AI methodology: § 15 .
Introduction
The classical cactus theorem of Dinitz, Karzanov, and Lomonosov from 1976 [48 ] says that the set of minimum edge-cuts of every connected graph has the structure of a cactus1 . The DKL cactus representation can be computed in near-linear time [69 ] , and hence can be used to enumerate all minimum cuts efficiently [KP09, LST21]. It is a key ingredient in algorithms for connectivity augmentation [Gab16, NGM97, CLP22] and for maintaining minimum cuts under edge insertions [Hen97, DW98, GHT18]. Vertex-connectivity encompasses edge-connectivity as a special case via the line-graph construction. However, the direct extension of the cactus theorem to arbitrary vertex-connectivity fails. The connection between vertex-connectivity and tree-decompositions makes this question particularly relevant for algorithmic applications.
The quest for such an analogue goes back at least to Tutte’s seminal work from 1966 [Tut66], even predating the DKL theorem. The search for a cactus-type representation of minimum vertex-separations2 motivated the work of Cohen, Di Battista, Kanevsky, and Tamassia [21 ] and subsequent structural and algorithmic investigations [91 , 106 , 112 ] . The broader problem of decomposing a k k k -connected3 graph into ( k + 1 ) (k+1) ( k + 1 ) -connected regions has been intensively studied in the last decade [25 , 65 Erd17, 42 , 7 ] . These results reveal substantially richer interactions between minimum vertex-separations than between minimum edge-cuts, making the formulation of a suitable global representation a challenge in its own right. In this paper, we establish an approximate cactus-decomposition theorem for arbitrary vertex-connectivity.
Our starting point is the SPQR decomposition – the algorithmic counterpart of Tutte’s 2-separator theorem – in the cactus-formulation of [EP14]. Informally, a cactus-decomposition of a graph G G G is defined like a tree-decomposition, except that the decomposition tree is replaced by a strong cactus C C C , some of whose nodes are designated as ‘P P P -nodes’. The decomposition displays separations of G G G through four types of cuts of C C C : a bridge, two edges on a common cycle, a cut consisting of edges incident with a single P P P -node, or the empty cut4 ; see Figure 1.1 . Each such cut defines a bipartition of the nodes of C C C , which we pull back to a separation of G G G by taking the union of the bags on each side. The resulting separation is displayed by the cactus-decomposition; see § 3.4 for the formal definitions. Cactus-decompositions extend the tree structure of nested separations in the simplest way, by allowing cycles to accommodate coherent crossings. The SPQR theorem is implied5 by the following variant [EP14]: Every 2-connected graph has a cactus-decomposition displaying all its proper 2-separations.
Figure 1.1. A 2-connected graph (black) and a cactus-decomposition (red), with each cactus node at the centre of its corresponding blue bag and the P P P -nodes drawn in yellow. Deleting a bridge or two edges of a common cycle splits the cactus into two components; the unions of their bags form the two sides of a displayed separation. The bags of the P P P -nodes, circled in green, form correspond to 2-separators whose deletion leaves more than two components: these may be partitioned arbitrarily into two groups, with the separator included on both sides. The cycles of the cactus encode the crossing patterns of minimum cuts, while the bridges encode the cuts that avoid every crossing.
SPQR trees have become a standard algorithmic tool, underlying classical algorithms for planarity testing and planar embeddings [71 , 100 ] , planar graph isomorphism [70 , 72 ] , and dynamic connectivity and planarity problems [21 , 39 , 40 ] .
For higher vertex-connectivity, exact cactus-decompositions are no longer possible in general: some graphs admit no cactus-decomposition simultaneously displaying all their minimum-order separations; this happens already for 3-separations, see [32 ] for a detailed discussion. For general k k k , cylindrical grids, such as the one in Figure 1.2 , provide examples of obstructions to an exact cactus-decomposition in which the necessary approximation error grows with k k k . Despite the linear global structure of these graphs, their k k k -separations admit too much local freedom to all be displayed by a single cactus-decomposition.
Figure 1.2. A 6-connected graph obtained from a cylindrical grid by adding both diagonals in alternate faces and attaching a large clique at each end. The top and bottom rows are identified. The three coloured separations pairwise cross. They cannot be displayed at a common P P P -node, since their separators are distinct, or on a common cycle, since each separator meets all four corners of the other two. Thus no cactus-decomposition displays all the 6-separations of this graph.
To address this for k = 3 k = 3 k = 3 , Carmesin and Kurkofka [32 ] introduced tri-separations , modified versions of 3-separations that retain their separating properties while removing the local ambiguity that prevents an exact cactus-decomposition. With this, they obtained a Tutte-type decomposition of 3-connected graphs. Kurkofka and Planken [84 ] subsequently introduced tetra-separations and proved a Tutte-type decomposition theorem for 4-connected graphs. These exact results for modified notions of separation can be viewed as approximate cactus-decomposition theorems for ordinary vertex-separations: for k = 3 , 4 k = 3, 4 k = 3 , 4 , every large enough k k k -connected graph G G G admits a cactus-decomposition displaying k k k -separations such that every k k k -separation of G G G has edit-distance at most k k k to a displayed k k k -separation; here, the edit-distance between separations ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) is measured by max { ∣ A △ C ∣ , ∣ B △ D ∣ } \max\{ |A \mathbin{\triangle} C|, |B \mathbin{\triangle} D| \} max { ∣ A △ C ∣ , ∣ B △ D ∣ } .
Our main result generalizes this approximate cactus picture to arbitrary vertex-connectivity:
Theorem 1.1 (Simplified version for k k k -connected graphs). For every k ∈ N k \in\mathbb{N} k ∈ N , there is d ∈ N d \in\mathbb{N} d ∈ N such that every k k k -connected graph G G G admits a cactus-decomposition displaying k k k -separations such that every k k k -separation of G G G has edit-distance at most d d d to a displayed separation.
A graph can have quadratically many distinct separators of minimum size [78 ] , and a single separator can give rise to exponentially many separations. For fixed k k k , our theorem gives a linear-size data structure that approximately represents all k k k -separations of a k k k -connected graph. These are represented implicitly by the cactus and its bags through bridges, cycles and P P P -nodes. Compact representations of minimum vertex separations have also been studied in connection with pairwise connectivity queries [KTDBC91, CDBKT93, PY21]; see § 2.3 .
Unbreakability is a standard relaxation of the notion of connectivity in parameterized algorithms, introduced in the context of recursive understanding [92 , 20 LRSZ18, 35 ] , which reduces many graph problems to suitable unbreakable instances. In this paper, we prove a strengthening of Theorem 1.1 with the assumption of ‘k k k -connectivity’ relaxed to ‘( k − 1 ) (k-1) ( k − 1 ) -unbreakability’. This generalisation allows us to apply our result iteratively.
Formally, a graph is ( q , k ) (q,k) ( q , k ) -unbreakable if every separation of order at most k k k has a proper side of size at most q q q . Thus, the k k k -connected graphs are exactly the ( 0 , k − 1 ) (0,k-1) ( 0 , k − 1 ) -unbreakable graphs with more than k k k vertices, while allowing q > 0 q>0 q > 0 gives an approximate version of k k k -connectivity. This relaxation allows us to look beyond the local obstructions imposed by low-degree vertices and capture the global connectivity of graphs such as square grids and cubic expanders. Our main result holds in this more general setting:
Theorem 1.2 (Main theorem). For all q , k ∈ N q,k \in\mathbb{N} q , k ∈ N , there is d ∈ N d \in\mathbb{N} d ∈ N such that every ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph G G G admits a cactus-decomposition displaying ( ≤ k ) (\le k) ( ≤ k ) -separations such that every ( ≤ k ) (\le k) ( ≤ k ) -separation of G G G has edit-distance at most d d d to a displayed separation.
The approximation bound d d d is effective and can be taken to be 2 ( q + k ) O ( 1 ) 2^{(q+k)^{O(1)}} 2 ( q + k ) O ( 1 ) . We leave the dependence on q q q and k k k implicit in our proof, prioritising simplicity over optimizing bounds.
Iterated decompositions have proved useful in structural graph theory [30 ] , graph enumeration [GNR13], and algorithms for connectivity [47 , 37 ] and graph isomorphism [DNTW10, DLN+22]. Theorem 1.2 can be applied recursively as follows. We first group together the bags whose nodes lie on the same cycle of the cactus, obtaining pieces with cycle-decompositions of bounded adhesion. We retain these cyclic pieces, as well as the bags at P P P -nodes (which have size at most k k k ). Each remaining bag can be described by a k k k -unbreakable minor of G G G , to which we apply Theorem 1.2 with k k k increased by one to reveal finer structure, see Section A.
Examples and the need for approximation
We now illustrate our theorem on a few families of graphs with a simple underlying structure, showing how the cactus-decomposition can recover this structure. Some of these examples also justify the need for approximation.
Already for k = 3 , 4 k = 3, 4 k = 3 , 4 , there are k k k -connected graphs whose k k k -separations cannot all be displayed by a single cactus-decomposition. Toroidal hexagonal and square grids provide examples for k = 3 k = 3 k = 3 and k = 4 k = 4 k = 4 , respectively. These examples extend to higher connectivity, for instance by adding universal vertices, or by replacing each vertex v v v by a clique K ( v ) K(v) K ( v ) and replacing each edge u w uw u w with all possible edges between K ( u ) K(u) K ( u ) and K ( w ) K(w) K ( w ) . However, these graphs are ( q , k ) (q, k) ( q , k ) -unbreakable for some q q q depending only on k k k , so every k k k -separation has a side of bounded size. It is therefore natural to accept such graphs as basic pieces of our decomposition, leaving them intact and focusing on sufficiently broad separations, meaning those with both sides sufficiently large.
To see why even these separations require approximation, consider a long cylindrical grid of width k k k , with a large clique attached at each end; see Figure 1.3 . This graph is ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable for some q q q depending only on k k k . Its vertical cross-sections define separations of order k k k . These separations admit local modifications that preserve their order but produce crossing patterns that cannot all be displayed by a single cactus-decomposition, for the same reason as in Figure 1.2 . Since the end cliques can be arbitrarily large, all the separations involved can be arbitrarily broad.
Figure 1.3. Both panels show a cylindrical grid of width 6, with the top and bottom rows identified. Left: the red 6-separation is close to the blue separation defined by a single vertical cross-section. Right: in the 7th power of the grid, the red vertices form a minimum separator. The corresponding separation is close to the blue separation, whose separator consists of seven consecutive cross-sections.
Nevertheless, as illustrated in Figure 1.3 , every sufficiently broad k k k -separation differs from a straight one only on a number of vertices bounded in terms of k k k , independently of the length of the cylinder. The cactus-decomposition constructed in our proof displays precisely the straight separations, thereby recovering the intrinsic linear structure of the graph.
Taking the r r r -th power of this graph yields a k ′ k' k ′ -connected example with k ′ = r k k' = rk k ′ = r k , provided r ≥ k r \ge k r ≥ k and the end cliques are sufficiently large. Now, the vertices of any r r r consecutive vertical cross-sections form a minimum separator. The cactus-decomposition constructed in our proof displays precisely these separations, again recovering the intrinsic linear structure of the graph. More generally, as illustrated in Figure 1.3 , a minimum separator consists of r r r consecutive vertices on each longitudinal path, with starting positions differing by at most one between neighbouring paths. These variable bands give rise to the same obstruction to an exact cactus representation, now among minimum-order separations of a k ′ k' k ′ -connected graph. Thus, strengthening unbreakability to k k k -connectivity does not eliminate the obstruction. Nevertheless, every minimum separation still differs from one displayed by our decomposition only on a bounded number of vertices.
One might hope to recover an exact theorem by generalising tri- and tetra-separations to a suitable notion of reduced mixed-separation, whose separator may contain both vertices and edges. In the cylinder examples, however, the crossing patterns can involve a number of vertices quadratic in the width of the cylinder, and it is not clear how a reduction favouring edges over vertices could resolve them.
A toroidal grid whose longitude is much longer than its meridian provides a cyclic analogue of the cylinder; see Figure 1.4 . Writing m m m for the length of the meridian and setting k = 2 m k = 2m k = 2 m , this graph is again ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable for some q q q depending only on k k k . Pairs of meridional cross-sections define separations of order k k k , and every sufficiently broad k k k -separation differs from one of these only on a bounded number of vertices, as shown in Figure 1.4 . Our construction displays precisely the separations defined by these pairs of cross-sections, thereby recovering the intrinsic cyclic structure of the graph.
Figure 1.4. A toroidal grid with meridians of length m = 16 m = 16 m = 16 and longitudes of length 40, so k = 2 m = 32 k = 2m = 32 k = 2 m = 32 . The two blue meridians together form the separator of a k k k -separation. The red cycles together form the separator of another k k k -separation, whose sides agree with those of the blue one outside the two separators.
Finally, consider the following generalisation of the wheels introduced by Geelen and Joeris [62 ] . Take a long cyclic sequence of arbitrary t t t -vertex trees, join consecutive trees by arbitrary perfect matchings, and add ℓ \ell ℓ hubs, each adjacent to a vertex in every tree; see Figure 1.5 for an example. For k = 2 t + ℓ k = 2t + \ell k = 2 t + ℓ , these graphs are again ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable for some q q q depending only on k k k . Pairs of trees together with all the hubs define separations of order k k k . Every sufficiently broad k k k -separation differs from one of these only on a number of vertices bounded in terms of k k k . Our cactus-decomposition recovers the cyclic arrangement of the trees, with the hubs forming a common core of the displayed separators.
Figure 1.5. Twelve trees with t = 4 t = 4 t = 4 vertices each, joined cyclically by perfect matchings, and ℓ = 2 \ell= 2 ℓ = 2 hubs, so k = 2 t + ℓ = 10 k = 2t + \ell= 10 k = 2 t + ℓ = 10 . The two blue trees and the hubs form the separator of a k k k -separation. The red vertices and the hubs form the separator of another k k k -separation, whose sides agree with those of the blue one outside the two separators.
A short outline of the proof of Theorem 1.1 (see Section 2 for a detailed overview).
We construct the cactus-decomposition of Theorem 1.1 in two steps. First, we build a tree-decomposition T \mathcal{T} T whose bags, roughly speaking, either have higher connectivity or possess a cyclic or star-like structure. We then refine T \mathcal{T} T into the desired cactus-decomposition C \mathcal{C} C by replacing each bag of the latter type with the corresponding cycle or star-decomposition, with the centres of the stars becoming P P P -nodes.
Step 1: Constructing the tree-decomposition T \mathcal{T} T . We construct our tree-decomposition from a maximal nested set of carefully selected separations and use maximality to obtain the desired approximation guarantee. A maximality argument also underlies a modern proof of the tangle-tree theorem [Car16], where the aim is to distinguish the highly connected regions of a graph G G G , formalised by its maximal tangles. We briefly recall the argument. Given a graph G G G , let S S S be the set of all separations efficiently distinguishing two tangles, and choose a maximal nested subset T ⊆ S \mathcal{T} \subseteq S T ⊆ S . The separations in T \mathcal{T} T induce a tree-decomposition T \mathcal{T} T : since they are nested, we can successively split G G G along them to obtain the bags of the decomposition. One then shows that the maximality of T \mathcal{T} T ensures that T \mathcal{T} T distinguishes all maximal tangles.
We now return to the proof of Theorem 1.1 and explain how we use maximality in this setting. Given k ∈ N k \in\mathbb{N} k ∈ N , choose constants β \beta β and γ \gamma γ depending only on k k k ; their precise values are not important for this overview. The approximation bound d d d will be larger than both.
Let G G G be a k k k -connected graph. We say that two k k k -separations ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) heavily cross if all four sets A ∩ C A \cap C A ∩ C , A ∩ D A \cap D A ∩ D , B ∩ C B \cap C B ∩ C and B ∩ D B \cap D B ∩ D have size at least γ \gamma γ . Intuitively, these are the separations that will be captured by the cyclic parts of the cactus, so we set them aside in the first step. Let S S S be the set of all k k k -separations ( A , B ) (A, B) ( A , B ) with ∣ A ∣ , ∣ B ∣ ≥ β |A|, |B| \ge\beta ∣ A ∣ , ∣ B ∣ ≥ β that are not heavily crossed. Choose a maximal nested subset T ⊆ S \mathcal{T} \subseteq S T ⊆ S , and let T \mathcal{T} T be the tree-decomposition induced by T \mathcal{T} T .
Using corner diagram arguments, we show that every separation in S S S has bounded edit-distance to some separation in T \mathcal{T} T . Thus, T \mathcal{T} T approximately displays all separations in S S S . Separations ( A , B ) (A, B) ( A , B ) with ∣ A ∣ < β |A| < \beta ∣ A ∣ < β or ∣ B ∣ < β |B| < \beta ∣ B ∣ < β are already within edit-distance β \beta β of ( ∅ , V ( G ) ) (\varnothing, V(G)) ( ∅ , V ( G )) or ( V ( G ) , ∅ ) (V(G), \varnothing) ( V ( G ) , ∅ ) , respectively, and are therefore covered by our choice of d d d . It remains to refine T \mathcal{T} T into a cactus-decomposition that also approximately displays the heavily crossed separations.
Construction of the cactus-decomposition C \mathcal{C} C
Step 2: Construction of the cactus-decomposition C \mathcal{C} C . We say that a separation ( A , B ) (A,B) ( A , B ) lives in a bag of a tree-decomposition if its separator is contained in that bag. Another corner diagram argument shows that every heavily crossed k k k -separation has bounded edit-distance to a separation that lives in a bag of T \mathcal{T} T and remains heavily crossed (for a slightly smaller value of γ \gamma γ ).
We can therefore build C \mathcal{C} C by treating the bags of T \mathcal{T} T separately. For each bag, we seek a cycle- or star-decomposition of its torso that approximately displays all the heavily crossed k k k -separations living there. Here, the torso is the graph induced by the bag, with each intersection with a neighbouring bag completed to a clique. Constructing these local decompositions is the heart of the proof and relies on the following new ideas:
Each separation ( A , B ) (A,B) ( A , B ) partitions the vertex set into three parts: A ∖ B A \setminus B A ∖ B , A ∩ B A \cap B A ∩ B and B ∖ A B \setminus A B ∖ A . Three separations therefore determine 27 regions. We show that when three k k k -separations pairwise heavily cross, their crossing pattern is much more constrained: many of these regions must be empty, and the whole configuration admits a cyclic description. In particular, heavily crossing separations cannot form truly three-dimensional crossing patterns, such as the one in Figure 1.6 .
Figure 1.6. Three separations cross three-dimensionally in the line-graph of the cube. Their separators are indicated in red, green and blue.
Families of separations with a common separator can be displayed by a star-decomposition. For the cyclic case, we use ideas from covering spaces in topology to combine the cyclic descriptions of triples in (i) into a single cycle-decomposition that approximately displays the family. Such cycle- and star-decompositions correspond to flowers from the matroid literature [105 , 11 ] ; Figure 1.7 illustrates the cyclic case.
Figure 1.7. A cycle-decomposition (blue bags, red decomposition cycle) of a graph G G G (black) that is a 4-flower
Several crossing families may occur in the same bag of T \mathcal{T} T , giving rise to several flowers. However, by choosing d d d sufficiently large compared to β \beta β and γ \gamma γ , we show that one flower per bag suffices: we can select a flower F \mathcal{F} F in each bag so that every k k k -separation living in this bag is within bounded edit-distance of a separation displayed by F \mathcal{F} F or T \mathcal{T} T . This requires a
systematic analysis of flowers which, together with the canonical constructions via coverings in (ii), extends the structural theory of flowers in graphs, and provides, for the first time, canonical constructions of flowers.
This completes the second step of the construction and thus our brief proof overview. We give a more detailed proof overview in Section 2, expanding on each of these three ideas.
A brief real-world example with a cyclic structure
Figure 1.8 shows an eight-petal flower in a contact graph obtained from the human nuclear-pore model PDB 7TBJ of Petrovic et al. [109 ] .
Figure 1.8. An eight-petal flower in a biological contact graph. Left: the largest component of the contact graph of the human nuclear-pore model PDB 7TBJ (Petrovic et al.). Vertices represent protein fragments; edges record contacts within 5 5 5 Å between non-hydrogen atoms. Vertex positions are obtained by projecting the fragment centres onto the x y xy x y -plane, so distinct vertices may overlap in the drawing. In particular, the six red vertices form a separator. This is difficult to discern in the projected drawing at the displayed scale, but becomes clear at 400% magnification. Right: the eight bags form a cycle with three-vertex overlaps; matching colours identify their interiors on the left. The red 6-separator leaves 469 vertices on each exclusive side.
Improving the constants in our main theorem would be interesting: with bounds small enough to apply here, an iterative application might eventually look like this: the graph has a unique 6-tangle that, for every separation ( A , B ) (A, B) ( A , B ) of order at most five, selects a side containing a vertex from each of the eight coloured petal interiors outside A ∩ B A \cap B A ∩ B . We would start to decompose iteratively with the block-cutvertex theorem, Tutte’s theorem at every 2-block and then the tri-separator theorem at each 3-connected torso, etc. Eventually, we would obtain a 5-unbreakable minor carrying this 6-tangle. Then we would apply the cactus decomposition for k = 6 k = 6 k = 6 , and the cactus structure it would compute would be similar to the cyclic decomposition of Figure 1.8 . The underlying computations have been worked out in detail, see [29 ] .
Applications and perspectives
Tangles and flowers. A tangle describes a highly connected region of a graph by indicating, for each separation of small order, on which side it lies. The tangle-tree theorem provides a tree-decomposition that distinguishes these regions [115 ] . We would like to extend this viewpoint to include large flowers as highly connected regions in a broader sense. This leads us to consider bramble orientations , a relaxation of tangles, and to group related orientations describing the same flower. We seek an analogue of the tangle-tree theorem that recognises these flower regions and distinguishes them from one another and from tangles. In a follow-up paper [13 ] , we build on the structural results developed here to obtain an approximate tangle-tree theorem that also detects sufficiently large flowers.
Connectivity augmentation. The Connectivity Augmentation problem asks whether a graph can be made k k k -connected by adding at most ℓ \ell ℓ edges from a prescribed set. If the graph is already ( k − 1 ) (k-1) ( k − 1 ) -connected, this amounts to selecting edges that cross every proper ( k − 1 ) (k-1) ( k − 1 ) -separation. For edge-connectivity, the DKL theorem reduces increasing the connectivity by one to augmentation on a cactus. This reduction has led to FPT and approximation algorithms for minimum-cost augmentation [101 , 124 ] .
For vertex-connectivity, Carmesin and Ramanujan [37 ] obtained FPT algorithms parameterized by ℓ \ell ℓ for k ≤ 4 k \le4 k ≤ 4 , using exact cactus representations. Korhonen and Thorup [93 ] extended this result to arbitrary k k k with running time 2 O ( ℓ log ( ℓ + k ) ) n O ( 1 ) 2^{O(\ell\log(\ell+k))}n^{O(1)} 2 O ( ℓ l o g ( ℓ + k )) n O ( 1 ) , using approximate structural arguments. For k = 2 k=2 k = 2 , Koana and Kumabe [80 ] recently obtained an FPT algorithm with single-exponential dependence on ℓ \ell ℓ for arbitrary input graphs. A natural question is whether our approximate cactus representation can help achieve such a dependence on ℓ \ell ℓ for larger fixed k k k .
Canonicity and symmetry. The canonicity of the Tutte-decomposition underlies linear-time and logspace algorithms for planar graph isomorphism [72 , 49 ] . Retaining cyclic torsos matters for canonicity: decomposing a long cycle into bounded-size bags necessarily breaks some of its symmetries [LPPS22]. The decomposition in Theorem 1.2 approximately represents the full family of separations of small order, which is canonical. Its displayed separation system is therefore preserved by every automorphism up to bounded edit-distance, yielding a form of approximate canonicity.
Since canonical decompositions are invariant under graph automorphisms, they are particularly useful for studying highly symmetric graphs. Using their canonical tetra-decomposition [84 ] , Kurkofka and Planken classified all finite connected vertex-transitive graphs of connectivity below five [83 ] . Our decomposition may help extend their classification to higher connectivity.
Cops and robber. In the classical helicopter cops and robber game, k k k cops win precisely when the graph has treewidth less than k k k [121 ] . Consider the randomized variant with an invisible robber, no repeated visits by the cops, and mixed strategies6 for both players. For k ≤ 4 k \le4 k ≤ 4 , Planken [107 ] uses the corresponding cactus-decomposition theorems to determine the maximum winning probability of k k k cops on k k k -connected graphs. Our cactus representation may support similar analyses at higher connectivity.
Local analogues. Local separators disconnect bounded-radius balls without necessarily separating the whole graph. Local analogues of graph decomposition theorems have been developed in [19 , 18 , 46 , 76 ] . This perspective has also led to structural characterizations and canonical decompositions of locally chordal graphs [5 , 6 , 75 ] . Local separators have applications in network analysis [27 ] and distributed approximation [15 ] .
Our construction of the cyclic parts of the decomposition uses covering graphs to combine local separators into global separators. This suggests a local analogue of our theorem that, under appropriate local connectivity assumptions, approximately represents local separations together with their crossing structure.
Related work
The tangle-tree theorem of Robertson and Seymour [115 ] states that every graph admits a tree-decomposition distinguishing all its maximal tangles. Canonical constructions and alternative proofs have been developed in [25 , 23 Car16, 31 ] . Refinements address the structure of the parts containing tangles [24 , 9 ] . Erde [Erd17] showed how to further decompose the parts containing no tangle of the prescribed order, and Albrechtsen [7 ] extended this approach to abstract separation systems. Related work studies k k k -blocks and tree-decompositions displaying them [22 , 28 , 8 ] .
The theory extends to profiles and abstract separation systems [42 , 41 EKT21] , and to infinite settings [EKT22, JK24]. Its algorithmic applications include computing with tangles [GS16] and clustering [52 , 79 ] ; see [Gro16c, Die24b] for broader accounts. Our emphasis in this paper is on an approximate representation of sufficiently broad separations together with their crossing structure.
Compact descriptions of separations also support data structures for pairwise connectivity queries. For edge-connectivity, a Gomory–Hu tree [GH61] encodes all pairwise minimum-cut values in O ( n ) O(n) O ( n ) space and can be preprocessed to answer queries in constant time. For vertex-connectivity, Cohen, Di Battista, Kanevsky, and Tamassia [21 ] , building on [KTDBC91], obtained an O ( k 3 n ) O(k^{3}n) O ( k 3 n ) -space representation for pairwise ( k + 1 ) (k+1) ( k + 1 ) -connectivity queries in k k k -connected graphs. Pettie and Yin [112 ] achieved O ( k n ) O(kn) O ( k n ) words of space and constant query time. Pettie, Saranurak, and Yin [PSY26] proved an Ω ( k n ) \Omega(kn) Ω ( k n ) -bit lower bound for general graphs, which Nutov [103 ] extended to k k k -connected graphs.
Our theorem complements these results by describing how separations of small order fit together in a cactus structure, up to bounded edit-distance.
Other structural descriptions of graphs at low connectivity include decompositions of 3-connected graphs [33 EGL24, Gro16a] and the characterization by Oporowski, Oxley, and Thomas [104 ] of the typical subgraphs of 3-connected and 4-connected graphs. A complementary question asks which substructures of a k k k -connected graph can be deleted or contracted while preserving k k k -connectivity. For 3-connectivity, this includes work of Ando, Enomoto, and Saito [2 ] and of Kriesell [85 , 86 , 87 , 88 , 89 , 90 ] . The corresponding questions for 4-connectivity are studied in [1 , 77 , 98 Mar90, PS96, 113 , 117 , 119 ] .
The structure of k k k -separations in k k k -connected matroids has also been studied extensively. Results for 2-connectivity appear in [4 , 26 , 118 ] . Oxley, Semple, and Whittle Oxley, Semple, and Whittle [105 ] introduced flowers to study the structure of 3-separations in 3-connected matroids. The case of 4-connectivity was studied in [12 , 16 ] . Aikin and Oxley Aikin and Oxley [11 ] generalized flowers to arbitrary orders and classified their structure. In a framework covering both graphs and matroids, Clark and Whittle Clark and Whittle [38 ] developed flower representations using an equivalence that can identify separations differing on arbitrarily many elements. Our theorem gives a uniform bound on the edit-distance to representatives.
For directed graphs, Bowler, Gut, Hatzel, Kawarabayashi, Muzi and Reich Bowler, Gut, Hatzel, Kawarabayashi, Muzi and Reich [14 ] proved a Tutte-type decomposition theorem for k = 1 k = 1 k = 1 .
Our decomposition is closely related to unbreakable tree-decompositions [35 , 34 ] , and lean tree-decompositions [RS90, Tho90]. Roughly speaking, our cactus-decompositions are to unbreakable tree-decompositions what Tutte decompositions are to ear decompositions. Indeed, deleting one edge from each cycle and adding its adhesion set to every bag along the resulting path converts our cactus-decompositions into unbreakable tree-decompositions of adhesion at most k k k . We do not know a direct converse construction recovering the cyclic structure.
Unbreakable tree-decompositions support parameterized algorithms for Minimum Bisection [35 ] , Fair Bisection [73 ] , and model checking [LRSZ18, PSS+22, SSS+24]. Their construction has seen substantial improvements [34 , 10 ] . Korhonen Korhonen [81 ] obtained a parameterized linear-time algorithm for computing the stronger lean tree-decompositions. This naturally suggests an analogous running-time bound for our cactus-decompositions.
Conjecture 1.3. There exist computable functions f , η f, \eta f , η such that, given q , k ∈ N q, k \in\mathbb{N} q , k ∈ N and an n n n -vertex ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable graph G G G , one can compute in time f ( q , k ) ( ∣ V ( G ) ∣ + ∣ E ( G ) ∣ ) f(q,k)(|V(G)| + |E(G)|) f ( q , k ) ( ∣ V ( G ) ∣ + ∣ E ( G ) ∣ ) a cactus-decomposition displaying ( ≤ k ) (\le k) ( ≤ k ) -separations such that every ( ≤ k ) (\le k) ( ≤ k ) -separation of G G G has edit-distance at most η ( q , k ) \eta(q,k) η ( q , k ) to a displayed separation.
Figure 1.
A detailed overview of the proof of Theorem 1.1
Definitions and statements in this section are often simplified – see the later sections for accurate information.
Strategy
Our proof is inspired by the pattern of Tutte’s decomposition, which we briefly recall. To decompose a k k k -connected graph for k = 0 k = 0 k = 0 and k = 1 k = 1 k = 1 , it suffices to cut along all k k k -separators of the graph. But already for k = 2 k = 2 k = 2 , this approach fails. Indeed, different 2-separators can cross in that they cut through each other, as happens in a cycle. Then if we cut at one of the 2-separators, we lose the other. Tutte realised that the solution is to cut the graph at all 2-separators that are totally-nested , those that are not crossed by any 2-separator, and indeed no conflicts arise when cutting in this way; see Figure 2.1 . The result is a tree-decomposition whose torsos are 3-connected, cycles or K 2 K_2 K 2 ’s. It captures every proper 2-separation s s s of the graph: either s s s is induced by an edge of the decomposition tree, or s s s stems from a 2-separation of a cyclic torso, or the separator of s s s is the bag of a node t t t with a K 2 K_2 K 2 -torso and s s s distributes the branches of the decomposition tree at t t t among its two sides. Nodes with K 2 K_2 K 2 -torsos become P P P -nodes in the cactus-formulation from the introduction (see Figure 1.1 for an example). The extensions of Tutte’s result to connectivity k ∈ { 3 , 4 } k \in\{3,4\} k ∈ { 3 , 4 } [32 , 84 ] follow the same strategy: they cut the graph at all totally-nested tri-separations or tetra-separations, respectively, the modified k k k -separations from the introduction. The resulting basic building blocks are quasi-( k + 1 ) (k + 1) ( k + 1 ) -connected graphs as well as graphs exhibiting a specific cyclic or star-like structure (see Figure 1.7 for the cyclic case). The latter are precisely the graphs in which every tri- or tetra-separation, respectively, is crossed by another one. For small k k k , such graphs can be captured with k k k -flowers, a central notion in matroid theory.
Figure 2.1. Left: The yellow and purple 2-separators cross, while the red 2-separators are totally-nested. Right: The Tutte-decomposition from a tree-decomposition perspective.
We show that this correspondence extends to all k k k , up to bounded errors. For this, we let G G G be a ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable graph for some q , k ∈ N q, k \in\mathbb{N} q , k ∈ N . Our goal is to obtain a cactus-decomposition of G G G (Theorem 1.2). The proof proceeds in two phases. In the first phase, which consists of
Steps 1–4 below and spans almost the entire paper, we construct a tree-decomposition of G G G along k k k -separations that are approximately totally-nested. Its torsos are ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable for a constant q ′ q' q ′ , or approximately have the structure of a k k k -flower. These are our analogues of the 3-connected torsos on the one hand, and of the cyclic torsos and P P P -nodes on the other, in Tutte’s decomposition. This is our technical main result, Theorem 2.6 , which we state at the end of Step 4. In the second phase, which consists of Step 5, we refine this tree-decomposition into the desired cactus-decomposition: by slightly adjusting the flowers, and forging them into the tree-decomposition.
The brief overview in § 1.2 constructs the tree-decomposition first and the flowers afterwards. Here we follow the order of the paper instead. Steps 1 and 2 expand on the first two ideas listed there: Step 1 analyses how k k k -separations heavily cross with the help of three-dimensional corner diagrams, and Step 2 uses coverings to assemble heavily crossing k k k -separations into flowers. Step 3 constructs the tree-decomposition, Step 4 expands on the third idea by analysing its torsos, and Step 5 refines it into the cactus-decomposition.
Approximate terminology
The breadth of a separation ( A , B ) (A,B) ( A , B ) is the minimum of ∣ A ∖ B ∣ |A \setminus B| ∣ A ∖ B ∣ and ∣ B ∖ A ∣ |B \setminus A| ∣ B ∖ A ∣ . We say that ( A , B ) (A,B) ( A , B ) is β \beta β -broad to mean that ( A , B ) (A,B) ( A , B ) has breadth ≥ β \ge\beta ≥ β . In these terms, a graph is ( q , k ) (q,k) ( q , k ) -unbreakable if and only if every ( ≤ k ) (\le k) ( ≤ k ) -separation of it has breadth ≤ q \le q ≤ q . Two separations ( A 1 , B 1 ) (A_1,B_1) ( A 1 , B 1 ) and ( A 2 , B 2 ) (A_2,B_2) ( A 2 , B 2 ) are ε \varepsilon ε -close if all four sets A 1 ∖ A 2 A_1 \setminus A_2 A 1 ∖ A 2 , A 2 ∖ A 1 A_2 \setminus A_1 A 2 ∖ A 1 , B 1 ∖ B 2 B_1 \setminus B_2 B 1 ∖ B 2 and B 2 ∖ B 1 B_2 \setminus B_1 B 2 ∖ B 1 have size ≤ ε \le\varepsilon ≤ ε . Closeness is our working version of the edit-distance from the introduction: ε \varepsilon ε -close separations have edit-distance at most 2 ε 2\varepsilon 2 ε (see Observation 4.11 ). Closeness is quasi-transitive: if r r r and s s s are ε \varepsilon ε -close and s s s and t t t are ε ′ \varepsilon' ε ′ -close, then r r r and t t t are ( ε + ε ′ ) (\varepsilon+\varepsilon') ( ε + ε ′ ) -close. Let S , S ′ S,S' S , S ′ be sets of separations of G G G . We say that S S S ε \varepsilon ε -approximates S ′ S' S ′ if every ε \varepsilon ε -broad s ′ ∈ S ′ s' \in S' s ′ ∈ S ′ is ε \varepsilon ε -close to some separation in S S S .
We write X ⊆ ε Y X \subseteq_{\varepsilon} Y X ⊆ ε Y for X , Y ⊆ V ( G ) X,Y \subseteq V(G) X , Y ⊆ V ( G ) to mean that all but ≤ ε \le\varepsilon ≤ ε vertices in X X X are contained in Y Y Y . We define a relation ≤ ε \le_{\varepsilon} ≤ ε on the set of separations of G G G by letting
( A , B ) ≤ ε ( C , D ) : ⟺ A ⊆ ε C and B ⊇ ε D . (A,B) \le_{\varepsilon} (C,D) \ :\Longleftrightarrow\ A \subseteq_{\varepsilon} C\ \text{ and }\ B \supseteq_{\varepsilon} D. ( A , B ) ≤ ε ( C , D ) :⟺ A ⊆ ε C and B ⊇ ε D .
Two separations ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) of G G G are ε \varepsilon ε -nested if, after possibly swapping the name A A A with B B B or C C C with D D D , we have ( A , B ) ≤ ε ( C , D ) (A,B) \le_{\varepsilon} (C,D) ( A , B ) ≤ ε ( C , D ) . For ε = 0 \varepsilon=0 ε = 0 , we drop the index: ( A , B ) ≤ ( C , D ) (A,B) \le(C,D) ( A , B ) ≤ ( C , D ) means A ⊆ C A \subseteq C A ⊆ C and B ⊇ D B \supseteq D B ⊇ D , and 0-nested separations are nested; separations that are not nested cross. A set of separations is nested if its elements are pairwise nested.
For the vast majority of the paper, we assume that G G G is additionally ( β , k ) (\beta,k) ( β , k ) -narrow for some β ∈ N \beta\in\mathbb{N} β ∈ N , meaning that for every β \beta β -broad k k k -separation ( A , B ) (A,B) ( A , B ) of G G G , every component of G − ( A ∩ B ) G-(A \cap B) G − ( A ∩ B ) has neighbourhood equal to A ∩ B A \cap B A ∩ B . This technical property allows us to almost treat G G G as if it were genuinely k k k -connected. We remove this additional assumption in the finale of the paper without too much effort (§ 12 and § 13 ). In particular, the key lemmas below assume that G G G is ( β , k ) (\beta,k) ( β , k ) -narrow, while Theorem 2.6 holds for all ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graphs G G G .
Step 1: Analysing how k k k -separations heavily cross
When two k k k -separations ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) of G G G cross, we can use a standard tool known as ‘crossing diagram’ to analyse the way in which the two separations cross; see Figure 2.2 . The crossing-diagram reveals, among other things, that ( C , D ) (C,D) ( C , D ) induces a tri-partition ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) of A ∩ B A\cap B A ∩ B into two links P , Q P,Q P , Q – namely ( A ∩ B ) ∩ ( C ∖ D ) (A\cap B)\cap(C\setminus D) ( A ∩ B ) ∩ ( C ∖ D ) and ( A ∩ B ) ∩ ( D ∖ C ) (A\cap B)\cap(D\setminus C) ( A ∩ B ) ∩ ( D ∖ C ) – and the centre X : = ( A ∩ B ) ∩ ( C ∩ D ) X := (A\cap B)\cap(C\cap D) X := ( A ∩ B ) ∩ ( C ∩ D ) .
Figure 2.2. A crossing-diagram for two separations ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) of a graph G G G . The four corners are yellow, the four links are blue, and the centre is red.
In general, we have little control over what tri-partition ( C , D ) (C,D) ( C , D ) induces on A ∩ B A\cap B A ∩ B . Curiously, we show that this changes drastically if ( C , D ) (C,D) ( C , D ) γ \gamma γ -crosses ( A , B ) (A,B) ( A , B ) with γ ≈ q + k \gamma\approx q+k γ ≈ q + k , where γ \gamma γ -crossing roughly means the same as crossing plus that all four corners have size ≥ γ \ge\gamma ≥ γ . So γ \gamma γ -crossing separations are essentially what we called heavily crossing in § 1.2 . A set S S S of separations of G G G is uniformly γ \gamma γ -crossing if every s = ( A , B ) ∈ S s=(A,B)\in S s = ( A , B ) ∈ S admits a tri-partition P ( s ) \mathcal{P}(s) P ( s ) of A ∩ B A\cap B A ∩ B such that every ( C , D ) ∈ S (C,D)\in S ( C , D ) ∈ S that γ \gamma γ -crosses s s s induces the tri-partition P ( s ) \mathcal{P}(s) P ( s ) on A ∩ B A\cap B A ∩ B . Let S 0 S_0 S 0 denote the set of all q q q -broad k k k -separations of G G G . We show that heavy crossings within S 0 S_0 S 0 are consistent: for every s ∈ S 0 s\in S_0 s ∈ S 0 , all separations in S 0 S_0 S 0 that γ \gamma γ -cross s s s induce the same centre and the same two links on the separator of s s s .
Key Lemma 2.1 (Rough version of Key Lemma 5.1). The set S 0 S_0 S 0 is uniformly γ \gamma γ -crossing .
To prove Key Lemma 2.1, we analyse how a given ( A , B ) ∈ S 0 (A,B)\in S_0 ( A , B ) ∈ S 0 can be γ \gamma γ -crossed simultaneously by two other separations from S 0 S_0 S 0 . For this, we introduce a new method – a three-dimensional corner diagram – and use the submodularity property of connectivity in a three-dimensional way. See Figure 1.6 for an illustrative example of how three separations can cross three-dimensionally.
Step 2: Finding the flowers
Let S S S be a nonempty set of separations of G G G that is symmetric in that it contains for every ( A , B ) ∈ S (A,B)\in S ( A , B ) ∈ S also its inverse ( A , B ) ∗ : = ( B , A ) (A,B)^* := (B,A) ( A , B ) ∗ := ( B , A ) . We say that S S S is γ \gamma γ -crossing-connected if the reflexive and transitive closure of the union of the relations ‘γ \gamma γ -crossing’ on S S S and { ( s , s ∗ ) : s ∈ S } \{(s,s^*):s\in S\} {( s , s ∗ ) : s ∈ S } has exactly one partition class. Every symmetric set of separations admits a unique partition into classes that are maximally γ \gamma γ -crossing-connected, and these classes are again symmetric. If the set is uniformly γ \gamma γ -crossing, then so are its classes, and we call them its γ \gamma γ -asterisks . By Key Lemma 2.1, this applies to S 0 S_0 S 0 (see Figure 2.3 for one of its γ \gamma γ -asterisks). A γ \gamma γ -asterisk is trivial precisely if it consists of only one separation and its inverse.
Figure 2.3. Bottom: The purple, orange and green separators of G − X G-X G − X stem from the same asterisk S S S of G G G . The indicated cycle-decomposition equals the daisy F ( S ) F(S) F ( S ) . Top: The tree-decomposition ( D , W ) (D,\mathcal{W}) ( D , W ) of G ^ \widehat{G} G .
We capture the nontrivial γ \gamma γ -asterisks with flowers. Flowers were introduced by Oxley, Semple and Whittle in their work on decomposing 3-connected matroids [105 ] . One of the striking discoveries of that paper is that the left-over pieces of such decompositions, which they called ‘flowers’, are in fact highly structured: by their dichotomy theorem, every flower is an ‘anemone’ or a ‘daisy’. Alikin and Oxley extended flowers from 3-separations to k k k -separations for arbitrary k k k [AOO8].
In graphs, flowers can be described in terms of cactus-decompositions, which we now define formally. A cactus-decomposition of G G G is a pair C = ( C , V ) \mathcal{C}=(C,\mathcal{V}) C = ( C , V ) of a strong cactus C C C , some of whose nodes are designated as P P P -nodes, and a family V = ( V t : t ∈ V ( C ) ) \mathcal{V}=(V_t:t\in V(C)) V = ( V t : t ∈ V ( C )) of sets of vertices of G G G , its bags , such that G = ⋃ t ∈ V ( C ) G [ V t ] G=\bigcup_{t\in V(C)}G[V_t] G = ⋃ t ∈ V ( C ) G [ V t ] and, for every vertex v v v of G G G , the nodes t t t with v ∈ V t v\in V_t v ∈ V t induce a connected subgraph of C C C . Every cut ( U , W ) (U,W) ( U , W ) of C C C , a bipartition of V ( C ) V(C) V ( C ) in which either part may be empty, induces the separation ( ⋃ t ∈ U V t , ⋃ t ∈ W V t ) \left(\bigcup_{t\in U}V_t,\bigcup_{t\in W}V_t\right) ( ⋃ t ∈ U V t , ⋃ t ∈ W V t ) of G G G . A separation of G G G is displayed by C C C if it is induced by a cut ( U , W ) (U,W) ( U , W ) of C C C such that E C ( U , W ) E_C(U,W) E C ( U , W ) is empty, consists of a single edge, of two edges on a common cycle of C C C , or of edges incident with a single P P P -node of C C C . Tree-decompositions are the cactus-decompositions whose cactus is a tree without P P P -nodes, so their displayed separations are those induced by single edges. A k k k -daisy of G G G is a cactus-decomposition of G G G whose cactus is a single cycle and all of whose displayed separations have order ≤ k \le k ≤ k ; see Figure 1.7 . A k k k -anemone of G G G is a cactus-decomposition of G G G whose cactus is a star with a P P P -node as its central node, such that the bag of the central node, the central bag , has size ≤ k \le k ≤ k and is included in all other bags, the leaf-bags ; so all displayed separations have order ≤ k \le k ≤ k . We call k k k -daisies and k k k -anemones k k k -flowers . (The formal definition of k k k -daisies in § 4 adds a technical condition.) In practice, k k k -daisies capture the situation where many k k k -separators cut through each other as happens in a cycle, while k k k -anemones capture the situation where a single k k k -separator of G G G separates three or more components. The centre of a k k k -flower is the intersection of all its bags; for a k k k -anemone, this is its central bag. For a node t t t of the cactus, the bag V t V_t V t is the petal of the k k k -flower at t t t , the displayed separation ( V t , ⋃ u ≠ t V u ) \left(V_t,\bigcup_{u\ne t}V_u\right) ( V t , ⋃ u = t V u ) that cuts it off is the petal-separation at t t t , and the petal-star of the k k k -flower is the set of all its petal-separations.
Key Lemma 2.2 (Rough version of Key Lemma 6.6). For every nontrivial γ \gamma γ -asterisk S S S of S 0 S_0 S 0 there is a k k k -flower F ( S ) F(S) F ( S ) of G G G such that every separation in S S S is ε F \varepsilon_F ε F -close to some displayed separation of F ( S ) F(S) F ( S ) . Moreover, F ( S ) F(S) F ( S ) is canonical: every automorphism φ \varphi φ of G G G maps F ( S ) F(S) F ( S ) onto F ( φ ( S ) ) F(\varphi(S)) F ( φ ( S )) , bag by bag. Here, ε F \varepsilon_F ε F depends only on q q q and k k k , not on G G G .
To prove Key Lemma 2.2, let S S S be any nontrivial γ \gamma γ -asterisk of S 0 S_0 S 0 . We choose s = ( A , B ) ∈ S s=(A,B)\in S s = ( A , B ) ∈ S arbitrarily. Let ( X , { P , Q } ) : = P ( s ) (X,\{P,Q\}):=\mathcal{P}(s) ( X , { P , Q }) := P ( s ) , so every ( C , D ) ∈ S (C,D)\in S ( C , D ) ∈ S that γ \gamma γ -crosses s s s induces the tri-partition ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) on A ∩ B A\cap B A ∩ B . The centre X X X is the same for every choice of s ∈ S s\in S s ∈ S (Lemma 6.7); we call it the common centre of S S S .
If X X X has size k k k , then we are in an easy case: here we take F ( S ) F(S) F ( S ) to be the X X X -anemone , meaning that F ( S ) F(S) F ( S ) is the k k k -anemone with X X X as its central bag and a leaf-bag V ( K ) ∪ X V(K)\cup X V ( K ) ∪ X for each component K K K of G − X G-X G − X . Standard arguments quickly verify that we are done in this case.
So we may assume that ∣ X ∣ < k |X|<k ∣ X ∣ < k . Then we will find a suitable k k k -daisy to take as F ( S ) F(S) F ( S ) . To find it, we build a covering p : G ^ → G − X p:\widehat{G}\to G-X p : G → G − X (as known from general topology), which is depicted in Figure 2.3 . Here G − X G-X G − X is the union of G [ A ] − X G[A]-X G [ A ] − X and G [ B ] − X G[B]-X G [ B ] − X , which intersect precisely in the two links P P P and Q Q Q , much like a circle is the union of two arcs that intersect precisely in their endpoints. To unroll this circle, we arrange infinitely many copies of G [ A ] − X G[A]-X G [ A ] − X and G [ B ] − X G[B]-X G [ B ] − X alternately along a two-way infinite path and glue consecutive copies together along copies of P P P and Q Q Q in turn; this is G ^ \widehat{G} G , and p p p sends every vertex of a copy to the vertex it is a copy of (Construction 6.12). It is not hard to show that the covering is independent of the choice of s s s . We recall from infinite graph theory that an end is an equivalence class of one-way infinite paths, where two such paths are equivalent if they are linked by infinitely many pairwise vertex-disjoint paths. Just as R \mathbb{R} R covers S 1 \mathbb{S}^{1} S 1 and R \mathbb{R} R is two-way infinite, also G ^ \widehat{G} G covers G − X G-X G − X and G ^ \widehat{G} G is ‘two-way infinite’: G ^ \widehat{G} G has exactly two ends. So, G − X G-X G − X ‘looks like S 1 \mathbb{S}^{1} S 1 ’ from the perspective of G ^ \widehat{G} G , and we will capture this circular structure of G − X G-X G − X with the help of a k k k -daisy.
By standard arguments, P P P and Q Q Q have the same size ℓ \ell ℓ , say. Moreover, every separation of G ^ \widehat{G} G that separates the two ends of G ^ \widehat{G} G has order ≥ ℓ \ge\ell ≥ ℓ . We let R R R be the (nonempty) set of all ℓ \ell ℓ -separations of G ^ \widehat{G} G that separate the two ends of G ^ \widehat{G} G . To each separation r ∈ R r \in R r ∈ R we assign its crossing-number cr R ( r ) \operatorname{cr}_{R}(r) cr R ( r ) which is the number of separations r ′ ∈ R r' \in R r ′ ∈ R that cross r r r . We let L L L be the set of all locally R R R -optimal separations r ∈ R r \in R r ∈ R : those with cr R ( r ′ ) > cr R ( r ) \operatorname{cr}_{R}(r') > \operatorname{cr}_{R}(r) cr R ( r ′ ) > cr R ( r ) for all r ′ ∈ R r' \in R r ′ ∈ R that cross r r r . This ensures that the separations in L L L are pairwise nested. A short proof shows that L L L is nonempty. Hence the set L L L is the recipe for a tree-decomposition ( D , W ) (D,\mathcal{W}) ( D , W ) of G ^ \widehat{G} G . Notably, the decomposition tree D D D is a double-ray: a two-way infinite path. We cannot directly project the tree-decomposition ( D , W ) (D,\mathcal{W}) ( D , W ) of G ^ \widehat{G} G along the covering map p p p to obtain a 2 ℓ 2\ell 2 ℓ -daisy of G − X G-X G − X . Instead, we use a folding technique from [46 ] which exploits the high symmetry of the covering (as encoded in its group of deck transformations) to ‘fold’ ( D , W ) (D,\mathcal{W}) ( D , W ) down to G − X G-X G − X where it becomes a 2 ℓ 2\ell 2 ℓ -daisy. For this step, it is crucial that ( D , W ) (D,\mathcal{W}) ( D , W ) is canonical, as this ensures that the group of deck transformations admits a well-defined action on the tree-decomposition ( D , W ) (D,\mathcal{W}) ( D , W ) . Finally, we add X X X to all bags of the 2 ℓ 2\ell 2 ℓ -daisy to obtain the desired k k k -daisy F ( S ) F(S) F ( S ) ; recall that ∣ X ∣ + 2 ℓ = k |X|+2\ell=k ∣ X ∣ + 2 ℓ = k .
Given an arbitrary s ′ ∈ S s' \in S s ′ ∈ S , we find an ε F \varepsilon_F ε F -close displayed separation of F ( S ) F(S) F ( S ) as follows. First, we may assume without loss of generality that s ′ = s s'=s s ′ = s , where s s s was used to construct the covering. Then, we lift s ′ s' s ′ to G ^ \hat{G} G ^ , where it splits into two ℓ \ell ℓ -separations s 1 ′ , s 2 ′ s'_1,s'_2 s 1 ′ , s 2 ′ . Each s i ′ s'_i s i ′ is contained in R R R , but might not be in L L L . If some s i ′ s'_i s i ′ is not in L L L , then it is crossed by a separation in R R R whose crossing number is not larger, and we replace s i ′ s'_i s i ′ by a corner-separation of the two that lies in R R R and has a lower crossing number than s i ′ s'_i s i ′ . We repeat this until s i ′ ∈ L s'_i \in L s i ′ ∈ L . We show that all crossing numbers in R R R are bounded, so that there are only boundedly many replacements, and that any two crossing separations in R R R cross with two small opposite corners, so that every replacement changes s i ′ s'_i s i ′ by only a bounded number of vertices. Hence the final ℓ \ell ℓ -separations s i ′ ∈ L s'_i \in L s i ′ ∈ L are close to the original ones. Combining the final s 1 ′ , s 2 ′ s'_1,s'_2 s 1 ′ , s 2 ′ and adding X X X to both sides yields a displayed separation of F ( S ) F(S) F ( S ) that is ε F \varepsilon_F ε F -close to s ′ s' s ′ .
Step 3: Finding the tree-decomposition
Let S 1 S_1 S 1 be the set of
all petal-separations of the k k k -flowers found for the nontrivial γ \gamma γ -asterisks of S 0 S_0 S 0 in Step 2, as well as their inverses, and
all k k k -separations contained in trivial γ \gamma γ -asterisks of S 0 S_0 S 0 .
This follows the pattern of Tutte’s decomposition, whose totally-nested 2-separations cut off the pieces attached to each cycle and leave the cycle as a torso (Figure 2.1 ). Similarly, many displayed separations of a flower cross each other, so we cannot cut along them; but every displayed separation is induced by a set of petals and hence nested with all petal-separations of the flower. So cutting along the petal-separations separates the petals from each other without cutting through any displayed separation, and the flower then fits into the bag of the resulting node (Step 4). It would ease our lives quite a lot if the separations in S 1 S_1 S 1 were pairwise nested, but unfortunately they usually are not. They do, however, satisfy the following approximate version of Tutte’s notion of totally-nested separations. A proper ( ≤ k ) (\le k) ( ≤ k ) -separation of G G G is totally-ε \varepsilon ε -nested if it is ε \varepsilon ε -nested with every proper ( ≤ k ) (\le k) ( ≤ k ) -separation of G G G . Indeed, we can use three-dimensional corner-diagrams to show that
Key Lemma 2.3 (Rough version of Key Lemma 7.1). The separations in S 1 S_1 S 1 are totally-γ \gamma γ -nested.
(For Key Lemma 2.3 to hold, we have to slightly increase γ \gamma γ first, which is no problem.) Since the separations in S 1 S_1 S 1 need not be pairwise nested, we look for the next best thing: a nested set N N N of k k k -separations of G G G such that N N N ε N \varepsilon_N ε N -approximates S 1 S_1 S 1 , where ε N \varepsilon_N ε N is some constant independent of G G G .
It is not clear to us if N N N can be found within S 1 S_1 S 1 , but we manage to find a suitable N N N within a slightly larger set. Let M M M be the set of all totally-γ \gamma γ -nested k k k -separations of G G G (so S 1 ⊆ M S_1 \subseteq M S 1 ⊆ M ). Up to the bounded errors, these are the k k k -separations that are not heavily crossed, from which § 1.2 built the tree-decomposition. Let M ′ M' M ′ be the set of all totally-γ ′ \gamma' γ ′ -nested k k k -separations of G G G that are ε N \varepsilon_N ε N -close to a separation in M M M , where γ ′ ≈ k ( k + γ ) ≈ ε N \gamma' \approx k(k+\gamma) \approx\varepsilon_N γ ′ ≈ k ( k + γ ) ≈ ε N . Then let N N N be a maximal nested subset of M ′ M' M ′ , and let ( T , V ) (T,\mathcal{V}) ( T , V ) be the tree-decomposition that is uniquely determined by N N N . We will show more generally that
Key Lemma 2.4 (Rough version of Key Lemma 8.1). N N N ε N \varepsilon_N ε N -approximates M ⊇ S 1 M \supseteq S_1 M ⊇ S 1 . Moreover, every ε N \varepsilon_N ε N -broad k k k -separation of G G G is ε N \varepsilon_N ε N -close to some k k k -separation of G G G that is nested with N N N .
Corollary 2.5 (Rough version of Corollary 8.2). N N N is 2 ε N 2\varepsilon_N 2 ε N -canonical.
Here, a set S S S of separations of G G G is ε \varepsilon ε -canonical if for every automorphism φ \varphi φ of G G G , the sets S S S and φ ( S ) \varphi(S) φ ( S ) ε \varepsilon ε -approximate each other; for ε = 0 \varepsilon= 0 ε = 0 , we call S S S canonical. Similarly, a tree-decomposition is ( ε − ) (\varepsilon-) ( ε − ) canonical if the set of separations induced by the edges of its decomposition tree is ( ε − ) (\varepsilon-) ( ε − ) canonical.
For a node t t t of T T T , each edge e e e of T T T at t t t induces the separation ( ⋃ u ∈ V ( T 1 ) V u , ⋃ u ∈ V ( T 2 ) V u ) (\bigcup_{u\in V(T_1)}V_u,\bigcup_{u\in V(T_2)}V_u) ( ⋃ u ∈ V ( T 1 ) V u , ⋃ u ∈ V ( T 2 ) V u ) of G G G , where T 2 T_2 T 2 is the component of T − e T-e T − e that contains t t t and T 1 T_1 T 1 is the other component; these separations form the t t t -star of ( T , V ) (T,\mathcal V) ( T , V ) . To show Key Lemma 2.4, let any ( A , B ) ∈ M (A,B)\in M ( A , B ) ∈ M be given. If ( A , B ) (A,B) ( A , B ) is γ \gamma γ -close to a separation in N N N , we are done. Otherwise, using that ( A , B ) (A,B) ( A , B ) is totally-γ \gamma γ -nested, we find a unique t t t -star σ \sigma σ of ( T , V ) (T,\mathcal V) ( T , V ) such that ( A , B ) (A,B) ( A , B ) almost-γ \gamma γ -interlaces σ \sigma σ , meaning that s ≤ γ ( A , B ) s\leq_\gamma(A,B) s ≤ γ ( A , B ) or s ≤ γ ( B , A ) s\leq_\gamma(B,A) s ≤ γ ( B , A ) for all s ∈ σ s\in\sigma s ∈ σ .
If ( A , B ) (A,B) ( A , B ) crosses at least about 3 k / 2 3k/2 3 k /2 of the separations in σ \sigma σ , then we deduce that A ∩ B A\cap B A ∩ B is essentially a totally-nested separator of G G G , and show with fairly standard arguments that ( A , B ) (A,B) ( A , B ) is close to a separation in N N N .
Otherwise, ( A , B ) (A,B) ( A , B ) crosses at most about 3 k / 2 3k/2 3 k /2 of the separations in σ \sigma σ . Then we ‘uncross’ ( A , B ) (A,B) ( A , B ) with those separations, which yields a k k k -separation ( A ′ , B ′ ) (A',B') ( A ′ , B ′ ) that is ε N \varepsilon_N ε N -close to the original ( A , B ) (A,B) ( A , B ) but satisfies s ≤ ( A ′ , B ′ ) s\leq(A',B') s ≤ ( A ′ , B ′ ) or s ≤ ( B ′ , A ′ ) s\leq(B',A') s ≤ ( B ′ , A ′ ) (without the γ \gamma γ -error) for every s ∈ σ s\in\sigma s ∈ σ . Essentially by maximality of N N N , we have ( A ′ , B ′ ) ∈ N (A',B')\in N ( A ′ , B ′ ) ∈ N . Hence N N N ε N \varepsilon_N ε N -approximates M ⊇ S 1 M\supseteq S_1 M ⊇ S 1 .
The ‘moreover’ part of Key Lemma 2.4 is proved by the same case analysis, applied to an arbitrary ε N \varepsilon_N ε N -broad k k k -separation ( A , B ) (A,B) ( A , B ) of G G G instead of to a separation in M M M . Since every separation in N N N is totally-γ ′ \gamma' γ ′ -nested, ( A , B ) (A,B) ( A , B ) is γ ′ \gamma' γ ′ -nested with every separation in N N N . What we can no longer conclude is that the resulting separation ( A ′ , B ′ ) (A',B') ( A ′ , B ′ ) lies in N N N : the maximality of N N N only applies to separations in M ′ M' M ′ , and ( A ′ , B ′ ) (A',B') ( A ′ , B ′ ) need not lie in M ′ M' M ′ as ( A , B ) (A,B) ( A , B ) need not lie in M M M . However, ( A ′ , B ′ ) (A',B') ( A ′ , B ′ ) will be nested with N N N .
Step 4: The structure of the torsos
Let σ ⊆ N \sigma\subseteq N σ ⊆ N be an arbitrary t t t -star of ( T , V ) (T,\mathcal V) ( T , V ) , and let H H H denote the torso of V t V_t V t . We distinguish between two cases.
In the first case, every k k k -separation ( A , B ) (A,B) ( A , B ) of G G G that ‘cuts through’ H H H is sufficiently close to some separation in σ \sigma σ , so that either A ∩ V ( H ) A\cap V(H) A ∩ V ( H ) or B ∩ V ( H ) B\cap V(H) B ∩ V ( H ) is small. In this case, we use standard arguments to deduce that the torso H H H is ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable for some constant q ′ q' q ′ that depends only on q q q and k k k .
In the second case, there exists a k k k -separation ( A , B ) (A,B) ( A , B ) of G G G such that A ∩ V ( H ) A\cap V(H) A ∩ V ( H ) and B ∩ V ( H ) B\cap V(H) B ∩ V ( H ) are large—large enough that ( A , B ) (A,B) ( A , B ) is not ε N \varepsilon_N ε N -close to any separation in σ \sigma σ , and hence not ε N \varepsilon_N ε N -close to any separation in N N N either. Then, since N N N ε N \varepsilon_N ε N -approximates M M M by Key Lemma 2.4, we know that ( A , B ) (A,B) ( A , B ) must be γ \gamma γ -crossed by some other k k k -separation of G G G , so ( A , B ) (A,B) ( A , B ) belongs to a nontrivial γ \gamma γ -asterisk S S S of S 0 S_0 S 0 . By Key Lemma 2.2, the k k k -flower F ( S ) F(S) F ( S ) has a displayed separation d d d that is ε F \varepsilon_F ε F -close to ( A , B ) (A,B) ( A , B ) . If d d d is close to some petal-separation of F ( S ) F(S) F ( S ) (which is contained in S 1 S_1 S 1 ), then N N N approximates d d d with some parameter slightly larger than ε N \varepsilon_N ε N (by Key Lemma 2.4 and quasi-transitivity of closeness), which allows us to push this outcome to the first case by tricking with the constants. So we may assume that d d d is sufficiently far from all petal-separations of F ( S ) F(S) F ( S ) . Then we employ the flower-structure of F ( S ) F(S) F ( S ) to construct another displayed separation d ′ d' d ′ of F ( S ) F(S) F ( S ) that δ \delta δ -crosses d d d , where δ ∈ N \delta\in\mathbb{N} δ ∈ N is much larger than all previous constants, and we define F t : = F ( S ) F_t := F(S) F t := F ( S ) . Thus F t F_t F t has spread at least δ \delta δ , meaning that two of its displayed separations δ \delta δ -cross. As we may choose δ \delta δ as large as we like, Step 5 will be able to exploit this.
It is not clear a priori that F t F_t F t is well-defined, as ( A , B ) (A,B) ( A , B ) was arbitrary. However, by heavily exploiting the fact that δ \delta δ is much larger than all previous constants, we can show that there is no competitor F ( S ′ ) F(S') F ( S ′ ) at the same node t t t . Finally, we show that F t F_t F t fits into the bag V t V_t V t in the following sense. A k k k -flower F F F ε \varepsilon ε -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t if the petal-star of F F F and the t t t -star of ( T , V ) (T,\mathcal{V}) ( T , V ) ε \varepsilon ε -approximate each other; see Figure 2.4 . We show that F t F_t F t ε \varepsilon ε -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t for some ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N that depends only on q q q and k k k .
Figure 2.4. G G G consists of the black vertices and edges together with the three yellow cliques K 237 K_{237} K 237 , K 238 K_{238} K 238 and K 239 K_{239} K 239 . A tree-decomposition is indicated on the left: the decomposition tree T T T is red and the bags are circled in blue. The bag V t V_t V t of the central node t t t is refined by a daisy, whose decomposition cycle H H H is indicated on the right. The 4-separation with the purple 4-separator is a displayed separation of that daisy.
Together, Steps 1–4 yield our technical main result, an approximate Tutte-type decomposition for all k k k .
Theorem 2.6. For every q , k ∈ N q,k \in\mathbb{N} q , k ∈ N there are γ , q ′ , ε , κ ∈ N \gamma,q',\varepsilon,\kappa\in\mathbb{N} γ , q ′ , ε , κ ∈ N such that, for every ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph G G G , there exists a nested set N N N of totally-γ \gamma γ -nested ( ≤ k ) (\le k) ( ≤ k ) -separations of G G G with corresponding tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) of G G G such that:
(i) every node t t t of T T T either has a ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable torso, or comes with a k k k -flower F t F_t F t of G G G such that F t F_t F t ε \varepsilon ε -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t ;
(ii) every ε \varepsilon ε -broad k k k -separation of G G G is ε \varepsilon ε -close to either a separation in N N N or a displayed separation of some F t F_t F t ;
(iii) the family of all flowers F t F_t F t is ( q , k ) (q,k) ( q , k ) -canonical, and both the tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) and the set N N N are κ \kappa κ -canonical.
Informally, a family of flowers is ( q , k ) (q,k) ( q , k ) -canonical if it is canonical up to a bounded number of vertices in each flower, all of which lie in small components of G G G minus the centre of the flower. This weakening is the price for dropping the narrowness assumption at the end of the paper; for narrow graphs, the flowers are canonical (Theorem 10.4 ). See Definition 12.5 for the formal definition. See Theorem 12.6 for a technical strengthening.
Step 5: Refining the tree-decomposition into a cactus-decomposition
By Theorem 2.6 (ii), every sufficiently broad k k k -separation of G G G is close to a separation in N N N or to a displayed separation of one of the flowers F t F_t F t . The separations in N N N are displayed by ( T , V ) (T,\mathcal{V}) ( T , V ) , but those of the flowers need not be, and it is in order to display these as well that we pass from a tree to a cactus: we would like to substitute, for every node t t t that comes with a flower F t F_t F t , the cactus of F t F_t F t for t t t , which would turn ( T , V ) (T,\mathcal{V}) ( T , V ) into a cactus-decomposition. The obstacle is that F t F_t F t is a flower of the whole graph G G G , so a displayed separation d d d of F t F_t F t need not almost-interlace the t t t -star σ \sigma σ of ( T , V ) (T,\mathcal{V}) ( T , V ) , that is, satisfy s ≤ d s \le d s ≤ d or s ≤ d ∗ s \le d^* s ≤ d ∗ for all s ∈ σ s \in\sigma s ∈ σ , and hence need not ‘live’ inside the bag V t V_t V t .
Call a displayed separation of F t F_t F t active if it is δ \delta δ -crossed by another displayed separation of F t F_t F t , and passive otherwise; here δ \delta δ is the large constant from Step 4. The passive ones cause no trouble. Indeed, those displayed by the empty cut are already displayed by ( T , V ) (T,\mathcal{V}) ( T , V ) . Every other passive separation is close to a petal-separation of F t F_t F t or its inverse, or δ \delta δ -crossed by another displayed separation of F t F_t F t (Lemma 3.13), so every passive one is close to a petal-separation or its inverse; and these are approximated by N N N already, as F t F_t F t ε \varepsilon ε -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t by Theorem 2.6 (i).
Since F t F_t F t has spread at least δ \delta δ , it has two active separations that δ \delta δ -cross each other, and we can show that all active separations lie in a single asterisk of the set of displayed separations of F t F_t F t (Lemma 9.9). This is where the spread bound from Step 4 comes in, and it makes the active separations crossing-connected. Being δ \delta δ -crossed, every active d d d is also ε N \varepsilon_N ε N -broad, so the ‘moreover’ part of Key Lemma 2.4 applies to it: there is a k k k -separation d ~ \tilde{d} d ~ of G G G that is ε N \varepsilon_N ε N -close to d d d and nested with N N N . We fix such a d ~ \tilde{d} d ~ for every active d d d so that d ↦ d ~ d \mapsto\tilde{d} d ↦ d ~ commutes with inversion, and we show that each d ~ \tilde{d} d ~ almost-interlaces σ \sigma σ and hence lives inside the bag V t V_t V t . As δ \delta δ is much larger than ε N \varepsilon_N ε N , the separations d ~ \tilde{d} d ~ are crossing-connected as well.
To bundle the k k k -separations d ~ \tilde{d} d ~ into a flower inside V t V_t V t , we let G ~ t \widetilde{G}_t G t be the graph obtained from G G G by making G [ A ] G[A] G [ A ] complete for every ( A , B ) ∈ σ (A,B) \in\sigma ( A , B ) ∈ σ . Then G ~ t \widetilde{G}_t G t again satisfies our assumptions on G G G , each d ~ \tilde{d} d ~ is still a separation of G ~ t \widetilde{G}_t G t (this is what we needed the almost-interlacing for), and now every separation of G ~ t \widetilde{G}_t G t almost-interlaces σ \sigma σ . So we may apply Key Lemma 2.2 a second time, to G ~ t \widetilde{G}_t G t in place of G G G and to the asterisk of G ~ t \widetilde{G}_t G t that contains the d ~ \tilde{d} d ~ , and obtain a k k k -flower F ~ t \widetilde{F}_t F t that approximates all the d ~ \tilde{d} d ~ and all of whose displayed separations almost-interlace σ \sigma σ . Every A A A with ( A , B ) ∈ σ (A,B) \in\sigma ( A , B ) ∈ σ is complete in G ~ t \widetilde{G}_t G t , so it lies in a bag of F ~ t \widetilde{F}_t F t , and we attach the branch of T T T that ( A , B ) (A,B) ( A , B ) cuts off at the node of that bag. Thus substituting F ~ t \widetilde{F}_t F t for t t t at every node t t t that comes with a flower, and intersecting the bags of F ~ t \widetilde{F}_t F t with V t V_t V t , now turns ( T , V ) (T,\mathcal{V}) ( T , V ) into the desired cactus-decomposition. Separations with a small side are close to a separation displayed by the empty cut. Chasing an arbitrary sufficiently broad k k k -separation of G G G through the dichotomy above – active separations via the corresponding d ~ \tilde{d} d ~ to a displayed separation of F ~ t \widetilde{F}_t F t , passive ones back into N N N – shows that the separations displayed by this cactus-decomposition approximate the set of all k k k -separations of G G G . This proves Theorem 1.2.
Organisation of the paper
In § 4 , we introduce flowers for graphs and prove a dichotomy theorem for them. In § 5 , we introduce three-dimensional corner-diagrams, which we then use in § 6 to show that broad k k k -separations are uniformly γ \gamma γ -crossing (Key Lemma 2.1). In § 7 , we employ coverings from general topology to construct canonical flowers for asteriskal sets of separations, and thereby prove Key Lemma 2.2. In § 8 , we show that every k k k -separation is γ \gamma γ -nested with the petal-separations of these flowers (Key Lemma 2.3). In § 9 , we construct the tree-decomposition T \mathcal{T} T for Theorem 2.6 and prove Key Lemma 2.4 and Corollary 2.5 . In § 10 , we analyse the structure of the torsos of T \mathcal{T} T , and show that each splitting star either has a torso of higher connectivity, in the sense of unbreakability, or is approximated by the petal-star of a flower. In § 11 , we prove Theorem 2.6 under the additional assumption that the graph is ( β , k ) (\beta,k) ( β , k ) -narrow and establish Theorem 1.2 in the narrow case. In § 12 , we remove this assumption and prove Theorem 1.2 in full generality. Finally, in § 13 , we remove this assumption to prove Theorem 2.6 in the general case. The appendix proves the unbreakable-minor lemma used to iterate our decomposition within minors of the original graph.
Separations and Flowers
In this section we shall define flowers and prove some basic properties thereof. Flowers were introduced by Oxley, Semple and Whittle when studying the structure of 3-separations in 3-connected matroids [105 ] . The case of 4-connectivity was studied in [12 , 16 ] . Aikin and Oxley [11 ] generalized flowers to arbitrary orders and classified their structure. Clark and Whittle extended this concept to k k k -separations in vertically k k k -connected matroids [38 ] . When specialised to graphs this definition reads as follows (modulo some tiny details which are not relevant here). Given k ∈ N k \in\mathbb{N} k ∈ N and a graph G G G , a k k k -flower (which we call a pseudo k k k -flower in this paper) is a cyclic family F = ( P i : i ∈ Z ℓ ) \mathcal{F} = (P_i : i \in\mathbb{Z}_\ell) F = ( P i : i ∈ Z ℓ ) such that for every interval7 I ⊆ Z ℓ I \subseteq\mathbb{Z}_\ell I ⊆ Z ℓ , we have that ( ⋃ i ∈ I P i , ⋃ j ∉ I P j ) (\bigcup_{i\in I} P_i,\bigcup_{j\notin I} P_j) ( ⋃ i ∈ I P i , ⋃ j ∈ / I P j ) is a ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G . Pseudo k k k -flowers were previously studied in the setting of k k k -connected graphs but in this paper we study them in the larger class of ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graphs (for some q ∈ N q \in\mathbb{N} q ∈ N ). Here we shall define flowers slightly differently than Whittle and Clark, and we shall show that these two definitions are equivalent for k k k -connected graphs in Observation 3.2 below, and in Example 3.1 we shall explain a technical difference between them.
In order to give our definition of flowers, we rely on the following standard notation.
Separations
Let G G G be a graph. A separation of G G G is a pair ( A , B ) (A,B) ( A , B ) with G [ A ] ∪ G [ B ] = G G[A]\cup G[B]=G G [ A ] ∪ G [ B ] = G (so there is no edge in G G G between A ∖ B A\setminus B A ∖ B and B ∖ A B\setminus A B ∖ A ). The separator of a separation s = ( A , B ) s=(A,B) s = ( A , B ) is the intersection sep ( s ) : = A ∩ B \operatorname{sep}(s):=A\cap B sep ( s ) := A ∩ B . We write s ∗ : = ( B , A ) s^*:=(B,A) s ∗ := ( B , A ) for the inverse of s s s . A set S S S of separations is symmetric if s ∗ ∈ S s^*\in S s ∗ ∈ S whenever s ∈ S s\in S s ∈ S . The breadth of a separation ( A , B ) (A,B) ( A , B ) is the minimum of ∣ A ∖ B ∣ |A\setminus B| ∣ A ∖ B ∣ and ∣ B ∖ A ∣ |B\setminus A| ∣ B ∖ A ∣ . We say that ( A , B ) (A,B) ( A , B ) is b b b -broad to mean that ( A , B ) (A,B) ( A , B ) has breadth ≥ b \geq b ≥ b . If S S S is a set of separations, we denote by S ∗ S^* S ∗ the set { s ∗ : s ∈ S } \{s^*:s\in S\} { s ∗ : s ∈ S } . A separation ( A 1 , A 2 ) (A_1,A_2) ( A 1 , A 2 ) of a graph G G G is proper if A i ∖ A 3 − i ≠ ∅ A_i \setminus A_{3-i} \ne\varnothing A i ∖ A 3 − i = ∅ for i ∈ { 1 , 2 } i \in\{1,2\} i ∈ { 1 , 2 } . Note that s s s is proper if and only if s s s is 1-broad. A proper separation ( A , B ) (A,B) ( A , B ) of G G G is half-connected if G [ A ∖ B ] G[A \setminus B] G [ A ∖ B ] or G [ B ∖ A ] G[B \setminus A] G [ B ∖ A ] is connected. A separation ( A 1 , A 2 ) (A_1,A_2) ( A 1 , A 2 ) of G G G is tight if for both i ∈ { 1 , 2 } i \in\{1,2\} i ∈ { 1 , 2 } there is a component K i K_i K i of G [ A i ∖ A 3 − i ] G[A_i \setminus A_{3-i}] G [ A i ∖ A 3 − i ] with N G ( K i ) = A 1 ∩ A 2 N_G(K_i)=A_1 \cap A_2 N G ( K i ) = A 1 ∩ A 2 . Note that tight separations are proper. A set S S S of separations of G G G is tight if every separation in S S S is tight. A separation ( A , B ) (A,B) ( A , B ) of G G G is narrow if every component of G − ( A ∩ B ) G-(A \cap B) G − ( A ∩ B ) has neighbourhood equal to A ∩ B A \cap B A ∩ B . Note that if ( A , B ) (A,B) ( A , B ) is narrow and proper, then it is tight. Finally, for b ∈ N b \in\mathbb{N} b ∈ N , G G G is ( b , k ) (b,k) ( b , k ) -narrow if every b b b -broad k k k -separation is narrow.
Given two separations s = ( A , B ) s=(A,B) s = ( A , B ) and t = ( C , D ) t=(C,D) t = ( C , D ) of a graph G G G , we define
s ∧ t : = ( A ∩ C , B ∪ D ) and s ∨ t : = ( A ∪ C , B ∩ D ) . s \wedge t := (A \cap C, B \cup D) \quad\text{and} \quad s \vee t := (A \cup C, B \cap D). s ∧ t := ( A ∩ C , B ∪ D ) and s ∨ t := ( A ∪ C , B ∩ D ) .
Note that s ∨ t = ( s ∗ ∧ t ∗ ) ∗ s \vee t=(s^* \wedge t^*)^* s ∨ t = ( s ∗ ∧ t ∗ ) ∗ . Note that both s ∧ t s \wedge t s ∧ t and s ∨ t s \vee t s ∨ t again are separations of G G G . Moreover, the set of all separations of G G G together with ∧ \wedge ∧ and ∨ \vee ∨ forms a lattice. Additionally, we define ⋁ s ∈ ∅ s = ( ∅ , V ( G ) ) \bigvee_{s \in\varnothing}s=(\varnothing,V(G)) ⋁ s ∈ ∅ s = ( ∅ , V ( G )) and ⋀ s ∈ ∅ s = ( V ( G ) , ∅ ) \bigwedge_{s \in\varnothing}s=(V(G),\varnothing) ⋀ s ∈ ∅ s = ( V ( G ) , ∅ ) .
Stars and torsos
Let G G G be a graph and S , σ S,\sigma S , σ be sets of separations of G G G . We say that σ \sigma σ is a star if s ≤ t ∗ s \leq t^* s ≤ t ∗ for all distinct s , t ∈ σ s,t \in\sigma s , t ∈ σ . Let s s s be a separation of G G G . We say that s s s interlaces σ \sigma σ if for every t ∈ σ t \in\sigma t ∈ σ , we have t < s t < s t < s or t < s ∗ t < s^* t < s ∗ . We say that s s s almost-interlaces σ \sigma σ if for every t ∈ σ t \in\sigma t ∈ σ , we have t ≤ s t \leq s t ≤ s or t ≤ s ∗ t \leq s^* t ≤ s ∗ . We say that S S S (almost-)interlaces σ \sigma σ if every separation s ∈ S s \in S s ∈ S (almost-)interlaces σ \sigma σ . We say that S S S is consistent if there do not exist s , t ∈ S s,t \in S s , t ∈ S with s ∗ ≤ t s^* \leq t s ∗ ≤ t . A star σ ⊆ S \sigma\subseteq S σ ⊆ S is splitting in S S S if it is consistent and no separation in S S S interlaces σ \sigma σ . The torso of a star σ = { ( A i , B i ) } i ∈ I \sigma=\{(A_i,B_i)\}_{i \in I} σ = {( A i , B i ) } i ∈ I is the graph obtained from G [ ⋂ i ∈ I B i ] G[\bigcap_{i \in I}B_i] G [ ⋂ i ∈ I B i ] by making every separator A i ∩ B i A_i \cap B_i A i ∩ B i into a clique.
Graph-decompositions
Let G G G and H H H be graphs, and let V = ( V h : h ∈ V ( H ) ) \mathcal{V}=(V_h:h \in V(H)) V = ( V h : h ∈ V ( H )) be a family of sets V h ⊆ V ( G ) V_h \subseteq V(G) V h ⊆ V ( G ) . Following [46 ] , the pair ( H , V ) (H,\mathcal{V}) ( H , V ) is called a graph-decomposition , or H H H -decomposition, of G G G if it satisfies the following two conditions:
(H1) G = ⋃ h ∈ H G [ V h ] G=\bigcup_{h \in H}G[V_h] G = ⋃ h ∈ H G [ V h ] ;
(H2) for every v ∈ V ( G ) v \in V(G) v ∈ V ( G ) the subgraph H [ { h ∈ V ( H ) : v ∈ V h } ] H[\{h \in V(H):v \in V_h\}] H [{ h ∈ V ( H ) : v ∈ V h }] is connected.
The sets V h V_h V h are the bags of the H H H -decomposition. For every edge e = h h ′ e=hh' e = h h ′ of H H H , the set V h ∩ V h ′ V_h \cap V_{h'} V h ∩ V h ′ is the adhesion-set at e e e . Whenever an H H H -decomposition is introduced as ( H , V ) (H,\mathcal{V}) ( H , V ) , we tacitly assume that V = ( V h : h ∈ V ( H ) ) \mathcal{V}=(V_h:h \in V(H)) V = ( V h : h ∈ V ( H )) . A cut of H H H is a bipartition ( U , W ) (U,W) ( U , W ) of V ( H ) V(H) V ( H ) , where either part is allowed to be empty. Every cut ( U , W ) (U,W) ( U , W ) of $H induces a separation ( A , B ) (A,B) ( A , B ) of G G G , where A = ⋃ h ∈ U V h A=\bigcup_{h \in U}V_h A = ⋃ h ∈ U V h and B = ⋃ h ∈ W V h B=\bigcup_{h \in W}V_h B = ⋃ h ∈ W V h , see [46 Lemma 3.4] . We also call ( A , B ) (A,B) ( A , B ) the separation induced by ( U , W ) (U,W) ( U , W ) or induced by U U U . Moreover, the separator of ( A , B ) (A,B) ( A , B ) equals ⋃ u w ∈ F ( V u ∩ V w ) \bigcup_{uw \in F}(V_u \cap V_w) ⋃ u w ∈ F ( V u ∩ V w ) where F = E H ( U , W ) F=E_H(U,W) F = E H ( U , W ) . We write V ∖ X : = ⋃ h ∈ V ( X ) V h V \setminus X:=\bigcup_{h \in V(X)}V_h V ∖ X := ⋃ h ∈ V ( X ) V h for all subgraphs X ⊆ H X \subseteq H X ⊆ H .
By abuse of notation, for every ℓ ≥ 2 \ell\geq2 ℓ ≥ 2 , we denote by Z ℓ Z_\ell Z ℓ the graph that is the ℓ \ell ℓ -cycle with vertices in Z ℓ Z_\ell Z ℓ arranged in the corresponding cyclic order. For ℓ = 2 \ell=2 ℓ = 2 , this consists of two vertices joined by two distinct parallel edges.
A cactus-decomposition of a graph G G G is a graph-decomposition ( H , V ) (H,\mathcal{V}) ( H , V ) of G G G where H H H is a strong cactus some of whose nodes are designated as P P P -nodes. Let G = ( H , V ) \mathcal{G}=(H,\mathcal{V}) G = ( H , V ) be a cactus-decomposition of a graph G G G . A separation of G G G is displayed by G \mathcal{G} G if it is induced by a cut ( U , W ) (U,W) ( U , W ) of H H H such that the set F : = E H ( U , W ) F:=E_H(U,W) F := E H ( U , W ) satisfies one of the following conditions:
∣ F ∣ ≤ 1 |F| \le1 ∣ F ∣ ≤ 1 ;
∣ F ∣ = 2 |F| = 2 ∣ F ∣ = 2 and both edges of F F F lie on a common cycle of H H H ;
∣ F ∣ ≥ 2 |F| \ge2 ∣ F ∣ ≥ 2 and all edges of F F F are incident with a common P P P -node of H H H .
We denote the set of all displayed separations of G G G by D ( G ) D(G) D ( G ) . Unless specified otherwise, we view tree-decompositions and cycle-decompositions as cactus-decompositions with no P P P -nodes. Note that, if H = Z ℓ H = Z_{\ell} H = Z ℓ then every displayed separation of G G G with both sides nonempty is induced by an interval of Z ℓ Z_{\ell} Z ℓ . Moreover, if H H H is a star whose centre is a P P P -node, every cut ( U , W ) (U, W) ( U , W ) of H H H induces a displayed separation of G G G . We make the following definition for all decomposition graphs H H H (but will only use it when H H H is a cycle or a star). For each node t ∈ V ( H ) t \in V(H) t ∈ V ( H ) , we call the separation of G G G induced by ( { t } , V ( H ) − { t } ) (\{t\}, V(H) - \{t\}) ({ t } , V ( H ) − { t }) the petal-separation of G G G at t t t . The set π ( G ) \pi(G) π ( G ) of all petal-separations of G G G forms a star of separations, the petal-star of G G G ; in particular, π ( G ) \pi(G) π ( G ) is a nested set. We refer to bag V t V_t V t as the petal at t t t .
Flowers
A k k k -flower in G G G is a cycle-decomposition O \mathcal{O} O of G G G such that every displayed separation has order ≤ k \le k ≤ k . A k k k -daisy in G G G is a k k k -flower D = ( O , P ) \mathcal{D} = (\mathcal{O}, \mathcal{P}) D = ( O , P ) such that every narrow proper k k k -separation induced by some subset of V ( O ) V(\mathcal{O}) V ( O ) is induced by the vertices of a path contained in O \mathcal{O} O . A weak k k k -anemone in G G G is a k k k -flower such that every induced separation has order ≤ k \le k ≤ k . A (strong) k k k -anemone in G G G is a k k k -flower A = ( O , P ) \mathcal{A} = (\mathcal{O}, \mathcal{P}) A = ( O , P ) such that the union of the adhesion-sets of A \mathcal{A} A has size ≤ k \le k ≤ k .
Example 3.1. Let G G G be the 4-wheel. We obtain a 3-flower structure on G G G by taking a cycle-decomposition with four bags that are the four triangles of this graph, with the appropriate cyclic ordering, see Figure 3.1 ; denote it by ( Z 4 , ( Δ i : i ∈ Z 4 ) ) (Z_4,(\Delta_i : i \in Z_4)) ( Z 4 , ( Δ i : i ∈ Z 4 )) . Now ( Δ i ∣ i ∈ Z 4 ) (\Delta_i \mid i \in Z_4) ( Δ i ∣ i ∈ Z 4 ) is also a pseudo 3-flower. If we delete the degree-four vertex from Δ 2 \Delta_2 Δ 2 and Δ 4 \Delta_4 Δ 4 , we still have a pseudo 3-flower, which is not a 3-flower.
Figure 3.1. A flower on the 4-wheel. Its bags Δ 1 \Delta_1 Δ 1 to Δ 4 \Delta_4 Δ 4 are highlighted in red.
The following says that the definition of pseudo-flowers, which is the original definition of flower by Clark and Whittle for matroids specialised to graphs (modulo minor details), is basically the same as the definition of flowers we use in this paper, modulo technicalities arising from examples like the one of Example 3.1 .
Observation 3.2. Given k ∈ N k \in\mathbb{N} k ∈ N , for every graph G G G the following holds:
If ( Z ℓ , ( P o : o ∈ Z ℓ ) ) (Z_\ell,(P_o : o \in Z_\ell)) ( Z ℓ , ( P o : o ∈ Z ℓ )) is a k k k -flower, then ( P o : o ∈ Z ℓ ) (P_o : o \in Z_\ell) ( P o : o ∈ Z ℓ ) is a pseudo k k k -flower.
If G G G is k k k -connected and ( P o : o ∈ Z ℓ ) (P_o : o \in Z_\ell) ( P o : o ∈ Z ℓ ) is a pseudo k k k -flower, then there is a k k k -flower ( Z ℓ , ( Q o : o ∈ Z ℓ ) ) (Z_\ell,(Q_o : o \in Z_\ell)) ( Z ℓ , ( Q o : o ∈ Z ℓ )) with P o ⊆ Q o P_o \subseteq Q_o P o ⊆ Q o for all o ∈ Z ℓ o \in Z_\ell o ∈ Z ℓ .
Proof. (i) holds by definition of k k k -flower.
(ii). Assume that G G G is k k k -connected and ( P o : o ∈ Z ℓ ) (P_o : o \in Z_\ell) ( P o : o ∈ Z ℓ ) is a pseudo k k k -flower. We obtain Q = ( Q o : o ∈ Z ℓ ) \mathcal{Q}=(Q_o : o \in Z_\ell) Q = ( Q o : o ∈ Z ℓ ) from ( P o : o ∈ Z ℓ ) (P_o : o \in Z_\ell) ( P o : o ∈ Z ℓ ) by greedily adding vertices to the bags P o P_o P o as long as the property of being a pseudo k k k -flower is preserved. It suffices to show that O = ( Z ℓ , Q ) \mathcal{O}=(Z_\ell,\mathcal{Q}) O = ( Z ℓ , Q ) is a graph-decomposition. First we show (H2). Suppose for a contradiction that there is a vertex v ∈ V ( G ) v \in V(G) v ∈ V ( G ) such that the subgraph H H H induced by { i ∈ Z ℓ : v ∈ Q i } \{i \in Z_\ell: v \in Q_i\} { i ∈ Z ℓ : v ∈ Q i } is not connected. Since H H H is not connected, the complement Z ℓ − H Z_\ell-H Z ℓ − H of the subgraph H H H in the cycle Z ℓ Z_\ell Z ℓ has at least two nonempty components; pick two such components and denote their vertex-sets by I I I and J J J .
Claim 3.2.1. One of ⋃ i ∈ I Q i \bigcup_{i\in I}Q_i ⋃ i ∈ I Q i and ⋃ j ∈ J Q j \bigcup_{j\in J}Q_j ⋃ j ∈ J Q j contains at most k − 1 k-1 k − 1 vertices.
Proof of the claim. Denote the two intervals that make up the complement of I ∪ J I \cup J I ∪ J in Z ℓ Z_\ell Z ℓ by K 1 K_1 K 1 and K 2 K_2 K 2 . By construction, for each of the four indices x x x in the boundary of K 1 K_1 K 1 and K 2 K_2 K 2 , the bag Q x Q_x Q x contains the vertex v v v . Consider the separations s = ( A , B ) s=(A,B) s = ( A , B ) and t = ( C , D ) t=(C,D) t = ( C , D ) induced by I ∪ K 1 I \cup K_1 I ∪ K 1 and I ∪ K 2 I \cup K_2 I ∪ K 2 , respectively. Since Q \mathcal{Q} Q is a pseudo k k k -flower, both these separations have order ≤ k \leq k ≤ k . By submodularity one of the separations s ∧ t s\wedge t s ∧ t and s ∨ t s\vee t s ∨ t has order ≤ k \leq k ≤ k , say it is s ∧ t = ( A ∩ C , B ∪ D ) s\wedge t=(A\cap C,B\cup D) s ∧ t = ( A ∩ C , B ∪ D ) (the other case is the same with the roles of ‘I I I ’ and ‘J J J ’ interchanged). The separation ( ⋃ i ∈ I Q i , ⋃ m ∉ I Q m ) \left(\bigcup_{i\in I}Q_i,\bigcup_{m\notin I}Q_m\right) ( ⋃ i ∈ I Q i , ⋃ m ∈ / I Q m ) induced by I I I satisfies that its side ⋃ i ∈ I Q i \bigcup_{i\in I}Q_i ⋃ i ∈ I Q i is contained in A ∩ C A\cap C A ∩ C , and its side ⋃ m ∉ I Q m \bigcup_{m\notin I}Q_m ⋃ m ∈ / I Q m is contained in B ∪ D B\cup D B ∪ D . However, the containment ⋃ i ∈ I Q i ⊆ A ∩ C \bigcup_{i\in I}Q_i\subseteq A\cap C ⋃ i ∈ I Q i ⊆ A ∩ C is proper since the vertex v v v is only in the latter set. Hence the separation induced by I I I has order ≤ k − 1 \leq k-1 ≤ k − 1 . Since G G G is k k k -connected, one of its sides contains at most k − 1 k-1 k − 1 vertices. Hence one of ⋃ i ∈ I Q i \bigcup_{i\in I}Q_i ⋃ i ∈ I Q i or ⋃ j ∈ J Q j \bigcup_{j\in J}Q_j ⋃ j ∈ J Q j contains at most k − 1 k-1 k − 1 vertices. ◊
By Claim 3.2.1 and by exchanging the roles of ‘I I I ’ and ‘J J J ’ if necessary, assume that ⋃ i ∈ I Q i \bigcup_{i\in I}Q_i ⋃ i ∈ I Q i contains at most k − 1 k-1 k − 1 vertices. We obtain R \mathcal{R} R from Q \mathcal{Q} Q by adding the vertex v v v to all Q i Q_i Q i with i ∈ I i \in I i ∈ I . To see that R \mathcal{R} R is a pseudo k k k -flower, let M M M be an interval of Z ℓ Z_\ell Z ℓ . If M M M and its complement are both not contained in I I I , then the separation induced by M M M is the same for R \mathcal{R} R and Q \mathcal{Q} Q . So assume that one of M M M or its complement, say M M M , is a subset of I I I . Then ⋃ m ∈ M R m \bigcup_{m\in M}R_m ⋃ m ∈ M R m is contained in { v } ∪ ⋃ i ∈ I Q i \{v\}\cup\bigcup_{i\in I}Q_i { v } ∪ ⋃ i ∈ I Q i and thus has size ≤ k \leq k ≤ k . Thus the separation induced by M M M for R \mathcal{R} R has order ≤ k \leq k ≤ k . Hence R \mathcal{R} R is a pseudo k k k -flower. This violates the maximality of Q \mathcal{Q} Q , as desired.
We now show (H1). Pick any i ∈ Z ℓ i \in Z_\ell i ∈ Z ℓ . Since ( P i , ⋃ j ≠ i P j ) (P_i,\bigcup_{j\ne i}P_j) ( P i , ⋃ j = i P j ) is a separation of G G G , it follows that every vertex is contained in some bag P o ⊆ Q o P_o \subseteq Q_o P o ⊆ Q o . Suppose that some edge u v uv uv lies in no bag Q o Q_o Q o . Let I u , I v I_u,I_v I u , I v be the set of indices of the bags Q o Q_o Q o containing u , v u,v u , v , respectively. By (H2), I u I_u I u and I v I_v I v are disjoint nonempty connected subgraphs of the cycle Z ℓ Z_\ell Z ℓ . The separation induced by I u I_u I u separates u u u from v v v , contradicting that u v uv uv is an edge.
Hence O = ( Z ℓ , Q ) \mathcal{O}=(Z_\ell,\mathcal{Q}) O = ( Z ℓ , Q ) is a graph-decomposition, and thus a k k k -flower.
Lemma 3.3. Let O = ( Z ℓ , P ) \mathcal{O} = (\mathbb{Z}_{\ell},\mathcal{P}) O = ( Z ℓ , P ) be a k k k -flower such that there is a partition of Z ℓ \mathbb{Z}_{\ell} Z ℓ into three nonempty intervals I 1 I_1 I 1 , I 2 I_2 I 2 and I 3 I_3 I 3 such that all their corresponding induced separations have the same separator X X X , and this separator has size exactly k k k . Then O \mathcal{O} O is a strong k k k -anemone.
Proof. It suffices to show that for every i ∈ Z ℓ i \in\mathbb{Z}_{\ell} i ∈ Z ℓ , the separator of the separation ( P i , ⋃ j ≠ i P j ) (P_i,\bigcup_{j\ne i} P_j) ( P i , ⋃ j = i P j ) induced by { i } \{i\} { i } is a subset of X X X . Suppose not for a contradiction. Then there is i ∈ Z ℓ i \in\mathbb{Z}_{\ell} i ∈ Z ℓ such that this separator contains a vertex v ∉ X v \notin X v ∈ / X . Since O \mathcal{O} O is a graph-decomposition, the vertex v v v is contained in one of the two bags adjacent to P i P_i P i ; these are the bags P i − 1 P_{i-1} P i − 1 and P i + 1 P_{i+1} P i + 1 . By reversing the cyclic orientation of Z ℓ \mathbb{Z}_{\ell} Z ℓ if necessary, we assume that v ∈ P i − 1 v \in P_{i-1} v ∈ P i − 1 . By relabelling if necessary, assume that i ∈ I 1 i \in I_1 i ∈ I 1 , and that I 2 I_2 I 2 is the interval just before I 1 I_1 I 1 , and I 3 I_3 I 3 is the one just after. Let M M M be the interval of Z ℓ \mathbb{Z}_{\ell} Z ℓ obtained from I 2 I_2 I 2 by adding the subinterval of I 1 I_1 I 1 from the initial element of I 1 I_1 I 1 all the way to i − 1 i-1 i − 1 (if i i i is the initial element of I 1 I_1 I 1 , this subinterval is empty). By construction, the vertex v v v is in the separator of the separation induced by M M M . Every vertex of X X X is in the separators of the separations induced by I 2 I_2 I 2 and I 3 I_3 I 3 by assumption. Hence X X X is contained in ⋃ j ∈ I 2 P j \bigcup_{j\in I_2} P_j ⋃ j ∈ I 2 P j and contained in ⋃ j ∈ I 3 P j \bigcup_{j\in I_3} P_j ⋃ j ∈ I 3 P j . Thus X X X is contained in the separator of the separation induced by M M M . Thus the separator of the separation induced by M M M contains X ∪ { v } X \cup\{v\} X ∪ { v } , and thus has size ≥ k + 1 \ge k+1 ≥ k + 1 . Since M M M is an interval, this contradicts the assumption that O \mathcal{O} O is a k k k -flower. Hence for every i ∈ Z ℓ i \in\mathbb{Z}_{\ell} i ∈ Z ℓ the separator of the separation ( P i , ⋃ j ≠ i P j ) (P_i,\bigcup_{j\ne i} P_j) ( P i , ⋃ j = i P j ) induced by { i } \{i\} { i } is a subset of X X X , and so O \mathcal{O} O is a strong anemone, as desired. □
Remark 3.4. Since in k k k -connected graphs, all k k k -separations are narrow and there are no proper separations of order less than k k k , the notion of daisy we give here specialises to the well-known notion of daisies in k k k -connected graphs in the sense of Clark and Whittle [38 ] ; below in Example 3.6 we explain in detail the motivation behind our definition of ‘daisy’. Thus the following theorem implies the well-known Flower-Dichotomy theorem [38 ] for graphs, showing that this central theorem lifts to our slightly more general setting.
Theorem 3.5 (Flower-Dichotomy Theorem). For every k k k , every k k k -flower is a strong k k k -anemone or a k k k -daisy.
Proof. Let a k k k -flower O = ( Z ℓ , ( P i ∣ i ∈ Z ℓ ) ) \mathcal{O} = (\mathbb{Z}_{\ell},(P_i \mid i \in\mathbb{Z}_{\ell})) O = ( Z ℓ , ( P i ∣ i ∈ Z ℓ )) be given, and assume that it is not a k k k -daisy. For every I ⊆ Z ℓ I \subseteq\mathbb{Z}_{\ell} I ⊆ Z ℓ , we define the interval catalogue M \mathcal{M} M of I I I to be the set of maximal intervals that are contained in one of I I I or its complement Z ℓ ∖ I \mathbb{Z}_{\ell} \setminus I Z ℓ ∖ I . Observe that the size of M \mathcal{M} M is twice the number of maximal intervals contained in I I I . Since O \mathcal{O} O is not a k k k -daisy, there is a narrow proper k k k -separation s s s induced by some I ⊆ Z ℓ I \subseteq\mathbb{Z}_{\ell} I ⊆ Z ℓ that is not induced by an interval. We let I ⊆ Z ℓ I \subseteq\mathbb{Z}_{\ell} I ⊆ Z ℓ be such that it induces s s s and its interval catalogue M \mathcal{M} M is of minimal size. Denote the separator of s s s by X X X . Since I I I is a non-interval, the family M \mathcal{M} M has size at least three (in fact at least four but we will not use that) and partitions Z ℓ \mathbb{Z}_{\ell} Z ℓ . We say that an interval M ∈ M M \in\mathcal{M} M ∈ M is bad if for all m ∈ M m \in M m ∈ M we have that the petal P m P_m P m is contained in X X X .
Claim 3.5.1. If M ∈ M M \in\mathcal{M} M ∈ M is not bad , then the separation it induces has separator equal to X X X .
Proof of the claim. Let M ∈ M M \in\mathcal{M} M ∈ M be given and let t M t_M t M be the separation induced by M M M . Without loss of generality, assume that M M M is contained in I I I . Our first aim is to show sep ( t M ) ⊆ X \operatorname{sep}(t_M) \subseteq X sep ( t M ) ⊆ X . Let v ∈ sep ( t M ) v \in\operatorname{sep}(t_M) v ∈ sep ( t M ) be arbitrary. Then v v v is in ⋃ i ∈ I P i \bigcup_{i\in I} P_i ⋃ i ∈ I P i . Moreover, there is some petal P j P_j P j with j ∉ M j\notin M j ∈ / M that contains v v v . If j ∉ I j\notin I j ∈ / I , then v v v is in X X X and we are done. So assume j ∈ I j\in I j ∈ I . Since O \mathcal{O} O is a graph-decomposition, by the maximality of the interval M M M , there is some j ′ ∈ Z ℓ ∖ I j'\in\mathbb{Z}_{\ell}\setminus I j ′ ∈ Z ℓ ∖ I with v ∈ P j ′ v\in P_{j'} v ∈ P j ′ . Hence v v v is in X X X . It follows that sep ( t M ) ⊆ X \operatorname{sep}(t_M)\subseteq X sep ( t M ) ⊆ X .
Since M M M is not bad, the set ⋃ m ∈ M P m \bigcup_{m\in M}P_m ⋃ m ∈ M P m is not a subset of X X X . Since s s s is proper, its side ⋃ j ∉ I P j \bigcup_{j\notin I}P_j ⋃ j ∈ / I P j is not a subset of X X X . Since M ⊂ I M\subset I M ⊂ I , the set ⋃ j ∉ M P j \bigcup_{j\notin M}P_j ⋃ j ∈ / M P j is not a subset of X X X . Together with sep ( t M ) ⊆ X \operatorname{sep}(t_M)\subseteq X sep ( t M ) ⊆ X it follows that t M t_M t M is proper. Since s s s is narrow, the separation t M t_M t M has the whole of X X X as its separator, as desired. ◇
Claim 3.5.2. If M ∈ M M\in\mathcal{M} M ∈ M is bad, then I △ M I\triangle M I △ M induces s s s .
Proof of the claim. Since s s s is proper, there are elements of M \mathcal{M} M , one contained in I I I , the other in the complement of I I I , that are not bad. By Claim 3.5.1 , these two elements of M \mathcal{M} M induce separations that have the whole of X X X in their separator. Thus, for every bad M ∈ M M\in\mathcal{M} M ∈ M , the symmetric difference I △ M I\triangle M I △ M induces the separation s s s induced by I I I . ◇
By Claim 3.5.2 , for every bad M ∈ M M\in\mathcal{M} M ∈ M , the index set I △ M I\triangle M I △ M induces s s s but has a strictly smaller interval catalogue. Therefore, no M ∈ M M\in\mathcal{M} M ∈ M is bad. Hence, since the separation induced by I I I is narrow, by Claim 3.5.1 the separator of every separation induced by some M ∈ M M\in\mathcal{M} M ∈ M is equal to X X X .
Thus we have constructed a partition of Z ℓ \mathbb{Z}_{\ell} Z ℓ into at least three intervals such that for each of the separations induced by them, their separators are equal to X X X . By merging adjacent intervals if necessary, we obtain such a family with exactly three elements. (Note that in a merge, the separator does not change as for each interval J J J in our partition, ( ⋃ j ∈ J P j ) ∖ X \left(\bigcup_{j\in J}P_j\right)\setminus X ( ⋃ j ∈ J P j ) ∖ X has the whole of X X X in its neighbourhood.)
Hence we can apply Lemma 3.3 to deduce that the flower O \mathcal{O} O is a strong anemone, as desired. □
Example 3.6. In this example we explain the motivation behind our definition of ‘daisy’ given above. We approached this by attempting to have the most expressive definition of daisy that still allows for a dichotomy between daisies and anemones, and so that it specialises to the established definition of daisy in the k k k -connected case. The cycle-decomposition depicted in Figure 3.2 is a 6-flower that is not an anemone, and thus is a 6-daisy by Flower-Dichotomy Theorem (3.5); in particular, the term ‘narrow’ in the definition of daisy cannot be dropped.
Figure 3.2. We depict a cycle-decomposition that is a 6-flower. The separation that has the three bags shaded green on one side and all other bags on the other side has order exactly six but is not narrow.
Example 3.7. The cycle-decomposition depicted in Figure 3.3 is a 3-flower that is not an anemone, and thus is a daisy. The figure shows that the condition ‘order exactly k k k ’ in the definition of daisy cannot be relaxed to ‘order at most k k k ’ with the Flower-Dichotomy Theorem (3.5) being true at the same time.
Figure 3.3. We depict a cycle-decomposition that is a 3-flower. The separation that has the two bags shaded green on one side and all other bags on the other side has order one and is narrow.
Lemma 3.8. Let O = ( Z ℓ , ( P i ∣ i ∈ Z ℓ ) ) \mathcal{O}=(\mathbb{Z}_{\ell},(P_i\mid i\in\mathbb{Z}_{\ell})) O = ( Z ℓ , ( P i ∣ i ∈ Z ℓ )) be a k k k -flower of a graph G G G and let s s s be a narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separation that almost-interlaces the star of petal-separations of O \mathcal{O} O . Then s s s is equal to a separation induced by some I ⊆ Z ℓ I\subseteq\mathbb{Z}_{\ell} I ⊆ Z ℓ .
Moreover, if s s s has order exactly k k k and O \mathcal{O} O is a daisy, then I I I can be chosen to be an interval.
Proof. Let s = : ( C , D ) s =: (C, D) s =: ( C , D ) . Let I I I be the set of all indices i ∈ Z ℓ i \in\mathbb{Z}_{\ell} i ∈ Z ℓ with P i ⊆ C P_i \subseteq C P i ⊆ C . By construction ⋃ i ∈ I P i ⊆ C \bigcup_{i\in I} P_i \subseteq C ⋃ i ∈ I P i ⊆ C . For every j ∈ Z ℓ ∖ I j \in\mathbb{Z}_{\ell} \setminus I j ∈ Z ℓ ∖ I , we have that P j P_j P j is not a subset of C C C . Since s s s almost-interlaces the star of petal-separations of O \mathcal{O} O , it follows that P j ⊆ D P_j \subseteq D P j ⊆ D . Hence ⋃ j ∉ I P j ⊆ D \bigcup_{j\notin I} P_j \subseteq D ⋃ j ∈ / I P j ⊆ D . It follows that the separation induced by I I I is equal to s s s up to possibly reassigning vertices of the separator C ∩ D C \cap D C ∩ D to one of the sides.
Since s s s is proper, the separation induced by I I I is proper too. Since s s s is narrow, the separator of the separation induced by I I I is equal to sep ( s ) \operatorname{sep}(s) sep ( s ) . Hence the separation induced by I I I is equal to s s s .
The ‘Moreover’-part is immediate from the definition of daisy. This completes the proof. □
Graphic-flowers
Given a k k k -flower F = ( Z ℓ , P ) \mathcal{F} = (\mathbb{Z}_{\ell},\mathcal{P}) F = ( Z ℓ , P ) , we are interested in all narrow proper k k k -separations that are induced by F \mathcal{F} F , call them S 0 \mathcal{S}_0 S 0 . In the case of daisies, they are easy to spot: they are induced by some interval of Z ℓ \mathbb{Z}_{\ell} Z ℓ . However, in the case of anemones, they are induced by arbitrary bipartitions of Z ℓ \mathbb{Z}_{\ell} Z ℓ and are therefore no longer displayed by the cycle-decomposition F \mathcal{F} F . Moreover, we could arrange the bags of P \mathcal{P} P in an arbitrary order along the cycle. Therefore, we have no canonical choice for constructing such k k k -anemones. To circumvent both issues, we instead display all separations in S 0 \mathcal{S}_0 S 0 by a star-decomposition S \mathcal{S} S in the following way.
Definition 3.9 (*Graphic-k k k -flower). A graphic-k k k -anemone of a graph G G G is a star-decomposition S = ( S , V ) \mathcal{S}=(S,\mathcal{V}) S = ( S , V ) such that the central bag V c V_c V c has size ≤ k \le k ≤ k , and every leaf-bag contains V c V_c V c . We define its centre c c c as a P P P -node, and declare that there are no other P P P -nodes. A graphic-k k k -daisy of G G G is a k k k -daisy of G G G . The term graphic-k k k -flower summarizes graphic-k k k -anemones and graphic-k k k -daisies.
For later use, we include the following definition:
Definition 3.10 (Centre). The centre of a graphic-k k k -flower is the intersection of all its bags.
Observation 3.11. Let O = ( Z ℓ , P ) \mathcal{O}=(\mathbb{Z}_{\ell},\mathcal{P}) O = ( Z ℓ , P ) be a k k k -daisy of a graph G G G . Then D = O \mathcal{D}=\mathcal{O} D = O is a graphic-k k k -daisy. Moreover, every narrow proper k k k -separation induced by O \mathcal{O} O is a displayed separation of D \mathcal{D} D .
Proof. That D \mathcal{D} D is a graphic-k k k -daisy is trivial. If s s s is a narrow proper k k k -separation induced by some I ⊆ Z ℓ I\subseteq\mathbb{Z}_{\ell} I ⊆ Z ℓ in O \mathcal{O} O , then I I I can be chosen to be an interval (by definition of ‘k k k -daisy’). Hence, s s s is a displayed separation of D \mathcal{D} D . □
Lemma 3.12. Let A = ( S , V ) \mathcal{A}=(S,\mathcal{V}) A = ( S , V ) be a graphic-k k k -anemone of a graph G G G . Then there is a k k k -anemone O \mathcal{O} O such that D ( O ) ⊆ D ( A ) \mathcal{D}(\mathcal{O})\subseteq\mathcal{D}(\mathcal{A}) D ( O ) ⊆ D ( A ) and π ( O ) = π ( A ) \pi(\mathcal{O})=\pi(\mathcal{A}) π ( O ) = π ( A ) . Moreover, given two displayed separations s , s ′ ∈ D ( A ) s,s'\in\mathcal{D}(\mathcal{A}) s , s ′ ∈ D ( A ) of A \mathcal{A} A , we can choose O \mathcal{O} O such that s , s ′ ∈ D ( O ) s,s'\in\mathcal{D}(\mathcal{O}) s , s ′ ∈ D ( O ) .
Proof. We let O \mathcal{O} O be a cycle on the nodes V ( S ) V(S) V ( S ) . Then O : = ( O , V ) \mathcal{O}:=(\mathcal{O},\mathcal{V}) O := ( O , V ) is a cycle-decomposition. It is easy to see that π ( O ) = π ( A ) \pi(\mathcal{O})=\pi(\mathcal{A}) π ( O ) = π ( A ) . Let s s s be a displayed separation of O \mathcal{O} O induced by a cut ( U , W ) (U,W) ( U , W ) of O \mathcal{O} O . Then, ( U , W ) (U,W) ( U , W ) induces s s s in A \mathcal{A} A , hence s s s is displayed by A \mathcal{A} A .
For the ‘Moreover’-part, let s , s ′ ∈ D ( A ) s,s'\in\mathcal{D}(\mathcal{A}) s , s ′ ∈ D ( A ) be displayed separations induced by cuts ( U , W ) (U,W) ( U , W ) and ( U ′ , W ′ ) (U',W') ( U ′ , W ′ ) of S S S , respectively. We arrange the nodes of S S S cyclically in four consecutive blocks U ∩ U ′ U\cap U' U ∩ U ′ , U ∩ W ′ U\cap W' U ∩ W ′ , W ∩ W ′ W\cap W' W ∩ W ′ , and W ∩ U ′ W\cap U' W ∩ U ′ , omitting empty blocks and ordering the nodes within each block arbitrarily. Then, s s s and s ′ s' s ′ are displayed by O \mathcal{O} O . □
In the following lemma, we prove that every separation displayed by a graphic-k k k -flower F F F with both sides nonempty is either close to a petal-separation or crossed significantly by a displayed separation of F F F .
Lemma 3.13. Let F F F be a graphic-k k k -flower of a graph G G G . Let ( A , B ) (A,B) ( A , B ) be a displayed separation of F F F with both sides nonempty. Let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N . Then we have at least one of the following:
( A , B ) (A,B) ( A , B ) is 2 ε 2\varepsilon 2 ε -close to some petal-separation of F F F or its inverse; or
F F F has a displayed separation that ( ε − k ) (\varepsilon-k) ( ε − k ) -crisses ( A , B ) (A,B) ( A , B ) .
Proof. By Lemma 3.12, F F F has a cycle decomposition ( O , V ) (\mathcal{O},\mathcal{V}) ( O , V ) such that every 2-edge cut of O \mathcal{O} O induces a separation displayed by F F F and there is a 2-edge-cut of O \mathcal{O} O that induces ( A , B ) (A,B) ( A , B ) . Recall that V ↾ X : = ⋃ x ∈ V ( X ) V x \mathcal{V}\mathord{\upharpoonright}X:=\bigcup_{x\in V(X)}V_x V ↾ X := ⋃ x ∈ V ( X ) V x for all subgraphs X ⊆ O X\subseteq\mathcal{O} X ⊆ O . Let P , Q ⊆ O P,Q\subseteq\mathcal{O} P , Q ⊆ O be paths with adjacent endvertices such that A = V ↾ P A=\mathcal{V}\mathord{\upharpoonright}P A = V ↾ P and B = V ↾ Q B=\mathcal{V}\mathord{\upharpoonright}Q B = V ↾ Q . If ∣ A ∣ ≤ ε |A|\le\varepsilon ∣ A ∣ ≤ ε , then ( A , B ) (A,B) ( A , B ) is ε \varepsilon ε -close to the petal-separation at any node of P P P . Similarly, if ∣ B ∣ ≤ ε |B|\le\varepsilon ∣ B ∣ ≤ ε , then ( A , B ) (A,B) ( A , B ) is ε \varepsilon ε -close to the inverse of the petal-separation at any node of Q Q Q . Hence, we can assume ∣ A ∣ , ∣ B ∣ > ε |A|,|B|>\varepsilon ∣ A ∣ , ∣ B ∣ > ε .
Let p p p be the first node on P P P such that ∣ V ↾ P p ∣ > ε |\mathcal{V}\mathord{\upharpoonright}Pp|>\varepsilon ∣ V ↾ P p ∣ > ε , which exists since ∣ A ∣ = ∣ V ↾ P ∣ > ε |A|=|\mathcal{V}\mathord{\upharpoonright}P|>\varepsilon ∣ A ∣ = ∣ V ↾ P ∣ > ε . If ∣ V ↾ p P ∣ ≤ ε |\mathcal{V}\mathord{\upharpoonright}pP|\le\varepsilon ∣ V ↾ pP ∣ ≤ ε , then ( A , B ) (A,B) ( A , B ) is 2 ε 2\varepsilon 2 ε -close to the petal-separation s p s_p s p of F F F at p p p , and so we have outcome (i). Hence we have ∣ V ↾ p P ˚ ∣ > ε |V \mathbin{\upharpoonright} \mathring{pP}| > \varepsilon ∣ V ↾ pP ˚ ∣ > ε in addition to ∣ V ↾ P p ∣ > ε |V \mathbin{\upharpoonright} Pp| > \varepsilon ∣ V ↾ P p ∣ > ε . Similarly, we find a node q ∈ V ( Q ) q \in V(Q) q ∈ V ( Q ) such that ∣ V ↾ Q q ∣ > ε |V \mathbin{\upharpoonright} Qq| > \varepsilon ∣ V ↾ Qq ∣ > ε and ∣ V ↾ q Q ˚ ∣ > ε |V \mathbin{\upharpoonright} \mathring{qQ}| > \varepsilon ∣ V ↾ q Q ˚ ∣ > ε . Let C : = V ↾ ( P p ∪ Q q ) C := V \mathbin{\upharpoonright} (Pp \cup Qq) C := V ↾ ( P p ∪ Qq ) and D : = V ↾ ( p P ˚ ∪ q Q ˚ ) D := V \mathbin{\upharpoonright} (\mathring{pP} \cup\mathring{qQ}) D := V ↾ ( pP ˚ ∪ q Q ˚ ) . Then ( C , D ) (C,D) ( C , D ) is a displayed separation of F F F , and therefore has order at most k k k . It remains to show that ( C , D ) (C,D) ( C , D ) ( ε − k ) (\varepsilon-k) ( ε − k ) -crisses ( A , B ) (A,B) ( A , B ) . By symmetry, it suffices to consider the corner for A C AC A C . The A C AC A C -corner is the proper side of the displayed separation of F F F corresponding to the path P p Pp P p of O O O . Since F F F is a graphic-k k k -flower, this separation has order at most k k k . Therefore, the A C AC A C -corner has size ≥ ∣ V ↾ P p ∣ − k > ε − k \ge|V \mathbin{\upharpoonright} Pp| - k > \varepsilon-k ≥ ∣ V ↾ P p ∣ − k > ε − k . This gives outcome (ii).
The 3D Corner Lemma
The aim of this section is to prove Key Lemma 4.14, which can be interpreted as a way of saying that the crossing pattern of separations is essentially 2-dimensional. In a sense it says that under certain assumptions the Venn-diagram of three crossing separations has lots of empty entries, which leads to a 2-dimensional representation of the Venn-diagram when empty regions of the Venn-diagram are omitted. Our proof strategy is to first prove Lemma 4.1, a purely combinatorial lemma about the cube, and then apply it to the Venn-diagram of three crossing separations to deduce Key Lemma 4.14.
Some combinatorial results on the cube
We remind the reader of the following standard notation. If G G G is a graph, a set R ⊆ V ( G ) R \subseteq V(G) R ⊆ V ( G ) is a vertex cover of G G G if every edge of G G G is incident with a vertex in R R R . The cube is the CW complex with vertex set { 0 , 1 } 3 \{0,1\}^{3} { 0 , 1 } 3 where for each d ∈ { 0 , 1 , 2 , 3 } d \in\{0,1,2,3\} d ∈ { 0 , 1 , 2 , 3 } , its faces of dimension d d d are the sets of vertices obtained by fixing 3 − d 3-d 3 − d coordinates and letting the remaining d d d coordinates vary freely over { 0 , 1 } \{0,1\} { 0 , 1 } . Throughout this section, we denote it by Q Q Q . Two vertices of the cube are diagonal if they differ in every coordinate; each vertex is diagonal to a unique vertex. Given a partition of V ( Q ) V(Q) V ( Q ) into two sets A A A and B B B , the cut c = E ( A , B ) c = E(A,B) c = E ( A , B ) between these two vertex sets consists of the edges with one endvertex in A A A and the other in B B B . A A A and B B B are the sides of this cut. A cut is atomic if A A A or B B B is a singleton. A cut c = E ( A , B ) c = E(A,B) c = E ( A , B ) of the cube is called a coordinate cut if there is a coordinate such that A A A consists of all vertices that take the value zero in that coordinate and B B B consists of all vertices taking the value one in that coordinate, or vice versa.
To state the next lemma, we further introduce the following notation. We fix a function λ \lambda λ from the 1D, 2D and 3D faces of the cube to non-negative integers. The λ \lambda λ -value of a cut c c c is given by the sum λ ( c ) = ∑ λ ( f ) \lambda(c) = \sum\lambda(f) λ ( c ) = ∑ λ ( f ) ranging over all faces f f f which contain both endpoints of an edge in c c c . For k ∈ N k \in\mathbb{N} k ∈ N , we say that λ \lambda λ is a ( k - ) (k\text{-}) ( k - ) flower if the following holds:
λ \lambda λ assigns the value zero to all 2D faces; and
there is a pair of two diagonal vertices such that λ \lambda λ assigns the value zero to all 1D faces incident with one of them; and
λ \lambda λ assigns the same value ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N with ℓ ≤ k / 2 \ell\le k/2 ℓ ≤ k /2 to all remaining 1D faces; and
λ \lambda λ assigns the value k − 2 ℓ k-2\ell k − 2 ℓ to the 3D face.
We say that λ \lambda λ is a k k k -corner-map if every coordinate cut has λ \lambda λ -value k k k .
Lemma 4.1. Let λ \lambda λ be a k k k -corner-map. Let R R R be a vertex cover of the cube, and assume that every cut with at least one vertex of R R R on each side has λ \lambda λ -value at least k k k . Then λ \lambda λ is a flower.
In order to prove this lemma, we first prove the following simple fact about the structure of vertex covers of the cube. A 6-cycle of the cube is anti-diagonal if its complement consists of a pair of diagonal vertices, see Figure 4.1 .
Figure 4.1. Left: a vertex cover of the cube by an independent set of size four. Right: a vertex cover of the cube by an anti-diagonal cycle.
Lemma 4.2. A set X ⊆ V ( Q ) X \subseteq V(Q) X ⊆ V ( Q ) is a vertex cover of the cube Q Q Q if and only if X X X contains an independent vertex set of size four, or induces an anti-diagonal 6-cycle.
Proof. The reverse implication is trivial. Conversely, assume that X ⊆ V ( Q ) X \subseteq V(Q) X ⊆ V ( Q ) is a vertex cover of Q Q Q . Since we are done otherwise, assume that some vertex of Q Q Q is not in X X X , call it v v v . Since X X X covers all edges of Q Q Q , all neighbours of v v v must be in X X X . Denote by w w w the unique vertex diagonal to v v v in Q Q Q . If w ∈ X w \in X w ∈ X then the three neighbours of v v v together with w w w form an independent set of size four. If w ∉ X w \notin X w ∈ / X then all its three neighbours are in X X X , and so X X X has size six and forms an anti-diagonal 6-cycle. □
Motivated by Lemma 4.2 , we prove Lemma 4.1 in the two special cases where R R R is an independent set of size four and an anti-diagonal 6-cycle. To prepare for this, we first prove the following. We abbreviate ∑ λ ( f ) \sum\lambda(f) ∑ λ ( f ) where the sum ranges over all faces f f f of dimension d d d by s ( d ) s(d) s ( d ) .
Lemma 4.3. Let λ \lambda λ be a k k k -corner-map. Then,
3 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) = 3 k . 3s(3) + 2s(2) + s(1) = 3k. 3 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) = 3 k . Proof. Let c 1 c_1 c 1 , c 2 c_2 c 2 , c 3 c_3 c 3 be the three coordinate cuts. By assumption, we have λ ( c 1 ) + λ ( c 2 ) + λ ( c 3 ) = 3 k \lambda(c_1) + \lambda(c_2) + \lambda(c_3) = 3k λ ( c 1 ) + λ ( c 2 ) + λ ( c 3 ) = 3 k . The 3D face contributes to all three terms on the left hand side, every 2D face contributes to two terms and every 1D face contributes to one term (see Figure 4.2 ). Hence the left hand side is equal to 3 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) 3s(3) + 2s(2) + s(1) 3 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) , as desired. □
Figure 4.2. Illustration of the proof of Lemma 4.3 . The cuts that we consider are displayed in red, blue and green. The 3-dimensional face contains an edge from each cut, every 2-dimensional face contains an edge from two of the cuts, and every 1-dimensional face contains an edge from a single cut.
Lemma 4.4. Let λ \lambda λ be a k k k -corner-map. Let R R R be an independent set of size four of the cube, and assume that every atomic cut of a vertex in R R R has λ \lambda λ -value at least k k k . Then λ \lambda λ is a flower.
Given a vertex r r r , we abbreviate the atomic cut c = E ( { r } , V − r ) c = E(\{r\}, V-r) c = E ({ r } , V − r ) by c ( r ) c(r) c ( r ) .
Proof of Lemma 4.4 . By assumption, we have ∑ r ∈ R λ ( c ( r ) ) ≥ 4 k \sum_{r\in R}\lambda(c(r)) \ge4k ∑ r ∈ R λ ( c ( r )) ≥ 4 k . The 3D face contributes to all four terms on the left hand side, every 2D face contributes to two terms and every 1D face contributes to one term (see Figure 4.3 ). Hence the left hand side is equal to 4 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) 4s(3)+2s(2)+s(1) 4 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) , so
Figure 4.3. Illustration of the proof of Lemma 4.4 . The vertices in R R R are displayed in red. We consider the atomic cuts at the vertices of R R R . The 3-dimensional face contains an edge from each of them, every 2-dimensional face contains an edge from exactly two of them, and every 1-dimensional face contains an edge from exactly one of them.
4 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) ≥ 4 k 4s(3)+2s(2)+s(1) \ge4k 4 s ( 3 ) + 2 s ( 2 ) + s ( 1 ) ≥ 4 k
Subtracting the equation from Lemma 4.3 gives s ( 3 ) ≥ k s(3) \ge k s ( 3 ) ≥ k . Since s ( 3 ) s(3) s ( 3 ) contributes to the λ \lambda λ -value of every coordinate cut, and the λ \lambda λ -values of coordinate cuts are equal to k k k , by non-negativity of λ \lambda λ , we get that s ( 3 ) = k s(3)=k s ( 3 ) = k and all other values contributing to the λ \lambda λ -value of some coordinate cut must be zero. Since every face contributes to some coordinate cut, λ \lambda λ assigns the value k k k to the 3D face and zero to all other faces, so λ \lambda λ is a flower. □
Lemma 4.5. Let λ \lambda λ be a k k k -corner-map. Let R R R be an anti-diagonal cycle of the cube, and assume that every cut with at least one vertex of R R R on each side has λ \lambda λ -value at least k k k . Then λ \lambda λ is a flower.
Proof. This proof follows the same strategy as the proof of Lemma 4.4 , with slightly more complicated computations. By assumption, we have ∑ r ∈ R λ ( c ( r ) ) ≥ 6 k \sum_{r\in R}\lambda(c(r)) \ge6k ∑ r ∈ R λ ( c ( r )) ≥ 6 k . The 3D face contributes to all six terms on the left hand side, every 2D face contributes to three terms, every 1D face incident to two vertices in R R R contributes to two terms, and all other 1D faces contribute to exactly one term since R R R is a vertex cover (see Figure 4.4 ). Let s + ( 1 ) = ∑ λ ( f ) s^{+}(1)=\sum\lambda(f) s + ( 1 ) = ∑ λ ( f ) where the sum ranges over all 1D faces with both endpoints in R R R , and s − ( 1 ) = ∑ λ ( f ) s^{-}(1)=\sum\lambda(f) s − ( 1 ) = ∑ λ ( f ) where the sum ranges over all other 1D faces. Hence, the left hand side is equal to 6 s ( 3 ) + 3 s ( 2 ) + 2 s + ( 1 ) + s − ( 1 ) 6s(3)+3s(2)+2s^{+}(1)+s^{-}(1) 6 s ( 3 ) + 3 s ( 2 ) + 2 s + ( 1 ) + s − ( 1 ) , so
6 s ( 3 ) + 3 s ( 2 ) + 2 s + ( 1 ) + s − ( 1 ) ≥ 6 k . (1) 6s(3)+3s(2)+2s^{+}(1)+s^{-}(1)\ge6k. \tag*{(1)} 6 s ( 3 ) + 3 s ( 2 ) + 2 s + ( 1 ) + s − ( 1 ) ≥ 6 k . (1) Subtracting twice the equation from Lemma 4.3 , and using that s ( 1 ) = s + ( 1 ) + s − ( 1 ) s(1)=s^{+}(1)+s^{-}(1) s ( 1 ) = s + ( 1 ) + s − ( 1 ) , we get
− s ( 2 ) − s − ( 1 ) ≥ 0 -s(2)-s^{-}(1)\ge0 − s ( 2 ) − s − ( 1 ) ≥ 0 Since λ \lambda λ is non-negative, it follows that s ( 2 ) = 0 s(2)=0 s ( 2 ) = 0 and s − ( 1 ) = 0 s^{-}(1)=0 s − ( 1 ) = 0 , so λ \lambda λ satisfies the first two conditions in the definition of a flower. Furthermore, (1) must be an equality. Hence, all the inequalities we used to derive (1) must also be equalities. Thus, for every r ∈ R r\in R r ∈ R , we have λ ( c ( r ) ) = k \lambda(c(r))=k λ ( c ( r )) = k . Denote the λ \lambda λ -value of the 3D face by x x x , and let k ′ : = k − x k':=k-x k ′ := k − x . Now enumerate the edges (that is, 1D faces) that are on the anti-diagonal cycle by e 0 , … , e 5 e_0,\ldots,e_5 e 0 , … , e 5 according to a cyclic orientation of that cycle. Since λ ( c ( r ) ) = k \lambda(c(r))=k λ ( c ( r )) = k for every r ∈ R r\in R r ∈ R , we get λ ( e i ) + λ ( e i + 1 ) = k ′ \lambda(e_i)+\lambda(e_{i+1})=k' λ ( e i ) + λ ( e i + 1 ) = k ′ for every i ∈ Z 6 i\in\mathbb{Z}_6 i ∈ Z 6 . Combining these equations at consecutive vertices yields λ ( e i ) = λ ( e i + 2 ) \lambda(e_i)=\lambda(e_{i+2}) λ ( e i ) = λ ( e i + 2 ) for all i ∈ Z 6 i\in\mathbb{Z}_6 i ∈ Z 6 . By renaming the edges if necessary, assume that λ ( e 0 ) ≤ λ ( e 1 ) \lambda(e_0)\le\lambda(e_1) λ ( e 0 ) ≤ λ ( e 1 ) . Let u , v ∈ V ( Q ) u,v\in V(Q) u , v ∈ V ( Q ) be the endpoints of e 1 e_1 e 1 , and consider the cut c : = E ( { u , v } , V ( Q ) ∖ { u , v } ) c:=E(\{u,v\},V(Q)\setminus\{u,v\}) c := E ({ u , v } , V ( Q ) ∖ { u , v }) . Both sides of c c c contain vertices of R R R , so λ ( c ) ≥ k \lambda(c)\ge k λ ( c ) ≥ k . Since all 2D faces and all 1D faces incident to a vertex outside of R R R have λ \lambda λ -value zero, we have λ ( c ) = x + λ ( e 0 ) + λ ( e 2 ) \lambda(c)=x+\lambda(e_0)+\lambda(e_2) λ ( c ) = x + λ ( e 0 ) + λ ( e 2 ) . Therefore, λ ( e 0 ) + λ ( e 2 ) ≥ k ′ \lambda(e_0)+\lambda(e_2)\ge k' λ ( e 0 ) + λ ( e 2 ) ≥ k ′ so λ ( e 0 ) ≥ k ′ / 2 \lambda(e_0)\ge k'/2 λ ( e 0 ) ≥ k ′ /2 . Thus λ ( e 0 ) = λ ( e 1 ) = k ′ / 2 = : ℓ \lambda(e_0)=\lambda(e_1)=k'/2=:\ell λ ( e 0 ) = λ ( e 1 ) = k ′ /2 =: ℓ . Hence λ \lambda λ assigns the same value ℓ ≤ k / 2 \ell\le k/2 ℓ ≤ k /2 at all edges of the anti-diagonal cycle, and value x = k − 2 ℓ x=k-2\ell x = k − 2 ℓ to the 3D face. This concludes the proof that λ \lambda λ is a flower. □
Figure 4.4. Illustration of the proof of Lemma 4.5 . The vertices in R R R are displayed in red. We consider the atomic cuts at the vertices of R R R . The 3-dimensional face contains an edge from each of them, every 2-dimensional face contains an edge from exactly three of them. Every 1-dimensional face incident to two vertices of R R R contains an edge from exactly two of them, and all other 1-dimensional faces contain an edge from exactly one of them.
We now have all we need to prove Lemma 4.1, which we restate for convenience.
Lemma 4.1. Let λ \lambda λ be a k k k -corner-map. Let R R R be a vertex cover of the cube, and assume that every cut with at least one vertex of R R R on each side has λ \lambda λ -value at least k k k . Then λ \lambda λ is a flower.
Proof. By Lemma 4.2 , it suffices to prove it when R R R is an independent set of size four or an anti-diagonal 6-cycle. This is done in Lemmas 4.4 and 4.5 , respectively. □
We now prove additional results in the same flavour, which we will later use to analyse how triples of separations can cross. A cut c = E ( A , B ) c=E(A,B) c = E ( A , B ) is a pair cut if either A A A or B B B consists only of two adjacent vertices of Q Q Q . We then say that these two vertices are contained in the pair cut. The statement of the following lemma is illustrated in Figure 4.5 .
Figure 4.5. Illustration of Lemma 4.6 . The vertices u u u and v v v are the endpoints of the red edge. The faces in red (plain or dashed) have λ \lambda λ -value 0 by assumption. Lemma 4.6 then states that the faces in blue (plain or dashed) also have λ \lambda λ -value 0.
Lemma 4.6. Let λ \lambda λ be a k k k -corner-map. Suppose that there exist adjacent vertices u , v ∈ V ( Q ) u,v\in V(Q) u , v ∈ V ( Q ) such that all 1D and 2D faces which contain u u u and v v v have λ \lambda λ -value 0, and all pair cuts which contain
exactly one of u , v u,v u , v have λ \lambda λ -value k k k . Let u ′ , v ′ ∈ V ( Q ) u',v' \in V(Q) u ′ , v ′ ∈ V ( Q ) be the vertices which are diagonal to u , v u,v u , v , respectively. Then, all 1D and 2D faces which contain u ′ u' u ′ and v ′ v' v ′ have λ \lambda λ -value 0.
Proof. Let C 1 C_1 C 1 be the set of all pair cuts which contain exactly one of u , v u,v u , v , and let C 2 C_2 C 2 be the multiset which contains the coordinate cut which contains the edge u v uv uv with multiplicity two, and the other two coordinate cuts with multiplicity one. By assumption, we have ∑ c ∈ C 1 λ ( c ) = 4 k = ∑ c ∈ C 2 λ ( c ) \sum_{c\in C_1}\lambda(c)=4k=\sum_{c\in C_2}\lambda(c) ∑ c ∈ C 1 λ ( c ) = 4 k = ∑ c ∈ C 2 λ ( c ) . We now analyse the contribution of each face to the left hand side and the right hand side, see Figure 4.6 . The 3D face contributes to four terms on both sides. The 2D faces which contain the edge u v uv uv have λ \lambda λ -value 0. The 2D faces f 1 , f 2 f_1,f_2 f 1 , f 2 which contain the edge u ′ v ′ u'v' u ′ v ′ contribute to two terms on the left hand side and to three terms on the right hand side. The other 2D faces contribute to two terms on both sides. The 1D face which contains u u u and v v v has λ \lambda λ -value 0. The 1D face f 3 f_3 f 3 which contains u ′ u' u ′ and v ′ v' v ′ does not contribute to the left hand side, and contributes to two terms on the right hand side. The 1D faces which are incident to exactly one of u , v , u ′ , v ′ u,v,u',v' u , v , u ′ , v ′ contribute to one term on both sides. The other 1D faces contribute to two terms on both sides. Thus, we get
Figure 4.6. Illustration of the proof of Lemma 4.6 . On the left, the colored edges correspond to the left side of the separations in C 1 \mathcal{C}_1 C 1 . On the right, the colored facets correspond to the left side of the separations in C 2 \mathcal{C}_2 C 2 . For each face f f f of the cube, we compare the number of colored edges that intersect f f f without containing f f f , with the number of colored facets that intersect f f f without containing f f f , counting the red facet twice. This number is the same, except for the zones which are highlighted in Figure 4.5 .
0 = λ ( f 1 ) + λ ( f 2 ) + 2 λ ( f 3 ) . 0 = \lambda(f_1) + \lambda(f_2) + 2\lambda(f_3). 0 = λ ( f 1 ) + λ ( f 2 ) + 2 λ ( f 3 ) .
Since λ \lambda λ is non-negative, this concludes the proof. □
Lemma 4.7. Let λ \lambda λ be a k k k -corner-map. Let u , v ∈ V ( Q ) u,v \in V(Q) u , v ∈ V ( Q ) be two vertices of Q Q Q which are at distance 2, and let w w w be one of their common neighbours. Let x x x be the third neighbour of w w w . Suppose that the atomic cuts c ( u ) c(u) c ( u ) and c ( v ) c(v) c ( v ) have λ \lambda λ -value at least k k k , and that the pair cuts which contain respectively u u u and w w w , and v v v and w w w have λ \lambda λ -value at most k k k . Then, all one- and two-dimensional faces which contain both w w w and x x x have λ \lambda λ -value 0.
Proof. Let C 1 = { c ( u ) , c ( v ) } \mathcal{C}_1 = \{c(u),c(v)\} C 1 = { c ( u ) , c ( v )} , and C 2 \mathcal{C}_2 C 2 contain the pair cuts which contain respectively u u u and w w w , and v v v and w w w . By assumption, we have ∑ c ∈ C 1 λ ( c ) ≥ 2 k ≥ ∑ c ∈ C 2 λ ( c ) \sum_{c\in\mathcal{C}_1}\lambda(c) \ge2k \ge\sum_{c\in\mathcal{C}_2}\lambda(c) ∑ c ∈ C 1 λ ( c ) ≥ 2 k ≥ ∑ c ∈ C 2 λ ( c ) . We now analyse the contribution of each face to the left hand side and the right hand side, see Figure 4.7 . The 3D face contributes to two terms on both sides. The 2D face which contains u u u and v v v contributes to two terms on both sides. The other 2D faces f 1 , f 2 f_1,f_2 f 1 , f 2 which contain x x x contribute to one term on the left hand side and to two terms on the right hand side. The 2D face which contains neither u u u nor v v v contributes to neither side. The last two 2D faces contribute to one term on both sides. The 1D face f 3 f_3 f 3 which contains w w w and x x x does not contribute to the left hand side, and contributes to two terms on the right hand side. The edges which are incident to one of u , v u,v u , v contribute to one term on both sides. The other edges contribute to neither side. Thus, we get
Figure 4.7. Illustration of the proof of Lemma 4.7. u u u and v v v are the colored vertices, w w w is the common endpoint of the colored edges. On the left, the colored vertices correspond to the left side of the separations in C 1 \mathcal{C}_1 C 1 . On the right, the colored edges correspond to the right side of the separations in C 2 \mathcal{C}_2 C 2 . For each face f f f of the cube, we compare the number of colored vertices that are contained in f f f , with the number of colored edges that intersect f f f without containing f f f .
0 ≥ λ ( f 1 ) + λ ( f 2 ) + 2 λ ( f 3 ) . 0 \ge\lambda(f_1)+\lambda(f_2)+2\lambda(f_3). 0 ≥ λ ( f 1 ) + λ ( f 2 ) + 2 λ ( f 3 ) .
Since λ \lambda λ is non-negative, this concludes the proof. □ \square □
Corner diagrams
Later, we will see how to apply the previous results to deduce properties on how three separations can cross. Before this, we review some notions about corner diagrams, which depict how two separations cross. Let us start by recalling some notions about corner diagrams. Consider a graph G G G and let r = ( A , B ) r=(A,B) r = ( A , B ) and s = ( C , D ) s=(C,D) s = ( C , D ) be two separations of G G G . These separations induce a partition of V ( G ) V(G) V ( G ) , which we represent using a corner diagram , see Figure 2.2 . The corresponding partition of V ( G ) V(G) V ( G ) has 9 parts, each of which corresponds to a zone of Figure 2.2 .
We now introduce the terminology that we will use to talk about these different parts. Since we will later work with corner diagrams for triples of separations, we need to introduce a new terminology to avoid ambiguity. Each part is of the form X ∩ Y X \cap Y X ∩ Y for X ∈ { A ∖ B , A ∩ B , B ∖ A } X \in\{A \setminus B,A \cap B,B \setminus A\} X ∈ { A ∖ B , A ∩ B , B ∖ A } and Y ∈ { C ∖ D , C ∩ D , D ∖ C } Y \in\{C \setminus D,C \cap D,D \setminus C\} Y ∈ { C ∖ D , C ∩ D , D ∖ C } . We then define the AC-corner to be the set ( A ∖ B ) ∩ ( C ∖ D ) (A \setminus B) \cap(C \setminus D) ( A ∖ B ) ∩ ( C ∖ D ) , and define the AD-corner , the BC-corner and the BD-corner analogously. The extended AC-corner is the set A ∩ C A \cap C A ∩ C , and the extended AD-corner , the extended BC-corner and the extended BD-corner are defined analogously. We define the rC-link to be the set ( A ∩ B ) ∩ ( C ∖ D ) (A \cap B) \cap(C \setminus D) ( A ∩ B ) ∩ ( C ∖ D ) , and define the rD-link , the As-link and the Bs-link analogously. Finally, we define the rs-centre as the set ( A ∩ B ) ∩ ( C ∩ D ) (A \cap B) \cap(C \cap D) ( A ∩ B ) ∩ ( C ∩ D ) .
We denote the size of the AC-corner by ∣ A C ∣ |AC| ∣ A C ∣ , and define ∣ A D ∣ |AD| ∣ A D ∣ , ∣ B C ∣ |BC| ∣ B C ∣ and ∣ B D ∣ |BD| ∣ B D ∣ similarly. We denote the size of the rC-link by ∣ r C ∣ |rC| ∣ r C ∣ , and define ∣ r D ∣ |rD| ∣ r D ∣ , ∣ A s ∣ |As| ∣ A s ∣ and ∣ B s ∣ |Bs| ∣ B s ∣ similarly. Finally, we denote the size of the rs-centre by ∣ r s ∣ |rs| ∣ r s ∣ . A corner separation is any separation of the form ( A ∩ C , B ∪ D ) (A \cap C,B \cup D) ( A ∩ C , B ∪ D ) , up to exchanging A A A and B B B , and/or C C C and D D D . Note that the corner separations are r ∧ s r \wedge s r ∧ s , r ∧ s ∗ r \wedge s^* r ∧ s ∗ , $r^* \wedge s and r ∗ ∧ s ∗ r^* \wedge s^* r ∗ ∧ s ∗ . For simplicity, we call the separation ( A ∩ C , B ∪ D ) (A \cap C, B \cup D) ( A ∩ C , B ∪ D ) the corner separation A C AC A C , and similarly for the other corner separations. The corners A C AC A C and B D BD B D are opposite , and the corners A D AD A D and B C BC B C as well. The following result is a fundamental property of the orders of opposite corner separations.
Lemma 4.8 (Submodularity, see e.g. [44 ] ). Let ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) be two separations of a graph. Then,
∣ ( A ∩ C ) ∩ ( B ∪ D ) ∣ + ∣ ( A ∪ C ) ∩ ( B ∩ D ) ∣ = ∣ A ∩ B ∣ + ∣ C ∩ D ∣ . |(A \cap C) \cap(B \cup D)| + |(A \cup C) \cap(B \cap D)| = |A \cap B| + |C \cap D|. ∣ ( A ∩ C ) ∩ ( B ∪ D ) ∣ + ∣ ( A ∪ C ) ∩ ( B ∩ D ) ∣ = ∣ A ∩ B ∣ + ∣ C ∩ D ∣. In particular, if ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) are two k k k -separations, then the orders of the opposite corner separations A C AC A C and B D BD B D sum to 2 k 2k 2 k .
Approximate notions on separations
Let G G G be a graph and ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N . We write X ⊆ ε Y X \subseteq_{\varepsilon} Y X ⊆ ε Y for X , Y ⊆ V ( G ) X,Y \subseteq V(G) X , Y ⊆ V ( G ) to mean that all but ≤ ε \leq\varepsilon ≤ ε vertices in X X X are contained in Y Y Y ; that is, ∣ X ∖ Y ∣ ≤ ε |X \setminus Y| \leq\varepsilon ∣ X ∖ Y ∣ ≤ ε . For ε = 0 \varepsilon=0 ε = 0 , this defaults to X ⊆ Y X \subseteq Y X ⊆ Y . We define X ≈ ε Y X \approx_{\varepsilon} Y X ≈ ε Y to mean X ⊆ ε Y X \subseteq_{\varepsilon} Y X ⊆ ε Y and Y ⊆ ε X Y \subseteq_{\varepsilon} X Y ⊆ ε X . We define a relation ≤ ε \leq_{\varepsilon} ≤ ε on the set of separations of G G G by letting
( A , B ) ≤ ε ( C , D ) : ⟺ A ⊆ ε C and B ⊇ ε D . (A, B) \leq_{\varepsilon} (C, D) \ :\Longleftrightarrow\ A \subseteq_{\varepsilon} C \text{ and } B \supseteq_{\varepsilon} D. ( A , B ) ≤ ε ( C , D ) :⟺ A ⊆ ε C and B ⊇ ε D .
As usual, we write ( A , B ) < ε ( C , D ) (A, B) <_{\varepsilon} (C, D) ( A , B ) < ε ( C , D ) if ( A , B ) ≤ ε ( C , D ) (A, B) \leq_{\varepsilon} (C, D) ( A , B ) ≤ ε ( C , D ) but ( C , D ) ̸ ≤ ε ( A , B ) (C, D) \not\leq_{\varepsilon} (A, B) ( C , D ) ≤ ε ( A , B ) . We write
( A , B ) ≈ ε ( C , D ) : ⟺ ( A , B ) ≤ ε ( C , D ) and ( C , D ) ≤ ε ( A , B ) . (A, B) \approx_{\varepsilon} (C, D) \ :\Longleftrightarrow\ (A, B) \leq_{\varepsilon} (C, D) \text{ and } (C, D) \leq_{\varepsilon} (A, B). ( A , B ) ≈ ε ( C , D ) :⟺ ( A , B ) ≤ ε ( C , D ) and ( C , D ) ≤ ε ( A , B ) .
If ( A , B ) ≈ ε ( C , D ) (A, B) \approx_{\varepsilon} (C, D) ( A , B ) ≈ ε ( C , D ) , then we say that ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) are ε \varepsilon ε -close. Note that this is equivalent to A ≈ ε C A \approx_{\varepsilon} C A ≈ ε C and B ≈ ε D B \approx_{\varepsilon} D B ≈ ε D . Note that neither ≤ ε \leq_{\varepsilon} ≤ ε nor ≈ ε \approx_{\varepsilon} ≈ ε nor < ε <_{\varepsilon} < ε needs to be transitive. However, we have:
Lemma 4.9. Let G G G be a graph. Let r ≤ ε s ≤ δ t r \leq_{\varepsilon} s \leq_{\delta} t r ≤ ε s ≤ δ t be separations of G G G where ε , δ ∈ N \varepsilon,\delta\in\mathbb{N} ε , δ ∈ N . Then r ≤ ε + δ t r \leq_{\varepsilon+\delta} t r ≤ ε + δ t .
Lemma 4.10. Let G G G be a graph. Let r , s , t r,s,t r , s , t be separations of G G G such that r ≤ ε s r \leq_{\varepsilon} s r ≤ ε s and r ≤ δ t r \leq_{\delta} t r ≤ δ t , where ε , δ ∈ N \varepsilon,\delta\in\mathbb{N} ε , δ ∈ N . Then r ≤ ε + δ s ∧ t r \leq_{\varepsilon+\delta} s \wedge t r ≤ ε + δ s ∧ t .
We say that ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) have edit-distance ≤ ε \leq\varepsilon ≤ ε if ∣ A △ C ∣ ≤ ε |A \mathbin{\triangle} C| \leq\varepsilon ∣ A △ C ∣ ≤ ε and ∣ B △ D ∣ ≤ ε |B \mathbin{\triangle} D| \leq\varepsilon ∣ B △ D ∣ ≤ ε .
Observation 4.11. Let s s s and t t t be two separations in a graph G G G , and let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N .
If s s s and t t t are ε \varepsilon ε -close, then s s s and t t t have edit-distance ≤ 2 ε \leq2\varepsilon ≤ 2 ε .
If s s s and t t t have edit-distance ≤ ε \leq\varepsilon ≤ ε , then s s s and t t t are ε \varepsilon ε -close.
Two separations ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) of G G G are ε \varepsilon ε -nested if, after possibly swapping the name A A A with B B B or C C C with D D D , we have ( A , B ) ≤ ε ( C , D ) (A, B) \leq_{\varepsilon} (C, D) ( A , B ) ≤ ε ( C , D ) . If ( A , B ) (A, B) ( A , B ) and ( C , D ) (C, D) ( C , D ) are not ε \varepsilon ε -nested, then they ε \varepsilon ε -cross. Note that for ε = 0 \varepsilon=0 ε = 0 , the notions of ε \varepsilon ε -nested and ε \varepsilon ε -cross coincide with the standard notions of nested and cross. A set S S S of separations of G G G is ε \varepsilon ε -nested if the separations in S S S are pairwise ε \varepsilon ε -nested.
3D corner diagrams
Let G G G be a graph and let ( A , B ) (A, B) ( A , B ) , ( C , D ) (C, D) ( C , D ) and ( E , F ) (E, F) ( E , F ) be three k k k -separations of G G G . Consider the cube on vertex set { A , B } × { C , D } × { E , F } \{A, B\} \times\{C, D\} \times\{E, F\} { A , B } × { C , D } × { E , F } . Label each face f f f of the cube by the set of letters in { A , B , C , D , E , F } \{A, B, C, D, E, F\} { A , B , C , D , E , F } which appear in the names of the vertices in f f f (for instance, the 1D face containing A C E ACE A C E and A C F ACF A C F has label A C E F ACEF A C E F ), see Figure 4.8 . Let φ \varphi φ be the function which maps every vertex v ∈ V ( G ) v \in V(G) v ∈ V ( G ) to the face with label { Z ∈ { A , B , C , D , E , F } : v ∈ Z } \{Z \in\{A, B, C, D, E, F\} : v \in Z\} { Z ∈ { A , B , C , D , E , F } : v ∈ Z } . For every 1D, 2D or 3D face f f f of the cube, let λ ( f ) = ∣ φ − 1 ( f ) ∣ \lambda(f) = |\varphi^{-1}(f)| λ ( f ) = ∣ φ − 1 ( f ) ∣ . We say that λ \lambda λ is the visualisation of ( A , B ) (A, B) ( A , B ) , ( C , D ) (C, D) ( C , D ) and ( E , F ) (E, F) ( E , F ) . We think of λ \lambda λ as a 3-dimensional generalization of corner diagrams. We now prove two simple but key properties of λ \lambda λ . The first one simply states that λ \lambda λ is a k k k -corner-map. The second one states that λ \lambda λ correctly measures the order of the separations that can arise from the 3D corner diagram.
Figure 4.8. The cube on vertex set { A , B } × { C , D } × { E , F } \{A, B\} \times\{C, D\} \times\{E, F\} { A , B } × { C , D } × { E , F } . The labels of some of the faces are indicated.
Lemma 4.12. Let ( A , B ) (A, B) ( A , B ) , ( C , D ) (C, D) ( C , D ) and ( E , F ) (E, F) ( E , F ) be three k k k -separations of a graph and let λ \lambda λ be their visualisation. Then, λ \lambda λ is a k k k -corner-map.
Proof. By definition, λ \lambda λ is a function from the 1D, 2D and 3D faces of the cube to the non-negative integers. By symmetry, it suffices to prove that the coordinate cut
c : = E ( { A C E , A C F , A D E , A D F } , { B C E , B C F , B D E , B D F } ) c := E(\{ACE, ACF, ADE, ADF\}, \{BCE, BCF, BDE, BDF\}) c := E ({ A C E , A C F , A D E , A D F } , { B C E , B C F , B D E , B D F })
has λ \lambda λ -value k k k . It follows immediately from the definition of λ \lambda λ and of c c c that λ ( c ) = ∣ A ∩ B ∣ = k \lambda(c) = |A \cap B| = k λ ( c ) = ∣ A ∩ B ∣ = k . □
Lemma 4.13. Let ( A , B ) (A, B) ( A , B ) , ( C , D ) (C, D) ( C , D ) and ( E , F ) (E, F) ( E , F ) be three separations of a graph and let λ \lambda λ be their visualisation. Let S ⊆ { A , B } × { C , D } × { E , F } S \subseteq\{A, B\} \times\{C, D\} \times\{E, F\} S ⊆ { A , B } × { C , D } × { E , F } , and let A ′ ⊆ V ( Q ) A' \subseteq V(Q) A ′ ⊆ V ( Q ) be the set of vertices of Q Q Q whose label is an element of S S S and let c = E ( A ′ , V ( Q ) ∖ A ′ ) c = E(A', V(Q) \setminus A') c = E ( A ′ , V ( Q ) ∖ A ′ ) . Then, the separation
s : = ( ⋃ ( X , Y , Z ) ∈ S X ∩ Y ∩ Z , ⋃ ( X , Y , Z ) ∉ S X ∩ Y ∩ Z ) s := \left(\bigcup_{(X,Y,Z)\in S} X \cap Y \cap Z,\ \bigcup_{(X,Y,Z)\notin S} X \cap Y \cap Z\right) s := ( X , Y , Z ) ∈ S ⋃ X ∩ Y ∩ Z , ( X , Y , Z ) ∈ / S ⋃ X ∩ Y ∩ Z has order λ ( c ) \lambda(c) λ ( c ) .
Proof. Let v ∈ V ( G ) v \in V(G) v ∈ V ( G ) . We have
v contributes to λ ( c ) ⟺ the face φ ( v ) of Q intersects both A ′ and V ( Q ) ∖ A ′ ⟺ there exist ( X 1 , Y 1 , Z 1 ) ∈ S such that v ∈ X 1 ∩ Y 1 ∩ Z 1 , and ( X 2 , Y 2 , Z 2 ) ∉ S such that v ∈ X 2 ∩ Y 2 ∩ Z 2 ⟺ v ∈ sep ( s ) . \begin{aligned}
v \text{ contributes to } \lambda(c) &\Longleftrightarrow\text{ the face } \varphi(v) \text{ of } Q \text{ intersects both } A' \text{ and } V(Q) \setminus A' \\
&\Longleftrightarrow\text{ there exist } (X_1,Y_1,Z_1) \in S \text{ such that } v \in X_1 \cap Y_1 \cap Z_1,\\
&\qquad\text{and } (X_2,Y_2,Z_2) \notin S \text{ such that } v \in X_2 \cap Y_2 \cap Z_2\\
&\Longleftrightarrow v \in\operatorname{sep}(s).
\end{aligned} v contributes to λ ( c ) ⟺ the face φ ( v ) of Q intersects both A ′ and V ( Q ) ∖ A ′ ⟺ there exist ( X 1 , Y 1 , Z 1 ) ∈ S such that v ∈ X 1 ∩ Y 1 ∩ Z 1 , and ( X 2 , Y 2 , Z 2 ) ∈ / S such that v ∈ X 2 ∩ Y 2 ∩ Z 2 ⟺ v ∈ sep ( s ) .
Furthermore, if v v v contributes to λ ( c ) \lambda(c) λ ( c ) then v v v contributes by exactly one. This concludes the proof. □
We now introduce the terminology that we will use to talk about the various zones in the Venn diagram of three separations. Let r : = ( A , B ) r := (A,B) r := ( A , B ) , s : = ( C , D ) s := (C,D) s := ( C , D ) and t : = ( E , F ) t := (E,F) t := ( E , F ) be three separations. We define the ACE-corner to be the set φ − 1 ( A C E ) = ( A ∖ B ) ∩ ( C ∖ D ) ∩ ( E ∖ F ) \varphi^{-1}(ACE) = (A \setminus B) \cap(C \setminus D) \cap(E \setminus F) φ − 1 ( A C E ) = ( A ∖ B ) ∩ ( C ∖ D ) ∩ ( E ∖ F ) , and define the 7 other corners analogously. The corners correspond to vertices of Q Q Q . We define the rCE(-edgy)-link to be the set φ − 1 ( r C E ) : = ( A ∩ B ) ∩ ( C ∖ D ) ∩ ( E ∖ F ) \varphi^{-1}(rCE) := (A \cap B) \cap(C \setminus D) \cap(E \setminus F) φ − 1 ( r C E ) := ( A ∩ B ) ∩ ( C ∖ D ) ∩ ( E ∖ F ) , and define the 11 other edgy links analogously. The edgy links correspond to edges of Q Q Q . For instance, the rCE-link corresponds to the edge A B C E ABCE A B C E . We define the rsE(-facial)-link to be the set φ − 1 ( r s E ) : = ( A ∩ B ) ∩ ( C ∩ D ) ∩ ( E ∖ F ) \varphi^{-1}(rsE) := (A \cap B) \cap(C \cap D) \cap(E \setminus F) φ − 1 ( r s E ) := ( A ∩ B ) ∩ ( C ∩ D ) ∩ ( E ∖ F ) , and define the other 5 facial links analogously. The facial links correspond to facets of Q Q Q . For instance, the rsE-link corresponds to the face labelled A B C D E ABCDE A B C D E . We define the rst-centre to be the set φ − 1 ( r s t ) : = ( A ∩ B ) ∩ ( C ∩ D ) ∩ ( E ∩ F ) \varphi^{-1}(rst) := (A \cap B) \cap(C \cap D) \cap(E \cap F) φ − 1 ( r s t ) := ( A ∩ B ) ∩ ( C ∩ D ) ∩ ( E ∩ F ) . The rst-centre corresponds to the entire cube Q Q Q .
A 3D corner separation is any separation of the form ( A ∩ C ∩ E , B ∪ D ∪ F ) (A \cap C \cap E, B \cup D \cup F) ( A ∩ C ∩ E , B ∪ D ∪ F ) , up to exchanging A A A and B B B , and/or C C C and D D D , and/or E E E and F F F . For simplicity, we call the separation ( A ∩ C ∩ E , B ∪ D ∪ F ) (A \cap C \cap E, B \cup D \cup F) ( A ∩ C ∩ E , B ∪ D ∪ F ) the 3D corner separation ACE , and similarly for the other 3D corner separations.
Proof of the 3D Corner Lemma
We now prove a result about how three well-behaved separations can cross in a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. To state this result, we first need some definitions.
Let G G G be a graph. Assume that two separations s = ( A , B ) s = (A,B) s = ( A , B ) and t = ( C , D ) t = (C,D) t = ( C , D ) of G G G cross. Let P P P denote the s C sC s C -link, let X X X denote the s t st s t -centre, and let Q Q Q denote the s D sD sD -link. We say that ( C , D ) (C,D) ( C , D ) induces ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) on A ∩ B A \cap B A ∩ B . By a slight abuse of notation, we call ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) a tripartition of A ∩ B A \cap B A ∩ B .
Let S S S be a set of separations of G G G , and let γ ∈ N \gamma\in\mathbb{N} γ ∈ N . We say that S S S is uniformly γ \gamma γ -crossing if for each ( A , B ) ∈ S (A,B) \in S ( A , B ) ∈ S there is a tripartition ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) of A ∩ B A \cap B A ∩ B such that, whenever a ( C , D ) ∈ S (C,D) \in S ( C , D ) ∈ S γ \gamma γ -crosses ( A , B ) (A,B) ( A , B ) , then ( C , D ) (C,D) ( C , D ) induces the tripartition ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) on A ∩ B A \cap B A ∩ B . We abbreviate uniformly 0-crossing to uniformly crossing .
If two k k k -separations cross, they cross equitably if all four corner separators have size exactly k k k . A set S S S of k k k -separations is equitable if whenever two elements of S S S cross, they cross equitably. Given ϵ ∈ N \epsilon\in\mathbb{N} ϵ ∈ N , we say that two separations ϵ \epsilon ϵ -cross if their four corners all have size more than ϵ \epsilon ϵ . A set S S S of separations is ϵ \epsilon ϵ -cross-compatible if whenever two separations in S S S cross, they ϵ \epsilon ϵ -cross.
The central statement of this section is the following.
Key Lemma 4.14 (3D Corner Lemma). Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let S S S be an equitable ( 2 q + k ) (2q+k) ( 2 q + k ) -cross-compatible triple of k k k -separations of G G G . Then, S S S is uniformly crossing.
We can prove that the same result holds under the weaker assumption that no two separations in S S S cross with two opposite corners of size at most 2 q + k 2q+k 2 q + k , but this requires some extra work and we do not need this stronger result. We note that in the statement of Key Lemma 4.14, we can drop the assumption of S S S being equitable (by Lemma 5.3).
We start with some preliminary results.
Lemma 4.15. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let S S S be an equitable ( 2 q + k ) (2q+k) ( 2 q + k ) -criss-compatible triple of q q q -broad k k k -separations of G G G . Let λ \lambda λ be the visualisation map of S S S and let R R R be the set of vertices X Y Z XYZ X Y Z of Q Q Q such that ∣ X ∩ Y ∩ Z ∣ ≥ q + k |X \cap Y \cap Z| \ge q+k ∣ X ∩ Y ∩ Z ∣ ≥ q + k . Suppose that the separations in S S S pairwise cross. Then, R R R is a vertex cover of Q Q Q .
Proof. By contradiction, suppose that some edge of Q Q Q is not covered by R R R . Without loss of generality, suppose that this edge is the edge ( A C E , A C F ) (ACE, ACF) ( A C E , A C F ) , see Figure 4.8 . Therefore, ∣ A ∩ C ∣ ≤ ∣ A ∩ C ∩ E ∣ + ∣ A ∩ C ∩ F ∣ ≤ 2 ( q + k ) |A \cap C| \le|A \cap C \cap E| + |A \cap C \cap F| \le2(q+k) ∣ A ∩ C ∣ ≤ ∣ A ∩ C ∩ E ∣ + ∣ A ∩ C ∩ F ∣ ≤ 2 ( q + k ) . Since S S S is equitable and ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) cross, the 2D corner separation ( A ∩ C , B ∪ D ) (A \cap C, B \cup D) ( A ∩ C , B ∪ D ) has order exactly k k k , so the A C AC A C -corner has size at most 2 q + k 2q+k 2 q + k , a contradiction. □ \square □
Lemma 4.16. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let S S S be a triple of k k k -separations of G G G . Let λ \lambda λ be the visualisation map of S S S and let R R R be the set of vertices X Y Z XYZ X Y Z of Q Q Q such that ∣ X ∩ Y ∩ Z ∣ ≥ q + k |X \cap Y \cap Z| \ge q+k ∣ X ∩ Y ∩ Z ∣ ≥ q + k . Then, every cut of Q Q Q with one vertex of R R R on each side has λ \lambda λ -value at least k k k .
Proof. Let c c c be a cut of Q Q Q with one vertex of R R R on each side. Then, c c c lifts to a separation s s s of G G G as in Lemma 4.13. By definition of R R R and since there is a vertex of R R R on each side of c c c , both sides of s s s have size at least q + k q+k q + k . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, s s s has order at least k k k . By Lemma 4.13, λ ( c ) \lambda(c) λ ( c ) is the order of this separation, so λ ( c ) ≥ k \lambda(c) \ge k λ ( c ) ≥ k . □ \square □
Lemma 4.17. Let G G G be a graph and let S S S be a triple of k k k -separations of G G G . Let λ \lambda λ be the visualisation map of S S S . If λ \lambda λ is a flower then S S S is uniformly crossing.
Proof. Write S = { ( A , B ) , ( C , D ) , ( E , F ) } S = \{(A,B),(C,D),(E,F)\} S = {( A , B ) , ( C , D ) , ( E , F )} and let r = ( A , B ) r=(A,B) r = ( A , B ) , s = ( C , D ) s=(C,D) s = ( C , D ) and t = ( E , F ) t=(E,F) t = ( E , F ) . Since λ \lambda λ is a flower, without loss of generality, we can assume that all the 1D faces incident to A C E ACE A C E or B D F BDF B D F have λ \lambda λ -value zero, that all other 1D faces have the same λ \lambda λ -value ℓ ≤ k / 2 \ell\le k/2 ℓ ≤ k /2 , that all the 2D faces have λ \lambda λ -value zero and that the 3D face has λ \lambda λ -value k − 2 ℓ k-2\ell k − 2 ℓ , see Figure 4.9 . By definition of λ \lambda λ , this means that the r s E rsE r s E -, r s F rsF r s F -, r C t rCt r C t -, r D t rDt r D t -, A s t Ast A s t - and B s t Bst B s t -links are empty, as well as the A s E AsE A s E -, A C t ACt A C t -, r C E rCE r C E -, r D F rDF r D F -, B s F BsF B s F - and B D t BDt B D t -links.
Figure 4.9. Illustration of the proof of Lemma 4.17 . Every 2D face has λ \lambda λ -value 0, as well as all blue 1D faces. All pairwise centres are then equal to the r s t rst r s t -centre. Furthermore, all pairwise links are equal to some non-blue edge.
Whenever two distinct separations in S S S cross, their centre is the union of the r s t rst r s t -centre and two facial links. However, we saw that all facial links are empty. Therefore, all three pairwise centres are equal to the r s t rst r s t -centre.
Similarly, for the crossing separations r = ( A , B ) r=(A,B) r = ( A , B ) and s = ( C , D ) s=(C,D) s = ( C , D ) , the r C rC r C -link is the union of the r C E rCE r C E -link, the r C F rCF r C F -link and the r C t rCt r C t -link, which is simply the r C F rCF r C F -link, and the r D rD r D -link is the union of the r D E rDE r D E -link, the r D F rDF r D F -link and the r D t rDt r D t -link, which is simply the r D E rDE r D E -link. For the crossing separations r = ( A , B ) r=(A,B) r = ( A , B ) and t = ( E , F ) t=(E,F) t = ( E , F ) , the r F rF r F -link is the union of the r C F rCF r C F -link, the r D F rDF r D F -link and the r s F rsF r s F -link, which is simply the r C F rCF r C F -link, and the r E rE r E -link is the union of the r C E rCE r C E -link, the r D E rDE r D E -link and the r s E rsE r s E -link, which is simply the r D E rDE r D E -link. This proves that the separations s s s and t t t uniformly cross the separation r r r . Similar arguments conclude the proof that S S S is uniformly crossing. □
Lemma 4.18. Let G G G be a graph and let S = { r , s , t } S = \{r,s,t\} S = { r , s , t } be an equitable set of k k k -separations of G G G . Assume that r , s r,s r , s cross, r , t r,t r , t cross and s ≤ t s \le t s ≤ t . Then, S S S is uniformly crossing.
Proof. Since the separations s s s and t t t are nested, it suffices to prove that the crossing separations r , s r,s r , s , and r , t r,t r , t induce the same tri-partition of the separator of r r r into one centre and two links.
Let r = : ( A , B ) r =: (A,B) r =: ( A , B ) , s = : ( C , D ) s =: (C,D) s =: ( C , D ) and t = : ( E , F ) t =: (E,F) t =: ( E , F ) . Let λ \lambda λ be the visualisation of S S S . By Lemma 4.12, λ \lambda λ is a k k k -corner-map. The vertices A C F ACF A C F and B C F BCF B C F are adjacent in Q Q Q (see Figure 4.8 ). Since C ⊆ E C \subseteq E C ⊆ E and F ⊆ D F \subseteq D F ⊆ D , we have that the r C t rCt r C t -link, the r s F rsF r s F -link and the r C F rCF r C F -link are empty. Thus, by definition of λ \lambda λ , the 1D and 2D faces which contain A C F ACF A C F and B C F BCF B C F have λ \lambda λ -value 0.
Since S S S is equitable, and ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) cross, and ( A , B ) (A,B) ( A , B ) and ( E , F ) (E,F) ( E , F ) cross, we have that the A C AC A C -, A F AF A F -, B C BC B C - and B F BF B F -corner separations have order exactly k k k . Thus, by Lemma 4.13, all pair cuts which contain exactly one of A C F ACF A C F , B C F BCF B C F have λ \lambda λ -value k k k . Therefore, by Lemma 4.6 , all 1D and 2D faces which contain A D E ADE A D E and B D E BDE B D E have λ \lambda λ -value 0. By definition of λ \lambda λ , this means that the r D t rDt r D t -link, the r s E rsE r s E -link and the r D E rDE r D E -link are also empty. See Figure 4.5 for an illustration of the setting.
Since the r C t rCt r C t -link, the r D t rDt r D t -link, the r s E rsE r s E -link and the r s F rsF r s F -link are empty, the two pairs of crossing separations ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) , and ( A , B ) (A,B) ( A , B ) and ( E , F ) (E,F) ( E , F ) have the same centre, namely the r s t rst r s t -centre.
Furthermore, since the r C t rCt r C t -link, the r s E rsE r s E -link, the r C F rCF r C F -link and the r D E rDE r D E -link are empty, it follows that r C E rCE r C E is both the r C rC r C -link of the crossing separations ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) , and the r E rE r E -link of the crossing separations ( A , B ) (A,B) ( A , B ) and ( E , F ) (E,F) ( E , F ) . Similarly, r D F rDF r D F is both the r D rD r D -link of the crossing separations ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) , and the r F rF r F -link of the crossing separations ( A , B ) (A,B) ( A , B ) and ( E , F ) (E,F) ( E , F ) . This concludes the proof. □
Combining these results, we can now prove Key Lemma 4.14, which we restate for convenience.
Key Lemma 4.14 (3D Corner Lemma). Let G G G be a ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable graph and let S S S be an equitable ( 2 q + k ) (2q + k) ( 2 q + k ) -criss-compatible triple of k k k -separations of G G G . Then, S S S is uniformly crossing.
Proof of Key Lemma 4.14. Write S = { ( A , B ) , ( C , D ) , ( E , F ) } S = \{(A, B), (C, D), (E, F)\} S = {( A , B ) , ( C , D ) , ( E , F )} . Let λ \lambda λ be the visualisation map of S S S . By Lemma 4.12, λ \lambda λ is a k k k -corner-map. If no separation in S S S crosses the other two, then we are done. Hence assume that one of them, say ( A , B ) (A, B) ( A , B ) , crosses the other two. Since S S S is ( 2 q + k ) (2q+k) ( 2 q + k ) -criss-compatible, it follows that all separations in S S S have breadth ≥ 4 q + 2 k ≥ q \ge4q + 2k \ge q ≥ 4 q + 2 k ≥ q .
Case 1: ( C , D ) (C, D) ( C , D ) and ( E , F ) (E, F) ( E , F ) cross. Then the three separations ( A , B ) (A, B) ( A , B ) , ( C , D ) (C, D) ( C , D ) and ( E , F ) (E, F) ( E , F ) are pairwise crossing. Let R R R be the set of vertices X Y Z XYZ X Y Z of Q Q Q such that ∣ X ∩ Y ∩ Z ∣ ≥ q + k \lvert X \cap Y \cap Z\rvert\ge q + k ∣ X ∩ Y ∩ Z ∣ ≥ q + k . Then, R R R is a vertex cover of Q Q Q by Lemma 4.15 and every cut of Q Q Q with one vertex of R R R on each side has λ \lambda λ -value at least k k k by Lemma 4.16 . Thus, λ \lambda λ is a flower by Lemma 4.1. Therefore, S S S is uniformly crossing by Lemma 4.17 .
Case 2: ( C , D ) (C, D) ( C , D ) and ( E , F ) (E, F) ( E , F ) are nested. Without loss of generality, suppose that C ⊆ E C \subseteq E C ⊆ E and F ⊆ D F \subseteq D F ⊆ D . Then, by Lemma 4.18 , S S S is uniformly crossing. □
Broad separations are uniformly crossing
We now use the 3D Corner Lemma to show that sufficiently strong crossings impose consistent centres and links. This consistency will allow us to organise crossing separations into asteriskal sets of separations in the next section.
The main goal of this section is to prove the following statement.
Key Lemma 5.1. Let G G G be a ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable graph, and define γ 5.1 ( q , k ) : = 3 q + 3 k \gamma_{5.1}(q, k) := 3q + 3k γ 5.1 ( q , k ) := 3 q + 3 k . Then the set of all q q q -broad k k k -separations of G G G is uniformly γ 5.1 \gamma_{5.1} γ 5.1 -crossing.
We will also prove some additional statements which will be useful in later sections.
Convention 5.2. Often we will define bounds of the form ε X ( q , k ) \varepsilon_X(q, k) ε X ( q , k ) where q , k ∈ N q, k \in\mathbb{N} q , k ∈ N determine the degree of ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakability and X X X is the number of a lemma or whatever that introduces the bound. When we use these bounds for some ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable graph, we sometimes drop ‘( q , k ) (q, k) ( q , k ) ’ and just write ε X \varepsilon_X ε X to reduce notational clutter, provided that q q q and k k k are clear from context.
Uniformly crossing sets of separations
Recall that a separation ( A , B ) (A, B) ( A , B ) of a graph G G G is β \beta β -broad if ∣ A ∖ B ∣ ≥ β \lvert A \setminus B\rvert\ge\beta ∣ A ∖ B ∣ ≥ β and ∣ B ∖ A ∣ ≥ β \lvert B \setminus A\rvert\ge\beta ∣ B ∖ A ∣ ≥ β . Moreover, recall that two separations s s s and t t t ε \varepsilon ε -criss if all four corners have size > ε > \varepsilon > ε .
Lemma 5.3. Let G G G be a ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable graph, and let s , s ′ s, s' s , s ′ be two q q q -crissing k k k -separations of G G G . Then s s s and s ′ s' s ′ cross equitably.
Proof. Assume s s s and s ′ s' s ′ do not cross equitably. Then, one of their corner separations has order less than k k k . By ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakability of G G G , the corresponding corner contains at most q q q vertices. So, s s s and s ′ s' s ′ do not q q q -criss. □
Observation 5.4. Let s = ( A , B ) s = (A, B) s = ( A , B ) and t = ( C , D ) t = (C, D) t = ( C , D ) be two ( ≤ k ) (\le k) ( ≤ k ) -separations in a graph G G G , and let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N .
If the A D AD A D -corner has size ≤ ε − ℓ \le\varepsilon-\ell ≤ ε − ℓ and both the A t At A t -link and s D sD sD -link have size ≤ ℓ \le\ell ≤ ℓ , then s ≤ ε t s\le_\varepsilon t s ≤ ε t .
If s s s and t t t ε \varepsilon ε -cross, then they ( ε − k ) (\varepsilon-k) ( ε − k ) -criss.
If s s s and t t t ε \varepsilon ε -cross and all links have size ≤ ℓ \le\ell ≤ ℓ , then s s s and t t t ( ε − ℓ ) (\varepsilon-\ell) ( ε − ℓ ) -criss.
If s s s and t t t ε \varepsilon ε -criss with ε > 0 \varepsilon>0 ε > 0 , then they also ε \varepsilon ε -cross.
If s s s and t t t ε \varepsilon ε -cross then ∣ A ∩ C ∣ , ∣ A ∩ D ∣ , ∣ B ∩ C ∣ , ∣ B ∩ D ∣ > ε |A\cap C|,|A\cap D|,|B\cap C|,|B\cap D|>\varepsilon ∣ A ∩ C ∣ , ∣ A ∩ D ∣ , ∣ B ∩ C ∣ , ∣ B ∩ D ∣ > ε .
Proof. {(citep)}(i). We have ∣ A ∖ C ∣ ≤ ∣ A ∖ ( B ∪ C ) ∣ + ∣ ( A ∩ B ) ∖ C ∣ ≤ ( ε − ℓ ) + ℓ = ε |A\setminus C|\le|A\setminus(B\cup C)|+|(A\cap B)\setminus C|\le(\varepsilon-\ell)+\ell=\varepsilon ∣ A ∖ C ∣ ≤ ∣ A ∖ ( B ∪ C ) ∣ + ∣ ( A ∩ B ) ∖ C ∣ ≤ ( ε − ℓ ) + ℓ = ε . Similarly, ∣ D ∖ B ∣ ≤ ε |D\setminus B|\le\varepsilon ∣ D ∖ B ∣ ≤ ε . It follows that A ⊆ ε C A\subseteq_\varepsilon C A ⊆ ε C and D ⊆ ε B D\subseteq_\varepsilon B D ⊆ ε B . Hence ( A , B ) ≤ ε ( C , D ) (A,B)\le_\varepsilon(C,D) ( A , B ) ≤ ε ( C , D ) .
{(citep)}(iii). If s = : ( A , B ) s=:(A,B) s =: ( A , B ) and t = : ( C , D ) t=:(C,D) t =: ( C , D ) do not ( ε − ℓ ) (\varepsilon-\ell) ( ε − ℓ ) -criss, then there is a corner of size ≤ ε − ℓ \le\varepsilon-\ell ≤ ε − ℓ , say the A D AD A D -corner. By {(citep)}(i), s ≤ ε t s\le_\varepsilon t s ≤ ε t , and they are ε \varepsilon ε -nested.
{(citep)}(ii) follows from {(citep)}(iii) with ℓ : = k \ell:=k ℓ := k .
{(citep)}(iv). We show the contrapositive. Let s = : ( A , B ) s=:(A,B) s =: ( A , B ) and t = : ( C , D ) t=:(C,D) t =: ( C , D ) be ε \varepsilon ε -nested, say ( A , B ) ≤ ε ( C , D ) (A,B)\le_\varepsilon(C,D) ( A , B ) ≤ ε ( C , D ) . Then ∣ ( A ∩ D ) ∖ ( B ∪ C ) ∣ ≤ ∣ A ∖ C ∣ ≤ ε |(A\cap D)\setminus(B\cup C)|\le|A\setminus C|\le\varepsilon ∣ ( A ∩ D ) ∖ ( B ∪ C ) ∣ ≤ ∣ A ∖ C ∣ ≤ ε . That is, ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) do not ε \varepsilon ε -criss.
{(citep)}(v). If ∣ A ∩ C ∣ ≤ ε |A\cap C|\le\varepsilon ∣ A ∩ C ∣ ≤ ε , say, then ∣ A ∖ D ∣ ≤ ε |A\setminus D|\le\varepsilon ∣ A ∖ D ∣ ≤ ε and ∣ C ∖ B ∣ ≤ ε |C\setminus B|\le\varepsilon ∣ C ∖ B ∣ ≤ ε . Then ( A , B ) ≤ ε ( D , C ) (A,B)\le_\varepsilon(D,C) ( A , B ) ≤ ε ( D , C ) , in particular they do not ε \varepsilon ε -cross. □
Lemma 5.5
Let s , t s,t s , t be two separations of a graph G G G with s ≤ γ t s\le_\gamma t s ≤ γ t . Then s ∧ t ≈ γ s s\wedge t\approx_\gamma s s ∧ t ≈ γ s and t ≈ γ s ∨ t t\approx_\gamma s\vee t t ≈ γ s ∨ t .
Proof. Write s = : ( A , B ) s=:(A,B) s =: ( A , B ) and t = : ( C , D ) t=:(C,D) t =: ( C , D ) . Then s ≤ γ t s\le_\gamma t s ≤ γ t implies ∣ A ∖ C ∣ ≤ γ |A\setminus C|\le\gamma ∣ A ∖ C ∣ ≤ γ and ∣ D ∖ B ∣ ≤ γ |D\setminus B|\le\gamma ∣ D ∖ B ∣ ≤ γ . It follows that ∣ A ∖ ( A ∩ C ) ∣ ≤ γ |A\setminus(A\cap C)|\le\gamma ∣ A ∖ ( A ∩ C ) ∣ ≤ γ and ∣ ( B ∪ D ) ∖ B ∣ ≤ γ |(B\cup D)\setminus B|\le\gamma ∣ ( B ∪ D ) ∖ B ∣ ≤ γ . Therefore, A ⊆ γ A ∩ C ⊆ A A\subseteq_\gamma A\cap C\subseteq A A ⊆ γ A ∩ C ⊆ A and B ⊇ γ B ∪ D ⊇ B B\supseteq_\gamma B\cup D\supseteq B B ⊇ γ B ∪ D ⊇ B . Hence, s ∧ t = ( A ∩ C , B ∪ D ) ≈ γ ( A , B ) = s s\wedge t=(A\cap C,B\cup D)\approx_\gamma(A,B)=s s ∧ t = ( A ∩ C , B ∪ D ) ≈ γ ( A , B ) = s .
Similarly, s ≤ γ t s\le_\gamma t s ≤ γ t implies t ∗ ≤ γ s ∗ t^*\le_\gamma s^* t ∗ ≤ γ s ∗ . By the first part, we have ( s ∨ t ) ∗ = s ∗ ∧ t ∗ ≈ γ t ∗ (s\vee t)^*=s^*\wedge t^*\approx_\gamma t^* ( s ∨ t ) ∗ = s ∗ ∧ t ∗ ≈ γ t ∗ . This implies t ≈ γ s ∨ t t\approx_\gamma s\vee t t ≈ γ s ∨ t . □
Lemma 5.6
If two ( ≤ k ) (\le k) ( ≤ k ) -separations of a graph G G G ε \varepsilon ε -cross, then both are 2 ( ε − k ) 2(\varepsilon-k) 2 ( ε − k ) -broad.
Proof. Let ( A , B ) (A,B) ( A , B ) and ( C , D ) (C,D) ( C , D ) be ( ≤ k ) (\le k) ( ≤ k ) -separations of G G G that ε \varepsilon ε -cross. Then every corner has size ≥ ε − k \ge\varepsilon-k ≥ ε − k by Observation 5.4, so ∣ A ∖ B ∣ , ∣ B ∖ A ∣ ≥ 2 ( ε − k ) |A\setminus B|,|B\setminus A|\ge2(\varepsilon-k) ∣ A ∖ B ∣ , ∣ B ∖ A ∣ ≥ 2 ( ε − k ) . □
Lemma 5.7
Let γ , ℓ ∈ N \gamma,\ell\in\mathbb{N} γ , ℓ ∈ N . Let G G G be a graph and let s , t s,t s , t be two ( ℓ + γ ) (\ell+\gamma) ( ℓ + γ ) -broad k k k -separations of G G G with s ≤ γ t s\le_\gamma t s ≤ γ t . Then both s ∧ t s\wedge t s ∧ t and s ∨ t s\vee t s ∨ t are ℓ \ell ℓ -broad. Moreover, if G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and ℓ > q \ell>q ℓ > q then both s ∧ t s\wedge t s ∧ t and s ∨ t s\vee t s ∨ t are ℓ \ell ℓ -broad k k k -separations of G G G .
Proof. Let s = : ( A , B ) s=:(A,B) s =: ( A , B ) , t = : ( C , D ) t=:(C,D) t =: ( C , D ) , s ∧ t = : ( X 1 , Y 1 ) = : r 1 s\wedge t=:(X_1,Y_1)=:r_1 s ∧ t =: ( X 1 , Y 1 ) =: r 1 and s ∨ t = : ( X 2 , Y 2 ) = : r 2 s\vee t=:(X_2,Y_2)=:r_2 s ∨ t =: ( X 2 , Y 2 ) =: r 2 . Since s ≤ γ t s\le_\gamma t s ≤ γ t , we have ∣ D ∖ B ∣ ≤ γ |D\setminus B|\le\gamma ∣ D ∖ B ∣ ≤ γ . So,
∣ Y 1 ∖ X 1 ∣ = ∣ ( B ∪ D ) ∖ ( A ∩ C ) ∣ ≥ ∣ B ∖ A ∣ ≥ ℓ + γ ≥ ℓ , |Y_1\setminus X_1|=|(B\cup D)\setminus(A\cap C)|\ge|B\setminus A|\ge\ell+\gamma\ge\ell, ∣ Y 1 ∖ X 1 ∣ = ∣ ( B ∪ D ) ∖ ( A ∩ C ) ∣ ≥ ∣ B ∖ A ∣ ≥ ℓ + γ ≥ ℓ ,
and
∣ X 1 ∖ Y 1 ∣ = ∣ ( A ∩ C ) ∖ ( B ∪ D ) ∣ = ∣ A ∖ B ∣ − ∣ ( A ∩ D ) ∖ B ∣ ≥ ∣ A ∖ B ∣ − ∣ D ∖ B ∣ ≥ ( ℓ + γ ) − γ = ℓ . |X_1\setminus Y_1|=|(A\cap C)\setminus(B\cup D)|=|A\setminus B|-|(A\cap D)\setminus B|\ge|A\setminus B|-|D\setminus B|\ge(\ell+\gamma)-\gamma=\ell. ∣ X 1 ∖ Y 1 ∣ = ∣ ( A ∩ C ) ∖ ( B ∪ D ) ∣ = ∣ A ∖ B ∣ − ∣ ( A ∩ D ) ∖ B ∣ ≥ ∣ A ∖ B ∣ − ∣ D ∖ B ∣ ≥ ( ℓ + γ ) − γ = ℓ .
Similarly, ∣ X 2 ∖ Y 2 ∣ ≥ ℓ |X_2\setminus Y_2|\ge\ell ∣ X 2 ∖ Y 2 ∣ ≥ ℓ and ∣ Y 2 ∖ X 2 ∣ ≥ ℓ |Y_2\setminus X_2|\ge\ell ∣ Y 2 ∖ X 2 ∣ ≥ ℓ . So, both r 1 r_1 r 1 and r 2 r_2 r 2 are ℓ \ell ℓ -broad.
Moreover, assume that G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable. Then, from ℓ > q \ell>q ℓ > q it follows that both r 1 r_1 r 1 and r 2 r_2 r 2 have order ≥ k \ge k ≥ k . From submodularity, it follows that the order of both r 1 r_1 r 1 and r 2 r_2 r 2 is exactly k k k . □
Lemma 5.8. Let G G G be a graph. Let s s s , s ′ s' s ′ , t t t be separations of G G G . Let γ , ε ∈ N \gamma, \varepsilon\in\mathbb{N} γ , ε ∈ N with γ ≥ ε \gamma\ge\varepsilon γ ≥ ε . Assume that s s s and s ′ s' s ′ γ \gamma γ -cross while s s s and t t t are ε \varepsilon ε -close. Then t t t and s ′ s' s ′ ( γ − ε ) (\gamma-\varepsilon) ( γ − ε ) -cross.
Proof. Since s s s and t t t are ε \varepsilon ε -close, we may assume without loss of generality that s ≤ ε t ≤ ε s s \le_\varepsilon t \le_\varepsilon s s ≤ ε t ≤ ε s . Assume for a contradiction that t t t and s ′ s' s ′ fail to ( γ − ε ) (\gamma-\varepsilon) ( γ − ε ) -cross. Then t ≤ γ − ε s ′ t \le_{\gamma-\varepsilon} s' t ≤ γ − ε s ′ , say. With s ≤ ε t s \le_\varepsilon t s ≤ ε t and Lemma 4.9 , we find that s ≤ γ s ′ s \le_\gamma s' s ≤ γ s ′ , contradicting that s s s and s ′ s' s ′ γ \gamma γ -cross. □
Lemma 5.9. Let s s s and t t t be separations of a graph G G G . Assume that s s s has order ≤ k \le k ≤ k , breadth β \beta β , and that s s s is ε \varepsilon ε -close to t t t . Then t t t has breadth β ′ \beta' β ′ with β − ε ≤ β ′ ≤ β + ε + k \beta-\varepsilon\le\beta' \le\beta+\varepsilon+k β − ε ≤ β ′ ≤ β + ε + k .
Proof. Let s = : ( A , B ) s =: (A, B) s =: ( A , B ) and t = : ( C , D ) t =: (C, D) t =: ( C , D ) . Without loss of generality, ∣ A ∖ B ∣ = β |A \setminus B| = \beta ∣ A ∖ B ∣ = β and ∣ B ∖ A ∣ ≥ β |B \setminus A| \ge\beta ∣ B ∖ A ∣ ≥ β . First, we show β ′ ≤ β + ε + k \beta' \le\beta+\varepsilon+k β ′ ≤ β + ε + k . Since s s s is ε \varepsilon ε -close to t t t , we have s ≈ ε t s \approx_\varepsilon t s ≈ ε t . Hence ∣ C ∖ A ∣ ≤ ε |C \setminus A| \le\varepsilon ∣ C ∖ A ∣ ≤ ε . From
C ∖ D ⊆ C ⊆ ( C ∖ A ) ∪ A ⊆ ( C ∖ A ) ∪ ( A ∩ B ) ∪ ( A ∖ B ) C \setminus D \subseteq C \subseteq(C \setminus A) \cup A \subseteq(C \setminus A) \cup(A \cap B) \cup(A \setminus B) C ∖ D ⊆ C ⊆ ( C ∖ A ) ∪ A ⊆ ( C ∖ A ) ∪ ( A ∩ B ) ∪ ( A ∖ B ) we obtain ∣ C ∖ D ∣ ≤ ε + k + β |C \setminus D| \le\varepsilon+k+\beta ∣ C ∖ D ∣ ≤ ε + k + β .
Now, we show β ′ ≥ β − ε \beta' \ge\beta-\varepsilon β ′ ≥ β − ε . From ∣ A ∖ B ∣ ≥ β |A \setminus B| \ge\beta ∣ A ∖ B ∣ ≥ β and ∣ D ∖ B ∣ ≤ ε |D \setminus B| \le\varepsilon ∣ D ∖ B ∣ ≤ ε , we obtain
∣ C ∖ D ∣ ≥ ∣ ( A ∖ B ) ∩ ( C ∖ D ) ∣ = ∣ A ∖ B ∣ − ∣ D ∖ B ∣ ≥ β − ε . |C \setminus D| \ge|(A \setminus B) \cap(C \setminus D)| = |A \setminus B| - |D \setminus B| \ge\beta-\varepsilon. ∣ C ∖ D ∣ ≥ ∣ ( A ∖ B ) ∩ ( C ∖ D ) ∣ = ∣ A ∖ B ∣ − ∣ D ∖ B ∣ ≥ β − ε . Similarly, from ∣ B ∖ A ∣ ≥ β |B \setminus A| \ge\beta ∣ B ∖ A ∣ ≥ β and ∣ C ∖ A ∣ ≤ ε |C \setminus A| \le\varepsilon ∣ C ∖ A ∣ ≤ ε we obtain ∣ D ∖ C ∣ ≥ β − ε |D \setminus C| \ge\beta-\varepsilon ∣ D ∖ C ∣ ≥ β − ε . Hence, t t t is ( β − ε ) (\beta-\varepsilon) ( β − ε ) -broad. □
A set S S S of separations of a graph G G G is ε \varepsilon ε -triple uniformly crossing if for every triple of elements of S S S , either the triple is uniformly crossing, or two separations do not ε \varepsilon ε -cross. The next result essentially says that to prove that a set of separations is uniformly γ \gamma γ -crossing, it suffices to prove that it is ε \varepsilon ε -triple uniformly crossing, for some ε < γ \varepsilon< \gamma ε < γ .
Lemma 5.10. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N and
γ ≥ γ 5.10 ( ε , q , k ) : = max { ε + k + q , 2 k } . \gamma\ge\gamma_{5.10}(\varepsilon,q,k) := \max\{\varepsilon+k+q, 2k\}. γ ≥ γ 5.10 ( ε , q , k ) := max { ε + k + q , 2 k } . Then, any ε \varepsilon ε -triple uniformly crossing set of k k k -separations of G G G is uniformly γ \gamma γ -crossing.
Proof. Let S S S be an ε \varepsilon ε -triple uniformly crossing set of k k k -separations of G G G , and consider an arbitrary separation r ∈ S r \in S r ∈ S . Let s , t ∈ S s,t \in S s , t ∈ S be two separations which γ \gamma γ -cross r r r . Our aim is to show that r r r is crossed by both of them with the same tri-partition of the separator sep ( r ) \operatorname{sep}(r) sep ( r ) into centre and links. Since S S S is ε \varepsilon ε -triple uniform, either the triple { r , s , t } \{r,s,t\} { r , s , t } is uniformly crossing, or two of its separations do not ε \varepsilon ε -cross. Since we are done in the first case, assume that two of its separations do not ε \varepsilon ε -cross. Since γ ≥ ε \gamma\ge\varepsilon γ ≥ ε and both s s s and t t t γ \gamma γ -cross r r r , we can assume that s s s and t t t are ε \varepsilon ε -nested, say s ≤ ε t s \le_\varepsilon t s ≤ ε t .
By Lemma 5.5, we have s ∧ t ≈ ε s s \wedge t \approx_\varepsilon s s ∧ t ≈ ε s . Since r r r γ \gamma γ -crosses s s s and t t t , all three separations are 2 ( γ − k ) 2(\gamma-k) 2 ( γ − k ) -broad by Lemma 5.6, with 2 ( γ − k ) ≥ γ 2(\gamma-k) \ge\gamma 2 ( γ − k ) ≥ γ (since γ ≥ 2 k \gamma\ge2k γ ≥ 2 k ). By Lemma 5.7 and since γ − ε > q \gamma-\varepsilon> q γ − ε > q , the separation s ∧ t s \wedge t s ∧ t is a ( γ − ε ) (\gamma-\varepsilon) ( γ − ε ) -broad separation of order exactly k k k .
By Lemma 5.8 , r r r and s ∧ t s \wedge t s ∧ t ( γ − ε ) (\gamma-\varepsilon) ( γ − ε ) -cross. By Observation 5.4, r r r and s ∧ t s \wedge t s ∧ t ( γ − ε − k ) (\gamma-\varepsilon-k) ( γ − ε − k ) -cross. By Lemma 5.3 and since γ − ε − k ≥ q \gamma-\varepsilon-k \ge q γ − ε − k ≥ q , the separations r r r and s ∧ t s \wedge t s ∧ t cross equitably.
Since s s s and t t t both ( γ − k ) (\gamma-k) ( γ − k ) -cross r r r (by Observation 5.4) with γ − k ≥ q \gamma-k \ge q γ − k ≥ q , by Lemma 5.3, the sets { r , s , s ∧ t } \{r,s,s \wedge t\} { r , s , s ∧ t } and { r , t , s ∧ t } \{r,t,s \wedge t\} { r , t , s ∧ t } are equitable. Since s s s and s ∧ t s \wedge t s ∧ t are nested, the triple { r , s , s ∧ t } \{r,s,s \wedge t\} { r , s , s ∧ t } is uniformly crossing by Lemma 4.18 . Similarly, { r , t , s ∧ t } \{r,t,s \wedge t\} { r , t , s ∧ t } is uniformly crossing.
Consider the tri-partition of the separator sep ( r ) \operatorname{sep}(r) sep ( r ) obtained when crossing r r r with s ∧ t s \wedge t s ∧ t . Since { r , s , s ∧ t } \{r, s, s \wedge t\} { r , s , s ∧ t } is uniformly crossing, this is the same as when crossing r r r with s s s , and since { r , t , s ∧ t } \{r, t, s \wedge t\} { r , t , s ∧ t } is uniformly crossing, this is the same as when crossing r r r with t t t . This completes the proof. □
Lemma 5.11. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, and define ε = ε ( q , k ) : = 2 q + 2 k \varepsilon= \varepsilon(q,k) := 2q + 2k ε = ε ( q , k ) := 2 q + 2 k . Then, the set of q q q -broad k k k -separations of G G G is ε \varepsilon ε -triple uniformly crossing.
Proof. Let S S S be a triple of q q q -broad k k k -separations of G G G . If two separations in S S S are ε \varepsilon ε -nested then we are done. Otherwise, every pair in S S S ε \varepsilon ε -crosses, hence ( ε − k ) (\varepsilon-k) ( ε − k ) -crisses by Observation 5.4 (with ε − k = 2 q + k \varepsilon-k=2q+k ε − k = 2 q + k ). Thus, S S S is equitable by Lemma 5.3. Then, S S S is uniformly crossing by Key Lemma 4.14. □
Proof of Key Lemma 5.1. This follows immediately from Lemma 5.10 and Lemma 5.11 . □
Empty links in 3D corner diagrams
We now show some additional results that will be used in § 8 .
Corollary 5.12. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let r = ( A , B ) r=(A,B) r = ( A , B ) , s = ( C , D ) s=(C,D) s = ( C , D ) and t = ( E , F ) t=(E,F) t = ( E , F ) be three k k k -separations of G G G such that the 3D corners A D F ADF A D F , B C F BCF B C F and B D E BDE B D E have size more than q q q . Then, the A s t Ast A s t -link, the r C t rCt r C t -link, the r s E rsE r s E -link, the A C t ACt A C t -link, the A s E AsE A s E -link and the r C E rCE r C E -link are empty.
Proof. Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, the 3D corner separations A D F ADF A D F , B C F BCF B C F and B D E BDE B D E have order at least k k k . Similarly, the 2D corner separations A D AD A D , A F AF A F , D F DF D F , B C BC B C , B F BF B F , C F CF C F , B D BD B D , B E BE B E and D E DE D E have order at least k k k . By submodularity, the 2D corner separations B C BC B C , B E BE B E , C E CE C E , A D AD A D , A E AE A E , D E DE D E , A C AC A C , A F AF A F and C F CF C F have order at most k k k . Let λ \lambda λ be the visualisation map of r r r , s s s and t t t . By Lemma 4.12, λ \lambda λ is a k k k -corner-map. By Lemma 4.13, the λ \lambda λ -values correspond to the order of the separations of G G G . See Figure 4.8 for an illustration of the setting. Applying Lemma 4.7 with u = A D F u=ADF u = A D F , v = B C F v=BCF v = B C F and w = A C F w=ACF w = A C F (which we can do since the A F AF A F and C F CF C F corner separations have order at most k k k ), we get that the A s t Ast A s t -link, the r C t rCt r C t -link and the A C t ACt A C t -link are empty. Applying Lemma 4.7 with u = A D F u=ADF u = A D F , v = B D E v=BDE v = B D E and w = A D E w=ADE w = A D E (which we can do since the A D AD A D and D E DE D E corner separations have order at most k k k ), we get that the A s t Ast A s t -link, the r s E rsE r s E -link and the A s E AsE A s E -link are empty. Finally, applying Lemma 4.7 with u = B C F u=BCF u = B C F , v = B D E v=BDE v = B D E and w = B C E w=BCE w = B C E (which we can do since the B C BC B C and B E BE B E corner separations have order at most k k k ), we get that the r C t rCt r C t -link, the r s E rsE r s E -link and the r C E rCE r C E -link are empty. □
Corollary 5.13. Let r = ( A , B ) r=(A,B) r = ( A , B ) , s = ( C , D ) s=(C,D) s = ( C , D ) and t = ( E , F ) t=(E,F) t = ( E , F ) be three k k k -separations of a graph, such that the 3D corner separations A D F ADF A D F , A C F ACF A C F , A C E ACE A C E and B C E BCE B C E have order at least k k k , and the 2D corner separations A C AC A C , A F AF A F and C E CE C E have order at most k k k . Then, the A s t Ast A s t -link, the r C t rCt r C t -link, the r s E rsE r s E -link, the r s F rsF r s F -link, the A s E AsE A s E -link and the r C F rCF r C F -link are empty.
Proof. Let λ \lambda λ be the visualisation map of r r r , s s s and t t t . By Lemma 4.12, λ \lambda λ is a k k k -corner-map. By Lemma 4.13, the λ \lambda λ -values correspond to the order of the separations of G G G . See Figure 4.8 for an illustration of the setting. Applying Lemma 4.7 with u = A C F u=ACF u = A C F , v = B C E v=BCE v = B C E and w = A C E w=ACE w = A C E (which we can do since the A C AC A C and C E CE C E corner separations have order at most k k k ), we get that the A s t Ast A s t -link,
the r s E rsE r s E -link and the A s E AsE A s E -link are empty. Similarly, applying Lemma 4.7 with u = A D F u=ADF u = A D F , v = A C E v=ACE v = A C E and w = A C F w=ACF w = A C F (which we can do since the A F AF A F and A C AC A C corner separations have order at most k k k ), we get that the r C t rCt r C t -link, the r s F rsF r s F -link and the r C F rCF r C F -link are empty. □
From asterisks to flowers via coverings
The purpose of this section is to prove Key Lemma 6.6. For this we need the following notation.
We say that a set S S S of separations is γ \gamma γ -crossing-connected if S ≠ ∅ S \ne\varnothing S = ∅ and the reflexive and transitive closure of the union of the relations ‘γ \gamma γ -crossing’ on S S S and { ( s , s ∗ ) : s ∈ S } \{(s,s^*):s\in S\} {( s , s ∗ ) : s ∈ S } has exactly one partition class. Equivalently, S S S is γ \gamma γ -crossing-connected if S ≠ ∅ S \ne\varnothing S = ∅ and for any two separations s , t ∈ S s,t\in S s , t ∈ S with s ≠ t ∗ s \ne t^* s = t ∗ there is a sequence s = s 0 , s 1 , … , s ℓ = t s=s_0,s_1,\ldots,s_\ell=t s = s 0 , s 1 , … , s ℓ = t of separations in S S S of some length ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N such that s i s_i s i γ \gamma γ -crosses s i − 1 s_{i-1} s i − 1 for every i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] . We call S S S γ \gamma γ -asteriskal if S S S is
uniformly γ \gamma γ -crossing,
γ \gamma γ -crossing-connected, and
symmetric;
see Figure 6.1 . A γ \gamma γ -asteriskal set S S S is trivial if it consists of a single separation and its inverse.
Figure 6.1. Each white chunk contains ≥ γ \ge\gamma ≥ γ vertices. The red, green and blue separations all include X X X in their separators. Together, they form a γ \gamma γ -asteriskal set of separations.
Assume now that S S S is uniformly γ \gamma γ -crossing and symmetric, but not necessarily γ \gamma γ -crossing-connected. Then S S S admits a unique partition into maximal γ \gamma γ -asteriskal subsets of S S S . We refer to these maximal subsets as the γ \gamma γ -asterisks of S S S .
Let ( H , V ) (H,\mathcal{V}) ( H , V ) be a graph-decomposition. For an edge e = t t ′ ∈ E ( H ) e=tt'\in E(H) e = t t ′ ∈ E ( H ) we write V e : = V t ∩ V t ′ V_e:=V_t\cap V_{t'} V e := V t ∩ V t ′ .
Definition 6.1 (Compatible). Let S = ( S , V ) \mathcal{S}=(S,\mathcal{V}) S = ( S , V ) be a graphic-k k k -anemone and let Z : = V c Z:=V_c Z := V c , where c c c is the centre of S \mathcal{S} S . A separation s = ( A , B ) s=(A,B) s = ( A , B ) is γ \gamma γ -compatible with S \mathcal{S} S if, whenever s s s γ \gamma γ -crosses a petal-separation t t t of S \mathcal{S} S then s s s and t t t cross with centre Z Z Z and four empty links.
Let D = ( O , ( V p ) p ∈ O ) \mathcal{D}=(O,(V_p)_{p\in O}) D = ( O , ( V p ) p ∈ O ) be a graphic-k k k -daisy of a graph G G G and let Z : = ⋂ p ∈ O V p Z:=\bigcap_{p\in O}V_p Z := ⋂ p ∈ O V p . A separation s = ( A , B ) s=(A,B) s = ( A , B ) is γ \gamma γ -compatible with D \mathcal{D} D if, whenever s s s γ \gamma γ -crosses a petal-separation t t t of D \mathcal{D} D , then all of the following assertions hold:
Z = V e ∩ V f Z = V_e \cap V_f Z = V e ∩ V f and Z Z Z is the s t st s t -centre, where e , f e, f e , f are the edges incident with the node of O O O corresponding to t t t ; and
V e ∖ Z V_e \setminus Z V e ∖ Z and V f ∖ Z V_f \setminus Z V f ∖ Z are the A t At A t -link and the B t Bt B t -link; and
all links of the two separations s , t s, t s , t are of the same size.
Intuitively, we require that whenever s s s γ \gamma γ -crosses a petal-separation t t t of the flower, this crossing happens in the way we expect it to, namely with the centre the “centre of the flower”, the t t t -links being V e ∖ V f V_e \setminus V_f V e ∖ V f and V f ∖ V e V_f \setminus V_e V f ∖ V e , and with all links of the same size.
Let D = ( O , V ) D = (O, \mathcal{V}) D = ( O , V ) be a graphic-k k k -daisy. Let e ≠ f e \ne f e = f be two edges of O O O , and let X , Y X, Y X , Y be the two connected components of O ∖ { e , f } O \setminus\{e, f\} O ∖ { e , f } . Let s s s be the separation induced by ( X , Y ) (X, Y) ( X , Y ) . Then, s s s is a displayed separation of D D D . We say that s s s is *displayed by { e , f } \{e, f\} { e , f } . Note that the separation s ∗ s^* s ∗ is also displayed by { e , f } \{e, f\} { e , f } .
Definition 6.2 (q q q -lean). Let F = ( H , V ) \mathcal{F} = (H, \mathcal{V}) F = ( H , V ) be a graphic-k k k -flower of a graph G G G . For every e ∈ E ( H ) e \in E(H) e ∈ E ( H ) , let V e V_e V e be the adhesion-set at e e e . Let Z = ⋂ h ∈ V ( H ) V h Z = \bigcap_{h \in V(H)} V_h Z = ⋂ h ∈ V ( H ) V h . If F \mathcal{F} F is a graphic-k k k -daisy, we say that F \mathcal{F} F is q q q -lean if for every ( > q ) (> q) ( > q ) -broad separation of G G G displayed by a set of edges { e , f } \{e, f\} { e , f } , the sets V e ∖ Z V_e \setminus Z V e ∖ Z and V f ∖ Z V_f \setminus Z V f ∖ Z are disjoint. If F \mathcal{F} F is a graphic-k k k -anemone, we say that F \mathcal{F} F is q q q -lean if ∣ Z ∣ = k |Z| = k ∣ Z ∣ = k .
Definition 6.3 (Clover-property). Let F = ( O , V ) \mathcal{F} = (O, \mathcal{V}) F = ( O , V ) be a graphic-k k k -daisy of a graph G G G . We say that F \mathcal{F} F has the clover-property if the centre X : = ⋂ o ∈ V ( O ) V o X := \bigcap_{o \in V(O)} V_o X := ⋂ o ∈ V ( O ) V o has size < k < k < k and F \mathcal{F} F has four adhesion-sets V e 0 , V e 1 , V e 2 , V e 3 V_{e_0}, V_{e_1}, V_{e_2}, V_{e_3} V e 0 , V e 1 , V e 2 , V e 3 with V e i ∩ V e j = X V_{e_i} \cap V_{e_j} = X V e i ∩ V e j = X for all i ≠ j i \ne j i = j .
Definition 6.4 (Canonical). Let S S S be a set of separations of a graph G G G . We say that S S S is canonical if for every automorphism φ \varphi φ of G G G we have φ ( S ) = S \varphi(S) = S φ ( S ) = S . Similarly, we call a tree-decomposition T = ( T , V ) \mathcal{T} = (T, \mathcal{V}) T = ( T , V ) canonical if the set of separations that are induced by an edge of T T T is canonical.
Definition 6.5 (ψ \psi ψ -isomorphic). Let H = ( H , V ) \mathcal{H} = (H, \mathcal{V}) H = ( H , V ) be a graph-decomposition of a graph G G G and H ′ = ( H ′ , V ′ ) \mathcal{H}' = (H', \mathcal{V}') H ′ = ( H ′ , V ′ ) be a graph-decomposition of a graph G ′ G' G ′ and let ψ : V ( G ) → V ( G ′ ) \psi: V(G) \to V(G') ψ : V ( G ) → V ( G ′ ) be an isomorphism between G G G and G ′ G' G ′ . We say that H \mathcal{H} H and H ′ \mathcal{H}' H ′ are ψ \psi ψ -isomorphic if there exists an isomorphism φ : V ( H ) → V ( H ′ ) \varphi: V(H) \to V(H') φ : V ( H ) → V ( H ′ ) such that for every h ∈ V ( H ) h \in V(H) h ∈ V ( H ) , we have ψ ( V h ) = V φ ( h ) ′ \psi(V_h) = V'_{\varphi(h)} ψ ( V h ) = V φ ( h ) ′ .
Key Lemma 6.6 (Flower Approximation Lemma). For all q , k ∈ N q, k \in\mathbb{N} q , k ∈ N , there exist constants β 6.6 ( q , k ) ≥ q \beta_{6.6}(q, k) \ge q β 6.6 ( q , k ) ≥ q , γ 6.6 ( q , k ) \gamma_{6.6}(q, k) γ 6.6 ( q , k ) and ε 6.6 ( q , k ) \varepsilon_{6.6}(q, k) ε 6.6 ( q , k ) such that the following holds. Let G G G be a ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable graph and γ > γ 6.6 \gamma> \gamma_{6.6} γ > γ 6.6 . Let S S S be a nontrivial γ \gamma γ -asteriskal set of separations of G G G that consists of β 6.6 \beta_{6.6} β 6.6 -broad narrow k k k -separations of G G G . Then there is a q q q -lean graphic-k k k -flower F : = F 6.6 ( S ) F := F_{6.6}(S) F := F 6.6 ( S ) of G G G such that:
every separation in S S S has edit-distance ≤ ε 6.6 \le\varepsilon_{6.6} ≤ ε 6.6 to some displayed separation of F F F ;
for every ε ≥ ε 6.6 \varepsilon\ge\varepsilon_{6.6} ε ≥ ε 6.6 , every k k k -separation of G G G that is ε \varepsilon ε -compatible with F F F is ε \varepsilon ε -nested with all the petal-separations of F F F ;
if F F F is a graphic-k k k -daisy, then F F F has the clover-property; and
The graphic-k k k -flower F 6.6 ( S ) F_{6.6}(S) F 6.6 ( S ) is canonical: whenever ψ \psi ψ is an automorphism of G G G , the graphic-k k k -flowers F 6.6 ( S ) F_{6.6}(S) F 6.6 ( S ) and F 6.6 ( ψ ( S ) ) F_{6.6}(\psi(S)) F 6.6 ( ψ ( S )) are ψ \psi ψ -isomorphic.
The constants will be (for q , k ∈ N q,k \in\mathbb{N} q , k ∈ N ):
β 6.6 ( q , k ) : = β 6.13 ( q , k ) \beta_{6.6}(q,k) := \beta_{6.13}(q,k) β 6.6 ( q , k ) := β 6.13 ( q , k )
γ 6.6 ( q , k ) : = max { γ 6.28 ( q , k ) , γ 6.8 ( q , k ) } \gamma_{6.6}(q,k) := \max\{\gamma_{6.28}(q,k),\gamma_{6.8}(q,k)\} γ 6.6 ( q , k ) := max { γ 6.28 ( q , k ) , γ 6.8 ( q , k )}
ε 6.6 ( q , k ) : = max { ε 6.32 ( q , ℓ ) : 0 ≤ ℓ ≤ k / 2 } \varepsilon_{6.6}(q,k) := \max\{\varepsilon_{6.32}(q,\ell):0 \leq\ell\leq k/2\} ε 6.6 ( q , k ) := max { ε 6.32 ( q , ℓ ) : 0 ≤ ℓ ≤ k /2 }
We start by establishing some basic properties of asteriskal sets of separations in § 7.1 . In particular, if S S S is a γ \gamma γ -asteriskal set of separations, then for each s ∈ S s \in S s ∈ S there is a tripartition ( X s , { P s , Q s } ) (X_s,\{P_s,Q_s\}) ( X s , { P s , Q s }) of sep ( s ) \operatorname{sep}(s) sep ( s ) such that every s ′ ∈ S s' \in S s ′ ∈ S that γ \gamma γ -crosses s s s induces ( X s , { P s , Q s } ) (X_s,\{P_s,Q_s\}) ( X s , { P s , Q s }) on sep ( s ) \operatorname{sep}(s) sep ( s ) . Furthermore, there exists a set X X X and an integer ℓ \ell ℓ such that for every s ∈ S s \in S s ∈ S we have X s = X X_s=X X s = X and ∣ P s ∣ = ℓ = ∣ Q s ∣ |P_s|=\ell=|Q_s| ∣ P s ∣ = ℓ = ∣ Q s ∣ .
When X X X is the separator of all separations in S S S , the graphic-X X X -anemone (Definition 6.9) satisfies the outcome of Key Lemma 6.6. This easy case is handled in § 7.2 .
The heart of the proof lies in the case where the separations in S S S do not satisfy X = sep ( s ) X=\operatorname{sep}(s) X = sep ( s ) . Then, we build a graphic-k k k -daisy that satisfies the outcome of Key Lemma 6.6. The construction roughly goes as follows.
First, we consider a covering ( G ^ , p ) (\widehat{G},p) ( G , p ) of G − X G-X G − X , where G ^ \widehat{G} G is an infinite graph and p : V ( G ^ ) → V ( G − X ) p:V(\widehat{G}) \to V(G-X) p : V ( G ) → V ( G − X ) is a local isomorphism. To construct this covering, we consider an arbitrary separation s = ( A , B ) ∈ S s=(A,B) \in S s = ( A , B ) ∈ S , and obtain G ^ \widehat{G} G by iteratively gluing copies of G [ A − X ] G[A-X] G [ A − X ] and G [ B − X ] G[B-X] G [ B − X ] on P s P_s P s and Q s Q_s Q s (see Figure 6.2 ). This construction is detailed in § 7.3 , where we also demonstrate some of its key properties. In particular, the covering does not depend on the separation s ∈ S s \in S s ∈ S we initially picked. Furthermore, the graph G ^ \widehat{G} G has exactly two ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 , and every separation of G ^ \widehat{G} G that distinguishes them has order at least ℓ \ell ℓ .
Figure 6.2. The covering p : G ^ → G − X p:\widehat{G}\to G-X p : G → G − X that is constructed in Construction 6.12. On the one hand, H A H_A H A consists of yellow, red and blue. On the other hand, H B H_B H B consists of turquoise, red and blue.
Then, denote by R R R the set of all ℓ \ell ℓ -separations ( U , W ) (U,W) ( U , W ) of G ^ \widehat{G} G such that ω 1 \omega_1 ω 1 lives in U U U and ω 2 \omega_2 ω 2 lives in W W W . Say that a separation s ∈ R s \in R s ∈ R is locally R R R -optimal if for every separation r ∈ R r \in R r ∈ R that crosses s s s , we have cr R ( s ) < cr R ( r ) \operatorname{cr}_R(s)<\operatorname{cr}_R(r) cr R ( s ) < cr R ( r ) , where cr R ( ⋅ ) \operatorname{cr}_R(\cdot) cr R ( ⋅ ) denotes the number of separations in R R R that cross a given separation. Let L L L be the set of all locally R R R -optimal separations. Note that the separations in L L L are pairwise nested. In § 7.4 , we prove that every separation in R R R is crossed by a bounded number of separations in R R R . From this, we deduce that every separation r ∈ R r \in R r ∈ R is at bounded edit-distance from a separation s ( r ) ∈ L s(r) \in L s ( r ) ∈ L .
Finally, since the separations in L L L are pairwise nested, they induce a unique (canonical) tree-decomposition of G ^ \widehat{G} G . This tree-decomposition of G ^ \widehat{G} G projects (via p p p ) to a (canonical) cycle-decomposition of G − X G-X G − X . In § 7.5 , we show that after adding X X X to all bags, this cycle-decomposition is a graphic-k k k -daisy of G G G that satisfies the outcome of Key Lemma 6.6.
Centres and links of asteriskal sets of separations
In this section, we prove some basic properties of asteriskal sets of separations.
Lemma 6.7. Let G G G be a graph and γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let S S S be a nontrivial γ \gamma γ -asteriskal set of separations of G G G . Then:
For every s ∈ S s \in S s ∈ S there is a unique tripartition ( X s , { P s , Q s } ) (X_s,\{P_s,Q_s\}) ( X s , { P s , Q s }) of sep ( s ) \operatorname{sep}(s) sep ( s ) such that every s ′ ∈ S s' \in S s ′ ∈ S that γ \gamma γ -crosses s s s induces ( X s , { P s , Q s } ) (X_s,\{P_s,Q_s\}) ( X s , { P s , Q s }) on sep ( s ) \operatorname{sep}(s) sep ( s ) .
(ii) There is a unique X ⊆ V ( G ) X \subseteq V(G) X ⊆ V ( G ) such that X = X s X = X_s X = X s for all s ∈ S s \in S s ∈ S .
Proof. (i). By ∣ S ∣ ≥ 2 |S| \ge2 ∣ S ∣ ≥ 2 and since S S S is γ \gamma γ -crossing-connected, for every s ∈ S s \in S s ∈ S there is some s ′ ∈ S s' \in S s ′ ∈ S that γ \gamma γ -crosses s s s . As S S S is uniformly γ \gamma γ -crossing, the tripartition ( X s , { P s , Q s } ) (X_s,\{P_s,Q_s\}) ( X s , { P s , Q s }) is unique. (ii) follows from (i) and from S S S being γ \gamma γ -crossing-connected. □
We refer to X X X as in Lemma 6.7 (ii) as the common centre of S S S .
Lemma 6.8. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let γ ∈ N \gamma\in\mathbb{N} γ ∈ N such that γ ≥ γ 6.8 ( q , k ) : = q + k \gamma\ge\gamma_{6.8}(q,k) := q+k γ ≥ γ 6.8 ( q , k ) := q + k . Let S S S be a nontrivial γ \gamma γ -asteriskal set of separations of G G G that consists of k k k -separations. For every s ∈ S s \in S s ∈ S let ( X s , { P s , Q s } ) (X_s,\{P_s,Q_s\}) ( X s , { P s , Q s }) be as in Lemma 6.7 (i). Then there is a unique ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N such that ∣ P s ∣ = ∣ Q s ∣ = ℓ |P_s|=|Q_s|=\ell ∣ P s ∣ = ∣ Q s ∣ = ℓ for all s ∈ S s \in S s ∈ S .
Proof. Let X X X be the common centre of S S S and define ℓ : = ( k − ∣ X ∣ ) / 2 \ell:= (k-|X|)/2 ℓ := ( k − ∣ X ∣ ) /2 . We show that ∣ P s ∣ = ∣ Q s ∣ = ℓ |P_s|=|Q_s|=\ell ∣ P s ∣ = ∣ Q s ∣ = ℓ for all s ∈ S s \in S s ∈ S . Assume otherwise, that there is a separation s ∈ S s \in S s ∈ S with ∣ Q s ∣ < ℓ < ∣ P s ∣ |Q_s|<\ell<|P_s| ∣ Q s ∣ < ℓ < ∣ P s ∣ , say. Let s ′ s' s ′ be a separation in S S S that γ \gamma γ -crosses s s s . Say, ∣ Q s ′ ∣ ≤ ℓ ≤ ∣ P s ′ ∣ |Q_{s'}|\le\ell\le|P_{s'}| ∣ Q s ′ ∣ ≤ ℓ ≤ ∣ P s ′ ∣ . Let K K K be the corner that is adjacent to Q s Q_s Q s and Q s ′ Q_{s'} Q s ′ , and let ( C , D ) (C,D) ( C , D ) be the K K K -corner-separation with K = C ∖ D K=C\setminus D K = C ∖ D , say. Then ( C , D ) (C,D) ( C , D ) has order ∣ Q s ∣ + ∣ Q s ′ ∣ + ∣ X ∣ < 2 ℓ + ∣ X ∣ ≤ k |Q_s|+|Q_{s'}|+|X|<2\ell+|X|\le k ∣ Q s ∣ + ∣ Q s ′ ∣ + ∣ X ∣ < 2 ℓ + ∣ X ∣ ≤ k . Since s s s and s ′ s' s ′ γ \gamma γ -cross, they ( γ − k ) (\gamma-k) ( γ − k ) -criss by Observation 5.4. Therefore, ∣ C ∖ D ∣ = ∣ K ∣ > γ − k ≥ q |C\setminus D|=|K|>\gamma-k\ge q ∣ C ∖ D ∣ = ∣ K ∣ > γ − k ≥ q and ∣ D ∖ C ∣ > 3 ( γ − k ) ≥ q |D\setminus C|>3(\gamma-k)\ge q ∣ D ∖ C ∣ > 3 ( γ − k ) ≥ q (using the assumption that γ ≥ q + k \gamma\ge q+k γ ≥ q + k ). Hence, ( C , D ) (C,D) ( C , D ) witnesses that G G G is not ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, a contradiction. □
We refer to ℓ \ell ℓ as in Lemma 6.8 as the common link size of the γ \gamma γ -asteriskal set of separations S S S . We remark that this definition requires γ ≥ γ 6.8 ( q , k ) \gamma\ge\gamma_{6.8}(q,k) γ ≥ γ 6.8 ( q , k ) .
Constructing the graphic-anemone
If the common link size is zero, all separators are equal to the common centre X X X . In this section, we show that the graphic-X X X -anemone displays every separation in S S S .
Definition 6.9 (Graphic-X X X -anemone). Let X X X be a set of vertices of a graph G G G and let C \mathcal{C} C be the set of components of G − X G-X G − X . The graphic-X X X -anemone of G G G is the graphic-∣ X ∣ |X| ∣ X ∣ -anemone A = ( S , V ) \mathcal{A}=(S,\mathcal{V}) A = ( S , V ) , where S S S is a star with C \mathcal{C} C as set of leaves, V C = X V_{\mathcal{C}}=X V C = X is the central bag, and V C = V ( C ) ∪ X V_C=V(C)\cup X V C = V ( C ) ∪ X for every component C ∈ C C\in\mathcal{C} C ∈ C .
Lemma 6.10. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let S S S be a nontrivial γ \gamma γ -asteriskal set of separations of G G G that consists of k k k -separations. Let X X X denote the common centre of S S S and assume that ∣ X ∣ = k |X|=k ∣ X ∣ = k . Let A \mathcal{A} A be the graphic-X X X -anemone. Then, A \mathcal{A} A is a q q q -lean graphic-k k k -anemone that displays every separation in S S S , and for every ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N , every separation which is ε \varepsilon ε -compatible with A \mathcal{A} A is ε \varepsilon ε -nested with all the petal-separations of A \mathcal{A} A .
Proof. A \mathcal{A} A is a q q q -lean graphic-k k k -anemone since ∣ X ∣ = k |X|=k ∣ X ∣ = k . Let s ∈ S s\in S s ∈ S . Then sep ( s ) = X \operatorname{sep}(s)=X sep ( s ) = X as X X X is the common centre of S S S , and has size k = ∣ sep ( s ) ∣ k=|\operatorname{sep}(s)| k = ∣ sep ( s ) ∣ . Hence s s s is displayed by A \mathcal{A} A .
Let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N and let s s s be a separation which is ε \varepsilon ε -compatible with A \mathcal{A} A . By contradiction, suppose that s s s ε \varepsilon ε -crosses some petal-separation t t t of A \mathcal{A} A . Since s s s is ε \varepsilon ε -compatible with A \mathcal{A} A , we have that s s s and t t t cross with centre X X X and four empty links. Thus, X = sep ( s ) X=\operatorname{sep}(s) X = sep ( s ) so s s s is displayed by A \mathcal{A} A , so cannot cross t t t . □
Remark 6.11. The graphic-k k k -anemone A = A ( S ) \mathcal{A}=\mathcal{A}(S) A = A ( S ) in Lemma 6.10 is canonical: whenever ψ \psi ψ is an automorphism of G G G , the graphic-k k k -anemones A ( S ) \mathcal{A}(S) A ( S ) and A ( ψ ( S ) ) \mathcal{A}(\psi(S)) A ( ψ ( S )) are ψ \psi ψ -isomorphic.
The covering construction
If the common link size is ℓ > 0 \ell> 0 ℓ > 0 , we build a covering ( G ^ , p ) (\widehat{G},p) ( G , p ) of G − X G-X G − X . In this section, we describe the construction of this covering (see Construction 6.12) and analyse some of its properties. We show that G ^ \widehat{G} G has exactly two ends and that every separation of G ^ \widehat{G} G that distinguishes them has order at least ℓ \ell ℓ (see Lemma 6.14 ). We also show that the covering ( G ^ , p ) (\widehat{G},p) ( G , p ) is canonical given ( G , S ) (G,\mathcal{S}) ( G , S ) (see Lemma 6.15 ).
A ray is a one-way infinite path. A tail of a ray means a subray. An end of a graph G G G is an equivalence class of rays in G G G , where two rays in G G G are considered equivalent if they are linked by infinitely many pairwise disjoint paths. We say that a ray of an infinite graph G G G lives in a set A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) if it has a tail contained in A A A . An end ω \omega ω of an infinite graph G G G lives in a set A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) if every ray contained in ω \omega ω lives in A A A . Note that, given a separation ( A , B ) (A,B) ( A , B ) of G G G of finite order, any end ω \omega ω of G G G lives either in A A A or in B B B , but not in both. A separation s s s of G G G distinguishes two ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 if they live in distinct sides of s s s . We say that s s s distinguishes ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 efficiently if s s s is of smallest order among all separations that distinguish ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 .
Construction 6.12 (Covering). This definition is supported by Figure 6.2 . Let ( A , B ) (A,B) ( A , B ) be a narrow proper k k k -separation of G G G . Let ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) be a tripartition of A ∩ B A \cap B A ∩ B such that ∣ X ∣ < k |X| < k ∣ X ∣ < k , ∣ P ∣ = ∣ Q ∣ |P|=|Q| ∣ P ∣ = ∣ Q ∣ and there is no edge in between P P P and Q Q Q . Let H A : = G [ A ] − X H_A:=G[A]-X H A := G [ A ] − X and H B : = G [ B ] − X H_B:=G[B]-X H B := G [ B ] − X . Let T T T be a double-ray … t ( − 1 ) t ( 0 ) t ( 1 ) … \ldots t(-1)t(0)t(1)\ldots … t ( − 1 ) t ( 0 ) t ( 1 ) … , viewed as a bipartite graph with classes T A T_A T A and T B T_B T B . For each t ∈ T A t \in T_A t ∈ T A let H t H_t H t be a copy of H A H_A H A , and for each t ∈ T B t \in T_B t ∈ T B let H t H_t H t be a copy of H B H_B H B . Let P t P_t P t and Q t Q_t Q t denote the copies of P P P and Q Q Q in H t H_t H t , respectively. We choose the graphs H t H_t H t so that they are pairwise disjoint except that they intersect as follows for every edge t ( i ) t ( i + 1 ) t(i)t(i+1) t ( i ) t ( i + 1 ) of T T T :
P t ( i ) = P t ( i + 1 ) = : L ( i ) P_{t(i)}=P_{t(i+1)}=:L(i) P t ( i ) = P t ( i + 1 ) =: L ( i ) if t ( i ) ∈ T A t(i)\in T_A t ( i ) ∈ T A and t ( i + 1 ) ∈ T B t(i+1)\in T_B t ( i + 1 ) ∈ T B , and
Q t ( i ) = Q t ( i + 1 ) = : L ( i ) Q_{t(i)}=Q_{t(i+1)}=:L(i) Q t ( i ) = Q t ( i + 1 ) =: L ( i ) if t ( i ) ∈ T B t(i)\in T_B t ( i ) ∈ T B and t ( i + 1 ) ∈ T A t(i+1)\in T_A t ( i + 1 ) ∈ T A .
Hence V ( H t ( i ) ) ∩ V ( H t ( i + 1 ) ) = L ( i ) V(H_{t(i)})\cap V(H_{t(i+1)})=L(i) V ( H t ( i ) ) ∩ V ( H t ( i + 1 ) ) = L ( i ) and V ( H t ( i ) ) ∩ V ( H t ( j ) ) = ∅ V(H_{t(i)})\cap V(H_{t(j)})=\varnothing V ( H t ( i ) ) ∩ V ( H t ( j ) ) = ∅ whenever ∣ i − j ∣ > 1 |i-j|>1 ∣ i − j ∣ > 1 . Let G ^ : = ⋃ t ∈ V ( T ) H t \widehat{G}:=\bigcup_{t\in V(T)}H_t G := ⋃ t ∈ V ( T ) H t . Note that G ^ \widehat{G} G is connected since ( A , B ) (A,B) ( A , B ) is narrow. Let V t : = V ( H t ) V_t:=V(H_t) V t := V ( H t ) . Then ( T , V ) (T,\mathcal{V}) ( T , V ) with V = ( V t : t ∈ T ) \mathcal{V}=(V_t:t\in T) V = ( V t : t ∈ T ) is a tree-decomposition of G ^ \widehat{G} G with pairwise disjoint adhesion-sets L ( i ) L(i) L ( i ) . Moreover, the map p : G ^ → G − X p:\widehat{G}\to G-X p : G → G − X that sends every vertex in a graph H t H_t H t to the vertex of G − X G-X G − X it stems from is a covering map (as there is no edge in between P P P and Q Q Q ). We call p p p the covering induced by ( A , B ) (A,B) ( A , B ) and ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) , as p p p is unique (up to equivalence of coverings). By a slight abuse of notation, we also call G ^ \widehat{G} G the induced covering . We call ( T , V ) (T,\mathcal{V}) ( T , V ) the standard decomposition of G ^ \widehat{G} G .
The following lemma will help us prove that every separation that distinguishes the two ends of G ^ \widehat{G} G has order at least ℓ \ell ℓ .
Lemma 6.13. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let β : = β 6.13 ( q , k ) : = 2 q + k \beta:=\beta_{6.13}(q,k):=2q+k β := β 6.13 ( q , k ) := 2 q + k . Let ( A , B ) (A,B) ( A , B ) be a β \beta β -broad k k k -separation of G G G and let ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) be a tripartition of A ∩ B A\cap B A ∩ B with ∣ P ∣ = ∣ Q ∣ = : ℓ |P|=|Q|=:\ell ∣ P ∣ = ∣ Q ∣ =: ℓ . Then there are ℓ \ell ℓ disjoint P P P –Q Q Q paths in G [ A ] − X G[A]-X G [ A ] − X .
Proof. Assume otherwise. Recall that H A = G [ A ] − X H_A=G[A]-X H A = G [ A ] − X . Then by Menger’s Theorem, there is a ( < ℓ ) (<\ell) ( < ℓ ) -separation ( C , D ) (C,D) ( C , D ) of H A H_A H A with P ⊆ C P\subseteq C P ⊆ C and Q ⊆ D Q\subseteq D Q ⊆ D . Then ( C 1 , D 1 ) : = ( C ∪ X , D ∪ B ) (C_1,D_1):=(C\cup X,D\cup B) ( C 1 , D 1 ) := ( C ∪ X , D ∪ B ) and ( C 2 , D 2 ) : = ( D ∪ X , C ∪ B ) (C_2,D_2):=(D\cup X,C\cup B) ( C 2 , D 2 ) := ( D ∪ X , C ∪ B ) are ( < k ) (<k) ( < k ) -separations of G G G . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and ∣ D i ∖ C i ∣ ≥ ∣ B ∖ A ∣ > q |D_i\setminus C_i|\ge|B\setminus A|>q ∣ D i ∖ C i ∣ ≥ ∣ B ∖ A ∣ > q , we must have ∣ C i ∖ D i ∣ ≤ q |C_i\setminus D_i|\le q ∣ C i ∖ D i ∣ ≤ q for i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } . Then, however,
∣ A ∖ B ∣ ≤ ∣ C 1 ∖ D 1 ∣ + ∣ C 2 ∖ D 2 ∣ + ∣ C ∩ D ∣ < 2 q + ℓ ≤ β , |A\setminus B|\le|C_1\setminus D_1|+|C_2\setminus D_2|+|C\cap D|<2q+\ell\le\beta, ∣ A ∖ B ∣ ≤ ∣ C 1 ∖ D 1 ∣ + ∣ C 2 ∖ D 2 ∣ + ∣ C ∩ D ∣ < 2 q + ℓ ≤ β ,
a contradiction.
Lemma 6.14. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let ( A , B ) (A,B) ( A , B ) be a β 6.13 ( q , k ) \beta_{6.13}(q,k) β 6.13 ( q , k ) -broad narrow k k k -separation of G G G and let ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) be a tripartition of A ∩ B A\cap B A ∩ B with ∣ X ∣ < k |X|<k ∣ X ∣ < k and ∣ P ∣ = ∣ Q ∣ = : ℓ |P|=|Q|=:\ell ∣ P ∣ = ∣ Q ∣ =: ℓ and no edge in between P P P and Q Q Q . Let G ^ \widehat{G} G be the covering induced by ( A , B ) (A,B) ( A , B ) and ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) . Then:
(i) G ^ \widehat{G} G has exactly two ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 ;
(ii) Every separation of G ^ \widehat{G} G distinguishing both ends has order ≥ ℓ \ge\ell ≥ ℓ .
Proof. (i). Since ( A , B ) (A,B) ( A , B ) is narrow, and in particular tight, there is a component of G [ A ∖ B ] G[A\setminus B] G [ A ∖ B ] with neighbourhood equal to A ∩ B A\cap B A ∩ B . Hence every two vertices in P ∪ Q P\cup Q P ∪ Q are joined by a path in H A : = G [ A ] − X H_A:=G[A]-X H A := G [ A ] − X , and similarly in H B : = G [ B ] − X H_B:=G[B]-X H B := G [ B ] − X . Therefore, if two rays R 1 , R 2 R_1,R_2 R 1 , R 2 in G ^ \widehat{G} G follow T T T in the same direction, there are infinitely many vertex-disjoint R 1 R_1 R 1 –R 2 R_2 R 2 paths in G ^ \widehat{G} G , so R 1 R_1 R 1 and R 2 R_2 R 2 lie in a common end. As every ray in G ^ \widehat{G} G follows T T T in some direction, G ^ \widehat{G} G has at most two ends.
Conversely, recall that ( A , B ) (A,B) ( A , B ) is tight and ∣ P ∣ = ∣ Q ∣ ≥ 1 |P|=|Q|\ge1 ∣ P ∣ = ∣ Q ∣ ≥ 1 . Hence, there are two rays R 1 R_1 R 1 and R 2 R_2 R 2 in G ^ \widehat{G} G that follow T T T in opposite directions. Then any adhesion-set L ( i ) L(i) L ( i ) of ( T , V ) (T,\mathcal{V}) ( T , V ) separates R 1 R_1 R 1 and R 2 R_2 R 2 . Hence, G ^ \widehat{G} G has at least two ends.
(ii). By Lemma 6.13, there are ℓ \ell ℓ vertex-disjoint P P P –Q Q Q paths contained in H A : = G [ A ] − X H_A:=G[A]-X H A := G [ A ] − X . Similarly, there are ℓ \ell ℓ vertex-disjoint P P P –Q Q Q paths in H B : = G [ B ] − X H_B:=G[B]-X H B := G [ B ] − X . The copies of these paths in G ^ \widehat{G} G combine to ℓ \ell ℓ vertex-disjoint double-rays between the two ends of G ^ \widehat{G} G . Hence every separator that distinguishes the two ends has order at least ℓ \ell ℓ .
Recall that two coverings p 1 : G ^ 1 → G p_1:\widehat{G}_1 \to G p 1 : G 1 → G and p 2 : G ^ 2 → G p_2:\widehat{G}_2 \to G p 2 : G 2 → G are equivalent if there is a graph-isomorphism f : G ^ 1 → G ^ 2 f:\widehat{G}_1 \to\widehat{G}_2 f : G 1 → G 2 such that p 1 = p 2 ∘ f p_1=p_2\circ f p 1 = p 2 ∘ f .
Lemma 6.15. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let S S S be a nontrivial γ \gamma γ -asteriskal set of separations of G G G that consists of narrow k k k -separations, where γ ≥ γ 6.8 ( q , k ) \gamma\ge\gamma_{6.8}(q,k) γ ≥ γ 6.8 ( q , k ) . For each s ∈ S s\in S s ∈ S , let ( X , { P s , Q s } ) (X,\{P_s,Q_s\}) ( X , { P s , Q s }) be as in Lemma 6.7 , where X X X is the common centre of S S S . Assume that ∣ X ∣ < k |X|<k ∣ X ∣ < k . Then:
Every s ∈ S s\in S s ∈ S and ( X , { P s , Q s } ) (X,\{P_s,Q_s\}) ( X , { P s , Q s }) induce a covering p s : G ^ s → G − X p_s:\widehat{G}_s\to G-X p s : G s → G − X .
All coverings p s p_s p s for s ∈ S s\in S s ∈ S are pairwise equivalent.
Proof. (i). Let s ∈ S s\in S s ∈ S . Since S S S is nontrivial, there is s ′ ∈ S s'\in S s ′ ∈ S that γ \gamma γ -crosses s s s . As S S S is uniformly γ \gamma γ -crossing, s ′ s' s ′ induces ( X , { P s , Q s } ) (X,\{P_s,Q_s\}) ( X , { P s , Q s }) on the separator of s s s . Since s ′ s' s ′ is a separation, it follows that G G G contains no edge in between P s P_s P s and Q s Q_s Q s . Therefore, the premises of Construction 6.12 are met, so s s s and ( X , { P s , Q s } ) (X,\{P_s,Q_s\}) ( X , { P s , Q s }) induce a covering of G − X G-X G − X .
(ii). Let s ≠ s ′ s\ne s' s = s ′ be separations in S S S . Since S S S is γ \gamma γ -crossing-connected, it suffices to prove it under the assumption that s s s and s ′ s' s ′ γ \gamma γ -cross. Write ( A , B ) : = s (A,B):=s ( A , B ) := s and ( C , D ) : = s ′ (C,D):=s' ( C , D ) := s ′ . Since S S S is uniformly γ \gamma γ -crossing, we have P s ′ ⊆ A ∖ B P_{s'}\subseteq A\setminus B P s ′ ⊆ A ∖ B and Q s ′ ⊆ B ∖ A Q_{s'}\subseteq B\setminus A Q s ′ ⊆ B ∖ A , say. Adding P s ′ P_{s'} P s ′ and Q s ′ Q_{s'} Q s ′ to Figure 6.2 , it is straightforward to see that the covering induced by s s s and ( X , { P s , Q s } ) (X,\{P_s,Q_s\}) ( X , { P s , Q s }) is equivalent to the covering induced by s ′ s' s ′ and ( X , { P s ′ , Q s ′ } ) (X,\{P_{s'},Q_{s'}\}) ( X , { P s ′ , Q s ′ }) . □
Let S S S and X X X be as in the assumptions of Lemma 6.15 . By Lemma 6.15 , each s ∈ S s\in S s ∈ S induces a covering of G − X G-X G − X , and all these coverings are equivalent. Hence, we may speak of the covering induced by the γ \gamma γ -asteriskal set of separations S S S .
Approximating by locally-optimal separations
Given a set S S S of separations of a graph G G G and a separation s ∈ S s\in S s ∈ S , we denote by cr S ( s ) \operatorname{cr}_S(s) cr S ( s ) the S S S -crossing-number of s s s , that is, the number of separations in S S S that cross s s s . Bounds for the S S S -crossing-number will usually use the letter χ \chi χ . A separation s ∈ S s\in S s ∈ S is locally S S S -optimal if cr S ( s ) < cr S ( r ) \operatorname{cr}_S(s)<\operatorname{cr}_S(r) cr S ( s ) < cr S ( r ) for every r ∈ S r\in S r ∈ S that crosses s s s .
Throughout the next two sections, we will work in the following setting.
Setting 6.16. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, and let γ ∈ N \gamma\in\mathbb{N} γ ∈ N such that γ ≥ γ 6.8 ( q , k ) \gamma\ge\gamma_{6.8}(q,k) γ ≥ γ 6.8 ( q , k ) . Let S S S be a nontrivial γ \gamma γ -asteriskal set of separations of G G G that consists of β 6.13 ( q , k ) \beta_{6.13}(q,k) β 6.13 ( q , k ) -broad narrow k k k -separations. Let X X X denote the common centre of S S S and assume that ∣ X ∣ < k |X|<k ∣ X ∣ < k . Let ℓ \ell ℓ denote the common link size of S S S . Let p : G ^ → G − X p:\widehat{G}\to G-X p : G → G − X denote the covering of G − X G-X G − X induced by S S S . By Lemma 6.14 , G ^ \widehat{G} G has exactly two ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 . Let R R R be the set of all ℓ \ell ℓ -separations ( U , W ) (U,W) ( U , W ) of G ^ \widehat{G} G such that ω 1 \omega_1 ω 1 lives in U U U and ω 2 \omega_2 ω 2 lives in W W W . Let L L L be the set of all locally R R R -optimal separations in R R R , which is nested by definition.
In this section, we show that every separation r ∈ R r\in R r ∈ R is at bounded edit-distance from some separation s ( r ) ∈ L s(r)\in L s ( r ) ∈ L (see Lemma 6.24 ). To do so, we first show that every separation r ∈ R r\in R r ∈ R is only crossed by a bounded number of separations r ′ ∈ R r'\in R r ′ ∈ R (see Lemma 6.22 , which immediately follows from Lemmas 6.20 and 6.21 ). Then, we argue that whenever two separations in R R R cross, they cross with two small opposite corners (see Lemma 6.23 ). To prove Lemma 6.24 , we then argue that starting from a separation r ∈ R r \in R r ∈ R , it suffices to “wiggle” it a bounded number of times by a small amount to obtain a separation in L L L .
We will use the following result of Elbracht, Kneip and Teegen. Recall that a (potentially infinite) graph is locally finite if every vertex has finite degree.
Lemma 6.17. [EKT22], Proof of Proposition 6.2. Let G G G be a locally finite connected graph . Let ( A , B ) (A,B) ( A , B ) be a tight finite-order separation of G G G , and let k ∈ N k \in\mathbb{N} k ∈ N . Then only finitely many tight separations of G G G of order ≤ k \le k ≤ k cross ( A , B ) (A,B) ( A , B ) .
We now show that R R R is closed under taking meets and joins, and that these operations can be used to create separations with smaller R R R -crossing-number.
Lemma 6.18 ([31 ] , Proof of Corollary 3.3). Let r r r be a separation of a graph G G G , let S S S be a set of separations of G G G and let s ∈ S s \in S s ∈ S . Assume that both cr S ( r ) \operatorname{cr}_{S}(r) cr S ( r ) and cr S ( s ) \operatorname{cr}_{S}(s) cr S ( s ) are finite and that r r r and s s s cross. Then
cr S ( r ∧ s ) + cr S ( r ∨ s ) < cr S ( r ) + cr S ( s ) . \operatorname{cr}_{S}(r \wedge s) + \operatorname{cr}_{S}(r \vee s) < \operatorname{cr}_{S}(r) + \operatorname{cr}_{S}(s). cr S ( r ∧ s ) + cr S ( r ∨ s ) < cr S ( r ) + cr S ( s ) .
Proof. If a separation t ∈ S t \in S t ∈ S satisfies t ≤ r t \le r t ≤ r or t ≤ s t \le s t ≤ s , then we have t ≤ r ∨ s t \le r \vee s t ≤ r ∨ s since r , s ≤ r ∨ s r,s \le r \vee s r , s ≤ r ∨ s . Similarly, r ≤ t r \le t r ≤ t or s ≤ t s \le t s ≤ t implies r ∧ s ≤ t r \wedge s \le t r ∧ s ≤ t since r ∧ s ≤ r , s r \wedge s \le r,s r ∧ s ≤ r , s . Hence, if a separation t ∈ S t \in S t ∈ S is nested with r r r or s s s , then it is nested with r ∧ s r \wedge s r ∧ s or r ∨ s r \vee s r ∨ s . Moreover, if t ∈ S t \in S t ∈ S is nested with both r r r and s s s , then it is nested with both r ∧ s r \wedge s r ∧ s and r ∨ s r \vee s r ∨ s (since r r r and s s s cross). It follows that cr S ( r ∧ s ) + cr S ( r ∨ s ) ≤ cr S ( r ) + cr S ( s ) \operatorname{cr}_{S}(r \wedge s) + \operatorname{cr}_{S}(r \vee s) \le\operatorname{cr}_{S}(r) + \operatorname{cr}_{S}(s) cr S ( r ∧ s ) + cr S ( r ∨ s ) ≤ cr S ( r ) + cr S ( s ) .
Since s ∈ S s \in S s ∈ S crosses r r r , the separation s s s is counted once on the right side of the inequality. Since s s s is nested with r ∧ s r \wedge s r ∧ s and r ∨ s r \vee s r ∨ s , it is not counted on the left side of the inequality. So, we have cr S ( r ∧ s ) + cr S ( r ∨ s ) < cr S ( r ) + cr S ( s ) \operatorname{cr}_{S}(r \wedge s) + \operatorname{cr}_{S}(r \vee s) < \operatorname{cr}_{S}(r) + \operatorname{cr}_{S}(s) cr S ( r ∧ s ) + cr S ( r ∨ s ) < cr S ( r ) + cr S ( s ) .
:::
Lemma 6.19. Assume Setting 6.16. Let s 1 , s 2 s_1,s_2 s 1 , s 2 be two separations in R R R . Let c 1 : = s 1 ∧ s 2 c_1 := s_1 \wedge s_2 c 1 := s 1 ∧ s 2 and c 2 : = s 1 ∨ s 2 c_2 := s_1 \vee s_2 c 2 := s 1 ∨ s 2 . Then both c 1 c_1 c 1 and c 2 c_2 c 2 are contained in R R R . Moreover, if s 1 s_1 s 1 and s 2 s_2 s 2 cross and cr R ( s 1 ) ≤ cr R ( s 2 ) \operatorname{cr}_{R}(s_1) \le\operatorname{cr}_{R}(s_2) cr R ( s 1 ) ≤ cr R ( s 2 ) , then there is i ∈ [ 2 ] i \in[2] i ∈ [ 2 ] such that cr R ( c i ) < cr R ( s 2 ) \operatorname{cr}_{R}(c_i) < \operatorname{cr}_{R}(s_2) cr R ( c i ) < cr R ( s 2 ) .
Proof. If s 1 s_1 s 1 and s 2 s_2 s 2 do not cross then we have s 1 ≤ s 2 s_1 \le s_2 s 1 ≤ s 2 or s 2 ≤ s 1 s_2 \le s_1 s 2 ≤ s 1 since ω 1 \omega_1 ω 1 lives on the left side of each s i s_i s i and ω 2 \omega_2 ω 2 lives on the right side of each s i s_i s i . Therefore, { c 1 , c 2 } = { s 1 , s 2 } ⊆ R \{c_1,c_2\} = \{s_1,s_2\} \subseteq R { c 1 , c 2 } = { s 1 , s 2 } ⊆ R and we are done.
Hence, we may assume that s 1 s_1 s 1 and s 2 s_2 s 2 cross. Both c 1 c_1 c 1 and c 2 c_2 c 2 distinguish ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 , hence both have order ≥ ℓ \ge\ell ≥ ℓ by Lemma 6.14 . Since c 1 c_1 c 1 and c 2 c_2 c 2 are opposite corner-separations, it follows from submodularity that both c i c_i c i have order equal to ℓ \ell ℓ . Moreover, ω 1 \omega_1 ω 1 lives on the left side of each c i c_i c i , and ω 2 \omega_2 ω 2 lives on the right side of each c i c_i c i . Hence both c i c_i c i are contained in R R R .
Assume now that cr R ( s 1 ) ≤ cr R ( s 2 ) \operatorname{cr}_{R}(s_1) \le\operatorname{cr}_{R}(s_2) cr R ( s 1 ) ≤ cr R ( s 2 ) . Note that every separation s ∈ R s \in R s ∈ R is tight since it efficiently distinguishes ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 . Thus, both cr R ( s i ) \operatorname{cr}_{R}(s_i) cr R ( s i ) are finite by Lemma 6.17 . By Lemma 6.18 , we have
cr R ( c 1 ) + cr R ( c 2 ) < cr R ( s 1 ) + cr R ( s 2 ) ≤ 2 cr R ( s 2 ) . \operatorname{cr}_{R}(c_1) + \operatorname{cr}_{R}(c_2) < \operatorname{cr}_{R}(s_1) + \operatorname{cr}_{R}(s_2) \le2\operatorname{cr}_{R}(s_2). cr R ( c 1 ) + cr R ( c 2 ) < cr R ( s 1 ) + cr R ( s 2 ) ≤ 2 cr R ( s 2 ) .
Hence there is i ∈ [ 2 ] i \in[2] i ∈ [ 2 ] such that cr R ( c i ) < cr R ( s 2 ) \operatorname{cr}_{R}(c_i) < \operatorname{cr}_{R}(s_2) cr R ( c i ) < cr R ( s 2 ) .
:::
The next result is our first step towards proving that cr R ( ⋅ ) \operatorname{cr}_{R}(\cdot) cr R ( ⋅ ) is bounded.
Lemma 6.20. Assume Setting 6.16. Let β ∈ N \beta\in\mathbb{N} β ∈ N and let χ 6.20 ( β , ℓ ) : = 3 2 β + 3 ℓ \chi_{6.20}(\beta,\ell) := 3^{2\beta+3\ell} χ 6.20 ( β , ℓ ) := 3 2 β + 3 ℓ . Let s = ( A , B ) ∈ R s = (A,B) \in R s = ( A , B ) ∈ R , and let R s R_s R s be the set of all ( C , D ) ∈ R (C,D) \in R ( C , D ) ∈ R that cross ( A , B ) (A,B) ( A , B ) . Then at least one of the following holds:
∣ R s ∣ < χ 6.20 ( β , ℓ ) |R_s| < \chi_{6.20}(\beta,\ell) ∣ R s ∣ < χ 6.20 ( β , ℓ ) , or
there is a ( < 2 ℓ ) (<2\ell) ( < 2 ℓ ) -separation ( U , W ) (U,W) ( U , W ) of G ^ \widehat{G} G such that ∣ U ∖ W ∣ ≥ β |U \setminus W| \ge\beta ∣ U ∖ W ∣ ≥ β and both ω i \omega_i ω i live in the side W W W .
We first give some intuition about the proof. To prove that R s R_s R s has bounded size, it suffices to prove that for every tripartition A ∩ B = P ⊔ X ⊔ Q A \cap B = P \sqcup X \sqcup Q A ∩ B = P ⊔ X ⊔ Q , the set R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) of all separations t = ( C , D ) ∈ R s t=(C,D) \in R_s t = ( C , D ) ∈ R s that cross s = ( A , B ) s=(A,B) s = ( A , B ) with P P P as the s D sD sD -link, X X X as the s t st s t -centre, and Q Q Q as the s C sC s C -link, has bounded size. Fix such a tripartition A ∩ B = P ⊔ X ⊔ Q A \cap B = P \sqcup X \sqcup Q A ∩ B = P ⊔ X ⊔ Q , and consider the set R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) . Consider the supremum and the infimum of all separations in R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) , and look at the region F F F between them (this is the union of the small sides of the green and yellow separations in Figure 6.3 ). Observe that each separation in R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) is entirely defined by its restriction to F F F . Thus, if ∣ F ∣ |F| ∣ F ∣ is small then ∣ R s ( P , X , Q ) ∣ |R_s(P,X,Q)| ∣ R s ( P , X , Q ) ∣ is small. If ∣ F ∣ |F| ∣ F ∣ is large, then one of these two separations has a small side which is large, and thus satisfies the second condition.
Figure 6.3. The situation in the proof of Lemma 6.20 . Here U 1 ∩ W 1 U_1 \cap W_1 U 1 ∩ W 1 is green and U 2 ∩ W 2 U_2 \cap W_2 U 2 ∩ W 2 is yellow.
Proof of Lemma 6.20 . The set R s R_s R s is finite by Lemma 6.17 . Let T T T denote the set of all tripartitions ( P , X , Q ) (P,X,Q) ( P , X , Q ) of A ∩ B A \cap B A ∩ B . For each ( P , X , Q ) ∈ T (P,X,Q) \in T ( P , X , Q ) ∈ T , let R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) denote the set of all t = ( C , D ) ∈ R s t=(C,D) \in R_s t = ( C , D ) ∈ R s that cross ( A , B ) (A,B) ( A , B ) with P P P as the s D sD sD -link, X X X as the s t st s t -centre, and Q Q Q as the s C sC s C -link. Now choose ( P , X , Q ) ∈ T (P,X,Q) \in T ( P , X , Q ) ∈ T with R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) of maximum size. If R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) is empty, then ∣ R s ∣ = 0 |R_s|=0 ∣ R s ∣ = 0 and we trivially have (i). Hence we may assume that R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) is not empty.
Let s 1 = ( X 1 , Y 1 ) s_1=(X_1,Y_1) s 1 = ( X 1 , Y 1 ) be the infimum of R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) and let s 2 = ( X 2 , Y 2 ) s_2=(X_2,Y_2) s 2 = ( X 2 , Y 2 ) be the supremum of R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) . By Lemma 6.19 , s 1 s_1 s 1 and s 2 s_2 s 2 are both in R R R . Moreover, for every s ∈ R s ( P , X , Q ) s \in R_s(P,X,Q) s ∈ R s ( P , X , Q ) we have s 1 ≤ s ≤ s 2 s_1 \le s \le s_2 s 1 ≤ s ≤ s 2 by definition.
Claim 6.20.1. Let i ∈ [ 2 ] i \in[2] i ∈ [ 2 ] and consider the two separations s i = ( X i , Y i ) s_i=(X_i,Y_i) s i = ( X i , Y i ) and s = ( A , B ) s=(A,B) s = ( A , B ) . Then the s Y i sY_i s Y i -link L ( Y i ) L(Y_i) L ( Y i ) equals P P P , the s s i ss_i s s i -centre equals X X X , and the s X i sX_i s X i -link L ( X i ) L(X_i) L ( X i ) equals Q Q Q .
Proof of the claim. Say i = 1 i=1 i = 1 . Since ( P , X , Q ) (P,X,Q) ( P , X , Q ) is a tripartition of A ∩ B A \cap B A ∩ B , it suffices to show the three inclusions
P ⊆ L ( Y 1 ) , X ⊆ X 1 ∩ Y 1 , and Q ⊆ L ( X 1 ) . P \subseteq L(Y_1), \qquad X \subseteq X_1 \cap Y_1, \qquad\text{and} \qquad Q \subseteq L(X_1). P ⊆ L ( Y 1 ) , X ⊆ X 1 ∩ Y 1 , and Q ⊆ L ( X 1 ) .
We have Q ⊆ X 1 ∖ Y 1 Q \subseteq X_1 \setminus Y_1 Q ⊆ X 1 ∖ Y 1 as s 1 s_1 s 1 is the infimum of R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) , and Q ⊆ A ∩ B Q \subseteq A \cap B Q ⊆ A ∩ B by definition. Hence Q ⊆ L ( X 1 ) Q \subseteq L(X_1) Q ⊆ L ( X 1 ) . Similarly, P ⊆ L ( Y 1 ) P \subseteq L(Y_1) P ⊆ L ( Y 1 ) . Finally, X X X is included in all separators of the separations in R s ( P , X , Q ) R_s(P,X,Q) R s ( P , X , Q ) , so X X X must also be included in the separator of the infimum of all these separations. Hence X ⊆ X 1 ∩ Y 1 X \subseteq X_1 \cap Y_1 X ⊆ X 1 ∩ Y 1 . ◇
The following definitions are supported by Figure 6.3 . Let
t 1 : = ( U 1 , W 1 ) : = s 1 ∗ ∧ s 2 ∧ s = ( Y 1 , X 1 ) ∧ ( X 2 , Y 2 ) ∧ ( A , B ) t_1 := (U_1,W_1) := s_1^* \wedge s_2 \wedge s = (Y_1,X_1) \wedge(X_2,Y_2) \wedge(A,B) t 1 := ( U 1 , W 1 ) := s 1 ∗ ∧ s 2 ∧ s = ( Y 1 , X 1 ) ∧ ( X 2 , Y 2 ) ∧ ( A , B )
and
t 2 : = ( U 2 , W 2 ) : = s 1 ∗ ∧ s 2 ∧ s ∗ = ( Y 1 , X 1 ) ∧ ( X 2 , Y 2 ) ∧ ( B , A ) . t_2 := (U_2,W_2) := s_1^* \wedge s_2 \wedge s^* = (Y_1,X_1) \wedge(X_2,Y_2) \wedge(B,A). t 2 := ( U 2 , W 2 ) := s 1 ∗ ∧ s 2 ∧ s ∗ = ( Y 1 , X 1 ) ∧ ( X 2 , Y 2 ) ∧ ( B , A ) .
Claim 6.20.2. For both i = 1 , 2 i=1,2 i = 1 , 2 we have that ( U i , W i ) (U_i,W_i) ( U i , W i ) is a ( < 2 ℓ ) (<2\ell) ( < 2 ℓ ) -separation such that both ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 live in W i W_i W i .
Proof of the claim. Recall that ω 1 \omega_1 ω 1 lives in X 1 X_1 X 1 and ω 2 \omega_2 ω 2 lives in Y 2 Y_2 Y 2 . Hence, both ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 live in X 1 ∪ Y 2 ⊆ W i X_1 \cup Y_2 \subseteq W_i X 1 ∪ Y 2 ⊆ W i . Assume for a contradiction that ∣ U 1 ∩ W 1 ∣ ≥ 2 ℓ |U_1 \cap W_1| \ge2\ell ∣ U 1 ∩ W 1 ∣ ≥ 2 ℓ , say. Let Z : = ( X 1 ∩ Y 1 ) ∪ ( X 2 ∩ Y 2 ) Z := (X_1 \cap Y_1) \cup(X_2 \cap Y_2) Z := ( X 1 ∩ Y 1 ) ∪ ( X 2 ∩ Y 2 ) . By the definition of infimum and supremum, we have
U 1 ∩ W 1 ⊆ A ∩ Z and U 2 ∩ W 2 ⊆ B ∩ Z , U_1 \cap W_1 \subseteq A \cap Z \quad\text{and} \quad U_2 \cap W_2 \subseteq B \cap Z, U 1 ∩ W 1 ⊆ A ∩ Z and U 2 ∩ W 2 ⊆ B ∩ Z ,
see Figure 6.3 . Hence
∣ U 1 ∩ W 1 ∣ + ∣ U 2 ∩ W 2 ∣ ≤ ∣ A ∩ Z ∣ + ∣ B ∩ Z ∣ = Claim 6.20.1 ∣ Z ∣ + ∣ X ∣ ≤ 2 ℓ . |U_1 \cap W_1| + |U_2 \cap W_2| \le|A \cap Z| + |B \cap Z| \mathrel{\overset{\text{Claim 6.20.1}}{=}} |Z| + |X| \le2\ell. ∣ U 1 ∩ W 1 ∣ + ∣ U 2 ∩ W 2 ∣ ≤ ∣ A ∩ Z ∣ + ∣ B ∩ Z ∣ = Claim 6.20.1 ∣ Z ∣ + ∣ X ∣ ≤ 2 ℓ .
Now ∣ U 1 ∩ W 1 ∣ ≥ 2 ℓ |U_1 \cap W_1| \ge2\ell ∣ U 1 ∩ W 1 ∣ ≥ 2 ℓ implies that U 2 ∩ W 2 U_2 \cap W_2 U 2 ∩ W 2 is empty. Since both ends of G ^ \hat{G} G ^ live on the side W 2 W_2 W 2 and since G ^ \hat{G} G ^ is connected, the side U 2 U_2 U 2 must be empty. By the definition of infimum we have
Y 1 ∩ X 2 ∩ B = U 2 = ∅ . (2) Y_1 \cap X_2 \cap B = U_2 = \varnothing. \tag*{(2)} Y 1 ∩ X 2 ∩ B = U 2 = ∅ . (2)
Let s ′ = ( I , J ) ∈ R s ( P , X , Q ) s' = (I,J) \in R_s(P,X,Q) s ′ = ( I , J ) ∈ R s ( P , X , Q ) . Then ( X 1 , Y 1 ) ≤ ( I , J ) ≤ ( X 2 , Y 2 ) (X_1,Y_1) \le(I,J) \le(X_2,Y_2) ( X 1 , Y 1 ) ≤ ( I , J ) ≤ ( X 2 , Y 2 ) by the definition of ( X i , Y i ) (X_i,Y_i) ( X i , Y i ) . Hence,
I ∩ J ⊆ Y 1 ∩ X 2 ⊆ A ∖ B . I \cap J \subseteq Y_1 \cap X_2 \subseteq A \setminus B. I ∩ J ⊆ Y 1 ∩ X 2 ⊆ A ∖ B .
We have ( A , B ) , ( I , J ) ∈ R (A,B),(I,J) \in R ( A , B ) , ( I , J ) ∈ R , so ω 1 \omega_1 ω 1 lives in the A I AI A I -corner and ω 2 \omega_2 ω 2 lives in the B J BJ B J -corner. Since I ∩ J ⊆ A ∖ B I \cap J \subseteq A \setminus B I ∩ J ⊆ A ∖ B , the s s ′ ss' s s ′ -centre and the B s ′ Bs' B s ′ -link are empty. The B J BJ B J -corner-separation separates ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 , so it has order ≥ ℓ \ge\ell ≥ ℓ by Lemma 6.14 . Therefore, the s J sJ s J -link has size ℓ \ell ℓ , so the s I sI s I -link is empty. Since G ^ \hat{G} G ^ is connected, this implies that the B I BI B I -corner is also empty, which contradicts that s s s and s ′ s' s ′ cross. ♢
By Claim 6.20.2, we may assume that both ( U i , W i ) (U_i,W_i) ( U i , W i ) (for i = 1 , 2 i = 1,2 i = 1 , 2 ) are ( < 2 ℓ ) (< 2\ell) ( < 2 ℓ ) -separations such that both ends ω 1 , ω 2 \omega_1,\omega_2 ω 1 , ω 2 live in W i W_i W i . If for some i ∈ [ 2 ] i \in[2] i ∈ [ 2 ] we have ∣ U i ∖ W i ∣ ≥ β |U_i \setminus W_i| \ge\beta ∣ U i ∖ W i ∣ ≥ β , then we have (ii) and are done. So we may assume that ∣ U i ∖ W i ∣ < β |U_i \setminus W_i| < \beta ∣ U i ∖ W i ∣ < β for both i ∈ [ 2 ] i \in[2] i ∈ [ 2 ] .
Let F : = Y 1 ∩ X 2 F := Y_1 \cap X_2 F := Y 1 ∩ X 2 . Then
F ⊆ ( X 1 ∩ Y 1 ) ∪ ( X 2 ∩ Y 2 ) ∪ ( U 1 ∖ W 1 ) ∪ ( U 2 ∖ W 2 ) F \subseteq(X_1 \cap Y_1) \cup(X_2 \cap Y_2) \cup(U_1 \setminus W_1) \cup(U_2 \setminus W_2) F ⊆ ( X 1 ∩ Y 1 ) ∪ ( X 2 ∩ Y 2 ) ∪ ( U 1 ∖ W 1 ) ∪ ( U 2 ∖ W 2 )
which implies ∣ F ∣ < 2 β + 2 ℓ |F| < 2\beta+ 2\ell ∣ F ∣ < 2 β + 2 ℓ . Every ( C , D ) ∈ R s ( P , X , Q ) (C,D) \in R_s(P,X,Q) ( C , D ) ∈ R s ( P , X , Q ) satisfies
( X 1 , Y 1 ) ≤ ( C , D ) ≤ ( X 2 , Y 2 ) . (X_1,Y_1) \le(C,D) \le(X_2,Y_2). ( X 1 , Y 1 ) ≤ ( C , D ) ≤ ( X 2 , Y 2 ) .
Hence ( C , D ) (C,D) ( C , D ) is uniquely determined by ( C ∩ F , D ∩ F ) (C \cap F,D \cap F) ( C ∩ F , D ∩ F ) . There are at most 3 ∣ F ∣ 3^{|F|} 3 ∣ F ∣ separations of G ^ [ F ] \widehat{G}[F] G [ F ] . Thus
∣ R s ( P , X , Q ) ∣ < 3 2 β + 2 ℓ |R_s(P,X,Q)| < 3^{2\beta+2\ell} ∣ R s ( P , X , Q ) ∣ < 3 2 β + 2 ℓ
We also have ∣ T ∣ ≤ 3 ℓ |T| \le3^\ell ∣ T ∣ ≤ 3 ℓ . Therefore,
∣ R s ∣ ≤ ∣ R s ( P , X , Q ) ∣ ⋅ ∣ T ∣ < 3 2 β + 3 ℓ = χ 6.20 ( β , ℓ ) |R_s| \le|R_s(P,X,Q)| \cdot|T| < 3^{2\beta+3\ell} = \chi_{6.20}(\beta,\ell) ∣ R s ∣ ≤ ∣ R s ( P , X , Q ) ∣ ⋅ ∣ T ∣ < 3 2 β + 3 ℓ = χ 6.20 ( β , ℓ )
which completes the proof. □
The next result is our second step towards proving that cr R ( ⋅ ) \operatorname{cr}_{R}(\cdot) cr R ( ⋅ ) is bounded.
Lemma 6.21. Assume Setting 6.16. Let ( U , W ) (U,W) ( U , W ) be a ( < 2 ℓ ) (< 2\ell) ( < 2 ℓ ) -separation of G ^ \widehat{G} G such that both ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 live in W W W . Then ∣ U ∖ W ∣ < β 6.21 ( q , ℓ ) : = 4 ℓ ( q + 2 ℓ ) |U \setminus W| < \beta_{6.21}(q,\ell) := 4\ell(q+2\ell) ∣ U ∖ W ∣ < β 6.21 ( q , ℓ ) := 4 ℓ ( q + 2 ℓ ) .
We first give a quick proof sketch. Let ( A , B ) ∈ S (A,B) \in S ( A , B ) ∈ S be any separation, and let ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) be its corresponding tripartition. We consider the standard decomposition of G ^ \widehat{G} G induced by ( A , B ) (A,B) ( A , B ) and ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) and analyse how U U U interacts with each bag V t V_t V t (recall that G ^ [ V t ] \widehat{G}[V_t] G [ V t ] is a copy either of G [ A − X ] G[A-X] G [ A − X ] or of G [ B − X ] G[B-X] G [ B − X ] , with a copy P t P_t P t of P P P and a copy Q t Q_t Q t of Q Q Q ). Let V t V_t V t be a bag such that U ∩ V t ≠ ∅ U \cap V_t \ne\varnothing U ∩ V t = ∅ . Consider the separation ( U t , W t ) = ( U ∩ V t , ( W ∩ V t ) ∪ P t ∪ Q t ) (U_t,W_t) = (U \cap V_t,(W \cap V_t) \cup P_t \cup Q_t) ( U t , W t ) = ( U ∩ V t , ( W ∩ V t ) ∪ P t ∪ Q t ) (this is the yellow separation in Figure 6.4 ). We first argue that this separation has order < 2 ℓ < 2\ell < 2 ℓ . This separation (of G [ A − X ] G[A-X] G [ A − X ] , say) gives a ( < k ) (< k) ( < k ) -separation of G G G by adding X X X to the separator and B ∖ A B \setminus A B ∖ A to the side of W t W_t W t . This then implies that U t = U ∩ V t U_t = U \cap V_t U t = U ∩ V t is small. Then, we argue that if U ∩ V t ≠ ∅ U \cap V_t \ne\varnothing U ∩ V t = ∅ then U ∩ W ∩ V t ≠ ∅ U \cap W \cap V_t \ne\varnothing U ∩ W ∩ V t = ∅ . Since non-consecutive bags of the tree-decomposition are disjoint, there can only be at most 2 k 2k 2 k bags V t V_t V t such that U ∩ V t ≠ ∅ U \cap V_t \ne\varnothing U ∩ V t = ∅ . This immediately implies that U U U has bounded size.
Figure 6.4. The interaction of ( U , W ) (U,W) ( U , W ) with V t V_t V t and the location of ( U t , W t ) (U_t,W_t) ( U t , W t )
Proof. Let ( A , B ) ∈ S (A,B) \in S ( A , B ) ∈ S be arbitrary and let ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) be the tripartition that all separations in S S S that γ \gamma γ -cross ( A , B ) (A,B) ( A , B ) induce on A ∩ B A \cap B A ∩ B . Consider the tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) of G ^ \widehat{G} G induced by ( A , B ) (A,B) ( A , B ) and ( X , { P , Q } ) (X,\{P,Q\}) ( X , { P , Q }) . Without loss of generality, assume T = Z T = \mathbb{Z} T = Z . For t ∈ V ( T ) t \in V(T) t ∈ V ( T ) , we define C t : = V t C_t := V_t C t := V t and D t : = ⋃ s ∈ V ( T − t ) V s D_t := \bigcup_{s \in V(T-t)} V_s D t := ⋃ s ∈ V ( T − t ) V s . Observe that ( C t , D t ) (C_t,D_t) ( C t , D t ) is a separation of G ^ \widehat{G} G of order ∣ P ∪ Q ∣ = 2 ℓ |P \cup Q| = 2\ell ∣ P ∪ Q ∣ = 2 ℓ for every t ∈ V ( T ) t \in V(T) t ∈ V ( T ) . We define ( U t , W t ) : = ( U , W ) ∧ ( C t , D t ) (U_t,W_t) := (U,W) \wedge(C_t,D_t) ( U t , W t ) := ( U , W ) ∧ ( C t , D t ) , see Figure 6.4 .
Moreover, we let E t : = ⋃ t ′ ≤ t V t ′ E_t := \bigcup_{t' \le t} V_{t'} E t := ⋃ t ′ ≤ t V t ′ and F t : = ⋃ t ′ > t V t ′ F_t := \bigcup_{t' > t} V_{t'} F t := ⋃ t ′ > t V t ′ . Observe that for every t ∈ T t \in T t ∈ T , ( E t , F t ) (E_t,F_t) ( E t , F t ) is an ℓ \ell ℓ -separation such that ω 1 \omega_1 ω 1 lives in E t E_t E t and ω 2 \omega_2 ω 2 lives in F t F_t F t . In particular, ( E t , F t ) ∈ R (E_t,F_t) \in R ( E t , F t ) ∈ R .
Claim 6.21.1. ( U t , W t ) (U_t,W_t) ( U t , W t ) has order < 2 ℓ < 2\ell < 2 ℓ for every t ∈ V ( T ) t \in V(T) t ∈ V ( T ) .
Proof of the claim. Consider the separation s 1 : = ( E t − 1 , F t − 1 ) ∧ ( W , U ) s_1 := (E_{t-1}, F_{t-1}) \wedge(W, U) s 1 := ( E t − 1 , F t − 1 ) ∧ ( W , U ) and let S 1 S_1 S 1 be its separator, see Figure 6.4 . Note that s 1 s_1 s 1 distinguishes ω 1 \omega_1 ω 1 from ω 2 \omega_2 ω 2 , so ∣ S 1 ∣ ≥ ℓ |S_1| \ge\ell ∣ S 1 ∣ ≥ ℓ by Lemma 6.14 . Similarly, the separation s 2 : = ( E t , F t ) ∨ ( U , W ) s_2 := (E_t, F_t) \vee(U, W) s 2 := ( E t , F t ) ∨ ( U , W ) with separator S 2 S_2 S 2 (see again Figure 6.4 ) distinguishes ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 , so ∣ S 2 ∣ ≥ ℓ |S_2| \ge\ell ∣ S 2 ∣ ≥ ℓ by Lemma 6.14 .
Consider ( X t , Y t ) : = ( U , W ) ∨ ( C t , D t ) (X_t, Y_t) := (U, W) \vee(C_t, D_t) ( X t , Y t ) := ( U , W ) ∨ ( C t , D t ) . The separator of ( X t , Y t ) (X_t, Y_t) ( X t , Y t ) is the disjoint union of S 1 S_1 S 1 and S 2 S_2 S 2 , see Figure 6.4 . Note that the union is indeed disjoint since S 1 ⊆ E t − 1 S_1 \subseteq E_{t-1} S 1 ⊆ E t − 1 and S 2 ⊆ F t S_2 \subseteq F_t S 2 ⊆ F t , and E t − 1 ∩ F t = ∅ E_{t-1} \cap F_t = \varnothing E t − 1 ∩ F t = ∅ . Hence, ∣ X t ∩ Y t ∣ = ∣ S 1 ∣ + ∣ S 2 ∣ ≥ 2 ℓ |X_t \cap Y_t| = |S_1| + |S_2| \ge2\ell ∣ X t ∩ Y t ∣ = ∣ S 1 ∣ + ∣ S 2 ∣ ≥ 2 ℓ . By submodularity, we get
∣ U t ∩ W t ∣ = ∣ U ∩ W ∣ + ∣ C t ∩ D t ∣ − ∣ X t ∩ Y t ∣ ≤ ( 2 ℓ − 1 ) + 2 ℓ − 2 ℓ < 2 ℓ |U_t \cap W_t| = |U \cap W| + |C_t \cap D_t| - |X_t \cap Y_t| \le(2\ell- 1) + 2\ell- 2\ell< 2\ell ∣ U t ∩ W t ∣ = ∣ U ∩ W ∣ + ∣ C t ∩ D t ∣ − ∣ X t ∩ Y t ∣ ≤ ( 2 ℓ − 1 ) + 2 ℓ − 2 ℓ < 2 ℓ
which completes the proof. ♢
Claim 6.21.2. For every t ∈ V ( T ) t \in V(T) t ∈ V ( T ) we either have U t = ∅ U_t = \varnothing U t = ∅ or U ∩ W ∩ V t ≠ ∅ U \cap W \cap V_t \ne\varnothing U ∩ W ∩ V t = ∅ .
Proof of the claim. We assume that U ∩ W ∩ V t = ∅ U \cap W \cap V_t = \varnothing U ∩ W ∩ V t = ∅ and show that U t = ∅ U_t = \varnothing U t = ∅ . By construction of G ^ \hat{G} G ^ , we either have p ( V t ) ⊆ A p(V_t) \subseteq A p ( V t ) ⊆ A or p ( V t ) ⊆ B p(V_t) \subseteq B p ( V t ) ⊆ B , say the former. Since ( A , B ) (A, B) ( A , B ) is tight, there is a component K K K of G [ A ∖ B ] G[A \setminus B] G [ A ∖ B ] with N G ( K ) = A ∩ B N_G(K) = A \cap B N G ( K ) = A ∩ B . Let K ′ K' K ′ be the induced subgraph of G ^ [ V t ] \hat{G}[V_t] G ^ [ V t ] such that p ( K ′ ) = K p(K') = K p ( K ′ ) = K . Similarly, let P ′ , Q ′ ⊆ V t P', Q' \subseteq V_t P ′ , Q ′ ⊆ V t be such that p ( P ′ ) = P p(P') = P p ( P ′ ) = P and p ( Q ′ ) = Q p(Q') = Q p ( Q ′ ) = Q . See Figure 6.5 for a depiction of the situation. Since U t ∩ W t = ( U ∩ V t ) ∩ ( W ∪ D t ) U_t \cap W_t = (U \cap V_t) \cap(W \cup D_t) U t ∩ W t = ( U ∩ V t ) ∩ ( W ∪ D t ) and U ∩ W ∩ V t = ∅ U \cap W \cap V_t = \varnothing U ∩ W ∩ V t = ∅ , it follows that U t ∩ W t = U ∩ V t ∩ D t ⊆ P ′ ∪ Q ′ U_t \cap W_t = U \cap V_t \cap D_t \subseteq P' \cup Q' U t ∩ W t = U ∩ V t ∩ D t ⊆ P ′ ∪ Q ′ . Since N G ( K ) = A ∩ B N_G(K) = A \cap B N G ( K ) = A ∩ B , we have N G ^ ( K ′ ) = P ′ ∪ Q ′ N_{\hat{G}}(K') = P' \cup Q' N G ^ ( K ′ ) = P ′ ∪ Q ′ . Since U ∩ W ∩ V t = ∅ U \cap W \cap V_t = \varnothing U ∩ W ∩ V t = ∅ , we either have V ( K ′ ) ⊆ U ∖ W V(K') \subseteq U \setminus W V ( K ′ ) ⊆ U ∖ W or V ( K ′ ) ⊆ W ∖ U V(K') \subseteq W \setminus U V ( K ′ ) ⊆ W ∖ U .
Figure 6.5. The two possible locations of K ′ K' K ′ . There are no vertices in the crossed out box as it equals U ∩ W ∩ V t U \cap W \cap V_t U ∩ W ∩ V t .
Case V ( K ′ ) ⊆ U ∖ W V(K') \subseteq U \setminus W V ( K ′ ) ⊆ U ∖ W . Since N G ^ ( K ′ ) = P ′ ∪ Q ′ N_{\hat{G}}(K') = P' \cup Q' N G ^ ( K ′ ) = P ′ ∪ Q ′ , we have P ′ ∪ Q ′ ⊆ U P' \cup Q' \subseteq U P ′ ∪ Q ′ ⊆ U so P ′ ∪ Q ′ ⊆ U t ∩ W t P' \cup Q' \subseteq U_t \cap W_t P ′ ∪ Q ′ ⊆ U t ∩ W t . With ∣ P ′ ∣ = ∣ P ∣ = ℓ = ∣ Q ∣ = ∣ Q ′ ∣ |P'| = |P| = \ell= |Q| = |Q'| ∣ P ′ ∣ = ∣ P ∣ = ℓ = ∣ Q ∣ = ∣ Q ′ ∣ , we get ∣ U t ∩ W t ∣ ≥ 2 ℓ |U_t \cap W_t| \ge2\ell ∣ U t ∩ W t ∣ ≥ 2 ℓ , contradicting Claim 6.21.1 .
Case V ( K ′ ) ⊆ W ∖ U V(K') \subseteq W \setminus U V ( K ′ ) ⊆ W ∖ U . Since N G ^ ( K ′ ) = P ′ ∪ Q ′ N_{\hat{G}}(K') = P' \cup Q' N G ^ ( K ′ ) = P ′ ∪ Q ′ , we have P ′ ∪ Q ′ ⊆ W P' \cup Q' \subseteq W P ′ ∪ Q ′ ⊆ W . Since we also have P ′ ∪ Q ′ ⊆ V t P' \cup Q' \subseteq V_t P ′ ∪ Q ′ ⊆ V t and since U ∩ W ∩ V t = ∅ U \cap W \cap V_t = \varnothing U ∩ W ∩ V t = ∅ , it follows that P ′ ∪ Q ′ ⊆ W ∖ U P' \cup Q' \subseteq W \setminus U P ′ ∪ Q ′ ⊆ W ∖ U . Hence, U t ∩ W t = ∅ U_t \cap W_t = \varnothing U t ∩ W t = ∅ . Since G ^ [ V t ] \hat{G}[V_t] G ^ [ V t ] is connected, U t ∖ W t U_t \setminus W_t U t ∖ W t is empty. Hence, U t = ∅ U_t = \varnothing U t = ∅ . ♢
We define Y : = { t ∈ T ∣ U t ≠ ∅ } Y := \{t \in T \mid U_t \ne\varnothing\} Y := { t ∈ T ∣ U t = ∅ } , and we let Y ′ ⊆ Y Y' \subseteq Y Y ′ ⊆ Y contain every other t ∈ Y t \in Y t ∈ Y (along the double-ray T T T ). In particular, ∣ Y ∣ ≤ 2 ∣ Y ′ ∣ |Y| \le2|Y'| ∣ Y ∣ ≤ 2∣ Y ′ ∣ , and for every t ≠ t ′ ∈ Y ′ t \ne t' \in Y' t = t ′ ∈ Y ′ we have V t ∩ V t ′ = ∅ V_t \cap V_{t'} = \varnothing V t ∩ V t ′ = ∅ . Since G G G is ( q , k − 1 ) (q, k - 1) ( q , k − 1 ) -unbreakable, every ( < k ) (< k) ( < k ) -separation in G G G has a proper side with ≤ q \le q ≤ q vertices. Note that for every t ∈ V ( T ) t \in V(T) t ∈ V ( T ) , the covering map p p p is injective on U t U_t U t . Moreover, there is a (unique) separation ( U t ′ ∪ X , W t ′ ∪ X ) (U'_t \cup X, W'_t \cup X) ( U t ′ ∪ X , W t ′ ∪ X ) of G G G such that p ( U t ) = U t ′ p(U_t) = U'_t p ( U t ) = U t ′ and p ( U t ∩ W t ) = U t ′ ∩ W t ′ p(U_t \cap W_t) = U'_t \cap W'_t p ( U t ∩ W t ) = U t ′ ∩ W t ′ . Note that ( U t ′ ∪ X , W t ′ ∪ X ) (U'_t \cup X, W'_t \cup X) ( U t ′ ∪ X , W t ′ ∪ X ) has order < 2 ℓ + ∣ X ∣ ≤ k < 2\ell+ |X| \le k < 2 ℓ + ∣ X ∣ ≤ k . By symmetry, suppose that p ( V t ) ⊆ A p(V_t) \subseteq A p ( V t ) ⊆ A . Then B ∖ A ⊆ W t ′ ∖ U t ′ B \setminus A \subseteq W'_t \setminus U'_t B ∖ A ⊆ W t ′ ∖ U t ′ and ∣ B ∖ A ∣ > q |B \setminus A| > q ∣ B ∖ A ∣ > q , so ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakability yields ∣ U t ′ ∖ W t ′ ∣ ≤ q |U'_t \setminus W'_t| \le q ∣ U t ′ ∖ W t ′ ∣ ≤ q . Hence ∣ U t ∣ = ∣ U t ′ ∣ < q + 2 ℓ |U_t| = |U'_t| < q + 2\ell ∣ U t ∣ = ∣ U t ′ ∣ < q + 2 ℓ . Moreover, by Claim 6.21.2 there is an injection from Y ′ Y' Y ′ to U ∩ W U \cap W U ∩ W , e.g. by mapping each t ∈ Y ′ t \in Y' t ∈ Y ′ to a vertex in U ∩ W ∩ V t U \cap W \cap V_t U ∩ W ∩ V t . In particular, ∣ Y ′ ∣ ≤ ∣ U ∩ W ∣ < 2 ℓ |Y'| \le|U \cap W| < 2\ell ∣ Y ′ ∣ ≤ ∣ U ∩ W ∣ < 2 ℓ . Putting all inequalities together, we get
∣ U ∖ W ∣ ≤ ∣ ⋃ t ∈ Y U t ∣ ≤ ( q + 2 ℓ ) ⋅ ∣ Y ∣ ≤ ( q + 2 ℓ ) ⋅ 2 ∣ Y ′ ∣ ≤ ( q + 2 ℓ ) ⋅ 2 ∣ U ∩ W ∣ < 4 ℓ ( q + 2 ℓ ) = β 6.21 ( q , ℓ ) . |U \setminus W| \le\left|\bigcup_{t \in Y} U_t\right| \le(q+2\ell)\cdot|Y| \le(q+2\ell)\cdot2|Y'| \le(q+2\ell)\cdot2|U \cap W| < 4\ell(q+2\ell) = \beta_{6.21}(q,\ell). ∣ U ∖ W ∣ ≤ t ∈ Y ⋃ U t ≤ ( q + 2 ℓ ) ⋅ ∣ Y ∣ ≤ ( q + 2 ℓ ) ⋅ 2∣ Y ′ ∣ ≤ ( q + 2 ℓ ) ⋅ 2∣ U ∩ W ∣ < 4 ℓ ( q + 2 ℓ ) = β 6.21 ( q , ℓ ) .
We can now prove that cr R ( ⋅ ) \operatorname{cr}_R(\cdot) cr R ( ⋅ ) is bounded.
Lemma 6.22. Assume Setting 6.16. Let χ 6.22 ( q , ℓ ) : = χ 6.20 ( β 6.21 ( q , ℓ ) , ℓ ) \chi_{6.22}(q,\ell) := \chi_{6.20}(\beta_{6.21}(q,\ell),\ell) χ 6.22 ( q , ℓ ) := χ 6.20 ( β 6.21 ( q , ℓ ) , ℓ ) . Then cr R ( s ) ≤ χ 6.22 ( q , ℓ ) \operatorname{cr}_R(s) \le\chi_{6.22}(q,\ell) cr R ( s ) ≤ χ 6.22 ( q , ℓ ) for all s ∈ R s \in R s ∈ R .
Proof. By Lemma 6.20 , applied with β : = β 6.21 ( q , ℓ ) \beta:= \beta_{6.21}(q,\ell) β := β 6.21 ( q , ℓ ) , we have at least one of the following two cases.
Case Lemma 6.20 (i): cr R ( s ) ≤ χ 6.20 ( β , ℓ ) \operatorname{cr}_R(s) \le\chi_{6.20}(\beta,\ell) cr R ( s ) ≤ χ 6.20 ( β , ℓ ) . In this case we are done.
Case Lemma 6.20 (ii): there is a ( < 2 ℓ ) (<2\ell) ( < 2 ℓ ) -separation ( U , W ) (U,W) ( U , W ) of G ^ \widehat{G} G with ∣ U ∖ W ∣ ≥ β |U \setminus W| \ge\beta ∣ U ∖ W ∣ ≥ β and both ω i \omega_i ω i live in the side W W W . In particular, we have ∣ U ∖ W ∣ ≥ β 6.21 ( q , ℓ ) |U \setminus W| \ge\beta_{6.21}(q,\ell) ∣ U ∖ W ∣ ≥ β 6.21 ( q , ℓ ) , contradicting Lemma 6.21 .
We now prove that any two crossing separations in R R R cross with two small opposite corners.
Lemma 6.23. Assume Setting 6.16. Let ( A 1 , A 2 ) , ( B 1 , B 2 ) ∈ R (A_1,A_2),(B_1,B_2) \in R ( A 1 , A 2 ) , ( B 1 , B 2 ) ∈ R be two crossing separations. Then, the A 1 B 2 A_1B_2 A 1 B 2 -corner and the A 2 B 1 A_2B_1 A 2 B 1 -corner both have size less than β 6.21 ( q , ℓ ) \beta_{6.21}(q,\ell) β 6.21 ( q , ℓ ) .
Proof. Let s : = ( A 1 , A 2 ) s := (A_1,A_2) s := ( A 1 , A 2 ) and t : = ( B 1 , B 2 ) t := (B_1,B_2) t := ( B 1 , B 2 ) . Let ( U 1 , W 1 ) : = ( A 1 , A 2 ) ∧ ( B 2 , B 1 ) (U_1,W_1) := (A_1,A_2) \wedge(B_2,B_1) ( U 1 , W 1 ) := ( A 1 , A 2 ) ∧ ( B 2 , B 1 ) and ( U 2 , W 2 ) : = ( A 2 , A 1 ) ∧ ( B 1 , B 2 ) (U_2,W_2) := (A_2,A_1) \wedge(B_1,B_2) ( U 2 , W 2 ) := ( A 2 , A 1 ) ∧ ( B 1 , B 2 ) . This proof is supported by Figure 6.6 .
Claim 6.23.1. For i ∈ { 1 , 2 } i \in\{1,2\} i ∈ { 1 , 2 } , ( U i , W i ) (U_i,W_i) ( U i , W i ) is a ( < 2 ℓ ) (<2\ell) ( < 2 ℓ ) -separation of G ^ \widehat{G} G such that both ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 live in W i W_i W i .
Proof of the claim. By symmetry, it suffices to prove it for i = 1 i=1 i = 1 . First, recall that ω 1 \omega_1 ω 1 lives in B 1 B_1 B 1 and ω 2 \omega_2 ω 2 lives in A 2 A_2 A 2 . Hence, both ω i \omega_i ω i live in B 1 ∪ A 2 = W 1 B_1 \cup A_2 = W_1 B 1 ∪ A 2 = W 1 .
Now, consider the crossing separations ( A 1 , A 2 ) (A_1,A_2) ( A 1 , A 2 ) and ( B 1 , B 2 ) (B_1,B_2) ( B 1 , B 2 ) . Let P P P be the s B 1 sB_1 s B 1 -link, let Q Q Q be the s B 2 sB_2 s B 2 -link, let P ′ P' P ′ be the A 2 t A_2t A 2 t -link, let Q ′ Q' Q ′ be the A 1 t A_1t A 1 t -link and let Y Y Y be the s t st s t -centre. Recall that ∣ P ∣ + ∣ Q ∣ + ∣ Y ∣ = ℓ = ∣ P ′ ∣ + ∣ Q ′ ∣ + ∣ Y ∣ |P|+|Q|+|Y|=\ell=|P'|+|Q'|+|Y| ∣ P ∣ + ∣ Q ∣ + ∣ Y ∣ = ℓ = ∣ P ′ ∣ + ∣ Q ′ ∣ + ∣ Y ∣ . The order of ( U 1 , W 1 ) (U_1,W_1) ( U 1 , W 1 ) is
∣ U 1 ∩ W 1 ∣ = ∣ Q ∣ + ∣ Q ′ ∣ + ∣ Y ∣ = 2 ℓ − ∣ P ∣ − ∣ P ′ ∣ − ∣ Y ∣ . |U_1 \cap W_1| = |Q| + |Q'| + |Y| = 2\ell- |P| - |P'| - |Y|. ∣ U 1 ∩ W 1 ∣ = ∣ Q ∣ + ∣ Q ′ ∣ + ∣ Y ∣ = 2 ℓ − ∣ P ∣ − ∣ P ′ ∣ − ∣ Y ∣. Hence, if P P P or P ′ P' P ′ or Y Y Y is nonempty, then ( U 1 , W 1 ) (U_1,W_1) ( U 1 , W 1 ) has order < 2 ℓ <2\ell < 2 ℓ . Otherwise, P = P ′ = Y = ∅ P=P'=Y=\varnothing P = P ′ = Y = ∅ . Since G ^ \widehat{G} G is connected, the A 2 B 1 A_2B_1 A 2 B 1 -corner is empty. But then ( A 1 , A 2 ) (A_1,A_2) ( A 1 , A 2 ) and ( B 1 , B 2 ) (B_1,B_2) ( B 1 , B 2 ) are nested, a contradiction.
Thus, for every i ∈ { 1 , 2 } i \in\{1,2\} i ∈ { 1 , 2 } , we have ∣ U i ∖ W i ∣ < β 6.21 ( q , ℓ ) |U_i \setminus W_i| < \beta_{6.21}(q,\ell) ∣ U i ∖ W i ∣ < β 6.21 ( q , ℓ ) by Lemma 6.21 , which concludes the proof.
We are now ready to prove the main result of this section, that every separation in R R R is at bounded edit-distance from a separation in L L L .
Lemma 6.24. Assume Setting 6.16. Let β : = β 6.21 ( q , ℓ ) \beta:= \beta_{6.21}(q,\ell) β := β 6.21 ( q , ℓ ) and ε 6.24 ( q , ℓ ) : = ( β + ℓ ) ⋅ χ 6.22 ( q , ℓ ) \varepsilon_{6.24}(q,\ell) := (\beta+\ell)\cdot\chi_{6.22}(q,\ell) ε 6.24 ( q , ℓ ) := ( β + ℓ ) ⋅ χ 6.22 ( q , ℓ ) . Then for every r ∈ R r \in R r ∈ R there is s ( r ) ∈ L s(r) \in L s ( r ) ∈ L such that r r r and s ( r ) s(r) s ( r ) have edit-distance ≤ ( β + ℓ ) ⋅ cr R ( r ) ≤ ε 6.24 ( q , ℓ ) \le(\beta+\ell)\cdot\operatorname{cr}_R(r) \le\varepsilon_{6.24}(q,\ell) ≤ ( β + ℓ ) ⋅ cr R ( r ) ≤ ε 6.24 ( q , ℓ ) .
Figure 6.6. The situation in the proof of Lemma 6.23 .
Proof. We proceed by induction on cr R ( r ) \operatorname{cr}_{R}(r) cr R ( r ) .
If cr R ( r ) = 0 \operatorname{cr}_{R}(r)=0 cr R ( r ) = 0 , then r ∈ L r\in L r ∈ L . And if r ∈ L r\in L r ∈ L , then we are done (with s ( r ) : = r s(r):=r s ( r ) := r ). Hence we may assume now that cr R ( r ) > 0 \operatorname{cr}_{R}(r)>0 cr R ( r ) > 0 and that some t ∈ R t\in R t ∈ R witnesses r ∉ L r\notin L r ∈ / L , i.e. that t ∈ R t\in R t ∈ R crosses r r r with cr R ( t ) ≤ cr R ( r ) \operatorname{cr}_{R}(t)\le\operatorname{cr}_{R}(r) cr R ( t ) ≤ cr R ( r ) .
By Lemma 6.19 , there are two opposite corner-separations c , d c,d c , d of r , t r,t r , t that are both contained in R R R . Moreover, the lemma yields cr R ( c ) < cr R ( r ) \operatorname{cr}_{R}(c)<\operatorname{cr}_{R}(r) cr R ( c ) < cr R ( r ) , say. By the induction hypothesis, there is s ∈ L s\in L s ∈ L such that s s s and c c c have edit-distance ≤ ( β + ℓ ) ⋅ cr R ( c ) \le(\beta+\ell)\cdot\operatorname{cr}_{R}(c) ≤ ( β + ℓ ) ⋅ cr R ( c ) .
Claim 6.24.1. The separations r r r and c c c have edit-distance < β + ℓ <\beta+\ell < β + ℓ .
Proof of the claim. Write r = : ( A , B ) r=:(A,B) r =: ( A , B ) , t = : ( C , D ) t=:(C,D) t =: ( C , D ) and c = : ( A ∩ C , B ∪ D ) c=:(A\cap C,B\cup D) c =: ( A ∩ C , B ∪ D ) , say. Consider how r r r and t t t cross. By Lemma 6.23 , the A D AD A D -corner has size < β 6.21 ( q , ℓ ) = β <\beta_{6.21}(q,\ell)=\beta < β 6.21 ( q , ℓ ) = β . The vertex-set A △ ( A ∩ C ) = A ∖ C A\mathbin{\triangle}(A\cap C)=A\setminus C A △ ( A ∩ C ) = A ∖ C is equal to the union of the A D AD A D -corner and the r D rD r D -link. Thus, ∣ A △ ( A ∩ C ) ∣ < β + ℓ \lvert A\mathbin{\triangle}(A\cap C)\rvert<\beta+\ell ∣ A △ ( A ∩ C )∣ < β + ℓ . Similarly, B △ ( B ∪ D ) = D ∖ B B\mathbin{\triangle}(B\cup D)=D\setminus B B △ ( B ∪ D ) = D ∖ B is equal to the union of the A D AD A D -corner and the A t At A t -link. Thus, ∣ B △ ( B ∪ D ) ∣ < β + ℓ \lvert B\mathbin{\triangle}(B\cup D)\rvert<\beta+\ell ∣ B △ ( B ∪ D )∣ < β + ℓ . Hence, r r r and c c c have edit-distance < β + ℓ <\beta+\ell < β + ℓ .
The edit-distance between r r r and s s s is at most the edit-distance between r r r and c c c plus the edit-distance between c c c and s s s , so by Claim 6.24.1 and the induction hypothesis this is strictly less than
( β + ℓ ) + ( β + ℓ ) ⋅ cr R ( c ) ≤ ( β + ℓ ) ⋅ cr R ( r ) ≤ ( β + ℓ ) ⋅ χ 6.22 ( q , ℓ ) = ε 6.24 ( q , ℓ ) , (\beta+\ell)+(\beta+\ell)\cdot\operatorname{cr}_{R}(c)\le(\beta+\ell)\cdot\operatorname{cr}_{R}(r)\le(\beta+\ell)\cdot\chi_{6.22}(q,\ell)=\varepsilon_{6.24}(q,\ell), ( β + ℓ ) + ( β + ℓ ) ⋅ cr R ( c ) ≤ ( β + ℓ ) ⋅ cr R ( r ) ≤ ( β + ℓ ) ⋅ χ 6.22 ( q , ℓ ) = ε 6.24 ( q , ℓ ) ,
where we used Lemma 6.22 implicitly.
Constructing the graphic-daisy
In this section, we show that the separations in L L L induce a path-decomposition of G ^ \hat{G} G ^ , which projects to a cycle-decomposition of G − X G-X G − X . Then, we argue that adding X X X to all bags of this cycle-decomposition yields a graphic-k k k -daisy D \mathcal{D} D of G G G , such that every separation s ∈ S s\in S s ∈ S is at bounded edit-distance from a separation displayed by D \mathcal{D} D .
Recall that a deck transformation of the covering p : G ^ → G p:\hat{G}\to G p : G ^ → G from Setting 6.16 is an automorphism φ \varphi φ of G ^ \hat{G} G ^ that commutes with p p p in that p = p ∘ φ p=p\circ\varphi p = p ∘ φ . The deck transformations form a subgroup of the automorphism group of G ^ \hat{G} G ^ , called the deck transformation group .
Construction 6.25. Let N N N be a nested symmetric set of proper separations of a graph G G G . We define a graph T T T by taking the splitting stars of N N N as the nodes of T T T , and joining two nodes σ 1 ≠ σ 2 \sigma_1\ne\sigma_2 σ 1 = σ 2 by an edge in T T T if there is ( A 1 , A 2 ) ∈ N (A_1,A_2)\in N ( A 1 , A 2 ) ∈ N such that ( A 3 − i , A i ) ∈ σ i (A_{3-i},A_i)\in\sigma_i ( A 3 − i , A i ) ∈ σ i for both i = 1 , 2 i=1,2 i = 1 , 2 . For each node σ \sigma σ we put V σ : = ⋂ { B : ( A , B ) ∈ σ } V_\sigma:=\bigcap\{B:(A,B)\in\sigma\} V σ := ⋂ { B : ( A , B ) ∈ σ } . Write T ( N ) : = ( T , V ) \mathcal{T}(N):=(T,\mathcal{V}) T ( N ) := ( T , V ) .
The following lemma was known to [EKT22], who provide all the proof ingredients. In the proof, we rely on [46 ] for convenience.
Lemma 6.26 ([EKT22]). Let G G G be a locally finite connected graph and k ∈ N k \in\mathbb{N} k ∈ N . Let N N N be a nested symmetric set of proper (≤ k \le k ≤ k )-separations of G G G . If G G G is infinite, we additionally assume that all separations in N N N are tight.
T ( N ) T(N) T ( N ) is a tree-decomposition of G G G of adhesion at most k k k .
There is a natural order-preserving bijection between the (directed) edges of T T T and ( N , ≤ ) (N, \le) ( N , ≤ ) .
The torso of a splitting star σ \sigma σ of N N N equals the torso of V σ V_\sigma V σ when viewed as a bag of the tree-decomposition T ( N ) T(N) T ( N ) .
Proof. We combine the following results from [46 ] : the discussion at the beginning of §7.1, Lemma 7.2 and Lemma 7.5. □
Lemma 6.27. Assume Setting 6.16. The set L L L with the usual ordering on separations is order-isomorphic to Z \mathbb{Z} Z .
Proof. The definition of locally-R R R -optimal immediately yields that L L L is nested. Since each ( A , B ) ∈ L (A, B) \in L ( A , B ) ∈ L has ω 1 \omega_1 ω 1 living in A A A and ω 2 \omega_2 ω 2 living in B B B , we find that ( L , ≤ ) (L, \le) ( L , ≤ ) is a linear order. Note that all separations in L ∪ L ∗ L \cup L^* L ∪ L ∗ are tight since they efficiently distinguish ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 . By Lemma 6.26 , L ∪ L ∗ L \cup L^* L ∪ L ∗ uniquely determines a tree-decomposition ( H , W ) (H, \mathcal{W}) ( H , W ) of G ^ \widehat{G} G . Moreover, ( L ∪ L ∗ , ≤ ) (L \cup L^*, \le) ( L ∪ L ∗ , ≤ ) is order-isomorphic to the set of directed edges of H H H . As ( L , ≤ ) (L, \le) ( L , ≤ ) is a linear order, H H H must be a (possibly infinite) path.
We claim that H H H is a double-ray, and assume for a contradiction that it is not. Then ( L , ≤ ) (L, \le) ( L , ≤ ) has a maximal element ( U , W ) (U, W) ( U , W ) , say. Let v ∈ U ∩ W v \in U \cap W v ∈ U ∩ W be arbitrary. Consider the standard decomposition ( T , V ) (T, \mathcal{V}) ( T , V ) of G ^ \widehat{G} G . Since the adhesion-sets of ( T , V ) (T, \mathcal{V}) ( T , V ) are pairwise disjoint, and since ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 correspond to the two ends of the double-ray T T T , a distance argument shows that there is an edge t 1 t 2 t_1t_2 t 1 t 2 of T T T whose induced separation ( A 1 , A 2 ) ∈ R (A_1, A_2) \in R ( A 1 , A 2 ) ∈ R satisfies U ∩ W ⊆ A 1 ∖ A 2 U \cap W \subseteq A_1 \setminus A_2 U ∩ W ⊆ A 1 ∖ A 2 . Let v ′ ∈ A 2 ∖ A 1 v' \in A_2 \setminus A_1 v ′ ∈ A 2 ∖ A 1 such that p ( v ′ ) = p ( v ) p(v') = p(v) p ( v ′ ) = p ( v ) . Note that v ′ ∈ W ∖ U v' \in W \setminus U v ′ ∈ W ∖ U since U ∩ W U \cap W U ∩ W does not separate v ′ v' v ′ from ω 2 \omega_2 ω 2 . Let φ \varphi φ be a deck transformation of p p p such that φ ( v ) = v ′ \varphi(v) = v' φ ( v ) = v ′ . Since φ \varphi φ is a deck transformation, it fixes ω i \omega_i ω i for both i i i by the construction of p p p . Hence ( U ′ , W ′ ) : = ( φ ( U ) , φ ( W ) ) ∈ L (U', W') := (\varphi(U), \varphi(W)) \in L ( U ′ , W ′ ) := ( φ ( U ) , φ ( W )) ∈ L . Since ( L , ≤ ) (L, \le) ( L , ≤ ) is a linear order, and since v ′ ∈ U ′ ∩ W ′ v' \in U' \cap W' v ′ ∈ U ′ ∩ W ′ satisfies v ′ ∈ W ∖ U v' \in W \setminus U v ′ ∈ W ∖ U , we must have ( U , W ) < ( U ′ , W ′ ) (U, W) < (U', W') ( U , W ) < ( U ′ , W ′ ) . This contradicts the maximality of ( U , W ) (U, W) ( U , W ) in ( L , ≤ ) (L, \le) ( L , ≤ ) . □
Lemma 6.28. Assume Setting 6.16 with
γ > γ 6.28 ( q , k ) : = max { 2 ε 6.24 ( q , k ′ ) + 2 k ′ + k + β 6.21 ( q , k ′ ) : 0 ≤ k ′ ≤ k / 2 } . \gamma> \gamma_{6.28}(q,k) := \max\left\{ 2\varepsilon_{6.24}(q,k') + 2k' + k + \beta_{6.21}(q,k') : 0 \le k' \le k/2 \right\}. γ > γ 6.28 ( q , k ) := max { 2 ε 6.24 ( q , k ′ ) + 2 k ′ + k + β 6.21 ( q , k ′ ) : 0 ≤ k ′ ≤ k /2 } . The set L L L with the usual ordering on separations is order-isomorphic to Z \mathbb{Z} Z by Lemma 6.27 , which yields an enumeration L = { t i : i ∈ Z } L = \{t_i : i \in\mathbb{Z}\} L = { t i : i ∈ Z } . Then all of the following hold.
For every nontrivial deck transformation φ \varphi φ of p p p , we have that φ ( t i ) = t j \varphi(t_i) = t_j φ ( t i ) = t j implies ∣ i − j ∣ ≥ 4 |i-j| \ge4 ∣ i − j ∣ ≥ 4 .
There are s 0 , s 1 , s 2 , s 3 ∈ L s_0,s_1,s_2,s_3 \in L s 0 , s 1 , s 2 , s 3 ∈ L such that the following holds. Let Γ ⋅ s j = { ⋯ < s j − 1 < s j 0 < s j 1 < ⋯ } \Gamma\cdot s_j = \{\cdots< s_j^{-1} < s_j^0 < s_j^1 < \cdots\} Γ ⋅ s j = { ⋯ < s j − 1 < s j 0 < s j 1 < ⋯ } with s j 0 = s j s_j^0 = s_j s j 0 = s j , where Γ \Gamma Γ is the deck transformation group of p p p . Then we have s 0 0 < s 1 0 < s 2 0 < s 3 0 < s 0 1 s_0^0 < s_1^0 < s_2^0 < s_3^0 < s_0^1 s 0 0 < s 1 0 < s 2 0 < s 3 0 < s 0 1 for every i ∈ Z i \in\mathbb{Z} i ∈ Z , and the separators of the s j i s_j^i s j i are pairwise disjoint.
Proof. By the assumptions of Setting 6.16, the covering p p p arises from a separation x ∈ S x \in S x ∈ S that is γ \gamma γ -crossed by another separation y ∈ S y \in S y ∈ S . The four links of x , y x,y x , y determine a two-way infinite sequence of separations r k = ( X k , Y k ) ∈ R r_k=(X_k,Y_k)\in R r k = ( X k , Y k ) ∈ R such that
⋯ < r − 1 < r 0 < r 1 < ⋯ \cdots< r_{-1} < r_0 < r_1 < \cdots ⋯ < r − 1 < r 0 < r 1 < ⋯ and
for every nontrivial deck transformation φ \varphi φ of p p p , we have that φ ( r k ) = r ℓ \varphi(r_k)=r_\ell φ ( r k ) = r ℓ implies ∣ k − ℓ ∣ ≥ 4 \lvert k-\ell\rvert\ge4 ∣ k − ℓ ∣ ≥ 4 .
By Lemma 6.24, for every r k r_k r k there is an s k = ( U k , W k ) ∈ L s_k=(U_k,W_k)\in L s k = ( U k , W k ) ∈ L such that r k r_k r k and s k s_k s k have edit-distance ≤ ε 6.24 ( q , ℓ ) \le\varepsilon_{6.24}(q,\ell) ≤ ε 6.24 ( q , ℓ ) . We choose the s k s_k s k so that, for every deck transformation φ \varphi φ of p p p , we have that φ ( r i ) = r j \varphi(r_i)=r_j φ ( r i ) = r j implies φ ( s i ) = s j \varphi(s_i)=s_j φ ( s i ) = s j .
Claim 6.28.1. We have
⋯ < s − 1 < s 0 < s 1 < ⋯ \cdots< s_{-1} < s_0 < s_1 < \cdots ⋯ < s − 1 < s 0 < s 1 < ⋯ That is to say that r k ↦ s k r_k\mapsto s_k r k ↦ s k is an order-preserving injection { r k : k ∈ Z } → L \{r_k:k\in\mathbb{Z}\}\to L { r k : k ∈ Z } → L . Moreover, the separators of two consecutive s k s_k s k are disjoint.
Proof of the claim. For the first part, assume for a contradiction that there are i < j i<j i < j with s i ≮ s j s_i\not< s_j s i < s j . Since ( L , ≤ ) (L,\le) ( L , ≤ ) is a linear order by Lemma 6.27, we have s j ≤ s i s_j\le s_i s j ≤ s i . In particular, ( U j , W j ) ≤ ( U i , W i ) (U_j,W_j)\le(U_i,W_i) ( U j , W j ) ≤ ( U i , W i ) . The following calculation is supported by Figure 6.7 . We obtain
Figure 6.7. The separations ( X i , Y i ) ≤ ( X j , Y j ) (X_i,Y_i)\le(X_j,Y_j) ( X i , Y i ) ≤ ( X j , Y j ) and ( U j , W j ) ≤ ( U i , W i ) (U_j,W_j)\le(U_i,W_i) ( U j , W j ) ≤ ( U i , W i )
Y i ∩ X j ⊆ ( U i ∖ X i ) ∪ ( W j ∖ Y j ) ∪ ( X i ∩ Y i ) ∪ ( X j ∩ Y j ) . Y_i\cap X_j\subseteq(U_i\setminus X_i)\cup(W_j\setminus Y_j)\cup(X_i\cap Y_i)\cup(X_j\cap Y_j). Y i ∩ X j ⊆ ( U i ∖ X i ) ∪ ( W j ∖ Y j ) ∪ ( X i ∩ Y i ) ∪ ( X j ∩ Y j ) .
Recall ∣ X i ∩ Y i ∣ = ℓ = ∣ X j ∩ Y j ∣ \lvert X_i\cap Y_i\rvert=\ell=\lvert X_j\cap Y_j\rvert ∣ X i ∩ Y i ∣ = ℓ = ∣ X j ∩ Y j ∣ . Moreover, r k r_k r k and s k s_k s k having edit-distance ≤ ε 6.24 ( q , ℓ ) \le\varepsilon_{6.24}(q,\ell) ≤ ε 6.24 ( q , ℓ ) implies ∣ U i ∖ X i ∣ ≤ ε 6.24 ( q , ℓ ) \lvert U_i\setminus X_i\rvert\le\varepsilon_{6.24}(q,\ell) ∣ U i ∖ X i ∣ ≤ ε 6.24 ( q , ℓ ) and ∣ W j ∖ Y j ∣ ≤ ε 6.24 ( q , ℓ ) \lvert W_j\setminus Y_j\rvert\le\varepsilon_{6.24}(q,\ell) ∣ W j ∖ Y j ∣ ≤ ε 6.24 ( q , ℓ ) . Hence ∣ Y i ∩ X j ∣ ≤ 2 ε 6.24 ( q , ℓ ) + 2 ℓ \lvert Y_i\cap X_j\rvert\le2\varepsilon_{6.24}(q,\ell)+2\ell ∣ Y i ∩ X j ∣ ≤ 2 ε 6.24 ( q , ℓ ) + 2 ℓ . But then one of the four corners of x , y x,y x , y has size ≤ 2 ε 6.24 ( q , ℓ ) + 2 ℓ ≤ γ 6.28 ( q , k ) − k < γ − k \le2\varepsilon_{6.24}(q,\ell)+2\ell\le\gamma_{6.28}(q,k)-k<\gamma-k ≤ 2 ε 6.24 ( q , ℓ ) + 2 ℓ ≤ γ 6.28 ( q , k ) − k < γ − k , contradicting that x , y x,y x , y ( γ − k ) (\gamma-k) ( γ − k ) -criss by Observation 5.4.
For the ‘moreover’-part, assume for a contradiction that the separators of two consecutive s k s_k s k intersect, say s 0 s_0 s 0 and s 1 s_1 s 1 . Recall that s 0 < s 1 s_0<s_1 s 0 < s 1 . Consider the separation s ′ : = s 0 ∨ s 1 ∗ s' := s_0\vee s_1^* s ′ := s 0 ∨ s 1 ∗ , which is a separation of order < ∣ sep ( s 0 ) ∣ + ∣ sep ( s 1 ) ∣ = 2 ℓ <\lvert\operatorname{sep}(s_0)\rvert+\lvert\operatorname{sep}(s_1)\rvert=2\ell < ∣ sep ( s 0 )∣ + ∣ sep ( s 1 )∣ = 2 ℓ . Observe that both ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 live in the left side of s ′ s' s ′ . By Lemma 6.21 applied to s ′ s' s ′ , we have ∣ ( W 0 ∖ U 0 ) ∩ ( U 1 ∖ W 1 ) ∣ < β 6.21 ( q , ℓ ) \lvert(W_0\setminus U_0)\cap(U_1\setminus W_1)\rvert<\beta_{6.21}(q,\ell) ∣( W 0 ∖ U 0 ) ∩ ( U 1 ∖ W 1 )∣ < β 6.21 ( q , ℓ ) . Therefore ∣ W 0 ∩ U 1 ∣ < β 6.21 ( q , ℓ ) + 2 ℓ \lvert W_0\cap U_1\rvert<\beta_{6.21}(q,\ell)+2\ell ∣ W 0 ∩ U 1 ∣ < β 6.21 ( q , ℓ ) + 2 ℓ . Since s i s_i s i and r i r_i r i have edit-distance ≤ ε 6.24 ( q , ℓ ) \le\varepsilon_{6.24}(q,\ell) ≤ ε 6.24 ( q , ℓ ) , they are ε 6.24 ( q , ℓ ) \varepsilon_{6.24}(q,\ell) ε 6.24 ( q , ℓ ) -close by Observation 4.11. So, we have ∣ Y 0 ∩ X 1 ∣ < 2 ε 6.24 ( q , ℓ ) + β 6.21 ( q , ℓ ) + 2 ℓ < γ − k \lvert Y_0\cap X_1\rvert<2\varepsilon_{6.24}(q,\ell)+\beta_{6.21}(q,\ell)+2\ell<\gamma-k ∣ Y 0 ∩ X 1 ∣ < 2 ε 6.24 ( q , ℓ ) + β 6.21 ( q , ℓ ) + 2 ℓ < γ − k . But then one of the four corners of x , y x,y x , y has size < γ − k <\gamma-k < γ − k , contradicting that x , y x,y x , y ( γ − k ) (\gamma-k) ( γ − k ) -criss by Observation 5.4. ◇
Claim 6.28.2. For every nontrivial deck transformation φ \varphi φ of p p p , we have that φ ( s i ) = s j \varphi(s_i)=s_j φ ( s i ) = s j implies ∣ i − j ∣ ≥ 4 |i-j|\ge4 ∣ i − j ∣ ≥ 4 .
Proof of the claim. Assume that φ ( s i ) = s j \varphi(s_i)=s_j φ ( s i ) = s j . Since φ \varphi φ is a deck transformation, φ \varphi φ fixes the two ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 of G ^ \widehat{G} G . In particular, φ \varphi φ acts on G ^ \widehat{G} G by ‘shifting’ but without ‘reflection’ with respect to the two ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 . Hence there is a unique index k ∈ Z k\in\mathbb{Z} k ∈ Z such that r k = φ ( r i ) r_k=\varphi(r_i) r k = φ ( r i ) . Then s k = φ ( s i ) s_k=\varphi(s_i) s k = φ ( s i ) by the choice of s k s_k s k . Combining this with the assumption that φ ( s i ) = s j \varphi(s_i)=s_j φ ( s i ) = s j yields s k = s j s_k=s_j s k = s j . Then k = j k=j k = j by Claim 6.28.1 . Hence φ ( r i ) = r k = r j \varphi(r_i)=r_k=r_j φ ( r i ) = r k = r j . Therefore, ∣ i − j ∣ ≥ 4 |i-j|\ge4 ∣ i − j ∣ ≥ 4 by (ii). ⋄ \diamond ⋄
Finally, assume that some nontrivial deck transformation φ \varphi φ of p p p satisfies φ ( t i ) = t j \varphi(t_i)=t_j φ ( t i ) = t j for some i , j ∈ Z i,j\in\mathbb{Z} i , j ∈ Z . By Claim 6.28.1 there are unique indices I , J ∈ Z I,J\in\mathbb{Z} I , J ∈ Z such that
s I ≤ t i < s I + 1 and s J ≤ t j < s J + 1 . s_I\le t_i<s_{I+1}\quad\text{and}\quad s_J\le t_j<s_{J+1}. s I ≤ t i < s I + 1 and s J ≤ t j < s J + 1 .
Since L L L is canonical, φ ( t i ) = t j \varphi(t_i)=t_j φ ( t i ) = t j implies φ ( s I ) = s J \varphi(s_I)=s_J φ ( s I ) = s J . Then ∣ I − J ∣ ≥ 4 |I-J|\ge4 ∣ I − J ∣ ≥ 4 by Claim 6.28.2 . Therefore, ∣ i − j ∣ ≥ 4 |i-j|\ge4 ∣ i − j ∣ ≥ 4 as desired. This gives (i). (ii) follows from Claim 6.28.1 . □ \square □
Corollary 6.29. Assume Setting 6.16 with γ > γ 6.28 ( q , k ) \gamma>\gamma_{6.28}(q,k) γ > γ 6.28 ( q , k ) . Let r 1 < r 2 < r 3 ∈ R r_1<r_2<r_3\in\mathbb{R} r 1 < r 2 < r 3 ∈ R be induced by edges of the standard tree-decomposition of G ^ \widehat{G} G . Assume that there is a deck transformation φ \varphi φ of p p p such that φ ( r 1 ) = r 3 \varphi(r_1)=r_3 φ ( r 1 ) = r 3 . Then there are s 1 , s 2 , s 3 ∈ L s_1,s_2,s_3\in L s 1 , s 2 , s 3 ∈ L such that:
(i) r i r_i r i and s i s_i s i have edit-distance ≤ ε 6.24 ( q , ℓ ) \le\varepsilon_{6.24}(q,\ell) ≤ ε 6.24 ( q , ℓ ) for every i ∈ [ 3 ] i\in[3] i ∈ [ 3 ] ,
(ii) s 1 < s 2 < s 3 s_1<s_2<s_3 s 1 < s 2 < s 3 , and
(iii) φ ( s 1 ) = s 3 \varphi(s_1)=s_3 φ ( s 1 ) = s 3 .
Proof. We use Lemma 6.24 to find s 1 s_1 s 1 and s 2 s_2 s 2 such that r i r_i r i and s i s_i s i have edit-distance ≤ ε 6.24 \le\varepsilon_{6.24} ≤ ε 6.24 for i = 1 , 2 i=1,2 i = 1 , 2 . We let s 3 : = φ ( s 1 ) s_3:=\varphi(s_1) s 3 := φ ( s 1 ) . Since φ \varphi φ is an automorphism, we find that r 3 r_3 r 3 and s 3 s_3 s 3 also have edit-distance ≤ ε 6.24 \le\varepsilon_{6.24} ≤ ε 6.24 . It remains to show that s 1 < s 2 < s 3 s_1<s_2<s_3 s 1 < s 2 < s 3 . For this, we closely follow the proof of Claim 6.28.1 . □ \square □
Lemma 6.30. Assume Setting 6.16 with γ > γ 6.28 ( q , k ) \gamma>\gamma_{6.28}(q,k) γ > γ 6.28 ( q , k ) . Let ( T , V ) : = T ( L ∪ L ∗ ) (T,\mathcal{V}):=\mathcal{T}(L\cup L^*) ( T , V ) := T ( L ∪ L ∗ ) be the tree-decomposition of G ^ \widehat{G} G that L ∪ L ∗ L\cup L^* L ∪ L ∗ defines by Lemma 6.26 . Note that T T T is a double-ray by Lemma 6.27 . Let t ∈ V ( T ) t\in V(T) t ∈ V ( T ) with neighbouring edges e , f e,f e , f . Assume that p ( V e ) p(V_e) p ( V e ) and p ( V f ) p(V_f) p ( V f ) are disjoint. Then p ↾ V t p\restriction V_t p ↾ V t is injective.
Proof. Assume for a contradiction that there are two distinct vertices u , v ∈ V t u,v\in V_t u , v ∈ V t such that p ( u ) = p ( v ) = : w p(u)=p(v)=:w p ( u ) = p ( v ) =: w . Let Γ \Gamma Γ denote the group of deck transformations of p p p . Let φ ∈ Γ \varphi\in\Gamma φ ∈ Γ such that φ ( u ) = v \varphi(u)=v φ ( u ) = v . Since ( T , V ) (T,\mathcal{V}) ( T , V ) is canonical, there is t ′ ∈ V ( T ) t'\in V(T) t ′ ∈ V ( T ) such that φ ( V t ) = V t ′ \varphi(V_t)=V_{t'} φ ( V t ) = V t ′ . By Lemma 6.28 , we have t ′ ≠ t t'\ne t t ′ = t .
Since Γ \Gamma Γ acts transitively on the fibres of p p p , and since ( T , V ) (T,\mathcal{V}) ( T , V ) is canonical, we have that as soon as one of u , v u,v u , v lies in some adhesion-set of ( T , V ) (T,\mathcal{V}) ( T , V ) , then so does the other. Hence, it suffices to distinguish the following cases:
Case 1: u ∈ V e u\in V_e u ∈ V e and v ∈ V f v\in V_f v ∈ V f (or vice versa). This case is not possible, as p ( V e ) p(V_e) p ( V e ) and p ( V f ) p(V_f) p ( V f ) are disjoint by assumption.
Case 2: u , v ∈ V f u,v\in V_f u , v ∈ V f , say. Let e ′ , f ′ e',f' e ′ , f ′ be the two edges of T T T that are incident to t ′ t' t ′ . Since ( T , V ) (T,\mathcal{V}) ( T , V ) is canonical, the deck transformation φ \varphi φ acts on T T T , and we name e ′ , f ′ e',f' e ′ , f ′ so that φ \varphi φ moves e e e to e ′ e' e ′ and f f f to f ′ f' f ′ . By the construction of p p p , all deck transformations fix ω i \omega_i ω i for both i i i , so we have e < f ≤ e ′ < f ′ e < f \leq e' < f' e < f ≤ e ′ < f ′ in the order of the double-ray T T T , say. Now, on the one hand, we have v ∈ V f v \in V_f v ∈ V f by assumption, and on the other hand, u ∈ V f u \in V_f u ∈ V f implies φ ( u ) ∈ V f ′ \varphi(u) \in V_{f'} φ ( u ) ∈ V f ′ and, in particular, v ∈ V f ′ v \in V_{f'} v ∈ V f ′ . But then v ∈ V e ′ v \in V_{e'} v ∈ V e ′ as f ≤ e ′ < f ′ f \leq e' < f' f ≤ e ′ < f ′ . Hence u ∈ V e u \in V_{e} u ∈ V e , in addition to u ∈ V f u \in V_f u ∈ V f . This contradicts our assumption that p ( V e ) p(V_e) p ( V e ) and p ( V f ) p(V_f) p ( V f ) are disjoint.
Case 3: u , v ∈ V t ∖ ( V e ∪ V f ) u,v \in V_t \setminus(V_e \cup V_f) u , v ∈ V t ∖ ( V e ∪ V f ) . This case is impossible. Indeed, since v = φ ( u ) v = \varphi(u) v = φ ( u ) we immediately find that t ′ = t t' = t t ′ = t is the only option, contradicting that we have t ′ ≠ t t' \neq t t ′ = t .
Lemma 6.31. Let A , B , C , D A, B, C, D A , B , C , D be four sets. Then
( A ∩ B ) △ ( C ∩ D ) ⊆ ( A △ C ) ∪ ( B △ D ) . (A \cap B) \mathbin{\triangle} (C \cap D) \subseteq(A \mathbin{\triangle} C) \cup(B \mathbin{\triangle} D). ( A ∩ B ) △ ( C ∩ D ) ⊆ ( A △ C ) ∪ ( B △ D ) .
Proof. Let v ∈ ( A ∩ B ) △ ( C ∩ D ) v \in(A \cap B) \mathbin{\triangle} (C \cap D) v ∈ ( A ∩ B ) △ ( C ∩ D ) . By symmetry, we may assume without loss of generality that v ∈ ( A ∩ B ) ∖ ( C ∩ D ) v \in(A \cap B) \setminus(C \cap D) v ∈ ( A ∩ B ) ∖ ( C ∩ D ) . Then v ∉ C v \notin C v ∈ / C or v ∉ D v \notin D v ∈ / D . If v ∉ C v \notin C v ∈ / C , then v ∈ A △ C v \in A \mathbin{\triangle} C v ∈ A △ C and we are done. If v ∉ D v \notin D v ∈ / D , then v ∈ B △ D v \in B \mathbin{\triangle} D v ∈ B △ D and we are done. □ \square □
Lemma 6.32. Assume Setting 6.16 with γ > γ 6.28 ( q , k ) \gamma> \gamma_{6.28}(q,k) γ > γ 6.28 ( q , k ) . Let ε 6.32 ( q , ℓ ) : = 2 ε 6.24 ( q , ℓ ) \varepsilon_{6.32}(q,\ell) := 2\varepsilon_{6.24}(q,\ell) ε 6.32 ( q , ℓ ) := 2 ε 6.24 ( q , ℓ ) . Then G G G has a q q q -lean graphic-k k k -daisy D = ( O , P ′ ) \mathcal{D} = (O,\mathcal{P}') D = ( O , P ′ ) with the following properties.
(i) D \mathcal{D} D has centre X X X and has the clover property.
(ii) Every separation in S S S has edit-distance ≤ ε 6.32 \leq\varepsilon_{6.32} ≤ ε 6.32 to some displayed separation of D \mathcal{D} D .
(iii) For every ε ≥ ε 6.32 \varepsilon\geq\varepsilon_{6.32} ε ≥ ε 6.32 , every k k k -separation of G G G that is ε \varepsilon ε -compatible with D \mathcal{D} D is ε \varepsilon ε -nested with all the petal-separations of D \mathcal{D} D .
Proof. The set L L L uniquely determines a tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) of G ^ \widehat{G} G such that L ∪ L ∗ L \cup L^* L ∪ L ∗ equals the set of separations induced by the oriented edges of T T T , see Lemma 6.26 . The tree T T T is a double-ray whose ends correspond to the ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 of G ^ \widehat{G} G , respectively, by Lemma 6.27 . We orient the edges of T T T from ω 1 \omega_1 ω 1 to ω 2 \omega_2 ω 2 . As ( T , V ) (T,\mathcal{V}) ( T , V ) is canonical, it defines via the covering map p p p a graph-decomposition O = ( O , P ) O = (O,\mathcal{P}) O = ( O , P ) of G − X G-X G − X with family P = ( P o : o ∈ V ( O ) ) \mathcal{P} = (P_o : o \in V(O)) P = ( P o : o ∈ V ( O )) of bags as in [46 Construction 3.8 and Lemma 3.9] .
Claim 6.32.1. O O O is a graphic-2 ℓ 2\ell 2 ℓ -daisy of G − X G-X G − X , and O O O has four pairwise disjoint adhesion-sets.
Proof of the claim. First, we show that O O O is a 2 ℓ 2\ell 2 ℓ -flower of G − X G-X G − X . Since T T T is a double-ray, O O O is a cycle of length ≥ 4 \geq4 ≥ 4 by Lemma 6.28 . Let ( C , D ) (C,D) ( C , D ) be a 2-edge-cut of O O O with cross-edges c i d i c_i d_i c i d i so that c i ∈ C c_i \in C c i ∈ C and d i ∈ D d_i \in D d i ∈ D for i = 1 , 2 i = 1,2 i = 1 , 2 . Let ( C ′ , D ′ ) (C',D') ( C ′ , D ′ ) be the displayed separation of O O O that is induced by ( C , D ) (C,D) ( C , D ) . Then
C ′ ∩ D ′ = ⋃ i = 1 , 2 ( P c i ∩ P d i ) C' \cap D' = \bigcup_{i=1,2}(P_{c_i} \cap P_{d_i}) C ′ ∩ D ′ = i = 1 , 2 ⋃ ( P c i ∩ P d i )
by [46 Lemma 3.4] . Each intersection P c i ∩ P d i P_{c_i} \cap P_{d_i} P c i ∩ P d i equals the projection (with regard to p p p ) of some adhesion-set of ( T , V ) (T,\mathcal{V}) ( T , V ) . Since every adhesion-set of ( T , V ) (T,\mathcal{V}) ( T , V ) has size ℓ \ell ℓ , the separator C ′ ∩ D ′ C' \cap D' C ′ ∩ D ′ has size ≤ 2 ℓ \leq2\ell ≤ 2 ℓ . Hence, O O O is a 2 ℓ 2\ell 2 ℓ -flower.
Now we show that O O O is a graphic-2 ℓ 2\ell 2 ℓ -daisy. By Theorem 3.5 and since every 2 ℓ 2\ell 2 ℓ -daisy is a graphic-2 ℓ 2\ell 2 ℓ -daisy, it suffices to show that O O O is not a 2 ℓ 2\ell 2 ℓ -anemone. From Lemma 6.28 (ii) we get that O O O has four pairwise disjoint adhesion-sets P e 0 , … , P e 3 P_{e_0},\ldots,P_{e_3} P e 0 , … , P e 3 . Since O O O has four pairwise disjoint adhesion-sets, the union of the adhesion-sets of O \mathcal{O} O has size ≥ 4 ℓ > 2 ℓ \ge4\ell> 2\ell ≥ 4 ℓ > 2 ℓ . So O \mathcal{O} O is not a 2 ℓ 2\ell 2 ℓ -anemone, and hence O \mathcal{O} O is a graphic-2 ℓ 2\ell 2 ℓ -daisy, which completes the proof. ♢
For each P o ∈ P P_o \in\mathcal{P} P o ∈ P let P o ′ : = P o ∪ X P'_o := P_o \cup X P o ′ := P o ∪ X , and let P ′ : = ( P o ′ : o ∈ V ( O ) ) \mathcal{P}' := (P'_o : o \in V(\mathcal{O})) P ′ := ( P o ′ : o ∈ V ( O )) . Put D : = ( O , P ′ ) \mathcal{D} := (\mathcal{O},\mathcal{P}') D := ( O , P ′ ) . Then D \mathcal{D} D is a graphic-k k k -daisy of G G G as O \mathcal{O} O is a graphic-2 ℓ 2\ell 2 ℓ -daisy of G − X G-X G − X by Claim 6.32.1.
Claim 6.32.2. D \mathcal{D} D is q q q -lean.
Proof of the claim. For an edge e e e of O \mathcal{O} O , we denote by P e P_e P e the adhesion-set of e e e in O \mathcal{O} O , and by P e ′ P'_e P e ′ the adhesion-set of e e e in D \mathcal{D} D . Let Z : = ⋂ o ∈ V ( O ) P o ′ Z := \bigcap_{o\in V(\mathcal{O})} P'_o Z := ⋂ o ∈ V ( O ) P o ′ . Let s = ( A , B ) s=(A,B) s = ( A , B ) be a ( > q ) (>q) ( > q ) -broad separation of G G G displayed by a set of edges { e , f } \{e,f\} { e , f } of O \mathcal{O} O in D \mathcal{D} D . We need to show that the sets P e ′ ∖ Z P'_e\setminus Z P e ′ ∖ Z and P f ′ ∖ Z P'_f\setminus Z P f ′ ∖ Z are disjoint.
Since O \mathcal{O} O has four pairwise disjoint adhesion-sets by Claim 6.32.1, we have ⋂ o ∈ V ( O ) P o = ∅ \bigcap_{o\in V(\mathcal{O})}P_o=\emptyset ⋂ o ∈ V ( O ) P o = ∅ and therefore Z = ⋂ o ∈ V ( O ) P o ′ = X Z=\bigcap_{o\in V(\mathcal{O})}P'_o=X Z = ⋂ o ∈ V ( O ) P o ′ = X . Let k ′ k' k ′ be the order of s s s , and note that k ′ ≥ k k'\ge k k ′ ≥ k since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable. Since X ⊆ sep ( s ) X\subseteq\operatorname{sep}(s) X ⊆ sep ( s ) , we have that s ′ : = ( A ∖ X , B ∖ X ) s':=(A\setminus X,B\setminus X) s ′ := ( A ∖ X , B ∖ X ) is a ( k ′ − ∣ X ∣ ) (k'-|X|) ( k ′ − ∣ X ∣ ) -separation of G − X G-X G − X that is displayed by { e , f } \{e,f\} { e , f } in O \mathcal{O} O . In particular, the order of s ′ s' s ′ is k ′ − ∣ X ∣ ≥ k − ∣ X ∣ = 2 ℓ k'-|X|\ge k-|X|=2\ell k ′ − ∣ X ∣ ≥ k − ∣ X ∣ = 2 ℓ . Since every adhesion-set of O \mathcal{O} O has size exactly ℓ \ell ℓ , the adhesion-sets for e e e and f f f in O \mathcal{O} O are disjoint. Hence, P e = P e ′ ∖ Z P_e=P'_e\setminus Z P e = P e ′ ∖ Z and P f = P f ′ ∖ Z P_f=P'_f\setminus Z P f = P f ′ ∖ Z are disjoint. ♢
Claim 6.32.3. Every ( A , B ) ∈ S (A,B)\in\mathcal{S} ( A , B ) ∈ S has edit-distance ≤ ϵ 6.32 ( q , ℓ ) \le\epsilon_{6.32}(q,\ell) ≤ ϵ 6.32 ( q , ℓ ) to some displayed separation of D \mathcal{D} D .
Proof of the claim. Let ( A , B ) ∈ S (A,B)\in\mathcal{S} ( A , B ) ∈ S be arbitrary. By Lemma 6.15, we may assume without loss of generality that p p p is induced by ( A , B ) (A,B) ( A , B ) and a unique tripartition Q \mathcal{Q} Q of A ∩ B A\cap B A ∩ B . Let ( H , W ) (H,\mathcal{W}) ( H , W ) denote the standard tree-decomposition of G ^ \widehat{G} G (induced by ( A , B ) (A,B) ( A , B ) and Q \mathcal{Q} Q ). Recall that H H H is a double-ray whose two ends reflect the two ends ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 of G ^ \widehat{G} G . We consider the edges of H H H as oriented from ω 1 \omega_1 ω 1 to ω 2 \omega_2 ω 2 .
Now let e , e ′ , e ′ ′ e,e',e'' e , e ′ , e ′′ be three consecutive edges on the oriented double-ray H H H . Let h h h be the node of H H H shared by e e e and e ′ e' e ′ . Let h ′ h' h ′ be the node of H H H shared by e ′ e' e ′ and e ′ ′ e'' e ′′ .
Let ( X e , Y e ) ∈ R (X_e,Y_e)\in R ( X e , Y e ) ∈ R denote the separation induced by e e e and ( H , W ) (H,\mathcal{W}) ( H , W ) , and define ( X e ′ , Y e ′ ) (X_{e'},Y_{e'}) ( X e ′ , Y e ′ ) and ( X e ′ ′ , Y e ′ ′ ) (X_{e''},Y_{e''}) ( X e ′′ , Y e ′′ ) similarly. Then e < e ′ < e ′ ′ e<e'<e'' e < e ′ < e ′′ implies
( X e , Y e ) < ( X e ′ , Y e ′ ) < ( X e ′ ′ , Y e ′ ′ ) . (X_e,Y_e)<(X_{e'},Y_{e'})<(X_{e''},Y_{e''}). ( X e , Y e ) < ( X e ′ , Y e ′ ) < ( X e ′′ , Y e ′′ ) .
Moreover, there is a deck transformation φ \varphi φ of p p p such that
( φ ( X e ) , φ ( Y e ) ) = ( X e ′ ′ , Y e ′ ′ ) (\varphi(X_e),\varphi(Y_e))=(X_{e''},Y_{e''}) ( φ ( X e ) , φ ( Y e )) = ( X e ′′ , Y e ′′ )
by the construction of p p p . So by Corollary 6.29, there are three separations
( U f , W f ) < ( U f ′ , W f ′ ) < ( U f ′ ′ , W f ′ ′ ) (U_f,W_f)<(U_{f'},W_{f'})<(U_{f''},W_{f''}) ( U f , W f ) < ( U f ′ , W f ′ ) < ( U f ′′ , W f ′′ )
in L L L induced by oriented edges f < f ′ < f ′ ′ ∈ E ( T ) f<f'<f''\in E(T) f < f ′ < f ′′ ∈ E ( T ) such that ( X e , Y e ) (X_e,Y_e) ( X e , Y e ) and ( U f , W f ) (U_f,W_f) ( U f , W f ) have edit-distance ≤ ϵ 6.24 ( q , ℓ ) \le\epsilon_{6.24}(q,\ell) ≤ ϵ 6.24 ( q , ℓ ) and so on, and such that
( φ ( U f ) , φ ( W f ) ) = ( U f ′ ′ , W f ′ ′ ) . (\varphi(U_f),\varphi(W_f))=(U_{f''},W_{f''}). ( φ ( U f ) , φ ( W f )) = ( U f ′′ , W f ′′ ) .
Without loss of generality, we have A=p(W_h)\mathbin{\mathaccent\cdot\cup}X and B=p(W_{h'})\mathbin{\mathaccent\cdot\cup}X . Let C ^ : = f ˚ T f ˚ ′ ⊆ V ( T ) \widehat{C}:=\mathring{f}T\mathring{f}'\subseteq V(T) C := f ˚ T f ˚ ′ ⊆ V ( T ) and D ^ : = f ˚ ′ T f ˚ ′ ′ ⊆ V ( T ) \widehat{D}:=\mathring{f}'T\mathring{f}''\subseteq V(T) D := f ˚ ′ T f ˚ ′′ ⊆ V ( T ) . Let Γ \Gamma Γ denote the deck transformation group of p p p , which acts on T T T as ( T , V ) (T,\mathcal{V}) ( T , V ) is canonical. By construction, V ( O ) V(O) V ( O ) consists of the orbits of the Γ \Gamma Γ -action on V ( T ) V(T) V ( T ) . Let
C : = { Γ ⋅ t : t ∈ C ^ } ⊆ V ( O ) and D : = { Γ ⋅ t : t ∈ D ^ } ⊆ V ( O ) . C := \{ \Gamma\cdot t : t \in\widehat{C} \} \subseteq V(O) \qquad\text{and}\qquad D := \{ \Gamma\cdot t : t \in\widehat{D} \} \subseteq V(O). C := { Γ ⋅ t : t ∈ C } ⊆ V ( O ) and D := { Γ ⋅ t : t ∈ D } ⊆ V ( O ) .
Then ( C , D ) (C,D) ( C , D ) is a 2-edge-cut of O O O with the two cross-edges Γ ⋅ f = Γ ⋅ f ′ ′ \Gamma\cdot f = \Gamma\cdot f^{\prime\prime} Γ ⋅ f = Γ ⋅ f ′′ and Γ ⋅ f ′ \Gamma\cdot f^{\prime} Γ ⋅ f ′ . Let ( C ′ , D ′ ) (C^{\prime},D^{\prime}) ( C ′ , D ′ ) denote the displayed ( ≤ k ) (\leq k) ( ≤ k ) -separation of D \mathcal{D} D that corresponds to ( C , D ) (C,D) ( C , D ) . We have
B ∖ X = p ( W h ′ ) = p ( Y e ′ ∩ X e ′ ′ ) (3) B \setminus X = p(W_{h^{\prime}}) = p(Y_{e^{\prime}} \cap X_{e^{\prime\prime}}) \tag*{(3)} B ∖ X = p ( W h ′ ) = p ( Y e ′ ∩ X e ′′ ) (3)
and
D ′ ∖ X = ⋃ o ∈ D P o = p ( ⋃ d ∈ D ^ V d ) = p ( W f ′ ∩ U f ′ ′ ) . (4) D^{\prime} \setminus X = \bigcup_{o\in D} P_o = p\left(\bigcup_{d\in\widehat{D}} V_d\right) = p(W_{f^{\prime}} \cap U_{f^{\prime\prime}}). \tag*{(4)} D ′ ∖ X = o ∈ D ⋃ P o = p d ∈ D ⋃ V d = p ( W f ′ ∩ U f ′′ ) . (4)
Combining X ⊆ B ∩ D ′ X \subseteq B \cap D^{\prime} X ⊆ B ∩ D ′ with (3) and (4) , we find that
∣ B △ D ′ ∣ = ∣ ( B ∖ X ) △ ( D ′ ∖ X ) ∣ = ∣ p ( Y e ′ ∩ X e ′ ′ ) △ p ( W f ′ ∩ U f ′ ′ ) ∣ ≤ ∣ ( Y e ′ ∩ X e ′ ′ ) △ ( W f ′ ∩ U f ′ ′ ) ∣ ≤ ( ∗ ) ∣ Y e ′ △ W f ′ ∣ + ∣ X e ′ ′ △ U f ′ ′ ∣ ≤ 2 ε 6.24 ( q , ℓ ) = ε 6.32 ( q , ℓ ) \begin{aligned}
|B\triangle D^{\prime}| &= |(B\setminus X)\triangle(D^{\prime}\setminus X)| \\
&= |p(Y_{e^{\prime}}\cap X_{e^{\prime\prime}})\triangle p(W_{f^{\prime}}\cap U_{f^{\prime\prime}})| \\
&\leq|(Y_{e^{\prime}}\cap X_{e^{\prime\prime}})\triangle(W_{f^{\prime}}\cap U_{f^{\prime\prime}})| \\
&\overset{(*)}{\leq} |Y_{e^{\prime}}\triangle W_{f^{\prime}}| + |X_{e^{\prime\prime}}\triangle U_{f^{\prime\prime}}| \\
&\leq2\varepsilon_{6.24}(q,\ell)=\varepsilon_{6.32}(q,\ell)
\end{aligned} ∣ B △ D ′ ∣ = ∣ ( B ∖ X ) △ ( D ′ ∖ X ) ∣ = ∣ p ( Y e ′ ∩ X e ′′ ) △ p ( W f ′ ∩ U f ′′ ) ∣ ≤ ∣ ( Y e ′ ∩ X e ′′ ) △ ( W f ′ ∩ U f ′′ ) ∣ ≤ ( ∗ ) ∣ Y e ′ △ W f ′ ∣ + ∣ X e ′′ △ U f ′′ ∣ ≤ 2 ε 6.24 ( q , ℓ ) = ε 6.32 ( q , ℓ )
where we use Lemma 6.31 for ( ∗ ) (*) ( ∗ ) . A similar argument shows that ∣ A △ C ′ ∣ ≤ ε 6.32 ( q , ℓ ) |A\triangle C^{\prime}| \leq\varepsilon_{6.32}(q,\ell) ∣ A △ C ′ ∣ ≤ ε 6.32 ( q , ℓ ) . Hence ( A , B ) (A,B) ( A , B ) and ( C ′ , D ′ ) (C^{\prime},D^{\prime}) ( C ′ , D ′ ) have edit-distance ≤ ε 6.32 ( q , ℓ ) \leq\varepsilon_{6.32}(q,\ell) ≤ ε 6.32 ( q , ℓ ) . ♢
Claim 6.32.4. For every ε ≥ ε 6.24 \varepsilon\geq\varepsilon_{6.24} ε ≥ ε 6.24 , every k k k -separation ( A , B ) (A,B) ( A , B ) of G G G that is ε \varepsilon ε -compatible with D \mathcal{D} D is ε \varepsilon ε -nested with all the petal-separations of D \mathcal{D} D .
Proof of the claim. Let ε ≥ ε 6.24 \varepsilon\geq\varepsilon_{6.24} ε ≥ ε 6.24 and let r : = ( A , B ) r := (A,B) r := ( A , B ) be a k k k -separation of G G G that is ε \varepsilon ε -compatible with D \mathcal{D} D . Assume for a contradiction that r r r ε \varepsilon ε -crosses a petal-separation s : = ( U , W ) s := (U,W) s := ( U , W ) of D \mathcal{D} D where U = P o ′ U=P_o^{\prime} U = P o ′ for some node o ∈ O o\in O o ∈ O . Let o − o^{-} o − and o + o^{+} o + be the two neighbours of o o o on O O O . Abbreviate e : = o o − e:=oo^{-} e := o o − and f : = o o + f:=oo^{+} f := o o + . Recall that P e ′ P_e^{\prime} P e ′ and P f ′ P_f^{\prime} P f ′ denote the adhesion-sets of D \mathcal{D} D at e e e and f f f , respectively. Then, by the definition of ε \varepsilon ε -compatible with D \mathcal{D} D , we have all of the following:
(i) X = P e ′ ∩ P f ′ X=P_e^{\prime}\cap P_f^{\prime} X = P e ′ ∩ P f ′ and X X X is the centre of ( A , B ) (A,B) ( A , B ) and ( U , W ) (U,W) ( U , W ) ;
(ii) P e ′ ∖ X P_e^{\prime}\setminus X P e ′ ∖ X and P f ′ ∖ X P_f^{\prime}\setminus X P f ′ ∖ X are the A A A s-link and the B B B s-link, respectively (after possibly swapping the names of A A A and B B B ); and
(iii) all links of the two separations r , s r,s r , s are of the same size. Since X X X is the centre of ( A , B ) (A,B) ( A , B ) and ( U , W ) (U,W) ( U , W ) , all links have size ℓ \ell ℓ .
Let A ′ : = A ∩ P o A^{\prime}:=A\cap P_o A ′ := A ∩ P o and B ′ : = B ∩ P o B^{\prime}:=B\cap P_o B ′ := B ∩ P o . Then ( A ′ , B ′ ) (A^{\prime},B^{\prime}) ( A ′ , B ′ ) is an ℓ \ell ℓ -separation of G [ P o ] G[P_o] G [ P o ] . Since X = P e ′ ∩ P f ′ X=P_e^{\prime}\cap P_f^{\prime} X = P e ′ ∩ P f ′ , we have that P e = P e ′ ∖ X P_e=P_e^{\prime}\setminus X P e = P e ′ ∖ X is the A A A s-link and P f = P f ′ ∖ X P_f=P_f^{\prime}\setminus X P f = P f ′ ∖ X is the B B B s-link. Therefore, we have P e ⊆ A ′ ∖ B ′ P_e\subseteq A^{\prime}\setminus B^{\prime} P e ⊆ A ′ ∖ B ′ and P f ⊆ B ′ ∖ A ′ P_f\subseteq B^{\prime}\setminus A^{\prime} P f ⊆ B ′ ∖ A ′ on the one hand, and we have that P e P_e P e and P f P_f P f are disjoint on the other hand.
Let t ∈ V ( T ) t\in V(T) t ∈ V ( T ) such that o = Γ ⋅ t o=\Gamma\cdot t o = Γ ⋅ t , which exists by the construction of O O O (which gave rise to D \mathcal{D} D by taking the union of every bag with X X X ) from ( T , V ) (T,\mathcal{V}) ( T , V ) by folding via p p p . Let t − , t + t^{-},t^{+} t − , t + be the two
neighbours of t t t in T T T with o ± = Γ ⋅ t ± o^{\pm}=\Gamma\cdot t^{\pm} o ± = Γ ⋅ t ± . Then let e ^ : = { t − , t } \hat{e}:=\{t^{-},t\} e ^ := { t − , t } and f ^ : = { t , t + } \hat{f}:=\{t,t^{+}\} f ^ := { t , t + } . We assume without loss of generality that the edge e ^ \hat{e} e ^ is oriented from t − t^{-} t − to t t t in the fixed orientation of the edges of the double ray T T T (from ω 1 \omega_{1} ω 1 to ω 2 \omega_{2} ω 2 ), and similarly that f ^ \hat{f} f ^ is oriented from t t t to t + t^{+} t + . Note that P o = p ( V t ) P_{o}=p(V_{t}) P o = p ( V t ) . Let A ^ : = p − 1 ( A ′ ) ∩ V t \hat{A}:=p^{-1}(A')\cap V_{t} A ^ := p − 1 ( A ′ ) ∩ V t and B ^ : = p − 1 ( B ′ ) ∩ V t \hat{B}:=p^{-1}(B')\cap V_{t} B ^ := p − 1 ( B ′ ) ∩ V t . Then ( A ^ , B ^ ) (\hat{A},\hat{B}) ( A ^ , B ^ ) is an ℓ \ell ℓ -separation of G ^ [ V t ] \hat{G}[V_{t}] G ^ [ V t ] with V e ^ ⊆ A ^ ∖ B ^ V_{\hat{e}}\subseteq\hat{A}\setminus\hat{B} V e ^ ⊆ A ^ ∖ B ^ and V f ^ ⊆ B ^ ∖ A ^ V_{\hat{f}}\subseteq\hat{B}\setminus\hat{A} V f ^ ⊆ B ^ ∖ A ^ , as p ↾ V t p\mathbin{\upharpoonright}V_{t} p ↾ V t is injective by Lemma 6.30 . So there is a unique ℓ \ell ℓ -separation r ˉ = ( A ˉ , B ˉ ) ∈ R \bar{r}=(\bar{A},\bar{B})\in R r ˉ = ( A ˉ , B ˉ ) ∈ R with A ˉ ∩ V t = A ^ \bar{A}\cap V_{t}=\hat{A} A ˉ ∩ V t = A ^ and B ˉ ∩ V t = B ^ \bar{B}\cap V_{t}=\hat{B} B ˉ ∩ V t = B ^ .
By Lemma 6.24 , there is s ( r ˉ ) = ( C , D ) ∈ L s(\bar{r})=(C,D)\in L s ( r ˉ ) = ( C , D ) ∈ L such that r ˉ \bar{r} r ˉ and s ( r ˉ ) s(\bar{r}) s ( r ˉ ) have edit-distance ≤ ε 6.24 ( q , ℓ ) \leq\varepsilon_{6.24}(q,\ell) ≤ ε 6.24 ( q , ℓ ) . Since s ( r ˉ ) ∈ L s(\bar{r})\in L s ( r ˉ ) ∈ L , we have that s ( r ˉ ) s(\bar{r}) s ( r ˉ ) is induced in ( T , V ) (T,\mathcal{V}) ( T , V ) by an (oriented) edge g ^ \hat{g} g ^ of T T T . Note that g ^ \hat{g} g ^ is oriented from ω 1 \omega_{1} ω 1 to ω 2 \omega_{2} ω 2 , since r ˉ ∈ R \bar{r}\in R r ˉ ∈ R has ω 1 \omega_{1} ω 1 living on its left side and ω 2 \omega_{2} ω 2 on its right side, and since r ˉ \bar{r} r ˉ and s ( r ˉ ) s(\bar{r}) s ( r ˉ ) have finite edit-distance.
We claim that we may assume without loss of generality that g ^ = e ^ \hat{g}=\hat{e} g ^ = e ^ , so that s ( r ˉ ) ≤ ( A ˉ , B ˉ ) s(\bar{r})\leq(\bar{A},\bar{B}) s ( r ˉ ) ≤ ( A ˉ , B ˉ ) . Assume first that g ^ ∉ { e ^ , f ^ } \hat{g}\notin\{\hat{e},\hat{f}\} g ^ ∈ / { e ^ , f ^ } . Then e ^ \hat{e} e ^ lies in between g ^ \hat{g} g ^ and f ^ \hat{f} f ^ on T T T , say, so g ^ < e ^ < f ^ \hat{g}<\hat{e}<\hat{f} g ^ < e ^ < f ^ in the fixed orientation of the edges of T T T . Hence s ( r ˉ ) ≤ s ( e ^ ) ≤ r ˉ s(\bar{r})\leq s(\hat{e})\leq\bar{r} s ( r ˉ ) ≤ s ( e ^ ) ≤ r ˉ , where s ( e ^ ) ∈ L s(\hat{e})\in L s ( e ^ ) ∈ L is the separation induced in ( T , V ) (T,\mathcal{V}) ( T , V ) by the oriented edge e ^ \hat{e} e ^ . Hence r ˉ \bar{r} r ˉ and s ( e ^ ) s(\hat{e}) s ( e ^ ) have edit-distance ≤ ε 6.24 ( q , ℓ ) \leq\varepsilon_{6.24}(q,\ell) ≤ ε 6.24 ( q , ℓ ) , so we may replace g ^ \hat{g} g ^ with e ^ \hat{e} e ^ in this case. The other cases work analogously or by renaming. Hence we may assume now that g ^ = e ^ \hat{g}=\hat{e} g ^ = e ^ .
Since s ( r ˉ ) s(\bar{r}) s ( r ˉ ) and r ˉ \bar{r} r ˉ have edit-distance ≤ ε 6.24 \leq\varepsilon_{6.24} ≤ ε 6.24 , and since A ^ ∖ V e ^ ⊆ C △ A ˉ \hat{A}\setminus V_{\hat{e}}\subseteq C\mathbin{\triangle}\bar{A} A ^ ∖ V e ^ ⊆ C △ A ˉ , we have ∣ A ^ ∖ V e ^ ∣ ≤ ε 6.24 |\hat{A}\setminus V_{\hat{e}}|\leq\varepsilon_{6.24} ∣ A ^ ∖ V e ^ ∣ ≤ ε 6.24 . Since P o = p ( V t ) P_{o}=p(V_{t}) P o = p ( V t ) and A ′ = p ( A ^ ) A'=p(\hat{A}) A ′ = p ( A ^ ) , we have ∣ A ′ ∖ P e ∣ ≤ ε 6.24 |A'\setminus P_{e}|\leq\varepsilon_{6.24} ∣ A ′ ∖ P e ∣ ≤ ε 6.24 . Therefore, the A U AU A U -corner (for r r r and ( U , W ) (U,W) ( U , W ) ) has size ≤ ε 6.24 − ℓ \leq\varepsilon_{6.24}-\ell ≤ ε 6.24 − ℓ (where we deduct ℓ \ell ℓ for the r U rU r U -link). But r r r and s = ( U , W ) s=(U,W) s = ( U , W ) ε \varepsilon ε -cross by assumption with all links of size ℓ \ell ℓ , so they ( ε − ℓ ) (\varepsilon-\ell) ( ε − ℓ ) -criss by Observation 5.4, with ε ≥ ε 6.24 \varepsilon\geq\varepsilon_{6.24} ε ≥ ε 6.24 , a contradiction. ◇
Claim 6.32.1 gives (i). Claim 6.32.3 gives (ii). Claim 6.32.4 implies (iii) since ε 6.32 ≥ ε 6.24 \varepsilon_{6.32}\geq\varepsilon_{6.24} ε 6.32 ≥ ε 6.24 . □
Remark 6.33. The graphic-k k k -daisy D = D ( S ) D=D(S) D = D ( S ) constructed in Lemma 6.32 is canonical: whenever ψ \psi ψ is an automorphism of G G G , the graphic-k k k -daisies D ( S ) D(S) D ( S ) and D ( ψ ( S ) ) D(\psi(S)) D ( ψ ( S )) are ψ \psi ψ -isomorphic.
We can now conclude the proof of Key Lemma 6.6, which we restate for convenience.
Key Lemma 6.6 (Flower Approximation Lemma). For all q , k ∈ N q,k\in\mathbb{N} q , k ∈ N , there exist constants β 6.6 ( q , k ) ≥ q \beta_{6.6}(q,k)\geq q β 6.6 ( q , k ) ≥ q , γ 6.6 ( q , k ) \gamma_{6.6}(q,k) γ 6.6 ( q , k ) and ε 6.6 ( q , k ) \varepsilon_{6.6}(q,k) ε 6.6 ( q , k ) such that the following holds. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and γ > γ 6.6 \gamma>\gamma_{6.6} γ > γ 6.6 . Let S S S be a nontrivial γ \gamma γ -asteriskal set of separations of G G G that consists of β 6.6 \beta_{6.6} β 6.6 -broad narrow k k k -separations of G G G . Then there is a q q q -lean graphic-k k k -flower F : = F 6.6 ( S ) F:=F_{6.6}(S) F := F 6.6 ( S ) of G G G such that:
(i) every separation in S S S has edit-distance ≤ ε 6.6 \leq\varepsilon_{6.6} ≤ ε 6.6 to some displayed separation of F F F ;
(ii) for every ε ≥ ε 6.6 \varepsilon\geq\varepsilon_{6.6} ε ≥ ε 6.6 , every k k k -separation of G G G that is ε \varepsilon ε -compatible with F F F is ε \varepsilon ε -nested with all the petal-separations of F F F ;
(iii) if F F F is a graphic-k k k -daisy, then F F F has the clover-property; and
(iv) The graphic-k k k -flower F 6.6 ( S ) F_{6.6}(S) F 6.6 ( S ) is canonical: whenever ψ \psi ψ is an automorphism of G G G , the graphic-k k k -flowers F 6.6 ( S ) F_{6.6}(S) F 6.6 ( S ) and F 6.6 ( ψ ( S ) ) F_{6.6}(\psi(S)) F 6.6 ( ψ ( S )) are ψ \psi ψ -isomorphic.
Proof of Key Lemma 6.6. The γ \gamma γ -asteriskal set of separations S S S has a common centre X X X by Lemma 6.7 , and it has a common link size ℓ \ell ℓ by Lemma 6.8. If ∣ X ∣ = k |X| = k ∣ X ∣ = k , then we apply Lemma 6.10 and Remark 6.11 to get a q q q -lean graphic-k k k -anemone A A A that displays every separation in S S S and such that every separation that is ε ≥ ε 6.6 \varepsilon\ge\varepsilon_{6.6} ε ≥ ε 6.6 -compatible with A A A is ε \varepsilon ε -nested with all the petal-separations of A A A .
If ∣ X ∣ < k |X| < k ∣ X ∣ < k , then we may assume Setting 6.16 and are done by Lemma 6.32 and Remark 6.33. □
Petal-separations are compatible
The aim of this section is to prove Key Lemma 7.1. In § 11 , we will apply this lemma to deduce that the petal-separations of the graphic-k k k -flowers constructed in Key Lemma 6.6 are totally-γ ′ \gamma' γ ′ -nested.
Let S , S ′ S,S' S , S ′ be two sets of separations of a graph G G G . We say that S S S ε \varepsilon ε -approximates S ′ S' S ′ if every ε \varepsilon ε -broad separation of S ′ S' S ′ is ε \varepsilon ε -close to a separation in S S S . A graphic-k k k -flower F F F of a graph G G G ε \varepsilon ε -approximates a set S S S of separations of G G G if the set of displayed separations of F F F ε \varepsilon ε -approximates S S S .
Key Lemma 7.1. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N and
γ ≥ γ 7.1 ( q , k , ε ) : = 2 q + 3 k + 2 ε + 1 and γ ′ ≥ γ 7.1 ′ ( q , k ) : = 3 q + 3 k + 1. \gamma\ge\gamma_{7.1}(q,k,\varepsilon) := 2q + 3k + 2\varepsilon+ 1 \quad\text{and} \quad\gamma' \ge\gamma'_{7.1}(q,k) := 3q + 3k + 1. γ ≥ γ 7.1 ( q , k , ε ) := 2 q + 3 k + 2 ε + 1 and γ ′ ≥ γ 7.1 ′ ( q , k ) := 3 q + 3 k + 1. Let A A A be a nontrivial γ \gamma γ -asteriskal set of k k k -separations of G G G . Let F F F be a q q q -lean graphic-k k k -flower of G G G that ε \varepsilon ε -approximates A A A . Then every k k k -separation of G G G is γ ′ \gamma' γ ′ -compatible with F F F .
We now give a high-level sketch of the proof. We focus here on the case where F F F is a graphic-daisy, which is the most difficult. Let r r r be a k k k -separation of G G G , and let p p p be a petal-separation of F F F that γ ′ \gamma' γ ′ -crosses r r r . Let e , f e,f e , f be the edges of O O O such that p p p is displayed by { e , f } \{e,f\} { e , f } . Since p p p and r r r γ ′ \gamma' γ ′ -cross, we have that p p p is 2 ( q + k + 1 ) 2(q+k+1) 2 ( q + k + 1 ) -broad (Lemma 5.6). Since A A A is nontrivial and γ \gamma γ -asteriskal and since F F F ε \varepsilon ε -approximates A A A , there are displayed separations s 1 , s 2 s_1,s_2 s 1 , s 2 of F F F which are ( 2 q + 2 k + 1 ) (2q+2k+1) ( 2 q + 2 k + 1 ) -crossing such that p = s 1 ∧ s 2 p=s_1\wedge s_2 p = s 1 ∧ s 2 holds (see Lemma 7.3 ). We then study the 3D-corner diagram of the separations s 1 , s 2 s_1,s_2 s 1 , s 2 and r r r to deduce that the p r pr p r -centre is the s 1 s 2 s_1s_2 s 1 s 2 -centre, and that the p p p -links are links of the crossing separations s 1 s_1 s 1 and s 2 s_2 s 2 (see Lemma 7.6 ). Finally, using that F F F is q q q -lean, we get that the s 1 s 2 s_1s_2 s 1 s 2 -centre is Z Z Z , and that the two aforementioned links are V e ∖ V f V_e \setminus V_f V e ∖ V f and V f ∖ V e V_f \setminus V_e V f ∖ V e (see Lemma 7.8 ).
Lemma 7.2. Let γ ≥ 2 ε \gamma\ge2\varepsilon γ ≥ 2 ε and let s s s and t t t be two γ \gamma γ -crossing separations of a graph G G G . Let s ′ s' s ′ and t ′ t' t ′ be two separations of G G G that are ε \varepsilon ε -close to s s s and t t t , respectively. Then s ′ s' s ′ and t ′ t' t ′ ( γ − 2 ε ) (\gamma-2\varepsilon) ( γ − 2 ε ) -cross.
Proof. Let γ ′ = γ − 2 ε \gamma'=\gamma-2\varepsilon γ ′ = γ − 2 ε , and suppose by contradiction that s ′ s' s ′ and t ′ t' t ′ are γ ′ \gamma' γ ′ -nested. Without loss of generality, assume that s ′ ≤ γ t ′ s' \le_\gamma t' s ′ ≤ γ t ′ . Then s ≈ ε s ′ ≤ γ ′ t ′ ≈ ε t s \approx_\varepsilon s' \le_{\gamma'} t' \approx_\varepsilon t s ≈ ε s ′ ≤ γ ′ t ′ ≈ ε t , so we have s ≤ γ ′ + 2 ε t s \le_{\gamma'+2\varepsilon} t s ≤ γ ′ + 2 ε t by Lemma 4.9 , which is a contradiction. □
Lemma 7.3. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, let β > q \beta>q β > q and γ ≥ q + k \gamma\ge q+k γ ≥ q + k . Let F F F be a graphic-k k k -flower of G G G and let s s s and t t t be two displayed separations of F F F that γ \gamma γ -cross. Let p p p be a β \beta β -broad petal-separation of F F F . Then there are two ( γ − k ) (\gamma-k) ( γ − k ) -broad displayed separations s ′ s' s ′ and t ′ t' t ′ of F F F that min { β − 1 , γ − k } \min\{\beta-1,\gamma-k\} min { β − 1 , γ − k } -cross, and such that p = s ′ ∧ t ′ p=s'\wedge t' p = s ′ ∧ t ′ .
Proof. By Lemma 3.12, there is a cycle-decomposition O = ( O , V ) \mathcal{O}=(O,\mathcal{V}) O = ( O , V ) of G G G together with two 2-edge-cuts of O O O that induce s s s and t t t in G G G , respectively. Without loss of generality, we can assume that O = Z ℓ O=Z_\ell O = Z ℓ , where ℓ = ∣ O ∣ \ell= |O| ℓ = ∣ O ∣ , and that
p = ( V 0 , ⋃ i ≠ 0 V i ) p = \left(V_0,\bigcup_{i\ne0} V_i\right) p = V 0 , i = 0 ⋃ V i
and
s = ( A , B ) = ( ⋃ i = i 2 i 1 − 1 V i , ⋃ i = i 1 i 2 − 1 V i ) s = (A,B) = \left(\bigcup_{i=i_2}^{i_1-1} V_i,\bigcup_{i=i_1}^{i_2-1} V_i\right) s = ( A , B ) = ( i = i 2 ⋃ i 1 − 1 V i , i = i 1 ⋃ i 2 − 1 V i )
and
t = ( C , D ) = ( ⋃ i = j 2 j 1 − 1 V i , ⋃ i = j 1 j 2 − 1 V i ) . t = (C,D) = \left(\bigcup_{i=j_2}^{j_1-1} V_i,\bigcup_{i=j_1}^{j_2-1} V_i\right). t = ( C , D ) = ( i = j 2 ⋃ j 1 − 1 V i , i = j 1 ⋃ j 2 − 1 V i ) .
with 0 < i 1 < j 1 < i 2 < j 2 ≤ ℓ 0 < i_1 < j_1 < i_2 < j_2 \le\ell 0 < i 1 < j 1 < i 2 < j 2 ≤ ℓ . Then we define
s ′ : = ( A ′ , B ′ ) : = ( ⋃ i = i 2 ℓ V i , ⋃ i = 1 i 2 − 1 V i ) s' := (A',B') := \left(\bigcup_{i=i_2}^{\ell} V_i,\bigcup_{i=1}^{i_2-1} V_i\right) s ′ := ( A ′ , B ′ ) := ( i = i 2 ⋃ ℓ V i , i = 1 ⋃ i 2 − 1 V i )
and
t ′ : = ( C ′ , D ′ ) : = ( ⋃ i = 0 j 1 − 1 V i , ⋃ i = j 1 ℓ − 1 V i ) . t' := (C',D') := \left(\bigcup_{i=0}^{j_1-1} V_i,\bigcup_{i=j_1}^{\ell-1} V_i\right). t ′ := ( C ′ , D ′ ) := ( i = 0 ⋃ j 1 − 1 V i , i = j 1 ⋃ ℓ − 1 V i ) .
Since p p p is β \beta β -broad, the A ′ C ′ A'C' A ′ C ′ -corner has size ≥ β \ge\beta ≥ β . The B C BC B C -corner has size > γ − k > \gamma-k > γ − k by Observation 5.4 (ii) and is contained in the B ′ C ′ B'C' B ′ C ′ -corner. Similarly, the A D AD A D -corner has size > γ − k > \gamma-k > γ − k by Observation 5.4 (ii) and is contained in the A ′ D ′ A'D' A ′ D ′ -corner. Moreover, the B D BD B D -corner is equal to the B ′ D ′ B'D' B ′ D ′ -corner, which has size > γ − k > \gamma-k > γ − k by Observation 5.4 (ii). Thus, all four corners of s ′ s' s ′ and t ′ t' t ′ have size > min { β − 1 , γ − k } > \min\{\beta-1,\gamma-k\} > min { β − 1 , γ − k } , so s ′ s' s ′ and t ′ t' t ′ min { β − 1 , γ − k } \min\{\beta-1,\gamma-k\} min { β − 1 , γ − k } -cross by Observation 5.4 (iv). In particular, all four corners of s ′ s' s ′ and t ′ t' t ′ have size > q >q > q . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, the corresponding corner separations all have order at least k k k . Since s ′ s' s ′ and t ′ t' t ′ both have order at most k k k , submodularity implies that all four corner separations have order exactly k k k . Moreover, p p p has order k k k since it is β \beta β -broad with β > q \beta>q β > q . The separations p p p and s ′ ∧ t ′ s'\wedge t' s ′ ∧ t ′ have the same second side, and the first side of p p p is contained in that of s ′ ∧ t ′ s'\wedge t' s ′ ∧ t ′ . This containment cannot be strict, otherwise p p p and s ′ ∧ t ′ s'\wedge t' s ′ ∧ t ′ would not have the same order. Thus, p = s ′ ∧ t ′ p=s'\wedge t' p = s ′ ∧ t ′ .
Finally, since the B ′ C ′ B'C' B ′ C ′ -corner and the A ′ D ′ A'D' A ′ D ′ -corner have size > γ − k >\gamma-k > γ − k , both s ′ s' s ′ and t ′ t' t ′ are ( γ − k ) (\gamma-k) ( γ − k ) -broad.
Lemma 7.4. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let s 1 = ( A 1 , B 1 ) s_1=(A_1,B_1) s 1 = ( A 1 , B 1 ) , s 2 = ( A 2 , B 2 ) s_2=(A_2,B_2) s 2 = ( A 2 , B 2 ) be two ( q + k + 1 ) (q+k+1) ( q + k + 1 ) -crossing k k k -separations of G G G . Denote by L i L_i L i the A i s 3 − i A_i s_{3-i} A i s 3 − i -link, let X X X denote the s 1 s 2 s_1s_2 s 1 s 2 -centre and let p = s 1 ∧ s 2 p=s_1\wedge s_2 p = s 1 ∧ s 2 . Let r = ( C , D ) r=(C,D) r = ( C , D ) be a k k k -separation of G G G such that the A 1 B 2 D A_1B_2D A 1 B 2 D -corner, the A 1 A 2 D A_1A_2D A 1 A 2 D -corner, the A 1 A 2 C A_1A_2C A 1 A 2 C -corner and the B 1 A 2 C B_1A_2C B 1 A 2 C -corner each have size > q >q > q . Then r r r induces ( X , { L 1 , L 2 } ) (X,\{L_1,L_2\}) ( X , { L 1 , L 2 }) on sep ( p ) \operatorname{sep}(p) sep ( p ) .
Proof. First, we show that the prerequisites for applying Corollary 5.13 are fulfilled for the separations s 1 s_1 s 1 , s 2 s_2 s 2 and r r r . Note that the 3D corner-separations A 1 B 2 D A_1B_2D A 1 B 2 D , A 1 A 2 D A_1A_2D A 1 A 2 D , A 1 A 2 C A_1A_2C A 1 A 2 C and B 1 A 2 C B_1A_2C B 1 A 2 C have order ≥ k \ge k ≥ k since the corresponding corners each have size > q >q > q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable. Now consider the two separations s 1 s_1 s 1 and r r r . Then the A 1 D A_1D A 1 D -corner and the B 1 C B_1C B 1 C -corner both have size > q >q > q . Hence, since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and by submodularity, the corresponding corner-separations have order exactly k k k . Similarly, the A 2 C A_2C A 2 C -corner-separation and the B 2 D B_2D B 2 D -corner-separation of the separations s 2 s_2 s 2 and r r r both have order exactly k k k . Now, consider the two separations s 1 s_1 s 1 and s 2 s_2 s 2 . Since they are ( q + k + 1 ) (q+k+1) ( q + k + 1 ) -crossing, each (2D) corner has size > q >q > q by Observation 5.4 (ii). By submodularity and since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, each 2D corner-separation of s 1 s_1 s 1 and s 2 s_2 s 2 has order exactly k k k . In particular, the 2D corner-separation for A 1 A 2 A_1A_2 A 1 A 2 has order exactly k k k .
So we can apply Corollary 5.13 to s 1 s_1 s 1 , s 2 s_2 s 2 and r r r . Then, we get that the A 1 s 2 r A_1s_2r A 1 s 2 r -link, the s 1 A 2 r s_1A_2r s 1 A 2 r -link, the s 1 s 2 C s_1s_2C s 1 s 2 C -link, the s 1 s 2 D s_1s_2D s 1 s 2 D -link, the A 1 s 2 C A_1s_2C A 1 s 2 C -link and the s 1 A 2 D s_1A_2D s 1 A 2 D -link are empty. Then, L 1 = ( A 2 ∩ B 2 ) ∖ B 1 L_1=(A_2\cap B_2)\setminus B_1 L 1 = ( A 2 ∩ B 2 ) ∖ B 1 is the union of the A 1 s 2 r A_1s_2r A 1 s 2 r -link, A 1 s 2 C A_1s_2C A 1 s 2 C -link and A 1 s 2 D A_1s_2D A 1 s 2 D -link, so L 1 L_1 L 1 is equal to the A 1 s 2 D A_1s_2D A 1 s 2 D -link and thus L 1 ⊆ D ⊂ C L_1\subseteq D\subset C L 1 ⊆ D ⊂ C . Similarly, L 2 = ( A 1 ∩ B 1 ) ∖ B 2 L_2=(A_1\cap B_1)\setminus B_2 L 2 = ( A 1 ∩ B 1 ) ∖ B 2 is the union of the s 1 A 2 r s_1A_2r s 1 A 2 r -link, s 1 A 2 C s_1A_2C s 1 A 2 C -link and s 1 A 2 D s_1A_2D s 1 A 2 D -link, so L 2 L_2 L 2 is equal to the s 1 A 2 C s_1A_2C s 1 A 2 C -link and thus L 2 ⊆ C ∖ D L_2 \subseteq C \setminus D L 2 ⊆ C ∖ D . Finally, X = A 1 ∩ B 1 ∩ A 2 ∩ B 2 X = A_1 \cap B_1 \cap A_2 \cap B_2 X = A 1 ∩ B 1 ∩ A 2 ∩ B 2 is the union of the s 1 s 2 r s_1s_2r s 1 s 2 r -centre, s 1 s 2 C s_1s_2C s 1 s 2 C -link and s 1 s 2 D s_1s_2D s 1 s 2 D -link, so X X X is equal to the s 1 s 2 r s_1s_2r s 1 s 2 r -centre and thus X ⊆ C ∩ D X \subseteq C \cap D X ⊆ C ∩ D . Therefore, r r r induces ( X , { L 1 , L 2 } ) (X, \{L_1,L_2\}) ( X , { L 1 , L 2 }) on sep ( p ) \operatorname{sep}(p) sep ( p ) .
Lemma 7.5. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let s 1 = ( A 1 , B 1 ) s_1=(A_1,B_1) s 1 = ( A 1 , B 1 ) , s 2 = ( A 2 , B 2 ) s_2=(A_2,B_2) s 2 = ( A 2 , B 2 ) be two ( q + k + 1 ) (q+k+1) ( q + k + 1 ) -crossing k k k -separations of G G G . Let r = ( C , D ) r=(C,D) r = ( C , D ) be a k k k -separation of G G G such that the A 1 A 2 D A_1A_2D A 1 A 2 D -corner, the A 1 B 2 C A_1B_2C A 1 B 2 C -corner, the B 1 A 2 C B_1A_2C B 1 A 2 C -corner, and the B 1 B 2 D B_1B_2D B 1 B 2 D -corner each have size > q >q > q . Then sep ( s 1 ) = sep ( s 2 ) = sep ( r ) \operatorname{sep}(s_1)=\operatorname{sep}(s_2)=\operatorname{sep}(r) sep ( s 1 ) = sep ( s 2 ) = sep ( r ) .
Proof. By Corollary 5.12 applied to the separations s 1 s_1 s 1 , s 2 s_2 s 2 and r r r , we get that the A 1 s 2 r A_1s_2r A 1 s 2 r -link, the s 1 A 2 r s_1A_2r s 1 A 2 r -link, the s 1 s 2 D s_1s_2D s 1 s 2 D -link, the A 1 A 2 r A_1A_2r A 1 A 2 r -link, the A 1 s 2 D A_1s_2D A 1 s 2 D -link and the s 1 A 2 D s_1A_2D s 1 A 2 D -link are empty. So, the neighbourhood of the A 1 A 2 D A_1A_2D A 1 A 2 D -corner is contained in the s 1 s 2 r s_1s_2r s 1 s 2 r -centre. Since the A 1 A 2 D A_1A_2D A 1 A 2 D -corner has size > q >q > q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, the A 1 A 2 D A_1A_2D A 1 A 2 D -corner-separation has order ≥ k \ge k ≥ k . Since s 1 s_1 s 1 , s 2 s_2 s 2 and r r r each have order k k k , it follows that sep ( s 1 ) = sep ( s 2 ) = sep ( r ) \operatorname{sep}(s_1)=\operatorname{sep}(s_2)=\operatorname{sep}(r) sep ( s 1 ) = sep ( s 2 ) = sep ( r ) .
Lemma 7.6. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let δ : = δ ( q , k ) : = 2 q + 2 k + 1 \delta:= \delta(q,k) := 2q+2k+1 δ := δ ( q , k ) := 2 q + 2 k + 1 and δ ′ : = δ ′ ( q , k ) : = 3 q + 3 k + 1 \delta' := \delta'(q,k) := 3q+3k+1 δ ′ := δ ′ ( q , k ) := 3 q + 3 k + 1 . Let s 1 = ( A 1 , B 1 ) s_1=(A_1,B_1) s 1 = ( A 1 , B 1 ) and s 2 = ( A 2 , B 2 ) s_2=(A_2,B_2) s 2 = ( A 2 , B 2 ) be two δ \delta δ -crossing k k k -separations of G G G . Let X X X be the s 1 s 2 s_1s_2 s 1 s 2 -centre and let L i L_i L i be the A i s 3 − i A_is_{3-i} A i s 3 − i -link. Assume that p : = s 1 ∧ s 2 p:=s_1\wedge s_2 p := s 1 ∧ s 2 has order k k k and is δ ′ \delta' δ ′ -crossed by a k k k -separation r = ( C , D ) r=(C,D) r = ( C , D ) of G G G . Then r r r induces ( X , { L 1 , L 2 } ) (X,\{L_1,L_2\}) ( X , { L 1 , L 2 }) on sep ( p ) \operatorname{sep}(p) sep ( p ) .
Proof. Since p p p and r r r ( q + k + 1 ) (q+k+1) ( q + k + 1 ) -cross, the A 1 A 2 C A_1A_2C A 1 A 2 C -corner and the A 1 A 2 D A_1A_2D A 1 A 2 D -corner both have size > q >q > q by Observation 5.4 (ii). Since s 1 s_1 s 1 and s 2 s_2 s 2 δ \delta δ -cross, the 2D corners for A 1 B 2 A_1B_2 A 1 B 2 and for B 1 A 2 B_1A_2 B 1 A 2 both have size ≥ δ − k \ge\delta-k ≥ δ − k . Since the A 1 B 2 r A_1B_2r A 1 B 2 r -link has size ≤ ∣ sep ( r ) ∣ ≤ k \leq|\operatorname{sep}(r)| \leq k ≤ ∣ sep ( r ) ∣ ≤ k , one of the corners for A 1 B 2 C A_1B_2C A 1 B 2 C or A 1 B 2 D A_1B_2D A 1 B 2 D has size ≥ ( δ − 2 k ) / 2 > q \geq(\delta-2k)/2>q ≥ ( δ − 2 k ) /2 > q . Similarly, one of the corners for B 1 A 2 C B_1A_2C B 1 A 2 C or B 1 A 2 D B_1A_2D B 1 A 2 D has size > q >q > q . Hence, we can distinguish the following cases.
Case 1: the A 1 B 2 D A_1B_2D A 1 B 2 D -corner and the B 1 A 2 C B_1A_2C B 1 A 2 C -corner have size > q >q > q . Then the claim follows from Lemma 7.4 .
Case 2: the A 1 B 2 C A_1B_2C A 1 B 2 C -corner and the B 1 A 2 D B_1A_2D B 1 A 2 D -corner have size > q >q > q . This case is symmetric to the previous case.
Case 3: the A 1 B 2 C A_1B_2C A 1 B 2 C -corner and the B 1 A 2 C B_1A_2C B 1 A 2 C -corner have size > q >q > q , and the A 1 B 2 D A_1B_2D A 1 B 2 D -corner and the B 1 A 2 D B_1A_2D B 1 A 2 D -corner have size ≤ q \leq q ≤ q . First, we check that the prerequisites for applying Lemma 7.5 are fulfilled. It remains to show that the B 1 B 2 D B_1B_2D B 1 B 2 D -corner has size > q >q > q . Since r r r δ ′ \delta' δ ′ -crosses p p p , each corresponding 2D corner has size ≥ δ ′ − k \geq\delta'-k ≥ δ ′ − k by Observation 5.4 (ii). Consider the two-dimensional ( B 1 ∪ B 2 ) D (B_1 \cup B_2)D ( B 1 ∪ B 2 ) D -corner, which is equal to the union of the three-dimensional A 1 B 2 D A_1B_2D A 1 B 2 D -corner, the B 1 A 2 D B_1A_2D B 1 A 2 D -corner, the B 1 B 2 D B_1B_2D B 1 B 2 D -corner, the B 1 s 2 D B_1s_2D B 1 s 2 D -link and the s 1 B 2 D s_1B_2D s 1 B 2 D -link. Both the B 1 s 2 D B_1s_2D B 1 s 2 D -link and the s 1 B 2 D s_1B_2D s 1 B 2 D -link have size ≤ k \leq k ≤ k , and both the A 1 B 2 D A_1B_2D A 1 B 2 D -corner and the B 1 A 2 D B_1A_2D B 1 A 2 D -corner have size ≤ q \leq q ≤ q . So, the B 1 B 2 D B_1B_2D B 1 B 2 D -corner has size ≥ δ ′ − k − 2 k − 2 q > q \geq\delta'-k-2k-2q>q ≥ δ ′ − k − 2 k − 2 q > q . Hence, we can apply Lemma 7.5 and get that sep ( s 1 ) = sep ( s 2 ) = sep ( r ) = X \operatorname{sep}(s_1)=\operatorname{sep}(s_2)=\operatorname{sep}(r)=X sep ( s 1 ) = sep ( s 2 ) = sep ( r ) = X . In particular, X = sep ( p ) X=\operatorname{sep}(p) X = sep ( p ) and L 1 = ∅ = L 2 L_1=\varnothing=L_2 L 1 = ∅ = L 2 and r r r induces ( X , { L 1 , L 2 } ) (X,\{L_1,L_2\}) ( X , { L 1 , L 2 }) on sep ( p ) \operatorname{sep}(p) sep ( p ) .
Case 4: the A 1 B 2 D A_1B_2D A 1 B 2 D -corner and the B 1 A 2 D B_1A_2D B 1 A 2 D -corner have size > q >q > q , and the A 1 B 2 C A_1B_2C A 1 B 2 C -corner and the B 1 A 2 C B_1A_2C B 1 A 2 C -corner have size ≤ q \leq q ≤ q . This case is symmetric to the previous case. □
Lemma 7.7. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let F F F be a q q q -lean graphic-k k k -anemone of G G G , with centre Z Z Z . Let p p p be a ( > q ) (>q) ( > q ) -broad petal-separation of F F F and let s 1 = ( A 1 , B 1 ) s_1=(A_1,B_1) s 1 = ( A 1 , B 1 ) and s 2 = ( A 2 , B 2 ) s_2=(A_2,B_2) s 2 = ( A 2 , B 2 ) be displayed separations of F F F that q q q -criss such that p = s 1 ∧ s 2 p=s_1\wedge s_2 p = s 1 ∧ s 2 . Then, the s 1 s 2 s_1s_2 s 1 s 2 -centre is Z Z Z .
Proof. Since F F F is q q q -lean, we have ∣ Z ∣ = k |Z|=k ∣ Z ∣ = k . Since s 1 s_1 s 1 and s 2 s_2 s 2 q q q -criss, they are both ( > q ) (>q) ( > q ) -broad. Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, s 1 s_1 s 1 and s 2 s_2 s 2 are k k k -separations. Since they are displayed by F F F , and since Z Z Z is the centre of F F F , we have sep ( s 1 ) = sep ( s 2 ) = Z \operatorname{sep}(s_1)=\operatorname{sep}(s_2)=Z sep ( s 1 ) = sep ( s 2 ) = Z . Thus, the s 1 s 2 s_1s_2 s 1 s 2 -centre is Z Z Z . □
Lemma 7.8. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let F = ( O , V ) F=(O,\mathcal{V}) F = ( O , V ) be a q q q -lean graphic-k k k -daisy of G G G . Let Z : = ⋂ o ∈ V ( O ) V o Z:=\bigcap_{o\in V(O)}V_o Z := ⋂ o ∈ V ( O ) V o be the centre of F F F . Let p p p be a ( > q ) (>q) ( > q ) -broad petal-separation of F F F and let s 1 = ( A 1 , B 1 ) s_1=(A_1,B_1) s 1 = ( A 1 , B 1 ) and s 2 = ( A 2 , B 2 ) s_2=(A_2,B_2) s 2 = ( A 2 , B 2 ) be displayed separations of F F F that q q q -criss such that p = s 1 ∧ s 2 p=s_1\wedge s_2 p = s 1 ∧ s 2 . Denote by L i L_i L i the A i S 3 − i A_iS_{3-i} A i S 3 − i -link. For every e ∈ E ( O ) e\in E(O) e ∈ E ( O ) , let V e V_e V e be the adhesion-set at e e e . Let e , f ∈ E ( O ) e,f\in E(O) e , f ∈ E ( O ) be the edges of O O O such that p p p is displayed by { e , f } \{e,f\} { e , f } . Then, the s 1 s 2 s_1s_2 s 1 s 2 -centre is V e ∩ V f = Z V_e\cap V_f=Z V e ∩ V f = Z and { L 1 , L 2 } = { V e ∖ Z , V f ∖ Z } \{L_1,L_2\}=\{V_e\setminus Z,V_f\setminus Z\} { L 1 , L 2 } = { V e ∖ Z , V f ∖ Z } .
Proof. Since s 1 s_1 s 1 and s 2 s_2 s 2 are displayed separations of F F F which cross, and p = s 1 ∧ s 2 p=s_1\wedge s_2 p = s 1 ∧ s 2 , there exist edges e ′ , f ′ ∈ E ( O ) e',f'\in E(O) e ′ , f ′ ∈ E ( O ) such that the edges e , f , e ′ , f ′ e,f,e',f' e , f , e ′ , f ′ appear in this order along O O O , s 1 s_1 s 1 is displayed by { e , e ′ } \{e,e'\} { e , e ′ } , and s 2 s_2 s 2 is displayed by { f , f ′ } \{f,f'\} { f , f ′ } . Since s 1 s_1 s 1 and s 2 s_2 s 2 q q q -criss, all their corner separations are ( > q ) (>q) ( > q ) -broad. These separations are displayed respectively by { e , f } \{e,f\} { e , f } , { f , e ′ } \{f,e'\} { f , e ′ } , { e ′ , f ′ } \{e',f'\} { e ′ , f ′ } , and { f ′ , e } \{f',e\} { f ′ , e } . Since s 1 s_1 s 1 and s 2 s_2 s 2 q q q -criss, they are ( > q ) (>q) ( > q ) -broad. Since F F F is q q q -lean, we then have that the sets V e ∖ Z V_e\setminus Z V e ∖ Z , V f ∖ Z V_f\setminus Z V f ∖ Z , V e ′ ∖ Z V_{e'}\setminus Z V e ′ ∖ Z and V f ′ ∖ Z V_{f'}\setminus Z V f ′ ∖ Z are pairwise disjoint. Thus, the s 1 s 2 s_1s_2 s 1 s 2 -centre is Z Z Z , and L 1 = V f ∖ Z L_1=V_f\setminus Z L 1 = V f ∖ Z and L 2 = V e ∖ Z L_2=V_e\setminus Z L 2 = V e ∖ Z . □
We now prove Key Lemma 7.1, which we restate for convenience.
Key Lemma 7.1. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N and
γ ≥ γ 7.1 ( q , k , ε ) : = 2 q + 3 k + 2 ε + 1 and γ ′ ≥ γ 7.1 ′ ( q , k ) : = 3 q + 3 k + 1. \gamma\ge\gamma_{7.1}(q,k,\varepsilon):=2q+3k+2\varepsilon+1 \quad\text{and}\quad\gamma'\ge\gamma'_{7.1}(q,k):=3q+3k+1. γ ≥ γ 7.1 ( q , k , ε ) := 2 q + 3 k + 2 ε + 1 and γ ′ ≥ γ 7.1 ′ ( q , k ) := 3 q + 3 k + 1. Let A A A be a nontrivial γ \gamma γ -asteriskal set of k k k -separations of G G G . Let F F F be a q q q -lean graphic-k k k -flower of G G G that ε \varepsilon ε -approximates A A A . Then every k k k -separation of G G G is γ ′ \gamma' γ ′ -compatible with F F F .
Proof. Let r r r be a k k k -separation of G G G . We show that r r r is γ ′ \gamma' γ ′ -compatible with F F F . Let p p p be a petal-separation of F F F which γ ′ \gamma' γ ′ -crosses r r r . Note that p p p is 2 ( γ ′ − k ) 2(\gamma'-k) 2 ( γ ′ − k ) -broad by Lemma 5.6. Since 2 ( γ ′ − k ) > q 2(\gamma'-k)>q 2 ( γ ′ − k ) > q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, p p p has order exactly k k k . Since A A A is nontrivial and γ \gamma γ -asteriskal, there are two separations r 1 r_1 r 1 and r 2 r_2 r 2 in A A A that γ \gamma γ -cross. Note that both r i r_i r i are 2 ( γ − k ) 2(\gamma-k) 2 ( γ − k ) -broad by Lemma 5.6, in particular they are ε \varepsilon ε -broad since 2 ( γ − k ) ≥ ε 2(\gamma-k)\ge\varepsilon 2 ( γ − k ) ≥ ε . Since F F F ε \varepsilon ε -approximates A A A , there are two displayed separations r 1 ′ r'_1 r 1 ′ and r 2 ′ r'_2 r 2 ′ of F F F that are ε \varepsilon ε -close to r 1 r_1 r 1 and r 2 r_2 r 2 , respectively. Then, r 1 ′ r'_1 r 1 ′ and r 2 ′ r'_2 r 2 ′ ( γ ′ − 2 ε ) (\gamma'-2\varepsilon) ( γ ′ − 2 ε ) -cross by Lemma 7.2 . By Lemma 7.3 , there are δ \delta δ -crossing displayed separations s 1 s_1 s 1 and s 2 s_2 s 2 of F F F such that p = s 1 ∧ s 2 p=s_1\wedge s_2 p = s 1 ∧ s 2 , where
δ = min { 2 γ ′ − 2 k − 1 , γ − 2 ε − k } ≥ 2 ( q + k ) + 1. \delta=\min\{2\gamma'-2k-1,\gamma-2\varepsilon-k\}\ge2(q+k)+1. δ = min { 2 γ ′ − 2 k − 1 , γ − 2 ε − k } ≥ 2 ( q + k ) + 1. Write s i = ( A i , B i ) s_i=(A_i,B_i) s i = ( A i , B i ) and let L i L_i L i be the A i S 3 − i A_iS_{3-i} A i S 3 − i -link of the two separations s 1 s_1 s 1 and s 2 s_2 s 2 and let X X X denote the s 1 s 2 s_1s_2 s 1 s 2 -centre. Recall that p p p is γ ′ \gamma' γ ′ -crossed by r r r , where γ ′ ≥ 3 q + 3 k + 1 \gamma'\ge3q+3k+1 γ ′ ≥ 3 q + 3 k + 1 . By Lemma 7.6 , r r r induces the tripartition ( X , { L 1 , L 2 } ) (X,\{L_1,L_2\}) ( X , { L 1 , L 2 }) on p p p . Since s 1 s_1 s 1 and s 2 s_2 s 2 are δ \delta δ -crossing, they are q q q -crissing.
If F F F is a graphic-k k k -anemone, let Z Z Z denote its centre. Then, by Lemma 7.7 , we get X = Z X = Z X = Z . Since F F F is q q q -lean, we have ∣ Z ∣ = k |Z| = k ∣ Z ∣ = k . Since p p p and r r r are k k k -separations, they cross with centre Z Z Z and four empty links. Thus, r r r is γ ′ \gamma' γ ′ -compatible with F F F .
If F F F is a graphic-k k k -daisy, write F = ( O , V ) F = (O,\mathcal{V}) F = ( O , V ) and let Z = ⋂ o ∈ V ( O ) V o Z = \bigcap_{o\in V(O)} V_o Z = ⋂ o ∈ V ( O ) V o . Let e , f e,f e , f be the edges of O O O such that p p p is displayed by { e , f } \{e,f\} { e , f } . By Lemma 7.8 , we have X = V e ∩ V f = Z X = V_e \cap V_f = Z X = V e ∩ V f = Z , and { L 1 , L 2 } = { V e ∖ Z , V f ∖ Z } \{L_1,L_2\} = \{V_e \setminus Z,V_f \setminus Z\} { L 1 , L 2 } = { V e ∖ Z , V f ∖ Z } . This proves the first two conditions of Definition 6.1 .
Since p p p and r r r γ ′ \gamma' γ ′ -cross, they ( γ ′ − k ) (\gamma' - k) ( γ ′ − k ) -criss, with γ ′ − k > q \gamma' - k > q γ ′ − k > q . Thus, their four corner separations are ( > q ) (> q) ( > q ) -broad. Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, these four separations have order ≥ k \ge k ≥ k . By submodularity, since p p p and r r r are k k k -separations, these four corner separations have order exactly k k k . This implies that all the links of the two separations p , r p,r p , r are of the same size. Therefore, r r r is γ ′ \gamma' γ ′ -compatible with F F F .
Approximation by a maximal nested set
Following the strategy outlined in the overview, we construct a tree-decomposition from an approximately nested family of separations. We allow bounded perturbations before choosing a maximal nested family, and show that this family still approximates the original separations.
Recall that a graph G G G is ( b , k ) (b,k) ( b , k ) -narrow if every b b b -broad k k k -separation is narrow.
Let S S S be a set of separations of a graph G G G . A separation s s s is S S S -totally-ε \varepsilon ε -nested if s ∈ S s \in S s ∈ S and s s s is ε \varepsilon ε -nested with every separation in S S S . A separation of G G G is totally-ε \varepsilon ε -nested if it is S S S -totally-ε \varepsilon ε -nested where S S S is the set of all proper ( ≤ k ) (\le k) ( ≤ k ) -separations of G G G . A separation of G G G is narrow-ε \varepsilon ε -nested if it is S S S -totally-ε \varepsilon ε -nested where S S S is the set of all narrow proper ( ≤ k ) (\le k) ( ≤ k ) -separations of G G G . Here, k k k will be clear from context. We abbreviate S S S -totally-0 0 0 -nested to S S S -totally-nested, and totally-0 0 0 -nested to totally-nested, and narrow-0 0 0 -nested to narrow-nested.
In this section, we prove the following result.
Key Lemma 8.1. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph and let γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let
δ : = δ 8.1 ( k , b , γ ) : = 3 δ 8.20 ( k , b , γ ) + 2 γ + 2 k and γ ′ : = γ 8.1 ′ ( k , b , γ ) : = δ + γ and β : = β 8.1 ( q , k , b ) : = β 8.20 ( q , k , b ) and ε : = ε 8.1 ( q , k , b , γ ) : = max { ε 8.20 ( q , k , b , γ ) , δ + β } and δ ′ : = δ 8.1 ′ ( k , b , γ ) : = δ 8.20 ( k , b , γ ′ ) and ε ′ : = ε 8.1 ′ ( q , k , b , γ ) : = ε 8.20 ( q , k , b , γ ′ ) . \begin{aligned}
\delta&:= \delta_{8.1}(k,b,\gamma) := 3\delta_{8.20}(k,b,\gamma) + 2\gamma+ 2k \quad\text{and}\\
\gamma' &:= \gamma'_{8.1}(k,b,\gamma) := \delta+ \gamma\quad\text{and}\\
\beta&:= \beta_{8.1}(q,k,b) := \beta_{8.20}(q,k,b) \quad\text{and}\\
\varepsilon&:= \varepsilon_{8.1}(q,k,b,\gamma) := \max\{\varepsilon_{8.20}(q,k,b,\gamma),\delta+\beta\} \quad\text{and}\\
\delta' &:= \delta'_{8.1}(k,b,\gamma) := \delta_{8.20}(k,b,\gamma') \quad\text{and}\\
\varepsilon' &:= \varepsilon'_{8.1}(q,k,b,\gamma) := \varepsilon_{8.20}(q,k,b,\gamma').
\end{aligned} δ γ ′ β ε δ ′ ε ′ := δ 8.1 ( k , b , γ ) := 3 δ 8.20 ( k , b , γ ) + 2 γ + 2 k and := γ 8.1 ′ ( k , b , γ ) := δ + γ and := β 8.1 ( q , k , b ) := β 8.20 ( q , k , b ) and := ε 8.1 ( q , k , b , γ ) := max { ε 8.20 ( q , k , b , γ ) , δ + β } and := δ 8.1 ′ ( k , b , γ ) := δ 8.20 ( k , b , γ ′ ) and := ε 8.1 ′ ( q , k , b , γ ) := ε 8.20 ( q , k , b , γ ′ ) . Let S S S be the set of all narrow-γ \gamma γ -nested k k k -separations of G G G . Let S ′ S' S ′ be the set of all half-connected narrow-γ ′ \gamma' γ ′ -nested β \beta β -broad k k k -separations of G G G that are δ \delta δ -close to some separation in S S S . Let N ⊆ S ′ N \subseteq S' N ⊆ S ′ be a maximal nested set of separations in S ′ S' S ′ .
Then N N N ε \varepsilon ε -approximates S S S . Moreover, for every ε ′ \varepsilon' ε ′ -broad k k k -separation of G G G there is a δ ′ \delta' δ ′ -close k k k -separation of G G G that is nested with N N N .
Let S S S be a set of separations of a graph G G G . We say that S S S is ε \varepsilon ε -canonical if for every automorphism φ \varphi φ of G G G , the sets S S S and φ ( S ) \varphi(S) φ ( S ) ε \varepsilon ε -approximate each other. So if S S S is ε \varepsilon ε -canonical for ε = 0 \varepsilon= 0 ε = 0 , then S S S is canonical. Similarly, we call a tree-decomposition T = ( T , V ) \mathcal{T} = (T,\mathcal{V}) T = ( T , V ) ε \varepsilon ε -canonical if the set of separations that are induced by an edge of T T T is ε \varepsilon ε -canonical.
Corollary 8.2. Assume the premises of Key Lemma 8.1. Then the set N N N is 2 ε 2\varepsilon 2 ε -canonical.
Proof. Let φ \varphi φ be an automorphism of G G G and let N ′ : = φ ( N ) N' := \varphi(N) N ′ := φ ( N ) . By Key Lemma 8.1, N N N ε \varepsilon ε -approximates S S S . Similarly, N ′ = φ ( N ) N' = \varphi(N) N ′ = φ ( N ) ε \varepsilon ε -approximates φ ( S ) = S \varphi(S) = S φ ( S ) = S . Since every separation in S ′ S' S ′ is δ \delta δ -close to some separation in S S S , the set S S S δ \delta δ -approximates S ′ S' S ′ . Recall that N ⊆ S ′ N \subseteq S' N ⊆ S ′ . Similarly, we have N ′ = φ ( N ) ⊆ φ ( S ′ ) = S ′ N' = \varphi(N) \subseteq\varphi(S') = S' N ′ = φ ( N ) ⊆ φ ( S ′ ) = S ′ . Hence, S S S δ \delta δ -approximates N N N and N ′ N' N ′ .
To summarize, N N N ε \varepsilon ε -approximates S S S , which in turn δ \delta δ -approximates N ′ N' N ′ . So, N N N ( ε + δ ) (\varepsilon+\delta) ( ε + δ ) -approximates N ′ N' N ′ by Lemma 4.9 . Similarly, N ′ N' N ′ ( ε + δ ) (\varepsilon+\delta) ( ε + δ ) -approximates N N N . It follows that N N N and N ′ N' N ′ ( ε + δ ) (\varepsilon+\delta) ( ε + δ ) -approximate each other, hence N N N is ( ε + δ ) (\varepsilon+\delta) ( ε + δ ) -canonical, so 2 ε 2\varepsilon 2 ε -canonical. □
A proper separation ( A , B ) (A,B) ( A , B ) of G G G is half-connected if G [ A ∖ B ] G[A \setminus B] G [ A ∖ B ] or G [ B ∖ A ] G[B \setminus A] G [ B ∖ A ] is connected.
We now give a high-level overview of the proof. Let s s s be a narrow-γ \gamma γ -nested separation of G G G . By definition, s s s is γ \gamma γ -nested with every separation in S ′ S' S ′ . Thus, either s s s is γ \gamma γ -close to some separation of N N N or there is a unique splitting star σ \sigma σ of N N N such that s s s γ \gamma γ -interlaces σ \sigma σ (Definition 8.5 and Lemma 8.6 ). In the first case, we are done, so suppose that s s s γ \gamma γ -interlaces σ \sigma σ . In that case, either s s s crosses few separations of σ \sigma σ , or each side of s s s has many connected components (Lemma 8.13 ). If s s s crosses few separations of σ \sigma σ then we can “wiggle” (see Lemma 8.15) s s s to a k k k -separation s ′ ∈ S ′ s' \in S' s ′ ∈ S ′ that almost-interlaces σ \sigma σ and is δ \delta δ -close to s s s . Then, s ′ s' s ′ is nested with every separation in N N N , so s ′ ∈ N s' \in N s ′ ∈ N by maximality of N N N . If each side of s s s has many connected components, then every half-connected separation s ′ s' s ′ with separator sep ( s ) \operatorname{sep}(s) sep ( s ) is not crossed by a narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G (Lemma 8.8), so it essentially belongs to N N N . Then, since s s s is narrow-γ \gamma γ -nested, it is δ \delta δ -close to some s ′ ∈ N s' \in N s ′ ∈ N (Lemma 3.13), and we are done.
Splitting stars and approximate interlacing
Lemma 8.3. Let γ ∈ N \gamma\in\mathbb{N} γ ∈ N , let G G G be a graph, and let s s s be a separation of G G G . If s ≤ γ s ∗ s \leq_{\gamma} s^* s ≤ γ s ∗ then s s s has breadth ≤ γ \leq\gamma ≤ γ .
Proof. Let s : = ( A , B ) s := (A,B) s := ( A , B ) . Then s ≤ γ s ∗ s \leq_{\gamma} s^* s ≤ γ s ∗ implies A ⊆ γ B A \subseteq_{\gamma} B A ⊆ γ B . Hence, ∣ A ∖ B ∣ ≤ γ |A \setminus B| \leq\gamma ∣ A ∖ B ∣ ≤ γ . □
Given a set N N N of separations of G G G , an orientation of N N N is a subset M ⊆ N M \subseteq N M ⊆ N such that for every s ∈ N s \in N s ∈ N exactly one of s s s and s ∗ s^* s ∗ is in M M M . The following result is folklore; we include a proof for completeness.
Lemma 8.4. Let N ⃗ \vec{N} N be a nested consistent finite set of proper separations of a graph G G G . Then the set of all ≤ \leq ≤ -maximal separations in N ⃗ \vec{N} N is the unique splitting star in N ⃗ \vec{N} N .
Proof. Let σ \sigma σ be the set of all ≤ \leq ≤ -maximal separations in N ⃗ \vec{N} N . First, we argue that σ \sigma σ is a splitting star in N ⃗ \vec{N} N . Let s ≠ t ∈ σ s \ne t \in\sigma s = t ∈ σ . Since N ⃗ \vec{N} N is nested, we have s ≤ t s \leq t s ≤ t or s ≤ t ∗ s \leq t^* s ≤ t ∗ or t ≤ s t \leq s t ≤ s or t ∗ ≤ s t^* \leq s t ∗ ≤ s . Since the elements of σ \sigma σ are ≤ \leq ≤ -maximal in N ⃗ \vec{N} N , we have either s ≤ t ∗ s \leq t^* s ≤ t ∗ or t ∗ ≤ s t^* \leq s t ∗ ≤ s . Since N ⃗ \vec{N} N is consistent, we must have s ≤ t ∗ s \le t^* s ≤ t ∗ , so σ \sigma σ is a star. Now we show that σ \sigma σ is splitting in N ⃗ \vec{N} N . Assume the contrary. Then there exists s ∈ N ⃗ s \in\vec{N} s ∈ N that interlaces σ \sigma σ . Thus, s ∉ σ s \notin\sigma s ∈ / σ , so s s s is not ≤ \le ≤ -maximal in N ⃗ \vec{N} N . Hence there exists t ∈ σ t \in\sigma t ∈ σ such that s < t s < t s < t . Since s s s interlaces σ \sigma σ , we have either t < s t < s t < s or t < s ∗ t < s^* t < s ∗ . The first outcome contradicts the ≤ \le ≤ -maximality of t t t , so we have t < s ∗ t < s^* t < s ∗ . Thus, s < t < s ∗ s < t < s^* s < t < s ∗ , which contradicts that s s s is proper.
Now, let τ \tau τ be a splitting star in N ⃗ \vec{N} N . We first show that σ ⊆ τ \sigma\subseteq\tau σ ⊆ τ . By contradiction, suppose that there exists s ∈ σ ∖ τ s \in\sigma\setminus\tau s ∈ σ ∖ τ . Since τ \tau τ is splitting, s s s does not interlace τ \tau τ so there exists t ∈ τ t \in\tau t ∈ τ such that t ≮ s t \not< s t < s and t ≮ s ∗ t \not< s^* t < s ∗ . Since N ⃗ \vec{N} N is nested, we have either s ≤ t s \le t s ≤ t or s ∗ ≤ t s^* \le t s ∗ ≤ t . The first outcome would contradict the ≤ \le ≤ -maximality of s s s and the second that N ⃗ \vec{N} N is consistent. Thus, σ ⊆ τ \sigma\subseteq\tau σ ⊆ τ . Now, we show that τ ⊆ σ \tau\subseteq\sigma τ ⊆ σ . By contradiction, suppose that there exists t ∈ τ ∖ σ t \in\tau\setminus\sigma t ∈ τ ∖ σ . Since t ∉ σ t \notin\sigma t ∈ / σ , there exists s ∈ σ s \in\sigma s ∈ σ such that t ≤ s t \le s t ≤ s . Since τ \tau τ is a star and s ∈ σ ⊆ τ s \in\sigma\subseteq\tau s ∈ σ ⊆ τ , we have s ≤ t ∗ s \le t^* s ≤ t ∗ . Thus, t ≤ s ≤ t ∗ t \le s \le t^* t ≤ s ≤ t ∗ , which contradicts that t t t is proper. Therefore, τ = σ \tau= \sigma τ = σ , which concludes the proof. □
It is folklore that if a proper separation s s s is nested with every separation from a finite symmetric nested set N N N of proper separations, either s ∈ N s \in N s ∈ N or s s s interlaces a unique splitting star of N N N . Lemma 8.6 is a generalization of this statement to the approximate setting. For this, we need the following definition.
Definition 8.5 (Approximate version of interlacing). Let σ \sigma σ be a star of separations of a graph G G G , and let s s s be a separation of G G G . We say that s s s ε \varepsilon ε -interlaces σ \sigma σ if for every t ∈ σ t \in\sigma t ∈ σ , we have t < ε s t <_{\varepsilon} s t < ε s or t < ε s ∗ t <_{\varepsilon} s^* t < ε s ∗ . We say that s s s almost-ε \varepsilon ε -interlaces σ \sigma σ if for every t ∈ σ t \in\sigma t ∈ σ , we have t ≤ ε s t \le_{\varepsilon} s t ≤ ε s or t ≤ ε s ∗ t \le_{\varepsilon} s^* t ≤ ε s ∗ . Now let S S S be a set of separations of G G G . We say that S S S (almost-)ε \varepsilon ε -interlaces σ \sigma σ if every separation in S S S (almost-)ε \varepsilon ε -interlaces σ \sigma σ .
Lemma 8.6 (Generalization of (CK25a, Corollary V.9)). Let G G G be a graph and let N N N be a symmetric nested finite set of proper separations of G G G . Let s s s be a ( 2 γ + 1 ) (2\gamma+ 1) ( 2 γ + 1 ) -broad separation of G G G that is γ \gamma γ -nested with every separation in N N N . Then either
(i) s s s is γ \gamma γ -close to some n ∈ N n \in N n ∈ N , or
(ii) O ( s ) : = { t ∈ N ∣ t ≤ γ s or t ≤ γ s ∗ } O(s) := \{t \in N \mid t \le_{\gamma} s \text{ or } t \le_{\gamma} s^*\} O ( s ) := { t ∈ N ∣ t ≤ γ s or t ≤ γ s ∗ } is a consistent orientation of N N N , and the set of ≤ \le ≤ -maximal elements in O ( s ) O(s) O ( s ) is the unique splitting star σ \sigma σ of N N N such that s s s γ \gamma γ -interlaces σ \sigma σ .
Proof. Assume that (i) fails, i.e. s s s is not γ \gamma γ -close to any n ∈ N n \in N n ∈ N .
Claim 8.6.1. For all n ∈ N n \in N n ∈ N , we have n ∈ O ( s ) n \in O(s) n ∈ O ( s ) if and only if n ∗ ∉ O ( s ) n^* \notin O(s) n ∗ ∈ / O ( s ) .
Proof of the claim. Note that since s s s is γ \gamma γ -nested with N N N , we either have n ∈ O ( s ) n \in O(s) n ∈ O ( s ) or n ∗ ∈ O ( s ) n^* \in O(s) n ∗ ∈ O ( s ) for every n ∈ N n \in N n ∈ N . On the other hand, assume both n ∈ O ( s ) n \in O(s) n ∈ O ( s ) and n ∗ ∈ O ( s ) n^* \in O(s) n ∗ ∈ O ( s ) for some n ∈ N n \in N n ∈ N . Then either n ≤ γ s n \le_{\gamma} s n ≤ γ s or n ≤ γ s ∗ n \le_{\gamma} s^* n ≤ γ s ∗ , and either n ∗ ≤ γ s n^* \le_{\gamma} s n ∗ ≤ γ s or n ∗ ≤ γ s ∗ n^* \le_{\gamma} s^* n ∗ ≤ γ s ∗ . Since s s s is not γ \gamma γ -close to n n n and not γ \gamma γ -close to n ∗ n^* n ∗ , we either have s ∗ ≤ γ n ≤ γ s s^* \le_{\gamma} n \le_{\gamma} s s ∗ ≤ γ n ≤ γ s or s ≤ γ n ≤ γ s ∗ s \le_{\gamma} n \le_{\gamma} s^* s ≤ γ n ≤ γ s ∗ . Without loss of generality, we can assume the latter. Then s ≤ 2 γ s ∗ s \le_{2\gamma} s^* s ≤ 2 γ s ∗ , which implies that s s s has breadth ≤ 2 γ \le2\gamma ≤ 2 γ by Lemma 8.3, a contradiction. ◇
Claim 8.6.2. If n 1 ∈ N n_1 \in N n 1 ∈ N and n 2 ∈ O ( s ) n_2 \in O(s) n 2 ∈ O ( s ) are such that n 1 ≤ n 2 n_1 \le n_2 n 1 ≤ n 2 , then n 1 ∈ O ( s ) n_1 \in O(s) n 1 ∈ O ( s ) .
Proof of the claim. Indeed, n 2 ∈ O ( s ) n_2 \in O(s) n 2 ∈ O ( s ) implies n 2 ≤ γ s n_2 \le_{\gamma} s n 2 ≤ γ s or n 2 ≤ γ s ∗ n_2 \le_{\gamma} s^* n 2 ≤ γ s ∗ . Since n 1 ≤ n 2 n_1 \le n_2 n 1 ≤ n 2 , we also have n 1 ≤ γ s n_1 \le_{\gamma} s n 1 ≤ γ s or n 1 ≤ γ s ∗ n_1 \le_{\gamma} s^* n 1 ≤ γ s ∗ . So, n 1 ∈ O ( s ) n_1 \in O(s) n 1 ∈ O ( s ) . ◇
Claim 8.6.3. Let σ \sigma σ be a splitting star of N N N . Then σ ⊆ O ( s ) \sigma\subseteq O(s) σ ⊆ O ( s ) if and only if s s s γ \gamma γ -interlaces σ \sigma σ .
Proof of the claim.
σ ⊆ O ( s ) ⟺ ∀ n ∈ σ , n ≼ γ s or n ≼ γ s ∗ ⟺ s γ -interlaces σ . \sigma\subseteq O(s) \Longleftrightarrow\forall n \in\sigma,\ n \preccurlyeq_{\gamma} s\text{ or }n \preccurlyeq_{\gamma} s^{*} \Longleftrightarrow s\ \gamma\text{-interlaces }\sigma. σ ⊆ O ( s ) ⟺ ∀ n ∈ σ , n ≼ γ s or n ≼ γ s ∗ ⟺ s γ -interlaces σ .
From Claims 8.6.1 and 8.6.2 , it follows that O ( s ) O(s) O ( s ) is consistent. By Lemma 8.4 , there is a unique splitting star σ \sigma σ in O ( s ) O(s) O ( s ) . By Claims 8.6.1 and 8.6.2 , it follows that σ \sigma σ is the only splitting star of N N N contained in O ( s ) O(s) O ( s ) . From Claim 8.6.3, we finally get that σ \sigma σ is the unique splitting star of N N N that is γ \gamma γ -interlaced by s s s .
We now establish two results about how narrow separations can cross other separations.
Lemma 8.7. Let s = ( A , B ) s=(A,B) s = ( A , B ) and t = ( C , D ) t=(C,D) t = ( C , D ) be crossing k k k -separations in a graph G G G . Assume that s s s is narrow and proper. Then either
the A t At A t -link is nonempty, or
sep ( s ) = sep ( t ) \operatorname{sep}(s)=\operatorname{sep}(t) sep ( s ) = sep ( t ) and G − sep ( t ) G-\operatorname{sep}(t) G − sep ( t ) has a component in the A C AC A C -corner and a component in the A D AD A D -corner.
Proof. Assume that (i) fails, i.e. the A t At A t -link is empty. First, we claim that both the A C AC A C -corner and the A D AD A D -corner are nonempty. Suppose for contradiction that the A C AC A C -corner is empty, say. Then the A D AD A D -corner is nonempty. Since s s s is narrow, the s C sC s C -link is empty. So s s s and t t t are nested, a contradiction.
Hence, both the A C AC A C -corner and the A D AD A D -corner are nonempty. Since ( A , B ) (A,B) ( A , B ) is narrow, both the s C sC s C -link and the s D sD sD -link are empty, so the s t st s t -centre has size k k k . Hence, sep ( s ) = sep ( t ) \operatorname{sep}(s)=\operatorname{sep}(t) sep ( s ) = sep ( t ) and we have outcome (ii).
Lemma 8.8. Let G G G be a graph and let s s s be a narrow half-connected proper separation of G G G . Assume that G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) has ≥ k + 1 \geq k+1 ≥ k + 1 components. Then s s s is not crossed by any narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G .
Proof. Let s = ( A , B ) s=(A,B) s = ( A , B ) and assume that there is a narrow ( ≤ k ) (\leq k) ( ≤ k ) -separation t = ( C , D ) t=(C,D) t = ( C , D ) that crosses s s s .
Claim 8.8.1. The s C sC s C -link or the s D sD sD -link is empty.
Proof of the claim. Assume for contradiction that there is a vertex v v v in the s C sC s C -link and a vertex w w w in the s D sD sD -link. Since s s s is narrow and G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) has ≥ k + 1 \geq k+1 ≥ k + 1 components, there are k + 1 k+1 k + 1 independent v v v –w w w paths in G G G . Each of those paths intersects the separator of t t t , contradicting ∣ sep ( t ) ∣ ≤ k |\operatorname{sep}(t)|\leq k ∣ sep ( t ) ∣ ≤ k .
Assume without loss of generality that the s C sC s C -link is empty. Consider any component K K K of C ∖ D C\setminus D C ∖ D . Say, K K K is contained in the A C AC A C -corner. Since t t t is narrow, we have N ( K ) = sep ( t ) N(K)=\operatorname{sep}(t) N ( K ) = sep ( t ) so the B t Bt B t -link is empty. Since s s s and t t t cross, the B C BC B C -corner must be nonempty. Since t t t is narrow, the A t At A t -link is empty, so the separator of t t t is contained in the s t st s t -centre; i.e. sep ( t ) ⊆ sep ( s ) \operatorname{sep}(t)\subseteq\operatorname{sep}(s) sep ( t ) ⊆ sep ( s ) . Since s s s is half-connected, either G [ A ∖ B ] G[A \setminus B] G [ A ∖ B ] or G [ B ∖ A ] G[B \setminus A] G [ B ∖ A ] is connected, say the former. Then the A D AD A D -corner is empty. Since s s s and t t t cross, there is a vertex u u u in the s D sD sD -link. Then, however, u u u does not have a neighbour in A ∖ B A \setminus B A ∖ B , contradicting that s s s is tight. □
We now demonstrate some properties of the operations ∧ \wedge ∧ and ∨ \vee ∨ , and how they interact with ≤ γ \leq_\gamma ≤ γ . The proofs are elementary.
Lemma 8.9. Let s = ( A , B ) s=(A,B) s = ( A , B ) , t = ( C , D ) t=(C,D) t = ( C , D ) and s ′ = ( E , F ) s'=(E,F) s ′ = ( E , F ) be separations of a graph G G G , and γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Assume s ≤ t ∗ s \leq t^* s ≤ t ∗ . Let ( X 1 , Y 1 ) : = s ∨ s ′ (X_1,Y_1):=s\vee s' ( X 1 , Y 1 ) := s ∨ s ′ and ( X 2 , Y 2 ) : = s ∗ ∧ s ′ (X_2,Y_2):=s^*\wedge s' ( X 2 , Y 2 ) := s ∗ ∧ s ′ . Then all of the following hold:
(i) the C F CF C F -corner is equal to the C Y 1 CY_1 C Y 1 -corner and equal to the C Y 2 CY_2 C Y 2 -corner;
(ii) the C E CE C E -corner is equal to the C X 1 CX_1 C X 1 -corner and equal to the C X 2 CX_2 C X 2 -corner;
(iii) If t ≤ γ s ′ t\leq_\gamma s' t ≤ γ s ′ , then both t ≤ γ s ∨ s ′ t\leq_\gamma s\vee s' t ≤ γ s ∨ s ′ and t ≤ γ s ∗ ∧ s ′ t\leq_\gamma s^*\wedge s' t ≤ γ s ∗ ∧ s ′ .
(iv) If t ≤ γ ( s ′ ) ∗ t\leq_\gamma(s')^* t ≤ γ ( s ′ ) ∗ , then both t ≤ γ ( s ∨ s ′ ) ∗ t\leq_\gamma(s\vee s')^* t ≤ γ ( s ∨ s ′ ) ∗ and t ≤ γ ( s ∗ ∧ s ′ ) ∗ t\leq_\gamma(s^*\wedge s')^* t ≤ γ ( s ∗ ∧ s ′ ) ∗ .
Proof. (i) and (ii). By symmetry, it suffices to show (i). Since s ≤ t ∗ s\leq t^* s ≤ t ∗ , we have A ⊆ D A\subseteq D A ⊆ D and C ⊆ B C\subseteq B C ⊆ B . Note that the C F CF C F -corner is ( C ∩ F ) ∖ ( D ∪ E ) (C\cap F)\setminus(D\cup E) ( C ∩ F ) ∖ ( D ∪ E ) . We have X 1 = A ∪ E X_1=A\cup E X 1 = A ∪ E and Y 1 = B ∩ F Y_1=B\cap F Y 1 = B ∩ F . Thus, the C Y 1 CY_1 C Y 1 -corner is
( C ∩ Y 1 ) ∖ ( D ∪ X 1 ) = ( C ∩ B ∩ F ) ∖ ( D ∪ A ∪ E ) = ( C ∩ F ) ∖ ( D ∪ E ) (since C ⊆ B and A ⊆ D ). \begin{aligned}
(C\cap Y_1)\setminus(D\cup X_1)=(C\cap B\cap F)\setminus(D\cup A\cup E) \\
&=(C\cap F)\setminus(D\cup E)\qquad\text{(since } C\subseteq B \text{ and } A\subseteq D\text{).}
\end{aligned} ( C ∩ Y 1 ) ∖ ( D ∪ X 1 ) = ( C ∩ B ∩ F ) ∖ ( D ∪ A ∪ E ) = ( C ∩ F ) ∖ ( D ∪ E ) (since C ⊆ B and A ⊆ D ).
We also have X 2 = B ∩ E X_2=B\cap E X 2 = B ∩ E and Y 2 = A ∪ F Y_2=A\cup F Y 2 = A ∪ F . Thus, the C Y 2 CY_2 C Y 2 -corner is
( C ∩ Y 2 ) ∖ ( D ∪ X 2 ) = ( C ∩ ( A ∪ F ) ) ∖ ( D ∪ ( B ∩ E ) ) = ( C ∩ F ) ∖ ( D ∪ ( B ∩ E ) ) (since A ⊆ D ). = ( C ∩ F ) ∖ ( D ∪ E ) (since C ⊆ B ). \begin{aligned}
(C\cap Y_2)\setminus(D\cup X_2)=(C\cap(A\cup F))\setminus(D\cup(B\cap E)) \\
&=(C\cap F)\setminus(D\cup(B\cap E))\qquad\text{(since } A\subseteq D\text{).} \\
&=(C\cap F)\setminus(D\cup E)\qquad\text{(since } C\subseteq B\text{).}
\end{aligned} ( C ∩ Y 2 ) ∖ ( D ∪ X 2 ) = ( C ∩ ( A ∪ F )) ∖ ( D ∪ ( B ∩ E )) = ( C ∩ F ) ∖ ( D ∪ ( B ∩ E )) (since A ⊆ D ). = ( C ∩ F ) ∖ ( D ∪ E ) (since C ⊆ B ).
(iii) and (iv). By symmetry, it suffices to show (iii). We have C ⊆ γ E ⊆ A ∪ E C\subseteq_\gamma E\subseteq A\cup E C ⊆ γ E ⊆ A ∪ E and D ⊇ γ F ⊇ B ∩ F D\supseteq_\gamma F\supseteq B\cap F D ⊇ γ F ⊇ B ∩ F . So, t = ( C , D ) ≤ γ ( A ∪ E , B ∩ F ) = s ∨ s ′ t=(C,D)\leq_\gamma(A\cup E,B\cap F)=s\vee s' t = ( C , D ) ≤ γ ( A ∪ E , B ∩ F ) = s ∨ s ′ . Similarly, with C ⊆ B C\subseteq B C ⊆ B and C ⊆ γ E C\subseteq_\gamma E C ⊆ γ E and D ⊇ γ F D\supseteq_\gamma F D ⊇ γ F and D ⊇ A D\supseteq A D ⊇ A , we get t = ( C , D ) ≤ γ ( B ∩ E , A ∪ F ) = s ∗ ∧ s ′ t=(C,D)\leq_\gamma(B\cap E,A\cup F)=s^*\wedge s' t = ( C , D ) ≤ γ ( B ∩ E , A ∪ F ) = s ∗ ∧ s ′ □
Lemma 8.10. Let σ \sigma σ be a star of separations of a graph G G G , let p ∈ σ p\in\sigma p ∈ σ and let γ ≥ 0 \gamma\geq0 γ ≥ 0 be an integer. If a separation s s s almost-γ \gamma γ -interlaces σ ∖ { p } \sigma\setminus\{p\} σ ∖ { p } , then both s ∨ p s\vee p s ∨ p and s ∧ p ∗ s\wedge p^* s ∧ p ∗ almost-γ \gamma γ -interlace σ \sigma σ .
Proof. Let t ∈ σ ∖ { p } t\in\sigma\setminus\{p\} t ∈ σ ∖ { p } . Since σ \sigma σ is a star, we have p ≤ t ∗ p\leq t^* p ≤ t ∗ . We have t ≤ γ s t\leq_\gamma s t ≤ γ s or t ≤ γ s ∗ t\leq_\gamma s^* t ≤ γ s ∗ since s s s almost-γ \gamma γ -interlaces σ ∖ { p } \sigma\setminus\{p\} σ ∖ { p } . Then Lemma 8.9 (iii) and (iv) imply t ≤ γ s ∨ p t\leq_\gamma s\vee p t ≤ γ s ∨ p or t ≤ γ ( s ∨ p ) ∗ t\leq_\gamma(s\vee p)^* t ≤ γ ( s ∨ p ) ∗ . Similarly, we get t ≤ γ s ∧ p ∗ t\leq_\gamma s\wedge p^* t ≤ γ s ∧ p ∗ or t ≤ γ ( s ∧ p ∗ ) ∗ t\leq_\gamma(s\wedge p^*)^* t ≤ γ ( s ∧ p ∗ ) ∗ . Moreover, we obviously have p ≤ s ∨ p p\leq s\vee p p ≤ s ∨ p and p ≤ ( s ∧ p ∗ ) ∗ p\leq(s\wedge p^*)^* p ≤ ( s ∧ p ∗ ) ∗ . This proves the statement. □
The following observation is a direct consequence of Lemma 4.9.
Observation 8.11. Let s s s and t t t be ( ≤ k ) (\leq k) ( ≤ k ) -separations of a graph G G G .
(i) If s s s is totally-γ \gamma γ -nested and t t t is proper and s ≈ ε t s\approx_\varepsilon t s ≈ ε t , then t t t is totally-( γ + ε ) (\gamma+\varepsilon) ( γ + ε ) -nested.
(ii) If s s s is narrow-γ \gamma γ -nested and t t t is narrow and proper and s ≈ ε t s\approx_\varepsilon t s ≈ ε t , then t t t is narrow-( γ + ε ) (\gamma+\varepsilon) ( γ + ε ) -nested.
Proof. (i). Note that t t t is a proper ( ≤ k ) (\leq k) ( ≤ k ) -separation. By Lemma 4.9 , t t t is ( γ + ε ) (\gamma+\varepsilon) ( γ + ε ) -nested with every proper ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G . Hence, t t t is totally-( γ + ε ) (\gamma+\varepsilon) ( γ + ε ) -nested.
(ii). Note that t t t is a narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separation. By Lemma 4.9 , t t t is ( γ + ε ) (\gamma+\varepsilon) ( γ + ε ) -nested with every narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G . Hence, t t t is narrow-( γ + ε ) (\gamma+\varepsilon) ( γ + ε ) -nested.
Lemma 8.12. Let ε ∈ N \varepsilon\in\mathbb{N} ε ∈ N , let G G G be a graph, let N N N be a nested set of separations of G G G and let σ ⊆ N \sigma\subseteq N σ ⊆ N be a splitting star of N N N . Let s s s be a separation of G G G . If s s s almost-ε \varepsilon ε -interlaces σ \sigma σ , then s s s is ε \varepsilon ε -nested with every separation in N N N . Moreover, if s s s ε \varepsilon ε -interlaces σ \sigma σ , then for every n ∈ N n \in N n ∈ N we have s < ε n s <_{\varepsilon} n s < ε n or s ∗ < ε n s^{*} <_{\varepsilon} n s ∗ < ε n or s < ε n ∗ s <_{\varepsilon} n^{*} s < ε n ∗ or s ∗ < ε n ∗ s^{*} <_{\varepsilon} n^{*} s ∗ < ε n ∗ .
Proof. Let n ∈ N n \in N n ∈ N be arbitrary. Since σ \sigma σ is a splitting star and N N N is nested, there is some n ′ ∈ σ n' \in\sigma n ′ ∈ σ such that n ≤ n ′ n \leq n' n ≤ n ′ or n ∗ ≤ n ′ n^{*} \leq n' n ∗ ≤ n ′ , say the former. Since s s s almost-ε \varepsilon ε -interlaces σ \sigma σ , we have n ′ ≤ ε s n' \leq_{\varepsilon} s n ′ ≤ ε s or n ′ ≤ ε s ∗ n' \leq_{\varepsilon} s^{*} n ′ ≤ ε s ∗ , say the former. Then, n ≤ n ′ ≤ ε s n \leq n' \leq_{\varepsilon} s n ≤ n ′ ≤ ε s implies n ≤ ε s n \leq_{\varepsilon} s n ≤ ε s by Lemma 4.9 . Hence, s s s is ε \varepsilon ε -nested with every separation in N N N .
For the ‘moreover’-part, it suffices to replace ‘≤ ε \leq_{\varepsilon} ≤ ε ’ by ‘< ε <_{\varepsilon} < ε ’ in the above argument.
Proof of Key Lemma 8.1
8.2. Proof of Key Lemma 8.1. We start by analysing how a separation can interact with a star of narrow separations.
Lemma 8.13. Let ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N . Let S S S be the set of narrow proper k k k -separations of a graph G G G . Let σ ⊆ S \sigma\subseteq S σ ⊆ S be a star and let s ∈ S s \in S s ∈ S . Then either
(i) s s s crosses < k + ℓ < k+\ell < k + ℓ many separations in σ \sigma σ , or
(ii) each proper side of s s s has ≥ ℓ \geq\ell ≥ ℓ components.
Proof. Let σ : = { ( A i , B i ) ∣ i ∈ I } \sigma:= \{(A_i,B_i) \mid i \in I\} σ := {( A i , B i ) ∣ i ∈ I } and let s : = ( C , D ) s := (C,D) s := ( C , D ) . Assume that (i) fails, i.e. s s s crosses ≥ k + ℓ \geq k+\ell ≥ k + ℓ many separations in σ \sigma σ . Let σ 1 \sigma_1 σ 1 be the set of all ( A i , B i ) ∈ σ (A_i,B_i) \in\sigma ( A i , B i ) ∈ σ such that the s A i sA_i s A i -link is nonempty. Since A j ∖ B j A_j \setminus B_j A j ∖ B j is disjoint from A j ′ ∖ B j ′ A_{j'} \setminus B_{j'} A j ′ ∖ B j ′ whenever j ≠ j ′ j \neq j' j = j ′ , the star σ 1 \sigma_1 σ 1 contains ≤ k \leq k ≤ k separations. Let σ 2 \sigma_2 σ 2 be the set of all ( A i , B i ) ∈ σ ∖ σ 1 (A_i,B_i) \in\sigma\setminus\sigma_1 ( A i , B i ) ∈ σ ∖ σ 1 that cross s s s . So, ∣ σ 2 ∣ ≥ ℓ |\sigma_2| \geq\ell ∣ σ 2 ∣ ≥ ℓ . By Lemma 8.7, every separation ( A i , B i ) ∈ σ 2 (A_i,B_i) \in\sigma_2 ( A i , B i ) ∈ σ 2 has separator A i ∩ B i = C ∩ D A_i \cap B_i = C \cap D A i ∩ B i = C ∩ D and G − ( C ∩ D ) G - (C \cap D) G − ( C ∩ D ) has a component in the A i C A_iC A i C -corner and a component in the A i D A_iD A i D -corner. Since the ( A i ∖ B i ) (A_i \setminus B_i) ( A i ∖ B i ) ’s are pairwise disjoint, each proper side of s s s has ≥ ℓ \geq\ell ≥ ℓ components.
The next two lemmas turn almost-γ \gamma γ -interlacing into almost-interlacing by “wiggling” a separation to accommodate the members of the star one at a time. These adjustments preserve its order while keeping its distance from the original separation bounded.
Lemma 8.14. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let r r r be a ( q + γ + 1 ) (q+\gamma+1) ( q + γ + 1 ) -broad k k k -separation and let p p p be a ( 2 q + k + 1 ) (2q+k+1) ( 2 q + k + 1 ) -broad k k k -separation. If p ≤ γ r p \leq_{\gamma} r p ≤ γ r or p ≤ γ r ∗ p \leq_{\gamma} r^{*} p ≤ γ r ∗ , then there is a k k k -separation r ′ ∈ { r ∨ p , r ∧ p ∗ } r' \in\{r \vee p,\ r \wedge p^{*}\} r ′ ∈ { r ∨ p , r ∧ p ∗ } that is γ \gamma γ -close to r r r . In particular, p ≤ r ′ p \leq r' p ≤ r ′ or p ≤ ( r ′ ) ∗ p \leq(r')^{*} p ≤ ( r ′ ) ∗ .
Proof. Suppose first that γ < q + k \gamma< q+k γ < q + k . Since 2 q + k + 1 ≥ q + γ + 1 2q+k+1 \geq q+\gamma+1 2 q + k + 1 ≥ q + γ + 1 , both p p p and r r r are ( q + γ + 1 ) (q+\gamma+1) ( q + γ + 1 ) -broad. After possibly reversing r r r , assume that p ≤ γ r p \leq_{\gamma} r p ≤ γ r . By Lemma 5.7, r ∨ p r \vee p r ∨ p has order k k k , and by Lemma 5.5, it is γ \gamma γ -close to r r r . Undoing any reversal of r r r gives the required separation in { r ∨ p , r ∧ p ∗ } \{r \vee p,\ r \wedge p^{*}\} { r ∨ p , r ∧ p ∗ } . Hence we may assume that γ ≥ q + k \gamma\geq q+k γ ≥ q + k . Write r : = ( A , B ) r := (A,B) r := ( A , B ) and p : = ( C , D ) p := (C,D) p := ( C , D ) . By symmetry, we may orient r r r so that p ≼ γ r p \preccurlyeq_{\gamma} r p ≼ γ r . We first show that we can further choose this orientation so that the A C AC A C -corner has size > q >q > q .
Suppose that the A C AC A C -corner has size at most q q q . Then, ∣ A ∖ D ∣ = ∣ A ∩ ( C ∖ D ) ∣ ≤ ∣ A ∩ B ∣ + ∣ ( A ∖ B ) ∩ ( C ∖ D ) ∣ ≤ k + q ≤ γ |A \setminus D| = |A \cap(C \setminus D)| \leq|A \cap B| + |(A \setminus B) \cap(C \setminus D)| \leq k+q \leq\gamma ∣ A ∖ D ∣ = ∣ A ∩ ( C ∖ D ) ∣ ≤ ∣ A ∩ B ∣ + ∣ ( A ∖ B ) ∩ ( C ∖ D ) ∣ ≤ k + q ≤ γ , and similarly ∣ C ∖ B ∣ ≤ γ |C \setminus B| \leq\gamma ∣ C ∖ B ∣ ≤ γ so p ≼ γ r ∗ p \preccurlyeq_{\gamma} r^* p ≼ γ r ∗ . Since p p p is ( 2 q + k + 1 ) (2q+k+1) ( 2 q + k + 1 ) -broad and r r r has order k k k , we also have that the B C BC B C -corner has size > q >q > q . Thus, replacing r r r by r ∗ r^* r ∗ gives the desired orientation.
Now, let r ′ : = r ∨ p r' := r \vee p r ′ := r ∨ p . Since p ≼ γ r p \preccurlyeq_{\gamma} r p ≼ γ r , r ′ r' r ′ is γ \gamma γ -close to r r r by Lemma 5.5. Since r r r is ( q + γ + 1 ) (q+\gamma+1) ( q + γ + 1 ) -broad, Lemma 5.9 implies that r ′ r' r ′ has breadth > q >q > q . The opposite corner-separation r ∧ p r \wedge p r ∧ p is also ( > q ) (>q) ( > q ) -broad: its first proper side is the A C AC A C -corner and its second proper side contains B ∖ A B \setminus A B ∖ A .
Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, it follows that both corner-separations have order at least k k k . By submodularity, they both have order exactly k k k . □ \square □
Lemma 8.15. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, γ , ℓ ∈ N \gamma,\ell\in\mathbb{N} γ , ℓ ∈ N with ℓ > q \ell>q ℓ > q . Let σ \sigma σ be a finite star of ( 2 q + k + 1 ) (2q+k+1) ( 2 q + k + 1 ) -broad k k k -separations of G G G . Let β : = ℓ + γ ⋅ ∣ σ ∣ \beta:= \ell+\gamma\cdot|\sigma| β := ℓ + γ ⋅ ∣ σ ∣ . Let s s s be a β \beta β -broad k k k -separation of G G G that almost-γ \gamma γ -interlaces σ \sigma σ . Then there is an ℓ \ell ℓ -broad k k k -separation s ′ s' s ′ that almost-interlaces σ \sigma σ and is ( γ ⋅ ∣ σ ∣ ) (\gamma\cdot|\sigma|) ( γ ⋅ ∣ σ ∣ ) -close to $s.
Moreover, if τ ⊇ σ \tau\supseteq\sigma τ ⊇ σ is a star and s s s almost-interlaces τ ∖ σ \tau\setminus\sigma τ ∖ σ , then s ′ s' s ′ can be chosen to almost-interlace τ \tau τ .
Proof. We prove the assertion involving τ \tau τ by induction on ∣ σ ∣ |\sigma| ∣ σ ∣ . The first assertion then follows by taking τ = σ \tau=\sigma τ = σ . If σ = ∅ \sigma=\varnothing σ = ∅ , then we take s ′ : = s s':=s s ′ := s . Since β = ℓ \beta=\ell β = ℓ , s ′ s' s ′ is ℓ \ell ℓ -broad and almost-interlaces τ = τ ∖ σ \tau=\tau\setminus\sigma τ = τ ∖ σ and is 0 0 0 -close to s s s .
Now suppose that σ ≠ ∅ \sigma\neq\varnothing σ = ∅ . We choose p ∈ σ p\in\sigma p ∈ σ and define σ ′ = σ ∖ { p } \sigma'=\sigma\setminus\{p\} σ ′ = σ ∖ { p } . Since β ≥ ℓ + γ > q + γ \beta\geq\ell+\gamma>q+\gamma β ≥ ℓ + γ > q + γ , Lemma 8.14 gives a k k k -separation r ∈ { s ∨ p , s ∧ p ∗ } r\in\{s\vee p,s\wedge p^*\} r ∈ { s ∨ p , s ∧ p ∗ } that is γ \gamma γ -close to s s s . By Lemma 5.9, r r r is ( β − γ ) (\beta-\gamma) ( β − γ ) -broad, where β − γ = ℓ + γ ⋅ ∣ σ ′ ∣ \beta-\gamma=\ell+\gamma\cdot|\sigma'| β − γ = ℓ + γ ⋅ ∣ σ ′ ∣ . Moreover, p ≼ r p\preccurlyeq r p ≼ r or p ≼ r ∗ p\preccurlyeq r^* p ≼ r ∗ . By Lemma 8.10, applied to σ \sigma σ (with parameter γ \gamma γ ) and to τ ∖ σ \tau\setminus\sigma τ ∖ σ (with parameter 0 0 0 ), r r r almost-γ \gamma γ -interlaces σ ′ \sigma' σ ′ and almost-interlaces τ ∖ σ ′ \tau\setminus\sigma' τ ∖ σ ′ .
The induction hypothesis applied to r r r , σ ′ \sigma' σ ′ and τ \tau τ now yields an ℓ \ell ℓ -broad k k k -separation s ′ s' s ′ that almost-interlaces τ \tau τ and satisfies
s ′ ≈ γ ∣ σ ′ ∣ r ≈ γ s . s' \approx_{\gamma|\sigma'|} r \approx_{\gamma} s . s ′ ≈ γ ∣ σ ′ ∣ r ≈ γ s . By Lemma 4.9, s ′ ≈ γ ∣ σ ∣ s s' \approx_{\gamma|\sigma|} s s ′ ≈ γ ∣ σ ∣ s , as required. □ \square □
Using the previous results, we now prove that if a broad k k k -separation s s s almost-γ \gamma γ -interlaces a star σ \sigma σ of broad separations, either s s s is close to a k k k -separation that almost-interlaces σ \sigma σ , or the broad petal-separations in the graphic-sep( s ) (s) ( s ) -anemone are narrow-nested.
Lemma 8.16. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph, γ ∈ N \gamma\in\mathbb{N} γ ∈ N and
β : = β 8.16 ( q , k , b , γ ) : = max { q + 1 + ⌊ 3 k / 2 ⌋ ⋅ γ , b } and \beta:= \beta_{8.16}(q,k,b,\gamma) := \max\{q+1+\lfloor3k/2\rfloor\cdot\gamma,b\}\quad\text{and} β := β 8.16 ( q , k , b , γ ) := max { q + 1 + ⌊ 3 k /2 ⌋ ⋅ γ , b } and β ′ : = β 8.16 ′ ( q , k , b ) : = max { 2 q + k + 1 , b } and \beta' := \beta'_{8.16}(q,k,b) := \max\{2q+k+1,b\}\quad\text{and} β ′ := β 8.16 ′ ( q , k , b ) := max { 2 q + k + 1 , b } and ε : = ε 8.16 ( k , γ ) : = ⌊ 3 k / 2 ⌋ ⋅ γ . \varepsilon:= \varepsilon_{8.16}(k,\gamma) := \lfloor3k/2\rfloor\cdot\gamma. ε := ε 8.16 ( k , γ ) := ⌊ 3 k /2 ⌋ ⋅ γ .
@page 79/137 starts_mid=1 ends_mid=1 Hence we may assume that γ ≥ q + k \gamma\geq q+k γ ≥ q + k . Write r : = ( A , B ) r := (A,B) r := ( A , B ) and p : = ( C , D ) p := (C,D) p := ( C , D ) . By symmetry, we may orient r r r so that p ≼ γ r p \preccurlyeq_{\gamma} r p ≼ γ r . We first show that we can further choose this orientation so that the A C AC A C -corner has size > q >q > q .
Suppose that the A C AC A C -corner has size at most q q q . Then, ∣ A ∖ D ∣ = ∣ A ∩ ( C ∖ D ) ∣ ≤ ∣ A ∩ B ∣ + ∣ ( A ∖ B ) ∩ ( C ∖ D ) ∣ ≤ k + q ≤ γ |A \setminus D| = |A \cap(C \setminus D)| \leq|A \cap B| + |(A \setminus B) \cap(C \setminus D)| \leq k+q \leq\gamma ∣ A ∖ D ∣ = ∣ A ∩ ( C ∖ D ) ∣ ≤ ∣ A ∩ B ∣ + ∣ ( A ∖ B ) ∩ ( C ∖ D ) ∣ ≤ k + q ≤ γ , and similarly ∣ C ∖ B ∣ ≤ γ |C \setminus B| \leq\gamma ∣ C ∖ B ∣ ≤ γ so p ≼ γ r ∗ p \preccurlyeq_{\gamma} r^* p ≼ γ r ∗ . Since p p p is ( 2 q + k + 1 ) (2q+k+1) ( 2 q + k + 1 ) -broad and r r r has order k k k , we also have that the B C BC B C -corner has size > q >q > q . Thus, replacing r r r by r ∗ r^* r ∗ gives the desired orientation.
Now, let r ′ : = r ∨ p r' := r \vee p r ′ := r ∨ p . Since p ≼ γ r p \preccurlyeq_{\gamma} r p ≼ γ r , r ′ r' r ′ is γ \gamma γ -close to r r r by Lemma 5.5. Since r r r is ( q + γ + 1 ) (q+\gamma+1) ( q + γ + 1 ) -broad, Lemma 5.9 implies that r ′ r' r ′ has breadth > q >q > q . The opposite corner-separation r ∧ p r \wedge p r ∧ p is also ( > q ) (>q) ( > q ) -broad: its first proper side is the A C AC A C -corner and its second proper side contains B ∖ A B \setminus A B ∖ A .
Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, it follows that both corner-separations have order at least k k k . By submodularity, they both have order exactly k k k . □ \square □
Lemma 8.15. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, γ , ℓ ∈ N \gamma,\ell\in\mathbb{N} γ , ℓ ∈ N with ℓ > q \ell>q ℓ > q . Let σ \sigma σ be a finite star of ( 2 q + k + 1 ) (2q+k+1) ( 2 q + k + 1 ) -broad k k k -separations of G G G . Let β : = ℓ + γ ⋅ ∣ σ ∣ \beta:= \ell+\gamma\cdot|\sigma| β := ℓ + γ ⋅ ∣ σ ∣ . Let s s s be a β \beta β -broad k k k -separation of G G G that almost-γ \gamma γ -interlaces σ \sigma σ . Then there is an ℓ \ell ℓ -broad k k k -separation s ′ s' s ′ that almost-interlaces σ \sigma σ and is ( γ ⋅ ∣ σ ∣ ) (\gamma\cdot|\sigma|) ( γ ⋅ ∣ σ ∣ ) -close to $s.
Moreover, if τ ⊇ σ \tau\supseteq\sigma τ ⊇ σ is a star and s s s almost-interlaces τ ∖ σ \tau\setminus\sigma τ ∖ σ , then s ′ s' s ′ can be chosen to almost-interlace τ \tau τ .
Proof. We prove the assertion involving τ \tau τ by induction on ∣ σ ∣ |\sigma| ∣ σ ∣ . The first assertion then follows by taking τ = σ \tau=\sigma τ = σ . If σ = ∅ \sigma=\varnothing σ = ∅ , then we take s ′ : = s s':=s s ′ := s . Since β = ℓ \beta=\ell β = ℓ , s ′ s' s ′ is ℓ \ell ℓ -broad and almost-interlaces τ = τ ∖ σ \tau=\tau\setminus\sigma τ = τ ∖ σ and is 0 0 0 -close to s s s .
Now suppose that σ ≠ ∅ \sigma\neq\varnothing σ = ∅ . We choose p ∈ σ p\in\sigma p ∈ σ and define σ ′ = σ ∖ { p } \sigma'=\sigma\setminus\{p\} σ ′ = σ ∖ { p } . Since β ≥ ℓ + γ > q + γ \beta\geq\ell+\gamma>q+\gamma β ≥ ℓ + γ > q + γ , Lemma 8.14 gives a k k k -separation r ∈ { s ∨ p , s ∧ p ∗ } r\in\{s\vee p,s\wedge p^*\} r ∈ { s ∨ p , s ∧ p ∗ } that is γ \gamma γ -close to s s s . By Lemma 5.9, r r r is ( β − γ ) (\beta-\gamma) ( β − γ ) -broad, where β − γ = ℓ + γ ⋅ ∣ σ ′ ∣ \beta-\gamma=\ell+\gamma\cdot|\sigma'| β − γ = ℓ + γ ⋅ ∣ σ ′ ∣ . Moreover, p ≼ r p\preccurlyeq r p ≼ r or p ≼ r ∗ p\preccurlyeq r^* p ≼ r ∗ . By Lemma 8.10, applied to σ \sigma σ (with parameter γ \gamma γ ) and to τ ∖ σ \tau\setminus\sigma τ ∖ σ (with parameter 0 0 0 ), r r r almost-γ \gamma γ -interlaces σ ′ \sigma' σ ′ and almost-interlaces τ ∖ σ ′ \tau\setminus\sigma' τ ∖ σ ′ .
The induction hypothesis applied to r r r , σ ′ \sigma' σ ′ and τ \tau τ now yields an ℓ \ell ℓ -broad k k k -separation s ′ s' s ′ that almost-interlaces τ \tau τ and satisfies
s ′ ≈ γ ∣ σ ′ ∣ r ≈ γ s . s' \approx_{\gamma|\sigma'|} r \approx_{\gamma} s . s ′ ≈ γ ∣ σ ′ ∣ r ≈ γ s . By Lemma 4.9, s ′ ≈ γ ∣ σ ∣ s s' \approx_{\gamma|\sigma|} s s ′ ≈ γ ∣ σ ∣ s , as required. □ \square □
Using the previous results, we now prove that if a broad k k k -separation s s s almost-γ \gamma γ -interlaces a star σ \sigma σ of broad separations, either s s s is close to a k k k -separation that almost-interlaces σ \sigma σ , or the broad petal-separations in the graphic-sep( s ) (s) ( s ) -anemone are narrow-nested.
Lemma 8.16. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph, γ ∈ N \gamma\in\mathbb{N} γ ∈ N and
β : = β 8.16 ( q , k , b , γ ) : = max { q + 1 + ⌊ 3 k / 2 ⌋ ⋅ γ , b } and \beta:= \beta_{8.16}(q,k,b,\gamma) := \max\{q+1+\lfloor3k/2\rfloor\cdot\gamma,b\}\quad\text{and} β := β 8.16 ( q , k , b , γ ) := max { q + 1 + ⌊ 3 k /2 ⌋ ⋅ γ , b } and β ′ : = β 8.16 ′ ( q , k , b ) : = max { 2 q + k + 1 , b } and \beta' := \beta'_{8.16}(q,k,b) := \max\{2q+k+1,b\}\quad\text{and} β ′ := β 8.16 ′ ( q , k , b ) := max { 2 q + k + 1 , b } and ε : = ε 8.16 ( k , γ ) : = ⌊ 3 k / 2 ⌋ ⋅ γ . \varepsilon:= \varepsilon_{8.16}(k,\gamma) := \lfloor3k/2\rfloor\cdot\gamma. ε := ε 8.16 ( k , γ ) := ⌊ 3 k /2 ⌋ ⋅ γ .
Let σ \sigma σ be a star of β ′ \beta' β ′ -broad k k k -separations of G G G . Let s s s be a β \beta β -broad k k k -separation that almost-γ \gamma γ -interlaces σ \sigma σ . Then we have one of the following two outcomes:
(i) s s s is ε \varepsilon ε -close to some k k k -separation s ′ s' s ′ that almost-interlaces σ \sigma σ .
(ii) s s s is a displayed separation of the graphic-X X X -anemone A \mathcal{A} A with X : = sep ( s ) X := \operatorname{sep}(s) X := sep ( s ) , and the narrow proper petal-separations of A \mathcal{A} A are narrow-nested. Moreover, s s s is not half-connected.
Proof. The graph G G G is ( b , k ) (b,k) ( b , k ) -narrow and every separation in σ \sigma σ is b b b -broad, so σ \sigma σ is a star of narrow k k k -separations of G G G . Similarly, s s s is a narrow k k k -separation of G G G . Let ℓ : = ⌈ ( k + 1 ) / 2 ⌉ \ell:= \lceil(k + 1)/2 \rceil ℓ := ⌈( k + 1 ) /2 ⌉ . We distinguish the following two cases.
Case (i): s s s crosses < k + ℓ < k + \ell < k + ℓ many separations in σ \sigma σ . If s s s is γ \gamma γ -close to some p ∈ σ ∪ σ ∗ p \in\sigma\cup\sigma^* p ∈ σ ∪ σ ∗ , then p p p almost-interlaces σ \sigma σ , so taking s ′ : = p s' := p s ′ := p gives outcome (i), since γ ≤ ε \gamma\le\varepsilon γ ≤ ε . Hence assume that no such p p p exists.
Let σ ′ \sigma' σ ′ consist of the separations in σ \sigma σ that cross s s s . Then ∣ σ ′ ∣ ≤ k + ℓ − 1 ≤ ⌊ 3 k / 2 ⌋ |\sigma'| \le k + \ell- 1 \le\lfloor3k/2 \rfloor ∣ σ ′ ∣ ≤ k + ℓ − 1 ≤ ⌊ 3 k /2 ⌋ . We show that s s s almost-interlaces σ ∖ σ ′ \sigma\setminus\sigma' σ ∖ σ ′ . Otherwise, some p ∈ σ ∖ σ ′ p \in\sigma\setminus\sigma' p ∈ σ ∖ σ ′ satisfies neither p ≤ s p \le s p ≤ s nor p ≤ s ∗ p \le s^* p ≤ s ∗ . Thus, s ≤ p s \le p s ≤ p or s ∗ ≤ p s^* \le p s ∗ ≤ p . Without loss of generality, assume the former. Since p ≤ γ s p \le_\gamma s p ≤ γ s or p ≤ γ s ∗ p \le_\gamma s^* p ≤ γ s ∗ , either s ≈ γ p s \approx_\gamma p s ≈ γ p , which contradicts our assumption, or s ≤ γ s ∗ s \le_\gamma s^* s ≤ γ s ∗ , which contradicts that s s s is β \beta β -broad by Lemma 8.3 and β > γ \beta> \gamma β > γ .
Every member of σ ′ \sigma' σ ′ is ( 2 q + k + 1 ) (2q + k + 1) ( 2 q + k + 1 ) -broad, and s s s is β \beta β -broad with β ≥ q + 1 + γ ∣ σ ′ ∣ \beta\ge q + 1 + \gamma|\sigma'| β ≥ q + 1 + γ ∣ σ ′ ∣ . So we can apply Lemma 8.15 to σ ′ \sigma' σ ′ (with ℓ = q + 1 \ell= q + 1 ℓ = q + 1 and τ = σ \tau= \sigma τ = σ ). Its ‘Moreover’-part yields a k k k -separation s ′ s' s ′ that almost-interlaces σ \sigma σ and is ( γ ∣ σ ′ ∣ ) (\gamma|\sigma'|) ( γ ∣ σ ′ ∣ ) -close to s s s . Since γ ∣ σ ′ ∣ ≤ ⌊ 3 k / 2 ⌋ ⋅ γ = ε \gamma|\sigma'| \le\lfloor3k/2 \rfloor\cdot\gamma= \varepsilon γ ∣ σ ′ ∣ ≤ ⌊ 3 k /2 ⌋ ⋅ γ = ε , this gives outcome (i).
Case (ii): not Case (i). By Lemma 8.13, each proper side of s s s has ≥ ℓ \ge\ell ≥ ℓ components. Let X : = sep ( s ) X := \operatorname{sep}(s) X := sep ( s ) , and let A \mathcal{A} A be the graphic-X X X -anemone. Since sep ( s ) = X \operatorname{sep}(s) = X sep ( s ) = X , the separation s s s is a displayed separation of A \mathcal{A} A , which is the first part of (ii). Let C 1 , C 2 , … , C m C_1, C_2, \ldots, C_m C 1 , C 2 , … , C m be the connected components of G − X G-X G − X . With ℓ = ⌈ ( k + 1 ) / 2 ⌉ \ell= \lceil(k + 1)/2 \rceil ℓ = ⌈( k + 1 ) /2 ⌉ , we have m ≥ 2 ℓ ≥ k + 1 m \ge2\ell\ge k + 1 m ≥ 2 ℓ ≥ k + 1 . Then, s s s is not half-connected, otherwise Lemma 8.8 would imply that s s s is nested with every separation in σ \sigma σ , contradicting that we are not in Case (i). We define s i : = ( V ( C i ) ∪ X , V ( G ) ∖ V ( C i ) ) s_i := (V(C_i) \cup X, V(G) \setminus V(C_i)) s i := ( V ( C i ) ∪ X , V ( G ) ∖ V ( C i )) , which is a half-connected k k k -separation of G G G with separator sep ( s i ) = X = sep ( s ) \operatorname{sep}(s_i) = X = \operatorname{sep}(s) sep ( s i ) = X = sep ( s ) . Every narrow proper petal-separation of A \mathcal{A} A is some s i s_i s i . By Lemma 8.8, no narrow proper s i s_i s i is crossed by a narrow proper ( ≤ k ) (\le k) ( ≤ k ) -separation of G G G . Hence, every narrow proper s i s_i s i is narrow-nested, which yields the second part of outcome (ii) and thus completes the proof. □
Lemma 8.17. Let s s s and t t t be two β \beta β -broad ( ≤ k ) (\le k) ( ≤ k ) -separations of a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph G G G , where β > 2 q + k \beta> 2q + k β > 2 q + k . Then there are two opposite corner-separations of order k k k and breadth ≥ ( β − k ) / 2 \ge(\beta-k)/2 ≥ ( β − k ) /2 .
Proof. Let s : = ( A , B ) s := (A,B) s := ( A , B ) and t : = ( C , D ) t := (C,D) t := ( C , D ) . First, we show that there are two opposite corners each of size ≥ ( β − k ) / 2 \ge(\beta-k)/2 ≥ ( β − k ) /2 . Assume the contrary. Then either the A C AC A C -corner or the B D BD B D -corner has size < ( β − k ) / 2 < (\beta-k)/2 < ( β − k ) /2 , say the former. Similarly, either the A D AD A D -corner or the B C BC B C -corner has size < ( β − k ) / 2 < (\beta-k)/2 < ( β − k ) /2 , say the former. Then ∣ A ∖ B ∣ < ( β − k ) / 2 + ( β − k ) / 2 + ∣ sep ( t ) ∣ ≤ β |A \setminus B| < (\beta-k)/2 + (\beta-k)/2 + |\operatorname{sep}(t)| \le\beta ∣ A ∖ B ∣ < ( β − k ) /2 + ( β − k ) /2 + ∣ sep ( t ) ∣ ≤ β , contradicting that s s s is β \beta β -broad.
Hence, there are two opposite corners each of size ≥ ( β − k ) / 2 \ge(\beta-k)/2 ≥ ( β − k ) /2 . So, their corresponding corner-separations have breadth ≥ ( β − k ) / 2 > q \ge(\beta-k)/2 > q ≥ ( β − k ) /2 > q . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, these separations have order ≥ k \ge k ≥ k . By submodularity, they have order exactly k k k . □
We shall use the following variant of [28 Lemma 4.4] ; we include a proof here for completeness.
Lemma 8.18. Whenever two narrow proper separations cross with one empty link, they have the same separator. In particular, none of them can be half-connected.
Proof. Let s = ( A , B ) s=(A,B) s = ( A , B ) and t = ( C , D ) t=(C,D) t = ( C , D ) be two narrow proper separations that cross with one empty link, say the s C sC s C -link. First, we show that there are two adjacent empty links. Suppose for a contradiction that no two adjacent links are empty. Then, the A t At A t -link and the B t Bt B t -link are nonempty, say they contain vertices v v v and w w w , respectively. Since t t t is proper, there exists a connected component K K K of G − sep ( t ) G-\operatorname{sep}(t) G − sep ( t ) such that V ( K ) ⊆ C ∖ D V(K)\subseteq C\setminus D V ( K ) ⊆ C ∖ D . Since t t t is narrow, N ( K ) = sep ( t ) N(K)=\operatorname{sep}(t) N ( K ) = sep ( t ) so there exists a path between v v v and w w w all of whose internal vertices are in C ∖ D C\setminus D C ∖ D . Since v ∈ A ∖ B v\in A\setminus B v ∈ A ∖ B and w ∈ B ∖ A w\in B\setminus A w ∈ B ∖ A , this path intersects sep ( s ) \operatorname{sep}(s) sep ( s ) at some vertex u ∈ C ∖ D u\in C\setminus D u ∈ C ∖ D . This contradicts that the s C sC s C -link is empty.
Hence there are two adjacent empty links, say the A t At A t -link and the s C sC s C -link. Since s s s and t t t cross, the A C AC A C -corner is nonempty. Let K K K be a component of the graph induced by the A C AC A C -corner. Now K K K is a component of G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) and of G − sep ( t ) G-\operatorname{sep}(t) G − sep ( t ) . Since s s s and t t t are narrow, we have sep ( s ) = N ( K ) = sep ( t ) \operatorname{sep}(s)=N(K)=\operatorname{sep}(t) sep ( s ) = N ( K ) = sep ( t ) .
To see the ‘In particular’-part, note that half-connected separations are nested with all separations with the same separator.
Lemma 8.19. Let q , k , β ∈ N q,k,\beta\in\mathbb{N} q , k , β ∈ N with β ≥ β 8.19 ( q , k ) : = 4 q + k + 1 \beta\ge\beta_{8.19}(q,k):=4q+k+1 β ≥ β 8.19 ( q , k ) := 4 q + k + 1 , and let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph with a β \beta β -broad narrow ( ≤ k ) (\leq k) ( ≤ k ) -separation s s s such that G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) has at least three components. Then every β \beta β -broad narrow ( ≤ k ) (\leq k) ( ≤ k ) -separation that crosses s s s has the same separator as s s s .
In particular, if s s s is half-connected then s s s is S S S -totally-nested, where S S S is the set of β \beta β -broad narrow ( ≤ k ) (\leq k) ( ≤ k ) -separations of G G G .
Proof. Let s : = ( A , B ) s:=(A,B) s := ( A , B ) and let t : = ( C , D ) t:=(C,D) t := ( C , D ) be a β \beta β -broad narrow ( ≤ k ) (\leq k) ( ≤ k ) -separation that crosses s s s . By symmetry assume that G [ A ∖ B ] G[A\setminus B] G [ A ∖ B ] has at least two components. By Lemma 8.17 (with β > 2 q + k \beta>2q+k β > 2 q + k ), there are two opposite corner-separations of order k k k and breadth ≥ ⌈ ( β − k ) / 2 ⌉ ≥ 2 q + 1 \geq\lceil(\beta-k)/2\rceil\geq2q+1 ≥ ⌈( β − k ) /2 ⌉ ≥ 2 q + 1 . Without loss of generality, these are the A C AC A C -corner-separation and the B D BD B D -corner-separation.
First we show that the corner-diagram of s , t s,t s , t has an empty link. Suppose for a contradiction that all four links are nonempty. Thus since s s s is narrow, every component of G [ A ∖ B ] G[A\setminus B] G [ A ∖ B ] intersects the A t At A t -link. Let K K K be a component of G [ A ∖ B ] G[A\setminus B] G [ A ∖ B ] whose intersection with the A C AC A C -corner has minimal size. We define s ′ : = ( A ′ , B ′ ) : = ( A ∖ K , B ∪ K ) s':=(A',B'):=(A\setminus K,B\cup K) s ′ := ( A ′ , B ′ ) := ( A ∖ K , B ∪ K ) , which is a separation of G G G with sep ( s ′ ) = sep ( s ) \operatorname{sep}(s')=\operatorname{sep}(s) sep ( s ′ ) = sep ( s ) . Note that since K K K intersects the A t At A t -link, the A ′ C A'C A ′ C -corner-separation has order < k <k < k . Since G [ A ∖ B ] G[A\setminus B] G [ A ∖ B ] has at least two components and since we chose K K K minimal, the A ′ C A'C A ′ C -corner still contains ≥ ⌈ ( 2 q + 1 ) / 2 ⌉ > q \geq\lceil(2q+1)/2\rceil>q ≥ ⌈( 2 q + 1 ) /2 ⌉ > q vertices. This is a contradiction to the assumption that G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable.
Hence the corner diagram of s s s and t t t has an empty link, so s s s and t t t have the same separator by Lemma 8.18 , as desired. The ‘In-particular’-part follows from the ‘In particular’-part of Lemma 8.18 .
Lemma 8.20. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph and let γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let
δ : = δ 8.20 ( k , b , γ ) : = max { ⌊ 3 k / 2 ⌋ ⋅ γ , 2 ( γ + k + b ) } \delta:= \delta_{8.20}(k,b,\gamma) := \max\{\lfloor3k/2 \rfloor\cdot\gamma, 2(\gamma+k+b)\} δ := δ 8.20 ( k , b , γ ) := max {⌊ 3 k /2 ⌋ ⋅ γ , 2 ( γ + k + b )} β : = β 8.20 ( q , k , b ) : = max { β 8.16 ( q , k , b ) , β 8.19 ( q , k ) } \beta:= \beta_{8.20}(q,k,b) := \max\{\beta_{8.16}(q,k,b),\beta_{8.19}(q,k)\} β := β 8.20 ( q , k , b ) := max { β 8.16 ( q , k , b ) , β 8.19 ( q , k )} ε : = ε 8.20 ( q , k , b , γ ) : = δ + β . \varepsilon:= \varepsilon_{8.20}(q,k,b,\gamma) := \delta+\beta. ε := ε 8.20 ( q , k , b , γ ) := δ + β . Let N N N be a nested set of half-connected β \beta β -broad k k k -separations of G G G . Let s s s be an ε \varepsilon ε -broad k k k -separation of G G G that is γ \gamma γ -nested with every separation in N N N . Then there is a k k k -separation s ′ s' s ′ of G G G that is nested with N N N and δ \delta δ -close to s s s .
Proof. Without loss of generality, we can assume that N N N is symmetric.
Recall that s s s is γ \gamma γ -nested with every separation in N N N . By Lemma 8.6 (which we can apply since s s s has breadth ≥ ε = δ + β ≥ 2 γ + 1 \geq\varepsilon=\delta+\beta\geq2\gamma+1 ≥ ε = δ + β ≥ 2 γ + 1 ), either s s s is γ \gamma γ -close to a separation in N N N or s s s γ \gamma γ -interlaces some splitting star σ \sigma σ of N N N . We are done in the former case (since δ ≥ γ \delta\geq\gamma δ ≥ γ ), so we can assume the latter case. Observe that s s s has breadth ε ≥ β 8.16 ( q , k , b , γ ) \varepsilon\geq\beta_{8.16}(q,k,b,\gamma) ε ≥ β 8.16 ( q , k , b , γ ) and δ ≥ ε 8.16 ( k , γ ) \delta\geq\varepsilon_{8.16}(k,\gamma) δ ≥ ε 8.16 ( k , γ ) .
Note that every separation in σ \sigma σ is β \beta β -broad, hence β 8.16 ( q , k , b ) \beta_{8.16}(q,k,b) β 8.16 ( q , k , b ) -broad. So, we can apply Lemma 8.16 to the separation s s s and the star σ \sigma σ , and have one of the following two outcomes.
Case (i): s s s is δ \delta δ -close to some k k k -separation s ′ s' s ′ that almost-interlaces σ \sigma σ . Since s ′ s' s ′ almost-interlaces σ \sigma σ , it is nested with every separation in N N N by Lemma 8.12 and we are done.
Case (ii): s s s is a displayed separation of the graphic-X X X -anemone A \mathcal{A} A with X : = sep ( s ) X:=\operatorname{sep}(s) X := sep ( s ) , and the narrow proper petal-separations of A \mathcal{A} A are narrow-nested. Let p 1 , p 2 , … , p m p_1,p_2,\ldots,p_m p 1 , p 2 , … , p m be the petal-separations of A \mathcal{A} A . By Lemma 3.13 applied to A \mathcal{A} A and s s s with ε ′ : = γ + k + b \varepsilon':=\gamma+k+b ε ′ := γ + k + b , we have one of the following two cases.
Case (ii) (i): s s s is 2 ε ′ 2\varepsilon' 2 ε ′ -close to some p i p_i p i or its inverse. Since δ ≥ 2 ( γ + k + b ) = 2 ε ′ \delta\geq2(\gamma+k+b)=2\varepsilon' δ ≥ 2 ( γ + k + b ) = 2 ε ′ , this separation is δ \delta δ -close to s s s . Moreover, the separation p i p_i p i has breadth ≥ ε − 2 ε ′ ≥ ε − δ ≥ β \geq\varepsilon-2\varepsilon'\geq\varepsilon-\delta\geq\beta ≥ ε − 2 ε ′ ≥ ε − δ ≥ β by Lemma 5.9 . Since p i p_i p i is β \beta β -broad, it is b b b -broad, hence narrow since G G G is ( b , k ) (b,k) ( b , k ) -narrow. Thus, p i p_i p i is narrow-nested by assumption. Furthermore, all separations in N N N are b b b -broad, and G G G is ( b , k ) (b,k) ( b , k ) -narrow, so all separations in N N N are narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separations of G G G . Therefore, p i p_i p i is nested with N N N , and thus p i ∗ p_i^* p i ∗ is also nested with N N N . We put s ′ : = p i s':=p_i s ′ := p i or s ′ : = p i ∗ s':=p_i^* s ′ := p i ∗ and are done.
Case (ii) (ii): A \mathcal{A} A has a displayed separation that ( ε ′ − k ) (\varepsilon'-k) ( ε ′ − k ) -crisses s s s . So, G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) has at least three components. Moreover, s s s is β 8.19 ( q , k ) \beta_{8.19}(q,k) β 8.19 ( q , k ) -broad and narrow (since it is b b b -broad and G G G is ( b , k ) (b,k) ( b , k ) -narrow). Similarly, every separation in N N N is β 8.19 ( q , k ) \beta_{8.19}(q,k) β 8.19 ( q , k ) -broad and narrow. By Lemma 8.19 and since every separation in N N N is half-connected, s s s is nested with N N N . We put s ′ : = s s':=s s ′ := s and are done. □
We can now prove Key Lemma 8.1, which we restate for convenience.
Key Lemma 8.1. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph and let γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let
δ : = δ 8.1 ( k , b , γ ) : = 3 δ 8.20 ( k , b , γ ) + 2 γ + 2 k and γ ′ : = γ 8.1 ′ ( k , b , γ ) : = δ + γ and β : = β 8.1 ( q , k , b ) : = β 8.20 ( q , k , b ) and ε : = ε 8.1 ( q , k , b , γ ) : = max { ε 8.20 ( q , k , b , γ ) , δ + β } and δ ′ : = δ 8.1 ′ ( k , b , γ ) : = δ 8.20 ( k , b , γ ′ ) and ε ′ : = ε 8.1 ′ ( q , k , b , γ ) : = ε 8.20 ( q , k , b , γ ′ ) . \begin{aligned}
\delta&:= \delta_{8.1}(k,b,\gamma) := 3\delta_{8.20}(k,b,\gamma) + 2\gamma+ 2k && \text{and}\\
\gamma' &:= \gamma'_{8.1}(k,b,\gamma) := \delta+ \gamma&& \text{and}\\
\beta&:= \beta_{8.1}(q,k,b) := \beta_{8.20}(q,k,b) && \text{and}\\
\varepsilon&:= \varepsilon_{8.1}(q,k,b,\gamma) := \max\{\varepsilon_{8.20}(q,k,b,\gamma),\delta+\beta\} && \text{and}\\
\delta' &:= \delta'_{8.1}(k,b,\gamma) := \delta_{8.20}(k,b,\gamma') && \text{and}\\
\varepsilon' &:= \varepsilon'_{8.1}(q,k,b,\gamma) := \varepsilon_{8.20}(q,k,b,\gamma').
\end{aligned} δ γ ′ β ε δ ′ ε ′ := δ 8.1 ( k , b , γ ) := 3 δ 8.20 ( k , b , γ ) + 2 γ + 2 k := γ 8.1 ′ ( k , b , γ ) := δ + γ := β 8.1 ( q , k , b ) := β 8.20 ( q , k , b ) := ε 8.1 ( q , k , b , γ ) := max { ε 8.20 ( q , k , b , γ ) , δ + β } := δ 8.1 ′ ( k , b , γ ) := δ 8.20 ( k , b , γ ′ ) := ε 8.1 ′ ( q , k , b , γ ) := ε 8.20 ( q , k , b , γ ′ ) . and and and and and Let S S S be the set of all narrow-γ \gamma γ -nested k k k -separations of G G G . Let S ′ S' S ′ be the set of all half-connected narrow-γ ′ \gamma' γ ′ -nested β \beta β -broad k k k -separations of G G G that are δ \delta δ -close to some separation in S S S . Let N ⊆ S ′ N \subseteq S' N ⊆ S ′ be a maximal nested set of separations in S ′ S' S ′ .
Then N N N ε \varepsilon ε -approximates S S S . Moreover, for every ε ′ \varepsilon' ε ′ -broad k k k -separation of G G G there is a δ ′ \delta' δ ′ -close k k k -separation of G G G that is nested with N N N .
Proof of Key Lemma 8.1. First we show that N N N ε \varepsilon ε -approximates S S S . Let s ∈ S s \in S s ∈ S be arbitrary. If s s s is not ε \varepsilon ε -broad there is nothing to show, so we can assume that s s s is ε \varepsilon ε -broad. The graph G G G is ( b , k ) (b,k) ( b , k ) -narrow and all separations in N N N are b b b -broad k k k -separations, so they are narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separations of G G G . Since s s s is narrow-γ \gamma γ -nested, it is γ \gamma γ -nested with every separation in N N N . By Lemma 8.20, there is a k k k -separation s ′ s' s ′ of G G G that is nested with N N N and δ 8.20 ( k , b , γ ) \delta_{8.20}(k,b,\gamma) δ 8.20 ( k , b , γ ) -close to s s s . By Lemma 5.9, s ′ s' s ′ has breadth at least ε − δ 8.20 ( k , b , γ ) ≥ ε − δ ≥ β ≥ b \varepsilon-\delta_{8.20}(k,b,\gamma) \geq\varepsilon-\delta\geq\beta\geq b ε − δ 8.20 ( k , b , γ ) ≥ ε − δ ≥ β ≥ b , so s ′ s' s ′ is narrow. By Observation 8.11, s ′ s' s ′ is narrow-( δ 8.20 ( k , b , γ ) + γ ) (\delta_{8.20}(k,b,\gamma)+\gamma) ( δ 8.20 ( k , b , γ ) + γ ) -nested. This implies that no k k k -separation of G G G ( δ 8.20 ( k , b , γ ) + γ ) (\delta_{8.20}(k,b,\gamma)+\gamma) ( δ 8.20 ( k , b , γ ) + γ ) -crosses s ′ s' s ′ . Indeed, if a k k k -separation t t t of G G G ( δ 8.20 ( k , b , γ ) + γ ) (\delta_{8.20}(k,b,\gamma)+\gamma) ( δ 8.20 ( k , b , γ ) + γ ) -crosses s ′ s' s ′ then, using that δ 8.20 ( k , b , γ ) ≥ b + k \delta_{8.20}(k,b,\gamma)\geq b+k δ 8.20 ( k , b , γ ) ≥ b + k , we have that t t t is b b b -broad by Lemma 5.6. But then t t t is narrow since G G G is ( b , k ) (b,k) ( b , k ) -narrow, which contradicts that s ′ s' s ′ is narrow-( δ 8.20 ( k , b , γ ) + γ ) (\delta_{8.20}(k,b,\gamma)+\gamma) ( δ 8.20 ( k , b , γ ) + γ ) -nested.
Let A \mathcal{A} A be the graphic-sep( s ′ ) (s') ( s ′ ) -anemone. By Lemma 3.13, there is a separation p ∈ π ( A ) ∪ π ( A ) ∗ p \in\pi(\mathcal{A}) \cup\pi(\mathcal{A})^* p ∈ π ( A ) ∪ π ( A ) ∗ that is 2 ( δ 8.20 ( k , b , γ ) + γ + k ) 2(\delta_{8.20}(k,b,\gamma)+\gamma+k) 2 ( δ 8.20 ( k , b , γ ) + γ + k ) -close to s ′ s' s ′ . If s ′ s' s ′ is half-connected, we choose p = s ′ p=s' p = s ′ . Note that p p p is half-connected and δ \delta δ -close to s ∈ S s \in S s ∈ S . Since s s s is ε \varepsilon ε -broad and ε ≥ δ + β \varepsilon\geq\delta+\beta ε ≥ δ + β , we have that p p p is β \beta β -broad by Lemma 5.9. Since β > q \beta>q β > q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, the separation p p p has order k k k . Since β ≥ b \beta\geq b β ≥ b and G G G is ( b , k ) (b,k) ( b , k ) -narrow, the separation p p p is narrow. By Observation 8.11 and with γ ′ = δ + γ \gamma'=\delta+\gamma γ ′ = δ + γ , p p p is narrow-γ ′ \gamma' γ ′ -nested. Hence, p ∈ S ′ p \in S' p ∈ S ′ . If s ′ s' s ′ is half-connected, then p = s ′ p=s' p = s ′ ; so p p p is nested with N N N . Otherwise, G − sep ( s ′ ) = G − sep ( p ) G-\operatorname{sep}(s')=G-\operatorname{sep}(p) G − sep ( s ′ ) = G − sep ( p ) has at least three components, so by Lemma 8.19 and since β ≥ β 8.19 ( q , k ) \beta\geq\beta_{8.19}(q,k) β ≥ β 8.19 ( q , k ) , the separation p p p is nested with N N N . Since N N N is maximal, we have p ∈ N p \in N p ∈ N . This completes the proof that N N N ε \varepsilon ε -approximates S S S .
We now show the ‘Moreover’-part. Let s s s be an ε ′ \varepsilon' ε ′ -broad k k k -separation of G G G . So s s s is b b b -broad and hence narrow since G G G is ( b , k ) (b,k) ( b , k ) -narrow. Since every separation in N N N is narrow-γ ′ \gamma' γ ′ -nested, s s s is γ ′ \gamma' γ ′ -nested with every separation in N N N . By Lemma 8.20 there is a k k k -separation s ′ s' s ′ of G G G that is nested with N N N and δ ′ \delta' δ ′ -close to s s s . ∎
The structure of the torsos
In this section, we locate the flowers within the tree-decomposition. We show that asterisks of sufficiently large spread determine distinct splitting stars and that each such star and its corresponding petal-star approximate each other. We then prove that the remaining torsos have higher connectivity.
This enables us to introduce and prove a technical strengthening (Theorem 9.31 and Theorem 9.32 ) of the main result (Theorem 2.6 ), under two additional assumptions: On the one hand, we assume that the graph G G G we are to decompose is ( b , k ) (b,k) ( b , k ) -narrow. On the other hand, we assume that virtually all key lemmas from the previous sections have been combined, with suitably chosen constants, to create a setting (9.5) in which we will work for the vast majority of this section. Assuming Setting 9.5 helps us keep the paper fairly modular. Later, in § 11 we will show how to establish Setting 9.5 using the key lemmas from the previous sections.
Preparations
We start by establishing an equivalence between totally-nested separations and totally-nested separators in Lemma 9.1 . Moreover, we strengthen Lemma 8.16 in the special case that the given star σ \sigma σ of separations is a petal-star of a graphic-flower (see Lemma 9.4 ).
Let G G G be a graph and let X ⊆ V ( G ) X \subseteq V(G) X ⊆ V ( G ) . Let S S S be a set of separations of G G G . We say that X X X is S S S -totally-nested if no separation in S S S separates two vertices of X X X .
Lemma 9.1. Let G G G be a graph, and let X ⊆ V ( G ) X \subseteq V(G) X ⊆ V ( G ) . Let S S S be the set of all narrow proper k k k -separations of G G G , and let S X ⊆ S S_X \subseteq S S X ⊆ S be the set of half-connected separations in S S S with separator X X X . Assume S X ≠ ∅ S_X \ne\varnothing S X = ∅ . Then the following assertions are equivalent:
(i) X X X is S S S -totally-nested;
(ii) every separation in S X S_X S X is S S S -totally-nested.
(iii) some separation in S X S_X S X is S S S -totally-nested.
Proof. (i) ⇒ \Rightarrow ⇒ (ii). Let X X X be S S S -totally-nested. Assume for a contradiction that there is a half-connected narrow proper separation s = ( A , B ) s=(A,B) s = ( A , B ) with separator X = A ∩ B X=A\cap B X = A ∩ B that is crossed by some separation t = ( C , D ) ∈ S t=(C,D)\in S t = ( C , D ) ∈ S . Assume without loss of generality that G [ A ∖ B ] G[A\setminus B] G [ A ∖ B ] is connected.
First, we show that the A t At A t -link is nonempty. Suppose that it is empty. Since G [ A ∖ B ] G[A\setminus B] G [ A ∖ B ] is connected, the A C AC A C -corner or the A D AD A D -corner is empty, say the former. Then the A D AD A D -corner is nonempty and there is a component of G − X G-X G − X contained in the A D AD A D -corner. Since s s s is narrow, the s C sC s C -link is empty. By the fact that the A t At A t -link, the s C sC s C -link and the A C AC A C -corner are empty, it follows that s s s and t t t are nested, a contradiction. Hence the A t At A t -link is nonempty.
Then, since s s s and t t t have the same order, either the s C sC s C -link or the s D sD sD -link is nonempty, say the s C sC s C -link. Since X X X is S S S -totally-nested, the s D sD sD -link is empty. Since s s s and t t t cross, either the B t Bt B t -link is nonempty or the B D BD B D -corner is nonempty. If the B t Bt B t -link is nonempty, every component of G [ D ∖ C ] G[D\setminus C] G [ D ∖ C ] has neighbourhood of size < k <k < k , contradicting that t t t is narrow. If the B t Bt B t -link is empty, there is a component of G [ D ∖ C ] G[D\setminus C] G [ D ∖ C ] contained in the B D BD B D -corner. This component has a neighbourhood of size < k <k < k , contradicting that t t t is narrow.
(ii) ⇒ \Rightarrow ⇒ (iii). This holds since S X ≠ ∅ S_X\ne\varnothing S X = ∅ . (iii) ⇒ \Rightarrow ⇒ (i). Suppose that X X X is not S S S -totally-nested. Let t = ( C , D ) ∈ S t=(C,D)\in S t = ( C , D ) ∈ S be a separation that separates two vertices of X X X . Let s s s be any narrow proper half-connected separation with separator X X X . Since the s C sC s C -link and the s D sD sD -link are both nonempty, the separations s s s and t t t cross. Hence, no separation in S X S_X S X is S S S -totally-nested. □ \square □
Lemma 9.2. Let G G G be a graph and let β ∈ N \beta\in\mathbb{N} β ∈ N . Let F F F be a graphic-k k k -flower in G G G that has a β \beta β -broad displayed separation. Then there is a k k k -flower F ′ F' F ′ such that the following holds:
(i) every separation induced by F ′ F' F ′ is induced by F F F ;
(ii) every petal-separation of F ′ F' F ′ is β \beta β -broad;
(iii) for every petal-separation t ′ = ( C , D ) t'=(C,D) t ′ = ( C , D ) of F ′ F' F ′ such that ∣ C ∖ D ∣ ≥ β + 2 ε \lvert C\setminus D\rvert\ge\beta+2\varepsilon ∣ C ∖ D ∣ ≥ β + 2 ε , there is a β \beta β -broad petal-separation t t t of F F F which is ε \varepsilon ε -close to t ′ t' t ′ , where ε : = ε 9.2 ( β , k ) : = β + k \varepsilon:=\varepsilon_{9.2}(\beta,k):=\beta+k ε := ε 9.2 ( β , k ) := β + k ; and
(iv) if a separation s s s of G G G almost-γ \gamma γ -interlaces the petal-star π \pi π of F F F for some γ ∈ N \gamma\in\mathbb{N} γ ∈ N , then s s s almost-γ ′ \gamma' γ ′ -interlaces the petal-star of F ′ F' F ′ for γ ′ : = γ 9.2 ′ ( β , k , γ ) : = max { 2 ε + β + k , ε + γ } \gamma':=\gamma'_{9.2}(\beta,k,\gamma):=\max\{2\varepsilon+\beta+k,\varepsilon+\gamma\} γ ′ := γ 9.2 ′ ( β , k , γ ) := max { 2 ε + β + k , ε + γ } .
Proof. Let t t t be a β \beta β -broad displayed separation of F F F . By Lemma 3.12, there is a cycle-decomposition O : = ( O , V ) \mathcal{O}:=(O,\mathcal{V}) O := ( O , V ) of G G G which displays t t t , such that every displayed separation of O \mathcal{O} O is a displayed separation of F F F , and every petal-separation of O \mathcal{O} O is a petal-separation of F F F . Let B ⊆ V ( O ) B\subseteq V(O) B ⊆ V ( O ) be the set of nodes of O O O whose corresponding petal-separation is β \beta β -broad. Our strategy is to ‘merge’ nodes of O O O that are not in B B B with adjacent nodes in order to obtain the desired k k k -flower F ′ F' F ′ , as follows.
We call a path Q ⊆ O Q\subseteq O Q ⊆ O good if the separation induced by ( Q , O − Q ) (Q,O-Q) ( Q , O − Q ) is β \beta β -broad, and Q Q Q does not have a proper subpath Q ′ Q' Q ′ such that the separation induced by ( Q ′ , O − Q ′ ) (Q',O-Q') ( Q ′ , O − Q ′ ) is β \beta β -broad. Note that every good path either consists of a single node of B B B or is contained in O − B O-B O − B .
Let Q \mathcal{Q} Q be a maximal set of pairwise disjoint good paths. Since O O O displays the β \beta β -broad separation t t t , and since t ∗ t^* t ∗ is a β \beta β -broad displayed separation of O O O too, we have ∣ Q ∣ ≥ 2 \lvert\mathcal{Q}\rvert\ge2 ∣ Q ∣ ≥ 2 . Then, we can write O = P 0 Q 0 P 1 Q 1 … P m − 1 Q m − 1 O=P_0Q_0P_1Q_1\ldots P_{m-1}Q_{m-1} O = P 0 Q 0 P 1 Q 1 … P m − 1 Q m − 1 with m : = ∣ Q ∣ ≥ 2 m:=\lvert\mathcal{Q}\rvert\ge2 m := ∣ Q ∣ ≥ 2 , such that V ( P i ) ⊆ V ( O − ⋃ Q ) V(P_i)\subseteq V(O-\bigcup\mathcal{Q}) V ( P i ) ⊆ V ( O − ⋃ Q ) where P i P_i P i is a (possibly empty) path and { Q i : i ∈ Z m } = Q \{Q_i:i\in\mathbb{Z}_m\}=\mathcal{Q} { Q i : i ∈ Z m } = Q . We define F ′ : = ( Z m , V ′ ) F':=(\mathbb{Z}_m,\mathcal{V}') F ′ := ( Z m , V ′ ) with V ′ : = ( V o ′ : o ∈ Z m ) \mathcal{V}':=(\mathcal{V}'_o:o\in\mathbb{Z}_m) V ′ := ( V o ′ : o ∈ Z m ) where V o ′ : = V ↾ ( P o ∪ Q o ) \mathcal{V}'_o:=\mathcal{V}\mathord{\upharpoonright}(P_o\cup Q_o) V o ′ := V ↾ ( P o ∪ Q o ) (that is, for each o o o we ‘merge’ the bags and nodes of O O O that correspond to the path P o ∪ Q o P_o\cup Q_o P o ∪ Q o ). Observe that F ′ F' F ′ is a k k k -flower. We show that F ′ F' F ′ satisfies (i)–(iv).
(i). By construction, every separation induced by F ′ F' F ′ is induced by O O O , hence by F F F .
(ii). This follows from m ≥ 2 m\ge2 m ≥ 2 .
(iii). Let t ′ = ( C , D ) t'=(C,D) t ′ = ( C , D ) be a petal-separation of F ′ F' F ′ such that ∣ C ∖ D ∣ ≥ β + 2 ε \lvert C\setminus D\rvert\ge\beta+2\varepsilon ∣ C ∖ D ∣ ≥ β + 2 ε , corresponding to some node o ∈ Z m o\in\mathbb{Z}_m o ∈ Z m , so C = V o ′ C=V'_o C = V o ′ . First, we show that Q o Q_o Q o consists of a single node of B B B . Suppose not for a contradiction, so Q o Q_o Q o is a path with at least two vertices. Then, there are nonempty paths Q o , 1 , Q o , 2 ⊆ Q o Q_{o,1},Q_{o,2}\subseteq Q_o Q o , 1 , Q o , 2 ⊆ Q o such that V ( Q o ) = V ( Q o , 1 ) ∪ V ( Q o , 2 ) V(Q_o)=V(Q_{o,1})\cup V(Q_{o,2}) V ( Q o ) = V ( Q o , 1 ) ∪ V ( Q o , 2 ) . Since Q o Q_o Q o is good, the separation induced by ( Q o , i , O − Q o , i ) (Q_{o,i},O-Q_{o,i}) ( Q o , i , O − Q o , i ) has breadth < β <\beta < β (on the left side, for i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } ). Furthermore, F F F is a graphic-k k k -flower and every displayed separation of O O O is a displayed separation of F F F , so the adhesion-set of the edge of Q o ⊆ O Q_o\subseteq O Q o ⊆ O joining Q o , 1 Q_{o,1} Q o , 1 and Q o , 2 Q_{o,2} Q o , 2 has size ≤ k \le k ≤ k . Thus, the separation induced by ( Q o , O − Q o ) (Q_o,O-Q_o) ( Q o , O − Q o ) has breadth < 2 β + k <2\beta+k < 2 β + k (on the left side). Moreover, the separation induced by ( P o , O − P o ) (P_o,O-P_o) ( P o , O − P o ) has breadth < β <\beta < β (on the left side), since otherwise we could find some good subpath of P o P_o P o , contradicting the maximality of Q \mathcal{Q} Q . Then it similarly follows that the separation induced by ( P O Q O , O − P O Q O ) (P_OQ_O,O-P_OQ_O) ( P O Q O , O − P O Q O ) has breadth (on the left side) < ( 2 β + k ) + β + k = 3 β + 2 k = β + 2 ε < (2\beta+k)+\beta+k=3\beta+2k=\beta+2\varepsilon < ( 2 β + k ) + β + k = 3 β + 2 k = β + 2 ε (where we use ε = ε 9.2 ( β , k ) = β + k \varepsilon=\varepsilon_{9.2}(\beta,k)=\beta+k ε = ε 9.2 ( β , k ) = β + k ), a contradiction with ∣ C ∖ D ∣ ≥ β + 2 ε |C\setminus D|\ge\beta+2\varepsilon ∣ C ∖ D ∣ ≥ β + 2 ε .
Thus, Q O Q_O Q O consists of a single node b ∈ B b\in B b ∈ B . Recall that the separation induced by ( P O , O − P O ) (P_O,O-P_O) ( P O , O − P O ) has breadth < β <\beta < β (on the left side) and the adhesion-set of the edge of O O O joining P O P_O P O and Q O Q_O Q O has size ≤ k \le k ≤ k , as before. Then it follows that the petal-separation t ′ t' t ′ is ε \varepsilon ε -close to the petal-separation of O O O (hence of F F F ) corresponding to the node b b b .
(iv). Let s s s be a separation of G G G that almost-γ \gamma γ -interlaces the petal-star π \pi π of F F F for some γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let t ′ = ( A ′ , B ′ ) t'=(A',B') t ′ = ( A ′ , B ′ ) be an arbitrary petal-separation of F ′ F' F ′ . We show that we have t ′ ≤ γ ′ s t'\le_{\gamma'}s t ′ ≤ γ ′ s or t ′ ≤ γ ′ s ∗ t'\le_{\gamma'}s^* t ′ ≤ γ ′ s ∗ . If ∣ A ′ ∖ B ′ ∣ < γ ′ − k |A'\setminus B'|<\gamma'-k ∣ A ′ ∖ B ′ ∣ < γ ′ − k , then this is trivially satisfied. Hence, suppose that ∣ A ′ ∖ B ′ ∣ ≥ γ ′ − k ≥ β + 2 ε |A'\setminus B'|\ge\gamma'-k\ge\beta+2\varepsilon ∣ A ′ ∖ B ′ ∣ ≥ γ ′ − k ≥ β + 2 ε .
By (iii), there is a petal-separation t t t of F F F which is ε \varepsilon ε -close to t ′ t' t ′ . Since s s s almost-γ \gamma γ -interlaces π \pi π , we have t ′ ≈ ε t ≤ γ s t'\approx_\varepsilon t\le_\gamma s t ′ ≈ ε t ≤ γ s or t ′ ≈ ε t ≤ γ s ∗ t'\approx_\varepsilon t\le_\gamma s^* t ′ ≈ ε t ≤ γ s ∗ . By Lemma 4.9 , we have t ′ ≤ γ + ε s t'\le_{\gamma+\varepsilon}s t ′ ≤ γ + ε s or t ′ ≤ γ + ε s ∗ t'\le_{\gamma+\varepsilon}s^* t ′ ≤ γ + ε s ∗ . So, s s s almost-γ ′ \gamma' γ ′ -interlaces the petal-star of F ′ F' F ′ .
Lemma 9.3. Let G G G be a graph, let F F F be a graphic-k k k -flower of G G G and let π \pi π be the petal-star of F F F . Let β ′ , γ ∈ N \beta',\gamma\in\mathbb{N} β ′ , γ ∈ N , and let β ≥ β 9.3 ( k , β ′ , γ ) : = max { β ′ + γ , 2 ( β ′ + k ) } \beta\ge\beta_{9.3}(k,\beta',\gamma):=\max\{\beta'+\gamma,2(\beta'+k)\} β ≥ β 9.3 ( k , β ′ , γ ) := max { β ′ + γ , 2 ( β ′ + k )} . Suppose that there exists a β \beta β -broad k k k -separation s s s that almost-γ \gamma γ -interlaces π \pi π . Then, F F F has a β ′ \beta' β ′ -broad displayed separation.
Proof. Write F = ( H , V ) F=(H,\mathcal{V}) F = ( H , V ) and s = ( A , B ) s=(A,B) s = ( A , B ) . We distinguish the following two cases.
Case 1: There exists a petal V h V_h V h with h ∈ V ( H ) h\in V(H) h ∈ V ( H ) such that ∣ V h ∣ ≥ β ′ + k |V_h|\ge\beta'+k ∣ V h ∣ ≥ β ′ + k . Let p = ( C , D ) ∈ π p=(C,D)\in\pi p = ( C , D ) ∈ π be the petal-separation at h h h , with C = V h C=V_h C = V h . Since s s s almost-γ \gamma γ -interlaces π \pi π , we may assume that ( C , D ) ≤ γ ( A , B ) (C,D)\le_\gamma(A,B) ( C , D ) ≤ γ ( A , B ) , so ∣ B ∖ D ∣ ≤ γ |B\setminus D|\le\gamma ∣ B ∖ D ∣ ≤ γ . Thus, ∣ D ∣ ≥ ∣ B ∣ − ∣ B ∖ D ∣ ≥ β + k − γ ≥ β ′ + k |D|\ge|B|-|B\setminus D|\ge\beta+k-\gamma\ge\beta'+k ∣ D ∣ ≥ ∣ B ∣ − ∣ B ∖ D ∣ ≥ β + k − γ ≥ β ′ + k . Since p p p is displayed by F F F , p p p has order at most k k k , so p p p is β ′ \beta' β ′ -broad.
Case 2: Otherwise, every petal V h V_h V h with h ∈ V ( H ) h\in V(H) h ∈ V ( H ) satisfies ∣ V h ∣ < β ′ + k |V_h|<\beta'+k ∣ V h ∣ < β ′ + k . Since s s s is β \beta β -broad, we have ∣ V ( G ) ∣ ≥ 2 β ≥ 4 ( β ′ + k ) |V(G)|\ge2\beta\ge4(\beta'+k) ∣ V ( G ) ∣ ≥ 2 β ≥ 4 ( β ′ + k ) . Let d = ( C , D ) d=(C,D) d = ( C , D ) be a displayed separation of F F F with ∣ C ∣ ≥ β ′ + k |C|\ge\beta'+k ∣ C ∣ ≥ β ′ + k and ∣ C ∣ |C| ∣ C ∣ minimum under these conditions. By minimality of ∣ C ∣ |C| ∣ C ∣ together with ∣ V h ∣ < β ′ + k |V_h|<\beta'+k ∣ V h ∣ < β ′ + k for every h ∈ V ( H ) h\in V(H) h ∈ V ( H ) , we have ∣ C ∣ < 2 ( β ′ + k ) |C|<2(\beta'+k) ∣ C ∣ < 2 ( β ′ + k ) . Thus, ∣ D ∣ ≥ 4 ( β ′ + k ) − 2 ( β ′ + k ) ≥ β ′ + k |D|\ge4(\beta'+k)-2(\beta'+k)\ge\beta'+k ∣ D ∣ ≥ 4 ( β ′ + k ) − 2 ( β ′ + k ) ≥ β ′ + k . Since d d d is displayed by F F F , d d d has order at most k k k , hence d d d is β ′ \beta' β ′ -broad. □
Lemma 9.4. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph and γ ∈ N \gamma\in\mathbb{N} γ ∈ N . Let
γ ′ : = γ 9.2 ′ ( β 8.16 ′ ( q , k , b ) , k , γ ) , \gamma' := \gamma'_{9.2}\bigl(\beta'_{8.16}(q,k,b),k,\gamma\bigr), γ ′ := γ 9.2 ′ ( β 8.16 ′ ( q , k , b ) , k , γ ) , ε : = ε 9.4 ( q , k , b , γ ) : = ε 8.16 ( k , γ ′ ) , \varepsilon:= \varepsilon_{9.4}(q,k,b,\gamma):=\varepsilon_{8.16}(k,\gamma'), ε := ε 9.4 ( q , k , b , γ ) := ε 8.16 ( k , γ ′ ) , β ≥ β 9.4 ( q , k , b , γ ) : = max { β 8.16 ( q , k , b , γ ′ ) , β 9.3 ( k , β 8.16 ′ ( q , k , b ) , γ ) , ε + q + b + 1 } . \beta\ge\beta_{9.4}(q,k,b,\gamma):=\max\bigl\{\beta_{8.16}(q,k,b,\gamma'),\beta_{9.3}\bigl(k,\beta'_{8.16}(q,k,b),\gamma\bigr),\varepsilon+q+b+1\bigr\}. β ≥ β 9.4 ( q , k , b , γ ) := max { β 8.16 ( q , k , b , γ ′ ) , β 9.3 ( k , β 8.16 ′ ( q , k , b ) , γ ) , ε + q + b + 1 } . Let F F F be a graphic-k k k -flower of G G G and let π \pi π be the petal-star of F F F . Let s s s be a β \beta β -broad k k k -separation that almost-γ \gamma γ -interlaces π \pi π . Then we have one of the following two outcomes:
s s s is ε \varepsilon ε -close to some displayed separation of F F F ;
X : = sep ( s ) X:=\operatorname{sep}(s) X := sep ( s ) is S S S -totally-nested with S S S the set of all narrow proper k k k -separations of G G G , and s s s is a displayed separation of the graphic-X X X -anemone. Moreover, s s s is not half-connected.
Proof. By Lemma 9.3 , F F F has a β 8.16 ′ ( q , k , b ) \beta'_{8.16}(q,k,b) β 8.16 ′ ( q , k , b ) -broad displayed separation. By Lemma 9.2 , there is a k k k -flower F ′ F' F ′ which satisfies the outcome of Lemma 9.2 . Let σ \sigma σ be the star of petal-separations
of F ′ F' F ′ . By Lemma 9.2 (ii), σ \sigma σ is a star of β 8.16 ′ ( q , k , b ) \beta'_{8.16}(q,k,b) β 8.16 ′ ( q , k , b ) -broad separations of G G G . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, all the separations in σ \sigma σ are k k k -separations. By Lemma 9.2 (iv), s s s almost-γ ′ \gamma' γ ′ -interlaces σ \sigma σ . Since β ≥ β 8.16 ( q , k , b , γ ′ ) \beta\ge\beta_{8.16}(q,k,b,\gamma') β ≥ β 8.16 ( q , k , b , γ ′ ) , we can apply Lemma 8.16 to s s s and σ \sigma σ and get one of the following two outcomes:
Case (i): s s s is ε \varepsilon ε -close to some k k k -separation s ′ s' s ′ that almost-interlaces σ \sigma σ . Since s s s is β \beta β -broad with β ≥ ε + q + b + 1 \beta\ge\varepsilon+q+b+1 β ≥ ε + q + b + 1 , we have that s ′ s' s ′ is a proper separation of G G G of breadth at least q + b + 1 q+b+1 q + b + 1 . Since G G G is ( b , k ) (b,k) ( b , k ) -narrow and ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, we get that s ′ s' s ′ is a narrow proper k k k -separation of G G G that almost-interlaces σ \sigma σ . By Lemma 3.8, s ′ s' s ′ is induced by F ′ F' F ′ , and hence by F F F by Lemma 9.2 (i). Since s ′ s' s ′ is a narrow proper k k k -separation, it is displayed by F F F : this follows from the definition of a daisy in the daisy case, while every induced separation of an anemone is displayed. So, we have outcome (i).
Case (ii): s s s is a displayed separation of the graphic-X X X -anemone A A A with X : = sep ( s ) X := \operatorname{sep}(s) X := sep ( s ) , and the narrow proper petal-separations of A A A are narrow-nested. Moreover, s s s is not half-connected. In particular, the narrow proper petal-separations of A A A are S S S -totally-nested. Let S X ⊆ S S_X \subseteq S S X ⊆ S be the set of half-connected separations in S S S with separator X X X . Since s s s is β \beta β -broad with β ≥ q + b + 1 \beta\ge q+b+1 β ≥ q + b + 1 , and since G G G is ( b , k ) (b,k) ( b , k ) -narrow and ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, we get that s s s is a narrow proper k k k -separation of G G G , with separator X X X . Thus, there exists a separation in S S S with separator X X X , so S X ≠ ∅ S_X \ne\varnothing S X = ∅ . Furthermore, every separation in S X S_X S X is a narrow proper petal-separation of A A A or its inverse. Hence, every separation in S X S_X S X is S S S -totally-nested. Thus, X X X is S S S -totally-nested by Lemma 9.1 . Then we have outcome (ii). □
Setting
We now introduce the setting in which we will work throughout this section, and under which assumptions we are essentially able to prove Theorem 2.6 (see Theorem 9.31 and Theorem 9.32 ). Later, in the proof of Theorem 10.4 , we will show how the key lemmas from the previous sections combine to obtain the conditions required for this setting: Key Lemma 5.1 will show that S 0 S_0 S 0 is uniformly γ A \gamma_A γ A -crossing, Key Lemma 6.6 constructs the flowers F A F_A F A , Key Lemma 7.1 essentially shows that the petal-separations of F A F_A F A are narrow-γ N \gamma_N γ N -nested, and Key Lemma 8.1 constructs the nested set N N N of separations.
Setting 9.5. Let q , k , b , β , γ N , ε N , γ A , ε F ∈ N q,k,b,\beta,\gamma_N,\varepsilon_N,\gamma_A,\varepsilon_F \in\mathbb{N} q , k , b , β , γ N , ε N , γ A , ε F ∈ N . Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph that is ( b , k ) (b,k) ( b , k ) -narrow. Let S 0 S_0 S 0 be the set of all β \beta β -broad k k k -separations of G G G . Assume that S 0 S_0 S 0 is uniformly γ A \gamma_A γ A -crossing. Let A t \mathcal{A}_t A t and A \mathcal{A} A consist of all trivial and nontrivial γ A \gamma_A γ A -asterisks of S 0 S_0 S 0 , respectively. Let ( F A : A ∈ A ) (F_A : A \in\mathcal{A}) ( F A : A ∈ A ) be a family of graphic-k k k -flowers. Let S 1 : = ⋃ { ( π ( F A ) ∪ ( π ( F A ) ) ∗ : A ∈ A } ∪ ⋃ A t S_1 := \bigcup\{(\pi(F_A) \cup(\pi(F_A))^* : A \in\mathcal{A}\} \cup\bigcup\mathcal{A}_t S 1 := ⋃ {( π ( F A ) ∪ ( π ( F A ) ) ∗ : A ∈ A } ∪ ⋃ A t . Let N N N be a nested symmetric set of k k k -separations of G G G . Assume that we have the following properties:
(9.5.1) β ≥ max { 2 q + k , q + 1 } \beta\ge\max\{2q+k,q+1\} β ≥ max { 2 q + k , q + 1 } and γ A ≥ β ≥ b \gamma_A \ge\beta\ge b γ A ≥ β ≥ b and ε N ≥ γ N ≥ β + k \varepsilon_N \ge\gamma_N \ge\beta+k ε N ≥ γ N ≥ β + k .
(9.5.2) If the common centre X X X of A ∈ A A \in\mathcal{A} A ∈ A has size k k k , then F A F_A F A is the graphic-X X X -anemone.
(9.5.3) For every A ∈ A A \in\mathcal{A} A ∈ A and s ∈ A s \in A s ∈ A there is d ∈ D ( F A ) d \in D(F_A) d ∈ D ( F A ) such that s s s and d d d are ε F \varepsilon_F ε F -close.
(9.5.4) Every k k k -separation of G G G is γ N \gamma_N γ N -nested with all the petal-separations of all the graphic-k k k -flowers F A F_A F A for A ∈ A A \in\mathcal{A} A ∈ A .
(9.5.5) The separations in N N N are narrow-γ N \gamma_N γ N -nested.
(9.5.6) Every narrow proper k k k -separation of G G G that is nested with every narrow proper k k k -separation of G G G is contained in N N N .
(9.5.7) N N N ε N \varepsilon_N ε N -approximates S 1 S_1 S 1 .
Observation 9.6. Assume Setting 9.5. Then every β \beta β -broad ( ≤ k ) (\le k) ( ≤ k ) -separation s s s of G G G is narrow and has order exactly k k k .
Proof. Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and β > q \beta>q β > q by (9.5.1), the order of s s s must be ≥ k \ge k ≥ k . Since G G G is ( b , k ) (b,k) ( b , k ) -narrow and β ≥ b \beta\ge b β ≥ b by (9.5.1), we have that s s s is narrow. □ \square □
High-spread asterisks interlace unique splitting stars
The main objective in this subsection is to prove Proposition 9.16, which roughly says that every asterisk A A A of large enough spread (that is, the largest integer γ \gamma γ such that two separations in A A A γ \gamma γ -cross) ε \varepsilon ε -interlaces a unique splitting star of N N N . One main step is to prove Lemma 9.9, which essentially says that although a lot of δ ′ \delta' δ ′ -asterisks of D ( F ) D(F) D ( F ) can ‘live’ in a given flower F F F , there is a unique δ \delta δ -asterisk of D ( F ) D(F) D ( F ) (with δ ≈ δ ′ / 2 \delta\approx\delta'/2 δ ≈ δ ′ /2 ) that contains all those δ ′ \delta' δ ′ -asterisks.
Lemma 9.7. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and γ ≥ γ 5.1 ( q , k ) \gamma\ge\gamma_{5.1}(q,k) γ ≥ γ 5.1 ( q , k ) . Let S S S be a set of ( ≤ k ) (\le k) ( ≤ k ) -separations of G G G . Then S ′ : = { s ∈ S : s is γ -crossed by some s ′ ∈ S } S' := \{s\in S:s\text{ is }\gamma\text{-crossed by some }s'\in S\} S ′ := { s ∈ S : s is γ -crossed by some s ′ ∈ S } is uniformly γ \gamma γ -crossing.
Proof. If s ∈ S ′ s\in S' s ∈ S ′ then s s s is γ \gamma γ -crossed by some s ′ ∈ S s'\in S s ′ ∈ S , so s s s is 2 ( γ − k ) 2(\gamma-k) 2 ( γ − k ) -broad by Lemma 5.6. Since γ ≥ γ 5.1 ( q , k ) = 3 q + 3 k \gamma\ge\gamma_{5.1}(q,k)=3q+3k γ ≥ γ 5.1 ( q , k ) = 3 q + 3 k , we have 2 ( γ − k ) > q 2(\gamma-k)>q 2 ( γ − k ) > q , so s s s is ( > q ) (>q) ( > q ) -broad. Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, it follows that s s s is a k k k -separation. Thus, S ′ S' S ′ is uniformly γ \gamma γ -crossing by Key Lemma 5.1. □ \square □
Definition 9.8 (A ( S , γ ) \mathcal{A}(S,\gamma) A ( S , γ ) ). Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and γ ≥ γ 5.1 ( q , k ) \gamma\ge\gamma_{5.1}(q,k) γ ≥ γ 5.1 ( q , k ) . Let S S S be a symmetric set of ( ≤ k ) (\le k) ( ≤ k ) -separations of G G G . We denote by A ( S , γ ) \mathcal{A}(S,\gamma) A ( S , γ ) the set of all γ \gamma γ -asterisks of { s ∈ S : s is γ -crossed by some s ′ ∈ S } \{s\in S:s\text{ is }\gamma\text{-crossed by some }s'\in S\} { s ∈ S : s is γ -crossed by some s ′ ∈ S } , which is well-defined by Lemma 9.7.
Lemma 9.9. Let δ , δ ′ ∈ N \delta,\delta'\in\mathbb{N} δ , δ ′ ∈ N such that δ ≥ γ 5.1 ( q , k ) \delta\ge\gamma_{5.1}(q,k) δ ≥ γ 5.1 ( q , k ) and δ ′ ≥ δ 9.9 ( k , δ ) : = 2 δ + 2 k \delta'\ge\delta_{9.9}(k,\delta):=2\delta+2k δ ′ ≥ δ 9.9 ( k , δ ) := 2 δ + 2 k . Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, and let F F F be a graphic-k k k -flower of G G G such that A ( D ( F ) , δ ′ ) ≠ ∅ \mathcal{A}(D(F),\delta')\ne\varnothing A ( D ( F ) , δ ′ ) = ∅ . Then there is a unique A ∈ A ( D ( F ) , δ ) A\in\mathcal{A}(D(F),\delta) A ∈ A ( D ( F ) , δ ) such that ⋃ A ( D ( F ) , δ ′ ) ⊆ A \bigcup\mathcal{A}(D(F),\delta')\subseteq A ⋃ A ( D ( F ) , δ ′ ) ⊆ A .
Proof. It suffices to show that for every A ≠ A ′ ∈ A ( D ( F ) , δ ′ ) A\ne A'\in\mathcal{A}(D(F),\delta') A = A ′ ∈ A ( D ( F ) , δ ′ ) , there is a displayed separation of F F F that δ \delta δ -crosses at least one separation in each of A A A and A ′ A' A ′ . Let A ≠ A ′ ∈ A ( D ( F ) , δ ′ ) A\ne A'\in\mathcal{A}(D(F),\delta') A = A ′ ∈ A ( D ( F ) , δ ′ ) . Let r 1 , s 1 ∈ A r_1,s_1\in A r 1 , s 1 ∈ A δ ′ \delta' δ ′ -cross and r 2 , s 2 ∈ A ′ r_2,s_2\in A' r 2 , s 2 ∈ A ′ δ ′ \delta' δ ′ -cross. Write r i : = ( A i , B i ) r_i:=(A_i,B_i) r i := ( A i , B i ) and s i : = ( C i , D i ) s_i:=(C_i,D_i) s i := ( C i , D i ) for i = 1 , 2 i=1,2 i = 1 , 2 . Then r i r_i r i and s i s_i s i ( δ ′ − k ) (\delta'-k) ( δ ′ − k ) -criss by Observation 5.4 for both i i i . If r 1 r_1 r 1 and r 2 r_2 r 2 δ \delta δ -cross, then we are done. Hence we may assume that r 1 ≤ δ r 2 r_1\le_\delta r_2 r 1 ≤ δ r 2 , say.
Assume first that F = ( O , V ) F=(O,\mathcal{V}) F = ( O , V ) is a graphic-k k k -daisy. See Figure 9.1 for an illustration of the setting. Since r 1 r_1 r 1 is displayed by F F F , there exist two subpaths P A 1 , P B 1 P_{A_1},P_{B_1} P A 1 , P B 1 of O O O such that V(O)=V(P_{A_1})\mathbin{\mathaccent\cdot\cup}V(P_{B_1}) , and A 1 = V ↾ P A 1 A_1=V\mathbin{\upharpoonright}P_{A_1} A 1 = V ↾ P A 1 and B 1 = V ↾ P B 1 B_1=V\mathbin{\upharpoonright}P_{B_1} B 1 = V ↾ P B 1 . Let P C 1 , P D 1 , P A 2 , P B 2 , P C 2 P_{C_1},P_{D_1},P_{A_2},P_{B_2},P_{C_2} P C 1 , P D 1 , P A 2 , P B 2 , P C 2 and P D 2 P_{D_2} P D 2 be defined similarly. Let e 1 , f 1 e_1,f_1 e 1 , f 1 be the only edges of O O O that are not edges of P C 1 ∪ P D 1 P_{C_1}\cup P_{D_1} P C 1 ∪ P D 1 . Since ( A 1 , B 1 ) (A_1,B_1) ( A 1 , B 1 ) and ( C 1 , D 1 ) (C_1,D_1) ( C 1 , D 1 ) cross, exactly one of them, say e 1 e_1 e 1 , is an edge of P A 1 P_{A_1} P A 1 . Similarly, let e 2 , f 2 e_2,f_2 e 2 , f 2 be the only edges of O O O that are not edges of P C 2 ∪ P D 2 P_{C_2}\cup P_{D_2} P C 2 ∪ P D 2 . Since ( A 2 , B 2 ) (A_2,B_2) ( A 2 , B 2 ) and ( C 2 , D 2 ) (C_2,D_2) ( C 2 , D 2 ) cross, exactly one of them, say f 2 f_2 f 2 , is an edge of P B 2 P_{B_2} P B 2 .
Figure 9.1. The situation in the proof of Lemma 9.9. Only one half of each s i s_i s i is drawn. Here, s s s is the turquoise separation.
We first show that f 2 ∈ E ( P B 1 ) f_2\in E(P_{B_1}) f 2 ∈ E ( P B 1 ) . Since r 1 ≤ δ r 2 r_1\le_\delta r_2 r 1 ≤ δ r 2 , we have ∣ A 1 ∩ B 2 ∣ ≤ ∣ A 1 ∖ A 2 ∣ + ∣ A 2 ∩ B 2 ∣ ≤ δ + k |A_1\cap B_2|\le|A_1\setminus A_2|+|A_2\cap B_2|\le\delta+k ∣ A 1 ∩ B 2 ∣ ≤ ∣ A 1 ∖ A 2 ∣ + ∣ A 2 ∩ B 2 ∣ ≤ δ + k . Let R 1 , R 2 R_1,R_2 R 1 , R 2 be the two components of P B 2 − { f 2 } P_{B_2}-\{f_2\} P B 2 − { f 2 } . Each V ↾ R j V\mathbin{\upharpoonright}R_j V ↾ R j contains one of the proper r 2 s 2 r_2s_2 r 2 s 2 -corners, so ∣ V ↾ R j ∣ > δ ′ − k |V\mathbin{\upharpoonright}R_j|>\delta'-k ∣ V ↾ R j ∣ > δ ′ − k by Observation 5.4 (ii). Suppose for a contradiction that f 2 ∉ E ( P B 1 ) f_2\notin E(P_{B_1}) f 2 ∈ / E ( P B 1 ) , so f 2 f_2 f 2 lies in P A 1 P_{A_1} P A 1 or is one of its boundary edges. If one of the endpoints of P B 2 P_{B_2} P B 2 lies in P A 1 P_{A_1} P A 1 then the corresponding R j R_j R j is contained in P A 1 P_{A_1} P A 1 , otherwise P A 1 P_{A_1} P A 1 is contained in P B 2 P_{B_2} P B 2 . In the former case, we have ∣ A 1 ∩ B 2 ∣ ≥ ∣ V ↾ R j ∣ > δ ′ − k ≥ δ + k |A_1 \cap B_2| \ge|V \restriction R_j| > \delta' - k \ge\delta+ k ∣ A 1 ∩ B 2 ∣ ≥ ∣ V ↾ R j ∣ > δ ′ − k ≥ δ + k . In the latter case, we have ∣ A 1 ∩ B 2 ∣ ≥ ∣ A 1 ∣ ≥ δ ′ > δ + k |A_1 \cap B_2| \ge|A_1| \ge\delta' > \delta+ k ∣ A 1 ∩ B 2 ∣ ≥ ∣ A 1 ∣ ≥ δ ′ > δ + k . In both cases we obtain a contradiction so f 2 ∈ E ( P B 1 ) f_2 \in E(P_{B_1}) f 2 ∈ E ( P B 1 ) . Symmetrically, we have e 1 ∈ P A 2 e_1 \in P_{A_2} e 1 ∈ P A 2 . In particular, e 1 ≠ f 2 e_1 \ne f_2 e 1 = f 2 .
Then, O − { e 1 , f 2 } O - \{e_1, f_2\} O − { e 1 , f 2 } has exactly two connected components, which are paths, call them Q 1 Q_1 Q 1 and Q 2 Q_2 Q 2 . Let s : = ( U , W ) : = ( V ↾ Q 1 , V ↾ Q 2 ) s := (U, W) := (V \restriction Q_1, V \restriction Q_2) s := ( U , W ) := ( V ↾ Q 1 , V ↾ Q 2 ) , and observe that s ∈ D ( F ) s \in D(\mathcal{F}) s ∈ D ( F ) .
Claim 9.9.1. s s s δ \delta δ -crosses both r 1 r_1 r 1 and r 2 r_2 r 2 .
Proof of the claim. Since f 2 ∈ E ( P B 1 ) f_2 \in E(P_{B_1}) f 2 ∈ E ( P B 1 ) and e 1 ∈ E ( P A 1 ) e_1 \in E(P_{A_1}) e 1 ∈ E ( P A 1 ) , the cut { e 1 , f 2 } \{e_1, f_2\} { e 1 , f 2 } splits P A 1 P_{A_1} P A 1 at e 1 e_1 e 1 , just as the cut defining s 1 s_1 s 1 does. Thus, one of the proper A 1 C 1 A_1C_1 A 1 C 1 -corner and the proper A 1 D 1 A_1D_1 A 1 D 1 -corner is contained in the proper A 1 U A_1U A 1 U -corner and the other is contained in the proper A 1 W A_1W A 1 W -corner. Thus, the proper A 1 U A_1U A 1 U -corner and the proper A 1 W A_1W A 1 W -corner both have size greater than δ ′ − k \delta' - k δ ′ − k by Observation 5.4 (ii).
Similarly, since e 1 ∈ P A 2 e_1 \in P_{A_2} e 1 ∈ P A 2 and f 2 ∈ P B 2 f_2 \in P_{B_2} f 2 ∈ P B 2 , one of the proper B 2 C 2 B_2C_2 B 2 C 2 -corner and the proper B 2 D 2 B_2D_2 B 2 D 2 -corner is contained in the proper B 2 U B_2U B 2 U -corner and the other is contained in the proper B 2 W B_2W B 2 W -corner. Thus, the proper B 2 U B_2U B 2 U -corner and the proper B 2 W B_2W B 2 W -corner both have size greater than δ ′ − k \delta' - k δ ′ − k .
Since r 1 ≤ δ r 2 r_1 \le_\delta r_2 r 1 ≤ δ r 2 , it follows that
∣ ( B 1 ∖ A 1 ) ∩ ( U ∖ W ) ∣ ≥ ∣ ( B 2 ∖ A 2 ) ∩ ( U ∖ W ) ∣ − ∣ A 1 ∖ A 2 ∣ > δ ′ − k − δ ≥ δ |(B_1 \setminus A_1) \cap(U \setminus W)| \ge|(B_2 \setminus A_2) \cap(U \setminus W)| - |A_1 \setminus A_2| > \delta' - k - \delta\ge\delta ∣ ( B 1 ∖ A 1 ) ∩ ( U ∖ W ) ∣ ≥ ∣ ( B 2 ∖ A 2 ) ∩ ( U ∖ W ) ∣ − ∣ A 1 ∖ A 2 ∣ > δ ′ − k − δ ≥ δ
by the above. Similarly, we also have ∣ ( B 1 ∖ A 1 ) ∩ ( W ∖ U ) ∣ > δ |(B_1 \setminus A_1) \cap(W \setminus U)| > \delta ∣ ( B 1 ∖ A 1 ) ∩ ( W ∖ U ) ∣ > δ . Thus, s s s and r 1 r_1 r 1 δ \delta δ -criss so they δ \delta δ -cross by Observation 5.4 (iv). Symmetrically, s s s and r 2 r_2 r 2 δ \delta δ -cross. ◇
Assume now that F = ( S , V ) F=(S,\mathcal{V}) F = ( S , V ) is a graphic-k k k -anemone with centre c c c . Note that ∣ ( A 1 ∩ B 2 ) ∖ ( A 2 ∪ B 1 ) ∣ ≤ ∣ A 1 ∖ A 2 ∣ ≤ δ \lvert(A_1\cap B_2)\setminus(A_2\cup B_1)\rvert\leq\lvert A_1\setminus A_2\rvert\leq\delta ∣( A 1 ∩ B 2 ) ∖ ( A 2 ∪ B 1 )∣ ≤ ∣ A 1 ∖ A 2 ∣ ≤ δ since r 1 ≤ δ r 2 r_1\leq_{\delta}r_2 r 1 ≤ δ r 2 . Let X = V c X=V_c X = V c , so the separator of every proper separation induced by F F F is equal to X X X . Let P : = { p ∈ V ( S ) : V p ⊆ ( A 1 ∩ C 1 ) ∪ ( B 2 ∩ D 2 ) } P:=\{p\in V(S):V_p\subseteq(A_1\cap C_1)\cup(B_2\cap D_2)\} P := { p ∈ V ( S ) : V p ⊆ ( A 1 ∩ C 1 ) ∪ ( B 2 ∩ D 2 )} and let s = : ( U , W ) : = ( ⋃ p ∈ P V p , ⋃ p ∉ P V p ) s=:(U,W):=(\bigcup_{p\in P}V_p,\bigcup_{p\notin P}V_p) s =: ( U , W ) := ( ⋃ p ∈ P V p , ⋃ p ∈ / P V p ) . Note that s s s is displayed by F F F .
Claim 9.9.2. s s s δ \delta δ -crosses both r 1 r_1 r 1 and r 2 r_2 r 2 .
Proof of the claim. By Observation 5.4 and since δ ≤ δ ′ − δ − 2 k \delta\leq\delta'-\delta-2k δ ≤ δ ′ − δ − 2 k , it suffices to prove that s s s ( δ ′ − δ − 2 k ) (\delta'-\delta-2k) ( δ ′ − δ − 2 k ) -crisses both r 1 r_1 r 1 and r 2 r_2 r 2 . We prove that s s s ( δ ′ − δ − 2 k ) (\delta'-\delta-2k) ( δ ′ − δ − 2 k ) -crisses r 1 r_1 r 1 ; the proof for r 2 r_2 r 2 is symmetric.
The separations r 1 r_1 r 1 and s 1 s_1 s 1 δ ′ \delta' δ ′ -cross, and both have their separator equal to X X X so ∣ ( A 1 ∩ C 1 ) ∖ X ∣ > δ ′ − k \lvert(A_1\cap C_1)\setminus X\rvert>\delta'-k ∣( A 1 ∩ C 1 ) ∖ X ∣ > δ ′ − k by Observation 5.4 (v). Let v ∈ ( A 1 ∩ C 1 ) ∖ X v\in(A_1\cap C_1)\setminus X v ∈ ( A 1 ∩ C 1 ) ∖ X . There exists a unique p ∈ V ( S ) p\in V(S) p ∈ V ( S ) such that v ∈ V p v\in V_p v ∈ V p . Since r 1 r_1 r 1 and s 1 s_1 s 1 are displayed by F F F , we have V p ⊆ A 1 ∩ C 1 V_p\subseteq A_1\cap C_1 V p ⊆ A 1 ∩ C 1 , so p ∈ P p\in P p ∈ P , and therefore v ∈ U v\in U v ∈ U . Since s s s is displayed by F F F , its separator is contained in X X X . Therefore, the A 1 U A_1U A 1 U -corner contains ( A 1 ∩ C 1 ) ∖ X (A_1\cap C_1)\setminus X ( A 1 ∩ C 1 ) ∖ X , so has size greater than δ ′ − k \delta'-k δ ′ − k .
As above, we have ∣ ( B 2 ∩ D 2 ) ∖ X ∣ > δ ′ − k \lvert(B_2\cap D_2)\setminus X\rvert>\delta'-k ∣( B 2 ∩ D 2 ) ∖ X ∣ > δ ′ − k . Let v ∈ ( B 2 ∩ D 2 ) ∖ X v\in(B_2\cap D_2)\setminus X v ∈ ( B 2 ∩ D 2 ) ∖ X . As before, we have v ∈ U v\in U v ∈ U . Furthermore, ∣ ( B 2 ∖ X ) ∖ B 1 ∣ ≤ ∣ ( A 1 ∩ B 2 ) ∖ X ∣ ≤ δ \lvert(B_2\setminus X)\setminus B_1\rvert\leq\lvert(A_1\cap B_2)\setminus X\rvert\leq\delta ∣( B 2 ∖ X ) ∖ B 1 ∣ ≤ ∣( A 1 ∩ B 2 ) ∖ X ∣ ≤ δ . Therefore, ∣ ( B 1 ∩ U ) ∖ X ∣ ≥ ∣ ( B 2 ∩ U ) ∖ X ∣ − ∣ ( B 2 ∖ X ) ∖ B 1 ∣ ≥ ∣ ( B 2 ∩ D 2 ) ∖ X ∣ − δ > δ ′ − δ − k \lvert(B_1\cap U)\setminus X\rvert\geq\lvert(B_2\cap U)\setminus X\rvert-\lvert(B_2\setminus X)\setminus B_1\rvert\geq\lvert(B_2\cap D_2)\setminus X\rvert-\delta>\delta'-\delta-k ∣( B 1 ∩ U ) ∖ X ∣ ≥ ∣( B 2 ∩ U ) ∖ X ∣ − ∣( B 2 ∖ X ) ∖ B 1 ∣ ≥ ∣( B 2 ∩ D 2 ) ∖ X ∣ − δ > δ ′ − δ − k . Therefore, the B 1 U B_1U B 1 U -corner has size greater than δ ′ − δ − k \delta'-\delta-k δ ′ − δ − k .
As above, we have ∣ ( A 1 ∩ D 1 ) ∖ X ∣ > δ ′ − k \lvert(A_1\cap D_1)\setminus X\rvert>\delta'-k ∣( A 1 ∩ D 1 ) ∖ X ∣ > δ ′ − k . Let v ∈ ( A 1 ∩ D 1 ) ∖ X v\in(A_1\cap D_1)\setminus X v ∈ ( A 1 ∩ D 1 ) ∖ X . Suppose further that v ∉ B 2 v\notin B_2 v ∈ / B 2 . There exists a unique p ∈ V ( S ) p\in V(S) p ∈ V ( S ) such that v ∈ V p v\in V_p v ∈ V p . Since v ∉ A 1 ∩ C 1 v\notin A_1\cap C_1 v ∈ / A 1 ∩ C 1 and v ∉ B 2 ∩ D 2 v\notin B_2\cap D_2 v ∈ / B 2 ∩ D 2 , we have p ∉ P p\notin P p ∈ / P , so v ∉ U v\notin U v ∈ / U so v ∈ W ∖ U v\in W\setminus U v ∈ W ∖ U . Therefore, the A 1 W A_1W A 1 W -corner has size at least ∣ ( A 1 ∩ D 1 ) ∖ ( X ∪ B 2 ) ∣ ≥ ∣ ( A 1 ∩ D 1 ) ∖ X ∣ − ∣ ( A 1 ∩ B 2 ) ∖ X ∣ > δ ′ − k − δ \lvert(A_1\cap D_1)\setminus(X\cup B_2)\rvert\geq\lvert(A_1\cap D_1)\setminus X\rvert-\lvert(A_1\cap B_2)\setminus X\rvert>\delta'-k-\delta ∣( A 1 ∩ D 1 ) ∖ ( X ∪ B 2 )∣ ≥ ∣( A 1 ∩ D 1 ) ∖ X ∣ − ∣( A 1 ∩ B 2 ) ∖ X ∣ > δ ′ − k − δ .
As above, we have ∣ ( B 2 ∩ C 2 ) ∖ X ∣ > δ ′ − k \lvert(B_2\cap C_2)\setminus X\rvert>\delta'-k ∣( B 2 ∩ C 2 ) ∖ X ∣ > δ ′ − k . Let v ∈ ( B 2 ∩ C 2 ) ∖ X v\in(B_2\cap C_2)\setminus X v ∈ ( B 2 ∩ C 2 ) ∖ X . Suppose further that v ∉ A 1 v\notin A_1 v ∈ / A 1 . There exists a unique p ∈ V ( S ) p\in V(S) p ∈ V ( S ) such that v ∈ V p v\in V_p v ∈ V p . Since v ∉ A 1 ∩ C 1 v\notin A_1\cap C_1 v ∈ / A 1 ∩ C 1 and v ∉ B 2 ∩ D 2 v\notin B_2\cap D_2 v ∈ / B 2 ∩ D 2 , we have p ∉ P p\notin P p ∈ / P , so v ∉ U v\notin U v ∈ / U so v ∈ W ∖ U v\in W\setminus U v ∈ W ∖ U . Therefore, ∣ ( B 1 ∩ W ) ∖ X ∣ ≥ ∣ ( B 1 ∩ B 2 ∩ C 2 ) ∖ X ∣ = ∣ ( B 2 ∩ C 2 ) ∖ X ∣ − ∣ ( B 2 ∩ C 2 ) ∖ ( B 1 ∪ X ) ∣ > δ ′ − k − ∣ B 2 ∖ B 1 ∣ ≥ δ ′ − k − δ \lvert(B_1\cap W)\setminus X\rvert\geq\lvert(B_1\cap B_2\cap C_2)\setminus X\rvert=\lvert(B_2\cap C_2)\setminus X\rvert-\lvert(B_2\cap C_2)\setminus(B_1\cup X)\rvert>\delta'-k-\lvert B_2\setminus B_1\rvert\geq\delta'-k-\delta ∣( B 1 ∩ W ) ∖ X ∣ ≥ ∣( B 1 ∩ B 2 ∩ C 2 ) ∖ X ∣ = ∣( B 2 ∩ C 2 ) ∖ X ∣ − ∣( B 2 ∩ C 2 ) ∖ ( B 1 ∪ X )∣ > δ ′ − k − ∣ B 2 ∖ B 1 ∣ ≥ δ ′ − k − δ .
Definition 9.10 (Spread). Let S S S be a set of separations of G G G . The spread of S S S is the maximum number γ ∈ N \gamma\in\mathbb{N} γ ∈ N for which some two separations in S S S γ \gamma γ -cross, and we set it to − ∞ -\infty − ∞ if no two separations in S S S cross. The spread of a graphic-k k k -flower is the spread of the set of its displayed separations.
Definition 9.11 (Relevant). Assume Setting 9.5. Let δ , δ ′ ∈ N \delta,\delta'\in\mathbb{N} δ , δ ′ ∈ N such that δ ≥ γ 5.1 ( q , k ) \delta\geq\gamma_{5.1}(q,k) δ ≥ γ 5.1 ( q , k ) and δ ′ ≥ δ 9.9 ′ ( k , δ ) \delta'\geq\delta'_{9.9}(k,\delta) δ ′ ≥ δ 9.9 ′ ( k , δ ) . We say that a nontrivial γ A \gamma_A γ A -asterisk A ∈ A A\in\mathcal{A} A ∈ A is δ ′ \delta' δ ′ -relevant if A ( D ( F A ) , δ ′ ) ≠ ∅ A(D(F_A),\delta')\neq\varnothing A ( D ( F A ) , δ ′ ) = ∅ , that is, if D ( F A ) D(F_A) D ( F A ) has spread ≥ δ ′ \geq\delta' ≥ δ ′ . In this event, we denote the unique nontrivial δ \delta δ -asterisk provided by Lemma 9.9 by A δ A^\delta A δ , where δ ′ \delta' δ ′ will be clear from context. Note that A δ ⊆ D ( F A ) A^\delta\subseteq D(F_A) A δ ⊆ D ( F A ) .
Lemma 9.12. Assume Setting 9.5. Let δ , δ ′ ∈ N \delta,\delta'\in\mathbb{N} δ , δ ′ ∈ N such that
δ > δ 9.12 ( q , k , γ A , ε F ) : = max { 2 k + 3 γ A + 3 ε F , γ 5.1 ( q , k ) } and δ ′ ≥ δ 9.9 ′ ( k , δ ) . \delta>\delta_{9.12}(q,k,\gamma_A,\varepsilon_F):=\max\{2k+3\gamma_A+3\varepsilon_F,\gamma_{5.1}(q,k)\}\qquad\text{and}\qquad\delta'\geq\delta'_{9.9}(k,\delta). δ > δ 9.12 ( q , k , γ A , ε F ) := max { 2 k + 3 γ A + 3 ε F , γ 5.1 ( q , k )} and δ ′ ≥ δ 9.9 ′ ( k , δ ) . Let A ∈ A A\in\mathcal{A} A ∈ A have spread ≥ ζ 9.12 ( δ ′ , ε F ) : = δ ′ + 2 ε F \geq\zeta_{9.12}(\delta',\varepsilon_F):=\delta'+2\varepsilon_F ≥ ζ 9.12 ( δ ′ , ε F ) := δ ′ + 2 ε F . Then:
A A A is δ ′ \delta' δ ′ -relevant and A δ A^\delta A δ exists;
:::`
(ii) whenever d 1 , d 2 ∈ D ( F A ) d_1,d_2 \in D(F_A) d 1 , d 2 ∈ D ( F A ) δ \delta δ -cross, we have d 1 , d 2 ∈ A d_1,d_2 \in A d 1 , d 2 ∈ A ;
(iii) A δ ⊆ A A^\delta\subseteq A A δ ⊆ A .
Proof. (i). Let s 1 ′ , s 2 ′ ∈ A s'_1,s'_2 \in A s 1 ′ , s 2 ′ ∈ A ( δ ′ + 2 ε F ) (\delta' + 2\varepsilon_F) ( δ ′ + 2 ε F ) -cross. Let s 1 , s 2 ∈ D ( F A ) s_1,s_2 \in D(F_A) s 1 , s 2 ∈ D ( F A ) be ε F \varepsilon_F ε F -close to s 1 ′ , s 2 ′ s'_1,s'_2 s 1 ′ , s 2 ′ , respectively, which we find with (9.5.3). Then s 1 s_1 s 1 and s 2 s_2 s 2 δ ′ \delta' δ ′ -cross by Lemma 7.2. Note that δ ≥ γ 5.1 ( q , k ) \delta\ge\gamma_{5.1}(q,k) δ ≥ γ 5.1 ( q , k ) . Hence A A A is δ ′ \delta' δ ′ -relevant, and A δ A^\delta A δ exists by Lemma 9.9.
(ii). Let d 1 , d 2 ∈ D ( F A ) d_1,d_2 \in D(F_A) d 1 , d 2 ∈ D ( F A ) δ \delta δ -cross. By symmetry, it suffices to show d 1 ∈ A d_1 \in A d 1 ∈ A . Since d 1 d_1 d 1 and d 2 d_2 d 2 δ \delta δ -cross, d 1 d_1 d 1 is ( δ − k ) (\delta-k) ( δ − k ) -broad by Lemma 5.6, with δ − k ≥ γ A ≥ β \delta-k \ge\gamma_A \ge\beta δ − k ≥ γ A ≥ β by (9.5.1). Furthermore, β > q \beta> q β > q by (9.5.1) and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable so d 1 d_1 d 1 is a β \beta β -broad k k k -separation of G G G so d 1 ∈ S 0 d_1 \in S_0 d 1 ∈ S 0 . If d 1 d_1 d 1 ( γ A + ε F ) (\gamma_A+\varepsilon_F) ( γ A + ε F ) -crosses s 1 s_1 s 1 or s 2 s_2 s 2 , then d 1 d_1 d 1 γ A \gamma_A γ A -crosses s 1 ′ s'_1 s 1 ′ or s 2 ′ s'_2 s 2 ′ by Lemma 5.8, so d 1 ∈ A d_1 \in A d 1 ∈ A . So we may assume that d 1 d_1 d 1 is ( γ A + ε F ) (\gamma_A+\varepsilon_F) ( γ A + ε F ) -nested with both s 1 s_1 s 1 and s 2 s_2 s 2 . Then s 1 , s 2 ≤ γ A + ε F d 1 s_1,s_2 \le_{\gamma_A+\varepsilon_F} d_1 s 1 , s 2 ≤ γ A + ε F d 1 , say. Write s i = ( A i , B i ) s_i=(A_i,B_i) s i = ( A i , B i ) and d i = ( C i , D i ) d_i=(C_i,D_i) d i = ( C i , D i ) for i ∈ [ 2 ] i \in[2] i ∈ [ 2 ] .
Claim 9.12.1. There is a displayed separation s = ( U , W ) s=(U,W) s = ( U , W ) of F A F_A F A with the following properties (up to renaming C 2 C_2 C 2 and D 2 D_2 D 2 ):
( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ⊆ U (A_1 \cap B_2) \cup(D_1 \cap C_2) \subseteq U ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ⊆ U ,
( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 ) ⊆ γ A + ε F W (A_1 \cap A_2) \cup(D_1 \cap D_2) \subseteq_{\gamma_A+\varepsilon_F} W ( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 ) ⊆ γ A + ε F W .
Proof of the claim. Suppose first that F A : = ( S , V ) F_A := (S,\mathcal{V}) F A := ( S , V ) is a graphic-k k k -anemone with centre c c c . Let X : = V c X := V_c X := V c , so the separator of every proper separation induced by F A F_A F A is equal to X X X . Let
P : = { p ∈ V ( S ) ∖ { c } ∣ V p ⊆ ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) } P := \{p \in V(S) \setminus\{c\} \mid V_p \subseteq(A_1 \cap B_2) \cup(D_1 \cap C_2)\} P := { p ∈ V ( S ) ∖ { c } ∣ V p ⊆ ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 )}
and let s : = ( ⋃ p ∈ P V p , ⋃ p ∉ P V p ) s := (\bigcup_{p\in P} V_p,\bigcup_{p\notin P} V_p) s := ( ⋃ p ∈ P V p , ⋃ p ∈ / P V p ) . Since s 1 s_1 s 1 and s 2 s_2 s 2 δ ′ \delta' δ ′ -cross, by Observation 5.4 (v) we have ∣ A 1 ∩ B 2 ∣ > δ ′ ≥ δ 9.9 ′ ( k , δ ) ≥ k |A_1 \cap B_2| > \delta' \ge\delta_{9.9}'(k,\delta) \ge k ∣ A 1 ∩ B 2 ∣ > δ ′ ≥ δ 9.9 ′ ( k , δ ) ≥ k , so there exists v ∈ ( A 1 ∩ B 2 ) ∖ X v \in(A_1 \cap B_2) \setminus X v ∈ ( A 1 ∩ B 2 ) ∖ X . Let p ∈ V ( S ) ∖ { c } p \in V(S) \setminus\{c\} p ∈ V ( S ) ∖ { c } be the unique node such that v ∈ V p v \in V_p v ∈ V p . Since s 1 s_1 s 1 and s 2 s_2 s 2 are displayed by F A F_A F A , we have V p ⊆ A 1 ∩ B 2 V_p \subseteq A_1 \cap B_2 V p ⊆ A 1 ∩ B 2 , so p ∈ P p \in P p ∈ P and P ≠ ∅ P \ne\varnothing P = ∅ . Thus, s s s is a proper displayed separation of F A F_A F A , so X = sep ( s ) X=\operatorname{sep}(s) X = sep ( s ) . Write s : = ( U , W ) s := (U,W) s := ( U , W ) . As above, if v ∈ ( ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ) ∖ X v \in((A_1 \cap B_2) \cup(D_1 \cap C_2)) \setminus X v ∈ (( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 )) ∖ X , the unique p ∈ V ( S ) p \in V(S) p ∈ V ( S ) such that v ∈ V p v \in V_p v ∈ V p satisfies p ∈ P p \in P p ∈ P , so v ∈ U v \in U v ∈ U . Thus, ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ⊆ U (A_1 \cap B_2) \cup(D_1 \cap C_2) \subseteq U ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ⊆ U . Then, let
N : = ∣ ( ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ) ∩ ( ( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 ) ) ∖ X ∣ . N := |((A_1 \cap B_2) \cup(D_1 \cap C_2)) \cap((A_1 \cap A_2) \cup(D_1 \cap D_2)) \setminus X|. N := ∣ (( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 )) ∩ (( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 )) ∖ X ∣.
Using that s 1 ≤ γ A + ε F d 1 s_1 \le_{\gamma_A+\varepsilon_F} d_1 s 1 ≤ γ A + ε F d 1 , we have
N = ∣ ( ( A 1 ∩ A 2 ∩ B 2 ) ∪ ( A 1 ∩ B 2 ∩ D 1 ∩ D 2 ) ∪ ( D 1 ∩ C 2 ∩ A 1 ∩ A 2 ) ∪ ( D 1 ∩ C 2 ∩ D 2 ) ) ∖ X ∣ ≤ ∣ ( A 2 ∩ B 2 ) ∖ X ∣ + ∣ ( A 1 ∩ D 1 ) ∖ X ∣ + ∣ ( C 2 ∩ D 2 ) ∖ X ∣ ≤ ∣ ( A 1 ∩ B 1 ) ∖ X ∣ + ∣ ( D 1 ∖ B 1 ) ∖ X ∣ ≤ γ A + ε F . \begin{aligned}
N = |((A_1 \cap A_2 \cap B_2) \cup(A_1 \cap B_2 \cap D_1 \cap D_2) \cup(D_1 \cap C_2 \cap A_1 \cap A_2) \cup(D_1 \cap C_2 \cap D_2)) \setminus X| \\
&\le|(A_2 \cap B_2) \setminus X| + |(A_1 \cap D_1) \setminus X| + |(C_2 \cap D_2) \setminus X| \\
&\le|(A_1 \cap B_1) \setminus X| + |(D_1 \setminus B_1) \setminus X| \\
&\le\gamma_A+\varepsilon_F.
\end{aligned} N = ∣ (( A 1 ∩ A 2 ∩ B 2 ) ∪ ( A 1 ∩ B 2 ∩ D 1 ∩ D 2 ) ∪ ( D 1 ∩ C 2 ∩ A 1 ∩ A 2 ) ∪ ( D 1 ∩ C 2 ∩ D 2 )) ∖ X ∣ ≤ ∣ ( A 2 ∩ B 2 ) ∖ X ∣ + ∣ ( A 1 ∩ D 1 ) ∖ X ∣ + ∣ ( C 2 ∩ D 2 ) ∖ X ∣ ≤ ∣ ( A 1 ∩ B 1 ) ∖ X ∣ + ∣ ( D 1 ∖ B 1 ) ∖ X ∣ ≤ γ A + ε F .
Thus, to prove that ( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 ) ⊆ γ A + ε F W (A_1 \cap A_2) \cup(D_1 \cap D_2) \subseteq_{\gamma_A+\varepsilon_F} W ( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 ) ⊆ γ A + ε F W , it suffices to prove that ( ( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 ) ) ∖ ( ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ) ⊆ W ((A_1 \cap A_2) \cup(D_1 \cap D_2)) \setminus((A_1 \cap B_2) \cup(D_1 \cap C_2)) \subseteq W (( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 )) ∖ (( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 )) ⊆ W , where we use X = sep ( s ) ⊆ W X=\operatorname{sep}(s) \subseteq W X = sep ( s ) ⊆ W . Let v ∈ ( ( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 ) ) ∖ ( ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) ) v \in((A_1 \cap A_2) \cup(D_1 \cap D_2)) \setminus((A_1 \cap B_2) \cup(D_1 \cap C_2)) v ∈ (( A 1 ∩ A 2 ) ∪ ( D 1 ∩ D 2 )) ∖ (( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 )) and let p ∈ V ( S ) ∖ { c } p \in V(S) \setminus\{c\} p ∈ V ( S ) ∖ { c } be the unique node such that v ∈ V p v \in V_p v ∈ V p . Since v ∉ ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) v \notin(A_1 \cap B_2) \cup(D_1 \cap C_2) v ∈ / ( A 1 ∩ B 2 ) ∪ ( D 1 ∩ C 2 ) , we have p ∉ P p \notin P p ∈ / P so v ∈ W v \in W v ∈ W .
Suppose now that F A : = ( O , V ) F_A := (O,\mathcal{V}) F A := ( O , V ) is a graphic-k k k -daisy. Given two edges e , e ′ ∈ E ( O ) e,e' \in E(O) e , e ′ ∈ E ( O ) , O − { e , e ′ } O-\{e,e'\} O − { e , e ′ } has exactly two connected components, which are paths, call them Q 1 Q_1 Q 1 and Q 2 Q_2 Q 2 . We say that the separations ( V ∖ Q 1 , V ∖ Q 2 ) (V \setminus Q_1,V \setminus Q_2) ( V ∖ Q 1 , V ∖ Q 2 ) and ( V ∖ Q 2 , V ∖ Q 1 ) (V \setminus Q_2,V \setminus Q_1) ( V ∖ Q 2 , V ∖ Q 1 ) are displayed by the edge-cut { e , e ′ } \{e,e'\} { e , e ′ } . There exist edges e 1 , e 2 , e 1 ′ , e 2 ′ , f 1 , f 2 , f 1 ′ , f 2 ′ e_1,e_2,e'_1,e'_2,f_1,f_2,f'_1,f'_2 e 1 , e 2 , e 1 ′ , e 2 ′ , f 1 , f 2 , f 1 ′ , f 2 ′ of O O O such that each s i s_i s i is displayed by the edge-cut { e i , e i ′ } \{e_i,e'_i\} { e i , e i ′ } and each d i d_i d i is displayed by the edge-cut { f i , f i ′ } \{f_i,f'_i\} { f i , f i ′ } . We arbitrarily fix one of the two orientations of the cycle O O O . For i ∈ { 1 , 2 } i \in\{1,2\} i ∈ { 1 , 2 } , up to exchanging A i A_i A i and B i B_i B i , we may assume that A i A_i A i is the union of the bags of the nodes on the arc from e i ′ e'_i e i ′ to e i e_i e i in this orientation, and B i B_i B i is the union of the bags of the nodes on the arc from e i e_i e i to e i ′ e'_i e i ′ in this orientation, see Figure 9.2 . Since s 1 s_1 s 1 and s 2 s_2 s 2 cross, we may assume that the edges e 2 , e 1 , e 2 ′ , e 1 ′ e_2,e_1,e'_2,e'_1 e 2 , e 1 , e 2 ′ , e 1 ′ appear in this order along O O O (following this orientation), as in Figure 9.2 . Up to renaming f 1 f_1 f 1 and f 1 ′ f'_1 f 1 ′ , we may assume that e 2 , f 1 , f 1 ′ e_2,f_1,f'_1 e 2 , f 1 , f 1 ′ appear in this order along O O O . Since d 1 d_1 d 1 and d 2 d_2 d 2 cross, up to renaming f 2 f_2 f 2 and f 2 ′ f'_2 f 2 ′ , we may assume that e 2 , f 1 , f 2 , f 2 ′ e_2,f_1,f_2,f'_2 e 2 , f 1 , f 2 , f 2 ′ appear in this order along O O O , again see Figure 9.2 . Let Q 1 , Q 2 Q_1,Q_2 Q 1 , Q 2 be the two connected components of O − { e 2 , f 2 } O-\{e_2,f_2\} O − { e 2 , f 2 } . By the above, up to renaming Q 1 Q_1 Q 1 and Q 2 Q_2 Q 2 , we may assume that f 1 ∈ E ( Q 1 ) f_1 \in E(Q_1) f 1 ∈ E ( Q 1 ) and f 1 ′ ∈ E ( Q 2 ) f'_1 \in E(Q_2) f 1 ′ ∈ E ( Q 2 ) . Let s : = ( V ↾ Q 1 , V ↾ Q 2 ) s := (V \mathbin{\upharpoonright} Q_1,V \mathbin{\upharpoonright} Q_2) s := ( V ↾ Q 1 , V ↾ Q 2 ) . Then, s s s is a displayed separation of F A F_A F A . Write s = : ( U , W ) s =: (U,W) s =: ( U , W ) . Since f 1 ∈ E ( Q 1 ) f_1 \in E(Q_1) f 1 ∈ E ( Q 1 ) and f 1 ′ ∈ E ( Q 2 ) f'_1 \in E(Q_2) f 1 ′ ∈ E ( Q 2 ) , we have D 1 ∩ C 2 ⊆ U D_1 \cap C_2 \subseteq U D 1 ∩ C 2 ⊆ U and D 1 ∩ D 2 ⊆ W D_1 \cap D_2 \subseteq W D 1 ∩ D 2 ⊆ W (up to renaming C 2 C_2 C 2 and D 2 D_2 D 2 ), see Figure 9.2 . We now argue that e 1 , f 2 , e 2 ′ e_1,f_2,e'_2 e 1 , f 2 , e 2 ′ appear in this order along O O O . If not, e 1 , e 2 ′ , f 2 e_1,e'_2,f_2 e 1 , e 2 ′ , f 2 appear in this order along O O O . Thus, either e 1 , e 2 ′ , f 2 , e 2 e_1,e'_2,f_2,e_2 e 1 , e 2 ′ , f 2 , e 2 appear in this order along O O O or e 1 , e 2 ′ , e 2 , f 2 e_1,e'_2,e_2,f_2 e 1 , e 2 ′ , e 2 , f 2 appear in this order along O O O . By symmetry, suppose that e 1 , e 2 ′ , f 2 , e 2 e_1,e'_2,f_2,e_2 e 1 , e 2 ′ , f 2 , e 2 appear in this order along O O O . Then, e 2 ′ , f 2 , f 1 ′ , e 2 e'_2,f_2,f'_1,e_2 e 2 ′ , f 2 , f 1 ′ , e 2 appear in this order along O O O . Combined with the fact that e 2 , f 1 , f 1 ′ e_2,f_1,f'_1 e 2 , f 1 , f 1 ′ appear in this order along O O O , this implies that the D 1 D 2 D_1D_2 D 1 D 2 -corner is contained in the D 1 A 2 D_1A_2 D 1 A 2 -corner (see Figure 9.2 ). However, d 1 d_1 d 1 and d 2 d_2 d 2 δ \delta δ -cross so they ( δ − k ) (\delta-k) ( δ − k ) -criss by Observation 5.4 (ii). Thus, the D 1 D 2 D_1D_2 D 1 D 2 -corner has size > δ − k ≥ γ A + ε F > \delta-k \geq\gamma_A+\varepsilon_F > δ − k ≥ γ A + ε F , so the D 1 A 2 D_1A_2 D 1 A 2 -corner also has size > γ A + ε F > \gamma_A+\varepsilon_F > γ A + ε F , so ∣ D 1 ∖ B 2 ∣ > γ A + ε F |D_1 \setminus B_2| > \gamma_A+\varepsilon_F ∣ D 1 ∖ B 2 ∣ > γ A + ε F , which contradicts s 2 ≤ γ A + ε F d 1 s_2 \leq\gamma_A+\varepsilon_F d_1 s 2 ≤ γ A + ε F d 1 . Thus, e 1 , f 2 , e 2 ′ e_1,f_2,e'_2 e 1 , f 2 , e 2 ′ indeed appear in this order along O O O . This immediately yields A 1 ∩ B 2 ⊆ U A_1 \cap B_2 \subseteq U A 1 ∩ B 2 ⊆ U and A 1 ∩ A 2 ⊆ W A_1 \cap A_2 \subseteq W A 1 ∩ A 2 ⊆ W . ◇
Figure 9.2. The situation in the proof of Lemma 9.12 . Here, the cycle O O O is oriented clockwise, s s s is the turquoise separation, and δ ′ ≫ δ ≫ γ A , k \delta' \gg\delta\gg\gamma_A, k δ ′ ≫ δ ≫ γ A , k .
Let s s s be the separation obtained from Claim 9.12.1 .
Claim 9.12.2. All extended corners of the separations s s s and d 1 d_1 d 1 , and of the separations s s s and s 1 s_1 s 1 have size greater than ( δ − 2 γ A − 2 ε F ) (\delta-2\gamma_A-2\varepsilon_F) ( δ − 2 γ A − 2 ε F ) .
Proof of the claim. We encourage the reader to check this statement by looking at Figure 9.2 , as the proof simply consists of computations.
Using Observation 5.4 (v) and that s s s was obtained from Claim 9.12.1 throughout, we have
∣ A 1 ∩ U ∣ ≥ ∣ A 1 ∩ B 2 ∣ > δ ′ |A_1 \cap U| \geq|A_1 \cap B_2| > \delta' ∣ A 1 ∩ U ∣ ≥ ∣ A 1 ∩ B 2 ∣ > δ ′
since s 1 s_1 s 1 and s 2 s_2 s 2 δ ′ \delta' δ ′ -cross. We have
∣ A 1 ∩ W ∣ ≥ ∣ A 1 ∩ A 2 ∩ W ∣ > δ ′ − ( γ A + ε F ) |A_1 \cap W| \geq|A_1 \cap A_2 \cap W| > \delta'-(\gamma_A+\varepsilon_F) ∣ A 1 ∩ W ∣ ≥ ∣ A 1 ∩ A 2 ∩ W ∣ > δ ′ − ( γ A + ε F )
since s 1 s_1 s 1 and s 2 s_2 s 2 δ ′ \delta' δ ′ -cross. We have
∣ B 1 ∩ W ∣ ≥ ∣ B 1 ∩ D 1 ∩ D 2 ∩ W ∣ ≥ ∣ D 1 ∩ D 2 ∩ W ∣ − ∣ D 1 ∖ B 1 ∣ > δ − ( γ A + ε F ) − ( γ A + ε F ) |B_1 \cap W| \geq|B_1 \cap D_1 \cap D_2 \cap W| \geq|D_1 \cap D_2 \cap W|-|D_1 \setminus B_1| > \delta-(\gamma_A+\varepsilon_F)-(\gamma_A+\varepsilon_F) ∣ B 1 ∩ W ∣ ≥ ∣ B 1 ∩ D 1 ∩ D 2 ∩ W ∣ ≥ ∣ D 1 ∩ D 2 ∩ W ∣ − ∣ D 1 ∖ B 1 ∣ > δ − ( γ A + ε F ) − ( γ A + ε F )
since d 1 d_1 d 1 and d 2 d_2 d 2 δ \delta δ -cross and s 1 ≤ γ A + ε F d 1 s_1 \leq\gamma_A+\varepsilon_F d_1 s 1 ≤ γ A + ε F d 1 . We have
∣ B 1 ∩ U ∣ ≥ ∣ B 1 ∩ D 1 ∩ C 2 ∣ ≥ ∣ D 1 ∩ C 2 ∣ − ∣ D 1 ∖ B 1 ∣ > δ − ( γ A + ε F ) |B_1 \cap U| \geq|B_1 \cap D_1 \cap C_2| \geq|D_1 \cap C_2|-|D_1 \setminus B_1| > \delta-(\gamma_A+\varepsilon_F) ∣ B 1 ∩ U ∣ ≥ ∣ B 1 ∩ D 1 ∩ C 2 ∣ ≥ ∣ D 1 ∩ C 2 ∣ − ∣ D 1 ∖ B 1 ∣ > δ − ( γ A + ε F )
since d 1 d_1 d 1 and d 2 d_2 d 2 δ \delta δ -cross and s 1 ≤ γ A + ε F d 1 s_1 \leq\gamma_A+\varepsilon_F d_1 s 1 ≤ γ A + ε F d 1 . We have
∣ C 1 ∩ U ∣ ≥ ∣ A 1 ∩ B 2 ∩ C 1 ∣ ≥ ∣ A 1 ∩ B 2 ∣ − ∣ A 1 ∖ C 1 ∣ > δ ′ − ( γ A + ε F ) |C_1 \cap U| \geq|A_1 \cap B_2 \cap C_1| \geq|A_1 \cap B_2|-|A_1 \setminus C_1| > \delta'-(\gamma_A+\varepsilon_F) ∣ C 1 ∩ U ∣ ≥ ∣ A 1 ∩ B 2 ∩ C 1 ∣ ≥ ∣ A 1 ∩ B 2 ∣ − ∣ A 1 ∖ C 1 ∣ > δ ′ − ( γ A + ε F )
since s 1 s_1 s 1 and s 2 s_2 s 2 δ ′ \delta' δ ′ -cross and s 1 ≤ γ A + ε F d 1 s_1 \le_{\gamma_A+\varepsilon_F} d_1 s 1 ≤ γ A + ε F d 1 . We have
∣ C 1 ∩ W ∣ ≥ ∣ A 1 ∩ A 2 ∩ C 1 ∩ W ∣ ≥ ∣ A 1 ∩ A 2 ∩ W ∣ − ∣ A 1 ∖ C 1 ∣ > δ ′ − ( γ A + ε F ) − ( γ A + ε F ) |C_1 \cap W| \ge|A_1 \cap A_2 \cap C_1 \cap W| \ge|A_1 \cap A_2 \cap W| - |A_1 \setminus C_1| > \delta' - (\gamma_A + \varepsilon_F) - (\gamma_A + \varepsilon_F) ∣ C 1 ∩ W ∣ ≥ ∣ A 1 ∩ A 2 ∩ C 1 ∩ W ∣ ≥ ∣ A 1 ∩ A 2 ∩ W ∣ − ∣ A 1 ∖ C 1 ∣ > δ ′ − ( γ A + ε F ) − ( γ A + ε F )
since s 1 s_1 s 1 and s 2 s_2 s 2 δ ′ \delta' δ ′ -cross and s 1 ≤ γ A + ε F d 1 s_1 \le_{\gamma_A+\varepsilon_F} d_1 s 1 ≤ γ A + ε F d 1 . We have
∣ D 1 ∩ W ∣ ≥ ∣ D 1 ∩ D 2 ∩ W ∣ > δ − ( γ A + ε F ) |D_1 \cap W| \ge|D_1 \cap D_2 \cap W| > \delta- (\gamma_A + \varepsilon_F) ∣ D 1 ∩ W ∣ ≥ ∣ D 1 ∩ D 2 ∩ W ∣ > δ − ( γ A + ε F )
since d 1 d_1 d 1 and d 2 d_2 d 2 δ \delta δ -cross. We have
∣ D 1 ∩ U ∣ ≥ ∣ D 1 ∩ C 2 ∣ > δ |D_1 \cap U| \ge|D_1 \cap C_2| > \delta ∣ D 1 ∩ U ∣ ≥ ∣ D 1 ∩ C 2 ∣ > δ
since d 1 d_1 d 1 and d 2 d_2 d 2 δ \delta δ -cross. ◇
Thus, the separations s s s and d 1 d_1 d 1 ( δ − 2 k − 2 γ A − 2 ε F ) (\delta- 2k - 2\gamma_A - 2\varepsilon_F) ( δ − 2 k − 2 γ A − 2 ε F ) -criss, so s s s has breadth at least 2 ( δ − 2 k − 2 γ A − 2 ε F ) ≥ γ A ≥ β 2(\delta- 2k - 2\gamma_A - 2\varepsilon_F) \ge\gamma_A \ge\beta 2 ( δ − 2 k − 2 γ A − 2 ε F ) ≥ γ A ≥ β , using that δ ≥ 2 k + 3 γ A + 3 ε F \delta\ge2k + 3\gamma_A + 3\varepsilon_F δ ≥ 2 k + 3 γ A + 3 ε F and (9.5.1). Furthermore, β > q \beta> q β > q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable so s ∈ S 0 s \in S_0 s ∈ S 0 . Since s 1 ′ s'_1 s 1 ′ is ε F \varepsilon_F ε F -close to s 1 s_1 s 1 , we have that s 1 ′ s'_1 s 1 ′ ( δ − 2 k − 2 γ A − 3 ε F ) (\delta- 2k - 2\gamma_A - 3\varepsilon_F) ( δ − 2 k − 2 γ A − 3 ε F ) -crosses s s s by Lemma 5.8 . Now s ∈ S 0 s \in S_0 s ∈ S 0 and s 1 ′ ∈ A s'_1 \in A s 1 ′ ∈ A with
δ − 2 k − 2 γ A − 3 ε F ≥ γ A \delta- 2k - 2\gamma_A - 3\varepsilon_F \ge\gamma_A δ − 2 k − 2 γ A − 3 ε F ≥ γ A
yields s ∈ A s \in A s ∈ A . Hence d 1 ∈ A d_1 \in A d 1 ∈ A as well.
(iii) follows from (i) and (ii) . □
Corollary 9.13. Assume Setting 9.5. Let δ > δ 9.12 ( q , k , γ A , ε F ) \delta> \delta_{9.12}(q,k,\gamma_A,\varepsilon_F) δ > δ 9.12 ( q , k , γ A , ε F ) and δ ′ ≥ δ 9.9 ′ ( k , δ ) \delta' \ge\delta'_{9.9}(k,\delta) δ ′ ≥ δ 9.9 ′ ( k , δ ) . Let A 1 ≠ A 2 ∈ A A_1 \ne A_2 \in\mathcal{A} A 1 = A 2 ∈ A have spread ≥ ζ 9.12 ( δ ′ , ε F ) \ge\zeta_{9.12}(\delta',\varepsilon_F) ≥ ζ 9.12 ( δ ′ , ε F ) . Then D ( F A 1 ) ≠ D ( F A 2 ) D(F_{A_1}) \ne D(F_{A_2}) D ( F A 1 ) = D ( F A 2 ) . In particular, F A 1 F_{A_1} F A 1 and F A 2 F_{A_2} F A 2 are not id G \operatorname{id}_G id G -isomorphic, and F A 1 ≠ F A 2 F_{A_1} \ne F_{A_2} F A 1 = F A 2 .
Proof. Assume for a contradiction that D ( F A 1 ) = D ( F A 2 ) D(F_{A_1}) = D(F_{A_2}) D ( F A 1 ) = D ( F A 2 ) . By Lemma 9.12 , both A i A_i A i are δ ′ \delta' δ ′ -relevant, their A i δ A_i^\delta A i δ exists, and A i δ ⊆ A i A_i^\delta\subseteq A_i A i δ ⊆ A i . Since D ( F A 1 ) = D ( F A 2 ) D(F_{A_1}) = D(F_{A_2}) D ( F A 1 ) = D ( F A 2 ) , we find that A 1 δ = A 2 δ A_1^\delta= A_2^\delta A 1 δ = A 2 δ by definition. Hence A 1 A_1 A 1 and A 2 A_2 A 2 intersect, contradicting that A 1 A_1 A 1 and A 2 A_2 A 2 are disjoint by definition. □
Lemma 9.14. Assume Setting 9.5. Let δ ∈ N \delta\in\mathbb{N} δ ∈ N such that δ ≥ γ N \delta\ge\gamma_N δ ≥ γ N . Let s , s ′ ∈ S 0 s,s' \in S_0 s , s ′ ∈ S 0 δ \delta δ -cross. Then neither s s s nor s ′ s' s ′ is ( δ − γ N ) (\delta-\gamma_N) ( δ − γ N ) -close to any t ∈ N t \in N t ∈ N .
Proof. By symmetry, it suffices to focus on s s s . Assume for a contradiction that s s s is ( δ − γ N ) (\delta-\gamma_N) ( δ − γ N ) -close to some t ∈ N t \in N t ∈ N . Then s ′ s' s ′ γ N \gamma_N γ N -crosses t t t by Lemma 5.8 . Recall that all separations in S 0 S_0 S 0 are narrow by Observation 9.6. This contradicts the assumption (9.5.5) that every separation in N N N is narrow-γ N \gamma_N γ N -nested. □
Lemma 9.15. Assume Setting 9.5. Let δ ∈ N \delta\in\mathbb{N} δ ∈ N such that δ ≥ δ 9.15 ( γ N , ε N ) : = γ N + 2 ε N \delta\ge\delta_{9.15}(\gamma_N,\varepsilon_N) := \gamma_N + 2\varepsilon_N δ ≥ δ 9.15 ( γ N , ε N ) := γ N + 2 ε N . Let A ⊆ S 0 A \subseteq S_0 A ⊆ S 0 be nontrivially δ \delta δ -asteriskal. For every s ∈ A s \in A s ∈ A let
O ( s ) : = { t ∈ N : t ≤ ε N s or t ≤ ε N s ∗ } (5) O(s) := \{t \in N : t \le_{\varepsilon_N} s \text{ or } t \le_{\varepsilon_N} s^*\} \tag*{(5)} O ( s ) := { t ∈ N : t ≤ ε N s or t ≤ ε N s ∗ } (5) and let σ ( s ) \sigma(s) σ ( s ) consist of the ≤ \le ≤ -maximal elements of O ( s ) O(s) O ( s ) . Then O ( s ) O(s) O ( s ) is a consistent orientation of N N N and σ ( s ) \sigma(s) σ ( s ) is the unique splitting star of N N N such that s s s ε N \varepsilon_N ε N -interlaces σ ( s ) \sigma(s) σ ( s ) . Moreover, O ( s ) = O ( s ′ ) O(s) = O(s') O ( s ) = O ( s ′ ) and σ ( s ) = σ ( s ′ ) \sigma(s) = \sigma(s') σ ( s ) = σ ( s ′ ) for every two s , s ′ ∈ A s,s' \in A s , s ′ ∈ A .
Proof. Since A A A is nontrivially δ \delta δ -asteriskal and s ∈ A s \in A s ∈ A , s s s δ \delta δ -crosses some separation of A A A . By Lemma 5.6, s s s is 2 ( δ − k ) 2(\delta-k) 2 ( δ − k ) -broad. Using that γ N ≥ β + k ≥ k + 1 \gamma_N \ge\beta+k \ge k+1 γ N ≥ β + k ≥ k + 1 by (9.5.1), it follows that s s s has breadth at least 2 ( γ N + 2 ε N − k ) ≥ 2 ε N + 1 2(\gamma_N+2\varepsilon_N-k) \ge2\varepsilon_N+1 2 ( γ N + 2 ε N − k ) ≥ 2 ε N + 1 .
Since δ ≥ γ N \delta\ge\gamma_N δ ≥ γ N and δ − γ N ≥ ε N \delta-\gamma_N \ge\varepsilon_N δ − γ N ≥ ε N by assumption, no t ∈ N t \in N t ∈ N is ε N \varepsilon_N ε N -close to s s s by Lemma 9.14. We have s ∈ A ⊆ S 0 s \in A \subseteq S_0 s ∈ A ⊆ S 0 so s s s is β \beta β -broad, hence b b b -broad by (9.5.1). Since G G G is ( b , k ) (b,k) ( b , k ) -narrow, this implies that s s s is narrow. By (9.5.1) and (9.5.5), s s s is ε N \varepsilon_N ε N -nested with every separation in N N N . By Lemma 8.6 , O ( s ) O(s) O ( s ) is a consistent orientation of N N N , and σ ( s ) \sigma(s) σ ( s ) is a splitting star of N N N . Note that s s s ε N \varepsilon_N ε N -interlaces σ ( s ) \sigma(s) σ ( s ) .
Since O ( s 1 ) = O ( s 2 ) O(s_1) = O(s_2) O ( s 1 ) = O ( s 2 ) implies σ ( s 1 ) = σ ( s 2 ) \sigma(s_1) = \sigma(s_2) σ ( s 1 ) = σ ( s 2 ) , it remains to show the former for every two s 1 , s 2 ∈ A s_1,s_2 \in A s 1 , s 2 ∈ A . For this, let s 1 ≠ s 2 ∈ A s_1 \ne s_2 \in A s 1 = s 2 ∈ A . Since A A A is δ \delta δ -crossing-connected, we may assume without loss of generality that s 1 s_1 s 1 and s 2 s_2 s 2 δ \delta δ -cross. Now assume for a contradiction that O ( s 1 ) ≠ O ( s 2 ) O(s_1) \ne O(s_2) O ( s 1 ) = O ( s 2 ) . Then we can assume without loss of generality that there exists t ∈ O ( s 2 ) ∖ O ( s 1 ) t \in O(s_2) \setminus O(s_1) t ∈ O ( s 2 ) ∖ O ( s 1 ) . Then, either s 1 ≤ ε N t s_1 \le_{\varepsilon_N} t s 1 ≤ ε N t or ( s 1 ) ∗ ≤ ε N t (s_1)^* \le_{\varepsilon_N} t ( s 1 ) ∗ ≤ ε N t , and either t ≤ ε N s 2 t \le_{\varepsilon_N} s_2 t ≤ ε N s 2 or t ≤ ε N ( s 2 ) ∗ t \le_{\varepsilon_N} (s_2)^* t ≤ ε N ( s 2 ) ∗ . In all cases, since δ ≥ 2 ε N \delta\ge2\varepsilon_N δ ≥ 2 ε N and with Lemma 4.9 , this contradicts our assumption that s 1 s_1 s 1 and s 2 s_2 s 2 δ \delta δ -cross. □
Proposition 9.16. Assume Setting 9.5. Let
δ > δ 9.16 ( q , k , γ A , ε F , γ N , ε N ) : = max { δ 9.12 ( q , k , γ A , ε F ) , δ 9.15 ( γ N , ε N ) } \delta> \delta_{9.16}(q,k,\gamma_A,\varepsilon_F,\gamma_N,\varepsilon_N) := \max\{\delta_{9.12}(q,k,\gamma_A,\varepsilon_F),\delta_{9.15}(\gamma_N,\varepsilon_N)\} δ > δ 9.16 ( q , k , γ A , ε F , γ N , ε N ) := max { δ 9.12 ( q , k , γ A , ε F ) , δ 9.15 ( γ N , ε N )} and δ ′ ≥ δ 9.9 ′ ( k , δ ) \delta' \ge\delta'_{9.9}(k,\delta) δ ′ ≥ δ 9.9 ′ ( k , δ ) . Let A ∈ A A \in\mathcal{A} A ∈ A have spread ≥ ζ 9.12 ( δ ′ , ε F ) \ge\zeta_{9.12}(\delta',\varepsilon_F) ≥ ζ 9.12 ( δ ′ , ε F ) . Then A A A is δ ′ \delta' δ ′ -relevant, A δ A^\delta A δ exists and A δ A^\delta A δ ε N \varepsilon_N ε N -interlaces a unique splitting star of N N N .
Proof. We combine Lemma 9.12 and Lemma 9.15.
Every splitting star is interlaced by at most one high-spread asterisk
Here, we prove Proposition 9.20: if A 1 , A 2 A_1,A_2 A 1 , A 2 are two distinct asterisks of large enough spread, then A 1 δ A_1^\delta A 1 δ and A 2 δ A_2^\delta A 2 δ interlace distinct splitting stars. For this, we first reduce to the case where A 1 δ A_1^\delta A 1 δ γ N \gamma_N γ N -interlaces the petal-star of F A 2 F_{A_2} F A 2 (see Lemma 9.17 ), which we then show is impossible. Later, we will combine Proposition 9.20 with Proposition 9.16 to obtain the flowers F t F_t F t in Theorem 2.6 .
Lemma 9.17. Assume Setting 9.5. Let
δ > δ 9.17 ( q , k , γ A , ε F , γ N , ε N ) : = max { δ 9.16 ( q , k , γ A , ε F , γ N , ε N ) , k + 2 γ N + ε N + 1 } \delta> \delta_{9.17}(q,k,\gamma_A,\varepsilon_F,\gamma_N,\varepsilon_N) := \max\{\delta_{9.16}(q,k,\gamma_A,\varepsilon_F,\gamma_N,\varepsilon_N), k+2\gamma_N+\varepsilon_N+1\} δ > δ 9.17 ( q , k , γ A , ε F , γ N , ε N ) := max { δ 9.16 ( q , k , γ A , ε F , γ N , ε N ) , k + 2 γ N + ε N + 1 } and δ ′ ≥ δ 9.9 ′ ( k , δ ) \delta' \ge\delta'_{9.9}(k,\delta) δ ′ ≥ δ 9.9 ′ ( k , δ ) . Let A 1 ≠ A 2 ∈ A A_1 \ne A_2 \in\mathcal{A} A 1 = A 2 ∈ A have spread ≥ ζ 9.12 ( δ ′ , ε F ) \ge\zeta_{9.12}(\delta',\varepsilon_F) ≥ ζ 9.12 ( δ ′ , ε F ) . So A 1 δ A_1^\delta A 1 δ and A 2 δ A_2^\delta A 2 δ exist by Proposition 9.16 and ε N \varepsilon_N ε N -interlace unique splitting stars σ 1 \sigma_1 σ 1 and σ 2 \sigma_2 σ 2 of N N N , respectively. Then either A 1 δ A_1^\delta A 1 δ γ N \gamma_N γ N -interlaces π ( F A 2 ) \pi(F_{A_2}) π ( F A 2 ) or σ 1 ≠ σ 2 \sigma_1 \ne\sigma_2 σ 1 = σ 2 .
Proof. Let π 2 \pi_2 π 2 denote the petal-star π ( F A 2 ) \pi(F_{A_2}) π ( F A 2 ) . Assume that A 1 δ A_1^\delta A 1 δ does not γ N \gamma_N γ N -interlace π 2 \pi_2 π 2 . Let r 1 ∈ A 1 δ r_1 \in A_1^\delta r 1 ∈ A 1 δ witness this. Let r 2 ∈ A 2 δ r_2 \in A_2^\delta r 2 ∈ A 2 δ . Recall that A i δ ⊆ D ( F A i ) A_i^\delta\subseteq D(F_{A_i}) A i δ ⊆ D ( F A i ) , so r i r_i r i is a displayed separation of F A i F_{A_i} F A i .
Since every separation in π 2 \pi_2 π 2 is γ N \gamma_N γ N -nested with r 1 r_1 r 1 by our assumption (9.5.4), there is p ∈ π 2 p \in\pi_2 p ∈ π 2 such that r 1 ≤ γ N p r_1 \leq_{\gamma_N} p r 1 ≤ γ N p , say. We also have p ≤ r 2 p \leq r_2 p ≤ r 2 , say, as p p p is a petal-separation of F A 2 F_{A_2} F A 2 while r 2 r_2 r 2 is a displayed separation of F A 2 F_{A_2} F A 2 . Hence r 1 ≤ γ N p ≤ r 2 r_1 \leq_{\gamma_N} p \leq r_2 r 1 ≤ γ N p ≤ r 2 . Since both r i r_i r i are 2 ( δ − k ) 2(\delta-k) 2 ( δ − k ) -broad by Lemma 5.6, and since 2 ( δ − k ) ≥ γ N + ε N 2(\delta-k) \geq\gamma_N+\varepsilon_N 2 ( δ − k ) ≥ γ N + ε N , we have that p p p is ε N \varepsilon_N ε N -broad. So p ∈ S 1 p \in S_1 p ∈ S 1 is ε N \varepsilon_N ε N -close to some t ∈ N t \in N t ∈ N by (9.5.7). Now r 1 ≤ γ N + ε N t ≤ ε N r 2 r_1 \leq_{\gamma_N+\varepsilon_N} t \leq_{\varepsilon_N} r_2 r 1 ≤ γ N + ε N t ≤ ε N r 2 by Lemma 4.9 .
Claim 9.17.1. We have r 1 < ε N t r_1 <_{\varepsilon_N} t r 1 < ε N t or r 1 ∗ < ε N t r_1^* <_{\varepsilon_N} t r 1 ∗ < ε N t .
Proof of the claim. Assume for a contradiction that neither holds. Since A 1 δ A_1^\delta A 1 δ contains r 1 r_1 r 1 and ε N \varepsilon_N ε N -interlaces σ 1 \sigma_1 σ 1 , and since σ 1 \sigma_1 σ 1 is a splitting star of N N N and t ∈ N t \in N t ∈ N , Lemma 8.12 implies that either t < ε N r 1 t <_{\varepsilon_N} r_1 t < ε N r 1 or t < ε N r 1 ∗ t <_{\varepsilon_N} r_1^* t < ε N r 1 ∗ .
Case t < ε N r 1 t <_{\varepsilon_N} r_1 t < ε N r 1 . Then r 1 r_1 r 1 is ( γ N + ε N ) (\gamma_N+\varepsilon_N) ( γ N + ε N ) -close to t t t . Since δ − γ N ≥ γ N + ε N \delta-\gamma_N \geq\gamma_N+\varepsilon_N δ − γ N ≥ γ N + ε N , this contradicts Lemma 9.14.
Case t < ε N r 1 ∗ t <_{\varepsilon_N} r_1^* t < ε N r 1 ∗ . Then r 1 ≤ 2 ε N + γ N r 1 ∗ r_1 \leq_{2\varepsilon_N+\gamma_N} r_1^* r 1 ≤ 2 ε N + γ N r 1 ∗ by Lemma 4.9 . Write ( U , W ) : = r 1 (U,W):=r_1 ( U , W ) := r 1 . Then ∣ U ∖ W ∣ ≤ 2 ε N + γ N |U\setminus W| \leq2\varepsilon_N+\gamma_N ∣ U ∖ W ∣ ≤ 2 ε N + γ N , so r 1 r_1 r 1 has breadth ≤ 2 ε N + γ N \leq2\varepsilon_N+\gamma_N ≤ 2 ε N + γ N . But since r 1 r_1 r 1 is δ \delta δ -crossed, r 1 r_1 r 1 has breadth ≥ 2 ( δ − k ) \geq2(\delta-k) ≥ 2 ( δ − k ) by Lemma 5.6. This yields a contradiction as 2 ( δ − k ) > 2 ε N + γ N 2(\delta-k)>2\varepsilon_N+\gamma_N 2 ( δ − k ) > 2 ε N + γ N . ◇
:::
Claim 9.17.2. We have t < ε N r 2 t <_{\varepsilon_N} r_2 t < ε N r 2 .
Proof of the claim. As we already have t ≤ ε N r 2 t \leq_{\varepsilon_N} r_2 t ≤ ε N r 2 , it remains to show r 2 ̸ ≤ ε N t r_2 \not\leq_{\varepsilon_N} t r 2 ≤ ε N t . We cannot have r 2 ≤ ε N t r_2 \leq_{\varepsilon_N} t r 2 ≤ ε N t , as then r 2 r_2 r 2 would be ε N \varepsilon_N ε N -close to t t t , contradicting Lemma 9.14 as δ − γ N ≥ ε N \delta-\gamma_N \geq\varepsilon_N δ − γ N ≥ ε N . ◇
By the two claims, O ( r 1 ) O(r_1) O ( r 1 ) and O ( r 2 ) O(r_2) O ( r 2 ) (as in Lemma 9.15) contain different elements of { t , t ∗ } \{t,t^*\} { t , t ∗ } . Hence σ 1 ≠ σ 2 \sigma_1 \ne\sigma_2 σ 1 = σ 2 . □
Let τ , σ \tau,\sigma τ , σ be two stars of separations of the same graph G G G . We let τ ≼ σ \tau\preccurlyeq\sigma τ ≼ σ if for every t ∈ τ t \in\tau t ∈ τ there is s ∈ σ s \in\sigma s ∈ σ such that t ≤ s t \leq s t ≤ s , as in [43 ] .
Lemma 9.18. Assume Setting 9.5. Let δ > δ 9.17 \delta>\delta_{9.17} δ > δ 9.17 and δ ′ ≥ δ 9.9 ′ ( k , δ ) \delta' \geq\delta'_{9.9}(k,\delta) δ ′ ≥ δ 9.9 ′ ( k , δ ) . Let A 1 ≠ A 2 ∈ A A_1 \ne A_2 \in\mathcal{A} A 1 = A 2 ∈ A both have spread ≥ ζ 9.12 ( δ ′ , ε F ) \geq\zeta_{9.12}(\delta',\varepsilon_F) ≥ ζ 9.12 ( δ ′ , ε F ) . So A 1 δ A_1^\delta A 1 δ and A 2 δ A_2^\delta A 2 δ exist by Proposition 9.16 and ε N \varepsilon_N ε N -interlace unique splitting
stars σ 1 \sigma_1 σ 1 and σ 2 \sigma_2 σ 2 of N N N , respectively. Let S S S be the set of all narrow proper k k k -separations of G G G . Assume that the common centre of A 1 A_1 A 1 is an S S S -totally-nested k k k -separator. Then σ 1 ≠ σ 2 \sigma_1 \ne\sigma_2 σ 1 = σ 2 .
Proof. Let F i : = F A i F_i := F_{A_i} F i := F A i for i = 1 , 2 i = 1, 2 i = 1 , 2 . Both A i A_i A i are δ ′ \delta' δ ′ -relevant by Proposition 9.16. Hence there are δ ′ \delta' δ ′ -crossing displayed separations r i , r i ′ ∈ D ( F i ) r_i, r'_i \in D(F_i) r i , r i ′ ∈ D ( F i ) . By Lemma 9.9, we have r i , r i ′ ∈ A i δ r_i, r'_i \in A_i^{\delta} r i , r i ′ ∈ A i δ . Moreover, by Lemma 5.6, r i r_i r i and r i ′ r'_i r i ′ have breadth ≥ 2 ( δ ′ − k ) ≥ 2 δ > 2 ε N \ge2(\delta' - k) \ge2\delta> 2\varepsilon_N ≥ 2 ( δ ′ − k ) ≥ 2 δ > 2 ε N . Let X X X denote the common centre of A 1 A_1 A 1 . By assumption, X X X is an S S S -totally-nested k k k -separator of G G G . Every separation in S 0 S_0 S 0 is a β \beta β -broad k k k -separation of G G G , so S 0 ⊆ S S_0 \subseteq S S 0 ⊆ S by Observation 9.6. By definition of X X X , there is a separation in S 0 ⊆ S S_0 \subseteq S S 0 ⊆ S with separator equal to X X X , so there is a half-connected separation in S S S with separator equal to X X X . Thus, every half-connected separation in S S S with separator equal to X X X is S S S -totally-nested by Lemma 9.1. Let F F F be the graphic-X X X -anemone, and let π X : = π ( F ) ∖ { ( X , V ( G ) ) } \pi_X := \pi(F) \setminus\{(X, V(G))\} π X := π ( F ) ∖ {( X , V ( G ))} be the star of proper petal-separations of F F F . Since there is a narrow separation with separator X X X , the neighbourhood of each component of G − X G - X G − X is equal to X X X . Thus, every separation in π X \pi_X π X is a narrow proper k k k -separation with separator X X X , so is S S S -totally-nested. Then π X ⊆ N \pi_X \subseteq N π X ⊆ N by (9.5.6). As both r 1 , r 1 ′ r_1, r'_1 r 1 , r 1 ′ have X X X as their separator, both r 1 , r 1 ′ r_1, r'_1 r 1 , r 1 ′ interlace π X \pi_X π X .
Assume for a contradiction that σ 1 = σ 2 \sigma_1 = \sigma_2 σ 1 = σ 2 . Then both r 2 , r 2 ′ r_2, r'_2 r 2 , r 2 ′ ε N \varepsilon_N ε N -interlace σ 1 \sigma_1 σ 1 . Both r 1 , r 1 ′ r_1, r'_1 r 1 , r 1 ′ ε N \varepsilon_N ε N -interlace σ 1 \sigma_1 σ 1 , and both interlace π X \pi_X π X . By the latter, we have π X ⊆ O ( r 1 ) = O ( r 1 ′ ) \pi_X \subseteq O(r_1) = O(r'_1) π X ⊆ O ( r 1 ) = O ( r 1 ′ ) (as in (5) ) by Lemma 9.15, which yields π X ≼ σ 1 \pi_X \preccurlyeq\sigma_1 π X ≼ σ 1 . Therefore, also r 2 , r 2 ′ r_2, r'_2 r 2 , r 2 ′ ε N \varepsilon_N ε N -interlace π X \pi_X π X . Let p ∈ π X p \in\pi_X p ∈ π X be arbitrary. Since r 2 r_2 r 2 ε N \varepsilon_N ε N -interlaces π X \pi_X π X , we have p < ε N r 2 p <_{\varepsilon_N} r_2 p < ε N r 2 or p < ε N ( r 2 ) ∗ p <_{\varepsilon_N} (r_2)^* p < ε N ( r 2 ) ∗ , say the former. Then, r 2 ̸ ε N p r_2 \not\mathrel{\varepsilon_N} p r 2 ε N p . Moreover, if ( r 2 ) ∗ ≤ ε N p (r_2)^* \le_{\varepsilon_N} p ( r 2 ) ∗ ≤ ε N p then we would obtain ( r 2 ) ∗ ≤ 2 ε N r 2 (r_2)^* \le_{2\varepsilon_N} r_2 ( r 2 ) ∗ ≤ 2 ε N r 2 by Lemma 4.9, which by Lemma 8.3 would contradict that r 2 r_2 r 2 has breadth > 2 ε N > 2\varepsilon_N > 2 ε N . Therefore, we have r 2 , r 2 ∗ ̸ ε N p r_2, r_2^* \not\mathrel{\varepsilon_N} p r 2 , r 2 ∗ ε N p , and similarly for r 2 ′ r'_2 r 2 ′ . Hence, since the separations in π X \pi_X π X are S S S -totally-nested, we even get that r 2 , r 2 ′ r_2, r'_2 r 2 , r 2 ′ interlace π X \pi_X π X (without ‘ε N \varepsilon_N ε N ’). Hence r 2 , r 2 ′ ∈ D ( F ) r_2, r'_2 \in D(F) r 2 , r 2 ′ ∈ D ( F ) . By (9.5.2), we have F = F 1 F = F_1 F = F 1 . Thus r 2 , r 2 ′ ∈ D ( F 1 ) r_2, r'_2 \in D(F_1) r 2 , r 2 ′ ∈ D ( F 1 ) . As r 2 , r 2 ′ r_2, r'_2 r 2 , r 2 ′ δ ′ \delta' δ ′ -cross, we have r 2 , r 2 ′ ∈ A 1 δ r_2, r'_2 \in A_1^\delta r 2 , r 2 ′ ∈ A 1 δ by Lemma 9.9. Since A 1 δ ⊆ A 1 A_1^\delta\subseteq A_1 A 1 δ ⊆ A 1 by Lemma 9.12, we have r 2 , r 2 ′ ∈ A 1 r_2, r'_2 \in A_1 r 2 , r 2 ′ ∈ A 1 , contradicting that A 1 ≠ A 2 A_1 \ne A_2 A 1 = A 2 . □
Lemma 9.19. Let s s s be a k k k -separation of G G G , and let t t t be a narrow proper k k k -separation of G G G that crosses s s s . Assume that t t t does not separate vertices of sep ( s ) \operatorname{sep}(s) sep ( s ) . Then sep ( t ) = sep ( s ) \operatorname{sep}(t) = \operatorname{sep}(s) sep ( t ) = sep ( s ) .
Proof. Let s : = ( A , B ) s := (A, B) s := ( A , B ) and t : = ( C , D ) t := (C, D) t := ( C , D ) . Assume for a contradiction that sep ( t ) ≠ sep ( s ) \operatorname{sep}(t) \ne\operatorname{sep}(s) sep ( t ) = sep ( s ) . Since s s s and t t t are both k k k -separations, we have that sep ( s ) ∖ sep ( t ) \operatorname{sep}(s) \setminus\operatorname{sep}(t) sep ( s ) ∖ sep ( t ) and sep ( t ) ∖ sep ( s ) \operatorname{sep}(t) \setminus\operatorname{sep}(s) sep ( t ) ∖ sep ( s ) are both nonempty. Without loss of generality, assume that the A t At A t -link and the s C sC s C -link are nonempty. Since t t t does not separate vertices of sep ( s ) \operatorname{sep}(s) sep ( s ) , the s D sD sD -link is empty. Since t t t is narrow and the A t At A t -link is nonempty, the B D BD B D -corner must be empty. Thus, the A D AD A D -corner is nonempty. Since t t t is narrow, this implies that the B t Bt B t -link is also empty, contradicting that s s s and t t t cross. □
Proposition 9.20. Assume Setting 9.5. Let δ > δ 9.17 \delta> \delta_{9.17} δ > δ 9.17 and
δ ′ > δ 9.20 ′ ( q , k , b , γ N , δ ) : = max { δ 9.9 ′ ( k , δ ) , β 9.4 ( q , k , b , γ N ) + k , δ + 2 ε 9.4 ( q , k , b , γ N ) } . \delta' > \delta'_{9.20}(q,k,b,\gamma_N,\delta) := \max\{\delta'_{9.9}(k,\delta),\ \beta_{9.4}(q,k,b,\gamma_N) + k,\ \delta+ 2\varepsilon_{9.4}(q,k,b,\gamma_N)\}. δ ′ > δ 9.20 ′ ( q , k , b , γ N , δ ) := max { δ 9.9 ′ ( k , δ ) , β 9.4 ( q , k , b , γ N ) + k , δ + 2 ε 9.4 ( q , k , b , γ N )} .
Let A 1 ≠ A 2 ∈ A A_1 \ne A_2 \in\mathcal{A} A 1 = A 2 ∈ A have spread ≥ ζ 9.12 ( δ ′ , ε F ) \ge\zeta_{9.12}(\delta',\varepsilon_F) ≥ ζ 9.12 ( δ ′ , ε F ) . So A 1 δ A_1^\delta A 1 δ and A 2 δ A_2^\delta A 2 δ exist by Proposition 9.16 and ε N \varepsilon_N ε N -interlace unique splitting stars σ 1 \sigma_1 σ 1 and σ 2 \sigma_2 σ 2 of N N N , respectively. Then σ 1 ≠ σ 2 \sigma_1 \ne\sigma_2 σ 1 = σ 2 .
Proof. Let F i : = F A i F_i := F_{A_i} F i := F A i for i = 1 , 2 i = 1, 2 i = 1 , 2 . Both A i A_i A i are δ ′ \delta' δ ′ -relevant by Proposition 9.16. Hence there are δ ′ \delta' δ ′ -crossing displayed separations r i , r i ′ ∈ D ( F i ) r_i, r'_i \in D(F_i) r i , r i ′ ∈ D ( F i ) . By Lemma 9.9, we have r i , r i ′ ∈ A i δ r_i, r'_i \in A_i^\delta r i , r i ′ ∈ A i δ . Let π 2 : = π ( F 2 ) \pi_2 := \pi(F_2) π 2 := π ( F 2 ) . By Lemma 9.17 , we may assume that both r 1 r_1 r 1 and r 1 ′ r'_1 r 1 ′ γ N \gamma_N γ N -interlace π 2 \pi_2 π 2 . Both r 1 , r 1 ′ r_1,r'_1 r 1 , r 1 ′ have breadth ≥ 2 ( δ ′ − k ) ≥ β 9.4 > b \ge2(\delta'-k)\ge\beta_{9.4}>b ≥ 2 ( δ ′ − k ) ≥ β 9.4 > b .
Note that both r 1 , r 1 ′ r_1,r'_1 r 1 , r 1 ′ are narrow as G G G is ( b , k ) (b,k) ( b , k ) -narrow. We apply Lemma 9.4 to both r 1 , r 1 ′ r_1,r'_1 r 1 , r 1 ′ , which leads to one of two outcomes. Let S S S be the set of all narrow proper k k k -separations of G G G .
Outcome 1: the separator of r 1 r_1 r 1 or r 1 ′ r'_1 r 1 ′ is an S S S -totally-nested k k k -separator of G G G . Without loss of generality, the separator X X X of r 1 r_1 r 1 is S S S -totally-nested. Since r 1 ′ r'_1 r 1 ′ crosses r 1 r_1 r 1 and both r 1 , r 1 ′ r_1,r'_1 r 1 , r 1 ′ have order k k k , the separator of r 1 ′ r'_1 r 1 ′ equals X X X by Lemma 9.19. Hence X X X is the common centre of A 1 A_1 A 1 . Now Lemma 9.18 yields σ 1 ≠ σ 2 \sigma_1\ne\sigma_2 σ 1 = σ 2 .
Outcome 2: r 1 , r 1 ′ r_1,r'_1 r 1 , r 1 ′ are ε 9.4 \varepsilon_{9.4} ε 9.4 -close to displayed separations d , d ′ ∈ D ( F 2 ) d,d'\in D(F_2) d , d ′ ∈ D ( F 2 ) , respectively. Since r 1 r_1 r 1 and r 1 ′ r'_1 r 1 ′ δ ′ \delta' δ ′ -cross, Lemma 7.2 yields that d d d and d ′ d' d ′ ( δ ′ − 2 ε 9.4 ) (\delta'-2\varepsilon_{9.4}) ( δ ′ − 2 ε 9.4 ) -cross. As δ ′ − 2 ε 9.4 ≥ δ \delta'-2\varepsilon_{9.4}\ge\delta δ ′ − 2 ε 9.4 ≥ δ , we find that d , d ′ ∈ D ( F 2 ) d,d'\in D(F_2) d , d ′ ∈ D ( F 2 ) δ \delta δ -cross. Hence d , d ′ ∈ A 2 d,d'\in A_2 d , d ′ ∈ A 2 by Lemma 9.12 . Since r 1 r_1 r 1 and r 1 ′ r'_1 r 1 ′ δ ′ \delta' δ ′ -cross and r 1 ′ r'_1 r 1 ′ is ε 9.4 \varepsilon_{9.4} ε 9.4 -close to d ′ d' d ′ , we have that r 1 r_1 r 1 and d ′ d' d ′ ( δ ′ − ε 9.4 ) (\delta'-\varepsilon_{9.4}) ( δ ′ − ε 9.4 ) -cross. Using
δ ′ − ε 9.4 ≥ δ ≥ δ 9.17 ≥ δ 9.16 ≥ δ 9.12 ≥ γ A , (6) \delta'-\varepsilon_{9.4}\ge\delta\ge\delta_{9.17}\ge\delta_{9.16}\ge\delta_{9.12}\ge\gamma_A, \tag*{(6)} δ ′ − ε 9.4 ≥ δ ≥ δ 9.17 ≥ δ 9.16 ≥ δ 9.12 ≥ γ A , (6)
we find that r 1 ∈ A 2 r_1\in A_2 r 1 ∈ A 2 . But also r 1 ∈ A 1 r_1\in A_1 r 1 ∈ A 1 by Lemma 9.12 . Hence A 1 = A 2 A_1=A_2 A 1 = A 2 , a contradiction. □
Interlacing separations and torso connectivity
We show that if a splitting star σ \sigma σ is ε N \varepsilon_N ε N -interlaced by a sufficiently broad k k k -separation s s s that is not too close to N N N , then there is a unique high-spread asterisk A A A such that A δ A^\delta A δ ε N \varepsilon_N ε N -interlaces σ \sigma σ . Moreover, s s s belongs to A A A and is ε F \varepsilon_F ε F -close to a displayed separation of F A F_A F A (see Key Lemma 9.22). If no such separation exists, we show that the torso of σ \sigma σ is ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable for a suitable q ′ q' q ′ (see Lemma 9.24).
Lemma 9.21. Let N N N be a symmetric nested set of proper separations of a graph G G G , and let σ ≠ σ ′ \sigma\ne\sigma' σ = σ ′ be two splitting stars of N N N . Let s , s ′ s,s' s , s ′ be two separations of G G G that almost-ε \varepsilon ε -interlace σ , σ ′ \sigma,\sigma' σ , σ ′ , respectively. Then there is a t ∈ σ ′ t\in\sigma' t ∈ σ ′ such that
(i) s ≤ ε t s\le_\varepsilon t s ≤ ε t or s ∗ ≤ ε t s^*\le_\varepsilon t s ∗ ≤ ε t , and
(ii) t ≤ ε s ′ t\le_\varepsilon s' t ≤ ε s ′ or t ≤ ε ( s ′ ) ∗ t\le_\varepsilon(s')^* t ≤ ε ( s ′ ) ∗ .
Proof. Since σ ≠ σ ′ \sigma\ne\sigma' σ = σ ′ , we can use Construction 6.25 and Lemma 6.26 to find r ∈ σ r\in\sigma r ∈ σ and r ′ ∈ σ ′ r'\in\sigma' r ′ ∈ σ ′ such that r ∗ ≤ r ′ r^*\le r' r ∗ ≤ r ′ . Since s s s almost-ε \varepsilon ε -interlaces σ \sigma σ , we have r ≤ ε s r\le_\varepsilon s r ≤ ε s or r ≤ ε s ∗ r\le_\varepsilon s^* r ≤ ε s ∗ . So, s ∗ ≤ ε r ∗ s^*\le_\varepsilon r^* s ∗ ≤ ε r ∗ or s ≤ ε r ∗ s\le_\varepsilon r^* s ≤ ε r ∗ . With Lemma 4.9 , we have s ∗ ≤ ε r ′ s^*\le_\varepsilon r' s ∗ ≤ ε r ′ or s ≤ ε r ′ s\le_\varepsilon r' s ≤ ε r ′ . Similarly, we get r ′ ≤ ε s ′ r'\le_\varepsilon s' r ′ ≤ ε s ′ or r ′ ≤ ε ( s ′ ) ∗ r'\le_\varepsilon(s')^* r ′ ≤ ε ( s ′ ) ∗ . So, for t : = r ′ t:=r' t := r ′ the statement follows. □
Key Lemma 9.22. Assume Setting 9.5. Let δ > δ 9.17 \delta>\delta_{9.17} δ > δ 9.17 , δ ′ > δ 9.20 ′ ( … , δ ) \delta'>\delta'_{9.20}(\ldots,\delta) δ ′ > δ 9.20 ′ ( … , δ ) ,
ζ ≥ ζ 9.22 ( k , δ ′ , ε F ) : = ζ 9.12 ( δ ′ + 2 k + 1 , ε F ) , \zeta\ge\zeta_{9.22}(k,\delta',\varepsilon_F):=\zeta_{9.12}(\delta'+2k+1,\varepsilon_F), ζ ≥ ζ 9.22 ( k , δ ′ , ε F ) := ζ 9.12 ( δ ′ + 2 k + 1 , ε F ) , and
ε 9.22 ( k , ε F , ε N , ζ ) : = 2 ζ + 3 ε F + ε N + 3 k . \varepsilon_{9.22}(k,\varepsilon_F,\varepsilon_N,\zeta):=2\zeta+3\varepsilon_F+\varepsilon_N+3k. ε 9.22 ( k , ε F , ε N , ζ ) := 2 ζ + 3 ε F + ε N + 3 k . Let σ \sigma σ be a splitting star of N N N . Assume that σ \sigma σ is ε N \varepsilon_N ε N -interlaced by an ε 9.22 \varepsilon_{9.22} ε 9.22 -broad k k k -separation s s s of G G G . Assume that s s s is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in N N N . Then we have all of the following:
There exists a unique A ∈ A A\in\mathcal{A} A ∈ A such that A A A has spread ≥ ζ \ge\zeta ≥ ζ , A A A is δ ′ \delta' δ ′ -relevant and A δ A^\delta A δ ε N \varepsilon_N ε N -interlaces σ \sigma σ . Moreover, F A F_A F A has spread ≥ ζ \ge\zeta ≥ ζ .
(ii) s ∈ A s \in A s ∈ A ; in particular, s s s is ε F \varepsilon_F ε F -close to a displayed separation of F A F_A F A by (9.5.3).
(iii) s s s is ζ \zeta ζ -crossed by some separation in A ∩ D ( F A ) A \cap D(F_A) A ∩ D ( F A ) .
Proof. First, we claim that s s s is γ A \gamma_A γ A -crossed by some separation in S 0 S_0 S 0 . Assume for a contradiction that it is not. Note that s ∈ S 0 s \in S_0 s ∈ S 0 as ε 9.22 ≥ ε N ≥ β \varepsilon_{9.22} \ge\varepsilon_N \ge\beta ε 9.22 ≥ ε N ≥ β by (9.5.1). Then { s , s ∗ } \{s,s^*\} { s , s ∗ } forms a trivial γ A \gamma_A γ A -asterisk of S 0 S_0 S 0 , so s ∈ S 1 s \in S_1 s ∈ S 1 by definition of S 1 S_1 S 1 . Since N N N ε N \varepsilon_N ε N -approximates S 1 ∋ s S_1 \ni s S 1 ∋ s by (9.5.7) and since s s s has breadth ≥ ε 9.22 ≥ ε N \ge\varepsilon_{9.22} \ge\varepsilon_N ≥ ε 9.22 ≥ ε N , there is t ∈ N t \in N t ∈ N that is ε N \varepsilon_N ε N -close to s s s . But since ε N ≤ ε 9.22 \varepsilon_N \le\varepsilon_{9.22} ε N ≤ ε 9.22 , this contradicts our assumption that s s s is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in N N N .
Since s ∈ S 0 s \in S_0 s ∈ S 0 is γ A \gamma_A γ A -crossed by some separation in S 0 S_0 S 0 , there is an A ∈ A A \in\mathcal{A} A ∈ A such that s ∈ A s \in A s ∈ A . Let d d d be a displayed separation of F A F_A F A that is ε F \varepsilon_F ε F -close to s s s , which exists by (9.5.3). We apply Lemma 3.13 with ξ : = ζ + ε F + k \xi:= \zeta+ \varepsilon_F + k ξ := ζ + ε F + k , which yields one of the following two outcomes:
Outcome 1: d d d is 2 ξ 2\xi 2 ξ -close to some petal-separation p p p of F A F_A F A or its inverse. We will show that this outcome is impossible under our assumptions. If p p p is ε N \varepsilon_N ε N -broad, then p p p is ε N \varepsilon_N ε N -close to some t ∈ N t \in N t ∈ N by (9.5.7), and then s s s is ( 2 ξ + ε N + ε F ) (2\xi+ \varepsilon_N + \varepsilon_F) ( 2 ξ + ε N + ε F ) -close to t t t or t ∗ t^* t ∗ , both of which belong to N N N since N N N is symmetric, by Lemma 4.9 , contradicting our assumptions as 2 ξ + ε N + ε F ≤ ε 9.22 2\xi+ \varepsilon_N + \varepsilon_F \le\varepsilon_{9.22} 2 ξ + ε N + ε F ≤ ε 9.22 . Hence we may assume that p p p has breadth < ε N < \varepsilon_N < ε N . Then s s s has breadth < 2 ξ + ε N + ε F + k ≤ ε 9.22 < 2\xi+ \varepsilon_N + \varepsilon_F + k \le\varepsilon_{9.22} < 2 ξ + ε N + ε F + k ≤ ε 9.22 by Lemma 5.9 , contradicting our assumption that s s s has breadth ≥ ε 9.22 \ge\varepsilon_{9.22} ≥ ε 9.22 .
Outcome 2: F A F_A F A has a displayed separation d ′ d' d ′ that ( ξ − k ) (\xi-k) ( ξ − k ) -crisses d d d . Then d d d and d ′ d' d ′ ( ξ − k ) (\xi-k) ( ξ − k ) -cross by Observation 5.4 (iv). Since ξ − k ≥ ζ \xi-k \ge\zeta ξ − k ≥ ζ , it follows that F A F_A F A has spread ≥ ζ \ge\zeta ≥ ζ . Note that d , d ′ ∈ S 0 d,d' \in S_0 d , d ′ ∈ S 0 since they are 2 ( ξ − k ) 2(\xi-k) 2 ( ξ − k ) -broad (with 2 ( ξ − k ) ≥ 2 ζ 9.12 ( δ ′ + 2 k + 1 , ε F ) ≥ 2 δ ′ ≥ 2 δ ≥ 2 γ N ≥ β 2(\xi-k) \ge2\zeta_{9.12}(\delta' + 2k + 1,\varepsilon_F) \ge2\delta' \ge2\delta\ge2\gamma_N \ge\beta 2 ( ξ − k ) ≥ 2 ζ 9.12 ( δ ′ + 2 k + 1 , ε F ) ≥ 2 δ ′ ≥ 2 δ ≥ 2 γ N ≥ β by (9.5.1)). As s s s is ε F \varepsilon_F ε F -close to d d d , we have that d ′ d' d ′ ( ξ − k − ε F ) (\xi-k-\varepsilon_F) ( ξ − k − ε F ) -crosses s s s by Lemma 5.8 . As
ξ − k − ε F ≥ ζ ≥ ζ 9.22 ≥ δ ′ > δ 9.20 ′ > δ > δ 9.17 ≥ δ 9.16 ≥ δ 9.12 ≥ γ A (7) \xi-k-\varepsilon_F \ge\zeta\ge\zeta_{9.22} \ge\delta' > \delta'_{9.20} > \delta> \delta_{9.17} \ge\delta_{9.16} \ge\delta_{9.12} \ge\gamma_A \tag*{(7)} ξ − k − ε F ≥ ζ ≥ ζ 9.22 ≥ δ ′ > δ 9.20 ′ > δ > δ 9.17 ≥ δ 9.16 ≥ δ 9.12 ≥ γ A (7)
and s ∈ A s \in A s ∈ A , we have d ′ ∈ A d' \in A d ′ ∈ A (which shows (iii)), and so A A A has spread ≥ ζ \ge\zeta ≥ ζ . Hence A A A is δ ′ \delta' δ ′ -relevant, A δ A^\delta A δ exists, and A δ A^\delta A δ ε N \varepsilon_N ε N -interlaces a unique splitting star σ ′ \sigma' σ ′ of N N N by Proposition 9.16 (using δ > δ 9.16 \delta> \delta_{9.16} δ > δ 9.16 from (7) ). Note that d , d ′ ∈ D ( F A ) d,d' \in D(F_A) d , d ′ ∈ D ( F A ) ( ξ − k ) (\xi-k) ( ξ − k ) -cross with ξ − k ≥ δ ′ \xi-k \ge\delta' ξ − k ≥ δ ′ as above, so both belong to A δ A^\delta A δ by definition.
We claim that σ ′ = σ \sigma' = \sigma σ ′ = σ . Assume for a contradiction that σ ′ ≠ σ \sigma' \ne\sigma σ ′ = σ . By Lemma 9.21 , there is t ∈ N t \in N t ∈ N such that s ≤ ε N t ≤ ε N d ′ s \le_{\varepsilon_N} t \le_{\varepsilon_N} d' s ≤ ε N t ≤ ε N d ′ , say. Hence s ≤ 2 ε N d ′ s \le_{2\varepsilon_N} d' s ≤ 2 ε N d ′ by Lemma 4.9 . But we have seen above that s s s and d ′ d' d ′ ( ξ − k − ε F ) (\xi-k-\varepsilon_F) ( ξ − k − ε F ) -cross, so they ζ \zeta ζ -cross and
ζ > δ 9.16 ≥ δ 9.15 ≥ 2 ε N , \zeta> \delta_{9.16} \ge\delta_{9.15} \ge2\varepsilon_N, ζ > δ 9.16 ≥ δ 9.15 ≥ 2 ε N ,
a contradiction. Therefore, σ ′ = σ \sigma' = \sigma σ ′ = σ .
The uniqueness of A A A in (i) is given by Proposition 9.20.
Definition 9.23 (Lift, induced separation). Let σ \sigma σ be a star of separations of a graph G G G . Let H H H be the torso of σ \sigma σ . Let s = ( A , B ) s=(A,B) s = ( A , B ) be a separation of H H H , and let s ^ = ( A ^ , B ^ ) \hat{s}=(\hat{A},\hat{B}) s ^ = ( A ^ , B ^ ) be a separation of G G G . Assume that A = A ^ ∩ V ( H ) A=\hat{A}\cap V(H) A = A ^ ∩ V ( H ) , B = B ^ ∩ V ( H ) B=\hat{B}\cap V(H) B = B ^ ∩ V ( H ) and sep ( s ) = sep ( s ^ ) \operatorname{sep}(s)=\operatorname{sep}(\hat{s}) sep ( s ) = sep ( s ^ ) . Then we call s ^ \hat{s} s ^ a lift of s s s (from H H H to G G G ). Conversely, we say that s s s is induced by s ^ \hat{s} s ^ (on the torso H H H ).
Lemma 9.24 Let σ \sigma σ be a star of k k k -separations of a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph G G G . Let q ′ , ε ∈ N q',\varepsilon\in\mathbb{N} q ′ , ε ∈ N such that q ′ ≥ ε + k q' \ge\varepsilon+k q ′ ≥ ε + k . Assume that there is no ( ≤ k ) (\le k) ( ≤ k ) -separation s s s of G G G such that s s s almost-interlaces σ \sigma σ ,
s s s has breadth > q ′ >q' > q ′ , and s s s is not ε \varepsilon ε -close to any separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ . Then the torso of σ \sigma σ is ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable.
Proof. Let H H H be the torso of σ \sigma σ . Assume for a contradiction that H H H fails to be ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable. Then H H H has a ( ≤ k ) (\leq k) ( ≤ k ) -separation ( A , B ) (A,B) ( A , B ) of breadth > q ′ >q' > q ′ . Let ( A ^ , B ^ ) (\widehat{A},\widehat{B}) ( A , B ) be a lift of ( A , B ) (A,B) ( A , B ) to G G G such that ( A ^ , B ^ ) (\widehat{A},\widehat{B}) ( A , B ) is a ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G that almost-interlaces σ \sigma σ , which exists by [32 2.6.4, its proof and the definition below] . The lift ( A ^ , B ^ ) (\widehat{A},\widehat{B}) ( A , B ) has breadth > q ′ >q' > q ′ by construction.
Claim 9.24.1. No r ∈ σ ∪ σ ∗ r \in\sigma\cup\sigma^* r ∈ σ ∪ σ ∗ is ε \varepsilon ε -close to ( A ^ , B ^ ) (\widehat{A},\widehat{B}) ( A , B ) .
Proof of the claim. Let r : = ( U , W ) ∈ σ r := (U,W) \in\sigma r := ( U , W ) ∈ σ . Since H H H is the torso of σ \sigma σ , we have V ( H ) ⊆ W V(H) \subseteq W V ( H ) ⊆ W . Then,
∣ A ^ ∖ U ∣ ≥ ∣ A ∖ U ∣ ≥ ∣ A ∩ W ∣ − ∣ U ∩ W ∣ ≥ ∣ A ∩ V ( H ) ∣ − ∣ U ∩ W ∣ > q ′ − k ≥ ε , |\widehat{A} \setminus U| \geq|A \setminus U| \geq|A \cap W| - |U \cap W| \geq|A \cap V(H)| - |U \cap W| > q' - k \geq\varepsilon, ∣ A ∖ U ∣ ≥ ∣ A ∖ U ∣ ≥ ∣ A ∩ W ∣ − ∣ U ∩ W ∣ ≥ ∣ A ∩ V ( H ) ∣ − ∣ U ∩ W ∣ > q ′ − k ≥ ε ,
and
∣ B ^ ∖ U ∣ ≥ ∣ B ∖ U ∣ ≥ ∣ B ∩ W ∣ − ∣ U ∩ W ∣ ≥ ∣ B ∩ V ( H ) ∣ − ∣ U ∩ W ∣ > q ′ − k ≥ ε . |\widehat{B} \setminus U| \geq|B \setminus U| \geq|B \cap W| - |U \cap W| \geq|B \cap V(H)| - |U \cap W| > q' - k \geq\varepsilon. ∣ B ∖ U ∣ ≥ ∣ B ∖ U ∣ ≥ ∣ B ∩ W ∣ − ∣ U ∩ W ∣ ≥ ∣ B ∩ V ( H ) ∣ − ∣ U ∩ W ∣ > q ′ − k ≥ ε .
The first inequality rules out ( A ^ , B ^ ) ̸ ≈ ε ( U , W ) (\widehat{A},\widehat{B}) \mathrel{\not\approx_{\varepsilon}} (U,W) ( A , B ) ≈ ε ( U , W ) and the second rules out ( A ^ , B ^ ) ̸ ≈ ε ( W , U ) (\widehat{A},\widehat{B}) \mathrel{\not\approx_{\varepsilon}} (W,U) ( A , B ) ≈ ε ( W , U ) .
Hence s : = ( A ^ , B ^ ) s := (\widehat{A},\widehat{B}) s := ( A , B ) contradicts our assumptions.
Closeness of petal-stars and splitting stars
The next two key lemmas (9.26 and 9.29) show that the petal-star of each constructed graphic-flower and its corresponding splitting star approximate each other.
Recall that when ( H , V ) (H,\mathcal{V}) ( H , V ) is an H H H -decomposition of a graph G G G and X ⊆ H X \subseteq H X ⊆ H , then we write V ↾ X \mathcal{V} \mathord{\upharpoonright} X V ↾ X for ⋃ x ∈ V ( X ) V x \bigcup_{x\in V(X)} V_x ⋃ x ∈ V ( X ) V x .
Lemma 9.25. Let δ ∈ N \delta\in\mathbb{N} δ ∈ N . Let ( O , V ) (O,\mathcal{V}) ( O , V ) be a cycle-decomposition of a graph G G G . Let P ⊆ O P \subseteq O P ⊆ O be a path and p ∈ P p \in P p ∈ P a vertex. Assume that ∣ V p ∣ > δ |V_p| > \delta ∣ V p ∣ > δ . Then we have at least one of the following:
There is an edge x y xy x y on P P P (with x x x before y y y ) such that both V ↾ P x \mathcal{V} \mathord{\upharpoonright} Px V ↾ P x and V ↾ y P \mathcal{V} \mathord{\upharpoonright} yP V ↾ y P have size > δ > \delta > δ .
Both V ↾ P p ˚ \mathcal{V} \mathord{\upharpoonright} P\mathring{p} V ↾ P p ˚ and V ↾ p ˚ P \mathcal{V} \mathord{\upharpoonright} \mathring{p}P V ↾ p ˚ P have size ≤ δ \leq\delta ≤ δ .
Proof. Assume that (ii) fails. By symmetry, we may assume without loss of generality that ∣ V ↾ p ˚ P ∣ > δ |\mathcal{V} \mathord{\upharpoonright} \mathring{p}P| > \delta ∣ V ↾ p ˚ P ∣ > δ . We also have ∣ V ↾ P p ∣ > δ |\mathcal{V} \mathord{\upharpoonright} Pp| > \delta ∣ V ↾ P p ∣ > δ as ∣ V p ∣ > δ |V_p| > \delta ∣ V p ∣ > δ by assumption. Let x : = p x := p x := p and let y y y be the successor of p p p on the path P P P . Then x y xy x y satisfies (i).
Key Lemma 9.26. Assume Setting 9.5. Let δ > δ 9.16 \delta> \delta_{9.16} δ > δ 9.16 and δ ′ > δ 9.9 ′ ( k , δ ) \delta' > \delta'_{9.9}(k,\delta) δ ′ > δ 9.9 ′ ( k , δ ) . Let ε 9.26 ( k , ε N , δ ′ ) : = 2 ( k + ε N + δ ′ ) \varepsilon_{9.26}(k,\varepsilon_N,\delta') := 2(k+\varepsilon_N+\delta') ε 9.26 ( k , ε N , δ ′ ) := 2 ( k + ε N + δ ′ ) . Let A ∈ A A \in\mathcal{A} A ∈ A have spread ≥ ζ 9.22 ( k , δ ′ , ε F ) \geq\zeta_{9.22}(k,\delta',\varepsilon_F) ≥ ζ 9.22 ( k , δ ′ , ε F ) . So A δ A^\delta A δ exists by Proposition 9.16 and ε N \varepsilon_N ε N -interlaces a unique splitting star σ \sigma σ of N N N . Let π \pi π denote the petal-star of F A F_A F A . Then σ \sigma σ ε 9.26 \varepsilon_{9.26} ε 9.26 -approximates π \pi π .
Proof. Choose s 1 , s 2 ∈ A s_1,s_2 \in A s 1 , s 2 ∈ A that ζ 9.22 ( k , δ ′ , ε F ) \zeta_{9.22}(k,\delta',\varepsilon_F) ζ 9.22 ( k , δ ′ , ε F ) -cross, that is, ( δ ′ + 2 k + 1 + 2 ε F ) (\delta' + 2k + 1 + 2\varepsilon_F) ( δ ′ + 2 k + 1 + 2 ε F ) -cross. By (9.5.3), there exist s 1 ′ , s 2 ′ ∈ D ( F A ) s'_1,s'_2 \in D(F_A) s 1 ′ , s 2 ′ ∈ D ( F A ) such that s i s_i s i and s i ′ s'_i s i ′ are ε F \varepsilon_F ε F -close for i ∈ { 1 , 2 } i \in\{1,2\} i ∈ { 1 , 2 } . Then, s 1 ′ s'_1 s 1 ′ and s 2 ′ s'_2 s 2 ′ ( δ ′ + 2 k + 1 ) (\delta' + 2k + 1) ( δ ′ + 2 k + 1 ) -cross by Lemma 7.2. If F A F_A F A is a graphic-k k k -anemone, let F : = ( O , V ) F := (O,\mathcal{V}) F := ( O , V ) be the k k k -anemone given by Lemma 3.12 that displays both s 1 ′ s'_1 s 1 ′ and s 2 ′ s'_2 s 2 ′ . Otherwise, F A F_A F A is a graphic-k k k -daisy, and we let ( O , V ) : = F : = F A (O,\mathcal{V}) := F := F_A ( O , V ) := F := F A . In both cases, we have D ( F ) ⊆ D ( F A ) D(F) \subseteq D(F_A) D ( F ) ⊆ D ( F A ) and s 1 ′ , s 2 ′ ∈ D ( F ) s'_1,s'_2 \in D(F) s 1 ′ , s 2 ′ ∈ D ( F ) . Moreover, s 1 ′ , s 2 ′ ∈ A δ s'_1,s'_2 \in A^\delta s 1 ′ , s 2 ′ ∈ A δ by Lemma 9.9.
Let p ∈ π p \in\pi p ∈ π be ε 9.26 \varepsilon_{9.26} ε 9.26 -broad. Let x ∈ O x \in O x ∈ O be the node that induces the petal-separation p p p . Let E δ E^\delta E δ be the set of all edges of O O O that are used by 2-edge-cuts of O O O to define displayed separations that contained in A δ A^\delta A δ . Note that ∣ E δ ∣ ≥ 4 |E^\delta| \ge4 ∣ E δ ∣ ≥ 4 as A δ A^\delta A δ is nontrivial. Let P P P denote the component of O − E δ O-E^\delta O − E δ that contains x x x , and note that P P P is a path. Let Q : = O − P Q:=O-P Q := O − P , which is a path with at least two edges as ∣ E δ ∣ ≥ 4 |E^\delta| \ge4 ∣ E δ ∣ ≥ 4 .
Claim 9.26.1. There are two edges u v uv uv , u ′ v ′ u'v' u ′ v ′ on Q Q Q (with u u u before v v v before u ′ u' u ′ before v ′ v' v ′ ) such that V ↾ Q u V \upharpoonright Qu V ↾ Q u , V ↾ v Q u ′ V \upharpoonright vQu' V ↾ v Q u ′ and V ↾ v ′ Q V \upharpoonright v'Q V ↾ v ′ Q all have size > δ ′ + k > \delta' + k > δ ′ + k .
Proof of the claim. By our choice of F F F , the separations s 1 ′ , s 2 ′ ∈ D ( F ) ∩ A δ s'_1,s'_2 \in D(F) \cap A^\delta s 1 ′ , s 2 ′ ∈ D ( F ) ∩ A δ ( δ ′ + 2 k + 1 ) (\delta' + 2k + 1) ( δ ′ + 2 k + 1 ) -cross. The corresponding edge-cuts of O O O therefore give four distinct edges e 1 , e 2 , e 3 , e 4 ∈ E δ e_1,e_2,e_3,e_4 \in E^\delta e 1 , e 2 , e 3 , e 4 ∈ E δ whose deletion splits O O O into four consecutive paths P 1 , … , P 4 P_1,\ldots,P_4 P 1 , … , P 4 such that V ↾ P i V \upharpoonright P_i V ↾ P i has size ≥ ( δ ′ + 2 k + 1 ) − k > δ ′ + k \ge(\delta' + 2k + 1)-k > \delta' + k ≥ ( δ ′ + 2 k + 1 ) − k > δ ′ + k by Observation 5.4.
Since all four edges e 1 , … , e 4 e_1,\ldots,e_4 e 1 , … , e 4 are contained in E δ E^\delta E δ , it follows that P 1 P_1 P 1 , say, includes the entire path P P P . The other three paths avoid P P P and hence are included in Q Q Q . Then we let u v uv uv be the edge on O O O in between P 2 P_2 P 2 and P 3 P_3 P 3 with u u u coming before v v v on Q Q Q , and let u ′ v ′ u'v' u ′ v ′ be the edge on O O O in between P 3 P_3 P 3 and P 4 P_4 P 4 with u ′ u' u ′ coming before v ′ v' v ′ on Q Q Q .
Consider the displayed separation s : = ( V ↾ P , V ↾ Q ) s := (V \upharpoonright P,V \upharpoonright Q) s := ( V ↾ P , V ↾ Q ) of F F F .
Claim 9.26.2. No displayed separation of F F F δ ′ \delta' δ ′ -crosses s s s .
Proof of the claim. Assume for a contradiction that some d ∈ D ( F ) d \in D(F) d ∈ D ( F ) δ ′ \delta' δ ′ -crosses s s s . Since s , d ∈ D ( F ) ⊆ D ( F A ) s,d \in D(F) \subseteq D(F_A) s , d ∈ D ( F ) ⊆ D ( F A ) and s , d s,d s , d δ ′ \delta' δ ′ -cross, Lemma 9.9 implies that s , d ∈ A δ s,d \in A^\delta s , d ∈ A δ . Consider the 2-edge-cuts of O O O that induce s s s and d d d , respectively. These must cross. For this, the 2-edge-cut that induces d d d must use an edge of P P P . But since d ∈ A δ d \in A^\delta d ∈ A δ , this edge also has to be contained in E δ E^\delta E δ , a contradiction.
Since x ∈ P x \in P x ∈ P we have p ≤ s p \le s p ≤ s . Since p p p has breadth > δ ′ + k > \delta' + k > δ ′ + k , we have ∣ V x ∣ > δ ′ + k |V_x| > \delta' + k ∣ V x ∣ > δ ′ + k . We apply Lemma 9.25 (with δ ′ + k \delta' + k δ ′ + k ), which has two possible outcomes.
Case 1: Lemma 9.25 yields outcome (i). We claim that this outcome is impossible. Assume for a contradiction that we get it, that is, there is an edge y z yz y z on P P P (with y y y before z z z on P P P ) such that both V ↾ P y V \upharpoonright Py V ↾ P y and V ↾ z P V \upharpoonright zP V ↾ z P have size > δ ′ + k > \delta' + k > δ ′ + k . Let u v ∈ E ( Q ) uv \in E(Q) uv ∈ E ( Q ) be as provided by Claim 9.26.1 . Let d ∈ D ( F ) d \in D(F) d ∈ D ( F ) be induced by the 2-edge-cut { y z , u v } \{yz,uv\} { y z , uv } of O O O . Then d d d δ ′ \delta' δ ′ -crisses s s s , and so d d d δ ′ \delta' δ ′ -crosses s s s by Observation 5.4. This contradicts Claim 9.26.2 .
Case 2: Lemma 9.25 yields outcome (ii). That is to say that V ↾ P x ˚ V \upharpoonright P\mathring{x} V ↾ P x ˚ and V ↾ x ˚ P V \upharpoonright\mathring{x}P V ↾ x ˚ P have size ≤ δ ′ + k \le\delta' + k ≤ δ ′ + k . Hence p ≤ s p \le s p ≤ s and s ≤ 2 ( δ ′ + k ) p s \le2(\delta' + k)p s ≤ 2 ( δ ′ + k ) p , so
p ≈ 2 ( δ ′ + k ) s . (8) p \approx_{2(\delta' + k)} s. \tag*{(8)} p ≈ 2 ( δ ′ + k ) s . (8)
As p ∈ S 1 p \in S_1 p ∈ S 1 has breadth ≥ ε 9.26 ≥ ε N \ge\varepsilon_{9.26} \ge\varepsilon_N ≥ ε 9.26 ≥ ε N , by (9.5.7) there is t ∈ N t \in N t ∈ N such that t t t is ε N \varepsilon_N ε N -close to p p p . If t ∈ σ t \in\sigma t ∈ σ , then we are done as ε N ≤ ε 9.26 \varepsilon_N \le\varepsilon_{9.26} ε N ≤ ε 9.26 . Hence we may assume that t ∉ σ t \notin\sigma t ∈ / σ .
We claim that there are d 1 , d 2 ∈ A δ d_1,d_2 \in A^\delta d 1 , d 2 ∈ A δ such that s = d 1 ∧ d 2 s=d_1 \wedge d_2 s = d 1 ∧ d 2 . By Claim 9.26.1 there are two edges u v , u ′ v ′ uv,u'v' uv , u ′ v ′ of Q Q Q (with u u u before v v v before u ′ u' u ′ before v ′ v' v ′ ) such that all of V ↾ Q u V \upharpoonright Qu V ↾ Q u , V ↾ v Q u ′ V \upharpoonright vQu' V ↾ v Q u ′ and V ↾ v ′ Q V \upharpoonright v'Q V ↾ v ′ Q have size > δ ′ + k > \delta' + k > δ ′ + k . Let y z , y ′ z ′ yz,y'z' y z , y ′ z ′ be the two edges that together induce the separation s s s , and are such that y , z , y ′ , z ′ , u , v , u ′ , v ′ y,z,y',z',u,v,u',v' y , z , y ′ , z ′ , u , v , u ′ , v ′ appear in this cyclic order along O O O . Then let d 1 d_1 d 1 be induced by { y z , u w } \{yz,uw\} { y z , u w } and let d 2 d_2 d 2 be induced by { y ′ z ′ , u ′ v ′ } \{y'z',u'v'\} { y ′ z ′ , u ′ v ′ } such that s = d 1 ∧ d 2 s=d_1 \wedge d_2 s = d 1 ∧ d 2 . Note that they δ ′ \delta' δ ′ -criss, hence they δ ′ \delta' δ ′ -cross by Observation 5.4. Hence, d 1 , d 2 ∈ A δ d_1,d_2 \in A^\delta d 1 , d 2 ∈ A δ . Since p ≤ s p \leq s p ≤ s , we have p ≤ d 1 , d 2 p \leq d_1,d_2 p ≤ d 1 , d 2 and, moreover, t ≤ ε N d 1 , d 2 t \leq_{\varepsilon_N} d_1,d_2 t ≤ ε N d 1 , d 2 . With the notations of Lemma 9.15, we have t ∈ O ( d 1 ) t \in O(d_1) t ∈ O ( d 1 ) , so there exists r ∈ σ ( d 1 ) r \in\sigma(d_1) r ∈ σ ( d 1 ) such that t ≤ r t \leq r t ≤ r . Since d 1 ∈ A δ d_1 \in A^\delta d 1 ∈ A δ , Lemma 9.15 implies that σ = σ ( d 1 ) \sigma= \sigma(d_1) σ = σ ( d 1 ) , so r ∈ σ r \in\sigma r ∈ σ . Since A δ A^\delta A δ ε N \varepsilon_N ε N -interlaces σ \sigma σ , we have r < ε N d i r <_{\varepsilon_N} d_i r < ε N d i or r < ε N d i ∗ r <_{\varepsilon_N} d_i^* r < ε N d i ∗ for i = 1 , 2 i=1,2 i = 1 , 2 . We claim that r < ε N d 1 , d 2 r <_{\varepsilon_N} d_1,d_2 r < ε N d 1 , d 2 , and assume for a contradiction that r < ε N d 1 ∗ r <_{\varepsilon_N} d_1^* r < ε N d 1 ∗ , say. Then
p ≤ ε N t ≤ r < ε N d 1 ∗ p \leq_{\varepsilon_N} t \leq r <_{\varepsilon_N} d_1^* p ≤ ε N t ≤ r < ε N d 1 ∗
which implies p ≤ 2 ε N d 1 ∗ p \leq_{2\varepsilon_N} d_1^* p ≤ 2 ε N d 1 ∗ by Lemma 4.9 . But also p ≤ d 1 p \leq d_1 p ≤ d 1 , which is equivalent to d 1 ∗ ≤ p ∗ d_1^* \leq p^* d 1 ∗ ≤ p ∗ , and combines with the previous sentence to p ≤ 2 ε N p ∗ p \leq_{2\varepsilon_N} p^* p ≤ 2 ε N p ∗ . Thus Lemma 8.3 yields that p p p has breadth ≤ 2 ε N \leq2\varepsilon_N ≤ 2 ε N , contradicting that p p p has breadth ≥ ε 9.26 > 2 ε N \geq\varepsilon_{9.26} > 2\varepsilon_N ≥ ε 9.26 > 2 ε N . Therefore, we have r < ε N d 1 , d 2 r <_{\varepsilon_N} d_1,d_2 r < ε N d 1 , d 2 as claimed. Hence r ≤ 2 ε N d 1 ∧ d 2 = s r \leq_{2\varepsilon_N} d_1 \wedge d_2 = s r ≤ 2 ε N d 1 ∧ d 2 = s by Lemma 4.10 . Therefore,
p ≤ ε N t ≤ r ≤ 2 ε N s ≤ 2 ( δ ′ + k ) p . (9) p \leq_{\varepsilon_N} t \leq r \leq_{2\varepsilon_N} s \leq_{2(\delta' + k)} p. \tag*{(9)} p ≤ ε N t ≤ r ≤ 2 ε N s ≤ 2 ( δ ′ + k ) p . (9)
Hence p p p is 2 ( δ ′ + k + ε N ) 2(\delta' + k + \varepsilon_N) 2 ( δ ′ + k + ε N ) -close to r ∈ σ r \in\sigma r ∈ σ by Lemma 4.9 . □
Lemma 9.27. Let ( A , B ) ≤ r ∧ s (A,B) \leq r \wedge s ( A , B ) ≤ r ∧ s , where r r r and s s s are k k k -separations and ( A , B ) (A,B) ( A , B ) is a separation of a graph G G G . Assume that r r r and s s s γ \gamma γ -cross. Then ∣ B ∖ A ∣ ≥ 3 ( γ − k ) |B \setminus A| \geq3(\gamma-k) ∣ B ∖ A ∣ ≥ 3 ( γ − k ) .
Proof. By Observation 5.4, r r r and s s s ( γ − k ) (\gamma-k) ( γ − k ) -criss, so all four corners have size ≥ γ − k \geq\gamma-k ≥ γ − k . Three of these corners are included in B ∖ A B \setminus A B ∖ A . □
Lemma 9.28. Let G G G be a ( b , k ) (b,k) ( b , k ) -narrow graph and X ⊆ V ( G ) X \subseteq V(G) X ⊆ V ( G ) of size ∣ X ∣ = k |X|=k ∣ X ∣ = k . Assume that the graphic-X X X -anemone A A A has a b b b -broad displayed separation. Then every petal-separation of A A A is narrow.
Proof. Let s s s be a b b b -broad displayed separation of the graphic-X X X -anemone A A A . Note that X = sep ( s ) X=\operatorname{sep}(s) X = sep ( s ) . Since G G G is ( b , k ) (b,k) ( b , k ) -narrow, the separation s s s is narrow. Hence, every component K K K of G − X G-X G − X has N ( K ) = X N(K)=X N ( K ) = X . In particular, every petal-separation of A A A is narrow. □
Key Lemma 9.29. Assume Setting 9.5. Let δ > δ 9.16 \delta> \delta_{9.16} δ > δ 9.16 , δ ′ > max { δ 9.9 ′ ( k , δ ) , δ + 2 ε 9.4 ( q , k , b , γ N ) } \delta' > \max\{\delta'_{9.9}(k,\delta),\delta+2\varepsilon_{9.4}(q,k,b,\gamma_N)\} δ ′ > max { δ 9.9 ′ ( k , δ ) , δ + 2 ε 9.4 ( q , k , b , γ N )} and
ε 9.29 ( q , k , b , γ N , ε N , δ ′ ) : = max { 3 ε 9.4 ( q , k , b , γ N ) + 2 γ N + 2 k , ε 9.26 + γ N + 1 , β 9.4 ( q , k , b , γ N ) } . \varepsilon_{9.29}(q,k,b,\gamma_N,\varepsilon_N,\delta') := \max\{3\varepsilon_{9.4}(q,k,b,\gamma_N)+2\gamma_N+2k,\varepsilon_{9.26}+\gamma_N+1,\beta_{9.4}(q,k,b,\gamma_N)\}. ε 9.29 ( q , k , b , γ N , ε N , δ ′ ) := max { 3 ε 9.4 ( q , k , b , γ N ) + 2 γ N + 2 k , ε 9.26 + γ N + 1 , β 9.4 ( q , k , b , γ N )} .
Let A ∈ A A \in\mathcal{A} A ∈ A have spread ≥ ζ 9.22 ( k , δ ′ , ε F ) \geq\zeta_{9.22}(k,\delta',\varepsilon_F) ≥ ζ 9.22 ( k , δ ′ , ε F ) . So A A A is δ ′ \delta' δ ′ -relevant, A δ A^\delta A δ exists by Proposition 9.16 and ε N \varepsilon_N ε N -interlaces a unique splitting star σ \sigma σ of N N N . Let π \pi π denote the petal-star of F A F_A F A . Then π \pi π ε 9.29 \varepsilon_{9.29} ε 9.29 -approximates σ \sigma σ .
Proof. Let s = : ( U , W ) ∈ σ s =: (U,W) \in\sigma s =: ( U , W ) ∈ σ . We may assume that s s s is ε 9.29 \varepsilon_{9.29} ε 9.29 -broad, as otherwise there is nothing to show. Since s ∈ σ ⊆ N s \in\sigma\subseteq N s ∈ σ ⊆ N , s s s is a k k k -separation of G G G so s s s is γ N \gamma_N γ N -nested with all the separations in π \pi π by (9.5.4). Hence we are in one of the following two cases:
Case 1: s s s almost-γ N \gamma_N γ N -interlaces π \pi π . We assumed earlier that s s s is ε 9.29 \varepsilon_{9.29} ε 9.29 -broad, and we have ε 9.29 ≥ β 9.4 \varepsilon_{9.29} \geq\beta_{9.4} ε 9.29 ≥ β 9.4 . We apply Lemma 9.4 , which has two possible outcomes:
Case 1a: s s s is ε 9.4 \varepsilon_{9.4} ε 9.4 -close to some displayed separation d d d of F A F_A F A . Since s s s is narrow-γ N \gamma_N γ N -nested by (9.5.5), and d d d is ε 9.4 \varepsilon_{9.4} ε 9.4 -close to s s s , Lemma 5.8 yields that d d d cannot be ( ε 9.4 + γ N ) (\varepsilon_{9.4}+\gamma_N) ( ε 9.4 + γ N ) -crossed by any narrow proper k k k -separation. Even more, since γ N ≥ β + k ≥ b \gamma_N \ge\beta+ k \ge b γ N ≥ β + k ≥ b by (9.5.1), and using Observation 5.4, we have that d d d cannot be ( ε 9.4 + γ N ) (\varepsilon_{9.4}+\gamma_N) ( ε 9.4 + γ N ) -crossed by any k k k -separation. So by Lemma 3.13, d d d is 2 ( ε 9.4 + γ N + k ) 2(\varepsilon_{9.4}+\gamma_N+k) 2 ( ε 9.4 + γ N + k ) -close to some p ∈ π p \in\pi p ∈ π or its inverse. Hence s s s is ( 3 ε 9.4 + 2 γ N + 2 k ) (3\varepsilon_{9.4}+2\gamma_N+2k) ( 3 ε 9.4 + 2 γ N + 2 k ) -close to some p ∈ π p \in\pi p ∈ π or its inverse by Lemma 4.9 .
Suppose for a contradiction that s s s is ( 3 ε 9.4 + 2 γ N + 2 k ) (3\varepsilon_{9.4}+2\gamma_N+2k) ( 3 ε 9.4 + 2 γ N + 2 k ) -close to some p ∗ = : ( Q , P ) ∈ π ∗ p^* =: (Q,P) \in\pi^* p ∗ =: ( Q , P ) ∈ π ∗ . Since A A A is δ ′ \delta' δ ′ -relevant, there exist two δ ′ \delta' δ ′ -crossing separations r 1 , r 2 ∈ A δ ⊆ D ( F A ) r_1,r_2 \in A^\delta\subseteq D(F_A) r 1 , r 2 ∈ A δ ⊆ D ( F A ) . Since A δ A^\delta A δ ε N \varepsilon_N ε N -interlaces σ \sigma σ and s ∈ σ s \in\sigma s ∈ σ , we can orient each r i r_i r i so that s ≤ ε N r i s \le_{\varepsilon_N} r_i s ≤ ε N r i . Write r 1 = : ( A 1 , B 1 ) r_1 =: (A_1,B_1) r 1 =: ( A 1 , B 1 ) and r 2 = : ( A 2 , B 2 ) r_2 =: (A_2,B_2) r 2 =: ( A 2 , B 2 ) . Since P P P meets at most one of their four corners, Q Q Q contains two corners that are disjoint from A 1 ∩ A 2 A_1 \cap A_2 A 1 ∩ A 2 . Each has size > δ ′ − k > \delta' - k > δ ′ − k by Observation 5.4 (ii). However, ∣ U ∖ ( A 1 ∩ A 2 ) ∣ ≤ ∣ U ∖ A 1 ∣ + ∣ U ∖ A 2 ∣ ≤ 2 ε N |U \setminus(A_1 \cap A_2)| \le|U \setminus A_1| + |U \setminus A_2| \le2\varepsilon_N ∣ U ∖ ( A 1 ∩ A 2 ) ∣ ≤ ∣ U ∖ A 1 ∣ + ∣ U ∖ A 2 ∣ ≤ 2 ε N . Thus, using that δ ′ ≥ δ + 2 ε 9.4 ≥ γ N + 2 ε N + 2 ε 9.4 ≥ γ N + ε N + 2 k + 2 ε 9.4 \delta' \ge\delta+ 2\varepsilon_{9.4} \ge\gamma_N + 2\varepsilon_N + 2\varepsilon_{9.4} \ge\gamma_N + \varepsilon_N + 2k + 2\varepsilon_{9.4} δ ′ ≥ δ + 2 ε 9.4 ≥ γ N + 2 ε N + 2 ε 9.4 ≥ γ N + ε N + 2 k + 2 ε 9.4 by (9.5.1), it follows that
3 ε 9.4 + 2 γ N + 2 k ≥ ∣ Q ∖ U ∣ ≥ ∣ Q ∖ ( A 1 ∩ A 2 ) ∣ − ∣ U ∖ ( A 1 ∩ A 2 ) ∣ > 2 ( δ ′ − k − ε N ) ≥ 3 ε 9.4 + 2 γ N + 2 k . 3\varepsilon_{9.4}+2\gamma_N+2k \ge|Q \setminus U| \ge|Q \setminus(A_1 \cap A_2)| - |U \setminus(A_1 \cap A_2)| > 2(\delta' - k - \varepsilon_N) \ge3\varepsilon_{9.4}+2\gamma_N+2k . 3 ε 9.4 + 2 γ N + 2 k ≥ ∣ Q ∖ U ∣ ≥ ∣ Q ∖ ( A 1 ∩ A 2 ) ∣ − ∣ U ∖ ( A 1 ∩ A 2 ) ∣ > 2 ( δ ′ − k − ε N ) ≥ 3 ε 9.4 + 2 γ N + 2 k .
a contradiction. Thus, s s s is ( 3 ε 9.4 + 2 γ N + 2 k ) (3\varepsilon_{9.4}+2\gamma_N+2k) ( 3 ε 9.4 + 2 γ N + 2 k ) -close, hence ε 9.29 \varepsilon_{9.29} ε 9.29 -close, to some p ∈ π p \in\pi p ∈ π .
Case 1b: X : = sep ( s ) X := \operatorname{sep}(s) X := sep ( s ) is S S S -totally-nested, where S S S is the set of all narrow proper k k k -separations of G G G , and s ∈ D ( F ) s \in D(F) s ∈ D ( F ) for the graphic-X X X -anemone F F F . Moreover, s s s is not half-connected. Let π X : = π ( F ) ∖ { ( X , V ( G ) ) } \pi_X := \pi(F) \setminus\{(X,V(G))\} π X := π ( F ) ∖ {( X , V ( G ))} be the star of proper petal-separations of F F F . Using that s s s is ε 9.29 \varepsilon_{9.29} ε 9.29 -broad with ε 9.29 > γ N ≥ β ≥ q \varepsilon_{9.29} > \gamma_N \ge\beta\ge q ε 9.29 > γ N ≥ β ≥ q by (9.5.1) and that G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, we have that ∣ X ∣ = k |X|=k ∣ X ∣ = k and F F F has a b b b -broad displayed separation, namely s s s . By Lemma 9.28, all the separations in π X \pi_X π X are narrow. Thus, π X ∪ π X ∗ \pi_X \cup\pi_X^* π X ∪ π X ∗ is the set of all narrow proper half-connected k k k -separations with separator X X X , and π X ≠ ∅ \pi_X \ne\varnothing π X = ∅ as s s s is a narrow proper k k k -separation with separator X X X . Since X X X is S S S -totally-nested, so are the separations in π X \pi_X π X by Lemma 9.1 . Then the separations in π X \pi_X π X are contained in N N N by (9.5.6).
We now show that π X ≼ σ \pi_X \preccurlyeq\sigma π X ≼ σ . Let p ∈ π X p \in\pi_X p ∈ π X . As s ∈ D ( F ) s \in D(F) s ∈ D ( F ) , we have that s s s almost-interlaces π X \pi_X π X , so p ≤ s p \le s p ≤ s or p ≤ s ∗ p \le s^* p ≤ s ∗ . If p ≤ s p \le s p ≤ s we are done, so assume p ≤ s ∗ p \le s^* p ≤ s ∗ . Since σ \sigma σ is a splitting star of N N N and p ∈ π X ⊆ N p \in\pi_X \subseteq N p ∈ π X ⊆ N , there exists t ∈ σ t \in\sigma t ∈ σ such that t ≰ p t \not\le p t ≤ p and t ≰ p ∗ t \not\le p^* t ≤ p ∗ . Since N N N is nested, we have p ≤ t p \le t p ≤ t or p ∗ ≤ t p^* \le t p ∗ ≤ t . If p ≤ t p \le t p ≤ t we are done, so assume p ∗ ≤ t p^* \le t p ∗ ≤ t . If t = s t=s t = s then s ≤ p ∗ ≤ t = s s \le p^* \le t=s s ≤ p ∗ ≤ t = s so s = p ∗ s=p^* s = p ∗ . This is impossible since p p p is half-connected and s s s is not. Thus t ≠ s t \ne s t = s , so, using that σ \sigma σ is a star, we have s ≤ p ∗ ≤ t ≤ s ∗ s \le p^* \le t \le s^* s ≤ p ∗ ≤ t ≤ s ∗ , which would contradict that s s s is proper. Therefore, we indeed have π X ≼ σ \pi_X \preccurlyeq\sigma π X ≼ σ .
By assumption, A δ A^\delta A δ exists. Let r 1 , r 2 ∈ A δ r_1,r_2 \in A^\delta r 1 , r 2 ∈ A δ δ \delta δ -cross. Thus, r 1 , r 2 r_1,r_2 r 1 , r 2 have breadth ≥ 2 ( δ − k ) \ge2(\delta-k) ≥ 2 ( δ − k ) by Lemma 5.6, with δ > δ 9.16 ≥ δ 9.12 ≥ γ A + k \delta> \delta_{9.16} \ge\delta_{9.12} \ge\gamma_A+k δ > δ 9.16 ≥ δ 9.12 ≥ γ A + k , so they have breadth ≥ γ A ≥ β \ge\gamma_A \ge\beta ≥ γ A ≥ β by (9.5.1). Then both r i r_i r i are narrow and have order exactly k k k , by Observation 9.6. Hence r 1 , r 2 ∈ S r_1,r_2 \in S r 1 , r 2 ∈ S . Both r i r_i r i ε N \varepsilon_N ε N -interlace σ \sigma σ by assumption. Since π X ≼ σ \pi_X \preccurlyeq\sigma π X ≼ σ , both r i r_i r i ε N \varepsilon_N ε N -interlace π X \pi_X π X as well. We claim that both r i r_i r i interlace π X \pi_X π X . Since the r i r_i r i ε N \varepsilon_N ε N -interlace π X \pi_X π X , neither r i r_i r i lies in π X \pi_X π X , so by symmetry it suffices to show that r 1 r_1 r 1 almost-interlaces π X \pi_X π X . For this, let ℓ ∈ π X \ell\in\pi_X ℓ ∈ π X . If ℓ ≤ r 1 \ell\le r_1 ℓ ≤ r 1 or ℓ ≤ r 1 ∗ \ell\le r_1^* ℓ ≤ r 1 ∗ , then we are done, so suppose that this fails. As the elements of π X \pi_X π X are S S S -totally-nested and both r i r_i r i lie in S S S , we then have r 1 ≤ ℓ r_1 \le\ell r 1 ≤ ℓ or r 1 ∗ ≤ ℓ r_1^* \le\ell r 1 ∗ ≤ ℓ . Since r 1 r_1 r 1 ε N \varepsilon_N ε N -interlaces π X \pi_X π X , we also have ℓ < ε N r 1 \ell<_{\varepsilon_N} r_1 ℓ < ε N r 1 or ℓ < ε N r 1 ∗ \ell<_{\varepsilon_N} r_1^* ℓ < ε N r 1 ∗ . The two cases r 1 ≤ ℓ < ε N r 1 r_1 \le\ell<_{\varepsilon_N} r_1 r 1 ≤ ℓ < ε N r 1 and r 1 ∗ ≤ ℓ < ε N r 1 ∗ r_1^* \le\ell<_{\varepsilon_N} r_1^* r 1 ∗ ≤ ℓ < ε N r 1 ∗ are clearly impossible. If r 1 ≤ ℓ < ε N r 1 ∗ r_1 \le\ell<_{\varepsilon_N} r_1^* r 1 ≤ ℓ < ε N r 1 ∗ , then r 1 r_1 r 1 has breadth ≤ ε N \le\varepsilon_N ≤ ε N by Lemma 8.3 , contradicting that r 1 r_1 r 1 is $2(\delta-k)$-broad with 2 ( δ − k ) > 2 ( δ 9.16 − k ) ≥ 2 ( δ 9.15 − k ) ≥ 2 ( 2 ε N − k ) ≥ 2 ε N ≥ ε N 2(\delta-k)>2(\delta_{9.16}-k)\ge2(\delta_{9.15}-k)\ge2(2\varepsilon_N-k)\ge2\varepsilon_N\ge\varepsilon_N 2 ( δ − k ) > 2 ( δ 9.16 − k ) ≥ 2 ( δ 9.15 − k ) ≥ 2 ( 2 ε N − k ) ≥ 2 ε N ≥ ε N (using ε N ≥ k \varepsilon_N \ge k ε N ≥ k by (9.5.1)). Similarly, the case r 1 ∗ ≤ ℓ < ε N r 1 r_1^* \le\ell<_{\varepsilon_N} r_1 r 1 ∗ ≤ ℓ < ε N r 1 yields a contradiction. Therefore, r 1 r_1 r 1 almost-interlaces π X \pi_X π X . This completes the proof of the claim that both r i r_i r i interlace $\pi_X.
Then r 1 r_1 r 1 and r 2 r_2 r 2 must cross with centre equal to X X X and r i ∈ D ( F ) r_i \in D(F) r i ∈ D ( F ) . Since A δ ⊆ A A^\delta\subseteq A A δ ⊆ A by Lemma 9.12, we have r 1 , r 2 ∈ A r_1,r_2 \in A r 1 , r 2 ∈ A , and so X X X is the common centre of A A A . By (9.5.2), we have F = F A F=F_A F = F A . So s ∈ D ( F ) = D ( F A ) s \in D(F)=D(F_A) s ∈ D ( F ) = D ( F A ) . Then the premise of Case 1a is satisfied, and so we conclude the argument as in Case 1a.
Case 2: there is p ∈ π p \in\pi p ∈ π such that s ≤ γ N p s \le_{\gamma_N} p s ≤ γ N p or s ∗ ≤ γ N p s^* \le_{\gamma_N} p s ∗ ≤ γ N p . We first claim that s ∗ ≤ γ N p s^* \le_{\gamma_N} p s ∗ ≤ γ N p is impossible, and assume for a contradiction that s ∗ ≤ γ N p s^* \le_{\gamma_N} p s ∗ ≤ γ N p . Then s ∗ ≤ ε N p s^* \le_{\varepsilon_N} p s ∗ ≤ ε N p since ε N ≥ γ N \varepsilon_N \ge\gamma_N ε N ≥ γ N by (9.5.1). Let r ∈ A δ r \in A^\delta r ∈ A δ . Then r ∈ A δ ⊆ A r \in A^\delta\subseteq A r ∈ A δ ⊆ A by Lemma 9.12. As r r r ε N \varepsilon_N ε N -interlaces σ \sigma σ , we have s ≤ ε N r s \le_{\varepsilon_N} r s ≤ ε N r or s ≤ ε N r ∗ s \le_{\varepsilon_N} r^* s ≤ ε N r ∗ , so s ∈ O ( r ) s \in O(r) s ∈ O ( r ) with O ( r ) O(r) O ( r ) from Lemma 9.15 (5). As r ∈ A δ ⊆ D ( F A ) r \in A^\delta\subseteq D(F_A) r ∈ A δ ⊆ D ( F A ) by Lemma 9.12, we also have p ≤ r p \le r p ≤ r or p ≤ r ∗ p \le r^* p ≤ r ∗ . Combining this with s ∗ ≤ ε N p s^* \le_{\varepsilon_N} p s ∗ ≤ ε N p yields s ∗ ≤ ε N r s^* \le_{\varepsilon_N} r s ∗ ≤ ε N r or s ∗ ≤ ε N r ∗ s^* \le_{\varepsilon_N} r^* s ∗ ≤ ε N r ∗ . Hence s ∗ ∈ O ( r ) s^* \in O(r) s ∗ ∈ O ( r ) . But s , s ∗ ∈ O ( r ) s,s^* \in O(r) s , s ∗ ∈ O ( r ) contradicts that O ( r ) O(r) O ( r ) is an orientation of N N N by Lemma 9.15.
Therefore, s ≤ γ N p s \le_{\gamma_N} p s ≤ γ N p . We claim that p p p is ε 9.26 \varepsilon_{9.26} ε 9.26 -broad. Let ( U , W ) : = p (U,W):=p ( U , W ) := p and ( C , D ) : = s (C,D):=s ( C , D ) := s . Since s ≤ γ N p s \le_{\gamma_N} p s ≤ γ N p and since s s s is ε 9.29 \varepsilon_{9.29} ε 9.29 -broad, we have
∣ U ∖ W ∣ ≥ ∣ ( U ∖ W ) ∩ ( C ∖ D ) ∣ = ∣ C ∖ D ∣ − ∣ W ∖ D ∣ ≥ ε 9.29 − γ N ≥ ε 9.26 . |U \setminus W| \ge|(U \setminus W) \cap(C \setminus D)| = |C \setminus D| - |W \setminus D| \ge\varepsilon_{9.29} - \gamma_N \ge\varepsilon_{9.26}. ∣ U ∖ W ∣ ≥ ∣ ( U ∖ W ) ∩ ( C ∖ D ) ∣ = ∣ C ∖ D ∣ − ∣ W ∖ D ∣ ≥ ε 9.29 − γ N ≥ ε 9.26 .
As A A A is δ ′ \delta' δ ′ -relevant by Proposition 9.16, there are r 1 , r 2 ∈ A δ r_1,r_2 \in A^\delta r 1 , r 2 ∈ A δ that δ ′ \delta' δ ′ -cross. Hence both r i r_i r i have breadth ≥ 2 ( δ ′ − k ) \ge2(\delta'-k) ≥ 2 ( δ ′ − k ) by Lemma 5.6. Since p ∈ π p \in\pi p ∈ π , we have p ≤ r 1 ∧ r 2 p \le r_1 \wedge r_2 p ≤ r 1 ∧ r 2 , say, so ∣ W ∖ U ∣ ≥ 3 ( δ ′ − k ) |W \setminus U| \ge3(\delta'-k) ∣ W ∖ U ∣ ≥ 3 ( δ ′ − k ) by Lemma 9.27. Using δ ′ > δ 9.9 ′ = 2 δ + 2 k \delta' > \delta'_{9.9}=2\delta+2k δ ′ > δ 9.9 ′ = 2 δ + 2 k , we find that
∣ W ∖ U ∣ ≥ 3 ( δ ′ − k ) ≥ 2 δ ′ + ( δ 9.9 ′ − 3 k ) ≥ 2 δ ′ + 2 δ − k ≥ 2 δ ′ + 2 δ 9.16 − k ≥ 2 δ ′ + δ 9.12 + δ 9.15 − k ≥ 2 δ ′ + k + 2 ε N + γ N ≥ 2 δ ′ + 2 k + 2 ε N = ε 9.26 . (10) \begin{aligned}
|W \setminus U| &\ge3(\delta'-k)\\
&\ge2\delta' + (\delta'_{9.9}-3k)\\
&\ge2\delta' + 2\delta-k\\
&\ge2\delta' + 2\delta_{9.16}-k\\
&\ge2\delta' + \delta_{9.12}+\delta_{9.15}-k\\
&\ge2\delta' + k+2\varepsilon_N+\gamma_N\\
&\ge2\delta' + 2k+2\varepsilon_N=\varepsilon_{9.26}. \tag*{(10)}
\end{aligned} ∣ W ∖ U ∣ ≥ 3 ( δ ′ − k ) ≥ 2 δ ′ + ( δ 9.9 ′ − 3 k ) ≥ 2 δ ′ + 2 δ − k ≥ 2 δ ′ + 2 δ 9.16 − k ≥ 2 δ ′ + δ 9.12 + δ 9.15 − k ≥ 2 δ ′ + k + 2 ε N + γ N ≥ 2 δ ′ + 2 k + 2 ε N = ε 9.26 . (10)
Therefore, p p p is ε 9.26 \varepsilon_{9.26} ε 9.26 -broad, as desired.
By Key Lemma 9.26, there is t ∈ σ t \in\sigma t ∈ σ that is ε 9.26 \varepsilon_{9.26} ε 9.26 -close to p p p . If s = t s=t s = t then we are done, so assume s ≠ t s \ne t s = t . Now s ≤ γ N p ≤ ε 9.26 t s \le_{\gamma_N}p \le_{\varepsilon_{9.26}}t s ≤ γ N p ≤ ε 9.26 t yields s ≤ γ N + ε 9.26 t s \le_{\gamma_N+\varepsilon_{9.26}}t s ≤ γ N + ε 9.26 t by Lemma 4.9. Since s , t ∈ σ s,t \in\sigma s , t ∈ σ and s ≠ t s \ne t s = t , by the definition of star we also have s ≤ t ∗ s \le t^* s ≤ t ∗ . As s ≤ γ N + ε 9.26 t s \le_{\gamma_N+\varepsilon_{9.26}}t s ≤ γ N + ε 9.26 t is equivalent to t ∗ ≤ γ N + ε 9.26 s ∗ t^* \le_{\gamma_N+\varepsilon_{9.26}}s^* t ∗ ≤ γ N + ε 9.26 s ∗ , we obtain s ≤ γ N + ε 9.26 s ∗ s \le_{\gamma_N+\varepsilon_{9.26}}s^* s ≤ γ N + ε 9.26 s ∗ . Hence s s s has breadth ≤ γ N + ε 9.26 \le\gamma_N+\varepsilon_{9.26} ≤ γ N + ε 9.26 by Lemma 8.3. But we assumed at the beginning of the proof that s s s has breadth ≥ ε 9.29 > γ N + ε 9.26 \ge\varepsilon_{9.29}>\gamma_N+\varepsilon_{9.26} ≥ ε 9.29 > γ N + ε 9.26 , a contradiction. □
The torso structure theorem
Now, we are able to prove the main results for this section (Theorem 9.31 and Theorem 9.32 ). This can be done by combining the key lemmas from the previous sections.
Lemma 9.30. Let N N N be a nested symmetric set of separations of a graph G G G . Let σ \sigma σ be a splitting star of N N N . Let s s s be a k k k -separation of G G G that ε \varepsilon ε -interlaces σ \sigma σ , but such that s s s is not ε ′ \varepsilon' ε ′ -close to any separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ . Assume that ε ′ ≥ ε \varepsilon' \ge\varepsilon ε ′ ≥ ε and that s s s has breadth at least β 9.30 ( ε , ε ′ ) : = ε ′ + ε + 1 \beta_{9.30}(\varepsilon,\varepsilon'):=\varepsilon'+\varepsilon+1 β 9.30 ( ε , ε ′ ) := ε ′ + ε + 1 . Then s s s is not ε ′ \varepsilon' ε ′ -close to any separation in N N N either.
Proof. Assume for a contradiction that s s s is ε ′ \varepsilon' ε ′ -close to some t ∈ N t \in N t ∈ N . Since σ \sigma σ is a splitting star of N N N , there is r ∈ σ r \in\sigma r ∈ σ such that t ≤ r t \leq r t ≤ r or t ∗ ≤ r t^* \leq r t ∗ ≤ r . Since s s s ε \varepsilon ε -interlaces σ \sigma σ , we have r < ε s r <_{\varepsilon} s r < ε s or r < ε s ∗ r <_{\varepsilon} s^* r < ε s ∗ . We now consider all four cases, and derive a contradiction in each of them.
Case t ≤ r t \leq r t ≤ r and r < ε s r <_{\varepsilon} s r < ε s . Here we have s ≤ ε ′ t ≤ r < ε s s \leq_{\varepsilon'} t \leq r <_{\varepsilon} s s ≤ ε ′ t ≤ r < ε s , so s s s and r r r are ε ′ \varepsilon' ε ′ -close (using ε ′ ≥ ε \varepsilon' \geq\varepsilon ε ′ ≥ ε ), a contradiction.
Case t ≤ r t \leq r t ≤ r and r < ε s ∗ r <_{\varepsilon} s^* r < ε s ∗ . Here we have s ≤ ε ′ t ≤ r < ε s ∗ s \leq_{\varepsilon'} t \leq r <_{\varepsilon} s^* s ≤ ε ′ t ≤ r < ε s ∗ , so s ≤ ε ′ + ε s ∗ s \leq_{\varepsilon'+\varepsilon} s^* s ≤ ε ′ + ε s ∗ by Lemma 4.9. By Lemma 8.3, s s s has breadth ≤ ε ′ + ε \leq\varepsilon' + \varepsilon ≤ ε ′ + ε . This contradicts that s s s is a β 9.30 \beta_{9.30} β 9.30 -broad separation.
Case t ∗ ≤ r t^* \leq r t ∗ ≤ r and r < ε s r <_{\varepsilon} s r < ε s . From the assumption t ≤ ε ′ s t \leq_{\varepsilon'} s t ≤ ε ′ s we obtain s ∗ ≤ ε ′ t ∗ s^* \leq_{\varepsilon'} t^* s ∗ ≤ ε ′ t ∗ . Then s ∗ ≤ ε ′ t ∗ ≤ r < ε s s^* \leq_{\varepsilon'} t^* \leq r <_{\varepsilon} s s ∗ ≤ ε ′ t ∗ ≤ r < ε s and we derive a contradiction in analogy to the previous case.
Case t ∗ ≤ r t^* \leq r t ∗ ≤ r and r < ε s ∗ r <_{\varepsilon} s^* r < ε s ∗ . We derive a contradiction as in the first case. □
Theorem 9.31. Assume Setting 9.5. Let δ > δ 9.17 \delta> \delta_{9.17} δ > δ 9.17 , δ ′ > δ 9.20 ′ ( … , δ ) \delta' > \delta'_{9.20}(\ldots,\delta) δ ′ > δ 9.20 ′ ( … , δ ) , ζ ≥ ζ 9.22 ( k , δ ′ , ε F ) \zeta\geq\zeta_{9.22}(k,\delta',\varepsilon_F) ζ ≥ ζ 9.22 ( k , δ ′ , ε F ) , β 9.31 : = β 9.30 ( ε N , ε 9.22 ) \beta_{9.31} := \beta_{9.30}(\varepsilon_N,\varepsilon_{9.22}) β 9.31 := β 9.30 ( ε N , ε 9.22 ) , and q 9.31 ′ : = max { q , ε 9.22 + k , β 9.31 } q'_{9.31} := \max\{q,\varepsilon_{9.22}+k,\beta_{9.31}\} q 9.31 ′ := max { q , ε 9.22 + k , β 9.31 } . Let σ \sigma σ be a splitting star of N N N . Then we have one of the following outcomes:
(9.31.1) There is a unique A ∈ A A \in\mathcal{A} A ∈ A of spread ≥ ζ \geq\zeta ≥ ζ such that A δ A^\delta A δ exists and ε N \varepsilon_N ε N -interlaces σ \sigma σ . The graphic-k k k -flower F A F_A F A has spread ≥ ζ \geq\zeta ≥ ζ . The splitting star σ \sigma σ and the petal-star of F A F_A F A ε 9.29 \varepsilon_{9.29} ε 9.29 -approximate each other.
(9.31.2) The torso of σ \sigma σ is ( q 9.31 ′ , k ) (q'_{9.31},k) ( q 9.31 ′ , k ) -unbreakable.
Theorem 9.32. Assume the premises of Theorem 9.31. Let S ( σ ) S(\sigma) S ( σ ) be the set of all β 9.31 \beta_{9.31} β 9.31 -broad k k k -separations of G G G that ε N \varepsilon_N ε N -interlace σ \sigma σ .
(i) If some s ∈ S ( σ ) s \in S(\sigma) s ∈ S ( σ ) is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ , then
σ \sigma σ has type (9.31.1);
s ∈ A s \in A s ∈ A and s s s is ε F \varepsilon_F ε F -close to some displayed separation of F A F_A F A ; and
s s s is ζ \zeta ζ -crossed by some s ′ ∈ D ( F A ) ∩ A s' \in D(F_A) \cap A s ′ ∈ D ( F A ) ∩ A .
(ii) Otherwise, σ \sigma σ has type (9.31.2).
Proof of Theorems 9.31 and 9.32. (i). Assume first that there is s ∈ S ( σ ) s \in S(\sigma) s ∈ S ( σ ) that is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ . Observe that ε 9.22 ≥ ε N \varepsilon_{9.22} \geq\varepsilon_N ε 9.22 ≥ ε N . Since σ \sigma σ is a splitting star of N N N , Lemma 9.30 yields that s s s is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in N N N . Then we are done by Key Lemma 9.22, Key Lemma 9.26 and Key Lemma 9.29.
(ii). Assume otherwise that every separation in S ( σ ) S(\sigma) S ( σ ) is ε 9.22 \varepsilon_{9.22} ε 9.22 -close to some separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ . Note that q 9.31 ′ ≥ ε 9.22 + k q'_{9.31} \geq\varepsilon_{9.22} + k q 9.31 ′ ≥ ε 9.22 + k . To show that the torso of σ \sigma σ is ( q 9.31 ′ , k ) (q'_{9.31},k) ( q 9.31 ′ , k ) -unbreakable, by Lemma 9.24 it suffices to show that there is no ( ≤ k ) (\leq k) ( ≤ k ) -separation s s s of G G G that almost-interlaces σ \sigma σ , has breadth > q 9.31 ′ >q'_{9.31} > q 9.31 ′ , and is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ . Assume for a contradiction that there is such a ( ≤ k ) (\leq k) ( ≤ k ) -separation s s s .
Since s s s has breadth > q 9.31 ′ ≥ q >q'_{9.31} \geq q > q 9.31 ′ ≥ q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, it follows that s s s has order k k k . The separation s s s has breadth ≥ β 9.31 \geq\beta_{9.31} ≥ β 9.31 as q 9.31 ′ ≥ β 9.31 q'_{9.31} \geq\beta_{9.31} q 9.31 ′ ≥ β 9.31 .
We claim that s s s ε N \varepsilon_N ε N -interlaces σ \sigma σ . For this, let r ∈ σ r \in\sigma r ∈ σ . As s s s almost-interlaces σ \sigma σ by assumption, we have r ≤ s r \leq s r ≤ s , say. Hence r ≤ ε N s r \leq_{\varepsilon_N} s r ≤ ε N s . We have to show that s ̸ ≤ ε N r s \not\leq_{\varepsilon_N} r s ≤ ε N r . If we had s ≤ ε N r s \leq_{\varepsilon_N} r s ≤ ε N r , then s s s would be ε N \varepsilon_N ε N -close to r ∈ σ r \in\sigma r ∈ σ . As ε N ≤ ε 9.22 \varepsilon_N \le\varepsilon_{9.22} ε N ≤ ε 9.22 , this would contradict our assumption that s s s is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ . Hence s s s ε N \varepsilon_N ε N -interlaces σ \sigma σ .
We have s ∈ S ( σ ) s \in S(\sigma) s ∈ S ( σ ) by the above claims. We chose s s s so that additionally s s s is not ε 9.22 \varepsilon_{9.22} ε 9.22 -close to any separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ . This now contradicts the assumption of (ii). Hence the torso of σ \sigma σ is ( q 9.31 ′ , k ) (q'_{9.31},k) ( q 9.31 ′ , k ) -unbreakable after all. □
Decompositions of narrow graphs
In this section, we prove Theorem 1.2 for narrow graphs, see Theorem 10.7 . To do so, in § 11.1 we first establish the corresponding Tutte-type statement (Theorem 2.6 ) for narrow graphs, see Theorem 10.4 . In § 11.2 we then refine the resulting tree-decomposition using the corresponding flowers to obtain the desired cactus-decomposition.
An approximate Tutte-type decomposition for narrow graphs
Before proving Theorem 10.4 , we just need one last result.
Lemma 10.1. Let q , k ∈ N q, k \in\mathbb{N} q , k ∈ N and let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let s s s be a narrow proper k k k -separation of G G G that is γ \gamma γ -nested with every narrow proper k k k -separation of G G G . Assume that γ ≥ q + k \gamma\ge q+k γ ≥ q + k . Then s s s is narrow-γ \gamma γ -nested.
Proof. Let r r r be an arbitrary narrow proper ( < k ) (<k) ( < k ) -separation of G G G . We have to show that s s s is γ \gamma γ -nested with r r r . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, r r r has breadth ≤ q \le q ≤ q . So some corner of r r r and s s s has size ≤ q \le q ≤ q . In particular, r r r and s s s do not q q q -criss. Hence r r r and s s s are ( q + k ) (q+k) ( q + k ) -nested by the contrapositive of Observation 5.4 (ii). □
:::
Definition 10.2 (ψ \psi ψ -isomorphic for sets of graph-decompositions). Let G G G be a graph, let G 1 , G 2 \mathcal{G}_1,\mathcal{G}_2 G 1 , G 2 be two sets of graph-decompositions of G G G and let ψ \psi ψ be an automorphism of G G G . We say that G 1 \mathcal{G}_1 G 1 and G 2 \mathcal{G}_2 G 2 are ψ \psi ψ -isomorphic if there exists a bijection Ψ : G 1 → G 2 \Psi:\mathcal{G}_1 \to\mathcal{G}_2 Ψ : G 1 → G 2 such that H \mathcal{H} H and Ψ ( H ) \Psi(\mathcal{H}) Ψ ( H ) are ψ \psi ψ -isomorphic (as in Definition 6.5) for every H ∈ G 1 \mathcal{H} \in\mathcal{G}_1 H ∈ G 1 .
Definition 10.3 (Canonical for sets of graph-decompositions). Let G G G be a graph, and let G \mathcal{G} G be a set of graph-decompositions of G G G . We say that G \mathcal{G} G is canonical if, for every automorphism ψ \psi ψ of G G G , G \mathcal{G} G is ψ \psi ψ -isomorphic to itself, i.e. there is a permutation Ψ \Psi Ψ of G \mathcal{G} G such that H \mathcal{H} H and Ψ ( H ) \Psi(\mathcal{H}) Ψ ( H ) are ψ \psi ψ -isomorphic (as in Definition 6.5) for every H ∈ G \mathcal{H} \in\mathcal{G} H ∈ G .
We can now prove Theorem 2.6 under the additional assumption that the graph to be decomposed is ( b , k ) (b,k) ( b , k ) -narrow. For a tree-decomposition T = ( T , V ) \mathcal{T}=(T,\mathcal{V}) T = ( T , V ) and a node t t t of T T T , the t t t -star is the set of all displayed separations of T \mathcal{T} T that are induced by cuts ( X , Y ) (X,Y) ( X , Y ) of T T T such that there is a unique X X X –Y Y Y edge in T T T , which is additionally incident to t ∈ Y t \in Y t ∈ Y . We say that a graphic-k k k -flower F F F ε \varepsilon ε -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t if the petal-star of F F F and the t t t -star of ( T , V ) (T,\mathcal{V}) ( T , V ) ε \varepsilon ε -approximate each other.
Theorem 10.4. For every q , k , b ∈ N q,k,b \in\mathbb{N} q , k , b ∈ N there are γ , κ , ℓ , ℓ ‾ ∈ N \gamma,\kappa,\ell,\overline{\ell} \in\mathbb{N} γ , κ , ℓ , ℓ ∈ N such that for all ζ ∈ N \zeta\in\mathbb{N} ζ ∈ N there are q ′ , κ ∈ N q',\kappa\in\mathbb{N} q ′ , κ ∈ N , such that the following holds. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph. There exists a set F \mathcal{F} F of graphic-k k k -flowers of spread ≥ ζ \ge\zeta ≥ ζ and a symmetric nested set N N N of narrow-γ \gamma γ -nested proper k k k -separations of G G G with corresponding tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) of G G G and a bipartition X ⊔ Y = V ( T ) X \sqcup Y = V(T) X ⊔ Y = V ( T ) such that:
(i) Every node t ∈ X t \in X t ∈ X has a ( q ′ , k ) (q', k) ( q ′ , k ) -unbreakable torso.
(ii) Every node t ∈ Y t \in Y t ∈ Y comes with a graphic-k k k -flower F t ∈ F F_t \in\mathcal{F} F t ∈ F of G G G such that F t F_t F t ϱ \varrho ϱ -refines ( T , V ) (T, \mathcal{V}) ( T , V ) at t t t . Furthermore, if F t F_t F t is a graphic-k k k -daisy, then F t F_t F t has the clover-property.
(iii) Every displayed separation of some F t F_t F t (for t ∈ Y t \in Y t ∈ Y ) that is ϱ \varrho ϱ -crossed by another displayed separation of F t F_t F t almost-ϱ \varrho ϱ -interlaces the splitting star of ( T , V ) (T, \mathcal{V}) ( T , V ) corresponding to t t t .
(iv) Every ε \varepsilon ε -broad k k k -separation of G G G either is ε \varepsilon ε -close to a separation in N N N or is ε \varepsilon ε -close to some displayed separation of some F t F_t F t .
(v) F \mathcal{F} F is canonical. Both the tree-decomposition ( T , V ) (T, \mathcal{V}) ( T , V ) and the set N N N are κ \kappa κ -canonical. Given N N N and ( T , V ) (T, \mathcal{V}) ( T , V ) , the bipartition X ⊔ Y = V ( T ) X \sqcup Y = V(T) X ⊔ Y = V ( T ) is canonical and the assignment Y ∋ t ↦ F t ∈ F Y \ni t \mapsto F_t \in\mathcal{F} Y ∋ t ↦ F t ∈ F is a canonical bijection.
(vi) D ( F ) ≠ D ( F ′ ) D(F) \ne D(F') D ( F ) = D ( F ′ ) for every two distinct graphic-k k k -flowers F , F ′ ∈ F F, F' \in\mathcal{F} F , F ′ ∈ F .
(vii) For every ε ˉ \bar{\varepsilon} ε ˉ -broad k k k -separation of G G G there is an ε ˉ \bar{\varepsilon} ε ˉ -close k k k -separation of G G G nested with N N N .
Proof. Let q q q , k k k , b b b , ζ ∈ N \zeta\in\mathbb{N} ζ ∈ N be given, and let G G G be a ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b, k) ( b , k ) -narrow graph. We first establish Setting 9.5. To this end, let
β : = max { β 6.6 ( q , k ) , b , q + 1 } ≥ 2 q + k \beta:= \max\{ \beta_{6.6}(q,k),\ b,\ q+1 \} \ge2q+k β := max { β 6.6 ( q , k ) , b , q + 1 } ≥ 2 q + k
ε F : = ε 6.6 ( q , k ) \varepsilon_{\mathcal{F}} := \varepsilon_{6.6}(q,k) ε F := ε 6.6 ( q , k )
γ A : = max { γ 5.1 ( q , k ) , γ 6.6 ( q , k ) + 1 , γ 7.1 ( q , k , ε F ) , β } . \gamma_A := \max\{ \gamma_{5.1}(q,k),\ \gamma_{6.6}(q,k)+1,\ \gamma_{7.1}(q,k,\varepsilon_{\mathcal{F}}),\ \beta\}. γ A := max { γ 5.1 ( q , k ) , γ 6.6 ( q , k ) + 1 , γ 7.1 ( q , k , ε F ) , β } .
The choice of β \beta β and γ A \gamma_A γ A establishes the first two points of (9.5.1). Let S 0 S_0 S 0 be the set of all β \beta β -broad k k k -separations of G G G . Since β ≥ b \beta\ge b β ≥ b and G G G is ( b , k ) (b,k) ( b , k ) -narrow, all separations in S 0 S_0 S 0 are narrow. By Key Lemma 5.1 and using β ≥ q \beta\ge q β ≥ q as well as γ A ≥ γ 5.1 \gamma_A \ge\gamma_{5.1} γ A ≥ γ 5.1 , we find that the set S 0 S_0 S 0 is uniformly γ A \gamma_A γ A -crossing.
Flowers. Let A t \mathcal{A}_t A t and A \mathcal{A} A consist of all trivial and nontrivial γ A \gamma_A γ A -asterisks of S 0 S_0 S 0 , respectively. For each A ∈ A A \in\mathcal{A} A ∈ A , let F A : = F 6.6 ( A ) F_A := F_{6.6}(A) F A := F 6.6 ( A ) be the q q q -lean graphic-k k k -flower found by Key Lemma 6.6. By Key Lemma 6.6 (iv), these graphic-k k k -flowers are canonical. When the common centre X X X of A A A has size k k k , we are in a special case that we considered in § 7.2 . In that special case, we defined F 6.6 ( A ) = F A F_{6.6}(A) = F_A F 6.6 ( A ) = F A to be the graphic-X X X -anemone, which establishes (9.5.2). In general, every separation in A A A is ε F \varepsilon_{\mathcal{F}} ε F -close to some displayed separation of F A F_A F A , by Key Lemma 6.6 (i) and Observation 4.11 , which establishes (9.5.3). We will later find F \mathcal{F} F as a subset of { F A : A ∈ A } \{F_A : A \in\mathcal{A}\} { F A : A ∈ A } .
Petal-separations. Next, we let
γ π : = max { γ 7.1 ′ ( q , k ) , γ A , ε 6.6 ( q , k ) , β + k } \gamma_{\pi} := \max\{ \gamma'_{7.1}(q,k),\ \gamma_A,\ \varepsilon_{6.6}(q,k),\ \beta+k \} γ π := max { γ 7.1 ′ ( q , k ) , γ A , ε 6.6 ( q , k ) , β + k }
γ N : = γ 8.1 ′ ( k , b , γ π ) ≥ γ π ≥ β + k ≥ q + k . \gamma_N := \gamma'_{8.1}(k,b,\gamma_{\pi}) \ge\gamma_{\pi} \ge\beta+k \ge q+k. γ N := γ 8.1 ′ ( k , b , γ π ) ≥ γ π ≥ β + k ≥ q + k .
The choice of γ N \gamma_N γ N ensures that it meets the requirements of (9.5.1). Let A ∈ A A \in\mathcal{A} A ∈ A be arbitrary. Then, A A A is a nontrivial γ A \gamma_A γ A -asteriskal set of k k k -separations of G G G , and F A F_A F A is a q q q -lean graphic-k k k -flower of G G G that ε F \varepsilon_{\mathcal{F}} ε F -approximates A A A by Key Lemma 6.6 (i). By Key Lemma 7.1, every k k k -separation of G G G is γ π \gamma_{\pi} γ π -compatible with F A F_A F A . By Key Lemma 6.6 (ii), since γ π ≥ ε 6.6 ( q , k ) \gamma_{\pi} \ge\varepsilon_{6.6}(q,k) γ π ≥ ε 6.6 ( q , k ) , we obtain that
every k -separation of G is γ π -nested with all the petal-separations of F A , (11) \text{every } k\text{-separation of } G \text{ is } \gamma_{\pi}\text{-nested with all the petal-separations of } F_A\text{,} \tag*{(11)} every k -separation of G is γ π -nested with all the petal-separations of F A , (11)
which establishes (9.5.4) since γ N ≥ γ π \gamma_N \ge\gamma_\pi γ N ≥ γ π .
Nested sets of separations. Let M M M be the set of all narrow-γ π \gamma_\pi γ π -nested k k k -separations of G G G . Let M ′ M' M ′ be the set of all half-connected narrow-γ N \gamma_N γ N -nested β M \beta_M β M -broad k k k -separations of G G G that are ϵ M \epsilon_M ϵ M -close to some separation in M M M for
β M : = β 8.1 ( q , k , b ) \beta_M := \beta_{8.1}(q,k,b) β M := β 8.1 ( q , k , b )
ϵ M : = δ 8.1 ( k , b , γ π ) . \epsilon_M := \delta_{8.1}(k,b,\gamma_\pi). ϵ M := δ 8.1 ( k , b , γ π ) .
Let N ′ ⊆ M ′ N' \subseteq M' N ′ ⊆ M ′ be a maximal nested subset. Let S S S be the set of all narrow proper k k k -separations of G G G . Note that N ′ ⊆ S N' \subseteq S N ′ ⊆ S , as every separation in N ′ N' N ′ is β M \beta_M β M -broad where β M ≥ b \beta_M \ge b β M ≥ b and G G G is ( b , k ) (b,k) ( b , k ) -narrow. We obtain N N N from N ′ N' N ′ by adding all the S S S -totally-nested k k k -separations of G G G . Then N N N is a nested symmetric set of proper k k k -separations of G G G and satisfies (9.5.6) by definition.
Every separation in N N N is either contained in N ′ ⊆ M ′ N' \subseteq M' N ′ ⊆ M ′ and so is narrow-γ N \gamma_N γ N -nested by definition of M ′ M' M ′ , or is S S S -totally-nested and so is narrow-γ N \gamma_N γ N -nested by Lemma 10.1 using γ N ≥ q + k \gamma_N \ge q+k γ N ≥ q + k . Hence all separations in N N N are narrow-γ N \gamma_N γ N -nested, so N N N satisfies (9.5.5).
By Key Lemma 8.1, N ′ N' N ′ ϵ N \epsilon_N ϵ N -approximates M M M for
ϵ N : = max { ϵ 8.1 ( q , k , b , γ π ) , γ N } ≥ q + 1 , b , γ π . \epsilon_N := \max\{ \epsilon_{8.1}(q,k,b,\gamma_\pi),\gamma_N \} \ge q+1,b,\gamma_\pi. ϵ N := max { ϵ 8.1 ( q , k , b , γ π ) , γ N } ≥ q + 1 , b , γ π .
Note that ϵ N ≥ γ N \epsilon_N \ge\gamma_N ϵ N ≥ γ N , so (9.5.1) is fully established now.
Let S 1 : = ⋃ { π ( F A ) ∪ ( π ( F A ) ) ∗ : A ∈ A } ∪ A t S_1 := \bigcup\{ \pi(F_A) \cup(\pi(F_A))^* : A \in\mathcal{A} \} \cup\mathcal{A}_t S 1 := ⋃ { π ( F A ) ∪ ( π ( F A ) ) ∗ : A ∈ A } ∪ A t .
Claim 10.4.1. N N N ϵ N \epsilon_N ϵ N -approximates S 1 S_1 S 1 .
Proof of the claim. For this, let s ∈ S 1 s \in S_1 s ∈ S 1 be ϵ N \epsilon_N ϵ N -broad. Since ϵ N > q \epsilon_N > q ϵ N > q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, s s s has order k k k . Since G G G is ( b , k ) (b,k) ( b , k ) -narrow and ϵ N ≥ b \epsilon_N \ge b ϵ N ≥ b , we have that s s s is narrow.
Assume first that s ∈ π ( F A ) ∪ ( π ( F A ) ) ∗ s \in\pi(F_A) \cup(\pi(F_A))^* s ∈ π ( F A ) ∪ ( π ( F A ) ) ∗ for some A ∈ A A \in\mathcal{A} A ∈ A . By (9), we have that s s s is γ π \gamma_\pi γ π -nested with every (narrow proper) k k k -separation of G G G . Hence s ∈ M s \in M s ∈ M by Lemma 10.1 and γ π ≥ q + k \gamma_\pi\ge q+k γ π ≥ q + k . Then some separation in N ′ ⊆ N N' \subseteq N N ′ ⊆ N is ϵ N \epsilon_N ϵ N -close to s s s .
We assume next that s ∈ A s \in A s ∈ A for some trivial A ∈ A t A \in\mathcal{A}_t A ∈ A t . We show that s ∈ M s \in M s ∈ M . If not, s s s is γ π \gamma_\pi γ π -crossed by some narrow proper ( ≤ k ) (\le k) ( ≤ k ) -separation r r r of G G G . As γ π ≥ γ A \gamma_\pi\ge\gamma_A γ π ≥ γ A and A A A is trivial, we cannot have r ∈ S 0 r \in S_0 r ∈ S 0 . To obtain a contradiction, we will show that r ∈ S 0 r \in S_0 r ∈ S 0 . Since s s s and r r r γ π \gamma_\pi γ π -cross, Lemma 5.6 yields that r r r has breadth ≥ 2 ( γ π − k ) ≥ β \ge2(\gamma_\pi-k) \ge\beta ≥ 2 ( γ π − k ) ≥ β . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable with β > q \beta> q β > q , r r r has order k k k , so r ∈ S 0 r \in S_0 r ∈ S 0 , a contradiction. Thus, s ∈ M s \in M s ∈ M , and as above some separation in N ′ ⊆ N N' \subseteq N N ′ ⊆ N is ϵ N \epsilon_N ϵ N -close to s s s . ♢
Claim 10.4.1 establishes (9.5.7). We now have completely established Setting 9.5.
Tree-decomposition. In order to apply Theorem 9.31 , we let
δ : = δ 9.17 ( q , k , γ A , ε F , γ N , ε N ) + 1 , \delta:= \delta_{9.17}(q,k,\gamma_{\mathcal{A}},\varepsilon_{\mathcal{F}},\gamma_N,\varepsilon_N)+1, δ := δ 9.17 ( q , k , γ A , ε F , γ N , ε N ) + 1 ,
δ ′ : = δ 9.20 ′ ( q , k , b , γ N , δ ) + 1 , \delta' := \delta'_{9.20}(q,k,b,\gamma_N,\delta)+1, δ ′ := δ 9.20 ′ ( q , k , b , γ N , δ ) + 1 ,
ϱ : = ε 9.29 ( q , k , b , γ N , ε N , δ ′ ) , \varrho:= \varepsilon_{9.29}(q,k,b,\gamma_N,\varepsilon_N,\delta'), ϱ := ε 9.29 ( q , k , b , γ N , ε N , δ ′ ) ,
ζ ′ : = max { ζ , ζ 9.22 ( k , δ ′ , ε F ) } , \zeta' := \max\{\zeta,\zeta_{9.22}(k,\delta',\varepsilon_{\mathcal{F}})\}, ζ ′ := max { ζ , ζ 9.22 ( k , δ ′ , ε F )} ,
q ′ : = q 9.31 ′ ( q , k , β , ε F , ε N , δ ′ , ζ ′ ) , q' := q'_{9.31}(q,k,\beta,\varepsilon_{\mathcal{F}},\varepsilon_N,\delta',\zeta'), q ′ := q 9.31 ′ ( q , k , β , ε F , ε N , δ ′ , ζ ′ ) ,
ε ′ : = ε 9.22 ( k , ε F , ε N , ζ ′ ) ≥ ε F , ε N , \varepsilon' := \varepsilon_{9.22}(k,\varepsilon_{\mathcal{F}},\varepsilon_N,\zeta') \ge\varepsilon_{\mathcal{F}},\varepsilon_N, ε ′ := ε 9.22 ( k , ε F , ε N , ζ ′ ) ≥ ε F , ε N ,
ε : = max { β 9.31 ( ε N , ε ′ ) , 2 γ N + 1 } ≥ ε F , ε N , ε ′ . \varepsilon:= \max\{\beta_{9.31}(\varepsilon_N,\varepsilon'),2\gamma_N+1\} \ge\varepsilon_{\mathcal{F}},\varepsilon_N,\varepsilon'. ε := max { β 9.31 ( ε N , ε ′ ) , 2 γ N + 1 } ≥ ε F , ε N , ε ′ .
Let ( T , V ) (T,\mathcal{V}) ( T , V ) be the tree-decomposition of G G G that is uniquely determined by N N N , see Lemma 6.26 . Setting γ : = γ N \gamma:=\gamma_N γ := γ N , we have that the separations in N N N are indeed narrow-γ \gamma γ -nested. Let σ ↦ t ( σ ) \sigma\mapsto t(\sigma) σ ↦ t ( σ ) be the natural bijection between the set of all splitting stars of N N N and V ( T ) V(T) V ( T ) .
Let Y N Y_N Y N be the set of all splitting stars of N N N that have type (9.31.1) for the parameter ζ ′ \zeta' ζ ′ , and let X N X_N X N be the set of all splitting stars of N N N that are not in Y N Y_N Y N . Let X , Y X,Y X , Y be the images of X N , Y N X_N,Y_N X N , Y N under the map σ ↦ t ( σ ) \sigma\mapsto t(\sigma) σ ↦ t ( σ ) . Given N N N and ( T , V ) (T,\mathcal{V}) ( T , V ) , the bipartition X ⊔ Y = V ( T ) X\sqcup Y=V(T) X ⊔ Y = V ( T ) is canonical.
Definition of F . \mathcal{F}. F . Let F \mathcal{F} F consist of all graphic-k k k -flowers F A F_A F A such that both A A A and F A F_A F A have spread ≥ ζ ′ \ge\zeta' ≥ ζ ′ .
Verification. First, observe that all the parameters are independent of ζ \zeta ζ , except ζ ′ \zeta' ζ ′ , q ′ q' q ′ , ε ′ \varepsilon' ε ′ and ε \varepsilon ε .
(i). For every t = t ( σ ) ∈ X t=t(\sigma)\in X t = t ( σ ) ∈ X , the torso of σ \sigma σ (which coincides with the torso of V t V_t V t by Lemma 6.26 ) is ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable by (9.31.2).
(ii). For each t = t ( σ ) ∈ Y t=t(\sigma)\in Y t = t ( σ ) ∈ Y , set F t : = F A F_t:=F_A F t := F A for the unique A ∈ A A\in\mathcal{A} A ∈ A provided for σ \sigma σ by (9.31.1). So, in particular, F t F_t F t has spread ≥ ζ ′ ≥ ζ \ge\zeta'\ge\zeta ≥ ζ ′ ≥ ζ , and F t ∈ F F_t\in\mathcal{F} F t ∈ F . Then F t F_t F t ϱ \varrho ϱ -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t by (9.31.1). F t F_t F t was constructed using Key Lemma 6.6, so the “Furthermore” part follows from Key Lemma 6.6 (iii).
(iii). Let t = t ( σ ) ∈ Y t=t(\sigma)\in Y t = t ( σ ) ∈ Y , let A ∈ A A\in\mathcal{A} A ∈ A be the asterisk provided for σ \sigma σ by (9.31.1). Recall that F t = F A F_t=F_A F t = F A . Note that
ϱ = ε 9.29 ( q , k , b , γ N , ε N , δ ′ ) ≥ ε 9.26 ( k , ε N , δ ′ ) = 2 ( k + ε N + δ ′ ) ≥ δ ′ , ε N . \varrho=\varepsilon_{9.29}(q,k,b,\gamma_N,\varepsilon_N,\delta')\ge\varepsilon_{9.26}(k,\varepsilon_N,\delta')=2(k+\varepsilon_N+\delta')\ge\delta',\varepsilon_N. ϱ = ε 9.29 ( q , k , b , γ N , ε N , δ ′ ) ≥ ε 9.26 ( k , ε N , δ ′ ) = 2 ( k + ε N + δ ′ ) ≥ δ ′ , ε N .
If s ∈ D ( F t ) s\in D(F_t) s ∈ D ( F t ) is ϱ \varrho ϱ -crossed by another separation in D ( F t ) D(F_t) D ( F t ) , then s ∈ ⋃ A ( D ( F A ) , δ ′ ) ⊆ A δ s\in\bigcup\mathcal{A}(D(F_A),\delta')\subseteq A^\delta s ∈ ⋃ A ( D ( F A ) , δ ′ ) ⊆ A δ by Lemma 9.9 and ϱ ≥ δ ′ \varrho\ge\delta' ϱ ≥ δ ′ . Since A δ ε N A^\delta\varepsilon_N A δ ε N -interlaces σ \sigma σ by (9.31.1), it follows that s s s almost-ϱ \varrho ϱ -interlaces σ \sigma σ .
(iv). For this, let s s s be an arbitrary ε \varepsilon ε -broad k k k -separation of G G G . In order to apply Lemma 8.6 , we need to show two things: that s s s is ( 2 γ N + 1 ) (2\gamma_N+1) ( 2 γ N + 1 ) -broad and that s s s is γ N \gamma_N γ N -nested with every separation in N N N . The former is immediate since s s s is ε \varepsilon ε -broad with ε ≥ 2 γ N + 1 \varepsilon\ge2\gamma_N+1 ε ≥ 2 γ N + 1 . For the latter, it suffices to show that s s s is narrow, as the separations in N N N are narrow-γ N \gamma_N γ N -nested. To show that s s s is narrow, it suffices to show that ε ≥ b \varepsilon\ge b ε ≥ b as G G G is ( b , k ) (b,k) ( b , k ) -narrow. And indeed, ε ≥ ε N ≥ b \varepsilon\ge\varepsilon_N\ge b ε ≥ ε N ≥ b .
By Lemma 8.6 , s s s either is γ N \gamma_N γ N -close to some separation in N N N (in which case we are done as γ N ≤ ε \gamma_N\le\varepsilon γ N ≤ ε ), or there is a unique splitting star σ \sigma σ of N N N such that s s s γ N \gamma_N γ N -interlaces σ \sigma σ (which we assume from now on). As ε N ≥ γ N \varepsilon_N\ge\gamma_N ε N ≥ γ N , we in particular have that s s s ε N \varepsilon_N ε N -interlaces σ \sigma σ (unless s s s is ε N \varepsilon_N ε N -close to some separation in σ ∪ σ ∗ ⊆ N \sigma\cup\sigma^*\subseteq N σ ∪ σ ∗ ⊆ N , in which case we are done as ε N ≤ ε \varepsilon_N\le\varepsilon ε N ≤ ε ). So we have s ∈ S ( σ ) s\in S(\sigma) s ∈ S ( σ ) with S ( σ ) S(\sigma) S ( σ ) as in Theorem 9.32 . If s s s is ε ′ \varepsilon' ε ′ -close to some separation in σ ∪ σ ∗ \sigma\cup\sigma^* σ ∪ σ ∗ , then we are done as σ ∪ σ ∗ ⊆ N \sigma\cup\sigma^*\subseteq N σ ∪ σ ∗ ⊆ N and ε ′ ≤ ε \varepsilon' \le\varepsilon ε ′ ≤ ε . Otherwise, Theorem 9.32 (i) applies (with ζ ′ \zeta' ζ ′ in place of ζ \zeta ζ ), by which s s s is ε F \varepsilon_F ε F -close (hence ε \varepsilon ε -close) to some displayed separation of F t F_t F t for t = t ( σ ) t=t(\sigma) t = t ( σ ) .
(v). To show that F \mathcal{F} F is canonical, let ψ \psi ψ be any automorphism of G G G . Then ψ \psi ψ induces a permutation of A \mathcal{A} A . By definition, for each F ∈ F F \in\mathcal{F} F ∈ F , there exists A ∈ A A \in\mathcal{A} A ∈ A of spread ≥ ζ ′ \ge\zeta' ≥ ζ ′ such that F = F A F=F_A F = F A . This A A A is unique by Corollary 9.13. Let Ψ \Psi Ψ be the permutation of F \mathcal{F} F that maps each F A F_A F A to Ψ ( F A ) : = F ψ ( A ) \Psi(F_A):=F_{\psi(A)} Ψ ( F A ) := F ψ ( A ) . Then F A F_A F A is ψ \psi ψ -isomorphic to Ψ ( F A ) \Psi(F_A) Ψ ( F A ) , as the graphic-k k k -flowers in F \mathcal{F} F are canonical by Key Lemma 6.6 (iv).
Given N N N and ( T , V ) (T,\mathcal{V}) ( T , V ) , we already argued that the bipartition X \mathbin{\mathaccent\cdot\cup} Y=V(T) is canonical. Moreover, the assignment Y ∋ t ↦ F t ∈ F Y\ni t\mapsto F_t\in\mathcal{F} Y ∋ t ↦ F t ∈ F is defined avoiding arbitrary decisions, hence it is also canonical. The assignment is injective by (9.31.1) and Proposition 9.20. To see surjectivity, let F A ∈ F F_A\in\mathcal{F} F A ∈ F . Recall that, by definition of F \mathcal{F} F , the asterisk A A A has spread ≥ ζ ′ ≥ ζ 9.22 ( k , δ ′ , ε F ) = ζ 9.12 ( δ ′ + 2 k + 1 , ε F ) = δ ′ + 2 k + 1 + 2 ε F ≥ ζ 9.12 ( δ ′ , ε F ) \ge\zeta' \ge\zeta_{9.22}(k,\delta',\varepsilon_F)=\zeta_{9.12}(\delta'+2k+1,\varepsilon_F)=\delta'+2k+1+2\varepsilon_F\ge\zeta_{9.12}(\delta',\varepsilon_F) ≥ ζ ′ ≥ ζ 9.22 ( k , δ ′ , ε F ) = ζ 9.12 ( δ ′ + 2 k + 1 , ε F ) = δ ′ + 2 k + 1 + 2 ε F ≥ ζ 9.12 ( δ ′ , ε F ) . So by Proposition 9.16, A δ A^\delta A δ exists and ε N \varepsilon_N ε N -interlaces a unique splitting star σ \sigma σ of N N N . By Key Lemma 9.26 and Key Lemma 9.29, σ \sigma σ and π ( F A ) \pi(F_A) π ( F A ) ϱ \varrho ϱ -approximate each other. By definition of F \mathcal{F} F , both A A A and F A F_A F A have spread ≥ ζ ′ \ge\zeta' ≥ ζ ′ , and Proposition 9.20 ensures that no other B ∈ A B\in\mathcal{A} B ∈ A of spread ≥ ζ ′ \ge\zeta' ≥ ζ ′ has B δ B^\delta B δ ε N \varepsilon_N ε N -interlacing σ \sigma σ . Together with the ϱ \varrho ϱ -approximation proved above, this shows that σ \sigma σ has type (9.31.1) for the parameter ζ ′ \zeta' ζ ′ . Thus σ ∈ Y N \sigma\in Y_N σ ∈ Y N and F t ( σ ) = F A F_{t(\sigma)}=F_A F t ( σ ) = F A . It follows that the assignment Y ∋ t ↦ F t ∈ F Y\ni t\mapsto F_t\in\mathcal{F} Y ∋ t ↦ F t ∈ F is surjective, and hence bijective.
Let κ : = 2 ε N \kappa:=2\varepsilon_N κ := 2 ε N . By Corollary 8.2 we have that N ′ N' N ′ is κ \kappa κ -canonical, and S \mathcal{S} S is canonical so the set of all S \mathcal{S} S -totally-nested k k k -separations of G G G too. Hence N N N is κ \kappa κ -canonical, so ( T , V ) (T,\mathcal{V}) ( T , V ) as well by definition.
(vi). This is immediate from Corollary 9.13.
(vii). We let ε ˉ : = max { ε 8.1 ( q , k , b , γ π ) , δ 8.1 ′ ( k , b , γ π ) + b } \bar{\varepsilon}:=\max\{\varepsilon_{8.1}(q,k,b,\gamma_\pi),\delta'_{8.1}(k,b,\gamma_\pi)+b\} ε ˉ := max { ε 8.1 ( q , k , b , γ π ) , δ 8.1 ′ ( k , b , γ π ) + b } . Let s s s be an ε ˉ \bar{\varepsilon} ε ˉ -broad k k k -separation of G G G . By the ‘moreover’-part of Key Lemma 8.1, there is a δ 8.1 ( k , b , γ π ) \delta_{8.1}(k,b,\gamma_\pi) δ 8.1 ( k , b , γ π ) -close k k k -separation s ′ s' s ′ of G G G that is nested with N ′ N' N ′ . Observe that s ′ s' s ′ is b b b -broad and hence narrow since G G G is ( b , k ) (b,k) ( b , k ) -narrow. So, s ′ s' s ′ is nested with every narrow-nested k k k -separation of G G G and hence nested with N N N . □
The approximate cactus-decomposition for narrow graphs
To combine the tree-decomposition and flowers from Theorem 10.4 into a cactus-decomposition, we first adjust each flower to the corresponding splitting star. We do this by perturbing its displayed separations to make them nested with the tree-decomposition, completing each attached branch to a clique, and then reconstructing a flower. The next lemmas ensure that these perturbations preserve crossing-connectedness and that each resulting clique is contained in a single bag.
Lemma 10.5. Let G G G be a graph and γ , ε ∈ N \gamma,\varepsilon\in\mathbb{N} γ , ε ∈ N with γ ≥ 2 ε \gamma\ge2\varepsilon γ ≥ 2 ε . Let S \mathcal{S} S be a symmetric γ \gamma γ -crossing-connected set of separations of G G G . For each separation s ∈ S s\in\mathcal{S} s ∈ S let s ′ s' s ′ be an ε \varepsilon ε -close separation of G G G such that ( s ′ ) ∗ = ( s ∗ ) ∗ (s')^*=(s^*)^* ( s ′ ) ∗ = ( s ∗ ) ∗ for every s ∈ S s\in\mathcal{S} s ∈ S . Denote S ′ : = { s ′ : s ∈ S } \mathcal{S}':=\{s':s\in\mathcal{S}\} S ′ := { s ′ : s ∈ S } . Then S ′ \mathcal{S}' S ′ is ( γ − 2 ε ) (\gamma-2\varepsilon) ( γ − 2 ε ) -crossing-connected.
Proof. It suffices to show that s ′ s' s ′ and t ′ t' t ′ ( γ − 2 ε ) (\gamma-2\varepsilon) ( γ − 2 ε ) -cross whenever s s s and t t t γ \gamma γ -cross. This follows from Lemma 5.8 . □
Lemma 10.6. Let G G G be a graph and let F = ( O , V ) F=(O,\mathcal{V}) F = ( O , V ) be a graphic-k k k -diacy of G G G that has the clover-property. Let X : = ⋂ O ∈ V ( O ) V O X:=\bigcap_{O\in V(O)}V_O X := ⋂ O ∈ V ( O ) V O be the centre of F F F . Let K K K be a clique of G G G with K ⊈ X K\not\subseteq X K ⊆ X . Then, the set of vertices x ∈ V ( O ) x\in V(O) x ∈ V ( O ) such that K ⊆ V x K\subseteq V_x K ⊆ V x induces a nonempty subpath of O O O . In particular, there is a vertex x ∈ V ( O ) x\in V(O) x ∈ V ( O ) whose bag V x V_x V x contains K K K .
Proof. Since F F F has the clover-property, F F F has four adhesion-sets V e 0 V_{e_0} V e 0 , V e 1 V_{e_1} V e 1 , V e 2 V_{e_2} V e 2 , V e 3 V_{e_3} V e 3 with V e i ∩ V e j = X V_{e_i} \cap V_{e_j} = X V e i ∩ V e j = X for all i ≠ j i \ne j i = j .
Fix an orientation of O O O , so every edge of O O O has a tail and a head. For every v ∈ V ( G ) v \in V(G) v ∈ V ( G ) , let O v O_v O v be the set of all nodes x ∈ V ( O ) x \in V(O) x ∈ V ( O ) such that v ∈ V x v \in V_x v ∈ V x . Let u ∈ K ∖ X u \in K \setminus X u ∈ K ∖ X . Without loss of generality, we may assume that the edges e 0 e_0 e 0 , e 1 e_1 e 1 , e 2 e_2 e 2 and e 3 e_3 e 3 appear in that cyclic order along O O O . Since u ∉ X u \notin X u ∈ / X we have that u u u belongs to at most one of the sets V e 0 V_{e_0} V e 0 , V e 1 V_{e_1} V e 1 , V e 2 V_{e_2} V e 2 , V e 3 V_{e_3} V e 3 . By symmetry, we may assume that all nodes in O u O_u O u lie between the head of e 0 e_0 e 0 and the tail of e 2 e_2 e 2 (both included) in O O O . Since K K K is complete, any two vertices of K K K must appear in a common bag. Thus, for every vertex v ∈ K ∖ X v \in K \setminus X v ∈ K ∖ X , all nodes in O v O_v O v lie between the head of e 0 e_0 e 0 and the tail of e 2 e_2 e 2 (both included) in O O O , or between the head of e 1 e_1 e 1 and the tail of e 3 e_3 e 3 (both included) in O O O , or between the head of e 3 e_3 e 3 and the tail of e 1 e_1 e 1 (both included) in O O O . Observe that the last two possibilities are incompatible: there do not exist v , v ′ ∈ K ∖ X v,v' \in K \setminus X v , v ′ ∈ K ∖ X such that all nodes in O v O_v O v lie between the head of e 1 e_1 e 1 and the tail of e 3 e_3 e 3 (both included) in O O O and all nodes in O v ′ O_{v'} O v ′ lie between the head of e 3 e_3 e 3 and the tail of e 1 e_1 e 1 (both included) in O O O . By symmetry, say that the last option does not happen for any v ∈ K v \in K v ∈ K . Then, the tail y y y of e 0 e_0 e 0 satisfies V y ∩ ( K ∖ X ) = ∅ V_y \cap(K \setminus X) = \varnothing V y ∩ ( K ∖ X ) = ∅ .
For every vertex u ∈ K ∖ X u \in K \setminus X u ∈ K ∖ X , O u O_u O u induces a connected subgraph of O − y O-y O − y , hence a subpath of O − y O-y O − y , and for all u ≠ v ∈ K ∖ X u \ne v \in K \setminus X u = v ∈ K ∖ X we have O u ∩ O v ≠ ∅ O_u \cap O_v \ne\varnothing O u ∩ O v = ∅ . By the Helly property of subpaths of a path, the set ⋂ v ∈ K ∖ X O v \bigcap_{v \in K \setminus X} O_v ⋂ v ∈ K ∖ X O v induces a nonempty subpath of O − y O-y O − y . Since X = ⋂ o ∈ V ( O ) V o X = \bigcap_{o \in V(O)} V_o X = ⋂ o ∈ V ( O ) V o , it follows that the set of all vertices x ∈ V ( O ) x \in V(O) x ∈ V ( O ) such that K ⊆ V x K \subseteq V_x K ⊆ V x induces a nonempty subpath of O − y O-y O − y . This concludes the proof.
A cactus-decomposition C C C of a graph G G G is clique-faithful if every clique of G G G is contained in a bag of a node of C C C that is not a P P P -node.
Theorem 10.7. For every q , k , b ∈ N q,k,b \in\mathbb{N} q , k , b ∈ N there exists η = η 10.7 ( q , k , b ) ∈ N \eta= \eta_{10.7}(q,k,b) \in\mathbb{N} η = η 10.7 ( q , k , b ) ∈ N such that every ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable ( b , k ) (b,k) ( b , k ) -narrow graph G G G has a clique-faithful cactus-decomposition C C C such that every displayed separation has order ≤ k \le k ≤ k and the set of displayed separations of C C C η \eta η -approximates the set of all k k k -separations of G G G .
Proof. Let q , k , b ∈ N q,k,b \in\mathbb{N} q , k , b ∈ N be arbitrary. Let γ , κ , ϱ , ε ˉ ∈ N \gamma,\kappa,\varrho,\bar{\varepsilon} \in\mathbb{N} γ , κ , ϱ , ε ˉ ∈ N be the constants guaranteed by Theorem 10.4 . We define
δ : = max { max { γ 5.1 ( q , k ) , γ 6.6 ( q , k ) } + 2 ε ˉ + 1 , max { b , β 6.6 ( q , k ) , γ + ϱ } + ε ˉ + k + 1 } . \delta:= \max\{\max\{\gamma_{5.1}(q,k),\gamma_{6.6}(q,k)\}+2\bar{\varepsilon}+1,\max\{b,\beta_{6.6}(q,k),\gamma+\varrho\}+\bar{\varepsilon}+k+1\}. δ := max { max { γ 5.1 ( q , k ) , γ 6.6 ( q , k )} + 2 ε ˉ + 1 , max { b , β 6.6 ( q , k ) , γ + ϱ } + ε ˉ + k + 1 } .
ζ : = δ 9.9 ′ ( k , δ ) . \zeta:= \delta'_{9.9}(k,\delta). ζ := δ 9.9 ′ ( k , δ ) .
Then by Theorem 10.4 (applied to these choices of q , k , b , ζ q,k,b,\zeta q , k , b , ζ ) there exist constants q ′ , ε ∈ N q',\varepsilon\in\mathbb{N} q ′ , ε ∈ N and a set F \mathcal{F} F of graphic-k k k -flowers of G G G of spread ≥ ζ \ge\zeta ≥ ζ and a symmetric nested set N N N of narrow-γ \gamma γ -nested proper k k k -separations of G G G with corresponding tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) of G G G and a bipartition X ⊔ Y = V ( T ) X \sqcup Y = V(T) X ⊔ Y = V ( T ) such that Theorem 10.4 (i)–(vii) hold. For every node t ∈ V ( T ) t \in V(T) t ∈ V ( T ) , we denote by σ t \sigma_t σ t the splitting star of N N N corresponding to t t t .
For a flower F t F_t F t (with t ∈ Y t \in Y t ∈ Y ) and α ∈ N \alpha\in\mathbb{N} α ∈ N , we define S t , α S_{t,\alpha} S t , α to be the set of displayed separations of F t F_t F t that are α \alpha α -crossed by some displayed separation of F t F_t F t . Moreover, we let P t , α : = D ( F t ) ∖ S t , α P_{t,\alpha} := D(F_t) \setminus S_{t,\alpha} P t , α := D ( F t ) ∖ S t , α be the set of displayed separations of F t F_t F t that are α \alpha α -nested with D ( F t ) D(F_t) D ( F t ) . For every α ∈ N \alpha\in\mathbb{N} α ∈ N with α ≥ γ 5.1 ( q , k ) \alpha\ge\gamma_{5.1}(q,k) α ≥ γ 5.1 ( q , k ) , the set S t , α S_{t,\alpha} S t , α is uniformly α \alpha α -crossing by Lemma 9.7. For every such α \alpha α , A ( D ( F t ) , α ) \mathcal{A}(D(F_t),\alpha) A ( D ( F t ) , α ) denotes the set of all α \alpha α -asterisks of S t , α S_{t,\alpha} S t , α , which is well-defined (see Definition 9.8) and partitions S t , α S_{t,\alpha} S t , α .
Claim 10.7.1. There is a unique nontrivial A t ∈ A ( D ( F t ) , δ ) A_t \in\mathcal{A}(D(F_t),\delta) A t ∈ A ( D ( F t ) , δ ) such that S t , ζ ⊆ A t S_{t,\zeta} \subseteq A_t S t , ζ ⊆ A t .
Proof of the claim. Since F t F_t F t has spread ≥ ζ \geq\zeta ≥ ζ , we have A ( D ( F t ) , ζ ) ≠ ∅ \mathcal{A}(D(F_t),\zeta) \neq\emptyset A ( D ( F t ) , ζ ) = ∅ . So we can apply Lemma 9.9 (since δ ≥ γ 5.1 ( q , k ) \delta\geq\gamma_{5.1}(q,k) δ ≥ γ 5.1 ( q , k ) and ζ ≥ δ 9.9 ′ ( k , δ ) \zeta\geq\delta'_{9.9}(k,\delta) ζ ≥ δ 9.9 ′ ( k , δ ) ) and get that there is a unique A t ∈ A ( D ( F t ) , δ ) A_t \in\mathcal{A}(D(F_t),\delta) A t ∈ A ( D ( F t ) , δ ) with S t , ζ = ⋃ A ( D ( F t ) , ζ ) ⊆ A t S_{t,\zeta} = \bigcup\mathcal{A}(D(F_t),\zeta) \subseteq A_t S t , ζ = ⋃ A ( D ( F t ) , ζ ) ⊆ A t . Since S t , ζ S_{t,\zeta} S t , ζ contains two ζ \zeta ζ -crossing separations, and ζ ≥ δ \zeta\geq\delta ζ ≥ δ , the δ \delta δ -asterisk A t A_t A t is nontrivial. ◇
Since A t A_t A t is a nontrivial δ \delta δ -asterisk, every separation in A t A_t A t is δ \delta δ -crossed by another separation in A t A_t A t . Therefore every separation in A t A_t A t is 2 ( δ − k ) 2(\delta-k) 2 ( δ − k ) -broad by Lemma 5.6, hence b b b -broad since δ ≥ b + k \delta\geq b+k δ ≥ b + k . Similarly, every separation in A t A_t A t has breadth > q >q > q since δ ≥ β 6.6 ( q , k ) + k + 1 ≥ q + k + 1 \delta\geq\beta_{6.6}(q,k)+k+1 \geq q+k+1 δ ≥ β 6.6 ( q , k ) + k + 1 ≥ q + k + 1 . Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, every separation in A t A_t A t is a k k k -separation, hence is narrow since G G G is ( b , k ) (b,k) ( b , k ) -narrow.
Additionally, every separation in A t A_t A t is ε ˉ \bar{\varepsilon} ε ˉ -broad since δ ≥ ε ˉ + k \delta\geq\bar{\varepsilon}+k δ ≥ ε ˉ + k . By Theorem 10.4 (vii), every separation s s s in A t A_t A t is ε ˉ \bar{\varepsilon} ε ˉ -close to some k k k -separation s ′ s' s ′ that is nested with N N N . For each s ∈ A t s \in A_t s ∈ A t we fix one such s ′ s' s ′ , chosen so that ( s ∗ ) ′ = ( s ′ ) ∗ (s^*)'=(s')^* ( s ∗ ) ′ = ( s ′ ) ∗ for all s ∈ A t s \in A_t s ∈ A t . We let A t ′ = { s ′ : s ∈ A t } A'_t=\{s':s \in A_t\} A t ′ = { s ′ : s ∈ A t } . Since A t A_t A t is a δ \delta δ -asterisk, it is δ \delta δ -crossing-connected, so A t ′ A'_t A t ′ is ( δ − 2 ε ˉ ) (\delta-2\bar{\varepsilon}) ( δ − 2 ε ˉ ) -crossing-connected by Lemma 10.5, with δ − 2 ε ˉ > γ 5.1 ( q , k ) , γ 6.6 ( q , k ) \delta-2\bar{\varepsilon}>\gamma_{5.1}(q,k),\gamma_{6.6}(q,k) δ − 2 ε ˉ > γ 5.1 ( q , k ) , γ 6.6 ( q , k ) . Moreover, note that every separation in A t ′ A'_t A t ′ is ( 2 δ − 2 k − ε ˉ ) (2\delta-2k-\bar{\varepsilon}) ( 2 δ − 2 k − ε ˉ ) -broad by Lemma 5.9 .
Claim 10.7.2. Let s ∈ A t s \in A_t s ∈ A t . Then s ′ s' s ′ almost-interlaces σ t \sigma_t σ t .
Proof of the claim. First, we argue that s s s almost-ϱ \varrho ϱ -interlaces σ t \sigma_t σ t . Since A t A_t A t is a nontrivial δ \delta δ -asterisk, there is a separation in A t A_t A t that δ \delta δ -crosses s s s . Since δ ≥ ϱ \delta\geq\varrho δ ≥ ϱ , Theorem 10.4 (iii) implies that s s s almost-ϱ \varrho ϱ -interlaces σ t \sigma_t σ t .
Suppose that s ′ s' s ′ does not almost-interlace σ t \sigma_t σ t . Since s ′ s' s ′ is nested with N N N , there is a separation n ∈ σ t n \in\sigma_t n ∈ σ t with s ′ ≤ n s' \leq n s ′ ≤ n or ( s ′ ) ∗ ≤ n (s')^* \leq n ( s ′ ) ∗ ≤ n , say the former. Since s s s almost-ϱ \varrho ϱ -interlaces σ t \sigma_t σ t , we have n ≤ ϱ s n \leq_{\varrho} s n ≤ ϱ s or n ≤ ϱ s ∗ n \leq_{\varrho} s^* n ≤ ϱ s ∗ . In the latter case, we have s ′ ≤ n ≤ ϱ s ∗ ≈ ε ˉ ( s ′ ) ∗ s' \leq n \leq_{\varrho} s^* \approx_{\bar{\varepsilon}} (s')^* s ′ ≤ n ≤ ϱ s ∗ ≈ ε ˉ ( s ′ ) ∗ , so s ′ ≤ ϱ + ε ˉ ( s ′ ) ∗ s' \leq_{\varrho+\bar{\varepsilon}} (s')^* s ′ ≤ ϱ + ε ˉ ( s ′ ) ∗ , which implies that s ′ s' s ′ has breadth ≤ ϱ + ε ˉ \leq\varrho+\bar{\varepsilon} ≤ ϱ + ε ˉ by Lemma 8.3 , contradicting that s ′ ∈ A t ′ s' \in A'_t s ′ ∈ A t ′ has breadth ≥ 2 δ − 2 k − ε ˉ > ϱ + ε ˉ \geq2\delta-2k-\bar{\varepsilon}>\varrho+\bar{\varepsilon} ≥ 2 δ − 2 k − ε ˉ > ϱ + ε ˉ . In the former case, we have s ≈ ε ˉ s ′ ≤ n ≤ ϱ s s \approx_{\bar{\varepsilon}} s' \leq n \leq_{\varrho} s s ≈ ε ˉ s ′ ≤ n ≤ ϱ s . In particular, s ≈ ϱ + ε ˉ n s \approx_{\varrho+\bar{\varepsilon}} n s ≈ ϱ + ε ˉ n . Recall that n ∈ N n \in N n ∈ N is narrow-γ \gamma γ -nested and that s s s is narrow and proper. By Observation 8.11 , s s s is narrow-( γ + ϱ + ε ˉ ) (\gamma+\varrho+\bar{\varepsilon}) ( γ + ϱ + ε ˉ ) -nested. With δ ≥ γ + ϱ + ε ˉ \delta\geq\gamma+\varrho+\bar{\varepsilon} δ ≥ γ + ϱ + ε ˉ , s s s is narrow-δ \delta δ -nested. However, since s ∈ A t s \in A_t s ∈ A t , there is some k k k -separation r ∈ A t r \in A_t r ∈ A t that δ \delta δ -crosses s s s . Recall that r r r is (proper and) narrow. This contradicts that s s s is narrow-δ \delta δ -nested. ◇
For t ∈ Y t \in Y t ∈ Y , we let G ~ t \widetilde{G}_t G t be the graph obtained from G G G by making G [ A ] G[A] G [ A ] into a clique for every ( A , B ) ∈ σ t (A,B) \in\sigma_t ( A , B ) ∈ σ t . So, G ~ t \widetilde{G}_t G t is again ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and ( b , k ) (b,k) ( b , k ) -narrow. Note that every separation in A t ′ A'_t A t ′ is a separation of G ~ t \widetilde{G}_t G t since it almost-interlaces σ t \sigma_t σ t by Claim 10.7.2 . Recall that every separation in A t ′ A'_t A t ′ is ( 2 δ − 2 k − ε ˉ ) (2\delta-2k-\bar{\varepsilon}) ( 2 δ − 2 k − ε ˉ ) -broad, and hence q q q -broad by our choice of δ \delta δ . Thus, A t ′ A'_t A t ′ is uniformly γ 5.1 ( q , k ) \gamma_{5.1}(q,k) γ 5.1 ( q , k ) -crossing by Key Lemma 5.1, hence uniformly ( δ − 2 ε ˉ ) (\delta-2\bar{\varepsilon}) ( δ − 2 ε ˉ ) -crossing since δ − 2 ε ˉ ≥ γ 5.1 ( q , k ) \delta-2\bar{\varepsilon} \geq\gamma_{5.1}(q,k) δ − 2 ε ˉ ≥ γ 5.1 ( q , k ) . Recall that the set A t ′ A'_t A t ′ is symmetric. Moreover, since A t A_t A t is nontrivial, there are δ \delta δ -crossing separations r , s r,s r , s in A t A_t A t . By Lemma 7.2 , the separations r ′ , s ′ r',s' r ′ , s ′ in A t ′ A'_t A t ′ ( δ − 2 ε ˉ ) (\delta-2\bar{\varepsilon}) ( δ − 2 ε ˉ ) -cross, so A t ′ A'_t A t ′ is nontrivial. Therefore, A t ′ A'_t A t ′ is nontrivial ( δ − 2 ε ˉ ) (\delta-2\bar{\varepsilon}) ( δ − 2 ε ˉ ) -asteriskal (both in G G G and in G ~ t \widetilde{G}_t G t ).
Now we apply Key Lemma 6.6 to G ~ t \widetilde{G}_t G t and A t ′ A'_t A t ′ with parameter δ − 2 ε > γ 6.6 ( q , k ) \delta-2\varepsilon>\gamma_{6.6}(q,k) δ − 2 ε > γ 6.6 ( q , k ) . We can do it since the separations in A t ′ A'_t A t ′ are ( 2 ( δ − k ) − ε ˉ ) (2(\delta-k)-\bar{\varepsilon}) ( 2 ( δ − k ) − ε ˉ ) -broad with 2 ( δ − k ) − ε ˉ ≥ β 6.6 ( q , k ) 2(\delta-k)-\bar{\varepsilon}\ge\beta_{6.6}(q,k) 2 ( δ − k ) − ε ˉ ≥ β 6.6 ( q , k ) , and hence also narrow. This gives a graphic-k k k -flower F ~ t \widetilde{F}_t F t such that Key Lemma 6.6 (i)–(iv) hold. In particular, every separation in A t ′ A'_t A t ′ is ε 6.6 \varepsilon_{6.6} ε 6.6 -close to some displayed separation of F ~ t \widetilde{F}_t F t by Key Lemma 6.6 (i). We choose
η ′ : = ε + max { ε ˉ + ε 6.6 , 2 ( ζ + k ) + ϱ } . \eta' := \varepsilon+\max\{\bar{\varepsilon}+\varepsilon_{6.6},\,2(\zeta+k)+\varrho\}. η ′ := ε + max { ε ˉ + ε 6.6 , 2 ( ζ + k ) + ϱ } .
Claim 10.7.3. The set N ∪ ⋃ t ∈ Y D ( F ~ t ) N\cup\bigcup_{t\in Y}D(\widetilde{F}_t) N ∪ ⋃ t ∈ Y D ( F t ) η ′ \eta' η ′ -approximates the set of all k k k -separations of G G G .
Proof. Let s s s be an η ′ \eta' η ′ -broad k k k -separation of G G G . We need to show that s s s is η ′ \eta' η ′ -close to a separation in N N N or to a displayed separation of some F ~ t \widetilde{F}_t F t . By Theorem 10.4 (iv), s s s is ε \varepsilon ε -close to a separation in N N N or ε \varepsilon ε -close to a displayed separation of some F t F_t F t . We are done in the former case (since η ′ ≥ ε \eta'\ge\varepsilon η ′ ≥ ε ), so assume the latter.
So, s s s is ε \varepsilon ε -close to some displayed separation d d d of some F t F_t F t (with t ∈ Y t\in Y t ∈ Y ). Recall that D ( F t ) = S t , ζ ⊔ P t , ζ D(F_t)=S_{t,\zeta}\sqcup P_{t,\zeta} D ( F t ) = S t , ζ ⊔ P t , ζ . If d ∈ S t , ζ ⊆ A t d\in S_{t,\zeta}\subseteq A_t d ∈ S t , ζ ⊆ A t , then there is a separation d ′ ∈ A t ′ d'\in A'_t d ′ ∈ A t ′ that is ε ˉ \bar{\varepsilon} ε ˉ -close to d d d . By Key Lemma 6.6 (i), there is a displayed separation d ′ ′ d'' d ′′ of F ~ t \widetilde{F}_t F t that is ε 6.6 \varepsilon_{6.6} ε 6.6 -close to d ′ d' d ′ . Thus, by Lemma 4.9 and with η ′ ≥ ε + ε ˉ + ε 6.6 \eta'\ge\varepsilon+\bar{\varepsilon}+\varepsilon_{6.6} η ′ ≥ ε + ε ˉ + ε 6.6 , s s s is η ′ \eta' η ′ -close to d ′ ′ d'' d ′′ and we are done. Otherwise, d d d is in P t , ζ P_{t,\zeta} P t , ζ . So, d d d is not ζ \zeta ζ -crossed (and hence not ζ \zeta ζ -crissed) by a displayed separation of F t F_t F t . Then d d d is 2 ( ζ + k ) 2(\zeta+k) 2 ( ζ + k ) -close to a petal-separation p p p of F t F_t F t or its inverse by Lemma 3.13. Then, p p p or p ∗ p^* p ∗ is ε + 2 ( ζ + k ) \varepsilon+2(\zeta+k) ε + 2 ( ζ + k ) -close to s s s by Lemma 4.9 , so both are ϱ \varrho ϱ -broad by Lemma 5.9 . Thus, p p p is ϱ \varrho ϱ -close to a separation of N N N by Theorem 10.4 (ii) (so its inverse too because N N N is symmetric). Therefore, by Lemma 4.9 and with η ′ ≥ ε + 2 ( ζ + k ) + ϱ \eta'\ge\varepsilon+2(\zeta+k)+\varrho η ′ ≥ ε + 2 ( ζ + k ) + ϱ , s s s is η ′ \eta' η ′ -close to a separation in N N N and we are done. □
We are now ready to build the desired cactus-decomposition C = ( C , W ) \mathcal{C}=(C,\mathcal{W}) C = ( C , W ) of G G G .
Claim 10.7.4. For every t ∈ Y t\in Y t ∈ Y and every separation n = ( A , B ) ∈ σ t n=(A,B)\in\sigma_t n = ( A , B ) ∈ σ t , there is a bag of F ~ t \widetilde{F}_t F t that contains A A A .
Proof of the claim. Let t ∈ Y t\in Y t ∈ Y and n = ( A , B ) ∈ σ t n=(A,B)\in\sigma_t n = ( A , B ) ∈ σ t . Recall that G ~ t [ A ] \widetilde{G}_t[A] G t [ A ] is complete and that F ~ t \widetilde{F}_t F t is a graphic-k k k -flower of G ~ t \widetilde{G}_t G t . First, we consider the case where F ~ t \widetilde{F}_t F t is a graphic-k k k -daisy, in which case F ~ t \widetilde{F}_t F t has the clover-property by Key Lemma 6.6 (iii). Then, some bag of F ~ t \widetilde{F}_t F t contains A A A : this is obvious if A A A is contained in the centre of F ~ t \widetilde{F}_t F t , and follows from Lemma 10.6 otherwise. Now, assume that F ~ t \widetilde{F}_t F t is a graphic-k k k -anemone. Since graphic-k k k -anemones are star-decompositions (hence tree-decompositions), it follows from the Helly property of subtrees of a tree that some bag of F ~ t \widetilde{F}_t F t contains A A A . ♢
Without loss of generality, we can assume that no two nodes of Y Y Y are adjacent in T T T (otherwise, we can subdivide this edge y y ′ yy' y y ′ of T T T , add the subdivision-node to X X X and define the bag of the subdivision-node to be the intersection V y ∩ V y ′ V_y\cap V_{y'} V y ∩ V y ′ ). For each t ∈ Y t\in Y t ∈ Y , we write F ~ t = ( H t , ( V x ′ : x ∈ V ( H t ) ) ) \widetilde{F}_t=(H_t,(V'_x:x\in V(H_t))) F t = ( H t , ( V x ′ : x ∈ V ( H t ))) . We construct the graph C C C from the disjoint union of the forest T [ X ] T[X] T [ X ] and the graphs H t H_t H t for t ∈ Y t\in Y t ∈ Y . For every t ∈ Y t\in Y t ∈ Y and every edge t t ′ tt' t t ′ of T T T (hence with t ′ ∈ X t'\in X t ′ ∈ X ), let n = ( A , B ) ∈ σ t n=(A,B)\in\sigma_t n = ( A , B ) ∈ σ t be the separation induced by t t ′ tt' t t ′ in ( T , V ) (T,\mathcal{V}) ( T , V ) . By Claim 10.7.4 , there is a node x ∈ V ( H t ) x\in V(H_t) x ∈ V ( H t ) with A ⊆ V x ′ A\subseteq V'_x A ⊆ V x ′ . Pick such a node arbitrarily and add the edge x t ′ xt' x t ′ to C C C . We define C C C to be the decomposition graph of C C C . Now, we define the bags of C C C . For a node t ∈ X t\in X t ∈ X , we define its bag to be W t : = V t W_t:=V_t W t := V t . For a node x x x of a decomposition-graph of some F ~ t \widetilde{F}_t F t with t ∈ Y t\in Y t ∈ Y , with bag V x ′ V'_x V x ′ in F ~ t \widetilde{F}_t F t , we define its bag to be W x : = V x ′ ∩ V t W_x:=V'_x\cap V_t W x := V x ′ ∩ V t . For each graphic-k k k -anemone F ~ t \widetilde{F}_t F t , we declare the centre of F ~ t \widetilde{F}_t F t as a P P P -node. There are no other P P P -nodes. This concludes the definition of C = ( C , W ) \mathcal{C} = (\mathcal{C}, \mathcal{W}) C = ( C , W ) .
Claim 10.7.5. C \mathcal{C} C is a cactus-decomposition of G G G .
Proof of the claim. Observe that C \mathcal{C} C is connected and that the only cycles in C \mathcal{C} C are the decomposition-cycles of the graphic-k k k -daisies. Hence C \mathcal{C} C is a strong cactus. It remains to verify (H1) and (H2).
(H1). For each t ∈ Y t \in Y t ∈ Y , restricting the bags of F ~ t \widetilde{F}_t F t to V t V_t V t gives a graph-decomposition of G [ V t ] G[V_t] G [ V t ] , since F ~ t \widetilde{F}_t F t is a graph-decomposition of G ~ t ⊇ G \widetilde{G}_t \supseteq G G t ⊇ G . Thus every vertex and every edge of G G G is contained in some W z W_z W z (with z ∈ V ( C ) z \in V(\mathcal{C}) z ∈ V ( C ) ), since this holds for the original tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) .
(H2). Let v ∈ V ( G ) v \in V(G) v ∈ V ( G ) be arbitrary. The nodes of T T T whose bag V t V_t V t contains v v v induce a subtree T v T_v T v of T T T . Each node t ∈ Y ∩ V ( T v ) t \in Y \cap V(T_v) t ∈ Y ∩ V ( T v ) is replaced by the nonempty connected subgraph of H t H_t H t induced by the nodes of H t H_t H t whose bag contains v v v . If an edge t t ′ tt' t t ′ of T v T_v T v (with t ∈ Y t \in Y t ∈ Y and t ′ ∈ X t' \in X t ′ ∈ X ) is replaced in C \mathcal{C} C by x t ′ xt' x t ′ for some x ∈ V ( H t ) x \in V(H_t) x ∈ V ( H t ) , then v ∈ V t ∩ V t ′ ⊆ A ⊆ V x ′ v \in V_t \cap V_{t'} \subseteq A \subseteq V'_x v ∈ V t ∩ V t ′ ⊆ A ⊆ V x ′ , where ( A , B ) ∈ σ t (A,B) \in\sigma_t ( A , B ) ∈ σ t is the separation induced by t t ′ tt' t t ′ . Thus both W x W_x W x and W t ′ W_{t'} W t ′ contain v v v . Every edge of T v T_v T v between two nodes of X X X is also an edge of C \mathcal{C} C . Therefore the nodes of C \mathcal{C} C whose bag contains v v v induce a connected subgraph of C \mathcal{C} C . ◇
Claim 10.7.6. Every displayed separation of C \mathcal{C} C has order ≤ k \leq k ≤ k .
Proof. Let ( A , B ) (A,B) ( A , B ) be a displayed separation of C \mathcal{C} C induced by a cut ( U , W ) (U,W) ( U , W ) of C \mathcal{C} C , and let F F F be the set of edges of C \mathcal{C} C with one end in U U U and the other in W W W . We distinguish the three types of displayed separations.
Case 1: ∣ F ∣ ≤ 1 |F| \leq1 ∣ F ∣ ≤ 1 . If F = ∅ F = \varnothing F = ∅ , then A ∩ B = ∅ A \cap B = \varnothing A ∩ B = ∅ , so the assertion holds. Assume now that ∣ F ∣ = 1 |F| = 1 ∣ F ∣ = 1 . If the unique edge of F F F arises from an edge of T T T , then ( A , B ) ∈ N (A,B) \in N ( A , B ) ∈ N and hence has order k k k . Otherwise, it lies in the star H t H_t H t of some graphic-k k k -anemone F ~ t \widetilde{F}_t F t , so A ∩ B A \cap B A ∩ B is contained in the central bag of F ~ t \widetilde{F}_t F t and hence has size ≤ k \leq k ≤ k .
Case 2: ∣ F ∣ = 2 |F| = 2 ∣ F ∣ = 2 and both edges of F F F lie on a common cycle of C \mathcal{C} C . This cycle is H t H_t H t for some graphic-k k k -daisy F ~ t \widetilde{F}_t F t . By (H2) and since W x ⊆ V x ′ W_x \subseteq V'_x W x ⊆ V x ′ for every x ∈ V ( H t ) x \in V(H_t) x ∈ V ( H t ) , we have
A ∩ B = ⋃ x y ∈ F ( W x ∩ W y ) ⊆ ⋃ x y ∈ F ( V x ′ ∩ V y ′ ) . A \cap B = \bigcup_{xy \in F}(W_x \cap W_y) \subseteq\bigcup_{xy \in F}(V'_x \cap V'_y). A ∩ B = x y ∈ F ⋃ ( W x ∩ W y ) ⊆ x y ∈ F ⋃ ( V x ′ ∩ V y ′ ) . The set on the right is the separator of the separation displayed by these two edges in F ~ t \widetilde{F}_t F t , and hence has size ≤ k \leq k ≤ k .
Case 3: ∣ F ∣ ≥ 2 |F| \geq2 ∣ F ∣ ≥ 2 and all edges are incident with a P P P -node of C \mathcal{C} C . Let c c c be this P P P -node. By definition of the P P P -nodes, c c c is the centre of the star H t H_t H t for some graphic-k k k -anemone F ~ t \widetilde{F}_t F t . By (H2), we have
A ∩ B = ⋃ c x ∈ F ( W c ∩ W x ) ⊆ W c ⊆ V c ′ . A \cap B = \bigcup_{cx \in F}(W_c \cap W_x) \subseteq W_c \subseteq V'_c. A ∩ B = c x ∈ F ⋃ ( W c ∩ W x ) ⊆ W c ⊆ V c ′ . Since V c ′ V'_c V c ′ is the central bag of a graphic-k k k -anemone, it has size ≤ k \leq k ≤ k . □
Claim 10.7.7. For every t ∈ Y t \in Y t ∈ Y and every separation d = ( A , B ) ∈ D ( F ~ t ) d = (A,B) \in D(\widetilde{F}_t) d = ( A , B ) ∈ D ( F t ) , there is a separation d ′ = ( A ′ , B ′ ) d' = (A',B') d ′ = ( A ′ , B ′ ) displayed by C \mathcal{C} C such that A ′ ⊆ A A' \subseteq A A ′ ⊆ A and B ′ ⊆ B B' \subseteq B B ′ ⊆ B . In particular, d d d and d ′ d' d ′ are k k k -close.
Proof of the claim. Let t ∈ Y t \in Y t ∈ Y and d = ( A , B ) ∈ D ( F ~ t ) d = (A, B) \in D(\widetilde{F}_t) d = ( A , B ) ∈ D ( F t ) . Write F ~ t = ( H , V ′ ) \widetilde{F}_t = (H, \mathcal{V}') F t = ( H , V ′ ) . If A = ∅ A = \varnothing A = ∅ or B = ∅ B = \varnothing B = ∅ , we may take d ′ = d d' = d d ′ = d . Hence assume that both sides are nonempty. By the construction of C C C , every component of C − E ( H ) C - E(H) C − E ( H ) contains exactly one node of H H H . For each x ∈ V ( H ) x \in V(H) x ∈ V ( H ) , let K x K_x K x denote the component containing x x x .
First, note that for every x ∈ V ( H ) x \in V(H) x ∈ V ( H ) , we have ⋃ z ∈ V ( K x ) W z ⊆ V x ′ \bigcup_{z \in V(K_x)} W_z \subseteq V'_x ⋃ z ∈ V ( K x ) W z ⊆ V x ′ . To see it, observe first that W x = V x ′ ∩ V t ⊆ V x ′ W_x = V'_x \cap V_t \subseteq V'_x W x = V x ′ ∩ V t ⊆ V x ′ . Then, every other node of K x K_x K x comes from a component of T − t T - t T − t corresponding to a separation n = ( A n , B n ) ∈ σ t n = (A_n, B_n) \in\sigma_t n = ( A n , B n ) ∈ σ t whose chosen attachment node in H H H is x x x . Each bag of ( C , W ) (C, \mathcal{W}) ( C , W ) coming from this component of T − t T - t T − t is contained in a bag of ( T , V ) (T, \mathcal{V}) ( T , V ) from that component, and hence in A n A_n A n . Since x x x was chosen as the attachment node in H H H for this component, we have A n ⊆ V x ′ A_n \subseteq V'_x A n ⊆ V x ′ , as desired.
Let ( U , W ) (U, W) ( U , W ) be a cut of H H H displaying d d d . Consider the cut ( U ′ , W ′ ) : = ( ⋃ x ∈ U V ( K x ) , ⋃ x ∈ W V ( K x ) ) (U', W') := \left(\bigcup_{x \in U} V(K_x), \bigcup_{x \in W} V(K_x)\right) ( U ′ , W ′ ) := ( ⋃ x ∈ U V ( K x ) , ⋃ x ∈ W V ( K x ) ) of C C C . The edges between U ′ U' U ′ and W ′ W' W ′ in C C C are exactly the edges between U U U and W W W in H H H . If H H H is a cycle, these are two edges on the same cycle, and if H H H is a star, these are edges incident to the centre of the star, which is a P P P -node. In both cases, the separation d ′ = ( A ′ , B ′ ) d' = (A', B') d ′ = ( A ′ , B ′ ) induced by ( U ′ , W ′ ) (U', W') ( U ′ , W ′ ) is displayed by C C C .
By the previous observation, we have A ′ = ⋃ x ∈ U ⋃ z ∈ V ( K x ) W z ⊆ ⋃ x ∈ U V x ′ = A A' = \bigcup_{x \in U} \bigcup_{z \in V(K_x)} W_z \subseteq\bigcup_{x \in U} V'_x = A A ′ = ⋃ x ∈ U ⋃ z ∈ V ( K x ) W z ⊆ ⋃ x ∈ U V x ′ = A and similarly B ′ ⊆ B B' \subseteq B B ′ ⊆ B . As d d d and d ′ d' d ′ have order at most k k k , it follows that they are k k k -close. ◇
Claim 10.7.8. Let η : = η ′ + k \eta:= \eta' + k η := η ′ + k . Then the set of displayed separations of C C C η \eta η -approximates the set of all k k k -separations of G G G .
Proof of the claim. Let s s s be an η \eta η -broad k k k -separation of G G G . By Claim 10.7.3 , s s s is η ′ \eta' η ′ -close to some separation in N ∪ ⋃ t ∈ Y D ( F ~ t ) N \cup\bigcup_{t \in Y} D(\widetilde{F}_t) N ∪ ⋃ t ∈ Y D ( F t ) . Note that every separation in N N N is displayed by C C C . So, if s s s is η ′ \eta' η ′ -close to a separation in N N N , it is η \eta η -close to a displayed separation of C C C . Otherwise, s s s is η ′ \eta' η ′ -close to some displayed separation d = ( A , B ) ∈ D ( F ~ t ) d = (A, B) \in D(\widetilde{F}_t) d = ( A , B ) ∈ D ( F t ) for some t ∈ Y t \in Y t ∈ Y . Then, d d d is k k k -close to a separation displayed by C C C by Claim 10.7.7 . Therefore, s s s is ( η ′ + k ) (\eta' + k) ( η ′ + k ) -close to a displayed separation of C C C . ◇
Claim 10.7.9. C C C is clique-faithful.
Proof of the claim. Let K K K be a clique of G G G . By the Helly property of subtrees of a tree, there is a node t t t of T T T such that K ⊆ V t K \subseteq V_t K ⊆ V t . If t ∈ X t \in X t ∈ X , then we are done since V t = W t V_t = W_t V t = W t and t t t is not a P P P -node. Otherwise, we have t ∈ Y t \in Y t ∈ Y . Recall that F ~ t \widetilde{F}_t F t is a graphic-k k k -flower of G ~ t ⊇ G \widetilde{G}_t \supseteq G G t ⊇ G . First assume that F ~ t = ( O , V ′ ) \widetilde{F}_t = (O, \mathcal{V}') F t = ( O , V ′ ) is a graphic-k k k -daisy. Then F ~ t \widetilde{F}_t F t has the clover-property. Therefore, there is a node x x x of O O O such that K ⊆ V x ′ K \subseteq V'_x K ⊆ V x ′ : this is obvious if K K K is contained in the centre of F ~ t \widetilde{F}_t F t , and follows from Lemma 10.6 otherwise. Moreover, x x x is not a P P P -node since F ~ t \widetilde{F}_t F t is a graphic-k k k -daisy. Since K ⊆ V t K \subseteq V_t K ⊆ V t , we have K ⊆ V x ′ ∩ V t = W x K \subseteq V'_x \cap V_t = W_x K ⊆ V x ′ ∩ V t = W x . Suppose now that F ~ t = ( S , V ′ ) \widetilde{F}_t = (S, \mathcal{V}') F t = ( S , V ′ ) is a graphic-k k k -anemone, and let c c c be the centre of S S S . Since F ~ t \widetilde{F}_t F t is a tree-decomposition, there is a node x x x of S S S such that K ⊆ V x ′ K \subseteq V'_x K ⊆ V x ′ . Furthermore, since V c ′ V'_c V c ′ is contained in all other bags of F ~ t \widetilde{F}_t F t , we can choose x x x distinct from c c c , so that x x x is not a P P P -node. Since K ⊆ V t K \subseteq V_t K ⊆ V t , we have K ⊆ V x ′ ∩ V t = W x K \subseteq V'_x \cap V_t = W_x K ⊆ V x ′ ∩ V t = W x . ◇
Claims 10.7.5, 10.7.6, 10.7.8 and 10.7.9 together conclude the proof of the theorem. □
The approximate cactus-decomposition
In this section, we prove the main result, Theorem 1.2, in full generality. Our strategy for this is to show that we can reduce to the case where G G G is ( β , k ) (\beta, k) ( β , k ) -narrow. We do this in § 12.1 , and then prove Theorem 1.2 in § 12.2 .
Reducing to a narrow graph
Let s s s be a separation of a graph G G G , and let K K K be a component of G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) . We say that the separation ( V ( K ) ∪ N ( K ) , V ( G − K ) ) (V(K)\cup N(K), V(G-K)) ( V ( K ) ∪ N ( K ) , V ( G − K )) is a componental separation at s s s . Note that s ∗ s^* s ∗ might not be componental. A separation s ′ s' s ′ is ( β , k ) (\beta, k) ( β , k ) -irrelevant if it is a componental ( < k ) (< k) ( < k ) -separation at some β \beta β -broad k k k -separation of G G G . Recall that G G G is ( β , k ) (\beta, k) ( β , k ) -narrow if every β \beta β -broad k k k -separation is narrow.
Our goal in this section is to show that we can easily reduce from the case where G G G is ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable to the case where G G G is additionally ( β , k ) (\beta, k) ( β , k ) -narrow. To do so, we show that the ( β , k ) (\beta, k) ( β , k ) -irrelevant separations form a star and that the torso obtained by cutting at all of them is ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable and ( β , k ) (\beta, k) ( β , k ) -narrow.
Key Lemma 11.1. Let G G G be a ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable graph and let β > β 11.5 ( q , k ) \beta> \beta_{11.5}(q,k) β > β 11.5 ( q , k ) . Then the set σ \sigma σ of all ( β , k ) (\beta, k) ( β , k ) -irrelevant separations forms a star. Moreover, the torso of σ \sigma σ is ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable and ( β , k ) (\beta, k) ( β , k ) -narrow.
The following lemma motivates the study of ( β , k ) (\beta, k) ( β , k ) -irrelevant separations.
Lemma 11.2. Let G G G be a graph. Then G G G is ( β , k ) (\beta, k) ( β , k ) -narrow if and only if it does not have a ( β , k ) (\beta, k) ( β , k ) -irrelevant separation.
Proof. If s ′ s' s ′ is a ( β , k ) (\beta, k) ( β , k ) -irrelevant separation, then it is a componental ( < k ) (< k) ( < k ) -separation at some β \beta β -broad k k k -separation s s s of G G G . Then s s s is not narrow, so G G G is not ( β , k ) (\beta, k) ( β , k ) -narrow.
On the other hand, if G G G is not ( β , k ) (\beta, k) ( β , k ) -narrow, then there is a β \beta β -broad k k k -separation s s s of G G G that is not narrow. Hence, there is a componental ( < k ) (< k) ( < k ) -separation s ′ s' s ′ at s s s . Then s ′ s' s ′ is ( β , k ) (\beta, k) ( β , k ) -irrelevant. □
We now prove some results about β \beta β -broad k k k -separations and ( β , k ) (\beta, k) ( β , k ) -irrelevant separations in ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable graphs.
Lemma 11.3. Let G G G be a ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable graph, let β > q \beta> q β > q and let ( C , D ) (C,D) ( C , D ) be a ( β , k ) (\beta, k) ( β , k ) -irrelevant separation of G G G . Then, ∣ C ∖ D ∣ ≤ q \lvert C\setminus D\rvert\leq q ∣ C ∖ D ∣ ≤ q .
Proof. Since ( C , D ) (C,D) ( C , D ) is a ( β , k ) (\beta, k) ( β , k ) -irrelevant separation of G G G , there exists a β \beta β -broad k k k -separation s = ( A , B ) s=(A,B) s = ( A , B ) of G G G such that ( C , D ) (C,D) ( C , D ) is a componental ( < k ) (< k) ( < k ) -separation at s s s . After possibly swapping the names A A A and B B B , we may assume without loss of generality that C ⊆ A C\subseteq A C ⊆ A , so D ⊇ B D\supseteq B D ⊇ B . Hence ∣ D ∖ C ∣ ≥ ∣ B ∖ A ∣ ≥ β > q \lvert D\setminus C\rvert\geq\lvert B\setminus A\rvert\geq\beta> q ∣ D ∖ C ∣ ≥ ∣ B ∖ A ∣ ≥ β > q . Since ( C , D ) (C,D) ( C , D ) is a ( < k ) (< k) ( < k ) -separation and G G G is ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable, we then have ∣ C ∖ D ∣ ≤ q \lvert C\setminus D\rvert\leq q ∣ C ∖ D ∣ ≤ q . □
Lemma 11.4. Let s s s and t t t be two β \beta β -broad k k k -separations of a ( q , k − 1 ) (q, k-1) ( q , k − 1 ) -unbreakable graph G G G . Let K K K be a component of G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) of size h h h . Assume β > 2 ( q + h ) + k \beta> 2(q+h)+k β > 2 ( q + h ) + k . Then V ( K ) V(K) V ( K ) is disjoint from sep ( t ) \operatorname{sep}(t) sep ( t ) .
Proof. Let s : = ( A , B ) s := (A, B) s := ( A , B ) and t : = ( C , D ) t := (C, D) t := ( C , D ) . Assume for a contradiction that V ( K ) V(K) V ( K ) is not disjoint from sep ( t ) \operatorname{sep}(t) sep ( t ) . Without loss of generality, K K K is a component of G [ A ∖ B ] G[A \setminus B] G [ A ∖ B ] and intersects the A t At A t -link. We define s ^ : = ( A ^ , B ^ ) : = ( A ∖ V ( K ) , B ∪ V ( K ) ) \hat{s} := (\hat{A}, \hat{B}) := (A \setminus V(K), B \cup V(K)) s ^ := ( A ^ , B ^ ) := ( A ∖ V ( K ) , B ∪ V ( K )) .
By Lemma 8.17 , there are two opposite corner-separations of s s s and t t t of order k k k and breadth ≥ ( β − k ) / 2 \ge(\beta-k)/2 ≥ ( β − k ) /2 , say the A C AC A C -corner-separation and the B D BD B D -corner-separation. Now, consider the two separations s ^ \hat{s} s ^ and t t t . Then the A ^ C \hat{A}C A ^ C -corner has size ≥ ( β − k ) / 2 − h \ge(\beta-k)/2-h ≥ ( β − k ) /2 − h . Hence, the corresponding corner-separation has breadth ≥ ( β − k ) / 2 − h > q \ge(\beta-k)/2-h>q ≥ ( β − k ) /2 − h > q . Since K K K intersects the A t At A t -link, the A ^ C \hat{A}C A ^ C -corner-separation has order < k <k < k . This contradicts that G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable. □
Corollary 11.5. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let β > β 11.5 ( q , k ) : = 4 q + k \beta>\beta_{11.5}(q,k):=4q+k β > β 11.5 ( q , k ) := 4 q + k . Let s ′ = ( A ′ , B ′ ) s'=(A',B') s ′ = ( A ′ , B ′ ) be a ( β , k ) (\beta,k) ( β , k ) -irrelevant separation, and let t t t be a β \beta β -broad k k k -separation of G G G . Then A ′ ∖ B ′ A'\setminus B' A ′ ∖ B ′ is disjoint from sep ( t ) \operatorname{sep}(t) sep ( t ) .
Proof. Since s ′ s' s ′ is ( β , k ) (\beta,k) ( β , k ) -irrelevant, there is a β \beta β -broad k k k -separation s s s such that s ′ s' s ′ is a componental ( < k ) (<k) ( < k ) -separation at s s s . That is, there is a component K K K of G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) such that A ′ ∖ B ′ = V ( K ) A'\setminus B'=V(K) A ′ ∖ B ′ = V ( K ) . Note that h : = ∣ V ( K ) ∣ ≤ q h:=|V(K)|\le q h := ∣ V ( K ) ∣ ≤ q since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable. Hence we have β > 4 q + k ≥ 2 ( q + h ) + k \beta>4q+k\ge2(q+h)+k β > 4 q + k ≥ 2 ( q + h ) + k and we can apply Lemma 11.4 to the separations s s s and t t t . Therefore, V ( K ) = A ′ ∖ B ′ V(K)=A'\setminus B' V ( K ) = A ′ ∖ B ′ is disjoint from sep ( t ) \operatorname{sep}(t) sep ( t ) . □
Lemma 11.6. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let β > β 11.5 ( q , k ) \beta>\beta_{11.5}(q,k) β > β 11.5 ( q , k ) . Then the set σ \sigma σ of all ( β , k ) (\beta,k) ( β , k ) -irrelevant separations of G G G forms a star.
Proof. Let s ′ ≠ t ′ s'\ne t' s ′ = t ′ be two ( β , k ) (\beta,k) ( β , k ) -irrelevant separations of G G G . That is, there are β \beta β -broad k k k -separations s : = ( A , B ) s:=(A,B) s := ( A , B ) and t : = ( C , D ) t:=(C,D) t := ( C , D ) of G G G , such that s ′ s' s ′ and t ′ t' t ′ are componental ( < k ) (<k) ( < k ) -separations at s s s and t t t , respectively. Let
( A ′ , B ′ ) : = s ′ = ( V ( K ) ∪ N ( K ) , V ( G − K ) ) (A',B'):=s'=(V(K)\cup N(K),V(G-K)) ( A ′ , B ′ ) := s ′ = ( V ( K ) ∪ N ( K ) , V ( G − K ))
and
( C ′ , D ′ ) : = t ′ = ( V ( L ) ∪ N ( L ) , V ( G − L ) ) (C',D'):=t'=(V(L)\cup N(L),V(G-L)) ( C ′ , D ′ ) := t ′ = ( V ( L ) ∪ N ( L ) , V ( G − L ))
where K K K is a component of G − sep ( s ) G-\operatorname{sep}(s) G − sep ( s ) and L L L is a component of G − sep ( t ) G-\operatorname{sep}(t) G − sep ( t ) . First, we show that V ( L ) V(L) V ( L ) is disjoint from V ( K ) ∪ N ( K ) V(K)\cup N(K) V ( K ) ∪ N ( K ) . Consider the two separations s ′ s' s ′ and t t t . By Corollary 11.5, V ( K ) V(K) V ( K ) is disjoint from sep ( t ) \operatorname{sep}(t) sep ( t ) . Since K K K is connected, we have without loss of generality V ( K ) ⊆ C ∖ D V(K)\subseteq C\setminus D V ( K ) ⊆ C ∖ D . Similarly, we have V ( L ) ⊆ A ∖ B V(L)\subseteq A\setminus B V ( L ) ⊆ A ∖ B , say. Hence K K K and L L L are components of G − ( sep ( s ) ∪ sep ( t ) ) G-(\operatorname{sep}(s)\cup\operatorname{sep}(t)) G − ( sep ( s ) ∪ sep ( t )) . If K K K and L L L intersect, then K = L K=L K = L , and so s ′ = t ′ s'=t' s ′ = t ′ , a contradiction. Therefore, K K K and L L L are disjoint. Corollary 11.5 yields V ( L ) ∩ N ( K ) ⊆ V ( L ) ∩ sep ( s ) = ∅ V(L)\cap N(K)\subseteq V(L)\cap\operatorname{sep}(s)=\varnothing V ( L ) ∩ N ( K ) ⊆ V ( L ) ∩ sep ( s ) = ∅ .
Since V ( L ) V(L) V ( L ) is disjoint from V ( K ) ∪ N ( K ) V(K)\cup N(K) V ( K ) ∪ N ( K ) , we have V ( L ) ⊆ B ′ ∖ A ′ V(L)\subseteq B'\setminus A' V ( L ) ⊆ B ′ ∖ A ′ and C ′ = V ( L ) ∪ N ( L ) ⊆ B ′ C'=V(L)\cup N(L)\subseteq B' C ′ = V ( L ) ∪ N ( L ) ⊆ B ′ . By symmetry, we also have A ′ ⊆ D ′ A'\subseteq D' A ′ ⊆ D ′ . Therefore, s ′ = ( A ′ , B ′ ) ≤ ( D ′ , C ′ ) = ( t ′ ) ∗ s'=(A',B')\leq(D',C')=(t')^* s ′ = ( A ′ , B ′ ) ≤ ( D ′ , C ′ ) = ( t ′ ) ∗ . Hence, σ \sigma σ is a star. □
Lemma 11.7. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph such that ∣ V ( G ) ∣ ≥ ν 11.7 ( q , k ) : = 3 q + k |V(G)|\ge\nu_{11.7}(q,k):=3q+k ∣ V ( G ) ∣ ≥ ν 11.7 ( q , k ) := 3 q + k . Let X X X be a set of less than k k k vertices of G G G . Then, the union of all connected components of G − X G-X G − X of size ≤ q \le q ≤ q has size ≤ q \le q ≤ q .
Proof. By contradiction, suppose that more than q q q vertices belong to a connected component of G − X G-X G − X of size at most q q q . Denote these connected components by C 1 , … , C c C_1,\ldots,C_c C 1 , … , C c . Let i ∈ [ c ] i\in[c] i ∈ [ c ] be minimum such that ∣ C 1 ∪ ⋯ ∪ C i ∣ > q |C_1 \cup\cdots\cup C_i| > q ∣ C 1 ∪ ⋯ ∪ C i ∣ > q (which exists since ∣ C 1 ∪ ⋯ ∪ C c ∣ > q |C_1 \cup\cdots\cup C_c| > q ∣ C 1 ∪ ⋯ ∪ C c ∣ > q ). Then, ∣ C 1 ∪ ⋯ ∪ C i ∣ ≤ 2 q |C_1 \cup\cdots\cup C_i| \le2q ∣ C 1 ∪ ⋯ ∪ C i ∣ ≤ 2 q by minimality of i i i . Thus,
( V ( C 1 ∪ ⋯ ∪ C i ) ∪ X , V ( G − ( C 1 ∪ ⋯ ∪ C i ) ) ) \left(V(C_1 \cup\cdots\cup C_i) \cup X, V(G - (C_1 \cup\cdots\cup C_i))\right) ( V ( C 1 ∪ ⋯ ∪ C i ) ∪ X , V ( G − ( C 1 ∪ ⋯ ∪ C i )) )
is a ( < k ) (< k) ( < k ) -separation of G G G which is ( > q ) (> q) ( > q ) -broad since ∣ V ( G ) ∣ ≥ 3 q + k |V(G)| \ge3q + k ∣ V ( G ) ∣ ≥ 3 q + k . This contradicts that G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable. □
The next lemma will be instrumental in the proofs of Key Lemma 11.1, Theorem 1.2 and later Theorem 12.6 . Roughly speaking, it provides a way to “lift” separations of the torso of a star of ( β , k ) (\beta,k) ( β , k ) -irrelevant separations to separations of the entire graph.
Definition 11.8 (Lift). Let σ \sigma σ be a star of separations of a graph G G G . Let Z Z Z be the torso of σ \sigma σ . Let ( C , V ) (C,\mathcal{V}) ( C , V ) be a clique-faithful cactus-decomposition of Z Z Z . We define a cactus-decomposition ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) of G G G as follows. Let s ∈ σ s \in\sigma s ∈ σ be arbitrary. Since σ \sigma σ is a star, we have sep ( s ) ⊆ V ( Z ) \operatorname{sep}(s) \subseteq V(Z) sep ( s ) ⊆ V ( Z ) . Since Z Z Z is the torso of σ \sigma σ , sep ( s ) \operatorname{sep}(s) sep ( s ) is a clique in Z Z Z . Thus, there exists a node x ∈ V ( C ) x \in V(C) x ∈ V ( C ) that is not a P P P -node such that sep ( s ) ⊆ V x \operatorname{sep}(s) \subseteq V_x sep ( s ) ⊆ V x . Choose an arbitrary such node x ( s ) ∈ V ( C ) x(s) \in V(C) x ( s ) ∈ V ( C ) . We obtain C ^ \hat{C} C ^ from C C C by adding a new leaf t ( s ) t(s) t ( s ) for each s ∈ σ s \in\sigma s ∈ σ , joining each leaf t ( s ) t(s) t ( s ) to the node x ( s ) x(s) x ( s ) we chose for it. Then we obtain V ^ \hat{\mathcal{V}} V ^ by extending V \mathcal{V} V to the set of new leaves t ( s ) t(s) t ( s ) by letting V ^ t ( s ) : = A \hat{V}_{t(s)} := A V ^ t ( s ) := A for ( A , B ) : = s (A,B) := s ( A , B ) := s . We declare that the P P P -nodes in ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) are the same as in ( C , V ) (C,\mathcal{V}) ( C , V ) . Then ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) is a cactus-decomposition of G G G .
We say that ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) is a lift of ( C , V ) (C,\mathcal{V}) ( C , V ) from Z Z Z to G G G . Let S S S be the set of separations displayed by ( C , V ) (C,\mathcal{V}) ( C , V ) and let S ^ \hat{S} S ^ be the set of separations displayed in ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) by the cuts of C ^ \hat{C} C ^ that do not separate any new leaf from its neighbour. We say that S ^ \hat{S} S ^ is the lift of S S S from Z Z Z to G G G corresponding to ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) .
Given a cut ( U , W ) (U,W) ( U , W ) of C C C , there is a unique cut ( U ^ , W ^ ) (\hat{U},\hat{W}) ( U ^ , W ^ ) of C ^ \hat{C} C ^ that does not separate any new leaf from its neighbour and such that U ^ ∩ V ( C ) = U \hat{U} \cap V(C) = U U ^ ∩ V ( C ) = U and W ^ ∩ V ( C ) = W \hat{W} \cap V(C) = W W ^ ∩ V ( C ) = W . We call ( U ^ , W ^ ) (\hat{U},\hat{W}) ( U ^ , W ^ ) the lift of ( U , W ) (U,W) ( U , W ) from C C C to C ^ \hat{C} C ^ .
This definition is justified by the following result.
Lemma 11.9. Let σ \sigma σ be a star of separations of a graph G G G . Let Z Z Z be the torso of σ \sigma σ . Let ( C , V ) (C,\mathcal{V}) ( C , V ) be a clique-faithful cactus-decomposition of Z Z Z . Let ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) be a lift of ( C , V ) (C,\mathcal{V}) ( C , V ) from Z Z Z to G G G . Let S S S be the set of separations displayed by ( C , V ) (C,\mathcal{V}) ( C , V ) and let S ^ \hat{S} S ^ be the lift of S S S from Z Z Z to G G G corresponding to ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) . The following properties hold:
The set of separations displayed by ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) equals S ^ ∪ σ ∪ σ ∗ \hat{S} \cup\sigma\cup\sigma^* S ^ ∪ σ ∪ σ ∗ .
Let ( U , W ) (U,W) ( U , W ) be a cut of C C C , let ( U ^ , W ^ ) (\hat{U},\hat{W}) ( U ^ , W ^ ) be the lift of ( U , W ) (U,W) ( U , W ) from C C C to C ^ \hat{C} C ^ . Let s s s be the separation induced by ( U , W ) (U,W) ( U , W ) in ( C , V ) (C,\mathcal{V}) ( C , V ) , and let s ^ \hat{s} s ^ be the separation induced by ( U ^ , W ^ ) (\hat{U},\hat{W}) ( U ^ , W ^ ) in ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) . Then s ^ \hat{s} s ^ is a lift of s s s .
If C C C is a tree and x ∈ V ( C ) x \in V(C) x ∈ V ( C ) , then the torso at x x x in ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) is equal to the torso at x x x in ( C , V ) (C,\mathcal{V}) ( C , V ) .
Proof. (i). By construction of ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) , every separation in S ^ ∪ σ ∪ σ ∗ \hat{S} \cup\sigma\cup\sigma^* S ^ ∪ σ ∪ σ ∗ is displayed by ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) . Conversely, let s ^ \hat{s} s ^ be a separation displayed by ( C ^ , V ^ ) (\hat{C},\hat{\mathcal{V}}) ( C ^ , V ^ ) , and let ( U ^ , W ^ ) (\hat{U},\hat{W}) ( U ^ , W ^ ) be a corresponding cut of C ^ \hat{C} C ^ . If ( U ^ , W ^ ) (\hat{U},\hat{W}) ( U ^ , W ^ ) does not separate any new leaf from its neighbour then s ^ ∈ S ^ \hat{s} \in\hat{S} s ^ ∈ S ^ . Otherwise, since no cycle separates a leaf from its neighbour and since the neighbours of the new leaves are not P P P -nodes, ( U ^ , W ^ ) (\hat{U},\hat{W}) ( U ^ , W ^ ) is the cut between a new leaf t ( s ) t(s) t ( s ) for some s ∈ σ s \in\sigma s ∈ σ and the rest of C ^ \hat{C} C ^ . In that case, we have s ^ = s \hat{s} = s s ^ = s or s ^ = s ∗ \hat{s} = s^* s ^ = s ∗ . (ii). This is immediate since C ^ \widehat{C} C is built from C C C by adding leaves attached to nodes of C C C but without changing the bags of the nodes in C C C , and since ( C , V ) (C,\mathcal{V}) ( C , V ) is a cactus-decomposition of Z Z Z .
(iii). Since C C C is a tree and C ^ \widehat{C} C is obtained by adding leaves, both decomposition graphs are trees. Let W ^ \widehat{W} W be the torso at x x x in ( C ^ , V ^ ) (\widehat{C},\widehat{\mathcal{V}}) ( C , V ) and W W W be the torso at x x x in ( C , V ) (C,\mathcal{V}) ( C , V ) . Since ( C ^ , V ^ ) (\widehat{C},\widehat{\mathcal{V}}) ( C , V ) is a lift of ( C , V ) (C,\mathcal{V}) ( C , V ) from Z Z Z to G G G , we have V ( W ^ ) = V ^ x = V x = V ( W ) V(\widehat{W})=\widehat{V}_x=V_x=V(W) V ( W ) = V x = V x = V ( W ) , so it suffices to argue that W W W and W ^ \widehat{W} W have the same edges.
First we show E ( W ) ⊆ E ( W ^ ) E(W) \subseteq E(\widehat{W}) E ( W ) ⊆ E ( W ) . For this, let u v ∈ E ( W ) uv \in E(W) uv ∈ E ( W ) . Then, either u v ∈ E ( Z ) uv \in E(Z) uv ∈ E ( Z ) or there exists a node x ′ ≠ x ∈ V ( C ) x' \ne x \in V(C) x ′ = x ∈ V ( C ) such that { u , v } ⊆ V x ′ \{u,v\} \subseteq V_{x'} { u , v } ⊆ V x ′ . In the second case, we also have { u , v } ⊆ V ^ x ′ \{u,v\} \subseteq\widehat{V}_{x'} { u , v } ⊆ V x ′ so u v ∈ E ( W ^ ) uv \in E(\widehat{W}) uv ∈ E ( W ) . In the first case, either u v ∈ E ( G ) uv \in E(G) uv ∈ E ( G ) or there exists ( A , B ) ∈ σ (A,B) \in\sigma ( A , B ) ∈ σ such that { u , v } ⊆ A ∩ B \{u,v\} \subseteq A \cap B { u , v } ⊆ A ∩ B . If u v ∈ E ( G ) uv \in E(G) uv ∈ E ( G ) , we have u v ∈ E ( W ^ ) uv \in E(\widehat{W}) uv ∈ E ( W ) . If there exists ( A , B ) ∈ σ (A,B) \in\sigma ( A , B ) ∈ σ such that { u , v } ⊆ A ∩ B \{u,v\} \subseteq A \cap B { u , v } ⊆ A ∩ B , then there exists a node x ′ ≠ x ∈ V ( C ^ ) x' \ne x \in V(\widehat{C}) x ′ = x ∈ V ( C ) such that { u , v } ⊆ V ^ x ′ \{u,v\} \subseteq\widehat{V}_{x'} { u , v } ⊆ V x ′ so u v ∈ E ( W ^ ) uv \in E(\widehat{W}) uv ∈ E ( W ) .
Now, we show E ( W ^ ) ⊆ E ( W ) E(\widehat{W}) \subseteq E(W) E ( W ) ⊆ E ( W ) . For this, let u v ∈ E ( W ^ ) uv \in E(\widehat{W}) uv ∈ E ( W ) . Then, either u v ∈ E ( G ) uv \in E(G) uv ∈ E ( G ) or there exists a node x ′ ≠ x ∈ V ( C ^ ) x' \ne x \in V(\widehat{C}) x ′ = x ∈ V ( C ) such that { u , v } ⊆ V ^ x ′ \{u,v\} \subseteq\widehat{V}_{x'} { u , v } ⊆ V x ′ . If u v ∈ E ( G ) uv \in E(G) uv ∈ E ( G ) then u v ∈ E ( Z ) uv \in E(Z) uv ∈ E ( Z ) so u v ∈ E ( W ) uv \in E(W) uv ∈ E ( W ) . Thus, we can assume that there exists a node x ′ ≠ x ∈ V ( C ^ ) x' \ne x \in V(\widehat{C}) x ′ = x ∈ V ( C ) such that { u , v } ⊆ V ^ x ′ \{u,v\} \subseteq\widehat{V}_{x'} { u , v } ⊆ V x ′ . If x ′ x' x ′ was already a node of C C C then we have V ^ x ′ = V x ′ \widehat{V}_{x'}=V_{x'} V x ′ = V x ′ so there exists a node x ′ ≠ x ∈ V ( C ) x' \ne x \in V(C) x ′ = x ∈ V ( C ) such that { u , v } ⊆ V x ′ \{u,v\} \subseteq V_{x'} { u , v } ⊆ V x ′ , and thus u v ∈ E ( W ) uv \in E(W) uv ∈ E ( W ) . Otherwise, there exists ( A , B ) ∈ σ (A,B) \in\sigma ( A , B ) ∈ σ such that V ^ x ′ = A \widehat{V}_{x'}=A V x ′ = A , and thus { u , v } ⊆ A ∩ B \{u,v\} \subseteq A \cap B { u , v } ⊆ A ∩ B so u v ∈ E ( Z ) uv \in E(Z) uv ∈ E ( Z ) so u v ∈ E ( W ) uv \in E(W) uv ∈ E ( W ) .
This finishes the proof that W = W ^ W=\widehat{W} W = W . □
We now prove Key Lemma 11.1, which we restate for convenience.
Key Lemma 11.1. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let β > β 11.5 ( q , k ) \beta>\beta_{11.5}(q,k) β > β 11.5 ( q , k ) . Then the set σ \sigma σ of all ( β , k ) (\beta,k) ( β , k ) -irrelevant separations forms a star. Moreover, the torso of σ \sigma σ is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and ( β , k ) (\beta,k) ( β , k ) -narrow.
Proof of Key Lemma 11.1. We have that σ \sigma σ is a star by Lemma 11.6.
Let Z Z Z be the torso of σ \sigma σ . By Lemma 11.9 (ii), every ( > q ) (>q) ( > q ) -broad separation of Z Z Z lifts to a ( > q ) (>q) ( > q ) -broad separation of G G G of the same order. Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, Z Z Z is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable too. It remains to show that Z Z Z is ( β , k ) (\beta,k) ( β , k ) -narrow. Otherwise, by Lemma 11.2 , there would exist a ( β , k ) (\beta,k) ( β , k ) -irrelevant separation r ′ = ( C , D ) r'=(C,D) r ′ = ( C , D ) of Z Z Z . That is, there is a β \beta β -broad k k k -separation r r r of Z Z Z such that r ′ r' r ′ is a componental ( < k ) (<k) ( < k ) -separation at r r r . Consider a lift r ^ \widehat{r} r of r r r to G G G as in Lemma 11.9 (ii). Then r ^ \widehat{r} r is a β \beta β -broad k k k -separation of G G G . Since r ′ r' r ′ is ( β , k ) (\beta,k) ( β , k ) -irrelevant, C ∖ D C\setminus D C ∖ D is the vertex set of a connected component of Z − sep ( r ) Z-\operatorname{sep}(r) Z − sep ( r ) , and N ( C ∖ D ) = sep ( r ′ ) ⊆ sep ( r ) = sep ( r ^ ) N(C\setminus D)=\operatorname{sep}(r')\subseteq\operatorname{sep}(r)=\operatorname{sep}(\widehat{r}) N ( C ∖ D ) = sep ( r ′ ) ⊆ sep ( r ) = sep ( r ) . Thus, all vertices in C ∖ D C\setminus D C ∖ D are in the same component K K K in G − sep ( r ^ ) G-\operatorname{sep}(\widehat{r}) G − sep ( r ) , and N ( K ) = sep ( r ′ ) ⊆ sep ( r ^ ) N(K)=\operatorname{sep}(r')\subseteq\operatorname{sep}(\widehat{r}) N ( K ) = sep ( r ′ ) ⊆ sep ( r ) . Now, consider the separation s ^ : = ( V ( K ) ∪ sep ( r ′ ) , V ( G − K ) ) \widehat{s}:=(V(K)\cup\operatorname{sep}(r'),V(G-K)) s := ( V ( K ) ∪ sep ( r ′ ) , V ( G − K )) . Therefore, s ^ \widehat{s} s is a componental ( < k ) (<k) ( < k ) -separation at r ^ \widehat{r} r , so s ^ ∈ σ \widehat{s}\in\sigma s ∈ σ . Since Z Z Z is the torso of σ \sigma σ , this implies V ( K ) ∩ V ( Z ) = ∅ V(K)\cap V(Z)=\varnothing V ( K ) ∩ V ( Z ) = ∅ , so C ∖ D = ∅ C\setminus D=\varnothing C ∖ D = ∅ , contradicting that r ′ r' r ′ is proper. □
Lifting the cactus-decomposition
Having established Key Lemma 11.1, we now turn to the proof of Theorem 1.2.
First, we prove that broad separations of G G G project to broad separations in the torso of a star σ \sigma σ of irrelevant separations. For all integers n , k ∈ N n,k\in\mathbb{N} n , k ∈ N , we set ( n ≤ k ) : = ∑ i = 0 k ( n i ) \binom{n}{\leq k}:=\sum_{i=0}^{k}\binom{n}{i} ( ≤ k n ) := ∑ i = 0 k ( i n ) , where we use the convention ( n i ) = 0 \binom{n}{i}=0 ( i n ) = 0 for i > n i>n i > n .
Lemma 11.10. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let β > β 11.5 ( q , k ) \beta>\beta_{11.5}(q,k) β > β 11.5 ( q , k ) . Let α ∈ N \alpha\in\mathbb{N} α ∈ N and let α ′ ≥ α 11.10 ′ ( q , k , α ) : = q ⋅ ( α ≤ k − 1 ) + α \alpha'\ge\alpha'_{11.10}(q,k,\alpha):=q\cdot\binom{\alpha}{\le k-1}+\alpha α ′ ≥ α 11.10 ′ ( q , k , α ) := q ⋅ ( ≤ k − 1 α ) + α . Let σ \sigma σ be a star of ( β , k ) (\beta,k) ( β , k ) -irrelevant separations of G G G and let Z Z Z denote the torso of σ \sigma σ . Let s : = ( A , B ) s:=(A,B) s := ( A , B ) be a β \beta β -broad k k k -separation of G G G with ∣ A ∣ > α ′ |A|>\alpha' ∣ A ∣ > α ′ . Then, ∣ A ∩ V ( Z ) ∣ > α |A\cap V(Z)|>\alpha ∣ A ∩ V ( Z ) ∣ > α .
Proof. First, since s s s is β \beta β -broad and β > β 11.5 ( q , k ) = 4 q + k \beta>\beta_{11.5}(q,k)=4q+k β > β 11.5 ( q , k ) = 4 q + k , we have ∣ V ( G ) ∣ ≥ 3 q + k ≥ ν 11.7 ( q , k ) |V(G)|\ge3q+k\ge\nu_{11.7}(q,k) ∣ V ( G ) ∣ ≥ 3 q + k ≥ ν 11.7 ( q , k ) .
By contradiction, suppose that ∣ A ∩ V ( Z ) ∣ ≤ α |A\cap V(Z)|\le\alpha ∣ A ∩ V ( Z ) ∣ ≤ α . Let v ∈ A ∖ V ( Z ) v\in A\setminus V(Z) v ∈ A ∖ V ( Z ) . There exists a unique t : = ( C , D ) ∈ σ t:=(C,D)\in\sigma t := ( C , D ) ∈ σ such that v ∈ C ∖ D v\in C\setminus D v ∈ C ∖ D . Since ( C , D ) (C,D) ( C , D ) is ( β , k ) (\beta,k) ( β , k ) -irrelevant, we have ∣ sep ( t ) ∣ < k |\operatorname{sep}(t)|<k ∣ sep ( t ) ∣ < k and G [ C ∖ D ] G[C\setminus D] G [ C ∖ D ] is a connected component of G − sep ( t ) G-\operatorname{sep}(t) G − sep ( t ) . By Lemma 11.3 , we have ∣ C ∖ D ∣ ≤ q |C\setminus D|\le q ∣ C ∖ D ∣ ≤ q . By Corollary 11.5, C ∖ D C\setminus D C ∖ D is disjoint from sep ( s ) \operatorname{sep}(s) sep ( s ) , so C ∖ D ⊆ A ∖ B C\setminus D\subseteq A\setminus B C ∖ D ⊆ A ∖ B since v ∈ A v\in A v ∈ A .
Since t t t is ( β , k ) (\beta,k) ( β , k ) -irrelevant, we have sep ( t ) = N ( C ∖ D ) \operatorname{sep}(t)=N(C\setminus D) sep ( t ) = N ( C ∖ D ) , so sep ( t ) ⊆ A \operatorname{sep}(t)\subseteq A sep ( t ) ⊆ A , and sep ( t ) ⊆ V ( Z ) \operatorname{sep}(t)\subseteq V(Z) sep ( t ) ⊆ V ( Z ) since σ \sigma σ is a star. Then, sep ( t ) \operatorname{sep}(t) sep ( t ) is a set of less than k k k vertices in A ∩ V ( Z ) A\cap V(Z) A ∩ V ( Z ) such that C ∖ D C\setminus D C ∖ D is a connected component of G − sep ( t ) G-\operatorname{sep}(t) G − sep ( t ) with ≤ q \le q ≤ q vertices. As ∣ A ∩ V ( Z ) ∣ ≤ α |A\cap V(Z)|\le\alpha ∣ A ∩ V ( Z ) ∣ ≤ α by assumption, there are at most ( α ≤ k − 1 ) \binom{\alpha}{\le k-1} ( ≤ k − 1 α ) sets X X X of less than k k k vertices in A ∩ V ( Z ) A\cap V(Z) A ∩ V ( Z ) , and for each such set X X X , at most q q q vertices belong to a connected component of G − X G-X G − X of size at most q q q by Lemma 11.7. Thus, ∣ A ∖ V ( Z ) ∣ ≤ q ⋅ ( α ≤ k − 1 ) |A\setminus V(Z)|\le q\cdot\binom{\alpha}{\le k-1} ∣ A ∖ V ( Z ) ∣ ≤ q ⋅ ( ≤ k − 1 α ) . Therefore, ∣ A ∣ ≤ q ⋅ ( α ≤ k − 1 ) + α ≤ α ′ |A|\le q\cdot\binom{\alpha}{\le k-1}+\alpha\le\alpha' ∣ A ∣ ≤ q ⋅ ( ≤ k − 1 α ) + α ≤ α ′ , a contradiction. □
Lemma 11.11. Let q , k , β , γ , δ ∈ N q,k,\beta,\gamma,\delta\in\mathbb{N} q , k , β , γ , δ ∈ N with β > q \beta>q β > q and δ : = δ 11.11 ( q , k , γ ) : = max { γ + q ( k + γ ≤ k − 1 ) , ν 11.7 ( q , k ) − 1 } \delta:=\delta_{11.11}(q,k,\gamma):=\max\{\gamma+q\binom{k+\gamma}{\le k-1},\nu_{11.7}(q,k)-1\} δ := δ 11.11 ( q , k , γ ) := max { γ + q ( ≤ k − 1 k + γ ) , ν 11.7 ( q , k ) − 1 } . Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, let σ \sigma σ be a star of ( β , k ) (\beta,k) ( β , k ) -irrelevant separations of G G G and let Z Z Z denote the torso of σ \sigma σ . Let s , t s,t s , t be two ( ≤ k ) (\le k) ( ≤ k ) -separations of Z Z Z , and let s ^ , t ^ \hat{s},\hat{t} s ^ , t ^ be lifts of s , t s,t s , t to G G G , respectively. If s ≤ γ t s\le_\gamma t s ≤ γ t then s ^ ≤ δ t ^ \hat{s}\le_\delta\hat{t} s ^ ≤ δ t ^ .
Proof. If ∣ V ( G ) ∣ < ν 11.7 ( q , k ) |V(G)|<\nu_{11.7}(q,k) ∣ V ( G ) ∣ < ν 11.7 ( q , k ) then s ^ ≤ δ t ^ \hat{s}\le_\delta\hat{t} s ^ ≤ δ t ^ , so suppose ∣ V ( G ) ∣ ≥ ν 11.7 ( q , k ) |V(G)|\ge\nu_{11.7}(q,k) ∣ V ( G ) ∣ ≥ ν 11.7 ( q , k ) . This allows us to apply Lemma 11.7 later.
Let s ^ : = ( A ^ , B ^ ) \hat{s}:=(\hat{A},\hat{B}) s ^ := ( A ^ , B ^ ) and t ^ : = ( C ^ , D ^ ) \hat{t}:=(\hat{C},\hat{D}) t ^ := ( C ^ , D ^ ) and note that s = ( A ^ ∩ V ( Z ) , B ^ ∩ V ( Z ) ) s=(\hat{A}\cap V(Z),\hat{B}\cap V(Z)) s = ( A ^ ∩ V ( Z ) , B ^ ∩ V ( Z )) and t = ( C ^ ∩ V ( Z ) , D ^ ∩ V ( Z ) ) t=(\hat{C}\cap V(Z),\hat{D}\cap V(Z)) t = ( C ^ ∩ V ( Z ) , D ^ ∩ V ( Z )) by definition of lift. By symmetry, it suffices to show that ∣ A ^ ∖ C ^ ∣ ≤ δ |\hat{A}\setminus\hat{C}|\le\delta ∣ A ^ ∖ C ^ ∣ ≤ δ . Let v ∈ ( A ^ ∖ C ^ ) ∖ V ( Z ) v\in(\hat{A}\setminus\hat{C})\setminus V(Z) v ∈ ( A ^ ∖ C ^ ) ∖ V ( Z ) . There exists a unique ( E , F ) ∈ σ (E,F)\in\sigma ( E , F ) ∈ σ such that v ∈ E ∖ F v\in E\setminus F v ∈ E ∖ F . Since ( E ∖ F ) ∩ V ( Z ) = ∅ (E\setminus F)\cap V(Z)=\varnothing ( E ∖ F ) ∩ V ( Z ) = ∅ and A ^ ∩ B ^ ⊆ V ( Z ) \hat{A}\cap\hat{B}\subseteq V(Z) A ^ ∩ B ^ ⊆ V ( Z ) and C ^ ∩ D ^ ⊆ V ( Z ) \hat{C}\cap\hat{D}\subseteq V(Z) C ^ ∩ D ^ ⊆ V ( Z ) by definition of lift, it follows that E ∖ F E\setminus F E ∖ F is disjoint from both A ^ ∩ B ^ \hat{A}\cap\hat{B} A ^ ∩ B ^ and C ^ ∩ D ^ \hat{C}\cap\hat{D} C ^ ∩ D ^ . Then, v ∈ ( A ^ ∖ B ^ ) ∩ ( D ^ ∖ C ^ ) v\in(\hat{A}\setminus\hat{B})\cap(\hat{D}\setminus\hat{C}) v ∈ ( A ^ ∖ B ^ ) ∩ ( D ^ ∖ C ^ ) . Since G [ E ∖ F ] G[E\setminus F] G [ E ∖ F ] is connected, this implies E ∖ F ⊆ ( A ^ ∖ B ^ ) ∩ ( D ^ ∖ C ^ ) E\setminus F\subseteq(\hat{A}\setminus\hat{B})\cap(\hat{D}\setminus\hat{C}) E ∖ F ⊆ ( A ^ ∖ B ^ ) ∩ ( D ^ ∖ C ^ ) . Then, E ∩ F = N ( E ∖ F ) ⊆ A ^ ∩ D ^ E\cap F=N(E\setminus F)\subseteq\hat{A}\cap\hat{D} E ∩ F = N ( E ∖ F ) ⊆ A ^ ∩ D ^ . Furthermore, we have E ∩ F ⊆ V ( Z ) E\cap F\subseteq V(Z) E ∩ F ⊆ V ( Z ) , so E ∩ F E\cap F E ∩ F is a set of less than k k k vertices in A ^ ∩ D ^ ∩ V ( Z ) \hat{A}\cap\hat{D}\cap V(Z) A ^ ∩ D ^ ∩ V ( Z ) . However, ∣ A ^ ∩ D ^ ∩ V ( Z ) ∣ ≤ ∣ A ^ ∩ B ^ ∩ V ( Z ) ∣ + ∣ ( D ^ ∖ B ^ ) ∩ V ( Z ) ∣ ≤ k + γ |\hat{A}\cap\hat{D}\cap V(Z)|\le|\hat{A}\cap\hat{B}\cap V(Z)|+|(\hat{D}\setminus\hat{B})\cap V(Z)|\le k+\gamma ∣ A ^ ∩ D ^ ∩ V ( Z ) ∣ ≤ ∣ A ^ ∩ B ^ ∩ V ( Z ) ∣ + ∣ ( D ^ ∖ B ^ ) ∩ V ( Z ) ∣ ≤ k + γ since s ≤ γ t s\le_\gamma t s ≤ γ t implies ∣ ( D ^ ∖ B ^ ) ∩ V ( Z ) ∣ ≤ γ |(\hat{D}\setminus\hat{B})\cap V(Z)|\le\gamma ∣ ( D ^ ∖ B ^ ) ∩ V ( Z ) ∣ ≤ γ . Furthermore, G [ E ∖ F ] G[E\setminus F] G [ E ∖ F ] is a connected component of G − ( E ∩ F ) G-(E\cap F) G − ( E ∩ F ) of size at most q q q by Lemma 11.3 . Thus, v v v is one of the at most q q q vertices that belongs to a connected component of G − ( E ∩ F ) G-(E\cap F) G − ( E ∩ F ) of size at most q q q by Lemma 11.7. This yields ∣ ( A ^ ∖ C ^ ) ∖ V ( Z ) ∣ ≤ q ( k + γ ≤ k − 1 ) |(\hat{A}\setminus\hat{C})\setminus V(Z)|\le q\binom{k+\gamma}{\le k-1} ∣ ( A ^ ∖ C ^ ) ∖ V ( Z ) ∣ ≤ q ( ≤ k − 1 k + γ ) . Thus, ∣ A ^ ∖ C ^ ∣ = ∣ ( A ^ ∖ C ^ ) ∩ V ( Z ) ∣ + ∣ ( A ^ ∖ C ^ ) ∖ V ( Z ) ∣ ≤ γ + q ( k + γ ≤ k − 1 ) ≤ δ |\hat{A}\setminus\hat{C}|=|(\hat{A}\setminus\hat{C})\cap V(Z)|+|(\hat{A}\setminus\hat{C})\setminus V(Z)|\le\gamma+q\binom{k+\gamma}{\le k-1}\le\delta ∣ A ^ ∖ C ^ ∣ = ∣ ( A ^ ∖ C ^ ) ∩ V ( Z ) ∣ + ∣ ( A ^ ∖ C ^ ) ∖ V ( Z ) ∣ ≤ γ + q ( ≤ k − 1 k + γ ) ≤ δ . □
This implies that two separations of G G G which project to two close separations in the torso of σ \sigma σ must be close in G G G .
Lemma 11.12. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and β > q \beta>q β > q . Let γ ∈ N \gamma\in\mathbb{N} γ ∈ N and let δ ≥ δ 11.12 ( q , k , γ ) : = δ 11.11 ( q , k , γ ) \delta\ge\delta_{11.12}(q,k,\gamma):=\delta_{11.11}(q,k,\gamma) δ ≥ δ 11.12 ( q , k , γ ) := δ 11.11 ( q , k , γ ) . Let σ \sigma σ be a star of ( β , k ) (\beta,k) ( β , k ) -irrelevant separations of G G G and let Z Z Z denote the torso of σ \sigma σ . Let s , t s,t s , t be two γ \gamma γ -close ( ≤ k ) (\le k) ( ≤ k ) -separations of Z Z Z and let s ^ \hat{s} s ^ (resp. t ^ \hat{t} t ^ ) be a lift of s s s (resp. t t t ). Then, s ^ \hat{s} s ^ and t ^ \hat{t} t ^ are δ \delta δ -close.
Proof. Follows from Lemma 11.11 . □
We now have all the tools to prove the main result of this paper, which we restate for convenience.
Theorem 1.2 (Main theorem). For all q , k ∈ N q,k \in\mathbb{N} q , k ∈ N , there is d ∈ N d \in\mathbb{N} d ∈ N such that every ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph G G G admits a cactus-decomposition displaying ( ≤ k ) (\leq k) ( ≤ k ) -separations such that every ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G has edit-distance at most d d d to a displayed separation.
Proof. Given q , k ∈ N q,k \in\mathbb{N} q , k ∈ N , set
β : = β 11.5 ( q , k ) + 1 , \beta:= \beta_{11.5}(q,k)+1, β := β 11.5 ( q , k ) + 1 ,
η : = η 10.7 ( q , k , β ) , \eta:= \eta_{10.7}(q,k,\beta), η := η 10.7 ( q , k , β ) ,
ε : = max { α 11.10 ′ ( q , k , η + k ) + 1 , β } , and \varepsilon:= \max\{\alpha'_{11.10}(q,k,\eta+k)+1,\beta\}, \quad\text{and} ε := max { α 11.10 ′ ( q , k , η + k ) + 1 , β } , and
d : = max { ε + k , 2 δ 11.12 ( q , k , η ) } . d := \max\{\varepsilon+k,2\delta_{11.12}(q,k,\eta)\}. d := max { ε + k , 2 δ 11.12 ( q , k , η )} .
Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let σ \sigma σ be the set of all ( β , k ) (\beta,k) ( β , k ) -irrelevant separations. By Key Lemma 11.1, σ \sigma σ forms a star and its torso Z Z Z is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and ( β , k ) (\beta,k) ( β , k ) -narrow. By Theorem 10.7 , Z Z Z has a clique-faithful cactus-decomposition C ˇ = ( C ˇ , V ˇ ) \check{\mathcal{C}}=(\check{C},\check{\mathcal{V}}) C ˇ = ( C ˇ , V ˇ ) such that every displayed separation has order ≤ k \leq k ≤ k and the set of displayed separations of C ˇ \check{\mathcal{C}} C ˇ η \eta η -approximates the set of all k k k -separations of Z Z Z . Let S ˇ \check{S} S ˇ be the set of separations displayed by ( C ˇ , V ˇ ) (\check{C},\check{\mathcal{V}}) ( C ˇ , V ˇ ) . Let ( C , V ) (C,\mathcal{V}) ( C , V ) be a lift of ( C ˇ , V ˇ ) (\check{C},\check{\mathcal{V}}) ( C ˇ , V ˇ ) from Z Z Z to G G G , and let S S S be the lift of S ˇ \check{S} S ˇ from Z Z Z to G G G corresponding to ( C , V ) (C,\mathcal{V}) ( C , V ) . Then, ( C , V ) (C,\mathcal{V}) ( C , V ) is a cactus-decomposition of G G G such that every displayed separation has order ≤ k \leq k ≤ k .
It remains to prove that every ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G has edit-distance at most d d d to a separation in S S S . Let s s s be a ( ≤ k ) (\leq k) ( ≤ k ) -separation of G G G . If s s s is not ε \varepsilon ε -broad then s s s has edit-distance at most d d d to ( V ( G ) , ∅ ) (V(G),\varnothing) ( V ( G ) , ∅ ) or ( ∅ , V ( G ) ) (\varnothing,V(G)) ( ∅ , V ( G )) , both of which belong to S S S . We may therefore assume that s s s is ε \varepsilon ε -broad. Since ε ≥ β > q \varepsilon\geq\beta> q ε ≥ β > q and G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, s s s has order k k k . It follows from Corollary 11.5 that sep ( s ) ⊆ V ( Z ) \operatorname{sep}(s) \subseteq V(Z) sep ( s ) ⊆ V ( Z ) , so s s s induces a k k k -separation s ˇ \check{s} s ˇ on Z Z Z (i.e. s s s is a lift of s ˇ \check{s} s ˇ from Z Z Z ). By Lemma 11.10 , since β > β 11.5 ( q , k ) \beta> \beta_{11.5}(q,k) β > β 11.5 ( q , k ) and ε ≥ max { α 11.10 ′ ( q , k , η + k ) + 1 , β } \varepsilon\geq\max\{\alpha'_{11.10}(q,k,\eta+k)+1,\beta\} ε ≥ max { α 11.10 ′ ( q , k , η + k ) + 1 , β } , it follows that s ˇ \check{s} s ˇ is an η \eta η -broad k k k -separation of Z Z Z . By Theorem 10.7 , s ˇ \check{s} s ˇ is η \eta η -close to a displayed separation s ˇ ′ \check{s}' s ˇ ′ of ( C ˇ , V ˇ ) (\check{C},\check{\mathcal{V}}) ( C ˇ , V ˇ ) . By Lemma 11.9 (ii), there exists a lift s ′ ∈ S s' \in S s ′ ∈ S of s ˇ ′ \check{s}' s ˇ ′ . By Lemma 11.12 , since β = β 11.5 ( q , k ) + 1 > q \beta=\beta_{11.5}(q,k)+1>q β = β 11.5 ( q , k ) + 1 > q , s s s and s ′ s' s ′ are δ 11.12 ( q , k , η ) \delta_{11.12}(q,k,\eta) δ 11.12 ( q , k , η ) -close. By Observation 4.11 , their edit-distance is therefore at most 2 δ 11.12 ( q , k , η ) ≤ d 2\delta_{11.12}(q,k,\eta)\leq d 2 δ 11.12 ( q , k , η ) ≤ d . □
An approximate Tutte-type decomposition
Finally, we extend Theorem 10.4 to all ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graphs by dropping the assumption that G G G is ( b , k ) (b,k) ( b , k ) -narrow. This establishes Theorem 2.6 .
Lifting flowers
We start by proving some results to lift graphic-k k k -flowers.
Lemma 12.1. Let G G G be a graph, let σ \sigma σ be a star of separations of G G G and let Z Z Z denote the torso of σ \sigma σ . Let F = ( O , V ) F=(O,\mathcal{V}) F = ( O , V ) be a graphic-k k k -daisy of Z Z Z that has the clover-property. Let X : = ⋂ v ∈ V ( O ) V v X:=\bigcap_{v\in V(O)}V_v X := ⋂ v ∈ V ( O ) V v be the centre of F F F . Let s = ( C , D ) ∈ σ s=(C,D)\in\sigma s = ( C , D ) ∈ σ be such that sep ( s ) ⊈ X \operatorname{sep}(s)\not\subseteq X sep ( s ) ⊆ X . Then, the set of vertices x ∈ V ( O ) x\in V(O) x ∈ V ( O ) such that sep ( s ) ⊆ V x \operatorname{sep}(s)\subseteq V_x sep ( s ) ⊆ V x induces a nonempty subpath of $O.
Proof. Since σ \sigma σ is a star, we have sep ( s ) ⊆ V ( Z ) \operatorname{sep}(s) \subseteq V(Z) sep ( s ) ⊆ V ( Z ) . Since Z Z Z is the torso of σ \sigma σ , we have that Z [ sep ( s ) ] Z[\operatorname{sep}(s)] Z [ sep ( s )] is a complete graph. The lemma then follows from Lemma 10.6. □ \square □
The next lemma (Lemma 12.4 ) is very close to Lemma 11.9 . Essentially, it says that graphic-k k k -flowers of the torso of σ \sigma σ can be lifted to graphic-k k k -flowers of G G G . When lifting a graphic-k k k -daisy, each component outside the torso whose neighbourhood lies in the centre can be attached to any bag. Taking all possible assignments yields a canonical family of lifts whose members differ only in the placement of these small components.
Definition 12.2 ((q , k q,k q , k )-coherent). Let q , k ∈ N q,k \in\mathbb{N} q , k ∈ N . Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. Let F \mathcal{F} F be a set of graphic-k k k -flowers of G G G , all with the same centre X X X . Let R R R be the set of all vertices of G G G that belong to some connected component K K K of G − X G-X G − X of size ∣ V ( K ) ∣ ≤ q |V(K)| \leq q ∣ V ( K ) ∣ ≤ q such that ∣ N G ( K ) ∣ < k |N_G(K)| < k ∣ N G ( K ) ∣ < k . For every F ∈ F F \in\mathcal{F} F ∈ F , let F ~ \widetilde{F} F be the graph-decomposition of G − R G-R G − R obtained from F F F by removing the vertices of R R R from the bags that contain them. We say that F \mathcal{F} F is ( q , k ) (q,k) ( q , k ) -coherent if F ~ = F ~ ′ \widetilde{F} = \widetilde{F}' F = F ′ for all F , F ′ ∈ F F,F' \in\mathcal{F} F , F ′ ∈ F .
Lemma 12.4. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph and let β > β 11.5 ( q , k ) \beta> \beta_{11.5}(q,k) β > β 11.5 ( q , k ) . Let σ \sigma σ denote the star of all ( β , k ) (\beta,k) ( β , k ) -irrelevant separations of G G G (see Key Lemma 11.1). Let Z Z Z denote the torso of σ \sigma σ . Let F F F be a graphic-k k k -flower of Z Z Z , such that F F F has the clover-property if it is a graphic-k k k -daisy. Then, there exists a nonempty ( q , k ) (q,k) ( q , k ) -coherent set F ^ \widehat{\mathcal{F}} F of graphic-k k k -flowers of G G G with the following properties:
(i) For every F ^ ∈ F ^ \widehat{F} \in\widehat{\mathcal{F}} F ∈ F , every s ^ ∈ D ( F ^ ) \widehat{s} \in D(\widehat{F}) s ∈ D ( F ) induces a separation s s s of Z Z Z . If s s s is proper then s ∈ D ( F ) s \in D(F) s ∈ D ( F ) . Moreover, if s ^ \widehat{s} s is a petal-separation of F ^ \widehat{F} F and s s s is proper, then s s s is a petal-separation of F F F .
(ii) For every F ^ ∈ F ^ \widehat{F} \in\widehat{\mathcal{F}} F ∈ F , every separation s ∈ D ( F ) s \in D(F) s ∈ D ( F ) has a lift s ^ ∈ D ( F ^ ) \widehat{s} \in D(\widehat{F}) s ∈ D ( F ) . Moreover, if s s s is a petal-separation of F F F , then we can choose s ^ \widehat{s} s as a petal-separation of F ^ \widehat{F} F .
(iii) F ^ = F ^ ( F ) \widehat{\mathcal{F}} = \widehat{\mathcal{F}}(F) F = F ( F ) is uniquely determined by F F F (given G G G ). Furthermore, if ψ \psi ψ is an automorphism of G G G and F ′ F' F ′ is a graphic-k k k -flower of Z Z Z such that F F F and F ′ F' F ′ are ψ ↾ Z \psi\!\upharpoonright_Z ψ ↾ Z -isomorphic, then F ^ ( F ) \widehat{\mathcal{F}}(F) F ( F ) and F ^ ( F ′ ) \widehat{\mathcal{F}}(F') F ( F ′ ) are ψ \psi ψ -isomorphic.
Proof. Suppose first that F = ( S , V ) F=(S,\mathcal{V}) F = ( S , V ) is a graphic-k k k -anemone. Let c c c be the centre of S S S . We define a graphic-k k k -anemone F ^ = ( S ^ , V ^ ) \widehat{F}=(\widehat{S},\widehat{\mathcal{V}}) F = ( S , V ) of G G G as follows. Start with ( S ^ , V ^ ) = ( S , V ) (\widehat{S},\widehat{\mathcal{V}})=(S,\mathcal{V}) ( S , V ) = ( S , V ) and let s : = ( C , D ) ∈ σ s:=(C,D)\in\sigma s := ( C , D ) ∈ σ . Since σ \sigma σ is a star, we have sep ( s ) ⊆ V ( Z ) \operatorname{sep}(s) \subseteq V(Z) sep ( s ) ⊆ V ( Z ) . Since Z Z Z is the torso of σ \sigma σ , Z [ sep ( s ) ] Z[\operatorname{sep}(s)] Z [ sep ( s )] is a complete graph. If sep ( s ) ⊆ V c \operatorname{sep}(s) \subseteq V_c sep ( s ) ⊆ V c , we add a new leaf t ( s ) t(s) t ( s ) to S ^ \widehat{S} S with c c c as its only neighbour, and set V t ( s ) = C ∪ V c V_{t(s)}=C\cup V_c V t ( s ) = C ∪ V c . Otherwise, there exists a unique leaf ℓ ( s ) ∈ V ( S ) \ell(s)\in V(S) ℓ ( s ) ∈ V ( S ) such that sep ( s ) ⊆ V ℓ ( s ) \operatorname{sep}(s)\subseteq V_{\ell(s)} sep ( s ) ⊆ V ℓ ( s ) , and we add C ∖ D C\setminus D C ∖ D to V ^ ℓ ( s ) \widehat{V}_{\ell(s)} V ℓ ( s ) . After iterating this process over all separations s ∈ σ s\in\sigma s ∈ σ , we eventually obtain a graphic-k k k -anemone F ^ = ( S ^ , V ^ ) \widehat{F}=(\widehat{S},\widehat{\mathcal{V}}) F = ( S , V ) of G G G . We set F ^ : = { F ^ } \widehat{\mathcal{F}}:=\{\widehat{F}\} F := { F } . The set F ^ \widehat{\mathcal{F}} F is trivially a nonempty ( q , k ) (q,k) ( q , k ) -coherent set of graphic-k k k -flowers of $G.
(i). Let s ^ = ( A ^ , B ^ ) ∈ D ( F ^ ) \widehat{s}=(\widehat{A},\widehat{B})\in D(\widehat{F}) s = ( A , B ) ∈ D ( F ) . There is a cut ( U ^ , W ^ ) (\widehat{U},\widehat{W}) ( U , W ) of S ^ \widehat{S} S such that A ^ = V ^ ↾ U ^ \widehat{A}=\widehat{\mathcal{V}}\upharpoonright\widehat{U} A = V ↾ U and B ^ = V ^ ↾ W ^ \widehat{B}=\widehat{\mathcal{V}}\upharpoonright\widehat{W} B = V ↾ W . If U ^ = ∅ \widehat{U}=\varnothing U = ∅ or W ^ = ∅ \widehat{W}=\varnothing W = ∅ , then s ^ \widehat{s} s induces ( ∅ , V ( Z ) ) (\varnothing,V(Z)) ( ∅ , V ( Z )) or ( V ( Z ) , ∅ ) (V(Z),\varnothing) ( V ( Z ) , ∅ ) , respectively, so the assertion holds. Hence assume that U ^ \widehat{U} U and W ^ \widehat{W} W are nonempty. Let U : = U ^ ∩ V ( S ) U := \widehat{U} \cap V(S) U := U ∩ V ( S ) and W : = W ^ ∩ V ( S ) W := \widehat{W} \cap V(S) W := W ∩ V ( S ) . Let A : = V ↾ U A := \mathcal{V} \mathord{\upharpoonright} U A := V ↾ U , B : = V ↾ W B := \mathcal{V} \mathord{\upharpoonright} W B := V ↾ W and s : = ( A , B ) s := (A,B) s := ( A , B ) . We distinguish two cases.
In the first case, both U U U and W W W are nonempty. Then s ^ \widehat{s} s induces ( A , B ) (A,B) ( A , B ) , which is displayed by F F F . Moreover, if s ^ \widehat{s} s is a petal-separation, then U ^ \widehat{U} U consists of a single node, and so does U U U , hence s s s is a petal-separation as well.
In the second case, U U U or W W W is empty, say U = ∅ U = \varnothing U = ∅ . Then U ^ = { t ( r ) : r ∈ σ ′ } \widehat{U} = \{t(r) : r \in\sigma'\} U = { t ( r ) : r ∈ σ ′ } for some subset σ ′ ⊆ σ \sigma' \subseteq\sigma σ ′ ⊆ σ . In this case, s ^ \widehat{s} s induces ( V c , V ( Z ) ) (V_c, V(Z)) ( V c , V ( Z )) , which is a petal-separation of F F F .
(ii). The proof is straightforward, as in Lemma 11.9 (ii).
(iii). This is immediate since the construction of F ^ \widehat{F} F involves no arbitrary choices.
Suppose now that F = ( O , V ) F = (O,\mathcal{V}) F = ( O , V ) is a graphic-k k k -daisy. Let O 1 O_1 O 1 be the 1-subdivision of the cycle O O O . For every edge e ∈ E ( O ) e \in E(O) e ∈ E ( O ) , denote by d ( e ) d(e) d ( e ) the corresponding vertex of O 1 O_1 O 1 . We extend V \mathcal{V} V to all vertices d ( e ) ∈ V ( O 1 ) ∖ V ( O ) d(e) \in V(O_1) \setminus V(O) d ( e ) ∈ V ( O 1 ) ∖ V ( O ) by letting V d ( e ) : = V e \mathcal{V}_{d(e)} := \mathcal{V}_e V d ( e ) := V e , where V e \mathcal{V}_e V e is the adhesion-set at e e e . Then, ( O 1 , V ) (O_1,\mathcal{V}) ( O 1 , V ) is a graphic-k k k -daisy of Z Z Z . Let X : = ⋂ x ∈ V ( O ) V x X := \bigcap_{x \in V(O)} V_x X := ⋂ x ∈ V ( O ) V x , and note that X = ⋂ x ∈ V ( O 1 ) V x X = \bigcap_{x \in V(O_1)} V_x X = ⋂ x ∈ V ( O 1 ) V x , so the centre remains unchanged. Since F F F has the clover-property, we have ∣ X ∣ < k |X| < k ∣ X ∣ < k .
We now describe an algorithm to build a graphic-k k k -daisy F ^ = ( O 1 , V ^ ) \widehat{F} = (O_1,\widehat{\mathcal{V}}) F = ( O 1 , V ) of G G G . Start with ( O 1 , V ^ ) = ( O 1 , V ) (O_1,\widehat{\mathcal{V}}) = (O_1,\mathcal{V}) ( O 1 , V ) = ( O 1 , V ) and let s : = ( C , D ) ∈ σ s := (C,D) \in\sigma s := ( C , D ) ∈ σ . Since σ \sigma σ is a star, we have sep ( s ) ⊆ V ( Z ) \operatorname{sep}(s) \subseteq V(Z) sep ( s ) ⊆ V ( Z ) . Since Z Z Z is the torso of σ \sigma σ , Z [ sep ( s ) ] Z[\operatorname{sep}(s)] Z [ sep ( s )] is a complete graph. If sep ( s ) ⊆ X \operatorname{sep}(s) \subseteq X sep ( s ) ⊆ X , let x ∈ V ( O 1 ) x \in V(O_1) x ∈ V ( O 1 ) be an arbitrary vertex, and add C ∖ D C \setminus D C ∖ D to V ^ x \widehat{\mathcal{V}}_x V x . Otherwise, by Lemma 12.1 , the set of vertices x ∈ V ( O ) x \in V(O) x ∈ V ( O ) such that sep ( s ) ⊆ V x \operatorname{sep}(s) \subseteq V_x sep ( s ) ⊆ V x induces a subpath of O O O . Thus, the set of vertices x ∈ V ( O 1 ) x \in V(O_1) x ∈ V ( O 1 ) such that sep ( s ) ⊆ V x \operatorname{sep}(s) \subseteq V_x sep ( s ) ⊆ V x induces a path in O 1 O_1 O 1 with an odd number of vertices. Let y y y be the vertex at the middle of that path, and add C ∖ D C \setminus D C ∖ D to V ^ y \widehat{\mathcal{V}}_y V y .
After iterating this process over all separations s ∈ σ s \in\sigma s ∈ σ , we eventually obtain a cycle-decomposition F ^ = ( O 1 , V ^ ) \widehat{F} = (O_1,\widehat{\mathcal{V}}) F = ( O 1 , V ) of G G G . The construction of F ^ \widehat{F} F involves some arbitrary choices. We then define F ^ \widehat{\mathcal{F}} F as the (nonempty) set of all graphic-k k k -flowers F ^ \widehat{F} F of G G G that can be built by this algorithm. Note that every F ^ ∈ F ^ \widehat{F} \in\widehat{\mathcal{F}} F ∈ F has centre X X X . Let R R R be the set of all vertices of G G G that belong to some connected component K K K of G − X G-X G − X of size ∣ V ( K ) ∣ ≤ q |V(K)| \leq q ∣ V ( K ) ∣ ≤ q such that ∣ N G ( K ) ∣ < k |N_G(K)| < k ∣ N G ( K ) ∣ < k . If s : = ( C , D ) ∈ σ s := (C,D) \in\sigma s := ( C , D ) ∈ σ is such that sep ( s ) ⊆ X \operatorname{sep}(s) \subseteq X sep ( s ) ⊆ X , then G [ C ∖ D ] G[C \setminus D] G [ C ∖ D ] is a connected component of G − X G-X G − X of size at most q q q by Lemma 11.3 , and ∣ N G ( C ∖ D ) ∣ ≤ ∣ X ∣ < k |N_G(C \setminus D)| \leq|X| < k ∣ N G ( C ∖ D ) ∣ ≤ ∣ X ∣ < k , so C ∖ D ⊆ R C \setminus D \subseteq R C ∖ D ⊆ R . Thus, we only make arbitrary choices regarding vertices in R R R . This shows that F ^ \widehat{\mathcal{F}} F is a nonempty ( q , k ) (q,k) ( q , k ) -coherent set of graphic-k k k -flowers of G G G .
The proofs of (i) and (ii) are similar to the proofs for the case where F F F is a graphic-k k k -anemone, so we do not reproduce them.
(iii). Since F ^ \widehat{\mathcal{F}} F contains all the graphic-k k k -flowers that can be built by the above algorithm, F ^ = F ^ ( F ) \widehat{\mathcal{F}} = \widehat{\mathcal{F}}(F) F = F ( F ) is uniquely determined by F F F (given G G G ). Let ψ \psi ψ be an automorphism of G G G and F ′ : = ( O ′ , V ′ ) F' := (O',\mathcal{V}') F ′ := ( O ′ , V ′ ) be a graphic-k k k -flower of Z Z Z such that F F F and F ′ F' F ′ are ψ ↾ Z \psi\mathbin{\upharpoonright} Z ψ ↾ Z -isomorphic. Let Φ : V ( O ) → V ( O ′ ) \Phi\colon V(O) \to V(O') Φ : V ( O ) → V ( O ′ ) be an isomorphism such that V Φ ( o ) ′ = ψ ( V o ) \mathcal{V}'_{\Phi(o)} = \psi(V_o) V Φ ( o ) ′ = ψ ( V o ) for every o ∈ V ( O ) o \in V(O) o ∈ V ( O ) . Then, Φ \Phi Φ naturally extends to Φ : V ( O 1 ) → V ( O 1 ′ ) \Phi\colon V(O_1) \to V(O'_1) Φ : V ( O 1 ) → V ( O 1 ′ ) . Note that the centre X ′ X' X ′ of F ′ F' F ′ satisfies X ′ = ψ ( X ) X' = \psi(X) X ′ = ψ ( X ) . Let F ^ : = ( O 1 , V ^ ) ∈ F ^ ( F ) \widehat{F} := (O_1,\widehat{\mathcal{V}}) \in\widehat{\mathcal{F}}(F) F := ( O 1 , V ) ∈ F ( F ) . We show that the graphic-k k k -flower F ^ ′ = ( O 1 ′ , V ^ ′ ) \widehat{F}' = (O'_1,\widehat{\mathcal{V}}') F ′ = ( O 1 ′ , V ′ ) of ψ ( G ) \psi(G) ψ ( G ) with V ^ Φ ( o ) ′ = ψ ( V ^ o ) \widehat{\mathcal{V}}'_{\Phi(o)} = \psi(\widehat{\mathcal{V}}_o) V Φ ( o ) ′ = ψ ( V o ) for every o ∈ V ( O 1 ) o \in V(O_1) o ∈ V ( O 1 ) belongs to F ^ ( F ′ ) \widehat{\mathcal{F}}(F') F ( F ′ ) . Let s : = ( C , D ) ∈ σ s := (C,D) \in\sigma s := ( C , D ) ∈ σ . Since σ \sigma σ is canonical, we have that s ′ : = ( ψ ( C ) , ψ ( D ) ) s' := (\psi(C),\psi(D)) s ′ := ( ψ ( C ) , ψ ( D )) is a ( β , k ) (\beta,k) ( β , k ) -irrelevant separation of ψ ( G ) \psi(G) ψ ( G ) .
If sep ( s ) ⊆ X \operatorname{sep}(s) \subseteq X sep ( s ) ⊆ X , then C ∖ D C \setminus D C ∖ D was added to an arbitrary bag V ^ o \widehat{V}_o V o by the algorithm. In that case, sep ( s ′ ) = ψ ( sep ( s ) ) ⊆ ψ ( X ) = X ′ \operatorname{sep}(s') = \psi(\operatorname{sep}(s)) \subseteq\psi(X) = X' sep ( s ′ ) = ψ ( sep ( s )) ⊆ ψ ( X ) = X ′ , so the algorithm can add ψ ( C ) ∖ ψ ( D ) \psi(C) \setminus\psi(D) ψ ( C ) ∖ ψ ( D ) to the bag V ^ Φ ( o ) ′ \widehat{V}'_{\Phi(o)} V Φ ( o ) ′ . If sep ( s ) ⊈ X \operatorname{sep}(s) \not\subseteq X sep ( s ) ⊆ X , then the set P s P_s P s of vertices x ∈ V ( O 1 ) x \in V(O_1) x ∈ V ( O 1 ) such that sep ( s ) ⊆ V x \operatorname{sep}(s) \subseteq V_x sep ( s ) ⊆ V x induces a path in O 1 O_1 O 1 with an odd number of vertices, with middle vertex y y y say. Then, we also have sep ( s ′ ) = ψ ( sep ( s ) ) ⊈ ψ ( X ) = X ′ \operatorname{sep}(s') = \psi(\operatorname{sep}(s)) \not\subseteq\psi(X) = X' sep ( s ′ ) = ψ ( sep ( s )) ⊆ ψ ( X ) = X ′ , and the set P s ′ ′ P_{s'}' P s ′ ′ of vertices x ′ ∈ V ( O 1 ′ ) x' \in V(O_1') x ′ ∈ V ( O 1 ′ ) such that sep ( s ′ ) ⊆ V x ′ ′ \operatorname{sep}(s') \subseteq V_{x'}' sep ( s ′ ) ⊆ V x ′ ′ induces a path in O 1 ′ O_1' O 1 ′ with an odd number of vertices, with middle vertex y ′ y' y ′ say. However, Φ : V ( O 1 ) → V ( O 1 ′ ) \Phi: V(O_1) \to V(O_1') Φ : V ( O 1 ) → V ( O 1 ′ ) satisfies V Φ ( o ) ′ = ψ ( V o ) V_{\Phi(o)}' = \psi(V_o) V Φ ( o ) ′ = ψ ( V o ) for every o ∈ V ( O ) o \in V(O) o ∈ V ( O ) , so P s ′ ′ = Φ ( P s ) P_{s'}' = \Phi(P_s) P s ′ ′ = Φ ( P s ) , and thus y ′ = Φ ( y ) y' = \Phi(y) y ′ = Φ ( y ) . In that case, the algorithm adds C ∖ D C \setminus D C ∖ D to V ^ y \widehat{V}_y V y to build F ^ \widehat{F} F from F F F , and adds ψ ( C ) ∖ ψ ( D ) \psi(C) \setminus\psi(D) ψ ( C ) ∖ ψ ( D ) to V ^ y ′ ′ \widehat{V}_{y'}' V y ′ ′ to build F ^ ′ \widehat{F}' F ′ from F ′ F' F ′ . Thus, the algorithm for F ′ F' F ′ can build the graphic-k k k -flower F ^ ′ : = ( O 1 ′ , V ^ ′ ) \widehat{F}' := (O_1', \widehat{\mathcal{V}}') F ′ := ( O 1 ′ , V ′ ) such that V ^ Φ ( o ) ′ = ψ ( V ^ o ) \widehat{V}_{\Phi(o)}' = \psi(\widehat{V}_o) V Φ ( o ) ′ = ψ ( V o ) for every o ∈ V ( O 1 ) o \in V(O_1) o ∈ V ( O 1 ) . Then, F ^ ∈ F ^ ( F ) \widehat{F} \in\widehat{\mathcal{F}}(F) F ∈ F ( F ) is ψ \psi ψ -isomorphic to F ^ ′ ∈ F ^ ( F ′ ) \widehat{F}' \in\widehat{\mathcal{F}}(F') F ′ ∈ F ( F ′ ) . It then follows immediately that F ^ ( F ) \widehat{\mathcal{F}}(F) F ( F ) and F ^ ( F ′ ) \widehat{\mathcal{F}}(F') F ( F ′ ) are ψ \psi ψ -isomorphic.
We refer to the set F ^ \widehat{\mathcal{F}} F in Lemma 12.4 as the set of lifts of F F F . This is justified by Lemma 12.4 (iii).
Lifting the approximate Tutte-type decomposition
We now have all the tools to prove the Tutte-type version of our results. We need one last definition to explain in what way the flowers in Theorem 2.6 are canonical.
Definition 12.5 ((q , k q,k q , k )-canonical). Let q , k ∈ N q,k \in\mathbb{N} q , k ∈ N . Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph, and let G \mathcal{G} G be a collection of sets of graphic-k k k -flowers of G G G . We say that G \mathcal{G} G is ( q , k ) (q,k) ( q , k ) -canonical if the following two conditions are satisfied:
Every set F ∈ G \mathcal{F} \in\mathcal{G} F ∈ G of graphic-k k k -flowers is ( q , k ) (q,k) ( q , k ) -coherent.
For every automorphism ψ \psi ψ of G G G , there is a permutation Φ \Phi Φ of G \mathcal{G} G such that F \mathcal{F} F and Φ ( F ) \Phi(\mathcal{F}) Φ ( F ) are ψ \psi ψ -isomorphic for every F ∈ G \mathcal{F} \in\mathcal{G} F ∈ G .
The following is a strengthening of Theorem 2.6 .
Theorem 12.6. For every q , k ∈ N q,k \in\mathbb{N} q , k ∈ N there are γ , κ , ϱ , q ′ , ε ∈ N \gamma,\kappa,\varrho,q',\varepsilon\in\mathbb{N} γ , κ , ϱ , q ′ , ε ∈ N such that the following holds. Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. There exists a collection G \mathcal{G} G of nonempty sets of graphic-k k k -flowers of G G G and a symmetric nested set M M M of totally-γ \gamma γ -nested proper ( ≤ k ) (\leq k) ( ≤ k ) -separations of G G G with corresponding tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) of G G G and a bipartition X ⊔ Y = V ( T ) X \sqcup Y = V(T) X ⊔ Y = V ( T ) such that:
Every node t ∈ X t \in X t ∈ X has a ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable torso.
Every node t ∈ Y t \in Y t ∈ Y comes with a set F t ∈ G \mathcal{F}_t \in\mathcal{G} F t ∈ G of graphic-k k k -flowers of G G G such that every F ∈ F t F \in\mathcal{F}_t F ∈ F t ϱ \varrho ϱ -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t .
For every ε \varepsilon ε -broad k k k -separation s s s of G G G , either s s s is ε \varepsilon ε -close to a separation in M M M or there is t ∈ Y t \in Y t ∈ Y such that for every F ∈ F t F \in\mathcal{F}_t F ∈ F t , s s s is ε \varepsilon ε -close to some displayed separation of F F F .
The set G \mathcal{G} G is ( q , k ) (q,k) ( q , k ) -canonical. Both the tree-decomposition ( T , V ) (T,\mathcal{V}) ( T , V ) and the set M M M are κ \kappa κ -canonical. Given M M M and ( T , V ) (T,\mathcal{V}) ( T , V ) , the bipartition X ⊔ Y = V ( T ) X \sqcup Y = V(T) X ⊔ Y = V ( T ) is canonical and the assignment Y ∋ t ↦ F t ∈ G Y \ni t \mapsto\mathcal{F}_t \in\mathcal{G} Y ∋ t ↦ F t ∈ G is a canonical bijection.
Proof. Let b : = β 11.5 ( q , k ) + 1 b := \beta_{11.5}(q,k) + 1 b := β 11.5 ( q , k ) + 1 . Let γ 0 \gamma_0 γ 0 , κ 0 \kappa_0 κ 0 , ϱ 0 \varrho_0 ϱ 0 , q 0 ′ q'_0 q 0 ′ , ε 0 ∈ N \varepsilon_0 \in\mathbb{N} ε 0 ∈ N be the constants obtained by applying Theorem 10.4 for q q q , k k k , b b b . Let
γ 1 : = max { γ 0 , b } , ε 1 : = max { q + 1 , ε 0 } , γ : = max { b + k , α 11.10 ′ ( q , k , γ 1 + k ) + k } , κ : = δ 11.12 ( q , k , κ 0 + k ) + 1 , ϱ : = max { b , α 11.10 ′ ( q , k , ϱ 0 + k ) + 1 , δ 11.12 ( q , k , ϱ 0 ) } , q ′ : = max { ν 11.7 ( q , k ) , q 0 ′ } , and ε : = max { ν 11.7 ( q , k ) , α 11.10 ′ ( q , k , ε 1 + k ) + 1 , δ 11.12 ( q , k , ε 0 ) } . \begin{aligned}
\gamma_1 &:= \max\{\gamma_0,b\},\\
\varepsilon_1 &:= \max\{q+1,\varepsilon_0\},\\
\gamma&:= \max\{b+k,\alpha'_{11.10}(q,k,\gamma_1+k)+k\},\\
\kappa&:= \delta_{11.12}(q,k,\kappa_0+k)+1,\\
\varrho&:= \max\{b,\alpha'_{11.10}(q,k,\varrho_0+k)+1,\delta_{11.12}(q,k,\varrho_0)\},\\
q' &:= \max\{\nu_{11.7}(q,k),q'_0\},\ \text{and}\\
\varepsilon&:= \max\{\nu_{11.7}(q,k),\alpha'_{11.10}(q,k,\varepsilon_1+k)+1,\delta_{11.12}(q,k,\varepsilon_0)\}.
\end{aligned} γ 1 ε 1 γ κ ϱ q ′ ε := max { γ 0 , b } , := max { q + 1 , ε 0 } , := max { b + k , α 11.10 ′ ( q , k , γ 1 + k ) + k } , := δ 11.12 ( q , k , κ 0 + k ) + 1 , := max { b , α 11.10 ′ ( q , k , ϱ 0 + k ) + 1 , δ 11.12 ( q , k , ϱ 0 )} , := max { ν 11.7 ( q , k ) , q 0 ′ } , and := max { ν 11.7 ( q , k ) , α 11.10 ′ ( q , k , ε 1 + k ) + 1 , δ 11.12 ( q , k , ε 0 )} .
Let G G G be a ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable graph. If ∣ V ( G ) ∣ < ν 11.7 ( q , k ) |V(G)| < \nu_{11.7}(q,k) ∣ V ( G ) ∣ < ν 11.7 ( q , k ) , then we set G = ∅ \mathcal{G} = \varnothing G = ∅ and M = ∅ M = \varnothing M = ∅ . The corresponding tree-decomposition consists of a single node t t t , with V t = V ( G ) V_t = V(G) V t = V ( G ) . We declare that t ∈ X t \in X t ∈ X . It is then straightforward to verify that the desired properties hold.
Suppose now that ∣ V ( G ) ∣ ≥ ν 11.7 ( q , k ) |V(G)| \ge\nu_{11.7}(q,k) ∣ V ( G ) ∣ ≥ ν 11.7 ( q , k ) . Let σ \sigma σ be the set of all ( b , k ) (b,k) ( b , k ) -irrelevant separations of G G G . By Key Lemma 11.1, σ \sigma σ is a star such that the torso Z Z Z of σ \sigma σ is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and ( b , k ) (b,k) ( b , k ) -narrow. Thus, Z Z Z satisfies the premises for Theorem 10.4 . Hence, there exist a set F ˘ \breve{\mathcal{F}} F ˘ of graphic-k k k -flowers of Z Z Z , a symmetric nested set N ˘ \breve{N} N ˘ of narrow-γ 0 \gamma_0 γ 0 -nested proper k k k -separations of Z Z Z with corresponding tree-decomposition ( T ˘ , V ˘ ) (\breve{T},\breve{\mathcal{V}}) ( T ˘ , V ˘ ) of Z Z Z , a bipartition \breve{X} \mathbin{\mathaccent\cdot\cup} \breve{Y} = V(\breve{T}) , and an assignment Y ˘ ∋ t ˘ ↦ F ˘ t ˘ ∈ F ˘ \breve{Y} \ni\breve{t} \mapsto\breve{F}_{\breve{t}} \in\breve{\mathcal{F}} Y ˘ ∋ t ˘ ↦ F ˘ t ˘ ∈ F ˘ , which satisfy the outcome of Theorem 10.4 .
For every F ˘ ∈ F ˘ \breve{F} \in\breve{\mathcal{F}} F ˘ ∈ F ˘ , let F ( F ˘ ) \mathcal{F}(\breve{F}) F ( F ˘ ) be the nonempty set of lifts of F ˘ \breve{F} F ˘ given by Lemma 12.4 (this set is uniquely determined by F ˘ \breve{F} F ˘ by Lemma 12.4 (iii)). Let G : = { F ( F ˘ ) : F ˘ ∈ F ˘ } \mathcal{G} := \{\mathcal{F}(\breve{F}) : \breve{F} \in\breve{\mathcal{F}}\} G := { F ( F ˘ ) : F ˘ ∈ F ˘ } .
Let ( T , V ) (T,\mathcal{V}) ( T , V ) be a lift of ( T ˘ , V ˘ ) (\breve{T},\breve{\mathcal{V}}) ( T ˘ , V ˘ ) from Z Z Z to G G G . Let N N N be the set of separations induced in ( T , V ) (T,\mathcal{V}) ( T , V ) by the oriented edges of T ˘ ⊆ T \breve{T} \subseteq T T ˘ ⊆ T , and put M : = N ∪ σ ∪ σ ∗ M := N \cup\sigma\cup\sigma^* M := N ∪ σ ∪ σ ∗ .
First, we show that the separations in M M M are totally-γ \gamma γ -nested. By contradiction, suppose that some separation s : = ( A , B ) ∈ M s := (A,B) \in M s := ( A , B ) ∈ M is γ \gamma γ -crossed by a proper ( ≤ k ) (\leq k) ( ≤ k ) -separation s ′ : = ( C , D ) s' := (C,D) s ′ := ( C , D ) of G G G . Then, s s s is 2 ( γ − k ) 2(\gamma-k) 2 ( γ − k ) -broad (by Lemma 5.6) with 2 ( γ − k ) ≥ b > β 11.5 ( q , k ) ≥ q 2(\gamma-k) \geq b > \beta_{11.5}(q,k) \geq q 2 ( γ − k ) ≥ b > β 11.5 ( q , k ) ≥ q , so s ∈ N s \in N s ∈ N by Lemma 11.3 . We have that s s s and s ′ s' s ′ ( γ − k ) (\gamma-k) ( γ − k ) -criss by Observation 5.4. Thus, the A C AC A C -, A D AD A D -, B C BC B C - and B D BD B D -corner separations of G G G are ( > γ − k ) (> \gamma-k) ( > γ − k ) -broad. Since G G G is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable and γ − k ≥ b > q \gamma-k \geq b > q γ − k ≥ b > q , they all have order at least k k k . By submodularity (Lemma 4.8 ), they all have order exactly k k k . By applying Lemma 11.10 to each of these four corner separations, while using γ − k ≥ b > β 11.5 ( q , k ) \gamma-k \geq b > \beta_{11.5}(q,k) γ − k ≥ b > β 11.5 ( q , k ) and γ − k ≥ α 11.10 ′ ( q , k , γ 1 + k ) \gamma-k \geq\alpha'_{11.10}(q,k,\gamma_1+k) γ − k ≥ α 11.10 ′ ( q , k , γ 1 + k ) , we get
∣ A ∩ C ∩ V ( Z ) ∣ , ∣ A ∩ D ∩ V ( Z ) ∣ , ∣ B ∩ C ∩ V ( Z ) ∣ , ∣ B ∩ D ∩ V ( Z ) ∣ > γ 1 + k . |A \cap C \cap V(Z)|,\ |A \cap D \cap V(Z)|,\ |B \cap C \cap V(Z)|,\ |B \cap D \cap V(Z)| > \gamma_1+k. ∣ A ∩ C ∩ V ( Z ) ∣ , ∣ A ∩ D ∩ V ( Z ) ∣ , ∣ B ∩ C ∩ V ( Z ) ∣ , ∣ B ∩ D ∩ V ( Z ) ∣ > γ 1 + k .
Therefore, the induced separations
s ˘ : = ( A ∩ V ( Z ) , B ∩ V ( Z ) ) and s ˘ ′ : = ( C ∩ V ( Z ) , D ∩ V ( Z ) ) \breve{s} := (A \cap V(Z), B \cap V(Z)) \qquad\text{and} \qquad\breve{s}' := (C \cap V(Z), D \cap V(Z)) s ˘ := ( A ∩ V ( Z ) , B ∩ V ( Z )) and s ˘ ′ := ( C ∩ V ( Z ) , D ∩ V ( Z ))
are γ 1 \gamma_1 γ 1 -criss, so they γ 1 \gamma_1 γ 1 -cross by Observation 5.4. Thus, the ( ≤ k ) (\leq k) ( ≤ k ) -separation s ˘ ′ \breve{s}' s ˘ ′ of Z Z Z is γ 1 \gamma_1 γ 1 -broad, hence b b b -broad, and thus narrow since Z Z Z is ( b , k ) (b,k) ( b , k ) -narrow. Therefore, s ˘ ′ \breve{s}' s ˘ ′ is a narrow proper ( ≤ k ) (\leq k) ( ≤ k ) -separation of Z Z Z which γ 0 \gamma_0 γ 0 -crosses s ~ \tilde{s} s ~ . However, by Lemma 11.9 (i) and Lemma 11.9 (ii), we have s ~ ∈ N ~ \tilde{s} \in\tilde{N} s ~ ∈ N ~ , so s ~ \tilde{s} s ~ is narrow-γ 0 \gamma_0 γ 0 -nested, which is a contradiction.
Since ( T , V ) (T,\mathcal{V}) ( T , V ) is a lift of ( T ~ , V ~ ) (\tilde{T},\tilde{\mathcal{V}}) ( T ~ , V ~ ) from Z Z Z to G G G , T T T can be obtained from T ~ \tilde{T} T ~ by adding leaves attached to nodes of T ~ \tilde{T} T ~ . We extend the bipartition ( X ~ , Y ~ ) (\tilde{X},\tilde{Y}) ( X ~ , Y ~ ) of V ( T ~ ) V(\tilde{T}) V ( T ~ ) to a bipartition ( X , Y ) (X,Y) ( X , Y ) of V ( T ) V(T) V ( T ) by declaring that all the newly added leaves are in X X X .
(i). Let t ∈ X t \in X t ∈ X . Suppose first that t t t is one of the leaves that were added to go from T ~ \tilde{T} T ~ to T T T . Since ( T , V ) (T,\mathcal{V}) ( T , V ) is a lift of ( T ~ , V ~ ) (\tilde{T},\tilde{\mathcal{V}}) ( T ~ , V ~ ) from Z Z Z to G G G , there exists ( C , D ) ∈ σ (C,D) \in\sigma ( C , D ) ∈ σ such that V t = C V_t=C V t = C . Thus, ∣ V t ∣ = ∣ C ∣ ≤ q + k |V_t|=|C|\leq q+k ∣ V t ∣ = ∣ C ∣ ≤ q + k by Lemma 11.3 , so the torso at t t t is indeed ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable since q ′ ≥ ν 11.7 ( q , k ) ≥ q + k q' \geq\nu_{11.7}(q,k) \geq q+k q ′ ≥ ν 11.7 ( q , k ) ≥ q + k . Otherwise, t t t was an original node of T ~ \tilde{T} T ~ and t ∈ X ~ t \in\tilde{X} t ∈ X ~ . By Lemma 11.9 (iii), the torso at t t t in ( T , V ) (T,\mathcal{V}) ( T , V ) is equal to the torso at t t t in ( T ~ , V ~ ) (\tilde{T},\tilde{\mathcal{V}}) ( T ~ , V ~ ) . We are then done since the latter is ( q ′ , k ) (q',k) ( q ′ , k ) -unbreakable by Theorem 10.4 (i).
(ii). Let t ∈ Y = Y ~ t \in Y=\tilde{Y} t ∈ Y = Y ~ and let F ~ t ∈ F ~ \tilde{F}_t \in\tilde{\mathcal{F}} F ~ t ∈ F ~ be the corresponding graphic-k k k -flower of Z Z Z . Let F t \mathcal{F}_t F t be the set of lifts of F ~ t \tilde{F}_t F ~ t given by Lemma 12.4 (this set is unique by Lemma 12.4 (iii)). By definition of G \mathcal{G} G , we have F t ∈ G \mathcal{F}_t \in\mathcal{G} F t ∈ G . Let F ∈ F t F \in\mathcal{F}_t F ∈ F t be arbitrary. We show that F F F ϱ \varrho ϱ -refines ( T , V ) (T,\mathcal{V}) ( T , V ) at t t t .
First, let s s s be a ϱ \varrho ϱ -broad petal-separation of F F F . By Lemma 11.10 , since s s s is ϱ \varrho ϱ -broad with ϱ ≥ b > β 11.5 ( q , k ) \varrho\geq b > \beta_{11.5}(q,k) ϱ ≥ b > β 11.5 ( q , k ) and ϱ > α 11.10 ′ ( q , k , ϱ 0 + k ) \varrho> \alpha'_{11.10}(q,k,\varrho_0+k) ϱ > α 11.10 ′ ( q , k , ϱ 0 + k ) , the separation s ~ \tilde{s} s ~ that s s s induces on Z Z Z is ϱ 0 \varrho_0 ϱ 0 -broad. By Lemma 12.4 (i), s ~ \tilde{s} s ~ is a petal-separation of F ~ t \tilde{F}_t F ~ t . By Theorem 10.4 (ii), there exists a separation s ~ ′ \tilde{s}' s ~ ′ of the t t t -star of ( T ~ , V ~ ) (\tilde{T},\tilde{\mathcal{V}}) ( T ~ , V ~ ) which is ϱ 0 \varrho_0 ϱ 0 -close to s ~ \tilde{s} s ~ . By Lemma 11.9 (ii), there is a lift s ′ s' s ′ of s ~ ′ \tilde{s}' s ~ ′ in the t t t -star of ( T , V ) (T,\mathcal{V}) ( T , V ) . By Lemma 11.12 , and since ϱ ≥ δ 11.12 ( q , k , ϱ 0 ) \varrho\geq\delta_{11.12}(q,k,\varrho_0) ϱ ≥ δ 11.12 ( q , k , ϱ 0 ) , we have that s ′ s' s ′ is ϱ \varrho ϱ -close to s s s .
Conversely, let s ′ s' s ′ be a ϱ \varrho ϱ -broad separation of the t t t -star of ( T , V ) (T,\mathcal{V}) ( T , V ) . By Lemma 11.3 and since ϱ ≥ b > q \varrho\geq b > q ϱ ≥ b > q , it follows that s ′ ∉ σ ∪ σ ∗ s' \notin\sigma\cup\sigma^* s ′ ∈ / σ ∪ σ ∗ . By Lemma 11.9 (i), we have s ′ ∈ N s' \in N s ′ ∈ N . Then, by Lemma 11.9 (ii), the separation s ~ ′ \tilde{s}' s ~ ′ that s ′ s' s ′ induces on Z Z Z belongs to the t t t -star of ( T ~ , V ~ ) (\tilde{T},\tilde{\mathcal{V}}) ( T ~ , V ~ ) . By Lemma 11.10 , since s ′ s' s ′ is ϱ \varrho ϱ -broad with ϱ ≥ b > β 11.5 ( q , k ) \varrho\geq b > \beta_{11.5}(q,k) ϱ ≥ b > β 11.5 ( q , k ) and ϱ ≥ α 11.10 ′ ( q , k , ϱ 0 + k ) \varrho\geq\alpha'_{11.10}(q,k,\varrho_0+k) ϱ ≥ α 11.10 ′ ( q , k , ϱ 0 + k ) , we get that s ~ ′ \tilde{s}' s ~ ′ is ϱ 0 \varrho_0 ϱ 0 -broad. By Theorem 10.4 (ii), there exists a petal-separation s ~ \tilde{s} s ~ of F ~ t \tilde{F}_t F ~ t which is ϱ 0 \varrho_0 ϱ 0 -close to s ~ ′ \tilde{s}' s ~ ′ . By Lemma 12.4 (ii), there is a petal-separation s s s of F F F that is a lift of s ~ \tilde{s} s ~ . By Lemma 11.12 , and since ϱ ≥ δ 11.12 ( q , k , ϱ 0 ) \varrho\geq\delta_{11.12}(q,k,\varrho_0) ϱ ≥ δ 11.12 ( q , k , ϱ 0 ) , we have that s s s is ϱ \varrho ϱ -close to s ′ s' s ′ .
(iii). Let s : = ( A , B ) s := (A,B) s := ( A , B ) be an ε \varepsilon ε -broad k k k -separation of G G G . Then,
∣ A ∣ , ∣ B ∣ > α 11.10 ′ ( q , k , ε 1 + k ) . |A|,|B| > \alpha'_{11.10}(q,k,\varepsilon_1+k). ∣ A ∣ , ∣ B ∣ > α 11.10 ′ ( q , k , ε 1 + k ) .
Let s ~ \tilde{s} s ~ be the separation that s s s induces on Z Z Z . By Lemma 11.10 , s ~ \tilde{s} s ~ is an ε 1 \varepsilon_1 ε 1 -broad (≤ k \leq k ≤ k )-separation of Z Z Z . Since s ~ \tilde{s} s ~ is ε 1 \varepsilon_1 ε 1 -broad with ε 1 > q \varepsilon_1 > q ε 1 > q , and Z Z Z is ( q , k − 1 ) (q,k-1) ( q , k − 1 ) -unbreakable, s ~ \tilde{s} s ~ is a k k k -separation of Z Z Z . Since ε 1 ≥ ε 0 \varepsilon_1 \geq\varepsilon_0 ε 1 ≥ ε 0 , by Theorem 10.4 (iv), s ~ \tilde{s} s ~ is either ε 0 \varepsilon_0 ε 0 -close to a separation in N ~ \tilde{N} N ~ or ε 0 \varepsilon_0 ε 0 -close to some displayed separation of some F ~ t \tilde{F}_t F ~ t . Call this separation s ′ s' s ′ . By Lemma 11.9 (ii) or Lemma 12.4 (ii), either there exists a lift s ′ ∈ N s' \in N s ′ ∈ N of s ′ s' s ′ , or every F ∈ F t F \in\mathcal{F}_t F ∈ F t has a displayed separation s ′ s' s ′ that is a lift of s ′ s' s ′ . By Lemma 11.12 , since ε ≥ δ 11.12 ( q , k , ε 0 ) \varepsilon\geq\delta_{11.12}(q,k,\varepsilon_0) ε ≥ δ 11.12 ( q , k , ε 0 ) , every such s ′ s' s ′ is ε \varepsilon ε -close to s s s .
(iv). Every F ( F ~ ) ∈ G \mathcal{F}(\tilde{F}) \in\mathcal{G} F ( F ~ ) ∈ G stems from Lemma 12.4 , and so is ( q , k ) (q,k) ( q , k ) -coherent. Let ψ \psi ψ be an automorphism of G G G . Since σ \sigma σ is canonical, ψ ∣ Z \psi|Z ψ ∣ Z is an automorphism of Z Z Z . Since F ~ \tilde{\mathcal{F}} F ~ is canonical by Theorem 10.4 (v), there is a permutation Φ ~ \tilde{\Phi} Φ ~ of F ~ \tilde{\mathcal{F}} F ~ such that F ~ \tilde{F} F ~ and Φ ~ ( F ~ ) \tilde{\Phi}(\tilde{F}) Φ ~ ( F ~ ) are ψ ∣ Z \psi|Z ψ ∣ Z -isomorphic for every F ~ ∈ F ~ \tilde{F} \in\tilde{\mathcal{F}} F ~ ∈ F ~ . By the proof of Corollary 9.13, distinct F ~ , F ~ ′ ∈ F ~ \tilde{F},\tilde{F}' \in\tilde{\mathcal{F}} F ~ , F ~ ′ ∈ F ~ have different sets of proper displayed separations. By Lemma 12.4 (i) and (ii), this set is recovered from any lift by projecting its displayed separations and retaining the proper ones. Hence the nonempty families F ( F ~ ) \mathcal{F}(\widetilde{F}) F ( F ) and F ( F ~ ′ ) \mathcal{F}(\widetilde{F}') F ( F ′ ) are disjoint, and thus distinct. In particular, F ~ ↦ F ( F ~ ) \widetilde{F} \mapsto\mathcal{F}(\widetilde{F}) F ↦ F ( F ) is a bijection F ~ → G \widetilde{\mathcal{F}} \to\mathcal{G} F → G . Hence Φ ~ \widetilde{\Phi} Φ induces a permutation Φ \Phi Φ of G \mathcal{G} G , namely the one which maps F ( F ~ ) \mathcal{F}(\widetilde{F}) F ( F ) to F ( Φ ~ ( F ~ ) ) \mathcal{F}(\widetilde{\Phi}(\widetilde{F})) F ( Φ ( F )) for every F ~ ∈ F ~ \widetilde{F} \in\widetilde{\mathcal{F}} F ∈ F . By Lemma 12.4 (iii), the sets F ( F ~ ) \mathcal{F}(\widetilde{F}) F ( F ) and F ( Φ ~ ( F ~ ) ) = Φ ( F ( F ~ ) ) \mathcal{F}(\widetilde{\Phi}(\widetilde{F})) = \Phi(\mathcal{F}(\widetilde{F})) F ( Φ ( F )) = Φ ( F ( F )) are ψ \psi ψ -isomorphic for every F ~ ∈ F ~ \widetilde{F} \in\widetilde{\mathcal{F}} F ∈ F . Hence G \mathcal{G} G is ( q , k ) (q,k) ( q , k ) -canonical.
Since σ \sigma σ is canonically determined by G \mathcal{G} G , we recover N = M ∖ ( σ ∪ σ ∗ ) N = M \setminus(\sigma\cup\sigma^*) N = M ∖ ( σ ∪ σ ∗ ) from M M M , and N ~ \widetilde{N} N by projecting N N N to Z Z Z . Removing the added leaves corresponding to σ \sigma σ recovers ( T ~ , V ~ ) (\widetilde{T},\widetilde{\mathcal{V}}) ( T , V ) . Hence, by Theorem 10.4 (v), the bipartition X ~ ⊔ Y ~ = V ( T ~ ) \widetilde{X} \sqcup\widetilde{Y} = V(\widetilde{T}) X ⊔ Y = V ( T ) and the assignment Y ~ ∋ t ↦ F ~ t ∈ F ~ \widetilde{Y} \ni t \mapsto\widetilde{F}_t \in\widetilde{\mathcal{F}} Y ∋ t ↦ F t ∈ F are canonical given M M M and ( T , V ) (T,\mathcal{V}) ( T , V ) . Since the added leaves are determined by σ \sigma σ , assigning them to X X X preserves canonicity of the bipartition. Moreover, assigning to each t ∈ Y t \in Y t ∈ Y the family F t = F ( F ~ t ) F_t = \mathcal{F}(\widetilde{F}_t) F t = F ( F t ) of lifts of F ~ t \widetilde{F}_t F t is canonical given M M M and ( T , V ) (T,\mathcal{V}) ( T , V ) by Lemma 12.4 (iii). It is bijective, because we have seen above that F ~ → G \widetilde{\mathcal{F}} \to\mathcal{G} F → G , F ~ ↦ F ( F ~ ) \widetilde{F} \mapsto\mathcal{F}(\widetilde{F}) F ↦ F ( F ) is a bijection.
Finally, we show that M M M and ( T , V ) (T,\mathcal{V}) ( T , V ) are κ \kappa κ -canonical. Let ψ \psi ψ be an automorphism of G \mathcal{G} G . Since σ \sigma σ is canonical, ψ ∣ Z \psi|Z ψ ∣ Z is an automorphism of Z Z Z . The set N ~ \widetilde{N} N is κ 0 \kappa_0 κ 0 -canonical by Theorem 10.4 (v), so the sets ψ ( N ~ ) \psi(\widetilde{N}) ψ ( N ) and N ~ \widetilde{N} N κ 0 \kappa_0 κ 0 -approximate each other. We show that ψ ( N ) \psi(N) ψ ( N ) κ \kappa κ -approximates N N N . Let s ′ ∈ N s' \in N s ′ ∈ N . By Lemma 11.9 (ii), s ′ s' s ′ is a lift of some s ~ ′ ∈ N ~ \widetilde{s}' \in\widetilde{N} s ′ ∈ N . If s ~ ′ \widetilde{s}' s ′ is not κ 0 \kappa_0 κ 0 -broad, then s ~ ′ \widetilde{s}' s ′ is ( κ 0 + k ) (\kappa_0 + k) ( κ 0 + k ) -close to the separation ( V ( Z ) , ∅ ) (V(Z),\varnothing) ( V ( Z ) , ∅ ) , say. Then, s ′ s' s ′ and ( V ( G ) , ∅ ) (V(\mathcal{G}),\varnothing) ( V ( G ) , ∅ ) are δ 11.12 ( q , k , κ 0 + k ) \delta_{11.12}(q,k,\kappa_0+k) δ 11.12 ( q , k , κ 0 + k ) -close by Lemma 11.12 , so s ′ s' s ′ is not κ \kappa κ -broad and we have nothing to show. If s ~ ′ \widetilde{s}' s ′ is κ 0 \kappa_0 κ 0 -broad, there exists s ~ ∈ N ~ \widetilde{s} \in\widetilde{N} s ∈ N such that ψ ( s ~ ) \psi(\widetilde{s}) ψ ( s ) and s ~ ′ \widetilde{s}' s ′ are κ 0 \kappa_0 κ 0 -close. There exists a lift s s s of s ~ \widetilde{s} s in N N N by Lemma 11.9 (ii). Then ψ ( s ) \psi(s) ψ ( s ) is a lift of ψ ( s ~ ) \psi(\widetilde{s}) ψ ( s ) , so ψ ( s ) \psi(s) ψ ( s ) and s ′ s' s ′ are κ \kappa κ -close by Lemma 11.12 . This proves that ψ ( N ) \psi(N) ψ ( N ) κ \kappa κ -approximates N N N . By symmetry, N N N also κ \kappa κ -approximates ψ ( N ) \psi(N) ψ ( N ) , so N N N is κ \kappa κ -canonical. Since σ \sigma σ is canonical and M = N ∪ σ ∪ σ ∗ M = N \cup\sigma\cup\sigma^* M = N ∪ σ ∪ σ ∗ , it follows that M M M is κ \kappa κ -canonical. Since ( T , V ) (T,\mathcal{V}) ( T , V ) is entirely determined by M M M , it is also κ \kappa κ -canonical. □
Proof of Theorem 2.6 . Theorem 12.6 implies Theorem 2.6 by picking a graphic-k k k -flower F t ∈ F t F_t \in\mathcal{F}_t F t ∈ F t for each t ∈ Y t \in Y t ∈ Y . □
Acknowledgements.
In 2012, Fabian Hundertmark mentioned to the second author the idea to investigate 3-dimensional corner diagrams, yet it is only now that we succeeded in applying this idea in a meaningful way. We are grateful to Nathan Bowler for spotting a masked error in an early version of our notion of totally-ε \varepsilon ε -nested. This led us to distinguish between the two notions totally-ε \varepsilon ε -nested and narrow-ε \varepsilon ε -nested. We thank Vera Traub for asking us at Bertinoro whether our initial decompositions, as in Theorem 2.6 , could be described in terms of cactus-decompositions. We thank Joseph Devine for discussions in the initial phase of the project. We thank Sandra Albrechtsen for the cactus illustration marking the final frontier of our extended abstract.
AI methodology
We developed the main structural ideas and original proof arguments without AI assistance. A complete draft containing the main results and their proofs was ready in April 2026 without any use of AI. During subsequent revisions, we used ChatGPT (OpenAI) and Claude (Anthropic) to examine and revise mathematical arguments, revise expository text, and proofread the manuscript.
The main technical revision concerned the parameter dependencies in the proof of Lemma 9.4 . AI identified incompatible breadth requirements in the application of an auxiliary lemma. The subsequent AI-assisted revision involved modifying the shifting argument in Lemmas 8.14 and 8.15 and strengthening Lemma 8.16 so that the required breadth of the star no longer depended on the approximation parameter.
Minor AI-assisted revisions concerned the lifting of flowers in Lemma 12.4 and its use in the canonicity argument or simplifying the argument in Lemma 9.24. AI also pointed out some small inaccuracies, such as not accounting for the separator size in an approximation bound, and some places where additional arguments were required, such as in Lemma 7.3 .
AI was also used to generate a draft proof of Lemma A.1. It proposed proof arguments, identified supporting results in the literature, and drafted the supplementary material in Section A. We checked this material and added it during the revision process.
AI was used to draft and revise portions of the introduction, the detailed proof overview, and explanatory passages throughout the manuscript. This assistance mostly consisted in refining the wording. It was also used to identify typographical errors, inconsistencies in notation, and other imprecisions.
We evaluated the suggestions and reviewed all AI-generated material incorporated into the manuscript. We checked the mathematical arguments and supporting references and take responsibility for the final text.
Unbreakable minors for iterated decompositions
In the recursive application of Theorem 1.2 described in the introduction, we retain the cyclic pieces and continue within the bags whose nodes lie on no cycle of the decomposition cactus. This appendix provides the auxiliary lemma used to represent these bags by unbreakable minors of the original graph. Working with minors preserves membership in any minor-closed class, whereas passing to a torso by completing adhesion sets into cliques need not do so.
A set C ⊆ V ( G ) C \subseteq V(G) C ⊆ V ( G ) is ( q , k ) (q,k) ( q , k ) -unbreakable in G G G if every separation ( A , B ) (A,B) ( A , B ) of G G G of order at most k k k satisfies min { ∣ C ∩ ( A ∖ B ) ∣ , ∣ C ∩ ( B ∖ A ) ∣ } ≤ q \min\{|C \cap(A \setminus B)|, |C \cap(B \setminus A)|\} \leq q min { ∣ C ∩ ( A ∖ B ) ∣ , ∣ C ∩ ( B ∖ A ) ∣ } ≤ q . A C C C -rooted minor of G G G is a graph H H H with C ⊆ V ( H ) C \subseteq V(H) C ⊆ V ( H ) that admits a minor model in G G G in which the branch set corresponding to each c ∈ C c \in C c ∈ C contains c c c .
The following lemma allows us to retain an unbreakable set as a set of distinct roots in an unbreakable minor.
Lemma A.1 (Unbreakable minor). There is a computable function F : N 0 2 → N 0 F \colon\mathbb{N}_0^2 \to\mathbb{N}_0 F : N 0 2 → N 0 such that, whenever C ⊆ V ( G ) C \subseteq V(G) C ⊆ V ( G ) is ( q , k ) (q,k) ( q , k ) -unbreakable in a finite graph G G G , there is a C C C -rooted minor H H H of G G G that is ( F ( q , k ) , k ) (F(q,k),k) ( F ( q , k ) , k ) -unbreakable.
To justify the recursive application described in the introduction, let d d d be the approximation bound for a cactus-decomposition given by Theorem 1.2, and let C C C be one of its bags whose node lies on no cycle. Every displayed separation has a side containing C C C , as does every improper separation. Consequently, every separation at edit-distance at most d d d from such a separation has at most d d d vertices of C C C on one of its proper sides. Thus C C C is ( d , k ) (d,k) ( d , k ) -unbreakable in G G G . By Lemma A.1, there is a C C C -rooted minor H H H of G G G that is ( F ( d , k ) , k ) (F(d,k),k) ( F ( d , k ) , k ) -unbreakable. We may therefore apply Theorem 1.2 to H H H with k k k replaced by k + 1 k+1 k + 1 .
We first prove the existence of the bound using the labelled Graph Minor Theorem [116 Theorem 1.7] , together with replacement and minimality arguments. We then establish its computability.
This background-material appendix was generated by AI and checked by the authors.
Bounded replacements and unbreakability
Fix k ∈ N 0 k \in\mathbb{N}_0 k ∈ N 0 . For a graph J J J with a specified set R R R of labelled roots and an ordered partition ( L , Z , U ) (L,Z,U) ( L , Z , U ) of R R R , let
λ J , R ( L , Z , U ) = min { ∣ X ∩ Y ∣ : ( X , Y ) is a separation of J , R ∩ ( X ∖ Y ) = L , R ∩ X ∩ Y = Z , R ∩ ( Y ∖ X ) = U } . \lambda_{J,R}(L,Z,U) = \min\left\{ \begin{array}{ll}
|X \cap Y| : & (X,Y)\text{ is a separation of }J,\\
& R \cap(X \setminus Y) = L,\quad R \cap X \cap Y = Z,\\
& R \cap(Y \setminus X) = U
\end{array} \right\}. λ J , R ( L , Z , U ) = min ⎩ ⎨ ⎧ ∣ X ∩ Y ∣ : ( X , Y ) is a separation of J , R ∩ ( X ∖ Y ) = L , R ∩ X ∩ Y = Z , R ∩ ( Y ∖ X ) = U ⎭ ⎬ ⎫ .
where the minimum of the empty set is ∞ \infty ∞ . The truncated separation profile of ( J , R ) (J,R) ( J , R ) is the function
π J , R k ( L , Z , U ) = min { k + 1 , λ J , R ( L , Z , U ) } . \pi^k_{J,R}(L,Z,U) = \min\{k+1,\lambda_{J,R}(L,Z,U)\}. π J , R k ( L , Z , U ) = min { k + 1 , λ J , R ( L , Z , U )} .
Profiles on different root sets are compared using the root labels.
Lemma A.2 (Bounded rooted replacement). For every r , k ∈ N 0 r,k \in\mathbb{N}_0 r , k ∈ N 0 there is b ( r , k ) b(r,k) b ( r , k ) such that every graph J J J with at most $r labelled roots R R R has an R R R -rooted minor J ′ J' J ′ satisfying
∣ V ( J ′ ) ∣ ≤ b ( r , k ) and π J ′ , R k = π J , R k . |V(J')| \leq b(r,k) \qquad\text{and}\qquad\pi^k_{J',R} = \pi^k_{J,R}. ∣ V ( J ′ ) ∣ ≤ b ( r , k ) and π J ′ , R k = π J , R k .
The function b b b can be chosen nondecreasing in its first argument.
Proof. Fix a set of s s s root labels. Graphs with these roots are well-quasi-ordered by the rooted minor relation, by the labelled Graph Minor Theorem [116 Theorem 1.7] . To obtain this rooted version, give the s s s roots distinct labels 1 , … , s 1,\ldots,s 1 , … , s , and give every other vertex label 0 0 0 . Order this finite label set by equality. In the resulting minor model, the branch set of each root must contain the correspondingly labelled root, as required.
There are at most ( k + 2 ) 3 s (k+2)^{3^s} ( k + 2 ) 3 s truncated separation profiles on these roots. For each profile, its minimal realizations under the rooted minor relation form a finite set, up to rooted isomorphism, since distinct minimal realizations are incomparable. Moreover, every realization contains a minimal realization of the same profile as a rooted minor: successive proper rooted minors strictly decrease ∣ V ∣ + ∣ E ∣ |V|+|E| ∣ V ∣ + ∣ E ∣ . Taking the maximum order of these minimal realizations, over all profiles and all s ≤ r s\le r s ≤ r , gives the desired b ( r , k ) b(r,k) b ( r , k ) . □ \square □
We next record the replacement property of the profiles. Suppose that Q = J ∪ K Q=J\cup K Q = J ∪ K , where V ( J ) ∩ V ( K ) = S ⊆ R ⊆ V ( J ) V(J)\cap V(K)=S\subseteq R\subseteq V(J) V ( J ) ∩ V ( K ) = S ⊆ R ⊆ V ( J ) , and that J ′ J' J ′ has the same truncated profile as J J J on R R R . Form Q ′ = J ′ ∪ K Q'=J'\cup K Q ′ = J ′ ∪ K by identifying the corresponding vertices of S S S , with all other vertices of J ′ J' J ′ taken to be new. Let D D D be a set of distinguished vertices of Q Q Q such that D ∩ V ( J ) ⊆ R D\cap V(J)\subseteq R D ∩ V ( J ) ⊆ R , and use the corresponding vertices as distinguished vertices of Q ′ Q' Q ′ . Then
π Q , D k = π Q ′ , D k . (12) \pi_{Q,D}^{k}=\pi_{Q',D}^{k}. \tag*{(12)} π Q , D k = π Q ′ , D k . (12)
Indeed, restrict a separation of Q Q Q of order at most k k k to J J J and K K K . Profile equality supplies a separation of J ′ J' J ′ of no greater order with the same assignment of the roots in R R R to the two exclusive sides and the separator. It agrees with the separation of K K K on S S S , so the two separations can be pasted. If their separator orders are a a a and b b b , and z z z vertices of S S S lie in the separator, the pasted separation has order a + b − z a+b-z a + b − z . Thus its order does not increase, and the assignment of D D D is unchanged. Interchanging J J J and J ′ J' J ′ proves the reverse implication, giving (10). If J ′ J' J ′ is an R R R -rooted minor of J J J , the replacement also gives a D D D -rooted minor of Q Q Q .
Proof of the existence assertion in Lemma A.1. Set F ( q , k ) = b ( q + k , k ) F(q,k)=b(q+k,k) F ( q , k ) = b ( q + k , k ) . Among all C C C -rooted minors of G G G having the same truncated separation profile on C C C as G G G , choose H H H with the fewest vertices. Such a minor exists, since G G G itself is eligible. Profile equality implies that C C C remains ( q , k ) (q,k) ( q , k ) -unbreakable in H H H .
Let ( A , B ) (A,B) ( A , B ) be a separation of H H H of order at most k k k . By symmetry, we may assume that ∣ C ∩ ( A ∖ B ) ∣ ≤ q |C\cap(A\setminus B)|\le q ∣ C ∩ ( A ∖ B ) ∣ ≤ q . Put
S = A ∩ B and R = S ∪ ( C ∩ A ) . S=A\cap B \qquad\text{and}\qquad R=S\cup(C\cap A). S = A ∩ B and R = S ∪ ( C ∩ A ) .
Then ∣ R ∣ ≤ k + q |R|\le k+q ∣ R ∣ ≤ k + q . By Lemma A.2, H [ A ] H[A] H [ A ] has an R R R -rooted minor J ′ J' J ′ with the same truncated profile on R R R and with at most b ( q + k , k ) b(q+k,k) b ( q + k , k ) vertices. Replace H [ A ] H[A] H [ A ] by J ′ J' J ′ and retain H [ B ] H[B] H [ B ] , identifying the corresponding vertices of S S S . The resulting graph H ′ H' H ′ is a C C C -rooted minor of H H H . By (10), it has the same truncated profile on C C C as H H H , and hence as G G G . The minimality of H H H now gives
∣ V ( H ) ∣ ≤ ∣ V ( H ′ ) ∣ = ∣ V ( H ) ∣ − ∣ A ∣ + ∣ V ( J ′ ) ∣ , |V(H)|\le|V(H')|=|V(H)|-|A|+|V(J')|, ∣ V ( H ) ∣ ≤ ∣ V ( H ′ ) ∣ = ∣ V ( H ) ∣ − ∣ A ∣ + ∣ V ( J ′ ) ∣ ,
so ∣ A ∣ ≤ ∣ V ( J ′ ) ∣ ≤ F ( q , k ) |A| \le|V(J')| \le F(q,k) ∣ A ∣ ≤ ∣ V ( J ′ ) ∣ ≤ F ( q , k ) . In particular,
min { ∣ A ∖ B ∣ , ∣ B ∖ A ∣ } ≤ F ( q , k ) , \min\{|A \setminus B|, |B \setminus A|\} \le F(q,k), min { ∣ A ∖ B ∣ , ∣ B ∖ A ∣ } ≤ F ( q , k ) ,
as required. □
Computability of the bound
It remains to show that b b b , and hence F F F , can be chosen computable. Well-quasi-ordering alone does not make the preceding choice of b b b effective. We use the following consequence of the irrelevant-vertex results for rooted topological minors.
Lemma A.3 (Bounded treewidth reduction). There is a computable function w ( r , k ) w(r,k) w ( r , k ) such that every graph J J J with r r r labelled roots R R R has an induced subgraph J 0 J_0 J 0 containing R R R with
tw ( J 0 ) ≤ w ( r , k ) and π J 0 , R k = π J , R k . \operatorname{tw}(J_0) \le w(r,k) \qquad\text{and}\qquad\pi^k_{J_0,R}=\pi^k_{J,R}. tw ( J 0 ) ≤ w ( r , k ) and π J 0 , R k = π J , R k .
Proof. Put δ = r + 2 ( k + 1 ) \delta=r+2(k+1) δ = r + 2 ( k + 1 ) . The rooted topological δ \delta δ -folio of J J J records its rooted topological minors P P P with ∣ E ( P ) ∣ + is ( P ) ≤ δ |E(P)|+\operatorname{is}(P)\le\delta ∣ E ( P ) ∣ + is ( P ) ≤ δ , where is ( P ) \operatorname{is}(P) is ( P ) is the number of isolated vertices of P P P ; see [FLP+20, Definition 4.4]. Rooted topological models map each root to the vertex with the same label.
We first show that this folio, together with J [ R ] J[R] J [ R ] , determines π J , R k \pi^k_{J,R} π J , R k . Consider an ordered partition ( L , Z , U ) (L,Z,U) ( L , Z , U ) of R R R . If ∣ Z ∣ > k |Z|>k ∣ Z ∣ > k , its profile value is k + 1 k+1 k + 1 . An edge between L L L and U U U also makes this value k + 1 k+1 k + 1 , since no separation with the prescribed root assignment exists. Otherwise, let τ \tau τ be the minimum size of a set T ⊆ V ( J ) ∖ R T\subseteq V(J)\setminus R T ⊆ V ( J ) ∖ R separating L L L from U U U in J − Z J-Z J − Z . Then
λ J , R ( L , Z , U ) = ∣ Z ∣ + τ . \lambda_{J,R}(L,Z,U)=|Z|+\tau. λ J , R ( L , Z , U ) = ∣ Z ∣ + τ .
By the vertex-capacitated form of Menger’s theorem, τ \tau τ is the maximum number of L L L –U U U paths with pairwise disjoint sets of nonroot vertices. The paths may share their endpoints. Trimming each path between two consecutive root vertices at which it passes from L L L to U U U , we may assume that no path has an internal root.
For 1 ≤ t ≤ k + 1 1\le t\le k+1 1 ≤ t ≤ k + 1 , the existence of t t t such paths is determined by the δ \delta δ -folio. Indeed, each path has an internal vertex, since there is no edge between L L L and U U U . Retain one internal vertex on each path and suppress the other internal vertices. Include all r r r roots as labelled vertices, adding unused roots as isolated vertices. The resulting rooted graph has 2 t 2t 2 t edges and at most r r r isolated vertices, so belongs to the δ \delta δ -folio. Conversely, a rooted topological model of such a graph supplies the required paths. Including every root prevents a root from occurring internally on any of these paths. Thus the folio determines min { k + 1 , τ } \min\{k+1,\tau\} min { k + 1 , τ } and hence the truncated profile.
The irrelevant-vertex results of Fomin, Lokshtanov, Panolan, Saurabh and Zehavi imply a computable threshold w = w ( r , k ) w=w(r,k) w = w ( r , k ) such that, if tw ( J ) > w \operatorname{tw}(J)>w tw ( J ) > w , some nonroot vertex can be deleted while preserving the rooted δ \delta δ -folio. Specifically, combine [FLP+20, Lemmas 3.1, 4.1 and 6.1], as in the proof of Lemma 8.1 there. These results give a bounded-treewidth, large-clique, or large-flat-wall alternative, with a nonroot irrelevant vertex in the latter two cases; they in fact preserve the extended folio. Their hypothesis on the number of roots holds since r ≤ δ ≤ 16 δ 2 r\le\delta\le16\delta^2 r ≤ δ ≤ 16 δ 2 . All these references are to the full version linked in the bibliography.
Repeatedly delete such vertices until the treewidth is at most w w w . Each deletion preserves both the folio and the graph induced by R R R , and hence the truncated separation profile. This yields J 0 J_0 J 0 . □
Proof of computability in Lemma A.1. Fix r r r , k k k and a set of r r r root labels, and compute w ( r , k ) w(r,k) w ( r , k ) from Lemma A.3 . For each possible truncated separation profile π \pi π , we compute a finite list L π \mathcal{L}_{\pi} L π of rooted graphs with profile π \pi π .
Start with the empty list. At each step, decide whether there is a graph of treewidth at most w ( r , k ) w(r,k) w ( r , k ) with these roots and profile π \pi π that contains no member of the current list as a rooted minor. This is a decidable monadic second-order satisfiability question on graphs of bounded treewidth; see [3 Theorem 2.3] . Indeed, for each root partition and each a ≤ k a \le k a ≤ k , the existence of a separation of order at most a a a with that root assignment is MSO-definable. Boolean combinations of these conditions specify the exact truncated profile. Containment of a fixed rooted minor is also MSO-definable, using disjoint connected branch sets with the prescribed roots in their corresponding branch sets. The root labels can be represented by existentially quantified distinct vertices. This is the bounded-treewidth MSO enumeration method used in [3 Section 3] .
If the answer is yes, find a witness by enumerating finite rooted graphs and add it to L π \mathcal{L}_{\pi} L π . Repeat until the answer is no. The procedure terminates: otherwise the successive witnesses would form an infinite sequence in which no earlier member is a rooted minor of a later member, contrary to labelled minor well-quasi-ordering. This also handles profiles having no realization, for which the list remains empty.
At termination, every graph of treewidth at most w ( r , k ) w(r,k) w ( r , k ) with profile π \pi π contains a member of L π \mathcal{L}_{\pi} L π as a rooted minor. Lemma A.3 extends this conclusion to every graph with profile π \pi π : first pass to the induced subgraph of bounded treewidth and then take the rooted minor from the list.
Perform this computation for every profile and every root count s ≤ r s \le r s ≤ r , and take the maximum order of the representatives in the resulting lists. This gives a computable choice of b ( r , k ) b(r,k) b ( r , k ) in Lemma A.2, nondecreasing in r r r . Using this choice in the existence proof gives the computable function F ( q , k ) = b ( q + k , k ) F(q,k)=b(q+k,k) F ( q , k ) = b ( q + k , k ) asserted in Lemma A.1. □
The minor in Lemma A.1 can also be found by a uniform algorithm. Enumerate the C C C -rooted minors of G G G , compute their truncated separation profiles on C C C , and choose one with the fewest vertices among those preserving the profile of G G G . The existence proof shows that this minor has the required unbreakability. No running-time bound is claimed.