and we define \tilde{D}_{j} as the \tilde{z}_{j} -cone of \tilde{C}_{1}^{j} . Since the center of mass of a curve varies continuously with that curve, we can assume \tilde{z}_{j} to be sufficiently close to z_{j} . In particular we can assume that ||z-z_{j}||<\tfrac{1}{2} . We then get that, as parametrized disks, ||\tilde{D}_{j}-D_{j}|| is arbitrarily small. Let \delta_{3}>0 . For \delta , \delta_{1} , \delta_{2} small enough we have ||\tilde{C}_{1}^{j}-C_{1}^{j}||,||\tilde{D}_{j}-D_{j}||<\delta_{3}. Now let j be an A -index and let l=h^{B}(j) , so C_{2}(\beta_{l})=z_{j} . Let i\in\alpha_{l} . Then C_{2}(b_{i})=z_{j}\in D_{j} , and this intersection is transverse, so C_{2}|_{H_{i}} and D_{j} have a transverse intersection. By Lemma 6.50 and Lemma 6.53 , for sufficiently small \delta_{3} , we have S_{i,j}\coloneqq\tilde{C}_{2}(H_{i})\cap\tilde{D}_{j}\neq\emptyset and that for S_{j}\coloneqq\bigcup\limits_{i\in\alpha_{l}}S_{i,j} \sup(\{||x-\tilde{z}_{j}||\mid x\in S_{j}\}),\mathrm{diam}(S_{j})<\frac{1}{2}. By construction of \{\tilde{b}_{i}\mid i\in\beta_{l}\} we have \mathrm{diam}(\tilde{C}_{2}(\tilde{\beta}_{l}))\leq\mathrm{diam}(S_{j}). Take \tilde{z}_{j} as the point in the convex hull of \tilde{C}_{1}^{j} . Note that by construction S_{j} is a subset of \tilde{D}_{j} , which is the \tilde{z}_{j} -cone of \tilde{C}_{1}^{j} , and note that \sup(\{||x-\tilde{z}_{j}||\mid x\in S_{j}\}<\tfrac{1}{2} . Let F(x)=\arcsin(2x)-x and consider C_{1}^{j}=C_{i_{1}}^{s_{1}}\frown C_{i_{2}}^{s_{2}}\frown\dots\frown C_{i_{c}}^{s_{c}}. For s_{k}=1 , let x_{k}=i_{k} and y_{k}=i_{k}+1 . For s_{k}=-1 , let x_{k}=i_{k}+1 and y_{k}=i_{k} . Then C_{i_{k}}^{s_{k}} has startpoint C_{1}(a_{x_{k}}) and endpoint C_{1}(a_{y_{k}}) . By Theorem 6.49 , we have \displaystyle\sum\limits_{i\in A_{j}}\mathrm{len}\left(\tilde{C}_{1}|_{\tilde{I}_{i}}\right)\geq \displaystyle 2\pi+2\mathrm{diam}(S_{j})-\sum_{k=1}^{c}\varphi(||\tilde{C}_{1}(\tilde{a}_{y_{k}})-\tilde{C}_{1}(\tilde{a}_{x_{k+1}})||) \displaystyle\geq \displaystyle 2\pi+2\mathrm{diam}(\tilde{C}_{2}(\beta_{h^{B}(j)}))-\sum_{k=1}^{c}\varphi(\mathrm{diam}(\tilde{C}_{1}(\alpha_{l(y_{k})}))) \displaystyle= \displaystyle 2\pi+2\mathrm{diam}(\tilde{C}_{2}(\beta_{h^{B}(j)}))-\sum_{\alpha_{l}\text{ along }K_{j}}\varphi(\mathrm{diam}(\tilde{C}_{1}(\alpha_{l}))) where we used that x_{k+1},y_{k}\in\alpha_{l(y_{k})} . Let q^{A}(l)=\mathrm{diam}(\tilde{C}_{1}(\alpha_{l})) , q^{B}(l)=\mathrm{diam}(\tilde{C}_{2}(\beta_{l})) , and let f^{A}(l)=\#\{j\mid\alpha_{l}\text{ along }C_{1}^{j}\} , f^{B}(l)=\#\{j\mid\beta_{l}\text{ along }C_{2}^{j}\} . Let n=\#\{A\text{-indices }j\} . By summing over all j , we get \mathrm{len}(\tilde{C}_{1})\geq 2m\pi+2\sum\limits_{j\text{ }A\text{-index }}q^{B}(h^{B}(j))-\sum_{l\text{ }A\text{-index}}f^{A}(l)\cdot\varphi(q^{A}(l)). Exactly analogously, for m=\#\{B\text{-indices }j\} , we have \mathrm{len}(\tilde{C}_{2})\geq 2m\pi+2\sum\limits_{j\text{ }B\text{-index }}q^{A}(h^{A}(j))-\sum_{l\text{ }B\text{-index}}f^{B}(l)\cdot\varphi(q^{B}(l)). In total, we get \displaystyle\mathrm{len}(\tilde{C}_{1})+\mathrm{len}(\tilde{C}_{2})\geq \displaystyle 2(n+m)\pi \displaystyle+ \displaystyle\sum_{l\text{ }A\text{-index}}2\#(h^{A})^{-1}(l)\cdot q^{A}(l)-f^{A}(j)\cdot\varphi(q^{A}(l)) \displaystyle+ \displaystyle\sum_{l\text{ }B\text{-index}}2\#(h^{B})^{-1}(l)\cdot q^{B}(l)-f^{B}(j)\cdot\varphi(q^{B}(l)) Since near zero we have that \varphi(x)\sim x to first order, for m=\min\left(\left\{\frac{2\#(h^{A})^{-1}(l)}{f^{A}(l)}\right\}\cup\left\{\frac{2\#(h^{B})^{-1}(l)}{f^{B}(l)}\right\}\right)>1, there is \delta_{4}>0 such that on [0,\delta_{4}] we have F(x)\leq mx with equality if and only if x=0 . In particular, since q^{A}(l),q^{B}(l)<\delta_{1} , for \delta_{1}<\delta_{4} , we have 2\#(h^{A})^{-1}(l)\cdot q^{A}(l)-f^{A}(l)\cdot\varphi(q^{A}(l))\geq 0, 2\#(h^{B})^{-1}(l)\cdot q^{B}(l)-f^{B}(l)\cdot\varphi(q^{B}(l))\geq 0, with equality if and only if q^{A}(l)=0 and q^{B}(l)=0 respectively. Thus, for d=||\tilde{L}-L|| small enough, we have \mathrm{len}(\tilde{L})=\mathrm{len}(\tilde{C}_{1})+\mathrm{len}(\tilde{C}_{2})\geq 2(n+m)\pi=\mathrm{len}(C_{1})+\mathrm{len}(C_{2})=\mathrm{len}(L). Therefore L=C_{1}\cup C_{2} is a local minimum. Assume now that \mathrm{len}(\tilde{L})=\mathrm{len}(L) . Claim: Both \tilde{C}_{1} and \tilde{C}_{2} circle-decompose with the same index sets as C_{1} and C_{2} respectively, and the linking scheme stays the same. Further, the circle decomposition of \tilde{C}_{1} and \tilde{C}_{2} is conjoined and suitable. We have equality \mathrm{len}(\tilde{C}_{1})+\mathrm{len}(\tilde{C}_{2})=2(n+m)\pi if and only if q^{A}(l)=0 for every A -index l and q^{B}(l)=0 for every B -index l . This implies that all clusters \alpha_{l} and \beta_{l} map to a single point. So the straight line interpolations used to define the curves \tilde{C}_{1}^{j} and \tilde{C}_{2}^{j} were unnecessary. Since every circle K_{j} of one of the curves was linked to some circle K_{j^{\prime}} of the other curve, and since \tilde{L} is close to L , every curve \tilde{C}_{1}^{j} will be linked to some curve \tilde{C}_{2}^{j^{\prime}} and vice-versa. Every curve \tilde{C}_{1}^{j} and every curve \tilde{C}_{2}^{j} needs to have length at least 2\pi , and through the global length constraint of \mathrm{len}(\tilde{L})=2(n+m)\pi we get that every curve \tilde{C}_{1}^{j} and every curve \tilde{C}_{2}^{j} has to have length exactly 2\pi , hence has to be a unit circle. This shows that \tilde{C}_{1} and \tilde{C}_{2} have circle decompositions with the same index sets as C_{1} and C_{2} respectively. That the linking scheme is the same is trivial since for every pair of indices (j,j^{\prime}) , the link \tilde{K}_{j}\cup\tilde{K}_{j^{\prime}} is close to K_{j}\cup K_{j^{\prime}} . It is easy to check that the circle decomposition of \tilde{C}_{1} and \tilde{C}_{2} is conjoined, with the same functions h^{A} and h^{B} and the same linking scheme as the conjoined circle decomposition of C_{1} and C_{2} , and that this circle decomposition is therefore suitable. ∎ We will now prove Theorem 6.44 . Proof. In the last theorem we showed that S is strictly locally minimizing. It remains to be shown that S is closed. Take a sequence (L_{n})_{n\in\mathbb{N}} in S converging to some thick link L=(C_{1},C_{2}) . Let \{A_{j}\},\{\alpha_{l}\},\{B_{j}\},\{\beta_{l}\} be the index sets and let h^{A},h^{B} be the index functions belonging to S . For each n\in\mathbb{N} , consider the sets of breaking points \{(a_{i})_{n}\},\{(b_{i})_{n}\}\subseteq S^{1} . Since S^{1} is compact, we can assume that all these sequences ((a_{i})_{n})_{n\in\mathbb{N}} and ((b_{i})_{n})_{n\in\mathbb{N}} converge, else take a subsequence of (L_{n})_{n\in\mathbb{N}} . We denote the limits by a_{i} and b_{i} respectively. Note that the \{(a_{i})_{n}\} and \{(b_{i})_{n}\} each have the same cyclic ordering for all n , hence the cyclic orderings of \{a_{i}\} and \{b_{i}\} are (weakly) preserved when taking limits (but a priori it seems possible that a_{i}=a_{i+1} or b_{i}=b_{i+1} for some i ). We will now show that (\{A_{j}\},\{\alpha_{l}\},\{a_{i}\}),(\{B_{j}\},\{b_{i}\},\{\beta_{l}\}) form a suitable conjoined circle decomposition of L . Claim: \{A_{j}\} , \{\alpha_{l}\} , \{a_{i}\} form a circle decomposition of C_{1} and \{B_{j}\} , \{b_{i}\} , \{\beta_{l}\} form a circle decompositon of C_{2} . We show this claim for C_{1} , the proof for C_{2} is the same. It is obvious that circles (C_{1}^{j})_{n} of (C_{1})_{n} converge to circles C_{1}^{j} of C_{1} . It is left to show that C_{1}(a_{i_{1}})=C_{1}(a_{i_{2}})\iff\text{ there is some }l\text{ with }i_{1},i_{2}\in\alpha_{l}. It is easy to see that we have i_{1},i_{2}\in\alpha_{l}\implies C_{1}(a_{i_{1}})=C_{1}(a_{i_{2}}). Now assume that i_{1}\in\alpha_{l_{1}} and i_{2}\in\alpha_{l_{2}} with l_{1}\neq l_{2} . Let K_{j_{1}} and K_{j_{2}} be the unit circles of C_{2} with z_{j_{1}}=C_{1}(\alpha_{l_{1}}) and z_{j_{2}}=C_{1}(\alpha_{l_{2}}) . Since l_{1}\neq l_{2} , we have that (C_{1})_{n}(\alpha_{l_{1}})\neq(C_{1})_{n}(\alpha_{l_{2}}) for every n\in\mathbb{N} . It follows that for every n\in\mathbb{N} we have (z_{j_{1}})_{n}\neq(z_{j_{2}})_{n} and therefore (K_{j_{1}})_{n}\neq(K_{j_{2}})_{n} . Since the circle decompositions of every L_{n} are suitable, there is some A -index j^{\prime} such that for every n the circle (K_{j^{\prime}})_{n} is non-trivially linked to (K_{j_{1}})_{n} , but trivially linked to (K_{j_{2}})_{n} . Taking limits, we get that K_{j^{\prime}} is non-trivially linked to K_{j_{1}} but trivially linked K_{j_{2}} , hence K_{j_{1}}\neq K_{j_{2}} . This implies z_{j_{1}}\neq z_{j_{2}} , since if two different circles K_{j_{1}} , K_{j_{2}} of C_{2} had the same center, it would be impossible for C_{1} to pierce their common center while keeping unit distance from both. Therefore C_{1}(a_{i_{1}})=z_{j_{1}}\neq z_{j_{2}}=C_{1}(a_{i_{2}}) . This shows that two A -indices i_{1},i_{2} are in the same point cluster if and only if C_{1}(a_{i_{1}})=C_{1}(a_{i_{2}}) . It is left to check that the intervals I_{i}=[a_{i},a_{i+1}] all have positive length, i.e. a_{i}\neq a_{i+1} . If i,i+1 are in different point clusters, then by what we have just shown C_{1}(a_{i})\neq C_{1}(a_{i+1}) and hence a_{i}\neq a_{i+1} . If i,i+1 are in the same \alpha_{l} , then there is some j with A_{j}=\{i\} . This means that (C_{1})_{n}([(a_{i})_{n},(a_{i+1})_{n}])=(K_{j})_{n} is a unit circle for each n . By taking limits, C_{1}([a_{i},a_{i+1}])=K_{j} is a unit circle, hence I_{i} has positive length. Claim: The circle decompositions of C_{1} and C_{2} are conjoined. We check the following three conditions 1. d(C_{1},C_{2})=1 , 2. K_{j} circle of C_{1} \implies z_{j}\in C_{2}(\{b_{i}\}) and every intersection of C_{2} and D_{j} is transverse 3. K_{j} circle of C_{2} \implies z_{j}\in C_{1}(\{a_{i}\}) and every intersection of C_{1} and D_{j} is transverse The first condition is obviously fulfilled. We will show that the second condition is fulfilled, the third one follows analogously. Let K_{j} be a circle of C_{1} . Let i be such that z_{j}^{n}=(C_{2})_{n}((b_{i})_{n}) for every n . Then by taking limits, z_{j}=C_{2}(b_{i}) . We omit the proof that every intersection is transverse after taking limits. The conjoined circle decomposition has the same index functions h^{A} and h^{B} as that of L . Note that taking limits preserves both thick Hopf links and thick unlinks, so we have K_{j}\text{ linked to }K_{j^{\prime}}\iff K_{j}^{n}\text{ linked to }K_{j^{\prime}}^{n}. This shows that L has the same linking scheme as each L_{n} . Claim: The conjoined circle decomposition of L is suitable. The conditions for suitability are 1. Every circle K_{j} of one of the curves is linked to exactly two circles K_{j_{1}},K_{j_{2}} of the other curve, 2. for every A -index l : \#\{j\text{ }A\text{-index }\mid\alpha_{l}\text{ along }K_{j}\}<2\#(h^{A})^{-1}(l) , 3. for every B -index l : \#\{j\text{ }B\text{-index }\mid\beta_{l}\text{ along }K_{j}\}<2\#(h^{B})^{-1}(l) , 4. for every \alpha_{l} , there are two distinguishable i_{1},i_{2}\in\alpha_{l} , 5. for every \beta_{l} , there are two distinguishable i_{1},i_{2}\in\beta_{l} . The first condition follows trivially from the linking scheme of L and L_{n} being the same. For the second condition, note that \{j\text{ }A\text{-index }\mid\alpha_{l}\text{ along }C_{1}^{j}\}=\{j\text{ }A\text{-index }\mid\alpha_{l}\text{ along }(C_{1}^{j})_{n}\}. Since h^{A},h^{B} are the index functions of both L and S , the second condition is fulfilled. The third condition holds analogously. The last two conditions follow by restating them as conditions about the linking scheme of L , which is the same as the linking scheme of S : Take two indices i_{1},i_{2}\in\alpha_{l} . Then i_{1} and i_{2} are distinguishable if and only if \{K_{j}\mid C_{1}(a_{i})\in K_{j}\}\cap\{K_{j}\mid C_{1}(a_{i_{2}})\in K_{j}\}=\emptyset . Since every circle K_{j} of C_{1} is linked to exactly two other circles of C_{2} , we have K_{j_{1}}=K_{j_{2}} if and only if K_{j_{1}} and K_{j_{2}} are linked to the same set of circles. ∎ We can now apply Proposition 3.20 . 6.57 Corollary . Every 2 -component thick link L that has a suitable circle decomposition is a local minimum for ropelength and a sink. 6.3 A Gordian unlink and other examples 6.58 Theorem . A two-component Gordian unlink exists. Proof. One carefully checks that the link shown in Figure 7 , parametrized as indicated in 9 has a suitable conjoined circle decomposition with linking number equal to zero. ∎ 6.59 Theorem . For every n\in\mathbb{N} , there exists a 2 -component thick link L with linking number n , which is a non-global, local minimum and a sink. In particular, there are Gordian pairs in every link homotopy class with 2 components. Proof. Let (C_{1},C_{2}) be the Gordian unlink from Theorem 6.58 and with parametrization indicated as in Figure 9 . If n=0 , we can just take L=(C_{1},C_{2}) and we are done, so let n\geq 1 . Let C_{1} be the curve with colors yellow, green, and blue, parametrized by starting at yellow–grey breaking point of the point cluster F and going along the grey segment. Let S_{1} be the curve parametrizing the grey unit circle around E of the sketch, parametrized with the same orientation it has as a part of C_{1} . Let C_{2} be the curve with the colors red, purple, and orange, parametrized by starting at the red–grey breaking point of the point cluster A and going along the grey segment. Let S_{2} be the curve parametrizing the grey unit circle around F , parametrized with the same orientation it has as a part of C_{1} . Now consider \tilde{C}_{1} defined by concatenating k copies of C_{1} and n copies of S_{1} , and consider \tilde{C}_{2} defined by concatenating k copies of C_{2} and one copy of S_{2} . Let \mathrm{link}(\cdot,\cdot) refer to the linking number of two curves. Then \displaystyle\mathrm{link}(\tilde{C}_{1},\tilde{C}_{2})= \displaystyle k^{2}\mathrm{link}(C_{1},C_{2})+nk\mathrm{link}(S_{1},C_{2})+k\mathrm{link}(C_{1},S_{2})+n\mathrm{link}(S_{1},S_{2}) \displaystyle= \displaystyle 0+0+0+n\mathrm{link}(S_{1},S_{2})=n where we used that \mathrm{link}(C_{1},C_{2})=\mathrm{link}(S_{1},C_{2})=\mathrm{link}(C_{1},S_{2})=0 . It is trivial to give (\tilde{C}_{1},\tilde{C}_{2}) a conjoined circle decomposition. As long as k\geq\max(2,n-1) , this circle decomposition will be suitable. This shows that L is a local minimum and a sink. For n\geq 1 , there is a thick link L^{\prime} of length \mathrm{len}(L^{\prime})=2(n+1)\pi<(24k+n+1)\pi=\mathrm{len}(L) in the link homotopy class of L , which shows that L is not a global minimum and that (L,L^{\prime}) is a Gordian pair. ∎