Introduction
Minimum Bisection asks whether the vertices of a graph can be divided into two parts of equal size, up to one vertex, with at most k k k edges between the two parts. Throughout the paper, G G G is a simple undirected graph and n = ∣ V ( G ) ∣ n = |V(G)| n = ∣ V ( G ) ∣ . For A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) , let δ G ( A ) : = { u v ∈ E ( G ) : ∣ { u , v } ∩ A ∣ = 1 } \delta_G(A) := \{uv \in E(G) : |\{u,v\} \cap A| = 1\} δ G ( A ) := { uv ∈ E ( G ) : ∣ { u , v } ∩ A ∣ = 1 } .
Minimum Bisection
Input: A simple undirected graph G G G and an integer k ≥ 0 k \ge0 k ≥ 0 .
Question: Is there a set A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) with ∣ A ∣ = ⌊ n / 2 ⌋ |A| = \lfloor n/2 \rfloor ∣ A ∣ = ⌊ n /2 ⌋ and ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \le k ∣ δ G ( A ) ∣ ≤ k ?
Parameter: k k k .
Kernighan and Lin [26 ] formulated the problem of partitioning the vertices of an edge-weighted graph into parts of prescribed sizes so as to minimize the weight of the crossing edges. The unweighted two-part case with equal part sizes is Minimum Bisection. Garey, Johnson, and Stockmeyer [16 ] proved that Minimum Bisection is NP-complete.
Cygan, Lokshtanov, Marcin Pilipczuk, Michał Pilipczuk, and Saurabh [13 ] gave the first fixed-parameter algorithm for Minimum Bisection, with running time O ∗ ( 2 O ( k 3 ) ) O^*(2^{O(k^3)}) O ∗ ( 2 O ( k 3 ) ) , and asked whether Minimum Bisection admits a single-exponential fixed-parameter algorithm.2 Cygan, Komosa, Lokshtanov, Marcin Pilipczuk, Michał Pilipczuk, Saurabh, and Wahlström [11 ] improved the running time to O ∗ ( 2 O ( k log k ) ) O^*(2^{O(k \log k)}) O ∗ ( 2 O ( k l o g k ) ) , making substantial progress toward this goal while leaving the question open.
In this paper, we answer the question posed in [13 ] affirmatively by giving a deterministic O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) -time algorithm for Minimum Bisection.
Existing framework for Minimum Bisection by [13 ] and [11 ] . Before describing our algorithm, we review the framework common to the two previous algorithms. Both algorithms first compute a tree decomposition with a certain property, called an unbreakable decomposition , and then perform bottom-up dynamic programming over the decomposition. Unbreakable decomposition itself was introduced in [13 ] as an algorithmically tractable relaxation of lean tree decompositions [42 , 6 ] .
For a graph H H H , a set X ⊆ V ( H ) X \subseteq V(H) X ⊆ V ( H ) is ( q , k ) (q,k) ( q , k ) -vertex-unbreakable if every separation ( L , R ) (L,R) ( L , R ) 3 of H H H with ∣ L ∩ R ∣ ≤ k |L \cap R| \le k ∣ L ∩ R ∣ ≤ k satisfies min { ∣ L ∩ X ∣ , ∣ R ∩ X ∣ } ≤ q \min\{|L \cap X|, |R \cap X|\} \le q min { ∣ L ∩ X ∣ , ∣ R ∩ X ∣ } ≤ q . A tree decomposition is ( q , k ) (q,k) ( q , k ) -vertex-unbreakable if every bag is ( q , k ) (q,k) ( q , k ) -vertex-unbreakable. [13 ] constructs in O ∗ ( 2 O ( k 2 ) ) O^*(2^{O(k^2)}) O ∗ ( 2 O ( k 2 ) ) time a ( 2 O ( k ) , k ) (2^{O(k)}, k) ( 2 O ( k ) , k ) -vertex-unbreakable decomposition with adhesion at most 2 O ( k ) 2^{O(k)} 2 O ( k ) , where the adhesion of a tree decomposition is the maximum size of the intersection of two adjacent bags. [11 ] substantially strengthens the decomposition guarantee and constructs in O ∗ ( 2 O ( k log k ) ) O^*(2^{O(k \log k)}) O ∗ ( 2 O ( k l o g k ) ) time a tree decomposition with adhesion at most k k k that is ( i , i ) (i,i) ( i , i ) -vertex-unbreakable for every 1 ≤ i ≤ k 1 \le i \le k 1 ≤ i ≤ k .
Both algorithms then perform bottom-up dynamic programming over their tree decompositions. At a node t t t with parent t ′ t' t ′ , a state records which vertices of the adhesion B t ∩ B t ′ B_t \cap B_{t'} B t ∩ B t ′ lie on one side of the cut, the number of vertices on that side in the subgraph below t t t , and the number of cut edges in that subgraph.
The difficult step is the transition at a large bag B t B_t B t . The unbreakability guarantee in [11 ] ensures that every set A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) with ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \le k ∣ δ G ( A ) ∣ ≤ k satisfies ∣ A ∩ B t ∣ ≤ k |A \cap B_t| \le k ∣ A ∩ B t ∣ ≤ k or ∣ B t ∖ A ∣ ≤ k |B_t \setminus A| \le k ∣ B t ∖ A ∣ ≤ k , but enumerating all such intersections takes n O ( k ) n^{O(k)} n O ( k ) time. Their algorithm avoids this enumeration by color coding. In particular, at most k k k child adhesions are split by the cut, and each has size at most k k k . Hence, the coloring must separate the smaller of A ∩ B t A \cap B_t A ∩ B t and B t ∖ A B_t \setminus A B t ∖ A from O ( k 2 ) O(k^2) O ( k 2 ) vertices on the other side of these adhesions. If each vertex is colored red with probability Θ ( 1 / k ) \Theta(1/k) Θ ( 1/ k ) , the probability that every vertex of the smaller set is red and all O ( k 2 ) O(k^2) O ( k 2 ) opposite-side vertices are blue is 2 − O ( k log k ) 2^{-O(k \log k)} 2 − O ( k l o g k ) . The color-coding step therefore contributes a factor of 2 O ( k log k ) 2^{O(k \log k)} 2 O ( k l o g k ) to the transition time.
Our techniques. In [11 ] , both the construction of the unbreakable decomposition and the transition at a large bag take O ∗ ( 2 O ( k log k ) ) O^*(2^{O(k \log k)}) O ∗ ( 2 O ( k l o g k ) ) time. We perform both steps in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time and thereby obtain a single-exponential algorithm.
We first consider the construction of the unbreakable decomposition. We do not know whether a vertex-unbreakable decomposition with the bounds of [11 ] , that is, adhesions of size at most k k k and ( i , i ) (i,i) ( i , i ) -vertex-unbreakable bags for every 1 ≤ i ≤ k 1 \le i \le k 1 ≤ i ≤ k , can be constructed in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. However, as observed in [11 ] , all applications considered there, including Minimum Bisection, require only the weaker notion of edge-unbreakability. For a graph H H H , a set X ⊆ V ( H ) X \subseteq V(H) X ⊆ V ( H ) is ( q , k ) (q,k) ( q , k ) -edge-unbreakable if min { ∣ A ∩ X ∣ , ∣ X ∖ A ∣ } ≤ q \min\{|A \cap X|, |X \setminus A|\} \le q min { ∣ A ∩ X ∣ , ∣ X ∖ A ∣ } ≤ q for every A ⊆ V ( H ) A \subseteq V(H) A ⊆ V ( H ) with ∣ δ H ( A ) ∣ ≤ k |\delta_H(A)| \le k ∣ δ H ( A ) ∣ ≤ k . A tree decomposition of H H H is ( q , k ) (q,k) ( q , k ) -edge-unbreakable if every bag is ( q , k ) (q,k) ( q , k ) -edge-unbreakable. We construct an edge-unbreakable decomposition in single-exponential time.
Theorem 1.1. Given a graph G G G and k k k , one can construct in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time a ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable decomposition with adhesion at most 2 k 2k 2 k .
We next consider the transition at a large bag. Unlike the previous algorithms, we do not use color coding. We instead apply algebraic operations to formal power series that represent the dynamic-programming tables, including inclusion–exclusion and formal logarithms and exponentials, all of which can be performed in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. These operations cannot be performed on Boolean existence values alone, so each state stores the number of partial solutions that it represents. Consequently, our algorithm counts all feasible solutions to Minimum Bisection rather than merely deciding feasibility.
Theorem 1.2. The feasible solutions to Minimum Bisection can be counted deterministically in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. If a feasible solution exists, one can recover one within the same running time.
Application. Our approach also applies to Matching Cut, parameterized by the number of crossing edges.
Matching Cut
Input: A simple undirected graph G G G and an integer k ≥ 0 k \ge0 k ≥ 0 .
Question: Is there a set ∅ ≠ A ⊂ V ( G ) \varnothing\ne A \subset V(G) ∅ = A ⊂ V ( G ) such that ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \le k ∣ δ G ( A ) ∣ ≤ k , and ∣ δ G ( A ) ∩ δ G ( { v } ) ∣ ≤ 1 |\delta_G(A) \cap\delta_G(\{v\})| \le1 ∣ δ G ( A ) ∩ δ G ({ v }) ∣ ≤ 1 for every v ∈ V ( G ) v \in V(G) v ∈ V ( G ) ?
Parameter: k k k .
Gomes and Sau [18 ] observed that Matching Cut parameterized by k k k reduces to Stable Cutset, for which Marx, O’Sullivan, and Razgon [35 ] gave a fixed-parameter algorithm using Courcelle’s theorem. Aravind and Saxena [4 ] later made the dependence on k k k explicit with an O ∗ ( 2 O ( k log k ) ) O^*(2^{O(k \log k)}) O ∗ ( 2 O ( k l o g k ) ) -time algorithm by applying the framework of [11 ] .
Our algorithm for Matching Cut combines the Minimum Bisection algorithm with color coding that assigns distinct colors to the endpoints of the cut. In particular, we obtain the following result.
Theorem 1.3. Matching Cut can be solved deterministically, and a witness can be recovered, in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Unlike our algorithm for Minimum Bisection, our algorithm for Matching Cut does not count feasible solutions, because the initial color-coding step preserves only the existence of a solution.
Remark. It is natural to ask whether our techniques also yield single-exponential algorithms for other applications of unbreakable decompositions. [11 ] applied the unbreakable-decomposition framework to Steiner Cut, a generalization of p p p -Way Cut, and to Steiner Multicut. As noted above, these applications use only edge-unbreakability, so our decomposition is sufficient for the same framework, up to constant changes in the parameters. This replacement alone, however, does not yield single-exponential algorithms, because the bottleneck is already present in the number of dynamic-programming states. For p p p -Way Cut, a solution may have up to p ≤ k + 1 p \le k+1 p ≤ k + 1 parts, and the direct dynamic program assigns each adhesion vertex to one of these parts, already resulting in p O ( k ) = 2 O ( k log k ) p^{O(k)} = 2^{O(k\log k)} p O ( k ) = 2 O ( k l o g k ) dynamic-programming states, whereas the corresponding number for Minimum Bisection is only 2 O ( k ) 2^{O(k)} 2 O ( k ) . We do not know whether these states can be compressed. Nevertheless, we believe that our construction provides a useful first step toward speeding up other applications of unbreakable decompositions, such as p p p -Way Cut, Steiner Cut, and Steiner Multicut.
Lower bound. Aravind and Saxena [4 ] gave a lower-bound result for Matching Cut parameterized by the number k k k of crossing edges. More quantitatively, let g : N → R ≥ 0 g:\mathbb{N}\to\mathbb{R}_{\ge0} g : N → R ≥ 0 be unbounded and nondecreasing, satisfy g ( x ) = O ( x ) g(x)=O(x) g ( x ) = O ( x ) , and satisfy g ( λ x ) = O ( g ( x ) ) g(\lambda x)=O(g(x)) g ( λ x ) = O ( g ( x )) for every fixed constant λ > 0 \lambda>0 λ > 0 . Then Matching Cut cannot be solved in O ∗ ( 2 o ( g ( k ) ) ) O^*(2^{o(g(k))}) O ∗ ( 2 o ( g ( k )) ) time unless 3-SAT can be solved in O ∗ ( 2 o ( g ( s ) ) ) O^*(2^{o(g(s))}) O ∗ ( 2 o ( g ( s )) ) time on instances of size s s s , where s s s is the total number of variables and clauses in the input 3-CNF formula.4
For Minimum Bisection, we prove Theorem 1.4.
Theorem 1.4. Let g : N → R ≥ 0 g:\mathbb{N}\to\mathbb{R}_{\ge0} g : N → R ≥ 0 be unbounded and nondecreasing, satisfy g ( x ) = O ( x ) g(x)=O(x) g ( x ) = O ( x ) , and satisfy g ( λ x ) = O ( g ( x ) ) g(\lambda x)=O(g(x)) g ( λ x ) = O ( g ( x )) for every fixed constant λ > 0 \lambda>0 λ > 0 . Suppose that Minimum Bisection can be solved in O ∗ ( 2 o ( g ( k ) ) ) O^*(2^{o(g(k))}) O ∗ ( 2 o ( g ( k )) ) time. Then 3-SAT can be solved in 2 o ( g ( q s ) ) + O ( s log q / q + log s ) 2^{o(g(qs))+O(s\log q/q+\log s)} 2 o ( g ( q s )) + O ( s l o g q / q + l o g s ) time for every integer q q q with 2 ≤ q ≤ s 2\le q\le s 2 ≤ q ≤ s , where s s s is the total number of variables and clauses in the input 3-CNF formula.5
In particular, the theorem yields three consequences.
For every fixed 0 < α < 1 0<\alpha<1 0 < α < 1 , if 3-SAT has no O ∗ ( 2 o ( s α ) ) O^*(2^{o(s^\alpha)}) O ∗ ( 2 o ( s α ) ) -time algorithm, then Minimum Bisection has no O ∗ ( 2 o ( k / ( log k ) α / ( 2 − α ) ) ) O^*(2^{o(k/(\log k)^{\alpha/(2-\alpha)})}) O ∗ ( 2 o ( k / ( l o g k ) α / ( 2 − α ) ) ) -time algorithm.
For every fixed c > 0 c>0 c > 0 , if 3-SAT has no O ∗ ( 2 o ( s / ( log s ) c ) ) O^*(2^{o(s/(\log s)^c)}) O ∗ ( 2 o ( s / ( l o g s ) c ) ) -time algorithm, then Minimum Bisection has no O ∗ ( 2 o ( k / ( ( log k ) 2 c log log k ) ) ) O^*(2^{o(k/((\log k)^{2c}\log\log k))}) O ∗ ( 2 o ( k / (( l o g k ) 2 c l o g l o g k )) ) -time algorithm.
If 3-SAT has no O ∗ ( 2 o ( s ) ) O^*(2^{o(s)}) O ∗ ( 2 o ( s ) ) -time algorithm, equivalently, if ETH holds, then Minimum Bisection has no O ∗ ( 2 o ( k ) ) O^*(2^{o(k)}) O ∗ ( 2 o ( k ) ) -time algorithm.
We remark that [11 ] asked whether an O ∗ ( 2 o ( k ) ) O^*(2^{o(k)}) O ∗ ( 2 o ( k ) ) -time algorithm for Minimum Bisection could be ruled out under ETH. The ETH consequence in Theorem 1.4 answers this question affirmatively. However, OpenAI and Anthropic recently released independent manuscripts claiming that ETH is false [38 , 12 ] . If either claim is correct, ETH no longer provides an algorithmic barrier. We nevertheless believe that Theorem 1.4 retains the spirit of the question in [11 ] , because the first and second items transfer near-exponential lower bounds for 3-SAT to near-single-exponential lower bounds for Minimum Bisection, with the exponent remaining arbitrarily close to linear.
Related work
Lean and unbreakable decompositions. A tree decomposition ( T , { B t } t ∈ V ( T ) ) (\mathcal{T}, \{B_t\}_{t\in V(\mathcal{T})}) ( T , { B t } t ∈ V ( T ) ) of a graph G G G is lean if, for all s , t ∈ V ( T ) s,t\in V(\mathcal{T}) s , t ∈ V ( T ) and all Z s ⊆ B s Z_s\subseteq B_s Z s ⊆ B s and Z t ⊆ B t Z_t\subseteq B_t Z t ⊆ B t with ∣ Z s ∣ = ∣ Z t ∣ = ℓ |Z_s|=|Z_t|=\ell ∣ Z s ∣ = ∣ Z t ∣ = ℓ , the graph G G G contains ℓ \ell ℓ pairwise vertex-disjoint Z s Z_s Z s –Z t Z_t Z t paths unless there is an adhesion of size less than ℓ \ell ℓ along the s s s –t t t path in T \mathcal{T} T . Thomas [42 ] introduced the notion of lean tree decompositions and proved that every graph G G G admits a lean tree decomposition of width tw ( G ) \operatorname{tw}(G) tw ( G ) .6 Bellenbaum and Diestel [6 ] gave a shorter proof of Thomas’s theorem. Carmesin, Diestel, Hamann, and Hundertmark [8 ] introduced k k k -lean tree decompositions, a variant of lean tree decompositions that has adhesion at most k k k and restricts the quantification over ℓ \ell ℓ to ℓ ≤ k + 1 \ell\leq k+1 ℓ ≤ k + 1 , and proved that every graph admits a k k k -lean tree decomposition for every k k k . Erde [14 ] developed a common framework for lean-decomposition existence theorems in abstract separation systems and proved that every matroid M M M admits a lean tree decomposition of width tw ( M ) \operatorname{tw}(M) tw ( M ) .
Cygan, Lokshtanov, Marcin Pilipczuk, Michał Pilipczuk, and Saurabh [13 ] introduced the algorithmic notion of an unbreakable decomposition as part of the first fixed-parameter algorithm for Minimum Bisection. Cygan, Komosa, Lokshtanov, Marcin Pilipczuk, Michał Pilipczuk, Saurabh, and Wahlström [11 ] subsequently gave an algorithm that computes a tree decomposition with adhesion at most k k k whose bags are ( i , i ) (i,i) ( i , i ) -vertex-unbreakable for every 1 ≤ i ≤ k 1\leq i\leq k 1 ≤ i ≤ k , and applied the decomposition to Minimum Bisection, Steiner Cut, and Steiner Multicut.
Lokshtanov, Saurabh, and Surianarayanan [LSS24] gave a polynomial-time algorithm that computes a ( ( k + 1 ) 5 , k ) ((k+1)^5,k) (( k + 1 ) 5 , k ) -edge-unbreakable decomposition with adhesion at most k k k , which they used in a parameterized approximation scheme for Minimum p p p -Way Cut. Inamdar, Lokshtanov, Saurabh, and Surianarayanan [ILS23] obtained a ( 9 k , k ) (9k,k) ( 9 k , k ) -edge-unbreakable decomposition with adhesion at most 8 k 8k 8 k and logarithmic depth, and used it to obtain an FPT-approximation algorithm for Fair Bisection. For every 0 < ϵ ≤ 1 0<\epsilon\leq1 0 < ϵ ≤ 1 , Anand, Lee, Li, Long, and Saranurak [2 ] give a randomized algorithm that computes with high probability a ( O ( k / ϵ ) , k ) (O(k/\epsilon),k) ( O ( k / ϵ ) , k ) -vertex-unbreakable decomposition with adhesion O ( k / ϵ ) O(k/\epsilon) O ( k / ϵ ) in 2 O ( ( k / ϵ ) log ( k / ϵ ) ) n 1 + ϵ 2^{O((k/\epsilon)\log(k/\epsilon))}n^{1+\epsilon} 2 O (( k / ϵ ) l o g ( k / ϵ )) n 1 + ϵ time. They use the decomposition to obtain the first close-to-linear-time FPT algorithm for Minimum p p p -Way Cut. Korhonen [28 ] gave the first algorithm for constructing a k k k -lean tree decomposition, which runs in k O ( k 2 ) ( n + m ) k^{O(k^2)}(n+m) k O ( k 2 ) ( n + m ) time. The decomposition is ( i , i ) (i,i) ( i , i ) -vertex-unbreakable for every 1 ≤ i ≤ k 1\leq i\leq k 1 ≤ i ≤ k and has adhesion at most k k k . The construction yields, for every fixed k k k , a linear-time algorithm for computing the k k k -edge-connected components, as well as algorithms for computing k k k -Gomory–Hu trees and testing k k k -vertex-connectivity. Li and Vaughn [33 ] gave a randomized near-linear-time construction of an ( O ( k log 2 n log log n ) , k ) (O(k\log^2 n\log\log n),k) ( O ( k log 2 n log log n ) , k ) -edge-unbreakable decomposition with adhesion O ( k log 2 n log log n ) O(k\log^2 n\log\log n) O ( k log 2 n log log n ) . They use the decomposition to obtain an improved FPT algorithm and a near-linear-time approximation algorithm for Minimum p p p -Way Cut. Most recently, Anand, Lee, Long, and Saranurak [3 ] gave a randomized algorithm that computes, with high probability, a ( 5 k , k ) (5k,k) ( 5 k , k ) -vertex-unbreakable decomposition with adhesion 4 k 4k 4 k in O ∗ ( 2 O ( k log k ) ) O^{*}(2^{O(k\log k)}) O ∗ ( 2 O ( k l o g k ) ) time, with near-linear dependence on the input size. Table 1 summarizes existing results on lean, vertex-unbreakable, and edge-unbreakable decompositions together with our result in Theorem 1.1.
Table 1. Comparison of lean and unbreakable decomposition guarantees. The results in the first two rows prove only existence and give no running-time bounds.
Algorithms for Minimum Bisection. Early algorithmic work on Minimum Bisection concerned heuristics [26 , 24 ] and average-case guarantees [5 ] . For approximation with worst-case guarantees, Krauthgamer and Feige [22 ] obtained approximation ratios O ( log 3 / 2 n ) O(\log^{3/2} n) O ( log 3/2 n ) for Minimum Bisection on general graphs and O ( log n ) O(\log n) O ( log n ) for every fixed-minor-free graph class. Räcke [40 ] improved the approximation ratio for general graphs to O ( log n ) O(\log n) O ( log n ) . Arora, Karger, and Karpinski [1 ] gave a polynomial-time approximation scheme (PTAS) for graphs with minimum degree linear in n n n . Fox, Klein, and Mozes [15 ] gave a bicriteria PTAS for weighted planar instances. Khot [23 ] ruled out a PTAS for Minimum Bisection on general graphs under the assumption N P ⊈ ⋂ ϵ > 0 B P T I M E ( 2 n ϵ ) \mathrm{NP} \not\subseteq\bigcap_{\epsilon>0}\mathrm{BPTIME}(2^{n^{\epsilon}}) NP ⊆ ⋂ ϵ > 0 BPTIME ( 2 n ϵ ) .
For exact computation, Scott and Sorkin [41 ] formulated Minimum Bisection as a polynomial-valued constraint-satisfaction problem and gave a polynomial-space O ∗ ( 2 19 m / 100 ) O^{*}(2^{19m/100}) O ∗ ( 2 19 m /100 ) -time algorithm that also supports counting. Hanaka, Kobayashi, and Sone [19 ] gave an O ( 2 t ( t n ) 2 ) O(2^{t(tn)^{2}}) O ( 2 t ( t n ) 2 ) -time algorithm on a given tree decomposition of width t t t . Cygan, Lokshtanov, Marcin Pilipczuk, Michał Pilipczuk, and Saurabh [13 ] gave the first fixed-parameter algorithm for Minimum Bisection parameterized by the number of cut edges, with running time O ∗ ( 2 O ( k 3 ) ) O^{*}(2^{O(k^{3})}) O ∗ ( 2 O ( k 3 ) ) . Cygan, Komosa, Lokshtanov, Marcin Pilipczuk, Michał Pilipczuk, Saurabh, and Wahlström [11 ] improved the running time to O ∗ ( 2 O ( k log k ) ) O^{*}(2^{O(k\log k)}) O ∗ ( 2 O ( k l o g k ) ) .
Algorithms for Matching Cut. Graham [17 ] introduced decomposable graphs , namely graphs admitting matching cuts, to solve a number-theoretic problem on cube numbering. Subsequent work studied the complexity of Matching Cut on graph classes such as bounded-degree graphs [10 ] , line graphs [36 ] , series-parallel graphs [39 ] , planar graphs, claw-free graphs, cographs, and graphs of bounded treewidth or clique-width [7 ] , bounded-diameter graphs, both general and bipartite [29 ] , bounded-radius graphs and P 6 P_6 P 6 -free graphs [30 ] , and P 15 P_{15} P 15 -free graphs [31 ] .
Exact and parameterized algorithms for general graphs appeared more recently. Kratsch and Le [27 ] gave the first branching algorithm, which runs in O ∗ ( 2 n / 2 ) \mathcal{O}^*(2^{n/2}) O ∗ ( 2 n /2 ) time, and a single-exponential algorithm parameterized by the vertex-cover number. Komusiewicz, Kratsch, and Le [25 ] gave a randomized exact algorithm running in O ∗ ( 1.3071 n ) \mathcal{O}^*(1.3071^n) O ∗ ( 1.307 1 n ) time, a quadratic kernel parameterized by distance to a cluster graph, a linear kernel parameterized by distance to a clique, and single-exponential algorithms parameterized by distance to a cluster graph and distance to a co-cluster graph. Gomes and Sau [18 ] proved that Matching Cut is fixed-parameter tractable with respect to the number k k k of crossing edges by combining the line-graph reduction to Stable Cutset with the fixed-parameter algorithm of Marx, O’Sullivan, and Razgon [35 ] for Stable Cutset parameterized by the cutset size. Aravind and Saxena [4 ] gave an O ∗ ( 2 O ( k log k ) ) \mathcal{O}^*(2^{\mathcal{O}(k\log k)}) O ∗ ( 2 O ( k l o g k ) ) -time algorithm.
Technical overview
A single-exponential algorithm for Minimum Bisection
Dynamic programming. Recall that Theorem 1.1 constructs a rooted tree decomposition whose adhesions have size at most 2 k 2k 2 k and whose bags are ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable. Fix such a decomposition T \mathcal{T} T . For a node t t t of T \mathcal{T} T , let ∂ G t \partial G_t ∂ G t be the adhesion between the bags corresponding to t t t and the parent of t t t , with ∂ G t : = ∅ \partial G_t := \varnothing ∂ G t := ∅ at the root.
We perform bottom-up dynamic programming on T \mathcal{T} T . We use the standard representation of a cut on an adhesion: for every P ⊆ ∂ G t P \subseteq\partial G_t P ⊆ ∂ G t , we introduce a state indexed by P P P , which stores information about cuts A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) satisfying A ∩ ∂ G t = P A \cap\partial G_t = P A ∩ ∂ G t = P . Since ∣ ∂ G t ∣ ≤ 2 k |\partial G_t| \leq2k ∣ ∂ G t ∣ ≤ 2 k , there are at most 2 2 k = 2 O ( k ) 2^{2k} = 2^{\mathcal{O}(k)} 2 2 k = 2 O ( k ) choices for P P P . More precisely, the dynamic programming keeps two additional scalar indices, which we suppress in this overview: the number of vertices on the chosen side, needed for the bisection constraint, and the number of cut edges. They have only polynomially many possible values and therefore increase the number of states only by a polynomial factor.
In the usual dynamic programming on a tree decomposition of bounded width, a single-exponential bound on the number of states typically yields a single-exponential running time. This is not the case here, because T \mathcal{T} T bounds only the adhesion size and its bags can be arbitrarily large. Hence we need to design the transitions carefully to obtain a single-exponential running time.
Consider computing the table at a node t t t from the already computed tables of its children, and let B t B_t B t denote the bag corresponding to t t t . A naive way to compute the transition is to enumerate the possible sets R : = A ∩ B t R := A \cap B_t R := A ∩ B t for a hypothetical cut A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) . Once R R R is fixed, every child s s s must use the state indexed by R ∩ ∂ G s R \cap\partial G_s R ∩ ∂ G s . Hence the contribution of cuts satisfying A ∩ B t = R A \cap B_t = R A ∩ B t = R can be computed from the child table entries indexed by R ∩ ∂ G s R \cap\partial G_s R ∩ ∂ G s and added to the entry at t t t indexed by R ∩ ∂ G t R \cap\partial G_t R ∩ ∂ G t . Figure 1 illustrates A A A , the set R = A ∩ B t R = A \cap B_t R = A ∩ B t , and the intersections of B t B_t B t with the adjacent bags.
Figure 1. A bag B t B_t B t , its parent bag B t ′ B_{t'} B t ′ , and three child bags B s 1 B_{s_1} B s 1 , B s 2 B_{s_2} B s 2 , B s 3 B_{s_3} B s 3 . The shaded regions enclose the three connected components of G [ A ] G[A] G [ A ] . Blue vertices form R = A ∩ B t R = A \cap B_t R = A ∩ B t .
We now analyze the number of possible sets R R R . A straightforward approach enumerates all R ⊆ B t R \subseteq B_t R ⊆ B t , which gives 2 ∣ B t ∣ 2^{|B_t|} 2 ∣ B t ∣ possibilities and can be as large as 2 n 2^n 2 n , which is far too large. Instead of enumerating every R ⊆ B t R \subseteq B_t R ⊆ B t , we observe that only cuts with at most k k k cut edges need to be considered, since a cut with more than k k k cut edges cannot be a feasible solution. For every such cut A A A , the ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakability of B t B_t B t gives min { ∣ A ∩ B t ∣ , ∣ B t ∖ A ∣ } ≤ 2 k \min\{|A \cap B_t|, |B_t \setminus A|\} \leq2k min { ∣ A ∩ B t ∣ , ∣ B t ∖ A ∣ } ≤ 2 k . Hence the number of possible sets R R R can be reduced to 2 ∑ i = 0 2 k ( ∣ B t ∣ i ) = n O ( k ) 2\sum_{i=0}^{2k}\binom{|B_t|}{i} = n^{\mathcal{O}(k)} 2 ∑ i = 0 2 k ( i ∣ B t ∣ ) = n O ( k ) . However, this is still too large.
Our key observation is that, we can break this n O ( k ) n^{\mathcal{O}(k)} n O ( k ) bound when we only consider cuts A A A for which G [ A ] G[A] G [ A ] is connected. More precisely, in O ∗ ( 2 O ( k ) ) \mathcal{O}^*(2^{\mathcal{O}(k)}) O ∗ ( 2 O ( k ) ) time we can compute a family C t ⊆ 2 B t \mathcal{C}_t \subseteq2^{B_t} C t ⊆ 2 B t of size O ∗ ( 2 O ( k ) ) \mathcal{O}^*(2^{\mathcal{O}(k)}) O ∗ ( 2 O ( k ) ) such that for every cut A A A for which G [ A ] G[A] G [ A ] is connected, ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \leq k ∣ δ G ( A ) ∣ ≤ k , and ∣ A ∩ B t ∣ ≤ 2 k |A \cap B_t| \leq2k ∣ A ∩ B t ∣ ≤ 2 k , the set R = A ∩ B t R = A \cap B_t R = A ∩ B t belongs to C t \mathcal{C}_t C t . To see this, we use important cuts, introduced by Marx [34 ] . For distinct vertices x , y x,y x , y in a graph H H H , a connected x x x -y y y cut C C C is important if no connected x x x -$y cut C ′ ⊋ C C' \supsetneq C C ′ ⊋ C satisfies ∣ δ H ( C ′ ) ∣ ≤ ∣ δ H ( C ) ∣ |\delta_H(C')| \leq|\delta_H(C)| ∣ δ H ( C ′ ) ∣ ≤ ∣ δ H ( C ) ∣ .7 It is known that every connected x x x –y y y cut C C C is contained in an important cut D D D with ∣ δ H ( D ) ∣ ≤ ∣ δ H ( C ) ∣ |\delta_H(D)| \leq|\delta_H(C)| ∣ δ H ( D ) ∣ ≤ ∣ δ H ( C ) ∣ , and that there are at most 4 ℓ 4^\ell 4 ℓ important x x x –y y y cuts with at most ℓ \ell ℓ cut edges (Lemma 2.1 ); see [9 Section 8.2] for further details.
To construct C t C_t C t , form H t H_t H t from G G G by adding a vertex r t r_t r t adjacent to every vertex of B t B_t B t . For each v ∈ B t v \in B_t v ∈ B t , enumerate the important connected v v v –r t r_t r t cuts D D D with at most 3 k 3k 3 k cut edges, and add every subset of D ∩ B t D \cap B_t D ∩ B t to C t C_t C t . There are at most ∣ B t ∣ 4 ⋅ 3 k |B_t|^{4\cdot3k} ∣ B t ∣ 4 ⋅ 3 k choices for D D D . Since r t ∉ D r_t \notin D r t ∈ / D and every vertex of D ∩ B t D \cap B_t D ∩ B t is adjacent to r t r_t r t , we have ∣ D ∩ B t ∣ ≤ ∣ δ H t ( D ) ∣ ≤ 3 k |D \cap B_t| \leq|\delta_{H_t}(D)| \leq3k ∣ D ∩ B t ∣ ≤ ∣ δ H t ( D ) ∣ ≤ 3 k . Thus C t C_t C t has size O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) and can be computed within the same time bound. Now let A A A satisfy G [ A ] G[A] G [ A ] connected, ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \leq k ∣ δ G ( A ) ∣ ≤ k , and 0 < ∣ A ∩ B t ∣ ≤ 2 k 0 < |A \cap B_t| \leq2k 0 < ∣ A ∩ B t ∣ ≤ 2 k , and choose v ∈ A ∩ B t v \in A \cap B_t v ∈ A ∩ B t . In H t H_t H t , the set A A A is a connected v v v –r t r_t r t cut with ∣ δ H t ( A ) ∣ = ∣ δ G ( A ) ∣ + ∣ A ∩ B t ∣ ≤ 3 k |\delta_{H_t}(A)| = |\delta_G(A)| + |A \cap B_t| \leq3k ∣ δ H t ( A ) ∣ = ∣ δ G ( A ) ∣ + ∣ A ∩ B t ∣ ≤ 3 k . Since every connected v v v –r t r_t r t cut is contained in an important v v v –r t r_t r t cut with no more cut edges, there is an important v v v –r t r_t r t cut D D D with A ⊆ D A \subseteq D A ⊆ D and ∣ δ H t ( D ) ∣ ≤ ∣ δ H t ( A ) ∣ ≤ 3 k |\delta_{H_t}(D)| \leq|\delta_{H_t}(A)| \leq3k ∣ δ H t ( D ) ∣ ≤ ∣ δ H t ( A ) ∣ ≤ 3 k . Hence D D D is enumerated in the construction of C t C_t C t , and A ∩ B t ⊆ D ∩ B t A \cap B_t \subseteq D \cap B_t A ∩ B t ⊆ D ∩ B t gives A ∩ B t ∈ C t A \cap B_t \in C_t A ∩ B t ∈ C t .
Formal logarithm. Before describing our use of the formal logarithm, we give a standard example of how the formal logarithm recovers a generating function for connected objects from one that counts arbitrary objects. The example is not used in our algorithm for Minimum Bisection: our algorithm uses a different generating polynomial and proves connected support by another argument. The example only provides the underlying intuition.
Let f n = 2 ( n 2 ) f_n = 2^{\binom{n}{2}} f n = 2 ( 2 n ) be the number of graphs on the labeled vertex set [ n ] [n] [ n ] , and let g n g_n g n be the number of connected such graphs. Every graph is uniquely determined by the partition of its vertex set into connected components and a connected graph on each part. Therefore, for every n ≥ 1 n \geq1 n ≥ 1 ,
f n = ∑ { B 1 , … , B r } ∏ j = 1 r g ∣ B j ∣ = n ! ∑ r = 1 n ( 1 r ! ∑ n 1 + ⋯ + n r = n n 1 , … , n r ≥ 1 ∏ j = 1 r g n j n j ! ) . f_n = \sum_{\{B_1,\ldots,B_r\}} \prod_{j=1}^{r} g_{|B_j|} = n! \sum_{r=1}^{n} \left( \frac{1}{r!} \sum_{\substack{n_1+\cdots+n_r=n \\ n_1,\ldots,n_r\geq1}} \prod_{j=1}^{r} \frac{g_{n_j}}{n_j!} \right). f n = { B 1 , … , B r } ∑ j = 1 ∏ r g ∣ B j ∣ = n ! r = 1 ∑ n r ! 1 n 1 + ⋯ + n r = n n 1 , … , n r ≥ 1 ∑ j = 1 ∏ r n j ! g n j .
where the sum ∑ { B 1 , … , B r } \sum_{\{B_1,\ldots,B_r\}} ∑ { B 1 , … , B r } ranges over all partitions { B 1 , … , B r } \{B_1,\ldots,B_r\} { B 1 , … , B r } of [ n ] [n] [ n ] . The factor n ! / ( n 1 ! ⋯ n r ! ) n!/(n_1! \cdots n_r!) n ! / ( n 1 ! ⋯ n r !) assigns the labeled vertices to components of the specified sizes, while 1 / r ! 1/r! 1/ r ! removes the ordering of the components.
Define the exponential generating functions
F ( x ) : = ∑ n ≥ 0 f n x n n ! , G ( x ) : = ∑ n ≥ 1 g n x n n ! . F(x) := \sum_{n\geq0} f_n \frac{x^n}{n!}, \qquad G(x) := \sum_{n\geq1} g_n \frac{x^n}{n!}. F ( x ) := n ≥ 0 ∑ f n n ! x n , G ( x ) := n ≥ 1 ∑ g n n ! x n .
Substituting the expression for f n f_n f n and reordering the sums gives
F ( x ) = 1 + ∑ n ≥ 1 ∑ r = 1 n ( 1 r ! ∑ n 1 + ⋯ + n r = n n 1 , … , n r ≥ 1 ∏ j = 1 r g n j n j ! ) x n = 1 + ∑ r ≥ 1 1 r ! ∑ n 1 , … , n r ≥ 1 ∏ j = 1 r g n j x n j n j ! = ∑ r ≥ 0 G ( x ) r r ! . \begin{aligned}
F(x) = 1 + \sum_{n\geq1}\sum_{r=1}^{n}\left(\frac{1}{r!}\sum_{\substack{n_1+\cdots+n_r=n\\n_1,\ldots,n_r\geq1}}\prod_{j=1}^{r}\frac{g_{n_j}}{n_j!}\right)x^n \\
&= 1 + \sum_{r\geq1}\frac{1}{r!}\sum_{n_1,\ldots,n_r\geq1}\prod_{j=1}^{r}\frac{g_{n_j}x^{n_j}}{n_j!} = \sum_{r\geq0}\frac{G(x)^r}{r!}.
\end{aligned} F ( x ) = 1 + n ≥ 1 ∑ r = 1 ∑ n r ! 1 n 1 + ⋯ + n r = n n 1 , … , n r ≥ 1 ∑ j = 1 ∏ r n j ! g n j x n = 1 + r ≥ 1 ∑ r ! 1 n 1 , … , n r ≥ 1 ∑ j = 1 ∏ r n j ! g n j x n j = r ≥ 0 ∑ r ! G ( x ) r .
The series ∑ r ≥ 0 G ( x ) r / r ! \sum_{r\geq0}G(x)^r/r! ∑ r ≥ 0 G ( x ) r / r ! is the Taylor series of the exponential function evaluated formally at G ( x ) G(x) G ( x ) . We write exp H : = ∑ r ≥ 0 H r / r ! \exp H := \sum_{r\geq0} H^r/r! exp H := ∑ r ≥ 0 H r / r ! and log H : = ∑ r ≥ 1 ( − 1 ) r + 1 ( H − 1 ) r / r \log H := \sum_{r\geq1}(-1)^{r+1}(H-1)^r/r log H := ∑ r ≥ 1 ( − 1 ) r + 1 ( H − 1 ) r / r for the formal exponential and logarithm, respectively. Then the equality F ( x ) = ∑ r ≥ 0 G ( x ) r / r ! F(x) = \sum_{r\geq0}G(x)^r/r! F ( x ) = ∑ r ≥ 0 G ( x ) r / r ! gives F ( x ) = exp G ( x ) F(x) = \exp G(x) F ( x ) = exp G ( x ) . Since the formal exponential and logarithm are inverse to each other, it follows that G ( x ) = log F ( x ) G(x) = \log F(x) G ( x ) = log F ( x ) . Thus, the formal logarithm recovers the generating function for connected labeled graphs from the generating function for all labeled graphs.
Now we return to Minimum Bisection. The construction of C t C_t C t computes a family in 2 B t 2^{B_t} 2 B t of size O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) containing A ∩ B t A\cap B_t A ∩ B t for every cut A A A such that G [ A ] G[A] G [ A ] is connected, ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)|\leq k ∣ δ G ( A ) ∣ ≤ k , and ∣ A ∩ B t ∣ ≤ 2 k |A\cap B_t|\leq2k ∣ A ∩ B t ∣ ≤ 2 k . To handle an arbitrary cut A A A , we need to convolve the contributions of the connected components of G [ A ] G[A] G [ A ] . A direct convolution again considers the n O ( k ) n^{O(k)} n O ( k ) possible unions R = A ∩ B t R=A\cap B_t R = A ∩ B t , so it does not yield a single-exponential running time. We instead use a technique combining the formal logarithm and degree truncation of the generating polynomial.
To focus on the algebraic intuition, we temporarily forget about dynamic programming and consider the special case in which the entire vertex set V ( G ) V(G) V ( G ) forms a single ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable bag. We also drop the bisection constraint and instead study a generating polynomial that represents every cut ( A , V ( G ) ∖ A ) (A,V(G)\setminus A) ( A , V ( G ) ∖ A ) . Indeed, if A ⊆ V ( G ) A\subseteq V(G) A ⊆ V ( G ) satisfies ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)|\leq k ∣ δ G ( A ) ∣ ≤ k and ∣ A ∣ = ∣ V ( G ) ∖ A ∣ |A|=|V(G)\setminus A| ∣ A ∣ = ∣ V ( G ) ∖ A ∣ , edge-unbreakability gives ∣ A ∣ ≤ 2 k |A|\leq2k ∣ A ∣ ≤ 2 k and hence ∣ V ( G ) ∣ ≤ 4 k |V(G)|\leq4k ∣ V ( G ) ∣ ≤ 4 k , but in this case, all 2 ∣ V ( G ) ∣ ≤ 2 4 k 2^{|V(G)|}\leq2^{4k} 2 ∣ V ( G ) ∣ ≤ 2 4 k choices of A A A can be enumerated directly, so there is nothing to do.
Introduce a variable y v y_v y v for every vertex v ∈ V ( G ) v\in V(G) v ∈ V ( G ) , and write y = ( y v ) v ∈ V ( G ) \mathbf{y}=(y_v)_{v\in V(G)} y = ( y v ) v ∈ V ( G ) . For every S ⊆ V ( G ) S\subseteq V(G) S ⊆ V ( G ) , define
Z S ( y , w ) : = ∑ A ⊆ S ( w ∣ δ G ( A ) ∣ ⋅ ∏ v ∈ A y v ) . Z_S(\mathbf{y},w) := \sum_{A\subseteq S}\left(w^{|\delta_G(A)|}\cdot\prod_{v\in A}y_v\right). Z S ( y , w ) := A ⊆ S ∑ ( w ∣ δ G ( A ) ∣ ⋅ v ∈ A ∏ y v ) .
Every term of Z S ( y , w ) Z_S(\mathbf{y},w) Z S ( y , w ) represents a set A ⊆ S A\subseteq S A ⊆ S . The vertex variables record A A A , and the exponent of w w w records ∣ δ G ( A ) ∣ |\delta_G(A)| ∣ δ G ( A ) ∣ . The coefficient of w c y a w^c y^a w c y a in the polynomial obtained from Z V ( G ) ( y , w ) Z_{V(G)}(\mathbf{y},w) Z V ( G ) ( y , w ) by the substitution y v ↦ y y_v\mapsto y y v ↦ y for every $v\in V(G) equals the number of sets A ⊆ V ( G ) A\subseteq V(G) A ⊆ V ( G ) with ∣ δ G ( A ) ∣ = c |\delta_G(A)|=c ∣ δ G ( A ) ∣ = c and ∣ A ∣ = a |A|=a ∣ A ∣ = a . The goal here is to compute all such coefficients.
We consider the formal logarithm of the polynomial Z V ( G ) ( y , w ) Z_{V(G)}(\mathbf{y},w) Z V ( G ) ( y , w ) , defined by
log Z V ( G ) ( y , w ) : = ∑ j ≥ 1 ( − 1 ) j + 1 j ( Z V ( G ) ( y , w ) − 1 ) j . (1) \log Z_{V(G)}(\mathbf{y},w) := \sum_{j\geq1}\frac{(-1)^{j+1}}{j}\left(Z_{V(G)}(\mathbf{y},w)-1\right)^j. \tag*{(1)} log Z V ( G ) ( y , w ) := j ≥ 1 ∑ j ( − 1 ) j + 1 ( Z V ( G ) ( y , w ) − 1 ) j . (1)
The key observation is that every monomial occurring in log Z V ( G ) ( y , w ) \log Z_{V(G)}(\mathbf{y},w) log Z V ( G ) ( y , w ) has connected support, where the support of w a ∏ v ∈ V ( G ) y v a v w^a\prod_{v\in V(G)}y_v^{a_v} w a ∏ v ∈ V ( G ) y v a v is defined as { v ∈ V ( G ) : a v > 0 } \{v\in V(G):a_v>0\} { v ∈ V ( G ) : a v > 0 } . To prove this, fix $S\subseteq V ( G ) V(G) V ( G ) . The coefficient of a monomial with support exactly S S S is the same in log Z V ( G ) ( y , w ) \log Z_{V(G)}(y,w) log Z V ( G ) ( y , w ) and log Z S ( y , w ) \log Z_S(y,w) log Z S ( y , w ) , since a product contributing to such a monomial cannot use a variable y v y_v y v with v ∉ S v \notin S v ∈ / S .
Let S 1 , … , S r S_1,\ldots,S_r S 1 , … , S r be the connected components of G [ S ] G[S] G [ S ] . For A ⊆ S A \subseteq S A ⊆ S , put A i : = A ∩ S i A_i := A \cap S_i A i := A ∩ S i . No edge joins distinct sets S i S_i S i , so A A A is the disjoint union of A 1 , … , A r A_1,\ldots,A_r A 1 , … , A r and δ G ( A ) \delta_G(A) δ G ( A ) is the disjoint union of δ G ( A 1 ) , … , δ G ( A r ) \delta_G(A_1),\ldots,\delta_G(A_r) δ G ( A 1 ) , … , δ G ( A r ) . Consequently,
w ∣ δ G ( A ) ∣ ∏ v ∈ A y v = ∏ i = 1 r ( w ∣ δ G ( A i ) ∣ ∏ v ∈ A i y v ) . w^{|\delta_G(A)|}\prod_{v\in A} y_v = \prod_{i=1}^{r}\left(w^{|\delta_G(A_i)|}\prod_{v\in A_i} y_v\right). w ∣ δ G ( A ) ∣ v ∈ A ∏ y v = i = 1 ∏ r ( w ∣ δ G ( A i ) ∣ v ∈ A i ∏ y v ) .
The sum of the monomial factorization over all choices A i ⊆ S i A_i \subseteq S_i A i ⊆ S i gives
Z S ( y , w ) = ∏ i = 1 r Z S i ( y , w ) . Z_S(y,w)=\prod_{i=1}^{r}Z_{S_i}(y,w). Z S ( y , w ) = i = 1 ∏ r Z S i ( y , w ) .
Taking formal logarithms turns the product into a sum:
log Z S ( y , w ) = ∑ i = 1 r log Z S i ( y , w ) . \log Z_S(y,w)=\sum_{i=1}^{r}\log Z_{S_i}(y,w). log Z S ( y , w ) = i = 1 ∑ r log Z S i ( y , w ) .
Suppose that G [ S ] G[S] G [ S ] is disconnected. Then r ≥ 2 r \ge2 r ≥ 2 and each S i S_i S i is a proper subset of S S S . Every monomial in log Z S i ( y , w ) \log Z_{S_i}(y,w) log Z S i ( y , w ) has support contained in S i S_i S i , so the coefficient of any monomial with support exactly S S S is zero in log Z S ( y , w ) \log Z_S(y,w) log Z S ( y , w ) . Since such coefficients are equal in log Z S ( y , w ) \log Z_S(y,w) log Z S ( y , w ) and log Z V ( G ) ( y , w ) \log Z_{V(G)}(y,w) log Z V ( G ) ( y , w ) , they are also zero in log Z V ( G ) ( y , w ) \log Z_{V(G)}(y,w) log Z V ( G ) ( y , w ) . Hence every monomial with a nonzero coefficient in log Z V ( G ) ( y , w ) \log Z_{V(G)}(y,w) log Z V ( G ) ( y , w ) has connected support.
Although every monomial in log Z V ( G ) ( y , w ) \log Z_{V(G)}(y,w) log Z V ( G ) ( y , w ) has connected support, the series may still have many distinct supports and, even worse, infinitely many terms with the same support, so we cannot compute it directly. To resolve this, we truncate the series by discarding every monomial whose w w w -degree exceeds k k k or whose total y y y -degree exceeds 2 k 2k 2 k . The goal is to show that only O ∗ ( 2 O ( k ) ) \mathcal{O}^*(2^{\mathcal{O}(k)}) O ∗ ( 2 O ( k ) ) monomial supports remain after truncation. To this end, we use the family C t \mathcal{C}_t C t constructed for the bag B t B_t B t . In the situation considered here, B t = V ( G ) B_t=V(G) B t = V ( G ) , so we denote C t \mathcal{C}_t C t simply by C \mathcal{C} C . As described earlier, the family C \mathcal{C} C can be computed in O ∗ ( 2 O ( k ) ) \mathcal{O}^*(2^{\mathcal{O}(k)}) O ∗ ( 2 O ( k ) ) time, has size O ∗ ( 2 O ( k ) ) \mathcal{O}^*(2^{\mathcal{O}(k)}) O ∗ ( 2 O ( k ) ) , and contains every A ⊆ V ( G ) A\subseteq V(G) A ⊆ V ( G ) such that G [ A ] G[A] G [ A ] is connected, ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)|\leq k ∣ δ G ( A ) ∣ ≤ k , and ∣ A ∣ ≤ 2 k |A|\leq2k ∣ A ∣ ≤ 2 k .
Consider a retained monomial m m m , and let S S S be its support. We show that S ∈ C S\in\mathcal{C} S ∈ C . We already know that G [ S ] G[S] G [ S ] is connected, so it remains to show that ∣ S ∣ ≤ 2 k |S|\leq2k ∣ S ∣ ≤ 2 k and ∣ δ G ( S ) ∣ ≤ k |\delta_G(S)|\leq k ∣ δ G ( S ) ∣ ≤ k . The truncation at total y y y -degree 2 k 2k 2 k gives ∣ S ∣ ≤ ∑ v ∈ V ( G ) deg y v ( m ) ≤ 2 k |S|\leq\sum_{v\in V(G)}\deg_{y_v}(m)\leq2k ∣ S ∣ ≤ ∑ v ∈ V ( G ) deg y v ( m ) ≤ 2 k . Moreover, every product that occurs in the expansion of a summand of Equation (1) and contributes to m m m counts every edge of δ G ( S ) \delta_G(S) δ G ( S ) at least once in its w w w -degree. Hence ∣ δ G ( S ) ∣ ≤ deg w ( m ) ≤ k |\delta_G(S)|\leq\deg_w(m)\leq k ∣ δ G ( S ) ∣ ≤ deg w ( m ) ≤ k . Thus S ∈ C S\in\mathcal{C} S ∈ C .
Consequently, we can compute the truncated logarithm as
⟨ log Z V ( G ) ( y , w ) ⟩ = ∑ T ∈ C L T ( y , w ) . \left\langle\log Z_{V(G)}(y,w)\right\rangle=\sum_{T\in\mathcal{C}}L_T(y,w). ⟨ log Z V ( G ) ( y , w ) ⟩ = T ∈ C ∑ L T ( y , w ) .
Here ⟨ F ⟩ \langle F\rangle ⟨ F ⟩ discards every monomial of F F F whose w w w -degree exceeds k k k or whose total y y y -degree exceeds 2 k 2k 2 k , and L T ( y , w ) L_T(y,w) L T ( y , w ) is the sum of the monomials in ⟨ log Z V ( G ) ( y , w ) ⟩ \left\langle\log Z_{V(G)}(y,w)\right\rangle ⟨ log Z V ( G ) ( y , w ) ⟩ with support exactly T T T . It is not hard to see that we can compute L T ( y , w ) L_T(y,w) L T ( y , w ) in O ∗ ( 2 O ( k ) ) \mathcal{O}^*(2^{\mathcal{O}(k)}) O ∗ ( 2 O ( k ) ) time using subset Möbius inversion. Since ∣ C ∣ = O ∗ ( 2 O ( k ) ) |\mathcal{C}|=\mathcal{O}^*(2^{\mathcal{O}(k)}) ∣ C ∣ = O ∗ ( 2 O ( k ) ) , the truncated logarithm is computable in O ∗ ( 2 O ( k ) ) \mathcal{O}^*(2^{\mathcal{O}(k)}) O ∗ ( 2 O ( k ) ) time.
Recall that the goal here is to compute the coefficients of w c y a w^c y^a w c y a for c ≤ k c\leq k c ≤ k and a ≤ 2 k a\leq2k a ≤ 2 k in the polynomial obtained from Z V ( G ) ( y , w ) Z_{V(G)}(y,w) Z V ( G ) ( y , w ) by substituting y v ↦ y y_v\mapsto y y v ↦ y for every v ∈ V ( G ) v\in V(G) v ∈ V ( G ) . To this end, we first make this substitution in the truncated logarithm. We then take the formal exponential, defined by
exp ( ⟨ log Z V ( G ) ( y , w ) ⟩ ) : = ∑ j ≥ 0 1 j ! ⟨ log Z V ( G ) ( y , w ) ⟩ j . \exp\left(\left\langle\log Z_{V(G)}(\mathbf{y},w)\right\rangle\right) := \sum_{j\ge0}\frac{1}{j!}\left\langle\log Z_{V(G)}(\mathbf{y},w)\right\rangle^j. exp ( ⟨ log Z V ( G ) ( y , w ) ⟩ ) := j ≥ 0 ∑ j ! 1 ⟨ log Z V ( G ) ( y , w ) ⟩ j .
Using the truncated logarithm–exponential identity
⟨ exp ( ⟨ log Z V ( G ) ( y , w ) ⟩ ) ⟩ = ⟨ Z V ( G ) ( y , w ) ⟩ , \left\langle\exp\left(\left\langle\log Z_{V(G)}(\mathbf{y},w)\right\rangle\right)\right\rangle= \left\langle Z_{V(G)}(\mathbf{y},w)\right\rangle, ⟨ exp ( ⟨ log Z V ( G ) ( y , w ) ⟩ ) ⟩ = ⟨ Z V ( G ) ( y , w ) ⟩ ,
we can compute all desired coefficients. The substitution y v ↦ y y_v \mapsto y y v ↦ y sends total y y y -degree to y y y -degree, so truncation does not discard any coefficient of w c y a w^c y^a w c y a with c ≤ k c \le k c ≤ k and a ≤ 2 k a \le2k a ≤ 2 k .
We now return to the dynamic-programming transition at a large bag B t B_t B t . In addition to B t B_t B t , the transition must account for the vertices contained in the child subtrees, so some additional care is needed. Nevertheless, the same technique as described above computes the transition in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time (§ 3 ). Hence the bottom-up dynamic programming solves Minimum Bisection in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Unbreakable decomposition
The remaining task is to construct a rooted ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable decomposition with adhesions of size at most 2 k 2k 2 k in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. Our strategy is as follows. If V ( G ) V(G) V ( G ) is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in G G G , we use V ( G ) V(G) V ( G ) as the only bag. Otherwise, we find a set B ⊊ V ( G ) B \subsetneq V(G) B ⊊ V ( G ) such that B B B is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in G G G and has adhesion at most 2 k 2k 2 k , use B B B as the root bag, and recurse on every component of G − B G-B G − B .
For this overview, we focus on constructing the root bag. Choose an arbitrary vertex v ∈ V ( G ) v \in V(G) v ∈ V ( G ) that must remain in the root bag, and start with B : = V ( G ) B := V(G) B := V ( G ) . If B B B is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable, we stop. Otherwise, we find a set B ′ B' B ′ such that v ∈ B ′ v \in B' v ∈ B ′ , ∣ B ′ ∣ < ∣ B ∣ |B'| < |B| ∣ B ′ ∣ < ∣ B ∣ , and B ′ B' B ′ has adhesion at most 2 k 2k 2 k , and replace B B B by B ′ B' B ′ . Each replacement decreases the bag size, so the process eventually produces the root bag.
Suppose that B B B is not ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable. Then, there is a cut with at most k k k cut edges and more than 2 k 2k 2 k vertices of B B B on each side. We choose the side A A A that does not contain v v v . The components of G [ A ] G[A] G [ A ] together contain more than 2 k 2k 2 k vertices of B B B , while their cut-edge sets contain at most k k k edges in total; hence some component K K K satisfies ∣ K ∩ B ∣ > ∣ δ G ( K ) ∣ |K \cap B| > |\delta_G(K)| ∣ K ∩ B ∣ > ∣ δ G ( K ) ∣ . We can prove that adding at most ∣ δ G ( K ) ∣ |\delta_G(K)| ∣ δ G ( K ) ∣ vertices to B ∖ K B \setminus K B ∖ K gives a set B ′ B' B ′ that contains v v v and has adhesion at most 2 k 2k 2 k . Since ∣ δ G ( K ) ∣ < ∣ K ∩ B ∣ |\delta_G(K)| < |K \cap B| ∣ δ G ( K ) ∣ < ∣ K ∩ B ∣ , we also have ∣ B ′ ∣ < ∣ B ∣ |B'| < |B| ∣ B ′ ∣ < ∣ B ∣ . Figure 2 illustrates A A A , B B B , and K K K .
Figure 2. A cut A A A and a component K K K of G [ A ] G[A] G [ A ] , for k = 3 k=3 k = 3 . The horizontal line separates A A A from V ( G ) ∖ A V(G) \setminus A V ( G ) ∖ A , the dashed line encloses the vertices of B B B , and the blue region marks K K K . The three orange edges form δ G ( A ) \delta_G(A) δ G ( A ) .
The remaining task is to find such a component K K K . Note that the witness cut A A A is not given, and finding it directly is already nontrivial. We instead use important cuts [34 ] . We can prove that, for some choice of r ∈ B r \in B r ∈ B , one of the important r r r –v v v cuts with at most k k k cut edges yields a suitable K K K . There are at most 4 k 4^k 4 k such cuts for each r r r . Thus we enumerate them for every r ∈ B r \in B r ∈ B and test the bag obtained from each cut. Whenever B B B is not ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable, at least one enumerated cut yields a valid smaller bag.
For a nonroot recursive call, the new root bag must contain the adhesion to its parent; otherwise, the bags containing an adhesion vertex may fail to form a connected subtree. To this end, instead of preserving the single vertex v v v , we preserve a set X X X consisting of the parent adhesion together with one vertex of the current component. Since the parent adhesion has size at most 2 k 2k 2 k , we have ∣ X ∣ ≤ 2 k + 1 |X| \le2k+1 ∣ X ∣ ≤ 2 k + 1 . The same shrinking procedure works with X X X in place of { v } \{v\} { v } , and recursion gives the complete decomposition. § 5 gives the details. The running time is dominated by the number of important cuts, which is O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) , and this proves Theorem 1.1.
Organization
The rest of this paper is organized as follows. § 2 collects the preliminaries. § 3 proves Theorem 1.2 by giving a single-exponential algorithm for Minimum Bisection, assuming Theorem 1.1. § 4 proves Theorem 1.3 by applying an algorithm similar to that in § 3 to Matching Cut. § 5 proves Theorem 1.1. § 6 proves Theorem 1.4.
Preliminaries
Graphs. All graphs in this paper are finite, simple, and undirected. A cut of a graph H H H is a set A ⊆ V ( H ) A \subseteq V(H) A ⊆ V ( H ) . Its cut-edge set δ H ( A ) \delta_H(A) δ H ( A ) consists of the edges with exactly one endpoint in A A A . The function A ↦ ∣ δ H ( A ) ∣ A \mapsto|\delta_H(A)| A ↦ ∣ δ H ( A ) ∣ is submodular: ∣ δ H ( A ) ∣ + ∣ δ H ( B ) ∣ ≥ ∣ δ H ( A ∩ B ) ∣ + ∣ δ H ( A ∪ B ) ∣ |\delta_H(A)| + |\delta_H(B)| \ge|\delta_H(A \cap B)| + |\delta_H(A \cup B)| ∣ δ H ( A ) ∣ + ∣ δ H ( B ) ∣ ≥ ∣ δ H ( A ∩ B ) ∣ + ∣ δ H ( A ∪ B ) ∣ for all A , B ⊆ V ( H ) A, B \subseteq V(H) A , B ⊆ V ( H ) . A vertex separation of H H H is a pair ( L , R ) (L, R) ( L , R ) with L ∪ R = V ( H ) L \cup R = V(H) L ∪ R = V ( H ) and no edge from L ∖ R L \setminus R L ∖ R to R ∖ L R \setminus L R ∖ L . Its separator is L ∩ R L \cap R L ∩ R . For vertex sets S , T ⊆ V ( H ) S, T \subseteq V(H) S , T ⊆ V ( H ) , an S S S –T T T vertex separator is a set that intersects every path from S S S to T T T ; it may contain vertices of S ∪ T S \cup T S ∪ T . Menger’s theorem states that the minimum size of an S S S –T T T vertex separator equals the maximum number of pairwise vertex-disjoint S S S –T T T paths. A minimum S S S –T T T vertex separator can be computed in polynomial time by vertex splitting and a maximum-flow computation.
Let S , T ⊆ V ( H ) S, T \subseteq V(H) S , T ⊆ V ( H ) be disjoint. A connected S S S –T T T cut is a cut C C C of H H H such that S ⊆ C ⊆ V ( H ) ∖ T S \subseteq C \subseteq V(H) \setminus T S ⊆ C ⊆ V ( H ) ∖ T and H [ C ] H[C] H [ C ] is connected. It is important if there is no connected S S S –T T T cut D ⊋ C D \supsetneq C D ⊋ C with ∣ δ H ( D ) ∣ ≤ ∣ δ H ( C ) ∣ |\delta_H(D)| \le|\delta_H(C)| ∣ δ H ( D ) ∣ ≤ ∣ δ H ( C ) ∣ . Every connected S S S –T T T cut C C C is contained in an important connected S S S –T T T cut D D D with ∣ δ H ( D ) ∣ ≤ ∣ δ H ( C ) ∣ |\delta_H(D)| \le|\delta_H(C)| ∣ δ H ( D ) ∣ ≤ ∣ δ H ( C ) ∣ . The following lemma bounds the number of important cuts.
Lemma 2.1 ([9 Theorems 8.11 and 8.13] ). Let H H H be an unweighted undirected graph, let S , T ⊆ V ( H ) S, T \subseteq V(H) S , T ⊆ V ( H ) be disjoint, and suppose that S S S is nonempty and connected. For every ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 , there are at most 4 ℓ 4^\ell 4 ℓ important connected S S S –T T T cuts D D D with ∣ δ H ( D ) ∣ ≤ ℓ |\delta_H(D)| \le\ell ∣ δ H ( D ) ∣ ≤ ℓ . All such cuts can be enumerated in O ∗ ( 4 ℓ ) O^*(4^\ell) O ∗ ( 4 ℓ ) time.
A tree decomposition of H H H is a tree T \mathcal{T} T together with bags B t ⊆ V ( H ) B_t \subseteq V(H) B t ⊆ V ( H ) for t ∈ V ( T ) t \in V(\mathcal{T}) t ∈ V ( T ) such that ⋃ t ∈ V ( T ) B t = V ( H ) \bigcup_{t \in V(\mathcal{T})} B_t = V(H) ⋃ t ∈ V ( T ) B t = V ( H ) , every edge of H H H has both endpoints in some bag, and, for every v ∈ V ( H ) v \in V(H) v ∈ V ( H ) , the nodes t t t with v ∈ B t v \in B_t v ∈ B t induce a connected subtree of T \mathcal{T} T . For t t ′ ∈ E ( T ) tt' \in E(\mathcal{T}) t t ′ ∈ E ( T ) , the set B t ∩ B t ′ B_t \cap B_{t'} B t ∩ B t ′ is the adhesion of t t ′ tt' t t ′ ; the adhesion of the decomposition is the maximum size of an adhesion, or zero if T \mathcal{T} T has no edge. A rooted tree decomposition additionally has a designated root.
Perfect hash families. For 1 ≤ r ≤ n 1 \le r \le n 1 ≤ r ≤ n , an ( n , r ) (n,r) ( n , r ) -perfect hash family is a family H \mathcal{H} H of functions h : [ n ] → [ r ] h : [n] \to[r] h : [ n ] → [ r ] such that, for every X ⊆ [ n ] X \subseteq[n] X ⊆ [ n ] with ∣ X ∣ ≤ r |X| \le r ∣ X ∣ ≤ r , some h ∈ H h \in\mathcal{H} h ∈ H is injective on X X X . An ( n , r ) (n,r) ( n , r ) -perfect hash family of size O ∗ ( 2 O ( r ) ) O^*(2^{O(r)}) O ∗ ( 2 O ( r ) ) can be constructed in O ∗ ( 2 O ( r ) ) O^*(2^{O(r)}) O ∗ ( 2 O ( r ) ) time [37 ] .
Formal power series. Let X \mathcal{X} X be a finite set of variables. A formal power series over X \mathcal{X} X is a possibly infinite Q \mathbb{Q} Q -linear combination of monomials in X \mathcal{X} X . We write Q [ [ X ] ] \mathbb{Q}[[\mathcal{X}]] Q [[ X ]] for the ring of such series. For a monomial m m m and a variable z z z , let deg z ( m ) \deg_z(m) deg z ( m ) denote the exponent of z z z in m m m . For u ⊆ X u \subseteq\mathcal{X} u ⊆ X , every F ∈ Q [ [ X ] ] F \in\mathbb{Q}[[\mathcal{X}]] F ∈ Q [[ X ]] has a unique expansion F = ∑ m F m m F = \sum_m F_m m F = ∑ m F m m , where m m m ranges over the monomials in u u u and F m ∈ Q [ [ X ∖ u ] ] F_m \in\mathbb{Q}[[\mathcal{X} \setminus u]] F m ∈ Q [[ X ∖ u ]] ; define [ m ] u F : = F m [m]_uF := F_m [ m ] u F := F m .
For a series F F F whose constant term is one, define
log F : = ∑ j ≥ 1 ( − 1 ) j + 1 j ( F − 1 ) j . (2) \log F := \sum_{j \ge1} \frac{(-1)^{j+1}}{j}(F-1)^j. \tag*{(2)} log F := j ≥ 1 ∑ j ( − 1 ) j + 1 ( F − 1 ) j . (2)
For a series F F F whose constant term is zero, define
exp F : = ∑ j ≥ 0 1 j ! F j . (3) \exp F := \sum_{j \ge0} \frac{1}{j!}F^j. \tag*{(3)} exp F := j ≥ 0 ∑ j ! 1 F j . (3)
The formal logarithm and exponential are inverse to each other.
For a series F F F and a nonnegative degree bound D D D , let ⟨ F ⟩ deg ≤ D \langle F\rangle_{\deg\le D} ⟨ F ⟩ d e g ≤ D denote the series obtained by discarding the monomials of degree greater than D D D .
Lemma 2.2. If the formal logarithm of Z Z Z is defined, then
⟨ exp ( ⟨ log Z ⟩ deg ≤ D ) ⟩ deg ≤ D = ⟨ Z ⟩ deg ≤ D . \left\langle\exp\left(\left\langle\log Z\right\rangle_{\deg\le D}\right)\right\rangle_{\deg\le D} = \langle Z\rangle_{\deg\le D}. ⟨ exp ( ⟨ log Z ⟩ d e g ≤ D ) ⟩ d e g ≤ D = ⟨ Z ⟩ d e g ≤ D . Proof. Every discarded monomial has degree greater than D D D , and multiplying it by any monomial cannot decrease its degree. Hence it cannot affect any retained coefficient in the exponential. The claim follows from exp ( log Z ) = Z \exp(\log Z) = Z exp ( log Z ) = Z .
Zeta and Möbius transforms. For a finite set Y Y Y and a function f : 2 Y → Q f : 2^Y \to\mathbb{Q} f : 2 Y → Q , its subset zeta transform is the function g : 2 Y → Q g : 2^Y \to\mathbb{Q} g : 2 Y → Q defined by g ( T ) : = ∑ J ⊆ T f ( J ) g(T) := \sum_{J \subseteq T} f(J) g ( T ) := ∑ J ⊆ T f ( J ) . The inverse operation is the subset Möbius transform , which recovers f f f by f ( S ) = ∑ T ⊆ S ( − 1 ) ∣ S ∣ − ∣ T ∣ g ( T ) f(S) = \sum_{T \subseteq S} (-1)^{|S|-|T|}g(T) f ( S ) = ∑ T ⊆ S ( − 1 ) ∣ S ∣ − ∣ T ∣ g ( T ) .
Minimum Bisection
In this section, we prove Theorem 1.2. We use the guarantee of Theorem 1.1, leaving the decomposition construction to § 5 .
Polynomial representation
Fix a rooted tree decomposition T \mathcal{T} T with bags B t B_t B t supplied by Theorem 1.1, and denote its root by t 0 t_0 t 0 . For a nonroot node t t t with parent t ′ t' t ′ , let G t G_t G t be the graph obtained from the subgraph of G G G induced by the union of the bags in the subtree rooted at t t t by deleting E ( G [ B t ∩ B t ′ ] ) E(G[B_t \cap B_{t'}]) E ( G [ B t ∩ B t ′ ]) ; at the root, set G t 0 : = G G_{t_0}:=G G t 0 := G . We call ∂ G t : = B t ∩ B t ′ \partial G_t := B_t \cap B_{t'} ∂ G t := B t ∩ B t ′ the parent boundary of G t G_t G t , and set ∂ G t 0 : = ∅ \partial G_{t_0}:=\emptyset ∂ G t 0 := ∅ . For a nonroot node t t t , deleting E ( G [ B t ∩ B t ′ ] ) E(G[B_t \cap B_{t'}]) E ( G [ B t ∩ B t ′ ]) prevents an edge with both endpoints in ∂ G t \partial G_t ∂ G t from being charged at both t t t and t ′ t' t ′ .
For A t ⊆ V ( G t ) A_t \subseteq V(G_t) A t ⊆ V ( G t ) , define the minimum completion cost of A t A_t A t by
μ t ( A t ) : = min { ∣ δ G ( A ) ∣ : A ∩ V ( G t ) = A t } . \mu_t(A_t) := \min\{|\delta_G(A)| : A \cap V(G_t) = A_t\}. μ t ( A t ) := min { ∣ δ G ( A ) ∣ : A ∩ V ( G t ) = A t } .
In other words, μ t ( A t ) \mu_t(A_t) μ t ( A t ) is the smallest number of cut edges obtainable when A t A_t A t is fixed inside V ( G t ) V(G_t) V ( G t ) and the vertices outside V ( G t ) V(G_t) V ( G t ) may be chosen freely. Call A t A_t A t k k k -extendable if μ t ( A t ) ≤ k \mu_t(A_t) \le k μ t ( A t ) ≤ k .
Our strategy is dynamic programming on T \mathcal{T} T . A natural way is to keep, for every node t t t , states indexed by the following data:
a set P ⊆ ∂ G t P \subseteq\partial G_t P ⊆ ∂ G t , which specifies A t ∩ ∂ G t A_t \cap\partial G_t A t ∩ ∂ G t ;
an integer b b b , which specifies ∣ A t ∖ ∂ G t ∣ |A_t \setminus\partial G_t| ∣ A t ∖ ∂ G t ∣ ;
an integer c ∈ { 0 , … , k } c \in\{0,\ldots,k\} c ∈ { 0 , … , k } , which specifies μ t ( A t ) \mu_t(A_t) μ t ( A t ) .
The state ( P , b , c ) (P,b,c) ( P , b , c ) corresponds to the sets A t ⊆ V ( G t ) A_t \subseteq V(G_t) A t ⊆ V ( G t ) satisfying A t ∩ ∂ G t = P A_t \cap\partial G_t = P A t ∩ ∂ G t = P , ∣ A t ∖ ∂ G t ∣ = b |A_t \setminus\partial G_t| = b ∣ A t ∖ ∂ G t ∣ = b , and μ t ( A t ) = c \mu_t(A_t) = c μ t ( A t ) = c . The answer to the Minimum Bisection instance can be recovered from the state values at the root t 0 t_0 t 0 .
We next decide what each state should store. The most natural way is to store a Boolean value indicating whether any such set A t A_t A t exists. However, Boolean state values are insufficient even if the goal is only to decide whether a solution exists: our algorithm applies algebraic operations, such as Möbius inversion and the formal logarithm, to the state values, which is impossible with existence information alone. We therefore store the number of sets A t A_t A t represented by each state. As our algorithm applies algebraic operations to these counts, it is convenient to encode them in a generating polynomial as follows.
Let y = ( y v ) v ∈ V ( G t ) \boldsymbol{y}=(y_v)_{v\in V(G_t)} y = ( y v ) v ∈ V ( G t ) . Define
Z t ( y , w ) : = ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ k ( w μ t ( A t ) ⋅ ∏ v ∈ A t y v ) . Z_t(\boldsymbol{y},w):=\sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\leq k}}\left(w^{\mu_t(A_t)}\cdot\prod_{v\in A_t}y_v\right). Z t ( y , w ) := A t ⊆ V ( G t ) μ t ( A t ) ≤ k ∑ ( w μ t ( A t ) ⋅ v ∈ A t ∏ y v ) .
The polynomial Z t Z_t Z t lists all k k k -extendable subsets of V ( G t ) V(G_t) V ( G t ) . For the monomial indexed by A t A_t A t , the exponent of w w w records μ t ( A t ) \mu_t(A_t) μ t ( A t ) .
In our dynamic-programming states, the vertices in A t ∖ ∂ G t A_t \setminus\partial G_t A t ∖ ∂ G t are not distinguished; each state records only ∣ A t ∖ ∂ G t ∣ = b |A_t \setminus\partial G_t|=b ∣ A t ∖ ∂ G t ∣ = b . In contrast, the monomial indexed by A t A_t A t in Z t Z_t Z t contains ∏ v ∈ A t y v \prod_{v\in A_t}y_v ∏ v ∈ A t y v , which distinguishes every vertex of A t A_t A t . To use Z t Z_t Z t for dynamic programming, we must forget the identities of the vertices in A t ∖ ∂ G t A_t \setminus\partial G_t A t ∖ ∂ G t . Throughout the algorithm, we implement such a forgetting operation by variable substitution. For X ⊆ V ( G t ) X\subseteq V(G_t) X ⊆ V ( G t ) , define the substitution ω t X ( y , x ) : = ( ω t , v X ) v ∈ V ( G t ) \omega_t^X(\boldsymbol{y},x):=(\omega_{t,v}^X)_{v\in V(G_t)} ω t X ( y , x ) := ( ω t , v X ) v ∈ V ( G t ) , where, for every v ∈ V ( G t ) v\in V(G_t) v ∈ V ( G t ) ,
ω t , v X : = { y v , v ∈ X , x , v ∈ V ( G t ) ∖ X . \omega_{t,v}^X :=
\begin{cases}
y_v, & v\in X,\\
x, & v\in V(G_t)\setminus X.
\end{cases} ω t , v X := { y v , x , v ∈ X , v ∈ V ( G t ) ∖ X .
We write Z t ( ω t X ( y , x ) , w ) Z_t(\omega_t^{X}(\mathbf{y},x),w) Z t ( ω t X ( y , x ) , w ) for the polynomial obtained by applying ω t X ( y , x ) \omega_t^{X}(\mathbf{y},x) ω t X ( y , x ) to Z t ( y , w ) Z_t(\mathbf{y},w) Z t ( y , w ) . Precisely, we have
Z t ( ω t ∂ G t ( y , x ) , w ) = ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ k ( x ∣ A t ∖ ∂ G t ∣ ⋅ ( ∏ v ∈ A t ∩ ∂ G t y v ) ⋅ w μ t ( A t ) ) . Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w)=\sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\leq k}}\left(x^{|A_t\setminus\partial G_t|}\cdot\left(\prod_{v\in A_t\cap\partial G_t}y_v\right)\cdot w^{\mu_t(A_t)}\right). Z t ( ω t ∂ G t ( y , x ) , w ) = A t ⊆ V ( G t ) μ t ( A t ) ≤ k ∑ ( x ∣ A t ∖ ∂ G t ∣ ⋅ ( v ∈ A t ∩ ∂ G t ∏ y v ) ⋅ w μ t ( A t ) ) .
Grouping the terms of Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) by P = A t ∩ ∂ G t P=A_t\cap\partial G_t P = A t ∩ ∂ G t gives
Z t ( ω t ∂ G t ( y , x ) , w ) = ∑ P ⊆ ∂ G t ( ( ∏ v ∈ P y v ) ⋅ ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ k , A t ∩ ∂ G t = P ( x ∣ A t ∖ ∂ G t ∣ ⋅ w μ t ( A t ) ) ) . Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w)=\sum_{P\subseteq\partial G_t}\left(\left(\prod_{v\in P}y_v\right)\cdot\sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\leq k,\ A_t\cap\partial G_t=P}}\left(x^{|A_t\setminus\partial G_t|}\cdot w^{\mu_t(A_t)}\right)\right). Z t ( ω t ∂ G t ( y , x ) , w ) = P ⊆ ∂ G t ∑ ( v ∈ P ∏ y v ) ⋅ A t ⊆ V ( G t ) μ t ( A t ) ≤ k , A t ∩ ∂ G t = P ∑ ( x ∣ A t ∖ ∂ G t ∣ ⋅ w μ t ( A t ) ) .
For P ⊆ ∂ G t P\subseteq\partial G_t P ⊆ ∂ G t , 0 ≤ b ≤ ∣ V ( G t ) ∖ ∂ G t ∣ 0\leq b\leq|V(G_t)\setminus\partial G_t| 0 ≤ b ≤ ∣ V ( G t ) ∖ ∂ G t ∣ , and 0 ≤ c ≤ k 0\leq c\leq k 0 ≤ c ≤ k , taking the coefficient of x b w c ∏ v ∈ P y v x^bw^c\prod_{v\in P}y_v x b w c ∏ v ∈ P y v gives
[ x b w c ∏ v ∈ P y v ] Z t ( ω t ∂ G t ( y , x ) , w ) = ∣ { A t ⊆ V ( G t ) : A t ∩ ∂ G t = P , ∣ A t ∖ ∂ G t ∣ = b , μ t ( A t ) = c } ∣ , \left[x^bw^c\prod_{v\in P}y_v\right]Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w)
=\left|\left\{A_t\subseteq V(G_t):A_t\cap\partial G_t=P,\ |A_t\setminus\partial G_t|=b,\ \mu_t(A_t)=c\right\}\right|, [ x b w c v ∈ P ∏ y v ] Z t ( ω t ∂ G t ( y , x ) , w ) = ∣ { A t ⊆ V ( G t ) : A t ∩ ∂ G t = P , ∣ A t ∖ ∂ G t ∣ = b , μ t ( A t ) = c } ∣ ,
which is exactly the number of sets represented by the dynamic-programming state ( P , b , c ) (P,b,c) ( P , b , c ) . Our algorithm computes Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) for every node t t t . Since 0 ≤ b ≤ n 0\leq b\leq n 0 ≤ b ≤ n , 0 ≤ c ≤ k 0\leq c\leq k 0 ≤ c ≤ k , and ∣ ∂ G t ∣ ≤ 2 k |\partial G_t|\leq2k ∣ ∂ G t ∣ ≤ 2 k , the polynomial Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) contains only O ∗ ( 2 ∣ ∂ G t ∣ ) = O ∗ ( 2 O ( k ) ) O^*(2^{|\partial G_t|})=O^*(2^{O(k)}) O ∗ ( 2 ∣ ∂ G t ∣ ) = O ∗ ( 2 O ( k ) ) monomials.
To recover the answer to the Minimum Bisection instance, consider the root t 0 t_0 t 0 . Since ∂ G t 0 = ∅ \partial G_{t_0}=\varnothing ∂ G t 0 = ∅ , ω t 0 ∂ G t 0 ( y , x ) \omega_{t_0}^{\partial G_{t_0}}(\mathbf{y},x) ω t 0 ∂ G t 0 ( y , x ) replaces every vertex variable by x x x . Moreover, G t 0 = G G_{t_0}=G G t 0 = G and μ t 0 ( A ) = ∣ δ G ( A ) ∣ \mu_{t_0}(A)=|\delta_G(A)| μ t 0 ( A ) = ∣ δ G ( A ) ∣ , so
Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w ) = ∑ A ⊆ V ( G ) ∣ δ G ( A ) ∣ ≤ k x ∣ A ∣ w ∣ δ G ( A ) ∣ . (4) Z_{t_0}(\omega_{t_0}^{\partial G_{t_0}}(\mathbf{y},x),w)=\sum_{\substack{A\subseteq V(G)\\ |\delta_G(A)|\leq k}}x^{|A|}w^{|\delta_G(A)|}. \tag*{(4)} Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w ) = A ⊆ V ( G ) ∣ δ G ( A ) ∣ ≤ k ∑ x ∣ A ∣ w ∣ δ G ( A ) ∣ . (4)
The number of feasible solutions to the Minimum Bisection instance is
∑ c = 0 k [ x ⌊ n / 2 ⌋ w c ] Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w ) . \sum_{c=0}^{k}[x^{\lfloor n/2\rfloor}w^c]Z_{t_0}(\omega_{t_0}^{\partial G_{t_0}}(\mathbf{y},x),w). c = 0 ∑ k [ x ⌊ n /2 ⌋ w c ] Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w ) .
Conceptual transition
We now consider how to compute Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) from the polynomials Z s ( ω s ∂ G s ( y , x ) , w ) Z_s(\omega_s^{\partial G_s}(\mathbf{y},x),w) Z s ( ω s ∂ G s ( y , x ) , w ) for the children s s s of t t t . We first consider a conceptually simple strategy. We enumerate every R ⊆ B t R\subseteq B_t R ⊆ B t , compute the contribution to Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) of the sets A t A_t A t satisfying A t ∩ B t = R A_t\cap B_t=R A t ∩ B t = R , and sum the contributions over all R ⊆ B t R\subseteq B_t R ⊆ B t .
Here, we should first remark that the strategy is only conceptual; our algorithm does not compute Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) by enumerating R ⊆ B t R\subseteq B_t R ⊆ B t because the strategy requires enumerating all 2 ∣ B t ∣ 2^{|B_t|} 2 ∣ B t ∣ subsets R ⊆ B t R\subseteq B_t R ⊆ B t , which cannot be done within the desired O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. Nevertheless, the strategy is useful for explaining the O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) -time algorithm, so we describe the strategy first.
Formally, Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) is represented as the following sum:
Z t ( ω t ∂ G t ( y , x ) , w ) = ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ k ( x ∣ A t ∖ ∂ G t ∣ ⋅ ( ∏ v ∈ A t ∩ ∂ G t y v ) ⋅ w μ t ( A t ) ) . = ∑ R ⊆ B t ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ k , A t ∩ B t = R ( x ∣ R ∖ ∂ G t ∣ ⋅ x ∣ A t ∖ B t ∣ ⋅ ( ∏ v ∈ R ∩ ∂ G t y v ) ⋅ w μ t ( A t ) ) = ∑ R ⊆ B t ( x ∣ R ∖ ∂ G t ∣ ⋅ ( ∏ v ∈ R ∩ ∂ G t y v ) ⋅ ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ k , A t ∩ B t = R ( x ∣ A t ∖ B t ∣ ⋅ w μ t ( A t ) ) ) (5) \begin{aligned}
Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w)=\sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\leq k}}\left(x^{|A_t\setminus\partial G_t|}\cdot\left(\prod_{v\in A_t\cap\partial G_t}y_v\right)\cdot w^{\mu_t(A_t)}\right). \\
&= \sum_{R\subseteq B_t} \sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\leq k,\ A_t\cap B_t=R}} \left(x^{|R\setminus\partial G_t|}\cdot x^{|A_t\setminus B_t|}\cdot\left(\prod_{v\in R\cap\partial G_t}y_v\right)\cdot w^{\mu_t(A_t)}\right) \\
&= \sum_{R\subseteq B_t} \left(x^{|R\setminus\partial G_t|}\cdot\left(\prod_{v\in R\cap\partial G_t}y_v\right)\cdot\sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\leq k,\ A_t\cap B_t=R}}\left(x^{|A_t\setminus B_t|}\cdot w^{\mu_t(A_t)}\right)\right) \tag*{(5)}
\end{aligned} Z t ( ω t ∂ G t ( y , x ) , w ) = A t ⊆ V ( G t ) μ t ( A t ) ≤ k ∑ ( x ∣ A t ∖ ∂ G t ∣ ⋅ ( v ∈ A t ∩ ∂ G t ∏ y v ) ⋅ w μ t ( A t ) ) . = R ⊆ B t ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ k , A t ∩ B t = R ∑ ( x ∣ R ∖ ∂ G t ∣ ⋅ x ∣ A t ∖ B t ∣ ⋅ ( v ∈ R ∩ ∂ G t ∏ y v ) ⋅ w μ t ( A t ) ) = R ⊆ B t ∑ x ∣ R ∖ ∂ G t ∣ ⋅ ( v ∈ R ∩ ∂ G t ∏ y v ) ⋅ A t ⊆ V ( G t ) μ t ( A t ) ≤ k , A t ∩ B t = R ∑ ( x ∣ A t ∖ B t ∣ ⋅ w μ t ( A t ) ) (5)
Now, the next task is to represent Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) using the child polynomials Z s ( ω s ∂ G s ( y , x ) , w ) Z_s(\omega_s^{\partial G_s}(\boldsymbol{y},x),w) Z s ( ω s ∂ G s ( y , x ) , w ) . Let child ( t ) \operatorname{child}(t) child ( t ) denote the set of children of t t t in T \mathcal{T} T . Fix R ⊆ B t R\subseteq B_t R ⊆ B t . For a child s s s of t t t , call a set A s ⊆ V ( G s ) A_s\subseteq V(G_s) A s ⊆ V ( G s ) R R R -consistent if A s ∩ ∂ G s = R ∩ ∂ G s A_s\cap\partial G_s=R\cap\partial G_s A s ∩ ∂ G s = R ∩ ∂ G s . By the connectedness axiom of the tree decomposition T \mathcal{T} T , every set A t ⊆ V ( G t ) A_t\subseteq V(G_t) A t ⊆ V ( G t ) with A t ∩ B t = R A_t\cap B_t=R A t ∩ B t = R is obtained uniquely by gluing R R R with one R R R -consistent set A s A_s A s for each child s s s of t t t in the form
A_t=R\mathbin{\mathaccent\cdot\cup}\mathop{\mathaccent\cdot\bigcup}_{s\in\operatorname{child}(t)}(A_s\setminus\partial G_s).
Consequently, Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) can be written entirely in terms of the sets A s A_s A s as
Q_{t,R}(x,w)=\sum\left(\left(\prod_{s\in\operatorname{child}(t)}x^{|A_s\setminus\partial G_s|}\right)\cdot w^{\mu_t\left(R\mathbin{\mathaccent\cdot\cup}_{s\in\operatorname{child}(t)}(A_s\setminus\partial G_s)\right)}\right) \tag*{(6)}
The sum is taken over all choices of an R R R -consistent set A s ⊆ V ( G s ) A_s\subseteq V(G_s) A s ⊆ V ( G s ) for every child s s s of t t t that satisfy
\mu_t\left(R\mathbin{\mathaccent\cdot\cup}\mathop{\mathaccent\cdot\bigcup}_{s\in\operatorname{child}(t)}(A_s\setminus\partial G_s)\right)\leq k.
We would like to rewrite the sum in Equation (6) as a product, over the children s s s of t t t , of sums indexed by R R R -consistent sets A s A_s A s . The x x x -weight already factors over the children, but the exponent of w w w is not additive over them. To do so, we correct the w w w -degree. For P ⊆ ∂ G t P\subseteq\partial G_t P ⊆ ∂ G t , define
η t ( P ) : = min { ∣ δ G − ( V ( G t ) ∖ ∂ G t ) ( A ) ∣ : A ⊆ V ( G ) ∖ ( V ( G t ) ∖ ∂ G t ) , A ∩ ∂ G t = P } . \eta_t(P):=\min\{|\delta_{G-(V(G_t)\setminus\partial G_t)}(A)|:A\subseteq V(G)\setminus(V(G_t)\setminus\partial G_t),\ A\cap\partial G_t=P\}. η t ( P ) := min { ∣ δ G − ( V ( G t ) ∖ ∂ G t ) ( A ) ∣ : A ⊆ V ( G ) ∖ ( V ( G t ) ∖ ∂ G t ) , A ∩ ∂ G t = P } .
The quantity η t ( P ) \eta_t(P) η t ( P ) is the minimum cut contribution outside G t G_t G t compatible with the boundary choice P P P . Thus, if A t ∩ ∂ G t = P A_t\cap\partial G_t=P A t ∩ ∂ G t = P , then
μ t ( A t ) = ∣ δ G t ( A t ) ∣ + η t ( P ) . (7) \mu_t(A_t)=|\delta_{G_t}(A_t)|+\eta_t(P). \tag*{(7)} μ t ( A t ) = ∣ δ G t ( A t ) ∣ + η t ( P ) . (7)
The next lemma corrects the w w w -degrees and gives an exact recurrence for Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) .
Lemma 3.1. For every R ⊆ B t R\subseteq B_t R ⊆ B t , the polynomial Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) satisfies
Q t , R ( x , w ) = ⟨ w ∣ δ G t [ B t ] ( R ) ∣ + η t ( R ∩ ∂ G t ) ⋅ ∏ s ∈ child ( t ) ( w − η s ( R ∩ ∂ G s ) ⋅ [ ∏ v ∈ R ∩ ∂ G s y v ] ⋅ Z s ( ω s ∂ G s ( y , x ) , w ) ) ⟩ deg w ≤ k . (8) Q_{t,R}(x,w)=\left\langle w^{|\delta_{G_t[B_t]}(R)|+\eta_t(R\cap\partial G_t)}
\cdot\prod_{s\in\operatorname{child}(t)}
\left(
w^{-\eta_s(R\cap\partial G_s)}
\cdot
\left[
\prod_{v\in R\cap\partial G_s}y_v
\right]
\cdot
Z_s(\omega_s^{\partial G_s}(\boldsymbol{y},x),w)
\right)
\right\rangle_{\deg_w\leq k}. \tag*{(8)} Q t , R ( x , w ) = ⟨ w ∣ δ G t [ B t ] ( R ) ∣ + η t ( R ∩ ∂ G t ) ⋅ s ∈ child ( t ) ∏ ( w − η s ( R ∩ ∂ G s ) ⋅ [ v ∈ R ∩ ∂ G s ∏ y v ] ⋅ Z s ( ω s ∂ G s ( y , x ) , w ) ) ⟩ d e g w ≤ k . (8)
Proof. Fix R ⊆ B t R \subseteq B_t R ⊆ B t . For every child s s s of t t t , the coefficient of the boundary monomial indexed by R ∩ ∂ G s R \cap\partial G_s R ∩ ∂ G s is
[ ∏ v ∈ R ∩ ∂ G s y v ] y Z s ∂ G s ( y , x , w ) = ∑ A s ⊆ V ( G s ) μ s ( A s ) ≤ k , A s ∩ ∂ G s = R ∩ ∂ G s ( x ∣ A s ∖ ∂ G s ∣ w μ s ( A s ) ) . \left[\prod_{v\in R\cap\partial G_s} y_v\right]_{\mathbf y} Z_s^{\partial G_s}(\mathbf y,x,w)=\sum_{\substack{A_s\subseteq V(G_s)\\ \mu_s(A_s)\leq k,\ A_s\cap\partial G_s=R\cap\partial G_s}}\left(x^{|A_s\setminus\partial G_s|}w^{\mu_s(A_s)}\right). [ v ∈ R ∩ ∂ G s ∏ y v ] y Z s ∂ G s ( y , x , w ) = A s ⊆ V ( G s ) μ s ( A s ) ≤ k , A s ∩ ∂ G s = R ∩ ∂ G s ∑ ( x ∣ A s ∖ ∂ G s ∣ w μ s ( A s ) ) .
Thus, expanding the product in Equation (8) amounts to choosing one R R R -consistent set A s A_s A s for every child s s s of t t t . For these choices, let
A_t:=R\mathbin{\mathaccent\cdot\cup}\bigcup_{s\in\operatorname{child}(t)}(A_s\setminus\partial G_s).
By Equation (6) , it suffices to verify that the right-hand side of Equation (8) assigns A t A_t A t the monomial x ∣ A t ∖ B t ∣ w μ t ( A t ) x^{|A_t\setminus B_t|}w^{\mu_t(A_t)} x ∣ A t ∖ B t ∣ w μ t ( A t ) if μ t ( A t ) ≤ k \mu_t(A_t)\leq k μ t ( A t ) ≤ k , and zero otherwise. For the x x x -weight,
∏ s ∈ child ( t ) x ∣ A s ∖ ∂ G s ∣ = x ∣ A t ∖ B t ∣ . \prod_{s\in\operatorname{child}(t)}x^{|A_s\setminus\partial G_s|}=x^{|A_t\setminus B_t|}. s ∈ child ( t ) ∏ x ∣ A s ∖ ∂ G s ∣ = x ∣ A t ∖ B t ∣ .
For the w w w -weight, the edge sets of G t [ B t ] G_t[B_t] G t [ B t ] and of the graphs G s G_s G s for the children s s s of t t t partition E ( G t ) E(G_t) E ( G t ) . Hence Equation (7) , applied at t t t and at every child s s s , gives
\begin{aligned}
\mu_t\left(R\mathbin{\mathaccent\cdot\cup}\bigcup_{s\in\operatorname{child}(t)}(A_s\setminus\partial G_s)\right)
&=|\delta_{G_t[B_t]}(R)|+\sum_{s\in\operatorname{child}(t)}|\delta_{G_s}(A_s)|+\eta_t(R\cap\partial G_t)\\
&=|\delta_{G_t[B_t]}(R)|+\eta_t(R\cap\partial G_t)+\sum_{s\in\operatorname{child}(t)}\left(\mu_s(A_s)-\eta_s(R\cap\partial G_s)\right).
\end{aligned}
Consequently, the selected terms contribute the required monomial x ∣ A t ∖ B t ∣ w μ t ( A t ) x^{|A_t\setminus B_t|}w^{\mu_t(A_t)} x ∣ A t ∖ B t ∣ w μ t ( A t ) . The truncation in Equation (8) keeps exactly the choices with μ t ( A t ) ≤ k \mu_t(A_t)\leq k μ t ( A t ) ≤ k . Conversely, every set A t A_t A t with μ t ( A t ) ≤ k \mu_t(A_t)\leq k μ t ( A t ) ≤ k is obtained from the sets A s = A t ∩ V ( G s ) A_s=A_t\cap V(G_s) A s = A t ∩ V ( G s ) , which satisfy μ s ( A s ) ≤ k \mu_s(A_s)\leq k μ s ( A s ) ≤ k for every child s s s because a minimum completion of A t A_t A t is also a completion of every A s A_s A s . Thus, Equation (6) shows that the retained sum is Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) . □
If ∣ B t ∣ ≤ 4 k + 1 |B_t|\leq4k+1 ∣ B t ∣ ≤ 4 k + 1 , we compute Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) from Equation (8) for every R ⊆ B t R\subseteq B_t R ⊆ B t and then compute Z t ∂ G t ( y , x , w ) Z_t^{\partial G_t}(\mathbf y,x,w) Z t ∂ G t ( y , x , w ) from Equation (5) . The computation takes O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
We now assume that ∣ B t ∣ > 4 k + 1 |B_t|>4k+1 ∣ B t ∣ > 4 k + 1 . Enumerating all 2 ∣ B t ∣ 2^{|B_t|} 2 ∣ B t ∣ subsets R ⊆ B t R\subseteq B_t R ⊆ B t is too expensive, so we need a different algorithm. By the edge-unbreakability guarantee for the bags of our tree decomposition, every k k k -extendable set A t A_t A t satisfies ∣ A t ∩ B t ∣ ≤ 2 k |A_t\cap B_t|\leq2k ∣ A t ∩ B t ∣ ≤ 2 k or ∣ B t ∖ A t ∣ ≤ 2 k |B_t\setminus A_t|\leq2k ∣ B t ∖ A t ∣ ≤ 2 k . Since ∣ B t ∣ > 4 k + 1 |B_t|>4k+1 ∣ B t ∣ > 4 k + 1 , the two inequalities are mutually exclusive. We compute the two contributions separately and add them. Define
Z t ≤ 2 k ( y , w ) : = ∑ A t ⊆ V ( G t ) , μ t ( A t ) ≤ k ∣ A t ∩ B t ∣ ≤ 2 k ( w μ t ( A t ) ⋅ ∏ v ∈ A t y v ) . Z_t^{\leq2k}(\mathbf y,w):=\sum_{\substack{A_t\subseteq V(G_t),\ \mu_t(A_t)\leq k\\ |A_t\cap B_t|\leq2k}}\left(w^{\mu_t(A_t)}\cdot\prod_{v\in A_t}y_v\right). Z t ≤ 2 k ( y , w ) := A t ⊆ V ( G t ) , μ t ( A t ) ≤ k ∣ A t ∩ B t ∣ ≤ 2 k ∑ ( w μ t ( A t ) ⋅ v ∈ A t ∏ y v ) .
Z t ≤ 2 k ( y , w ) Z_t^{\leq2k}(\mathbf y,w) Z t ≤ 2 k ( y , w ) is the restriction of Z t ( y , w ) Z_t(\mathbf y,w) Z t ( y , w ) to the sets A t A_t A t satisfying ∣ A t ∩ B t ∣ ≤ 2 k |A_t\cap B_t|\leq2k ∣ A t ∩ B t ∣ ≤ 2 k . The next lemma represents the restriction of Z t ( y , w ) Z_t(\mathbf y,w) Z t ( y , w ) to the sets A t A_t A t satisfying ∣ B t ∖ A t ∣ ≤ 2 k |B_t\setminus A_t|\leq2k ∣ B t ∖ A t ∣ ≤ 2 k in terms of Z t ≤ 2 k ( y , w ) Z_t^{\leq2k}(\mathbf y,w) Z t ≤ 2 k ( y , w ) , and hence expresses Z t ( y , w ) Z_t(\mathbf y,w) Z t ( y , w ) as the sum of the two restrictions.
Lemma 3.2. Let y − 1 : = ( y v − 1 ) v ∈ V ( G t ) \mathbf{y}^{-1}:=(y_v^{-1})_{v\in V(G_t)} y − 1 := ( y v − 1 ) v ∈ V ( G t ) . Then
Z t ( y , w ) = Z t ≤ 2 k ( y , w ) + ( ∏ v ∈ V ( G t ) y v ) Z t ≤ 2 k ( y − 1 , w ) . Z_t(\mathbf{y},w)=Z_t^{\le2k}(\mathbf{y},w)+\left(\prod_{v\in V(G_t)}y_v\right)Z_t^{\le2k}(\mathbf{y}^{-1},w). Z t ( y , w ) = Z t ≤ 2 k ( y , w ) + v ∈ V ( G t ) ∏ y v Z t ≤ 2 k ( y − 1 , w ) . Proof. Edge-unbreakability and ∣ B t ∣ > 4 k + 1 |B_t|>4k+1 ∣ B t ∣ > 4 k + 1 imply that every set A t A_t A t with μ t ( A t ) ≤ k \mu_t(A_t)\le k μ t ( A t ) ≤ k satisfies exactly one of ∣ A t ∩ B t ∣ ≤ 2 k |A_t\cap B_t|\le2k ∣ A t ∩ B t ∣ ≤ 2 k and ∣ B t ∖ A t ∣ ≤ 2 k |B_t\setminus A_t|\le2k ∣ B t ∖ A t ∣ ≤ 2 k . The polynomial Z t ≤ 2 k ( y , w ) Z_t^{\le2k}(\mathbf{y},w) Z t ≤ 2 k ( y , w ) contains exactly the terms indexed by the sets satisfying ∣ A t ∩ B t ∣ ≤ 2 k |A_t\cap B_t|\le2k ∣ A t ∩ B t ∣ ≤ 2 k .
Fix a set A t ⊆ V ( G t ) A_t\subseteq V(G_t) A t ⊆ V ( G t ) satisfying μ t ( A t ) ≤ k \mu_t(A_t)\le k μ t ( A t ) ≤ k and ∣ B t ∖ A t ∣ ≤ 2 k |B_t\setminus A_t|\le2k ∣ B t ∖ A t ∣ ≤ 2 k . Since ( V ( G t ) ∖ A t ) ∩ B t = B t ∖ A t (V(G_t)\setminus A_t)\cap B_t=B_t\setminus A_t ( V ( G t ) ∖ A t ) ∩ B t = B t ∖ A t and μ t ( V ( G t ) ∖ A t ) = μ t ( A t ) \mu_t(V(G_t)\setminus A_t)=\mu_t(A_t) μ t ( V ( G t ) ∖ A t ) = μ t ( A t ) , we have
( ∏ v ∈ V ( G t ) y v ) ( w μ t ( A t ) ∏ v ∈ V ( G t ) ∖ A t y v − 1 ) = w μ t ( A t ) ∏ v ∈ A t y v . \left(\prod_{v\in V(G_t)}y_v\right)\left(w^{\mu_t(A_t)}\prod_{v\in V(G_t)\setminus A_t}y_v^{-1}\right)=w^{\mu_t(A_t)}\prod_{v\in A_t}y_v. v ∈ V ( G t ) ∏ y v w μ t ( A t ) v ∈ V ( G t ) ∖ A t ∏ y v − 1 = w μ t ( A t ) v ∈ A t ∏ y v . Summing this identity over all such A t A_t A t gives
( ∏ v ∈ V ( G t ) y v ) Z t ≤ 2 k ( y − 1 , w ) = ∑ A t ⊆ V ( G t ) , μ t ( A t ) ≤ k ∣ B t ∖ A t ∣ ≤ 2 k w μ t ( A t ) ∏ v ∈ A t y v . \left(\prod_{v\in V(G_t)}y_v\right)Z_t^{\le2k}(\mathbf{y}^{-1},w)=\sum_{\substack{A_t\subseteq V(G_t),\ \mu_t(A_t)\le k\\ |B_t\setminus A_t|\le2k}}w^{\mu_t(A_t)}\prod_{v\in A_t}y_v. v ∈ V ( G t ) ∏ y v Z t ≤ 2 k ( y − 1 , w ) = A t ⊆ V ( G t ) , μ t ( A t ) ≤ k ∣ B t ∖ A t ∣ ≤ 2 k ∑ w μ t ( A t ) v ∈ A t ∏ y v . Together with the terms satisfying ∣ A t ∩ B t ∣ ≤ 2 k |A_t\cap B_t|\le2k ∣ A t ∩ B t ∣ ≤ 2 k , this proves the lemma.
Applying Lemma 3.2 after the substitution ω t ∂ G t ( y , x ) \omega_t^{\partial G_t}(\mathbf{y},x) ω t ∂ G t ( y , x ) gives
Z t ( ω t ∂ G t ( y , x ) , w ) = Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) + x ∣ V ( G t ) ∖ ∂ G t ∣ ( ∏ v ∈ ∂ G t y v ) Z t ≤ 2 k ( ω t ∂ G t ( y − 1 , x − 1 ) , w ) . \begin{aligned}
Z_t(\omega_t^{\partial G_t}(\mathbf{y},x),w)
&=Z_t^{\le2k}(\omega_t^{\partial G_t}(\mathbf{y},x),w)\\
&\quad+x^{|V(G_t)\setminus\partial G_t|}\left(\prod_{v\in\partial G_t}y_v\right)Z_t^{\le2k}(\omega_t^{\partial G_t}(\mathbf{y}^{-1},x^{-1}),w).
\end{aligned} Z t ( ω t ∂ G t ( y , x ) , w ) = Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) + x ∣ V ( G t ) ∖ ∂ G t ∣ ( v ∈ ∂ G t ∏ y v ) Z t ≤ 2 k ( ω t ∂ G t ( y − 1 , x − 1 ) , w ) .
The right-hand side can be computed directly from the coefficients of Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) . Thus, we now focus on computing Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) .
Large-bag transition via formal logarithms
By the previous subsection, it remains to compute Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) . Consider the substitution ω t B t ( y , x ) \omega_t^{B_t}(\mathbf{y},x) ω t B t ( y , x ) . By definition,
ω t , v B t ( y , x ) = { y v , v ∈ B t , x , v ∈ V ( G t ) ∖ B t . ( v ∈ V ( G t ) ) . \omega_{t,v}^{B_t}(\mathbf{y},x)=
\begin{cases}
y_v, & v\in B_t,\\
x, & v\in V(G_t)\setminus B_t.
\end{cases}
\qquad(v\in V(G_t)). ω t , v B t ( y , x ) = { y v , x , v ∈ B t , v ∈ V ( G t ) ∖ B t . ( v ∈ V ( G t )) .
It gives
Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = ∑ A t ⊆ V ( G t ) , μ t ( A t ) ≤ k ∣ A t ∩ B t ∣ ≤ 2 k x ∣ A t ∖ B t ∣ ( ∏ v ∈ A t ∩ B t y v ) w μ t ( A t ) . Z_t^{\le2k}(\omega_t^{B_t}(\mathbf{y},x),w)
=
\sum_{\substack{A_t\subseteq V(G_t),\ \mu_t(A_t)\le k\\ |A_t\cap B_t|\le2k}}
x^{|A_t\setminus B_t|}
\left(\prod_{v\in A_t\cap B_t}y_v\right)w^{\mu_t(A_t)}. Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = A t ⊆ V ( G t ) , μ t ( A t ) ≤ k ∣ A t ∩ B t ∣ ≤ 2 k ∑ x ∣ A t ∖ B t ∣ ( v ∈ A t ∩ B t ∏ y v ) w μ t ( A t ) .
If we can compute Z t ≤ 2 k ( ω t B t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{B_t}(\mathbf{y},x),w) Z t ≤ 2 k ( ω t B t ( y , x ) , w ) , then substituting y v = x y_v=x y v = x for every v ∈ B t ∖ ∂ G t v\in B_t\setminus\partial G_t v ∈ B t ∖ ∂ G t yields Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(\mathbf{y},x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) . However, computing all coefficients of this polynomial appears infeasible: in the worst case, for every R ⊆ B t R\subseteq B_t R ⊆ B t with ∣ R ∣ ≤ 2 k |R|\le2k ∣ R ∣ ≤ 2 k , it contains a monomial x b w ℓ ∏ v ∈ R y v x^b w^\ell\prod_{v\in R}y_v x b w ℓ ∏ v ∈ R y v for some b ≥ 0 b \ge0 b ≥ 0 and 0 ≤ ℓ ≤ k 0 \le\ell\le k 0 ≤ ℓ ≤ k . It can therefore have at least ∑ i = 0 2 k ( ∣ B t ∣ i ) = n Θ ( k ) \sum_{i=0}^{2k}\binom{|B_t|}{i}=n^{\Theta(k)} ∑ i = 0 2 k ( i ∣ B t ∣ ) = n Θ ( k ) distinct monomials, so even writing all its coefficients is too expensive.
The key idea is to consider instead its formal logarithm log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) \log Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w) log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) , viewed as a formal power series. The polynomial has constant term 1, and its formal logarithm is well-defined; indeed, the set A t = ∅ A_t=\emptyset A t = ∅ contributes the term 1, whereas every other set contributes a monomial of positive degree.
For a monomial m = x b w ℓ ∏ v ∈ V ( G t ) y v a v m=x^b w^\ell\prod_{v\in V(G_t)}y_v^{a_v} m = x b w ℓ ∏ v ∈ V ( G t ) y v a v , call { v ∈ B t : a v > 0 } \{v\in B_t:a_v>0\} { v ∈ B t : a v > 0 } the B t B_t B t -support of m m m . The next lemma shows that, after restricting to monomials of w w w -degree at most k k k and B t B_t B t -support size at most 2 k 2k 2 k , only O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) distinct supports occur. It also shows that these supports can be enumerated in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Lemma 3.3. In O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time, one can enumerate a downward-closed family C t ⊆ 2 B t C_t\subseteq2^{B_t} C t ⊆ 2 B t of size O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) such that every member has size at most 2 k 2k 2 k and
[ x b w ℓ ∏ v ∈ B t y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ≠ 0 ⟹ { v ∈ B t : a v > 0 } ∈ C t \left[x^b w^\ell\prod_{v\in B_t}y_v^{a_v}\right]\log Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)\ne0\quad\Longrightarrow\quad\{v\in B_t:a_v>0\}\in C_t [ x b w ℓ v ∈ B t ∏ y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = 0 ⟹ { v ∈ B t : a v > 0 } ∈ C t whenever ℓ ≤ k \ell\le k ℓ ≤ k and ∣ { v ∈ B t : a v > 0 } ∣ ≤ 2 k |\{v\in B_t:a_v>0\}|\le2k ∣ { v ∈ B t : a v > 0 } ∣ ≤ 2 k .
We prove Lemma 3.3 in § 3.4 . For each S ⊆ B t S\subseteq B_t S ⊆ B t , let L t , S ( y , w ) L_{t,S}(\boldsymbol{y},w) L t , S ( y , w ) be the formal power series consisting of the terms of log Z t ≤ 2 k ( y , w ) \log Z_t^{\le2k}(\boldsymbol{y},w) log Z t ≤ 2 k ( y , w ) whose monomials have B t B_t B t -support S S S . Then
log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = ∑ S ⊆ B t L t , S ( ω t B t ( y , x ) , w ) . \log Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)=\sum_{S\subseteq B_t}L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w). log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = S ⊆ B t ∑ L t , S ( ω t B t ( y , x ) , w ) .
We would like to compute the left-hand side from the series on the right-hand side. Since these series are generally infinite, we truncate both sides to obtain:
⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) = ∑ S ⊆ B t ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) . \left\langle\log Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)\right\rangle_{\le(k,n,2k)}
=
\sum_{S\subseteq B_t}\left\langle L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)\right\rangle_{\le(k,n,2k)}. ⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) = S ⊆ B t ∑ ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) .
where, for a formal power series F F F ,
⟨ F ⟩ ≤ ( k , n , 2 k ) : = ⟨ F ⟩ deg w ≤ k , deg x ≤ n , ∑ v ∈ B t deg y v ≤ 2 k . \langle F\rangle_{\le(k,n,2k)}:=\langle F\rangle_{\deg_w\le k,\ \deg_x\le n,\ \sum_{v\in B_t}\deg_{y_v}\le2k}. ⟨ F ⟩ ≤ ( k , n , 2 k ) := ⟨ F ⟩ d e g w ≤ k , d e g x ≤ n , ∑ v ∈ B t d e g y v ≤ 2 k .
Applying Lemma 3.3 to the right-hand side of the preceding identity, we see that every summand indexed by S ∉ C t S\notin C_t S ∈ / C t is zero. Consequently, we obtain
⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) = ∑ S ∈ C t ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) . (9) \left\langle\log Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)\right\rangle_{\le(k,n,2k)}
=
\sum_{S\in C_t}\left\langle L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)\right\rangle_{\le(k,n,2k)}. \tag*{(9)} ⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) = S ∈ C t ∑ ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) . (9)
Fix S ∈ C t S\in C_t S ∈ C t . The next lemma computes ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) \left\langle L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)\right\rangle_{\le(k,n,2k)} ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) from the polynomials Q t , R Q_{t,R} Q t , R with R ⊆ S R\subseteq S R ⊆ S .
Lemma 3.4. All coefficients of ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) \left\langle L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},\boldsymbol{x}),w)\right\rangle_{\le(k,n,2k)} ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) can be computed in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Proof. For T ⊆ S T\subseteq S T ⊆ S , define the substitution ρ t T ( y , x ) : = ( ρ t , v T ) v ∈ V ( G t ) \rho_t^T(\boldsymbol{y},\boldsymbol{x}):=(\rho_{t,v}^T)_{v\in V(G_t)} ρ t T ( y , x ) := ( ρ t , v T ) v ∈ V ( G t ) , where
ρ t , v T = { y v , v ∈ T , 0 , v ∈ B t ∖ T , x , v ∈ V ( G t ) ∖ B t . ( v ∈ V ( G t ) ) . \rho_{t,v}^T=
\begin{cases}
y_v, & v\in T,\\
0, & v\in B_t\setminus T,\\
x, & v\in V(G_t)\setminus B_t.
\end{cases}
\qquad(v\in V(G_t)). ρ t , v T = ⎩ ⎨ ⎧ y v , 0 , x , v ∈ T , v ∈ B t ∖ T , v ∈ V ( G t ) ∖ B t . ( v ∈ V ( G t )) .
The substitution ρ t T ( y , x ) \rho_t^T(\boldsymbol{y},x) ρ t T ( y , x ) is obtained from ω t B t ( y , x ) \omega_t^{B_t}(\boldsymbol{y},x) ω t B t ( y , x ) by additionally setting y v = 0 y_v=0 y v = 0 for every v ∈ B t ∖ T v\in B_t\setminus T v ∈ B t ∖ T . Since L t , J L_{t,J} L t , J consists of the terms of log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) \log Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},x),w) log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) with B t B_t B t -support exactly J J J , this substitution leaves L t , J ( ω t B t ( y , x ) , w ) L_{t,J}(\omega_t^{B_t}(\boldsymbol{y},x),w) L t , J ( ω t B t ( y , x ) , w ) unchanged when J ⊆ T J\subseteq T J ⊆ T and annihilates L t , J ( ω t B t ( y , x ) , w ) L_{t,J}(\omega_t^{B_t}(\boldsymbol{y},x),w) L t , J ( ω t B t ( y , x ) , w ) when J ⊈ T J\not\subseteq T J ⊆ T . Thus,
log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) = ∑ J ⊆ T L t , J ( ω t B t ( y , x ) , w ) . \log Z_t^{\le2k}(\rho_t^T(\boldsymbol{y},x),w)=\sum_{J\subseteq T}L_{t,J}(\omega_t^{B_t}(\boldsymbol{y},x),w). log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) = J ⊆ T ∑ L t , J ( ω t B t ( y , x ) , w ) .
This is the subset zeta transform of the family L t , J ( ω t B t ( y , x ) , w ) L_{t,J}(\omega_t^{B_t}(\boldsymbol{y},x),w) L t , J ( ω t B t ( y , x ) , w ) . Hence Möbius inversion gives
L t , S ( ω t B t ( y , x ) , w ) = ∑ T ⊆ S ( − 1 ) ∣ S ∣ − ∣ T ∣ log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) . (10) L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},x),w)=\sum_{T\subseteq S}(-1)^{|S|-|T|}\log Z_t^{\le2k}(\rho_t^T(\boldsymbol{y},x),w). \tag*{(10)} L t , S ( ω t B t ( y , x ) , w ) = T ⊆ S ∑ ( − 1 ) ∣ S ∣ − ∣ T ∣ log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) . (10)
Truncating Equation (10) gives
⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) = ∑ T ⊆ S ( − 1 ) ∣ S ∣ − ∣ T ∣ ⟨ log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) . (11) \left\langle L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},x),w)\right\rangle_{\le(k,n,2k)}
=\sum_{T\subseteq S}(-1)^{|S|-|T|}\left\langle\log Z_t^{\le2k}(\rho_t^T(\boldsymbol{y},x),w)\right\rangle_{\le(k,n,2k)}. \tag*{(11)} ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) = T ⊆ S ∑ ( − 1 ) ∣ S ∣ − ∣ T ∣ ⟨ log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) . (11)
For each T ⊆ S T\subseteq S T ⊆ S , we compute the summand in Equation (11) as follows. As in § 3.2 , we group the terms by R = A t ∩ B t R=A_t\cap B_t R = A t ∩ B t :
Z t ≤ 2 k ( ρ t T ( y , x ) , w ) = ∑ R ⊆ B t ∣ R ∣ ≤ 2 k ( ∏ v ∈ R ρ t , v T ) ⋅ Q t , R ( x , w ) = ∑ R ⊆ T ( ∏ v ∈ R y v ) ⋅ Q t , R ( x , w ) , (12) Z_t^{\le2k}(\rho_t^T(\boldsymbol{y},x),w)=\sum_{\substack{R\subseteq B_t\\ |R|\le2k}}\left(\prod_{v\in R}\rho_{t,v}^T\right)\cdot Q_{t,R}(x,w)=\sum_{R\subseteq T}\left(\prod_{v\in R}y_v\right)\cdot Q_{t,R}(x,w), \tag*{(12)} Z t ≤ 2 k ( ρ t T ( y , x ) , w ) = R ⊆ B t ∣ R ∣ ≤ 2 k ∑ ( v ∈ R ∏ ρ t , v T ) ⋅ Q t , R ( x , w ) = R ⊆ T ∑ ( v ∈ R ∏ y v ) ⋅ Q t , R ( x , w ) , (12)
where the second equality follows because ρ t , v T = 0 \rho_{t,v}^T=0 ρ t , v T = 0 for v ∈ B t ∖ T v\in B_t\setminus T v ∈ B t ∖ T . By Lemma 3.1, Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) can be computed from the polynomials Z s ( ω s ∂ G s ( y , x ) , w ) Z_s(\omega_s^{\partial G_s}(\boldsymbol{y},x),w) Z s ( ω s ∂ G s ( y , x ) , w ) for the children s s s of t t t , for every R ⊆ T R\subseteq T R ⊆ T . Hence all coefficients of Z t ≤ 2 k ( ρ t T ( y , x ) , w ) Z_t^{\le2k}(\rho_t^T(\boldsymbol{y},x),w) Z t ≤ 2 k ( ρ t T ( y , x ) , w ) can be computed. For every T ⊆ S T\subseteq S T ⊆ S , we take its formal logarithm and truncate it to degrees ≤ ( k , n , 2 k ) \le(k,n,2k) ≤ ( k , n , 2 k ) . The sum in Equation (11) then computes
⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) . \left\langle L_{t,S}(\omega_t^{B_t}(\boldsymbol{y},x),w)\right\rangle_{\le(k,n,2k)}. ⟨ L t , S ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) .
We analyze the running time. For each T ⊆ S T\subseteq S T ⊆ S , the sum over R ⊆ T R\subseteq T R ⊆ T in Equation (12) has at most 2 2 k 2^{2k} 2 2 k summands because ∣ T ∣ ≤ ∣ S ∣ ≤ 2 k |T|\le|S|\le2k ∣ T ∣ ≤ ∣ S ∣ ≤ 2 k . By Lemma 3.1, each Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) can be computed in polynomial time. Hence Z t ≤ 2 k ( ρ t T ( y , x ) , w ) Z_t^{\le2k}(\rho_t^T(\boldsymbol{y},x),w) Z t ≤ 2 k ( ρ t T ( y , x ) , w ) can be computed from Equation (12) in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. A term with monomial x b w ℓ ∏ v ∈ T y v a v x^bw^\ell\prod_{v\in T}y_v^{a_v} x b w ℓ ∏ v ∈ T y v a v in ⟨ log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) \left\langle\log Z_t^{\le2k}(\rho_t^T(\boldsymbol{y},x),w)\right\rangle_{\le(k,n,2k)} ⟨ log Z t ≤ 2 k ( ρ t T ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) satisfies 0 ≤ b ≤ n 0\le b\le n 0 ≤ b ≤ n , 0 ≤ ℓ ≤ k 0\le\ell\le k 0 ≤ ℓ ≤ k , and ∑ v ∈ T a v ≤ 2 k \sum_{v\in T}a_v\le2k ∑ v ∈ T a v ≤ 2 k . Hence b b b has n + 1 n+1 n + 1 choices, ℓ \ell ℓ has k + 1 k+1 k + 1 choices, and ( a v ) v ∈ T (a_v)_{v\in T} ( a v ) v ∈ T has ( ∣ T ∣ + 2 k ∣ T ∣ ) \binom{|T|+2k}{|T|} ( ∣ T ∣ ∣ T ∣ + 2 k ) choices; the binomial coefficient counts the nonnegative integer vectors whose coordinate sum is at most 2 k 2k 2 k . Therefore, the number of retained terms is at most ( n + 1 ) ( k + 1 ) ( ∣ T ∣ + 2 k ∣ T ∣ ) = O ∗ ( 2 O ( k ) ) (n+1)(k+1)\binom{|T|+2k}{|T|}=O^*(2^{O(k)}) ( n + 1 ) ( k + 1 ) ( ∣ T ∣ ∣ T ∣ + 2 k ) = O ∗ ( 2 O ( k ) ) because ∣ T ∣ ≤ 2 k |T|\le2k ∣ T ∣ ≤ 2 k . Using Equation (2) with truncation after every multiplication computes all coefficients of the truncated logarithm in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. Finally, ∣ S ∣ ≤ 2 k |S|\le2k ∣ S ∣ ≤ 2 k , so computing the truncated logarithm for all 2 ∣ S ∣ 2^{|S|} 2 ∣ S ∣ sets T ⊆ S T\subseteq S T ⊆ S and evaluating Equation (11) takes O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. □
By Equation (9) and Lemma 3.4 , we can compute in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time the formal logarithm of Z t ≤ 2 k ( ω t B t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},x),w) Z t ≤ 2 k ( ω t B t ( y , x ) , w ) truncated by the bounds ( k , n , 2 k ) (k,n,2k) ( k , n , 2 k ) . We want to recover Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(\boldsymbol{y},x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) from the truncated logarithm. Intuitively, this requires two operations: changing ω t B t ( y , x ) \omega_t^{B_t}(\boldsymbol{y},x) ω t B t ( y , x ) to ω t ∂ G t ( y , x ) \omega_t^{\partial G_t}(\boldsymbol{y},x) ω t ∂ G t ( y , x ) by the substitution
y v ⟼ { y v , v ∈ ∂ G t , x , v ∈ B t ∖ ∂ G t ( v ∈ B t ) . y_v\longmapsto
\begin{cases}
y_v, & v\in\partial G_t,\\
x, & v\in B_t\setminus\partial G_t
\end{cases}
\qquad(v\in B_t). y v ⟼ { y v , x , v ∈ ∂ G t , v ∈ B t ∖ ∂ G t ( v ∈ B t ) .
and taking the exponential using the relation
⟨ exp ( ⟨ log Z t ≤ 2 k ⟩ ≤ ( k , n , 2 k ) ) ⟩ ≤ ( k , n , 2 k ) = Z t ≤ 2 k \left\langle\exp\left(\left\langle\log Z_t^{\le2k}\right\rangle_{\le(k,n,2k)}\right)\right\rangle_{\le(k,n,2k)} = Z_t^{\le2k} ⟨ exp ( ⟨ log Z t ≤ 2 k ⟩ ≤ ( k , n , 2 k ) ) ⟩ ≤ ( k , n , 2 k ) = Z t ≤ 2 k
which follows from Lemma 2.2 .
A naive approach is to first take the exponential and then replace ω t B t ( y , x ) \omega_t^{B_t}(y,x) ω t B t ( y , x ) with ω t ∂ G t ( y , x ) \omega_t^{\partial G_t}(y,x) ω t ∂ G t ( y , x ) . However, this approach does not work because it must first compute Z t ≤ 2 k ( ω t B t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{B_t}(y,x),w) Z t ≤ 2 k ( ω t B t ( y , x ) , w ) , which can contain ∑ i = 0 2 k ( ∣ B t ∣ i ) = n Θ ( k ) \sum_{i=0}^{2k}\binom{|B_t|}{i}=n^{\Theta(k)} ∑ i = 0 2 k ( i ∣ B t ∣ ) = n Θ ( k ) distinct monomials.
Alternatively, one could substitute x x x for y v y_v y v for every v ∈ B t ∖ ∂ G t v\in B_t\setminus\partial G_t v ∈ B t ∖ ∂ G t to obtain log Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) \log Z_t^{\le2k}(\omega_t^{\partial G_t}(y,x),w) log Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) and then take the exponential. However, substituting before exponentiation also fails because the substitution moves the degrees of y v y_v y v for v ∈ B t ∖ ∂ G t v\in B_t\setminus\partial G_t v ∈ B t ∖ ∂ G t into the x x x -degree. Consequently, a term that should be discarded because its total degree in the variables y v y_v y v with v ∈ B t v\in B_t v ∈ B t exceeds 2 k 2k 2 k may satisfy all three degree bounds ≤ ( k , n , 2 k ) \le(k,n,2k) ≤ ( k , n , 2 k ) after the substitution and contribute to the exponential.
To avoid these issues, we use an intermediate substitution. Introduce a fresh variable x ′ x' x ′ and define
σ t ( y , x , x ′ ) : = ( σ t , v ) v ∈ V ( G t ) , σ t , v : = { y v , v ∈ ∂ G t , x ′ , v ∈ B t ∖ ∂ G t , x , v ∈ V ( G t ) ∖ B t . \sigma_t(y,x,x') := (\sigma_{t,v})_{v\in V(G_t)}, \qquad
\sigma_{t,v} :=
\begin{cases}
y_v, & v\in\partial G_t,\\
x', & v\in B_t\setminus\partial G_t,\\
x, & v\in V(G_t)\setminus B_t.
\end{cases} σ t ( y , x , x ′ ) := ( σ t , v ) v ∈ V ( G t ) , σ t , v := ⎩ ⎨ ⎧ y v , x ′ , x , v ∈ ∂ G t , v ∈ B t ∖ ∂ G t , v ∈ V ( G t ) ∖ B t .
The substitution ω t B t ( y , x ) ↦ σ t ( y , x , x ′ ) \omega_t^{B_t}(y,x)\mapsto\sigma_t(y,x,x') ω t B t ( y , x ) ↦ σ t ( y , x , x ′ ) fixes x x x , w w w , and every y v y_v y v with v ∈ ∂ G t v\in\partial G_t v ∈ ∂ G t , and replaces y v y_v y v by x ′ x' x ′ for every v ∈ B t ∖ ∂ G t v\in B_t\setminus\partial G_t v ∈ B t ∖ ∂ G t . The important observation is that if m ′ m' m ′ is the image of a monomial m m m under this substitution, then
∑ v ∈ B t deg y v ( m ) = deg x ′ ( m ′ ) + ∑ v ∈ ∂ G t deg y v ( m ′ ) . \sum_{v\in B_t}\deg_{y_v}(m)=\deg_{x'}(m')+\sum_{v\in\partial G_t}\deg_{y_v}(m'). v ∈ B t ∑ deg y v ( m ) = deg x ′ ( m ′ ) + v ∈ ∂ G t ∑ deg y v ( m ′ ) .
That is, the original total degree in the variables y v y_v y v with v ∈ B t v\in B_t v ∈ B t is preserved as the sum of the x ′ x' x ′ -degree and the total degree in the remaining variables y v y_v y v with v ∈ ∂ G t v\in\partial G_t v ∈ ∂ G t . We first substitute σ t ( y , x , x ′ ) \sigma_t(y,x,x') σ t ( y , x , x ′ ) for ω t B t ( y , x ) \omega_t^{B_t}(y,x) ω t B t ( y , x ) to obtain ⟨ log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) ⟩ ≤ ( k , n , 2 k ) \left\langle\log Z_t^{\le2k}(\sigma_t(y,x,x'),w)\right\rangle_{\le(k,n,2k)} ⟨ log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) ⟩ ≤ ( k , n , 2 k ) , where, by a slight abuse of notation, the third coordinate of ≤ ( k , n , 2 k ) \le(k,n,2k) ≤ ( k , n , 2 k ) is deg x ′ + ∑ v ∈ ∂ G t deg y v \deg_{x'}+\sum_{v\in\partial G_t}\deg_{y_v} deg x ′ + ∑ v ∈ ∂ G t deg y v instead of ∑ v ∈ B t deg y v \sum_{v\in B_t}\deg_{y_v} ∑ v ∈ B t deg y v . We next take the exponential and apply ⟨ ⋅ ⟩ ≤ ( k , n , 2 k ) \langle\cdot\rangle_{\le(k,n,2k)} ⟨ ⋅ ⟩ ≤ ( k , n , 2 k ) , and finally substitute x x x for x ′ x' x ′ . The next lemma proves that the polynomial obtained after the substitution x ′ ↦ x x'\mapsto x x ′ ↦ x is Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(y,x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) and can be computed in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Lemma 3.5. The polynomial Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(y,x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) can be computed in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Proof. First, apply the substitution σ t ( y , x , x ′ ) \sigma_t(y,x,x') σ t ( y , x , x ′ ) to ⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) \left\langle\log Z_t^{\le2k}(\omega_t^{B_t}(y,x),w)\right\rangle_{\le(k,n,2k)} ⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) . We can claim that this substitution yields ⟨ log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) ⟩ ≤ ( k , n , 2 k ) \left\langle\log Z_t^{\le2k}(\sigma_t(y,x,x'),w)\right\rangle_{\le(k,n,2k)} ⟨ log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) ⟩ ≤ ( k , n , 2 k ) . Indeed, let
m : = x b w ℓ ∏ v ∈ B t y v a v , m ′ : = x b ( x ′ ) ∑ v ∈ B t ∖ ∂ G t a v w ℓ ∏ v ∈ ∂ G t y v a v m:=x^bw^\ell\prod_{v\in B_t}y_v^{a_v},\qquad
m':=x^b(x')^{\sum_{v\in B_t\setminus\partial G_t}a_v}w^\ell\prod_{v\in\partial G_t}y_v^{a_v} m := x b w ℓ v ∈ B t ∏ y v a v , m ′ := x b ( x ′ ) ∑ v ∈ B t ∖ ∂ G t a v w ℓ v ∈ ∂ G t ∏ y v a v
be corresponding monomials of log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) \log Z_t^{\le2k}(\omega_t^{B_t}(y,x),w) log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) and log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) \log Z_t^{\le2k}(\sigma_t(y,x,x'),w) log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) , respectively. Since
( deg w ( m ′ ) , deg x ( m ′ ) , deg x ′ ( m ′ ) + ∑ v ∈ ∂ G t deg y v ( m ′ ) ) = ( ℓ , b , ∑ v ∈ B t a v ) = ( deg w ( m ) , deg x ( m ) , ∑ v ∈ B t deg y v ( m ) ) , \left(\deg_w(m'),\deg_x(m'),\deg_{x'}(m')+\sum_{v\in\partial G_t}\deg_{y_v}(m')\right)
=
\left(\ell,b,\sum_{v\in B_t}a_v\right)
=
\left(\deg_w(m),\deg_x(m),\sum_{v\in B_t}\deg_{y_v}(m)\right), ( deg w ( m ′ ) , deg x ( m ′ ) , deg x ′ ( m ′ ) + v ∈ ∂ G t ∑ deg y v ( m ′ ) ) = ( ℓ , b , v ∈ B t ∑ a v ) = ( deg w ( m ) , deg x ( m ) , v ∈ B t ∑ deg y v ( m ) ) ,
m m m survives the truncation under ω t B t ( y , x ) \omega_t^{B_t}(\boldsymbol{y},x) ω t B t ( y , x ) if and only if m ′ m' m ′ survives the truncation under σ t ( y , x , x ′ ) \sigma_t(\boldsymbol{y},x,x') σ t ( y , x , x ′ ) .
Next, take the exponential and truncate again to ≤ ( k , n , 2 k ) \le(k,n,2k) ≤ ( k , n , 2 k ) . Since every term of Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) Z_t^{\le2k}(\sigma_t(\boldsymbol{y},x,x'),w) Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) satisfies these bounds, Lemma 2.2 gives
⟨ exp ( ⟨ log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) ⟩ ≤ ( k , n , 2 k ) ) ⟩ ≤ ( k , n , 2 k ) = Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) . \left\langle\exp\left(\left\langle\log Z_t^{\le2k}(\sigma_t(\boldsymbol{y},x,x'),w)\right\rangle_{\le(k,n,2k)}\right)\right\rangle_{\le(k,n,2k)}=Z_t^{\le2k}(\sigma_t(\boldsymbol{y},x,x'),w). ⟨ exp ( ⟨ log Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) ⟩ ≤ ( k , n , 2 k ) ) ⟩ ≤ ( k , n , 2 k ) = Z t ≤ 2 k ( σ t ( y , x , x ′ ) , w ) .
Finally, set x ′ = x x'=x x ′ = x . Since σ t ( y , x , x ) = ω t ∂ G t ( y , x ) \sigma_t(\boldsymbol{y},x,x)=\omega_t^{\partial G_t}(\boldsymbol{y},x) σ t ( y , x , x ) = ω t ∂ G t ( y , x ) , we obtain Z t ≤ 2 k ( σ t ( y , x , x ) , w ) = Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\sigma_t(\boldsymbol{y},x,x),w)=Z_t^{\le2k}(\omega_t^{\partial G_t}(\boldsymbol{y},x),w) Z t ≤ 2 k ( σ t ( y , x , x ) , w ) = Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) .
We now analyze the running time. By Equation (9) and Lemma 3.4 , we can compute the truncated polynomial ⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) \left\langle\log Z_t^{\le2k}(\omega_t^{B_t}(\boldsymbol{y},x),w)\right\rangle_{\le(k,n,2k)} ⟨ log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ⟩ ≤ ( k , n , 2 k ) in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. Substituting σ t ( y , x , x ′ ) \sigma_t(\boldsymbol{y},x,x') σ t ( y , x , x ′ ) only relabels variables and combines terms, so it does not increase the number of terms and also takes O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
We evaluate the exponential using Equation (3) and truncate after each multiplication. Every intermediate polynomial retains at most
( n + 1 ) ( k + 1 ) ( ∣ ∂ G t ∣ + 2 k + 1 ∣ ∂ G t ∣ + 1 ) = O ∗ ( 2 O ( k ) ) , (n+1)(k+1)\binom{|\partial G_t|+2k+1}{|\partial G_t|+1}=O^*(2^{O(k)}), ( n + 1 ) ( k + 1 ) ( ∣ ∂ G t ∣ + 1 ∣ ∂ G t ∣ + 2 k + 1 ) = O ∗ ( 2 O ( k ) ) ,
terms. Hence the truncated exponential can be computed in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. The substitution x ′ = x x'=x x ′ = x only combines coefficients and takes the same time.
The next lemma proves that the transitions compute the required polynomials.
Lemma 3.6. For every node t t t , given Z s ( ω s ∂ G s ( y , x ) , w ) Z_s(\omega_s^{\partial G_s}(\boldsymbol{y},x),w) Z s ( ω s ∂ G s ( y , x ) , w ) for every child s s s of t t t , the algorithm computes Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\boldsymbol{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Proof. If ∣ B t ∣ ≤ 4 k + 1 |B_t|\le4k+1 ∣ B t ∣ ≤ 4 k + 1 , then Lemma 3.1 computes Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) for every R ⊆ B t R\subseteq B_t R ⊆ B t , and substituting these polynomials into Equation (5) gives Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\boldsymbol{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) . There are 2 ∣ B t ∣ = 2 O ( k ) 2^{|B_t|}=2^{O(k)} 2 ∣ B t ∣ = 2 O ( k ) choices of R R R , so this case takes O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
If ∣ B t ∣ > 4 k + 1 |B_t|>4k+1 ∣ B t ∣ > 4 k + 1 , then Lemma 3.5 computes Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) Z_t^{\le2k}(\omega_t^{\partial G_t}(\boldsymbol{y},x),w) Z t ≤ 2 k ( ω t ∂ G t ( y , x ) , w ) in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. Applying Lemma 3.2 under ω t ∂ G t ( y , x ) \omega_t^{\partial G_t}(\boldsymbol{y},x) ω t ∂ G t ( y , x ) recovers Z t ( ω t ∂ G t ( y , x ) , w ) Z_t(\omega_t^{\partial G_t}(\boldsymbol{y},x),w) Z t ( ω t ∂ G t ( y , x ) , w ) by a coefficientwise transformation, which takes the same time.
Now we prove Theorem 1.2.
Theorem 1.2. The feasible solutions to Minimum Bisection can be counted deterministically in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. If a feasible solution exists, one can recover one within the same running time.
Proof. Let G 1 , … , G r G_1,\ldots,G_r G 1 , … , G r be the connected components of G G G , and let t i t_i t i be the root of the decomposition of G i G_i G i . For every i ∈ [ r ] i\in[r] i ∈ [ r ] , we compute Z t i ( ω t i ∂ G t i ( y , x ) , w ) Z_{t_i}(\omega_{t_i}^{\partial G_{t_i}}(\boldsymbol{y},x),w) Z t i ( ω t i ∂ G t i ( y , x ) , w ) using Theorem 1.1 and Lemma 3.6 . By Equation (4) and additivity over components, the number of feasible solutions is
∑ c = 0 k [ x ⌊ n / 2 ⌋ w c ] ∏ i = 1 r Z t i ( ω t i ∂ G t i ( y , x ) , w ) . \sum_{c=0}^{k}\left[x^{\lfloor n/2\rfloor}w^c\right]\prod_{i=1}^{r}Z_{t_i}(\omega_{t_i}^{\partial G_{t_i}}(\boldsymbol{y},x),w). c = 0 ∑ k [ x ⌊ n /2 ⌋ w c ] i = 1 ∏ r Z t i ( ω t i ∂ G t i ( y , x ) , w ) . The decompositions contain at most n n n bags in total, each transition takes O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time, and the root products take polynomial time. The intermediate numbers have polynomial bit length, so the total running time is O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) .
For witness recovery, first find the minimum feasible cost c ≤ k c\le k c ≤ k . While c > 0 c>0 c > 0 , delete an edge e e e for which the resulting instance is feasible with cost c − 1 c-1 c − 1 , and decrement c c c ; an edge of a minimum bisection always has this property. At cost zero, subset sum over the sizes of the connected components recovers a bisection. This uses only polynomially many calls to the decision algorithm and preserves the running-time bound.
Proof of Lemma 3.3
In this subsection, we prove Lemma 3.3. The proof has two main steps. The first step considers log Z t ( y , w ) \log Z_t(\boldsymbol{y},w) log Z t ( y , w ) , where y v y_v y v is a distinct variable for each v ∈ V ( G t ) v \in V(G_t) v ∈ V ( G t ) . Take any monomial with nonzero coefficient in log Z t ( y , w ) \log Z_t(\boldsymbol{y},w) log Z t ( y , w ) , write it as w c ∏ v ∈ V ( G t ) y v a v w^c\prod_{v\in V(G_t)}y_v^{a_v} w c ∏ v ∈ V ( G t ) y v a v , and let S : = { v ∈ V ( G t ) : a v > 0 } S := \{v \in V(G_t) : a_v > 0\} S := { v ∈ V ( G t ) : a v > 0 } . We show that S S S is c c c -extendable to a connected set; that is, there is a connected set A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) such that A ∩ V ( G t ) = S A \cap V(G_t) = S A ∩ V ( G t ) = S and ∣ δ G ( A ) ∣ ≤ c |\delta_G(A)| \le c ∣ δ G ( A ) ∣ ≤ c (Lemma 3.10 ).
For the specialized series log Z t ≤ 2 k ( ω B t ( k ) ( y , x ) , w ) \log Z_t^{\le2k}(\boldsymbol{\omega}_{B_t}^{(k)}(\boldsymbol{y},x),w) log Z t ≤ 2 k ( ω B t ( k ) ( y , x ) , w ) in Lemma 3.3, we can state that such an A A A satisfies ∣ A ∩ B t ∣ ≤ 2 k |A \cap B_t| \le2k ∣ A ∩ B t ∣ ≤ 2 k . The second step is to use important cuts to enumerate in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time a family C t \mathcal{C}_t C t that contains A ∩ B t A \cap B_t A ∩ B t for every connected set A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) satisfying ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \le k ∣ δ G ( A ) ∣ ≤ k and ∣ A ∩ B t ∣ ≤ 2 k |A \cap B_t| \le2k ∣ A ∩ B t ∣ ≤ 2 k (Lemma 3.11 ). Hence S = A ∩ B t ∈ C t S = A \cap B_t \in\mathcal{C}_t S = A ∩ B t ∈ C t .
We now prove the first step. We would like to choose a minimum-cost extension of S S S and factor the part of Z t ( y , w ) Z_t(\boldsymbol{y},w) Z t ( y , w ) supported on subsets of S S S as a product over its connected components, so that taking the formal logarithm turns the product into a sum. The y y y -part of each monomial factors automatically over these components. The difficulty lies in the w w w -part. The exponent of w w w in the term indexed by R ⊆ S R \subseteq S R ⊆ S is μ t ( R ) \mu_t(R) μ t ( R ) , and since μ t ( R ) \mu_t(R) μ t ( R ) is defined by minimizing over extensions outside G t G_t G t , it is not immediate that these exponents are additive over the components.
To establish this additivity, for each S ⊆ V ( G t ) S \subseteq V(G_t) S ⊆ V ( G t ) , let Φ t ( S ) \Phi_t(S) Φ t ( S ) be the inclusion-wise largest member of
{ A ⊆ V ( G ) : A ∩ V ( G t ) = S , ∣ δ G ( A ) ∣ = μ t ( S ) } . \{A \subseteq V(G) : A \cap V(G_t) = S,\quad|\delta_G(A)| = \mu_t(S)\}. { A ⊆ V ( G ) : A ∩ V ( G t ) = S , ∣ δ G ( A ) ∣ = μ t ( S )} .
We first prove that Φ t ( S ) \Phi_t(S) Φ t ( S ) is well-defined for every S ⊆ V ( G t ) S \subseteq V(G_t) S ⊆ V ( G t ) .
Lemma 3.7. For every S ⊆ V ( G t ) S \subseteq V(G_t) S ⊆ V ( G t ) , the set Φ t ( S ) \Phi_t(S) Φ t ( S ) is well-defined.
Proof. The family defining Φ t ( S ) \Phi_t(S) Φ t ( S ) is nonempty by the definition of μ t ( S ) \mu_t(S) μ t ( S ) . Let A , A ′ ⊆ V ( G ) A,A' \subseteq V(G) A , A ′ ⊆ V ( G ) satisfy A ∩ V ( G t ) = A ′ ∩ V ( G t ) = S A \cap V(G_t) = A' \cap V(G_t) = S A ∩ V ( G t ) = A ′ ∩ V ( G t ) = S and ∣ δ G ( A ) ∣ = ∣ δ G ( A ′ ) ∣ = μ t ( S ) |\delta_G(A)| = |\delta_G(A')| = \mu_t(S) ∣ δ G ( A ) ∣ = ∣ δ G ( A ′ ) ∣ = μ t ( S ) . Both A ∩ A ′ A \cap A' A ∩ A ′ and A ∪ A ′ A \cup A' A ∪ A ′ intersect V ( G t ) V(G_t) V ( G t ) in S S S , and thus ∣ δ G ( A ∩ A ′ ) ∣ ≥ μ t ( S ) |\delta_G(A \cap A')| \ge\mu_t(S) ∣ δ G ( A ∩ A ′ ) ∣ ≥ μ t ( S ) and ∣ δ G ( A ∪ A ′ ) ∣ ≥ μ t ( S ) |\delta_G(A \cup A')| \ge\mu_t(S) ∣ δ G ( A ∪ A ′ ) ∣ ≥ μ t ( S ) . By submodularity of the cut function,
2 μ t ( S ) = ∣ δ G ( A ) ∣ + ∣ δ G ( A ′ ) ∣ ≥ ∣ δ G ( A ∩ A ′ ) ∣ + ∣ δ G ( A ∪ A ′ ) ∣ ≥ 2 μ t ( S ) . 2\mu_t(S) = |\delta_G(A)| + |\delta_G(A')| \ge|\delta_G(A \cap A')| + |\delta_G(A \cup A')| \ge2\mu_t(S). 2 μ t ( S ) = ∣ δ G ( A ) ∣ + ∣ δ G ( A ′ ) ∣ ≥ ∣ δ G ( A ∩ A ′ ) ∣ + ∣ δ G ( A ∪ A ′ ) ∣ ≥ 2 μ t ( S ) .
Thus, both inequalities are equalities. In particular, A ∪ A ′ A \cup A' A ∪ A ′ also satisfies ∣ δ G ( A ∪ A ′ ) ∣ = μ t ( S ) |\delta_G(A \cup A')| = \mu_t(S) ∣ δ G ( A ∪ A ′ ) ∣ = μ t ( S ) , and hence the family defining Φ t ( S ) \Phi_t(S) Φ t ( S ) is closed under unions, which implies that its inclusion-wise largest member Φ t ( S ) \Phi_t(S) Φ t ( S ) is unique. □ \square □
We also have the following.
Lemma 3.8. If S ′ ⊆ S ⊆ V ( G t ) S' \subseteq S \subseteq V(G_t) S ′ ⊆ S ⊆ V ( G t ) , then Φ t ( S ′ ) ⊆ Φ t ( S ) \Phi_t(S') \subseteq\Phi_t(S) Φ t ( S ′ ) ⊆ Φ t ( S ) .
Proof. Let S ′ ⊆ S S' \subseteq S S ′ ⊆ S . The sets Φ t ( S ′ ) ∩ Φ t ( S ) \Phi_t(S') \cap\Phi_t(S) Φ t ( S ′ ) ∩ Φ t ( S ) and Φ t ( S ′ ) ∪ Φ t ( S ) \Phi_t(S') \cup\Phi_t(S) Φ t ( S ′ ) ∪ Φ t ( S ) intersect V ( G t ) V(G_t) V ( G t ) in S ′ S' S ′ and S S S , respectively. Hence
μ t ( S ′ ) + μ t ( S ) = ∣ δ G ( Φ t ( S ′ ) ) ∣ + ∣ δ G ( Φ t ( S ) ) ∣ ≥ ∣ δ G ( Φ t ( S ′ ) ∩ Φ t ( S ) ) ∣ + ∣ δ G ( Φ t ( S ′ ) ∪ Φ t ( S ) ) ∣ ≥ μ t ( S ′ ) + μ t ( S ) . \begin{aligned}
\mu_t(S') + \mu_t(S) = |\delta_G(\Phi_t(S'))| + |\delta_G(\Phi_t(S))| \\
&\ge|\delta_G(\Phi_t(S') \cap\Phi_t(S))| + |\delta_G(\Phi_t(S') \cup\Phi_t(S))| \ge\mu_t(S') + \mu_t(S).
\end{aligned} μ t ( S ′ ) + μ t ( S ) = ∣ δ G ( Φ t ( S ′ )) ∣ + ∣ δ G ( Φ t ( S )) ∣ ≥ ∣ δ G ( Φ t ( S ′ ) ∩ Φ t ( S )) ∣ + ∣ δ G ( Φ t ( S ′ ) ∪ Φ t ( S )) ∣ ≥ μ t ( S ′ ) + μ t ( S ) .
Thus Φ t ( S ′ ) ∪ Φ t ( S ) \Phi_t(S') \cup\Phi_t(S) Φ t ( S ′ ) ∪ Φ t ( S ) is a minimizer in the definition of μ t ( S ) \mu_t(S) μ t ( S ) . The maximality of Φ t ( S ) \Phi_t(S) Φ t ( S ) gives Φ t ( S ′ ) ∪ Φ t ( S ) = Φ t ( S ) \Phi_t(S') \cup\Phi_t(S) = \Phi_t(S) Φ t ( S ′ ) ∪ Φ t ( S ) = Φ t ( S ) , and hence Φ t ( S ′ ) ⊆ Φ t ( S ) \Phi_t(S') \subseteq\Phi_t(S) Φ t ( S ′ ) ⊆ Φ t ( S ) . □ \square □
For S ⊆ V ( G t ) S \subseteq V(G_t) S ⊆ V ( G t ) , let K S \mathcal{K}_S K S be the family of connected components of G [ Φ t ( S ) ] G[\Phi_t(S)] G [ Φ t ( S )] . The next lemma shows that every member of K S \mathcal{K}_S K S intersects V ( G t ) V(G_t) V ( G t ) .
Lemma 3.9. For every S ⊆ V ( G t ) S \subseteq V(G_t) S ⊆ V ( G t ) and every K ∈ K S K \in\mathcal{K}_S K ∈ K S , we have K ∩ V ( G t ) ≠ ∅ K \cap V(G_t) \ne\varnothing K ∩ V ( G t ) = ∅ .
Proof. Suppose that K ∩ V ( G t ) = ∅ K \cap V(G_t) = \varnothing K ∩ V ( G t ) = ∅ for some K ∈ K S K \in\mathcal{K}_S K ∈ K S . The set Φ t ( S ) ∖ K \Phi_t(S) \setminus K Φ t ( S ) ∖ K still intersects V ( G t ) V(G_t) V ( G t ) in S S S . Since K ∩ V ( G t ) = ∅ K \cap V(G_t) = \varnothing K ∩ V ( G t ) = ∅ and V ( G t ) ≠ ∅ V(G_t) \ne\varnothing V ( G t ) = ∅ , we have K ≠ V ( G ) K \ne V(G) K = V ( G ) . Since G G G is connected, ∣ δ G ( K ) ∣ > 0 |\delta_G(K)| > 0 ∣ δ G ( K ) ∣ > 0 . Moreover, there is no edge between K K K and Φ t ( S ) ∖ K \Phi_t(S) \setminus K Φ t ( S ) ∖ K , so
∣ δ G ( Φ t ( S ) ∖ K ) ∣ = ∣ δ G ( Φ t ( S ) ) ∣ − ∣ δ G ( K ) ∣ < ∣ δ G ( Φ t ( S ) ) ∣ = μ t ( S ) . |\delta_G(\Phi_t(S) \setminus K)| = |\delta_G(\Phi_t(S))| - |\delta_G(K)| < |\delta_G(\Phi_t(S))| = \mu_t(S). ∣ δ G ( Φ t ( S ) ∖ K ) ∣ = ∣ δ G ( Φ t ( S )) ∣ − ∣ δ G ( K ) ∣ < ∣ δ G ( Φ t ( S )) ∣ = μ t ( S ) .
This contradicts the definition of μ t ( S ) \mu_t(S) μ t ( S ) .
The next lemma is the goal of the first step toward Lemma 3.3.
Lemma 3.10. Suppose that c ≤ k c \le k c ≤ k and that the coefficient of w c ∏ v ∈ V ( G t ) y v a v w^c \prod_{v \in V(G_t)} y_v^{a_v} w c ∏ v ∈ V ( G t ) y v a v in log Z t ( y , w ) \log Z_t(y,w) log Z t ( y , w ) is nonzero. For S : = { v ∈ V ( G t ) : a v > 0 } S := \{v \in V(G_t) : a_v > 0\} S := { v ∈ V ( G t ) : a v > 0 } , the set Φ t ( S ) \Phi_t(S) Φ t ( S ) is connected and ∣ δ G ( Φ t ( S ) ) ∣ ≤ c |\delta_G(\Phi_t(S))| \le c ∣ δ G ( Φ t ( S )) ∣ ≤ c .
Proof. Define the substitution π t S ( y ) : = ( π t , v S ) v ∈ V ( G t ) \pi_t^S(y) := (\pi_{t,v}^S)_{v \in V(G_t)} π t S ( y ) := ( π t , v S ) v ∈ V ( G t ) , where, for each v ∈ V ( G t ) v \in V(G_t) v ∈ V ( G t ) ,
π t , v S : = { y v , v ∈ S , 0 , v ∈ V ( G t ) ∖ S . \pi_{t,v}^S :=
\begin{cases}
y_v, & v \in S,\\
0, & v \in V(G_t) \setminus S.
\end{cases} π t , v S := { y v , 0 , v ∈ S , v ∈ V ( G t ) ∖ S .
Consider log Z t ( π t S ( y ) , w ) \log Z_t(\pi_t^S(y),w) log Z t ( π t S ( y ) , w ) . Setting y v = 0 y_v = 0 y v = 0 for every v ∉ S v \notin S v ∈ / S leaves only the monomials of log Z t ( y , w ) \log Z_t(y,w) log Z t ( y , w ) whose support is contained in S S S . In particular, the coefficient of w c ∏ v ∈ V ( G t ) y v a v w^c \prod_{v \in V(G_t)} y_v^{a_v} w c ∏ v ∈ V ( G t ) y v a v is unchanged.
Let K S \mathcal{K}_S K S be the family of connected components of G [ Φ t ( S ) ] G[\Phi_t(S)] G [ Φ t ( S )] . We will prove below that the polynomial Z t ( π t S ( y ) , w ) Z_t(\pi_t^S(y),w) Z t ( π t S ( y ) , w ) has the following factorization:
Z t ( π t S ( y ) , w ) = ⟨ ∏ K ∈ K S ( ∑ R ⊆ K ∩ V ( G t ) ( w μ t ( R ) ⋅ ∏ v ∈ R y v ) ) ⟩ deg w ≤ k . (13) Z_t(\pi_t^S(y),w)
=
\left\langle
\prod_{K \in\mathcal{K}_S}
\left(
\sum_{R \subseteq K \cap V(G_t)}
\left(
w^{\mu_t(R)} \cdot\prod_{v \in R} y_v
\right)
\right)
\right\rangle_{\deg_w \le k}. \tag*{(13)} Z t ( π t S ( y ) , w ) = ⟨ K ∈ K S ∏ R ⊆ K ∩ V ( G t ) ∑ ( w μ t ( R ) ⋅ v ∈ R ∏ y v ) ⟩ d e g w ≤ k . (13)
Taking formal logarithms in Equation (13) and retaining only terms of w w w -degree at most k k k gives
⟨ log Z t ( π t S ( y ) , w ) ⟩ deg w ≤ k = ⟨ ∑ K ∈ K S log ( ∑ R ⊆ K ∩ V ( G t ) ( w μ t ( R ) ⋅ ∏ v ∈ R y v ) ) ⟩ deg w ≤ k . \left\langle\log Z_t(\pi_t^S(y),w) \right\rangle_{\deg_w \le k}
=
\left\langle
\sum_{K \in\mathcal{K}_S}
\log\left(
\sum_{R \subseteq K \cap V(G_t)}
\left(
w^{\mu_t(R)} \cdot\prod_{v \in R} y_v
\right)
\right)
\right\rangle_{\deg_w \le k}. ⟨ log Z t ( π t S ( y ) , w ) ⟩ d e g w ≤ k = ⟨ K ∈ K S ∑ log R ⊆ K ∩ V ( G t ) ∑ ( w μ t ( R ) ⋅ v ∈ R ∏ y v ) ⟩ d e g w ≤ k .
Each summand on the right-hand side uses variables from only one member of K S \mathcal{K}_S K S . If K S \mathcal{K}_S K S has at least two members, then Lemma 3.9 implies that every member intersects S S S , so no summand has support S S S ; if K S \mathcal{K}_S K S is empty, the right-hand side is zero. Thus, unless K S \mathcal{K}_S K S is a singleton, the coefficient of w c ∏ v ∈ V ( G t ) y v a v w^c \prod_{v \in V(G_t)} y_v^{a_v} w c ∏ v ∈ V ( G t ) y v a v is zero, contradicting the hypothesis and its preservation under π t S \pi_t^S π t S .
It remains to prove Equation (13) . It suffices to show that, for every R ⊆ S R \subseteq S R ⊆ S ,
μ t ( R ) = ∑ K ∈ K S μ t ( R ∩ K ) . (14) \mu_t(R) = \sum_{K \in\mathcal{K}_S} \mu_t(R \cap K). \tag*{(14)} μ t ( R ) = K ∈ K S ∑ μ t ( R ∩ K ) . (14)
Indeed, if this is proved, then, since the sets K ∩ V ( G t ) K \cap V(G_t) K ∩ V ( G t ) for K ∈ K S K \in\mathcal{K}_S K ∈ K S partition S S S , we have
Z t ( π t S ( y ) , w ) = ⟨ ∑ R ⊆ S ( w μ t ( R ) ⋅ ∏ v ∈ R y v ) ⟩ deg w ≤ k = ⟨ ∏ K ∈ K S ( ∑ R ⊆ K ∩ V ( G t ) ( w μ t ( R ) ⋅ ∏ v ∈ R y v ) ) ⟩ deg w ≤ k . \begin{aligned}
Z_t(\pi_t^S(y),w)
=
\left\langle
\sum_{R \subseteq S}
\left(
w^{\mu_t(R)} \cdot\prod_{v \in R} y_v
\right)
\right\rangle_{\deg_w \le k} \\
&=
\left\langle
\prod_{K \in\mathcal{K}_S}
\left(
\sum_{R \subseteq K \cap V(G_t)}
\left(
w^{\mu_t(R)} \cdot\prod_{v \in R} y_v
\right)
\right)
\right\rangle_{\deg_w \le k}.
\end{aligned} Z t ( π t S ( y ) , w ) = ⟨ R ⊆ S ∑ ( w μ t ( R ) ⋅ v ∈ R ∏ y v ) ⟩ d e g w ≤ k = ⟨ K ∈ K S ∏ R ⊆ K ∩ V ( G t ) ∑ ( w μ t ( R ) ⋅ v ∈ R ∏ y v ) ⟩ d e g w ≤ k .
which is Equation (13) .
Now we prove Equation (14) . By Lemma 3.8 , Φ t ( R ) ⊆ Φ t ( S ) \Phi_t(R) \subseteq\Phi_t(S) Φ t ( R ) ⊆ Φ t ( S ) . The sets Φ t ( R ) ∩ K \Phi_t(R) \cap K Φ t ( R ) ∩ K for K ∈ K S K \in\mathcal{K}_S K ∈ K S partition Φ t ( R ) \Phi_t(R) Φ t ( R ) , and there are no edges between distinct members of K S \mathcal{K}_S K S . Hence
μ t ( R ) = ∑ K ∈ K S ∣ δ G ( Φ t ( R ) ∩ K ) ∣ ≥ ∑ K ∈ K S μ t ( R ∩ K ) . \mu_t(R) = \sum_{K\in\mathcal{K}_S} \left|\delta_G(\Phi_t(R)\cap K)\right| \ge\sum_{K\in\mathcal{K}_S} \mu_t(R\cap K). μ t ( R ) = K ∈ K S ∑ ∣ δ G ( Φ t ( R ) ∩ K ) ∣ ≥ K ∈ K S ∑ μ t ( R ∩ K ) .
Conversely, fix K ∈ K S K \in\mathcal{K}_S K ∈ K S . We show that Φ t ( R ∩ K ) ⊆ K \Phi_t(R\cap K) \subseteq K Φ t ( R ∩ K ) ⊆ K (Figure 3 ). By Lemma 3.8 , Φ t ( R ∩ K ) ⊆ Φ t ( S ) \Phi_t(R\cap K) \subseteq\Phi_t(S) Φ t ( R ∩ K ) ⊆ Φ t ( S ) . By Lemma 3.9 , every connected component of G [ Φ t ( R ∩ K ) ] G[\Phi_t(R\cap K)] G [ Φ t ( R ∩ K )] contains a vertex of V ( G t ) V(G_t) V ( G t ) . Since Φ t ( R ∩ K ) ∩ V ( G t ) = R ∩ K \Phi_t(R\cap K)\cap V(G_t)=R\cap K Φ t ( R ∩ K ) ∩ V ( G t ) = R ∩ K , this vertex lies in R ∩ K ⊆ K R\cap K \subseteq K R ∩ K ⊆ K . Since K K K is a connected component of G [ Φ t ( S ) ] G[\Phi_t(S)] G [ Φ t ( S )] , every connected component of G [ Φ t ( R ∩ K ) ] G[\Phi_t(R\cap K)] G [ Φ t ( R ∩ K )] lies in K K K . Hence Φ t ( R ∩ K ) ⊆ K \Phi_t(R\cap K) \subseteq K Φ t ( R ∩ K ) ⊆ K . Consequently, ⋃ K ∈ K S Φ t ( R ∩ K ) \bigcup_{K\in\mathcal{K}_S}\Phi_t(R\cap K) ⋃ K ∈ K S Φ t ( R ∩ K ) intersects V ( G t ) V(G_t) V ( G t ) exactly in R R R . Moreover, since no edge joins distinct members of K S \mathcal{K}_S K S , the edge sets δ G ( Φ t ( R ∩ K ) ) \delta_G(\Phi_t(R\cap K)) δ G ( Φ t ( R ∩ K )) for K ∈ K S K\in\mathcal{K}_S K ∈ K S are pairwise disjoint. Therefore,
Figure 3. The gray regions are the components K 1 , K 2 K_1,K_2 K 1 , K 2 of G [ Φ t ( S ) ] G[\Phi_t(S)] G [ Φ t ( S )] . The darker regions show Φ t ( R ∩ K 1 ) \Phi_t(R\cap K_1) Φ t ( R ∩ K 1 ) and Φ t ( R ∩ K 2 ) \Phi_t(R\cap K_2) Φ t ( R ∩ K 2 ) , each with two connected components, all meeting R R R .
μ t ( R ) ≤ ∣ δ G ( ⋃ K ∈ K S Φ t ( R ∩ K ) ) ∣ = ∑ K ∈ K S ∣ δ G ( Φ t ( R ∩ K ) ) ∣ = ∑ K ∈ K S μ t ( R ∩ K ) , \mu_t(R) \le\left|\delta_G\left(\bigcup_{K\in\mathcal{K}_S}\Phi_t(R\cap K)\right)\right| = \sum_{K\in\mathcal{K}_S}\left|\delta_G(\Phi_t(R\cap K))\right| = \sum_{K\in\mathcal{K}_S}\mu_t(R\cap K), μ t ( R ) ≤ δ G ( K ∈ K S ⋃ Φ t ( R ∩ K ) ) = K ∈ K S ∑ ∣ δ G ( Φ t ( R ∩ K )) ∣ = K ∈ K S ∑ μ t ( R ∩ K ) ,
which proves Equation (14) . This proves Equation (13) , and hence Φ t ( S ) \Phi_t(S) Φ t ( S ) is connected.
Now we prove ∣ δ G ( Φ t ( S ) ) ∣ ≤ c \left|\delta_G(\Phi_t(S))\right| \le c ∣ δ G ( Φ t ( S )) ∣ ≤ c . By the definitions of Z t Z_t Z t and the formal logarithm,
log Z t ( y , w ) = ∑ j ≥ 1 ( − 1 ) j + 1 j ( ∑ ∅ ≠ A t ⊆ V ( G t ) μ t ( A t ) ≤ k ( w μ t ( A t ) ⋅ ∏ v ∈ A t y v ) ) j . \log Z_t(y,w)=\sum_{j\ge1}\frac{(-1)^{j+1}}{j}\left(\sum_{\substack{\varnothing\ne A_t\subseteq V(G_t)\\ \mu_t(A_t)\le k}}\left(w^{\mu_t(A_t)}\cdot\prod_{v\in A_t}y_v\right)\right)^j. log Z t ( y , w ) = j ≥ 1 ∑ j ( − 1 ) j + 1 ∅ = A t ⊆ V ( G t ) μ t ( A t ) ≤ k ∑ ( w μ t ( A t ) ⋅ v ∈ A t ∏ y v ) j .
Since the coefficient of w c ∏ v ∈ V ( G t ) y v a v w^c\prod_{v\in V(G_t)}y_v^{a_v} w c ∏ v ∈ V ( G t ) y v a v is nonzero, this monomial occurs in the expansion of one of the powers on the right-hand side. Hence, for some j ≥ 1 j\ge1 j ≥ 1 , there are nonempty sets A t , 1 , … , A t , j ⊆ V ( G t ) A_{t,1},\ldots,A_{t,j}\subseteq V(G_t) A t , 1 , … , A t , j ⊆ V ( G t ) such that
w c ∏ v ∈ V ( G t ) y v a v = ∏ ℓ = 1 j ( w μ t ( A t , ℓ ) ⋅ ∏ v ∈ A t , ℓ y v ) = w ∑ ℓ = 1 j μ t ( A t , ℓ ) ∏ v ∈ V ( G t ) y v ∣ { ℓ : v ∈ A t , ℓ } ∣ . w^c\prod_{v\in V(G_t)}y_v^{a_v}=\prod_{\ell=1}^{j}\left(w^{\mu_t(A_{t,\ell})}\cdot\prod_{v\in A_{t,\ell}}y_v\right)=w^{\sum_{\ell=1}^{j}\mu_t(A_{t,\ell})}\prod_{v\in V(G_t)}y_v^{|\{\ell:v\in A_{t,\ell}\}|}. w c v ∈ V ( G t ) ∏ y v a v = ℓ = 1 ∏ j w μ t ( A t , ℓ ) ⋅ v ∈ A t , ℓ ∏ y v = w ∑ ℓ = 1 j μ t ( A t , ℓ ) v ∈ V ( G t ) ∏ y v ∣ { ℓ : v ∈ A t , ℓ } ∣ .
Comparing the exponents of w w w and the variables y v y_v y v gives
∑ ℓ = 1 j μ t ( A t , ℓ ) = c , ⋃ ℓ = 1 j A t , ℓ = S . \sum_{\ell=1}^{j} \mu_t(A_{t,\ell}) = c,\qquad\bigcup_{\ell=1}^{j} A_{t,\ell} = S. ℓ = 1 ∑ j μ t ( A t , ℓ ) = c , ℓ = 1 ⋃ j A t , ℓ = S .
Using Φ t ( A t , ℓ ) ∩ V ( G t ) = A t , ℓ \Phi_t(A_{t,\ell}) \cap V(G_t) = A_{t,\ell} Φ t ( A t , ℓ ) ∩ V ( G t ) = A t , ℓ , we obtain
( ⋃ ℓ = 1 j Φ t ( A t , ℓ ) ) ∩ V ( G t ) = ⋃ ℓ = 1 j A t , ℓ = S . \left(\bigcup_{\ell=1}^{j} \Phi_t(A_{t,\ell})\right) \cap V(G_t) = \bigcup_{\ell=1}^{j} A_{t,\ell} = S. ( ℓ = 1 ⋃ j Φ t ( A t , ℓ ) ) ∩ V ( G t ) = ℓ = 1 ⋃ j A t , ℓ = S .
Moreover,
δ G ( ⋃ ℓ = 1 j Φ t ( A t , ℓ ) ) ⊆ ⋃ ℓ = 1 j δ G ( Φ t ( A t , ℓ ) ) . \delta_G\left(\bigcup_{\ell=1}^{j} \Phi_t(A_{t,\ell})\right) \subseteq\bigcup_{\ell=1}^{j} \delta_G(\Phi_t(A_{t,\ell})). δ G ( ℓ = 1 ⋃ j Φ t ( A t , ℓ ) ) ⊆ ℓ = 1 ⋃ j δ G ( Φ t ( A t , ℓ )) .
Hence
μ t ( S ) ≤ ∣ δ G ( ⋃ ℓ = 1 j Φ t ( A t , ℓ ) ) ∣ ≤ ∑ ℓ = 1 j ∣ δ G ( Φ t ( A t , ℓ ) ) ∣ = ∑ ℓ = 1 j μ t ( A t , ℓ ) = c . \mu_t(S) \leq\left|\delta_G\left(\bigcup_{\ell=1}^{j} \Phi_t(A_{t,\ell})\right)\right| \leq\sum_{\ell=1}^{j} |\delta_G(\Phi_t(A_{t,\ell}))| = \sum_{\ell=1}^{j} \mu_t(A_{t,\ell}) = c. μ t ( S ) ≤ δ G ( ℓ = 1 ⋃ j Φ t ( A t , ℓ ) ) ≤ ℓ = 1 ∑ j ∣ δ G ( Φ t ( A t , ℓ )) ∣ = ℓ = 1 ∑ j μ t ( A t , ℓ ) = c .
By the definition of Φ t \Phi_t Φ t , we have ∣ δ G ( Φ t ( S ) ) ∣ = μ t ( S ) ≤ c |\delta_G(\Phi_t(S))| = \mu_t(S) \leq c ∣ δ G ( Φ t ( S )) ∣ = μ t ( S ) ≤ c . □ \square □
The next lemma supplies the second part toward Lemma 3.3.
Lemma 3.11. In O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time, one can enumerate a downward-closed family C t ⊆ 2 B t C_t \subseteq2^{B_t} C t ⊆ 2 B t of size O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) such that every member has size at most 2 k 2k 2 k and A ∩ B t ∈ C t A \cap B_t \in C_t A ∩ B t ∈ C t for every connected A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) with ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \leq k ∣ δ G ( A ) ∣ ≤ k and ∣ A ∩ B t ∣ ≤ 2 k |A \cap B_t| \leq2k ∣ A ∩ B t ∣ ≤ 2 k .
Proof. Let H t H_t H t be obtained from G G G by adding a new vertex r t r_t r t adjacent to every vertex of B t B_t B t . For every C ⊆ V ( G ) C \subseteq V(G) C ⊆ V ( G ) , we have ∣ δ H t ( C ) ∣ = ∣ δ G ( C ) ∣ + ∣ C ∩ B t ∣ |\delta_{H_t}(C)| = |\delta_G(C)| + |C \cap B_t| ∣ δ H t ( C ) ∣ = ∣ δ G ( C ) ∣ + ∣ C ∩ B t ∣ . For each v ∈ B t v \in B_t v ∈ B t , use Lemma 2.1 to enumerate all important { v } \{v\} { v } –{ r t } \{r_t\} { r t } cuts D D D in H t H_t H t satisfying ∣ δ H t ( D ) ∣ ≤ 3 k |\delta_{H_t}(D)| \leq3k ∣ δ H t ( D ) ∣ ≤ 3 k . Let D t D_t D t be the union over v ∈ B t v \in B_t v ∈ B t of the enumerated cut families and set
C t : = { ∅ } ∪ { T : T ⊆ D ∩ B t , ∣ T ∣ ≤ 2 k , D ∈ D t } . (15) C_t := \{\emptyset\} \cup\{T : T \subseteq D \cap B_t,\ |T| \leq2k,\ D \in D_t\}. \tag*{(15)} C t := { ∅ } ∪ { T : T ⊆ D ∩ B t , ∣ T ∣ ≤ 2 k , D ∈ D t } . (15)
By definition, C t C_t C t is downward closed. Moreover, Lemma 2.1 gives ∣ D t ∣ ≤ ∣ B t ∣ 4 3 k |D_t| \leq|B_t|4^{3k} ∣ D t ∣ ≤ ∣ B t ∣ 4 3 k , and each D ∈ D t D \in D_t D ∈ D t contributes at most 2 ∣ D ∩ B t ∣ ≤ 2 3 k 2^{|D \cap B_t|} \leq2^{3k} 2 ∣ D ∩ B t ∣ ≤ 2 3 k members to C t C_t C t . Hence ∣ C t ∣ ≤ 1 + ∣ B t ∣ 4 3 k 2 3 k = O ∗ ( 2 O ( k ) ) |C_t| \leq1 + |B_t|4^{3k}2^{3k} = O^*(2^{O(k)}) ∣ C t ∣ ≤ 1 + ∣ B t ∣ 4 3 k 2 3 k = O ∗ ( 2 O ( k ) ) .
Let A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) satisfy the hypotheses of the lemma and set T : = A ∩ B t T := A \cap B_t T := A ∩ B t . If T = ∅ T = \emptyset T = ∅ , then T ∈ C t T \in C_t T ∈ C t by Equation (15) . Otherwise choose v ∈ T v \in T v ∈ T . Then ∣ δ H t ( A ) ∣ = ∣ δ G ( A ) ∣ + ∣ T ∣ ≤ 3 k |\delta_{H_t}(A)| = |\delta_G(A)| + |T| \leq3k ∣ δ H t ( A ) ∣ = ∣ δ G ( A ) ∣ + ∣ T ∣ ≤ 3 k . Hence A A A is a connected { v } \{v\} { v } –{ r t } \{r_t\} { r t } cut and is contained in an important such cut D D D with ∣ δ H t ( D ) ∣ ≤ ∣ δ H t ( A ) ∣ |\delta_{H_t}(D)| \leq|\delta_{H_t}(A)| ∣ δ H t ( D ) ∣ ≤ ∣ δ H t ( A ) ∣ . The enumeration contains D D D by Lemma 2.1 . Since T ⊆ D ∩ B t T \subseteq D \cap B_t T ⊆ D ∩ B t and ∣ T ∣ ≤ 2 k |T| \leq2k ∣ T ∣ ≤ 2 k , Equation (15) gives T ∈ C t T \in C_t T ∈ C t . □ \square □
Now we prove Lemma 3.3.
Lemma 3.3. In O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time, one can enumerate a downward-closed family C t ⊆ 2 B t C_t \subseteq2^{B_t} C t ⊆ 2 B t of size O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) such that every member has size at most 2 k 2k 2 k and
[ x b w ℓ ∏ v ∈ B t y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ≠ 0 ⟹ { v ∈ B t : a v > 0 } ∈ C t \left[x^b w^\ell\prod_{v \in B_t} y_v^{a_v}\right]\log Z_t^{\leq2k}\left(\omega_t^{B_t}(\mathbf{y},x),w\right) \ne0 \quad\Longrightarrow\quad\{v \in B_t : a_v > 0\} \in C_t [ x b w ℓ v ∈ B t ∏ y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = 0 ⟹ { v ∈ B t : a v > 0 } ∈ C t whenever ℓ ≤ k \ell\leq k ℓ ≤ k and ∣ { v ∈ B t : a v > 0 } ∣ ≤ 2 k |\{v \in B_t : a_v > 0\}| \leq2k ∣ { v ∈ B t : a v > 0 } ∣ ≤ 2 k .
Proof. Let C t C_t C t be the family produced by Lemma 3.11 . We show that it satisfies the statement of Lemma 3.3. The running time, downward-closedness, the size of C t C_t C t , and the bound on the size of its members follow immediately from Lemma 3.11 . It remains to show that the B t B_t B t -support of every monomial satisfying the hypothesis belongs to C t C_t C t .
Choose nonnegative integers b , ℓ b,\ell b , ℓ and ( a v ) v ∈ B t (a_v)_{v\in B_t} ( a v ) v ∈ B t such that
[ x b w ℓ ∏ v ∈ B t y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) ≠ 0 , ℓ ≤ k , ∣ { v ∈ B t : a v > 0 } ∣ ≤ 2 k . \left[x^b w^\ell\prod_{v\in B_t} y_v^{a_v}\right]\log Z_t^{\le2k}(\omega_t^{B_t}(y,x),w)\ne0,\qquad\ell\le k,\qquad\lvert\{v\in B_t:a_v>0\}\rvert\le2k. [ x b w ℓ v ∈ B t ∏ y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = 0 , ℓ ≤ k , ∣{ v ∈ B t : a v > 0 }∣ ≤ 2 k .
Set T : = { v ∈ B t : a v > 0 } T:=\{v\in B_t:a_v>0\} T := { v ∈ B t : a v > 0 } . We claim that the exponents ( a v ) v ∈ B t (a_v)_{v\in B_t} ( a v ) v ∈ B t can be extended to ( a v ) v ∈ V ( G t ) (a_v)_{v\in V(G_t)} ( a v ) v ∈ V ( G t ) such that
[ w ℓ ∏ v ∈ V ( G t ) y v a v ] log Z t ≤ 2 k ( y , w ) ≠ 0. (16) \left[w^\ell\prod_{v\in V(G_t)}y_v^{a_v}\right]\log Z_t^{\le2k}(y,w)\ne0. \tag*{(16)} w ℓ v ∈ V ( G t ) ∏ y v a v log Z t ≤ 2 k ( y , w ) = 0. (16)
Given such an extension, set S : = { v ∈ V ( G t ) : a v > 0 } S:=\{v\in V(G_t):a_v>0\} S := { v ∈ V ( G t ) : a v > 0 } and A : = Φ t ( S ) A:=\Phi_t(S) A := Φ t ( S ) . Then A ∩ B t = S ∩ B t = T A\cap B_t=S\cap B_t=T A ∩ B t = S ∩ B t = T , and Lemma 3.10 shows that A A A is connected and ∣ δ G ( A ) ∣ ≤ ℓ \lvert\delta_{G}(A)\rvert\le\ell ∣ δ G ( A )∣ ≤ ℓ . Since ℓ ≤ k \ell\le k ℓ ≤ k and ∣ T ∣ ≤ 2 k \lvert T\rvert\le2k ∣ T ∣ ≤ 2 k , Lemma 3.11 gives T = A ∩ B t ∈ C t T=A\cap B_t\in C_t T = A ∩ B t ∈ C t , which proves the lemma.
We now construct the extension in Equation (16) . Since ω t B t ( y , x ) \omega_t^{B_t}(y,x) ω t B t ( y , x ) replaces every y v y_v y v with v ∈ V ( G t ) ∖ B t v\in V(G_t)\setminus B_t v ∈ V ( G t ) ∖ B t by x x x , we have
[ x b w ℓ ∏ v ∈ B t y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = ∑ [ w ℓ ∏ v ∈ V ( G t ) y v a v ] log Z t ≤ 2 k ( y , w ) , \left[x^b w^\ell\prod_{v\in B_t}y_v^{a_v}\right]\log Z_t^{\le2k}(\omega_t^{B_t}(y,x),w)=\sum\left[w^\ell\prod_{v\in V(G_t)}y_v^{a_v}\right]\log Z_t^{\le2k}(y,w), [ x b w ℓ v ∈ B t ∏ y v a v ] log Z t ≤ 2 k ( ω t B t ( y , x ) , w ) = ∑ w ℓ v ∈ V ( G t ) ∏ y v a v log Z t ≤ 2 k ( y , w ) ,
where the sum is taken over all choices of a v ∈ Z ≥ 0 a_v\in\mathbb{Z}_{\ge0} a v ∈ Z ≥ 0 for v ∈ V ( G t ) ∖ B t v\in V(G_t)\setminus B_t v ∈ V ( G t ) ∖ B t satisfying ∑ v ∈ V ( G t ) ∖ B t a v = b \sum_{v\in V(G_t)\setminus B_t}a_v=b ∑ v ∈ V ( G t ) ∖ B t a v = b . The left-hand side is nonzero by the hypothesis, so at least one summand on the right-hand side is nonzero. Fix an extension ( a v ) v ∈ V ( G t ) (a_v)_{v\in V(G_t)} ( a v ) v ∈ V ( G t ) indexing a nonzero summand. Then,
[ w ℓ ∏ v ∈ V ( G t ) y v a v ] log Z t ≤ 2 k ( y , w ) ≠ 0. \left[w^\ell\prod_{v\in V(G_t)}y_v^{a_v}\right]\log Z_t^{\le2k}(y,w)\ne0. w ℓ v ∈ V ( G t ) ∏ y v a v log Z t ≤ 2 k ( y , w ) = 0.
Set S : = { v ∈ V ( G t ) : a v > 0 } S:=\{v\in V(G_t):a_v>0\} S := { v ∈ V ( G t ) : a v > 0 } . Define the substitution π t S ( y ) : = ( π t , v S ) v ∈ V ( G t ) \pi_t^S(y):=(\pi_{t,v}^S)_{v\in V(G_t)} π t S ( y ) := ( π t , v S ) v ∈ V ( G t ) , where, for each v ∈ V ( G t ) v\in V(G_t) v ∈ V ( G t ) ,
π t , v S : = { y v , v ∈ S , 0 , v ∈ V ( G t ) ∖ S . \pi_{t,v}^S:=\begin{cases}
y_v,&v\in S,\\
0,&v\in V(G_t)\setminus S.
\end{cases} π t , v S := { y v , 0 , v ∈ S , v ∈ V ( G t ) ∖ S .
By the definition of S S S , the monomial in Equation (16) has support S S S . The substitution π t S \pi_t^S π t S fixes y v y_v y v for every v ∈ S v\in S v ∈ S and sends only the variables outside S S S to zero, so it preserves the coefficient of this monomial in log Z t ≤ 2 k ( y , w ) \log Z_t^{\le2k}(y,w) log Z t ≤ 2 k ( y , w ) . Thus,
[ w ℓ ∏ v ∈ V ( G t ) y v a v ] log Z t ≤ 2 k ( π t S ( y ) , w ) ≠ 0. \left[w^\ell\prod_{v\in V(G_t)}y_v^{a_v}\right]\log Z_t^{\le2k}(\pi_t^S(y),w)\ne0. w ℓ v ∈ V ( G t ) ∏ y v a v log Z t ≤ 2 k ( π t S ( y ) , w ) = 0.
Every term of Z t ( π t S ( y ) , w ) Z_t(\pi_t^S(y),w) Z t ( π t S ( y ) , w ) is indexed by a set A t ⊆ S A_t\subseteq S A t ⊆ S and hence satisfies ∣ A t ∩ B t ∣ ≤ ∣ S ∩ B t ∣ = ∣ T ∣ ≤ 2 k \lvert A_t\cap B_t\rvert\le\lvert S\cap B_t\rvert=\lvert T\rvert\le2k ∣ A t ∩ B t ∣ ≤ ∣ S ∩ B t ∣ = ∣ T ∣ ≤ 2 k . Therefore,
Z t ≤ 2 k ( π t S ( y ) , w ) = Z t ( π t S ( y ) , w ) . Z_t^{\le2k}(\pi_t^S(y),w)=Z_t(\pi_t^S(y),w). Z t ≤ 2 k ( π t S ( y ) , w ) = Z t ( π t S ( y ) , w ) .
Taking the formal logarithm of both sides gives
log Z t ≤ 2 k ( π t S ( y ) , w ) = log Z t ( π t S ( y ) , w ) . \log Z_t^{\le2k}(\pi_t^S(y),w)=\log Z_t(\pi_t^S(y),w). log Z t ≤ 2 k ( π t S ( y ) , w ) = log Z t ( π t S ( y ) , w ) .
The substitution π t S \pi_t^S π t S likewise preserves the coefficient of the monomial in Equation (16) in log Z t ( y , w ) \log Z_t(\mathbf{y},w) log Z t ( y , w ) . Hence,
[ w ℓ ∏ v ∈ V ( G t ) y v a v ] log Z t ( y , w ) = [ w ℓ ∏ v ∈ V ( G t ) y v a v ] log Z t ≤ 2 k ( y , w ) ≠ 0. \left[w^\ell\prod_{v\in V(G_t)}y_v^{a_v}\right]\log Z_t(\mathbf{y},w)=\left[w^\ell\prod_{v\in V(G_t)}y_v^{a_v}\right]\log Z_t^{\le2k}(\mathbf{y},w)\ne0. w ℓ v ∈ V ( G t ) ∏ y v a v log Z t ( y , w ) = w ℓ v ∈ V ( G t ) ∏ y v a v log Z t ≤ 2 k ( y , w ) = 0.
Matching Cut
In this section, we consider Matching Cut. We use color coding to reduce Matching Cut to an auxiliary problem called Colorful Cut, and solve Colorful Cut using almost the same algorithm as for Minimum Bisection in § 3 . Without loss of generality, assume throughout the section that G G G is connected.
For an integer q ≥ 1 q\ge1 q ≥ 1 , a coloring h : V ( G ) → [ 2 q ] h:V(G)\to[2q] h : V ( G ) → [ 2 q ] , and a cut A ⊆ V ( G ) A\subseteq V(G) A ⊆ V ( G ) , let χ h ( A ) : = { h ( v ) : v ∈ V ( G ) , δ G ( A ) ∩ δ G ( { v } ) ≠ ∅ } \chi_h(A):=\{h(v):v\in V(G),\ \delta_G(A)\cap\delta_G(\{v\})\ne\varnothing\} χ h ( A ) := { h ( v ) : v ∈ V ( G ) , δ G ( A ) ∩ δ G ({ v }) = ∅ } be the set of colors that appear at endpoints of edges in δ G ( A ) \delta_G(A) δ G ( A ) .
Colorful Cut
Input: A simple undirected graph G G G , a positive integer q q q , and a coloring h : V ( G ) → [ 2 q ] h:V(G)\to[2q] h : V ( G ) → [ 2 q ] .
Question: Is there a set ∅ ⊊ A ⊊ V ( G ) \varnothing\subsetneq A\subsetneq V(G) ∅ ⊊ A ⊊ V ( G ) such that ∣ δ G ( A ) ∣ = q |\delta_G(A)|=q ∣ δ G ( A ) ∣ = q and χ h ( A ) = [ 2 q ] \chi_h(A)=[2q] χ h ( A ) = [ 2 q ] ?
Parameter: q q q .
The next lemma reduces Matching Cut to solving O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) instances of Colorful Cut.
Lemma 4.1. A Matching Cut instance ( G , k ) (G,k) ( G , k ) with n ≥ 2 n\ge2 n ≥ 2 and k ≥ 1 k\ge1 k ≥ 1 can be solved by solving O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) instances of Colorful Cut.
Proof. For every q ∈ [ min { k , ⌊ n / 2 ⌋ } ] q\in[\min\{k,\lfloor n/2\rfloor\}] q ∈ [ min { k , ⌊ n /2 ⌋}] , construct as above an ( n , 2 q ) (n,2q) ( n , 2 q ) -perfect hash family H q \mathcal{H}_q H q of colorings h : V ( G ) → [ 2 q ] h:V(G)\to[2q] h : V ( G ) → [ 2 q ] , and add ( G , q , h ) (G,q,h) ( G , q , h ) to I \mathcal{I} I for every h ∈ H q h\in\mathcal{H}_q h ∈ H q . The construction time and the number of instances are O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) .
Suppose that G G G admits a matching cut A A A with ∣ δ G ( A ) ∣ = q ≤ k |\delta_G(A)|=q\le k ∣ δ G ( A ) ∣ = q ≤ k . Since G G G is connected, q ≥ 1 q\ge1 q ≥ 1 . Let X X X be the set of endpoints of δ G ( A ) \delta_G(A) δ G ( A ) . Then ∣ X ∣ = 2 q ≤ n |X|=2q\le n ∣ X ∣ = 2 q ≤ n , so some h ∈ H q h\in\mathcal{H}_q h ∈ H q is injective on X X X . The equality h ( X ) = [ 2 q ] h(X)=[2q] h ( X ) = [ 2 q ] shows that A A A is feasible for ( G , q , h ) (G,q,h) ( G , q , h ) .
Conversely, take a Colorful Cut instance ( G , q , h ) ∈ I (G,q,h)\in\mathcal{I} ( G , q , h ) ∈ I , and suppose that A A A is a feasible solution for ( G , q , h ) (G,q,h) ( G , q , h ) . Let X X X be the set of endpoints of δ G ( A ) \delta_G(A) δ G ( A ) . Since ∣ δ G ( A ) ∣ = q |\delta_G(A)|=q ∣ δ G ( A ) ∣ = q , we have ∣ X ∣ ≤ 2 q |X|\le2q ∣ X ∣ ≤ 2 q . Since χ h ( A ) = h ( X ) = [ 2 q ] \chi_h(A)=h(X)=[2q] χ h ( A ) = h ( X ) = [ 2 q ] , the set X X X contains a vertex of each of the 2 q 2q 2 q colors, so ∣ X ∣ ≥ 2 q |X|\ge2q ∣ X ∣ ≥ 2 q . Therefore ∣ X ∣ = 2 q |X|=2q ∣ X ∣ = 2 q , and no two edges of δ G ( A ) \delta_G(A) δ G ( A ) share an endpoint. Thus δ G ( A ) \delta_G(A) δ G ( A ) is a matching and ∣ δ G ( A ) ∣ = q ≤ k |\delta_G(A)|=q\le k ∣ δ G ( A ) ∣ = q ≤ k .
We now solve Colorful Cut by slightly modifying the algorithm for Minimum Bisection in § 3 . Intuitively, it suffices to record in each dynamic-programming state the set of colors that occur at the endpoints of cut edges. To apply the algorithm of § 3 , we express χ h ( A ) = [ 2 q ] \chi_h(A)=[2q] χ h ( A ) = [ 2 q ] as a condition on the polynomial. Fix an instance ( G , q , h ) (G,q,h) ( G , q , h ) , and introduce one variable u i u_i u i for each color i ∈ [ 2 q ] i\in[2q] i ∈ [ 2 q ] . Write u = ( u 1 , … , u 2 q ) \mathbf{u}=(u_1,\ldots,u_{2q}) u = ( u 1 , … , u 2 q ) . For every node t t t , define the polynomial over Q \mathbb{Q} Q by
Z t ( y , w , u ) : = ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ q ( ∏ u v ∈ δ G t ( A t ) u h ( u ) u h ( v ) ) ⋅ w μ t ( A t ) ⋅ ∏ v ∈ A t y v . Z_t(\mathbf{y},w,\mathbf{u}):=\sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\le q}}\left(\prod_{uv\in\delta_{G_t}(A_t)}u_{h(u)}u_{h(v)}\right)\cdot w^{\mu_t(A_t)}\cdot\prod_{v\in A_t}y_v. Z t ( y , w , u ) := A t ⊆ V ( G t ) μ t ( A t ) ≤ q ∑ uv ∈ δ G t ( A t ) ∏ u h ( u ) u h ( v ) ⋅ w μ t ( A t ) ⋅ v ∈ A t ∏ y v .
To compute the transition, the only nontrivial change from § 3 is the definition of Q t , R Q_{t,R} Q t , R .
Use the tree decomposition T \mathcal{T} T and bottom-up computation of § 3 with q q q in place of k k k . Define the contribution of the sets A t A_t A t satisfying A t ∩ B t = R A_t \cap B_t = R A t ∩ B t = R by
Q t , R ( x , w , u ) : = ∑ A t ⊆ V ( G t ) μ t ( A t ) ≤ q , A t ∩ B t = R ( ( ∏ u v ∈ δ G t ( A t ) u h ( u ) u h ( v ) ) ⋅ x ∣ A t ∖ B t ∣ ⋅ w μ t ( A t ) ) . Q_{t,R}(x,w,\mathbf{u}) := \sum_{\substack{A_t\subseteq V(G_t)\\ \mu_t(A_t)\leq q,\ A_t\cap B_t=R}} \left(\left(\prod_{uv\in\delta_{G_t}(A_t)}u_{h(u)}u_{h(v)}\right)\cdot x^{\lvert A_t\setminus B_t\rvert}\cdot w^{\mu_t(A_t)}\right). Q t , R ( x , w , u ) := A t ⊆ V ( G t ) μ t ( A t ) ≤ q , A t ∩ B t = R ∑ uv ∈ δ G t ( A t ) ∏ u h ( u ) u h ( v ) ⋅ x ∣ A t ∖ B t ∣ ⋅ w μ t ( A t ) .
By Equation (7) , the monomial of either Z t ( y , w , u ) Z_t(y,w,\mathbf{u}) Z t ( y , w , u ) or Q t , R ( x , w , u ) Q_{t,R}(x,w,\mathbf{u}) Q t , R ( x , w , u ) indexed by A t A_t A t has total u \mathbf{u} u -degree 2 ∣ δ G t ( A t ) ∣ 2\lvert\delta_{G_t}(A_t)\rvert 2 ∣ δ G t ( A t )∣ and w w w -degree μ t ( A t ) \mu_t(A_t) μ t ( A t ) . Consequently, every monomial m m m in Z t ( y , w , u ) Z_t(y,w,\mathbf{u}) Z t ( y , w , u ) or Q t , R ( x , w , u ) Q_{t,R}(x,w,\mathbf{u}) Q t , R ( x , w , u ) satisfies ∑ i = 1 2 q deg u i ( m ) ≤ 2 deg w ( m ) \sum_{i=1}^{2q}\deg_{u_i}(m)\leq2\deg_w(m) ∑ i = 1 2 q deg u i ( m ) ≤ 2 deg w ( m ) . Hence the w w w -degree truncation at q q q also bounds the total u \mathbf{u} u -degree by 2 q 2q 2 q throughout the algorithm, even after taking the formal logarithm or exponential. The number of monomials in u \mathbf{u} u of total degree at most 2 q 2q 2 q is ( 4 q 2 q ) = 2 O ( q ) \binom{4q}{2q}=2^{O(q)} ( 2 q 4 q ) = 2 O ( q ) . Thus, managing the variables u \mathbf{u} u adds only a 2 O ( q ) 2^{O(q)} 2 O ( q ) factor to the total running time.
For R = A t ∩ B t R=A_t\cap B_t R = A t ∩ B t and A s = A t ∩ V ( G s ) A_s=A_t\cap V(G_s) A s = A t ∩ V ( G s ) , we have
δ G t ( A t ) = δ G t [ B t ] ( R ) ∪ ˙ ⋃ s ∈ child ( t ) δ G s ( A s ) . \delta_{G_t}(A_t)=\delta_{G_t[B_t]}(R)\mathbin{\dot\cup}\bigcup_{s\in\operatorname{child}(t)}\delta_{G_s}(A_s). δ G t ( A t ) = δ G t [ B t ] ( R ) ∪ ˙ s ∈ child ( t ) ⋃ δ G s ( A s ) .
Therefore, Equation (8) becomes
Q t , R ( x , w , u ) = ⟨ ( ∏ u v ∈ δ G t [ B t ] ( R ) u h ( u ) u h ( v ) ) ⋅ w ∣ δ G t [ B t ] ( R ) ∣ + η t ( R ∩ ∂ G t ) ∏ s ∈ child ( t ) ( w − η s ( R ∩ ∂ G s ) ⋅ [ ∏ v ∈ R ∩ ∂ G s y v ] Z s ( ω ∂ G s ( y , x ) , w , u ) ) ⟩ deg w ≤ q (17) Q_{t,R}(x,w,\mathbf{u})=
\left\langle
\left(\prod_{uv\in\delta_{G_t[B_t]}(R)}u_{h(u)}u_{h(v)}\right)\cdot w^{\lvert\delta_{G_t[B_t]}(R)\rvert+\eta_t(R\cap\partial G_t)}
\prod_{s\in\operatorname{child}(t)}
\left(
w^{-\eta_s(R\cap\partial G_s)}
\cdot
\left[\prod_{v\in R\cap\partial G_s}y_v\right]
Z_s\left(\omega^{\partial G_s}(y,x),w,\mathbf{u}\right)
\right)
\right\rangle_{\deg_w\leq q} \tag*{(17)} Q t , R ( x , w , u ) = ⟨ uv ∈ δ G t [ B t ] ( R ) ∏ u h ( u ) u h ( v ) ⋅ w ∣ δ G t [ B t ] ( R )∣ + η t ( R ∩ ∂ G t ) s ∈ child ( t ) ∏ ( w − η s ( R ∩ ∂ G s ) ⋅ [ v ∈ R ∩ ∂ G s ∏ y v ] Z s ( ω ∂ G s ( y , x ) , w , u ) ) ⟩ d e g w ≤ q (17)
Apart from Equation (17) , use the algorithm of § 3 without change, with Z t ( y , w , u ) Z_t(y,w,\mathbf{u}) Z t ( y , w , u ) and Q t , R ( x , w , u ) Q_{t,R}(x,w,\mathbf{u}) Q t , R ( x , w , u ) in place of Z t ( y , w ) Z_t(y,w) Z t ( y , w ) and Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) . Accept exactly when
[ x b w q ∏ i = 1 2 q u i ] Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w , u ) > 0 (18) \left[x^b w^q\prod_{i=1}^{2q}u_i\right]Z_{t_0}\left(\omega_{t_0}^{\partial G_{t_0}}(y,x),w,\mathbf{u}\right)>0 \tag*{(18)} [ x b w q i = 1 ∏ 2 q u i ] Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w , u ) > 0 (18)
for some 1 ≤ b ≤ n − 1 1\leq b\leq n-1 1 ≤ b ≤ n − 1 .
The next lemma proves the correctness and running time of the Colorful Cut algorithm.
Lemma 4.2. Colorful Cut can be decided deterministically in O ∗ ( 2 O ( q ) ) O^*(2^{O(q)}) O ∗ ( 2 O ( q ) ) time.
Proof. Let Z t ≤ 2 q ( y , w , u ) Z_t^{\leq2q}(y,w,\mathbf{u}) Z t ≤ 2 q ( y , w , u ) denote the restriction of Z t ( y , w , u ) Z_t(y,w,\mathbf{u}) Z t ( y , w , u ) to the terms indexed by sets A t A_t A t with ∣ A t ∩ B t ∣ ≤ 2 q \lvert A_t\cap B_t\rvert\leq2q ∣ A t ∩ B t ∣ ≤ 2 q . For S ⊆ B t S\subseteq B_t S ⊆ B t , let L t , S ( y , w , u ) L_{t,S}(y,w,\mathbf{u}) L t , S ( y , w , u ) consist of the terms in log Z t ≤ 2 q ( y , w , u ) \log Z_t^{\leq2q}(y,w,\mathbf{u}) log Z t ≤ 2 q ( y , w , u ) whose B t B_t B t -support is S S S .
In Lemma 3.1, replace Q t , R ( x , w ) Q_{t,R}(x,w) Q t , R ( x , w ) by Q t , R ( x , w , u ) Q_{t,R}(x,w,\mathbf{u}) Q t , R ( x , w , u ) , adjoin u \mathbf{u} u as the last argument of every child polynomial Z s Z_s Z s , and replace Equation (8) by Equation (17) . For R = A t ∩ B t R=A_t\cap B_t R = A t ∩ B t and A s = A t ∩ V ( G s ) A_s=A_t\cap V(G_s) A s = A t ∩ V ( G s ) , the edge partition used in Lemma 3.1 gives
( ∏ u v ∈ δ G t [ B t ] ( R ) u h ( u ) u h ( v ) ) ∏ s ∈ child ( t ) ( ∏ u v ∈ δ G s ( A s ) u h ( u ) u h ( v ) ) = ∏ u v ∈ δ G t ( A t ) u h ( u ) u h ( v ) . \left(\prod_{uv\in\delta_{G_t[B_t]}(R)}u_{h(u)}u_{h(v)}\right)
\prod_{s\in\operatorname{child}(t)}
\left(\prod_{uv\in\delta_{G_s}(A_s)}u_{h(u)}u_{h(v)}\right)
=
\prod_{uv\in\delta_{G_t}(A_t)}u_{h(u)}u_{h(v)}. uv ∈ δ G t [ B t ] ( R ) ∏ u h ( u ) u h ( v ) s ∈ child ( t ) ∏ uv ∈ δ G s ( A s ) ∏ u h ( u ) u h ( v ) = uv ∈ δ G t ( A t ) ∏ u h ( u ) u h ( v ) . All factors independent of u \mathbf{u} u are identical in Equations (8) and (17). Hence the proof of Lemma 3.1 requires no other change.
In Equation (5) , adjoin u u u as the last argument of Q t , R Q_{t,R} Q t , R and Z t Z_t Z t . The derivation of Equation (5) partitions the terms according to A t ∩ B t = R A_t \cap B_t = R A t ∩ B t = R , so Equation (5) holds coefficientwise in u u u .
In Lemma 3.2 , replace Z t ( y , w ) Z_t(y,w) Z t ( y , w ) and Z t ≤ 2 q ( y , w ) Z_t^{\leq2q}(y,w) Z t ≤ 2 q ( y , w ) by Z t ( y , w , u ) Z_t(y,w,u) Z t ( y , w , u ) and Z t ≤ 2 q ( y , w , u ) Z_t^{\leq2q}(y,w,u) Z t ≤ 2 q ( y , w , u ) , respectively. Since δ G t ( V ( G t ) ∖ A t ) = δ G t ( A t ) \delta_{G_t}(V(G_t) \setminus A_t) = \delta_{G_t}(A_t) δ G t ( V ( G t ) ∖ A t ) = δ G t ( A t ) , the proof of Lemma 3.2 requires no other change.
In Lemma 3.3, replace Z t ≤ 2 q ( y , w ) Z_t^{\leq2q}(y,w) Z t ≤ 2 q ( y , w ) by Z t ≤ 2 q ( y , w , u ) Z_t^{\leq2q}(y,w,u) Z t ≤ 2 q ( y , w , u ) and the coefficient ring Q [ x , w ] \mathbb{Q}[x,w] Q [ x , w ] by Q [ x , w , u ] \mathbb{Q}[x,w,u] Q [ x , w , u ] . The proof of Lemma 3.3 is unchanged: Equation (13) holds over Q [ x , w , u ] \mathbb{Q}[x,w,u] Q [ x , w , u ] , and the B t B_t B t -support depends only on the exponents of the variables y v y_v y v .
In Lemma 3.4 , adjoin u u u as the last argument of L t , S L_{t,S} L t , S , Q t , R Q_{t,R} Q t , R , and Z t ≤ 2 q Z_t^{\leq2q} Z t ≤ 2 q . The subset inversion in Equation (11) holds coefficientwise for every monomial in u u u . Therefore, Lemma 3.4 holds without modification.
In Lemma 3.5 , replace Z t ≤ 2 q ( y , w ) Z_t^{\leq2q}(y,w) Z t ≤ 2 q ( y , w ) by Z t ≤ 2 q ( y , w , u ) Z_t^{\leq2q}(y,w,u) Z t ≤ 2 q ( y , w , u ) . The substitutions σ t ( y , x , x ′ ) \sigma_t(y,x,x') σ t ( y , x , x ′ ) and x ′ ↦ x x' \mapsto x x ′ ↦ x fix u u u , and Equations (2) and (3) remain valid over Q [ u ] \mathbb{Q}[u] Q [ u ] . Therefore, Lemma 3.5 holds without modification.
Finally, in Lemma 3.6 , adjoin u u u as the last argument of every polynomial Z t Z_t Z t . Applying Lemmas 3.1 to 3.5 with Q t , R ( x , w , u ) Q_{t,R}(x,w,u) Q t , R ( x , w , u ) , Z t ( y , w , u ) Z_t(y,w,u) Z t ( y , w , u ) , and Z t ≤ 2 q ( y , w , u ) Z_t^{\leq2q}(y,w,u) Z t ≤ 2 q ( y , w , u ) proves the small-bag and large-bag cases of Lemma 3.6 . Thus, the bottom-up computation obtains Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w , u ) Z_{t_0}(\omega_{t_0}^{\partial G_{t_0}}(y,x),w,u) Z t 0 ( ω t 0 ∂ G t 0 ( y , x ) , w , u ) .
At the root, a cut A A A with ∣ δ G ( A ) ∣ = q |\delta_G(A)| = q ∣ δ G ( A ) ∣ = q contributes a monomial of total u u u -degree 2 q 2q 2 q . The equality χ h ( A ) = [ 2 q ] \chi_h(A) = [2q] χ h ( A ) = [ 2 q ] holds if and only if every variable u i u_i u i for i ∈ [ 2 q ] i \in[2q] i ∈ [ 2 q ] occurs in the monomial. Hence every such variable has exponent one. Therefore, Equation (18) accepts exactly the feasible cuts of Colorful Cut.
Throughout the computation, the total u u u -degree is at most 2 q 2q 2 q . At most ( 4 q 2 q ) = 2 O ( q ) \binom{4q}{2q} = 2^{O(q)} ( 2 q 4 q ) = 2 O ( q ) monomials in u u u occur. Thus, each operation from § 3 acquires only a 2 O ( q ) 2^{O(q)} 2 O ( q ) factor, and the total running time remains O ∗ ( 2 O ( q ) ) O^*(2^{O(q)}) O ∗ ( 2 O ( q ) ) .
Now we prove Theorem 1.3.
Theorem 1.3. Matching Cut can be solved deterministically, and a witness can be recovered, in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time.
Proof. Construct the family I \mathcal{I} I from Lemma 4.1 , and apply Lemma 4.2 to every instance in I \mathcal{I} I . By Lemma 4.1 , some instance in I \mathcal{I} I is a yes-instance if and only if ( G , k ) (G,k) ( G , k ) is a yes-instance of Matching Cut. For a yes-instance ( G , q , h ) ∈ I (G,q,h) \in\mathcal{I} ( G , q , h ) ∈ I , vertex self-reduction using the decision algorithm of Lemma 4.2 recovers a feasible cut with at most n n n additional calls. By Lemma 4.1 , the feasible cut defines the required matching cut. The family I \mathcal{I} I has O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) instances. Since q ≤ k q \leq k q ≤ k for every ( G , q , h ) ∈ I (G,q,h) \in\mathcal{I} ( G , q , h ) ∈ I , each instance takes O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time. Therefore, the total running time is O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) .
Unbreakable decomposition
In this section, we prove Theorem 1.1, completing the decomposition construction used by the algorithms in § 3 and § 4 .
For a graph H H H and U ⊆ V ( H ) U \subseteq V(H) U ⊆ V ( H ) , let N H ( U ) : = { v ∈ V ( H ) ∖ U : u v ∈ E ( H ) for some u ∈ U } N_H(U) := \{v \in V(H) \setminus U : uv \in E(H) \text{ for some } u \in U\} N H ( U ) := { v ∈ V ( H ) ∖ U : uv ∈ E ( H ) for some u ∈ U } denote the open neighbourhood of U U U in H H H . For T ⊆ V ( H ) T \subseteq V(H) T ⊆ V ( H ) , let adh H ( T ) \operatorname{adh}_H(T) adh H ( T ) be the maximum of ∣ N H ( C ) ∣ |N_H(C)| ∣ N H ( C ) ∣ over the components C C C of H − T H-T H − T , or zero if H − T H-T H − T is empty.
We construct the tree decomposition from the root down. If V ( G ) V(G) V ( G ) is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in G G G , we output the one-node tree decomposition with bag V ( G ) V(G) V ( G ) . Otherwise, we find a set B ⊆ V ( G ) B \subseteq V(G) B ⊆ V ( G ) that is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in G G G and satisfies adh G ( B ) ≤ 2 k \operatorname{adh}_G(B) \leq2k adh G ( B ) ≤ 2 k , use B B B as the root bag, and recurse for every connected component of G − B G-B G − B .
Consider a recursive call on H C : = H [ C ∪ N H ( C ) ] H_C := H[C \cup N_H(C)] H C := H [ C ∪ N H ( C )] , where C C C is a component created by the parent call. We terminate if V ( H C ) V(H_C) V ( H C ) is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in H C H_C H C . Otherwise, let X C : = N H ( C ) ∪ { v C } X_C := N_H(C) \cup\{v_C\} X C := N H ( C ) ∪ { v C } , where v C v_C v C is an arbitrary vertex of C = V ( H C ) ∖ N H ( C ) C = V(H_C) \setminus N_H(C) C = V ( H C ) ∖ N H ( C ) . We construct a bag B C B_C B C with X C ⊆ B C X_C \subseteq B_C X C ⊆ B C and recurse for every connected component of H C − B C H_C - B_C H C − B C . We require N H ( C ) ⊆ B C N_H(C) \subseteq B_C N H ( C ) ⊆ B C because every vertex of N H ( C ) N_H(C) N H ( C ) occurs in the parent bag, and the bags containing any fixed vertex must form a connected subtree. We require v C ∈ B C v_C \in B_C v C ∈ B C to rule out B C = N H ( C ) B_C = N_H(C) B C = N H ( C ) , which would repeat the same recursive call.
To find such a bag B C B_C B C in single-exponential time, we repeatedly use the next lemma.
Lemma 5.1. Let X ⊆ B ⊆ V ( H ) X \subseteq B \subseteq V(H) X ⊆ B ⊆ V ( H ) satisfy ∣ X ∣ ≤ 2 k + 1 |X| \leq2k+1 ∣ X ∣ ≤ 2 k + 1 and adh H ( B ) ≤ 2 k \operatorname{adh}_H(B) \leq2k adh H ( B ) ≤ 2 k . In O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time one can either find B ′ ⊆ V ( H ) B' \subseteq V(H) B ′ ⊆ V ( H ) such that X ⊆ B ′ X \subseteq B' X ⊆ B ′ , ∣ B ′ ∣ < ∣ B ∣ |B'| < |B| ∣ B ′ ∣ < ∣ B ∣ , and adh H ( B ′ ) ≤ 2 k \operatorname{adh}_H(B') \leq2k adh H ( B ′ ) ≤ 2 k , or certify that B B B is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable.
Proof. If ∣ B ∣ ≤ 4 k + 1 |B| \leq4k+1 ∣ B ∣ ≤ 4 k + 1 , every bipartition of B B B has a side of size at most 2 k 2k 2 k , so B B B is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable. Hence assume ∣ B ∣ > 4 k + 1 |B| > 4k+1 ∣ B ∣ > 4 k + 1 . Since ∣ X ∣ ≤ 2 k + 1 |X| \leq2k+1 ∣ X ∣ ≤ 2 k + 1 , we have B ∖ X ≠ ∅ B \setminus X \neq\varnothing B ∖ X = ∅ .
Guess a set Z ⊆ X Z \subseteq X Z ⊆ X with ∣ X ∖ Z ∣ ≤ k |X \setminus Z| \leq k ∣ X ∖ Z ∣ ≤ k and a vertex r ∈ B ∖ X r \in B \setminus X r ∈ B ∖ X . There are O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) choices for ( Z , r ) (Z,r) ( Z , r ) . For each guess, first delete X ∖ Z X \setminus Z X ∖ Z from H H H and set H ′ : = H − ( X ∖ Z ) H' := H - (X \setminus Z) H ′ := H − ( X ∖ Z ) . Then enumerate in H ′ H' H ′ the important { r } \{r\} { r } –Z Z Z cuts D D D with ∣ δ H ′ ( D ) ∣ ≤ k |\delta_{H'}(D)| \leq k ∣ δ H ′ ( D ) ∣ ≤ k using Lemma 2.1 .
The key point is that if B B B is not ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable, then the enumeration finds a cut D D D containing more vertices of B B B than the number of edges leaving D D D in H ′ H' H ′ , which helps us construct a smaller separator.
Claim 5.2. If B B B is not ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable, then for some guess ( Z , r ) (Z,r) ( Z , r ) , the enumeration finds a cut D D D satisfying ∣ D ∩ B ∣ > ∣ δ H ′ ( D ) ∣ |D \cap B| > |\delta_{H'}(D)| ∣ D ∩ B ∣ > ∣ δ H ′ ( D ) ∣ .
Proof of Claim 5.2. We construct a guess ( Z , r ) (Z,r) ( Z , r ) for which the enumeration finds D D D . Since B B B is not ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in H H H , there is a set A ⊆ V ( H ) A \subseteq V(H) A ⊆ V ( H ) such that ∣ δ H ( A ) ∣ ≤ k |\delta_H(A)| \leq k ∣ δ H ( A ) ∣ ≤ k , ∣ A ∩ B ∣ > 2 k |A \cap B| > 2k ∣ A ∩ B ∣ > 2 k , and ∣ B ∖ A ∣ > 2 k |B \setminus A| > 2k ∣ B ∖ A ∣ > 2 k . Since ∣ X ∣ ≤ 2 k + 1 |X| \leq2k+1 ∣ X ∣ ≤ 2 k + 1 , we may assume without loss of generality that ∣ A ∩ X ∣ ≤ k |A \cap X| \leq k ∣ A ∩ X ∣ ≤ k . Set Z : = X ∖ A Z := X \setminus A Z := X ∖ A . See Figure 4 for an illustration.
Figure 4. The choice of K K K and D D D . Here K K K is a component of H [ A ∖ X ] H[A \setminus X] H [ A ∖ X ] , and D ⊇ K D \supseteq K D ⊇ K is an important cut.
Let K \mathcal{K} K be the family of connected components of H [ A ∖ X ] H[A \setminus X] H [ A ∖ X ] . Observe that A ∖ X A \setminus X A ∖ X is the part of A A A that remains in H ′ H' H ′ . Since the sets in K \mathcal{K} K partition A ∖ X A \setminus X A ∖ X ,
∑ K ∈ K ∣ K ∩ B ∣ = ∣ ( A ∖ X ) ∩ B ∣ ≥ ∣ A ∩ B ∣ − ∣ A ∩ X ∣ > k . \sum_{K \in\mathcal{K}} |K \cap B| = |(A \setminus X) \cap B| \geq|A \cap B| - |A \cap X| > k. K ∈ K ∑ ∣ K ∩ B ∣ = ∣ ( A ∖ X ) ∩ B ∣ ≥ ∣ A ∩ B ∣ − ∣ A ∩ X ∣ > k .
For every K ∈ K K \in\mathcal{K} K ∈ K , we have δ H ′ ( K ) ⊆ δ H ( A ) \delta_{H'}(K) \subseteq\delta_H(A) δ H ′ ( K ) ⊆ δ H ( A ) . Moreover, the sets δ H ′ ( K ) \delta_{H'}(K) δ H ′ ( K ) for K ∈ K K \in\mathcal{K} K ∈ K are pairwise disjoint. Hence
∑ K ∈ K ∣ δ H ′ ( K ) ∣ ≤ ∣ δ H ( A ) ∣ ≤ k . \sum_{K \in\mathcal{K}} |\delta_{H'}(K)| \leq|\delta_H(A)| \leq k. K ∈ K ∑ ∣ δ H ′ ( K ) ∣ ≤ ∣ δ H ( A ) ∣ ≤ k .
Therefore some K ∈ K K \in\mathcal{K} K ∈ K satisfies ∣ K ∩ B ∣ > ∣ δ H ′ ( K ) ∣ |K \cap B| > |\delta_{H'}(K)| ∣ K ∩ B ∣ > ∣ δ H ′ ( K ) ∣ .
Fix such a component K K K and choose r ∈ K ∩ B r \in K \cap B r ∈ K ∩ B . Since K ⊆ A ∖ X K \subseteq A \setminus X K ⊆ A ∖ X , we have r ∈ B ∖ X r \in B \setminus X r ∈ B ∖ X , so ( Z , r ) (Z,r) ( Z , r ) is among the guesses. Also, K ∩ Z = ∅ K \cap Z = \varnothing K ∩ Z = ∅ and H ′ [ K ] H'[K] H ′ [ K ] is connected, so K K K is a connected { r } \{r\} { r } –Z Z Z cut in H ′ H' H ′ . Choose an important connected { r } \{r\} { r } –Z Z Z cut D D D such that K ⊆ D K \subseteq D K ⊆ D (possibly D = K D = K D = K ). Then, ∣ δ H ′ ( D ) ∣ ≤ ∣ δ H ′ ( K ) ∣ ≤ k |\delta_{H'}(D)| \leq|\delta_{H'}(K)| \leq k ∣ δ H ′ ( D ) ∣ ≤ ∣ δ H ′ ( K ) ∣ ≤ k . By Lemma 2.1 , the enumeration contains D D D . Finally, ∣ D ∩ B ∣ ≥ ∣ K ∩ B ∣ > ∣ δ H ′ ( K ) ∣ ≥ ∣ δ H ′ ( D ) ∣ |D \cap B| \geq|K \cap B| > |\delta_{H'}(K)| \geq|\delta_{H'}(D)| ∣ D ∩ B ∣ ≥ ∣ K ∩ B ∣ > ∣ δ H ′ ( K ) ∣ ≥ ∣ δ H ′ ( D ) ∣ .
For a cut D D D guaranteed by Claim 5.2, construct B ′ B' B ′ as follows. Let Y Y Y be the set of vertices of D D D incident with an edge of δ H ′ ( D ) \delta_{H'}(D) δ H ′ ( D ) , and let P ⊆ D P \subseteq D P ⊆ D be a minimum Y Y Y –( D ∩ B ) (D \cap B) ( D ∩ B ) vertex separator in H [ D ] H[D] H [ D ] . Set B ′ : = ( B ∖ D ) ∪ P B' := (B \setminus D) \cup P B ′ := ( B ∖ D ) ∪ P . The sets B B B , D D D , Y Y Y , P P P , and B ′ B' B ′ are illustrated in Figure 5 .
Figure 5. The construction of B ′ = ( B \ D ) ∪ P B' = (B \backslash D) \cup P B ′ = ( B \ D ) ∪ P , where P P P is a minimum Y Y Y –( D ∩ B ) (D \cap B) ( D ∩ B ) vertex separator in H [ D ] H[D] H [ D ] .
The set D D D is disjoint from X X X , so X ⊆ B ∖ D ⊆ B ′ X \subseteq B \setminus D \subseteq B' X ⊆ B ∖ D ⊆ B ′ . Both Y Y Y and P P P are Y Y Y –( D ∩ B ) (D \cap B) ( D ∩ B ) vertex separators in H [ D ] H[D] H [ D ] , so the minimality of P P P gives ∣ P ∣ ≤ ∣ Y ∣ |P| \leq|Y| ∣ P ∣ ≤ ∣ Y ∣ . Since ∣ Y ∣ ≤ ∣ δ H ′ ( D ) ∣ |Y| \leq|\delta_{H'}(D)| ∣ Y ∣ ≤ ∣ δ H ′ ( D ) ∣ , Claim 5.2 gives ∣ Y ∣ ≤ ∣ δ H ′ ( D ) ∣ < ∣ D ∩ B ∣ |Y| \leq|\delta_{H'}(D)| < |D \cap B| ∣ Y ∣ ≤ ∣ δ H ′ ( D ) ∣ < ∣ D ∩ B ∣ . Hence ∣ P ∣ < ∣ D ∩ B ∣ |P| < |D \cap B| ∣ P ∣ < ∣ D ∩ B ∣ . Moreover, P ⊆ D P \subseteq D P ⊆ D , so
∣ B ′ ∣ = ∣ B ∖ D ∣ + ∣ P ∣ = ∣ B ∣ − ∣ D ∩ B ∣ + ∣ P ∣ < ∣ B ∣ . |B'| = |B \setminus D| + |P| = |B| - |D \cap B| + |P| < |B|. ∣ B ′ ∣ = ∣ B ∖ D ∣ + ∣ P ∣ = ∣ B ∣ − ∣ D ∩ B ∣ + ∣ P ∣ < ∣ B ∣.
We now prove that adh H ( B ′ ) ≤ 2 k \operatorname{adh}_H(B') \leq2k adh H ( B ′ ) ≤ 2 k . Let K K K be a component of H − B ′ H - B' H − B ′ . It suffices to show that ∣ N H ( K ) ∣ ≤ 2 k |N_H(K)| \leq2k ∣ N H ( K ) ∣ ≤ 2 k . Suppose first that K ∩ B ≠ ∅ K \cap B \neq\varnothing K ∩ B = ∅ . We first show that K ⊆ D K \subseteq D K ⊆ D . Since B ∖ D ⊆ B ′ B \setminus D \subseteq B' B ∖ D ⊆ B ′ and K ∩ B ′ = ∅ K \cap B' = \varnothing K ∩ B ′ = ∅ , every vertex of K ∩ B K \cap B K ∩ B belongs to D D D . Moreover, P ⊆ B ′ P \subseteq B' P ⊆ B ′ gives K ∩ P = ∅ K \cap P = \varnothing K ∩ P = ∅ . Suppose that K ⊈ D K \nsubseteq D K ⊈ D . Choose a path Q Q Q in H [ K ] H[K] H [ K ] from K ∩ B K \cap B K ∩ B to K ∖ D K \setminus D K ∖ D , and let u v uv uv be the first edge of Q Q Q with u ∈ D u \in D u ∈ D and v ∉ D v \notin D v ∈ / D . Since X ∖ Z ⊆ B ′ X \setminus Z \subseteq B' X ∖ Z ⊆ B ′ , we have v ∉ X ∖ Z v \notin X \setminus Z v ∈ / X ∖ Z . Thus u v ∈ δ H ′ ( D ) uv \in\delta_{H'}(D) uv ∈ δ H ′ ( D ) and u ∈ Y u \in Y u ∈ Y . The initial subpath of Q Q Q ending at u u u lies in H [ D ] − P H[D] - P H [ D ] − P and joins D ∩ B D \cap B D ∩ B to Y Y Y , contradicting that P P P separates Y Y Y from D ∩ B D \cap B D ∩ B . Hence K ⊆ D K \subseteq D K ⊆ D .
Since H [ K ] H[K] H [ K ] is connected and K ∩ P = ∅ K \cap P = \varnothing K ∩ P = ∅ , the separator P P P also gives K ∩ Y = ∅ K \cap Y = \varnothing K ∩ Y = ∅ . Every neighbor of K K K in D D D belongs to B ′ ∩ D = P B' \cap D = P B ′ ∩ D = P . Every neighbor v ∉ D v \notin D v ∈ / D of K K K belongs to X ∖ Z X \setminus Z X ∖ Z , since any edge u v ∈ E ( H ) uv \in E(H) uv ∈ E ( H ) with u ∈ K u \in K u ∈ K and v ∉ D ∪ ( X ∖ Z ) v \notin D \cup(X \setminus Z) v ∈ / D ∪ ( X ∖ Z ) belongs to δ H ′ ( D ) \delta_{H'}(D) δ H ′ ( D ) , contrary to K ∩ Y = ∅ K \cap Y = \varnothing K ∩ Y = ∅ . Hence N H ( K ) ⊆ P ∪ ( X ∖ Z ) N_H(K) \subseteq P \cup(X \setminus Z) N H ( K ) ⊆ P ∪ ( X ∖ Z ) , and ∣ N H ( K ) ∣ ≤ ∣ P ∣ + ∣ X ∖ Z ∣ ≤ 2 k |N_H(K)| \leq|P| + |X \setminus Z| \leq2k ∣ N H ( K ) ∣ ≤ ∣ P ∣ + ∣ X ∖ Z ∣ ≤ 2 k .
Suppose now that K ∩ B = ∅ K \cap B = \varnothing K ∩ B = ∅ , and let K 0 K_0 K 0 be the component of H − B H - B H − B that contains K K K . We show that ∣ N H ( K ) ∣ ≤ ∣ N H ( K 0 ) ∣ ≤ adh H ( B ) |N_H(K)| \leq|N_H(K_0)| \leq\operatorname{adh}_H(B) ∣ N H ( K ) ∣ ≤ ∣ N H ( K 0 ) ∣ ≤ adh H ( B ) by constructing an injection from N H ( K ) N_H(K) N H ( K ) into N H ( K 0 ) N_H(K_0) N H ( K 0 ) . Figure 6 illustrates an injection from N H ( K ) N_H(K) N H ( K ) into N H ( K 0 ) N_H(K_0) N H ( K 0 ) when K ∩ B = ∅ K \cap B = \varnothing K ∩ B = ∅ .
Figure 6. The injection from N H ( K ) N_H(K) N H ( K ) into N H ( K 0 ) N_H(K_0) N H ( K 0 ) when K ∩ B = ∅ K \cap B = \varnothing K ∩ B = ∅ . The green Y Y Y –( D ∩ B ) (D \cap B) ( D ∩ B ) paths contain subpaths connecting the vertices of N H ( K ) ∩ ( P \ B ) N_H(K) \cap(P \backslash B) N H ( K ) ∩ ( P \ B ) to distinct vertices of N H ( K 0 ) \ B ′ N_H(K_0) \backslash B' N H ( K 0 ) \ B ′ .
By Menger’s theorem, there are ∣ P ∣ |P| ∣ P ∣ vertex-disjoint Y Y Y –( D ∩ B ) (D \cap B) ( D ∩ B ) paths in H [ D ] H[D] H [ D ] . Truncate each path at its first vertex in D ∩ B D \cap B D ∩ B . Since the paths are vertex-disjoint and every path intersects P P P , each path contains exactly one vertex of P P P , and every vertex of P P P lies on one path. For each p ∈ P ∖ B p \in P \setminus B p ∈ P ∖ B , let q p ∈ ( D ∩ B ) ∖ P q_p \in(D \cap B) \setminus P q p ∈ ( D ∩ B ) ∖ P be the endpoint of the path through p p p . The vertices q p q_p q p are distinct.
Since K K K is a component of H − B ′ H - B' H − B ′ , every vertex of N H ( K ) N_H(K) N H ( K ) belongs to B ′ B' B ′ . Together with B ′ = ( B ∖ D ) ∪ P B' = (B \setminus D) \cup P B ′ = ( B ∖ D ) ∪ P , this shows that every vertex of N H ( K ) N_H(K) N H ( K ) belongs either to N H ( K ) ∩ B N_H(K) \cap B N H ( K ) ∩ B or to N H ( K ) ∩ ( P ∖ B ) N_H(K) \cap(P \setminus B) N H ( K ) ∩ ( P ∖ B ) , and these two events are mutually exclusive. For x ∈ N H ( K ) ∩ B x \in N_H(K) \cap B x ∈ N H ( K ) ∩ B , the vertex x x x has a neighbor in K ⊆ K 0 K \subseteq K_0 K ⊆ K 0 , so x ∈ N H ( K 0 ) x \in N_H(K_0) x ∈ N H ( K 0 ) .
Now fix p ∈ N H ( K ) ∩ ( P ∖ B ) p \in N_H(K) \cap(P \setminus B) p ∈ N H ( K ) ∩ ( P ∖ B ) . Both p p p and K K K avoid B B B , and p p p has a neighbor in K K K , so p p p belongs to K 0 K_0 K 0 . On the path through p p p , the subpath from p p p to q p q_p q p contains no vertex of B B B except q p q_p q p , because q p q_p q p is the first vertex of D ∩ B D \cap B D ∩ B on the path. Hence all vertices of this subpath other than q p q_p q p belong to K 0 K_0 K 0 , and therefore q p ∈ N H ( K 0 ) q_p \in N_H(K_0) q p ∈ N H ( K 0 ) . Moreover, q p ∉ B ′ q_p \notin B' q p ∈ / B ′ because q p ∈ ( D ∩ B ) ∖ P q_p \in(D \cap B) \setminus P q p ∈ ( D ∩ B ) ∖ P and B ′ = ( B ∖ D ) ∪ P B' = (B \setminus D) \cup P B ′ = ( B ∖ D ) ∪ P .
Define f : N H ( K ) → N H ( K 0 ) f : N_H(K) \to N_H(K_0) f : N H ( K ) → N H ( K 0 ) by f ( x ) = x f(x) = x f ( x ) = x for x ∈ N H ( K ) ∩ B x \in N_H(K) \cap B x ∈ N H ( K ) ∩ B and f ( p ) = q p f(p) = q_p f ( p ) = q p for p ∈ N H ( K ) ∩ ( P ∖ B ) p \in N_H(K) \cap(P \setminus B) p ∈ N H ( K ) ∩ ( P ∖ B ) . The vertices q p q_p q p are distinct because the paths are vertex-disjoint. Moreover, f ( N H ( K ) ∩ B ) ⊆ B ′ f(N_H(K) \cap B) \subseteq B' f ( N H ( K ) ∩ B ) ⊆ B ′ , whereas f ( N H ( K ) ∩ ( P ∖ B ) ) ⊆ N H ( K 0 ) ∖ B ′ f(N_H(K) \cap(P \setminus B)) \subseteq N_H(K_0) \setminus B' f ( N H ( K ) ∩ ( P ∖ B )) ⊆ N H ( K 0 ) ∖ B ′ . Hence no image of a vertex in N H ( K ) ∩ B N_H(K) \cap B N H ( K ) ∩ B equals an image of a vertex in N H ( K ) ∩ ( P ∖ B ) N_H(K) \cap(P \setminus B) N H ( K ) ∩ ( P ∖ B ) . Thus f f f is injective, and hence ∣ N H ( K ) ∣ ≤ ∣ N H ( K 0 ) ∣ ≤ adh H ( B ) |N_H(K)| \leq|N_H(K_0)| \leq\operatorname{adh}_H(B) ∣ N H ( K ) ∣ ≤ ∣ N H ( K 0 ) ∣ ≤ adh H ( B ) .
Since ∣ P ∣ ≤ ∣ δ H ′ ( D ) ∣ ≤ k |P| \leq|\delta_{H'}(D)| \leq k ∣ P ∣ ≤ ∣ δ H ′ ( D ) ∣ ≤ k and ∣ X ∖ Z ∣ ≤ k |X \setminus Z| \leq k ∣ X ∖ Z ∣ ≤ k , the bounds for K ∩ B ≠ ∅ K \cap B \neq\varnothing K ∩ B = ∅ and K ∩ B = ∅ K \cap B = \varnothing K ∩ B = ∅ give
adh H ( B ′ ) ≤ max { adh H ( B ) , ∣ P ∪ ( X ∖ Z ) ∣ } ≤ 2 k . \operatorname{adh}_H(B') \leq\max\{\operatorname{adh}_H(B), |P \cup(X \setminus Z)|\} \leq2k. adh H ( B ′ ) ≤ max { adh H ( B ) , ∣ P ∪ ( X ∖ Z ) ∣ } ≤ 2 k .
The minimum separator P P P and the set B ′ B' B ′ are computable in polynomial time, so the total running time is O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) .
Now we prove Theorem 1.1.
Theorem 1.1. Given a graph G G G and k k k , one can construct in O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) time a ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable decomposition with adhesion at most 2 k 2k 2 k .
Proof. To construct the root bag, choose an arbitrary vertex v ∈ V ( G ) v \in V(G) v ∈ V ( G ) , set X : = { v } X := \{v\} X := { v } and B : = V ( G ) B := V(G) B := V ( G ) , and apply Lemma 5.1 to ( X , B ) (X,B) ( X , B ) . Whenever the lemma returns B ′ B' B ′ , replace B B B by B ′ B' B ′ and apply the lemma to ( X , B ) (X,B) ( X , B ) again. Each replacement satisfies X ⊆ B X \subseteq B X ⊆ B , adh G ( B ) ≤ 2 k \operatorname{adh}_G(B) \leq2k adh G ( B ) ≤ 2 k , and ∣ B ∣ |B| ∣ B ∣ strictly decreases. Hence, after at most ∣ V ( G ) ∣ |V(G)| ∣ V ( G ) ∣ applications, the lemma certifies that B B B is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in G G G .
We recurse on the components as follows. At the root, set H : = G H := G H := G and let B B B be the bag obtained by the repeated applications of Lemma 5.1 . After constructing a bag B B B in an induced subgraph H H H of G G G , create a node with bag B B B . For every component C C C of H − B H - B H − B , choose an arbitrary vertex v C ∈ C v_C \in C v C ∈ C , define H C : = H [ C ∪ N H ( C ) ] H_C := H[C \cup N_H(C)] H C := H [ C ∪ N H ( C )] and X C : = N H ( C ) ∪ { v C } X_C := N_H(C) \cup\{v_C\} X C := N H ( C ) ∪ { v C } . Since ∣ N H ( C ) ∣ ≤ adh H ( B ) ≤ 2 k |N_H(C)| \leq\operatorname{adh}_H(B) \leq2k ∣ N H ( C ) ∣ ≤ adh H ( B ) ≤ 2 k , we have ∣ X C ∣ ≤ 2 k + 1 |X_C| \leq2k + 1 ∣ X C ∣ ≤ 2 k + 1 . Starting with B C : = V ( H C ) B_C := V(H_C) B C := V ( H C ) , the same repeated application of Lemma 5.1 as for the root produces a set B C B_C B C such that X C ⊆ B C X_C \subseteq B_C X C ⊆ B C , adh H C ( B C ) ≤ 2 k \operatorname{adh}_{H_C}(B_C) \leq2k adh H C ( B C ) ≤ 2 k , and B C B_C B C is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in H C H_C H C . Create a child node with bag B C B_C B C , attach it below the node with bag B B B , and repeat the construction for every component of H C − B C H_C - B_C H C − B C . Since v C ∈ X C ⊆ B C v_C \in X_C \subseteq B_C v C ∈ X C ⊆ B C , the vertex v C v_C v C belongs to no descendant component; thus the recursion terminates after at most ∣ V ( G ) ∣ |V(G)| ∣ V ( G ) ∣ recursive steps.
We verify that the procedure actually yields a tree decomposition of G G G . Every vertex outside B B B belongs to some component C C C of H − B H-B H − B and hence to V ( H C ) V(H_C) V ( H C ) , so the recursion covers all vertices. For an edge x y ∈ E ( H ) xy \in E(H) x y ∈ E ( H ) , either x , y ∈ B x,y \in B x , y ∈ B , or both endpoints lie in C ∪ N H ( C ) = V ( H C ) C \cup N_H(C) = V(H_C) C ∪ N H ( C ) = V ( H C ) for some component C C C of H − B H-B H − B ; thus some bag contains both endpoints.
Fix x ∈ V ( H ) x \in V(H) x ∈ V ( H ) . If x ∉ B x \notin B x ∈ / B , it occurs only in the child instance for the component containing x x x . If x ∈ B x \in B x ∈ B , every child instance containing x x x has x ∈ N H ( C ) ⊆ X C ⊆ B C x \in N_H(C) \subseteq X_C \subseteq B_C x ∈ N H ( C ) ⊆ X C ⊆ B C . Induction therefore shows that the bags containing x x x form a connected subtree.
For the parent–child edge associated with C C C , since N H ( C ) ⊆ X C ⊆ B C N_H(C) \subseteq X_C \subseteq B_C N H ( C ) ⊆ X C ⊆ B C and B ∩ V ( H C ) = N H ( C ) B \cap V(H_C) = N_H(C) B ∩ V ( H C ) = N H ( C ) , we have B ∩ B C = N H ( C ) B \cap B_C = N_H(C) B ∩ B C = N H ( C ) . Thus the adhesion between B B B and B C B_C B C has size ∣ N H ( C ) ∣ ≤ adh H ( B ) ≤ 2 k |N_H(C)| \leq\operatorname{adh}_H(B) \leq2k ∣ N H ( C ) ∣ ≤ adh H ( B ) ≤ 2 k .
It remains to verify that every bag is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in G G G . Consider a bag B B B created in a recursive step on H = G [ W ] H = G[W] H = G [ W ] . For every A ⊆ V ( G ) A \subseteq V(G) A ⊆ V ( G ) with ∣ δ G ( A ) ∣ ≤ k |\delta_G(A)| \leq k ∣ δ G ( A ) ∣ ≤ k , we have δ H ( A ∩ W ) ⊆ δ G ( A ) \delta_H(A \cap W) \subseteq\delta_G(A) δ H ( A ∩ W ) ⊆ δ G ( A ) . The ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakability of B B B in H H H therefore gives min { ∣ A ∩ B ∣ , ∣ B ∖ A ∣ } ≤ 2 k \min\{|A \cap B|, |B \setminus A|\} \leq2k min { ∣ A ∩ B ∣ , ∣ B ∖ A ∣ } ≤ 2 k . Hence every bag is ( 2 k , k ) (2k,k) ( 2 k , k ) -edge-unbreakable in G G G . There are at most ∣ V ( G ) ∣ |V(G)| ∣ V ( G ) ∣ recursive steps, and each step applies Lemma 5.1 at most ∣ V ( H ) ∣ |V(H)| ∣ V ( H ) ∣ times, so the total running time is O ∗ ( 2 O ( k ) ) O^*(2^{O(k)}) O ∗ ( 2 O ( k ) ) .
Reduction from 3-SAT
We prove Theorem 1.4 by combining a bounded-degree reduction from 3-SAT to Max Cut with a reduction from Max Cut to Minimum Bisection.
Max Cut
Input: A simple undirected graph H H H and an integer t ≥ 0 t \ge0 t ≥ 0 .
Question: Is there a set X ⊆ V ( H ) X \subseteq V(H) X ⊆ V ( H ) such that ∣ δ H ( X ) ∣ ≥ t |\delta_H(X)| \ge t ∣ δ H ( X ) ∣ ≥ t ?
Garey, Johnson, and Stockmeyer [16 ] proved the NP-completeness of Max Cut via a reduction from 3-SAT. Here we need a reduction whose constructed Max Cut instance has linear size and bounded degree. We therefore give a reduction satisfying both requirements.
Lemma 6.1. There is a polynomial-time reduction from 3-SAT to Max Cut that maps a 3-CNF formula φ \varphi φ with a total of s s s variables and clauses to an instance ( H , t ) (H,t) ( H , t ) such that φ \varphi φ is satisfiable if and only if ( H , t ) (H,t) ( H , t ) is a yes-instance, H H H has maximum degree O ( 1 ) O(1) O ( 1 ) , and ∣ V ( H ) ∣ + ∣ E ( H ) ∣ = O ( s ) |V(H)| + |E(H)| = O(s) ∣ V ( H ) ∣ + ∣ E ( H ) ∣ = O ( s ) .
Proof. Introduce a variable z z z and, for every clause C = ( a ∨ b ∨ c ) C = (a \lor b \lor c) C = ( a ∨ b ∨ c ) of φ \varphi φ , a variable w C w_C w C . Replace C C C by the two constraints NAE( a , b , w C ) (a,b,w_C) ( a , b , w C ) and NAE( ¬ w C , c , z ) (\lnot w_C,c,z) ( ¬ w C , c , z ) , where NAE requires its three arguments not to be all equal. Let ψ \psi ψ be the conjunction of these constraints. Complementing every variable preserves all constraints of ψ \psi ψ . Hence, if ψ \psi ψ is satisfiable, it has a satisfying assignment with z = 0 z = 0 z = 0 . If c = 1 c = 1 c = 1 , one can choose w C w_C w C so that NAE( a , b , w C ) (a,b,w_C) ( a , b , w C ) holds, while NAE( ¬ w C , c , z ) (\lnot w_C,c,z) ( ¬ w C , c , z ) holds for either value of w C w_C w C . If c = 0 c = 0 c = 0 and a ∨ b = 1 a \lor b = 1 a ∨ b = 1 , setting w C = 0 w_C = 0 w C = 0 satisfies both constraints. If a = b = c = 0 a = b = c = 0 a = b = c = 0 , NAE( ¬ w C , c , z ) (\lnot w_C,c,z) ( ¬ w C , c , z ) forces w C = 0 w_C = 0 w C = 0 , which violates NAE( a , b , w C ) (a,b,w_C) ( a , b , w C ) . Hence, φ \varphi φ is satisfiable if and only if ψ \psi ψ is satisfiable.
Let r r r be the number of constraints of ψ \psi ψ . Construct a simple undirected graph H H H with one vertex for every literal occurrence in ψ \psi ψ . For each NAE constraint, add the three edges joining the vertices representing its three literal occurrences.
We next add a chain for every variable x x x of ψ \psi ψ . List the occurrence vertices of literals x x x and ¬ x \lnot x ¬ x in an arbitrary order u 1 , … , u d u_1,\ldots,u_d u 1 , … , u d . For every 1 ≤ h < d 1 \le h < d 1 ≤ h < d , add a path from u h u_h u h to u h + 1 u_{h+1} u h + 1 whose internal vertices are new. The path has length two if the two occurrences have the same sign and length three otherwise. Let L L L be the total length of all these paths and set t : = 2 r + L t := 2r + L t := 2 r + L .
The three edges added for any NAE constraint contribute at most two edges to any cut, and the paths between consecutive occurrences contain L L L edges in total. Hence no cut has more than t t t edges.
Assume first that φ \varphi φ is satisfiable. Then ψ \psi ψ has a satisfying assignment. For every literal occurrence, place its representing vertex in X X X exactly when the literal is true under this assignment. For each path joining two consecutive occurrences of a variable, its endpoints are on the same side of X X X if the path has length two and on opposite sides if the path has length three. We can therefore add its internal vertices to X X X so that every path edge crosses the cut. For each NAE constraint, its three literals are not all equal, so exactly two of the three edges among their vertices cross the cut. Hence ∣ δ H ( X ) ∣ = 2 r + L = t |\delta_H(X)| = 2r + L = t ∣ δ H ( X ) ∣ = 2 r + L = t .
Conversely, assume that X ⊆ V ( H ) X \subseteq V(H) X ⊆ V ( H ) satisfies ∣ δ H ( X ) ∣ ≥ t |\delta_H(X)| \ge t ∣ δ H ( X ) ∣ ≥ t . The preceding upper bound implies that equality holds, every path edge crosses the cut, and exactly two of the three edges added for each NAE constraint cross the cut. It follows along the chain for each variable x x x that any two occurrences with the same sign lie on the same side of X X X , whereas any two occurrences with opposite signs lie on opposite sides. Hence there is a truth value for x x x such that each occurrence of x x x or ¬ x \lnot x ¬ x is true exactly when its vertex belongs to X X X . Finally, the three vertices corresponding to any NAE constraint are not all on the same side of X X X , because exactly two of the three edges among them cross the cut. Thus every constraint of ψ \psi ψ is satisfied, and hence φ \varphi φ is satisfiable.
Each vertex representing a literal occurrence has degree at most four, and each internal path vertex has degree two. The number of constraints, literal occurrences, internal path vertices, and edges is O ( s ) O(s) O ( s ) . Therefore, H H H has maximum degree O ( 1 ) O(1) O ( 1 ) and ∣ V ( H ) ∣ + ∣ E ( H ) ∣ = O ( s ) |V(H)| + |E(H)| = O(s) ∣ V ( H ) ∣ + ∣ E ( H ) ∣ = O ( s ) .
For an instance ( H , t ) (H,t) ( H , t ) , write n : = ∣ V ( H ) ∣ n := |V(H)| n := ∣ V ( H ) ∣ and m : = ∣ E ( H ) ∣ m := |E(H)| m := ∣ E ( H ) ∣ . We transform a Max Cut instance ( H , t ) (H,t) ( H , t ) with 0 ≤ t ≤ m 0 \le t \le m 0 ≤ t ≤ m into a Minimum Bisection instance. Let q ≥ 2 q \ge2 q ≥ 2 be arbitrary. Partition V ( H ) V(H) V ( H ) into nonempty sets V 0 , … , V ⌈ n / q ⌉ − 1 V_0,\ldots,V_{\lceil n/q\rceil-1} V 0 , … , V ⌈ n / q ⌉ − 1 of size at most q q q . For every 0 ≤ i < ⌈ n / q ⌉ 0 \le i < \lceil n/q\rceil 0 ≤ i < ⌈ n / q ⌉ , define D i : = 1 + max v ∈ V i deg H ( v ) D_i := 1 + \max_{v\in V_i}\deg_H(v) D i := 1 + max v ∈ V i deg H ( v ) . Set
K : = ∑ i = 0 ⌈ n / q ⌉ − 1 ( D i ⋅ ∣ V i ∣ ⋅ ( ∣ V i ∣ − 1 ) ) + 2 ( m − t ) . K := \sum_{i=0}^{\lceil n/q\rceil-1} \left(D_i \cdot|V_i| \cdot(|V_i|-1)\right) + 2(m-t). K := i = 0 ∑ ⌈ n / q ⌉ − 1 ( D i ⋅ ∣ V i ∣ ⋅ ( ∣ V i ∣ − 1 ) ) + 2 ( m − t ) .
For every v ∈ V i v \in V_i v ∈ V i and a ∈ { 0 , 1 } a \in\{0,1\} a ∈ { 0 , 1 } , create a clique P ( v , a ) P(v,a) P ( v , a ) of size ( K + n + 2 ) ( q + 1 ) i (K+n+2)(q+1)^i ( K + n + 2 ) ( q + 1 ) i . Add edges between these cliques as follows.
For every i i i , every unordered pair of distinct vertices u , v ∈ V i u,v \in V_i u , v ∈ V i , and every a , b ∈ { 0 , 1 } a,b \in\{0,1\} a , b ∈ { 0 , 1 } , add D i D_i D i edges with distinct endpoint pairs between P ( u , a ) P(u,a) P ( u , a ) and P ( v , b ) P(v,b) P ( v , b ) .
For every u v ∈ E ( H ) uv \in E(H) uv ∈ E ( H ) , add an arbitrary edge between P ( u , 0 ) P(u,0) P ( u , 0 ) and P ( v , 1 ) P(v,1) P ( v , 1 ) and an arbitrary edge between P ( u , 1 ) P(u,1) P ( u , 1 ) and P ( v , 0 ) P(v,0) P ( v , 0 ) , distinct from all edges already added.
These choices are always possible: it adds at most n + 1 < K + n + 2 n+1 < K+n+2 n + 1 < K + n + 2 edges between any pair of cliques. The resulting graph G G G is simple, and the Minimum Bisection instance is ( G , K ) (G,K) ( G , K ) . The graph G G G has an even number of vertices because every vertex of H H H contributes two cliques of the same size. Figure 7 illustrates the construction of G G G .
Figure 7. The construction for the six-vertex graph H H H with q = 2 q = 2 q = 2 and t = 5 t = 5 t = 5 . The blocks are V 0 = { u , v } V_0 = \{u,v\} V 0 = { u , v } , V 1 = { w , x } V_1 = \{w,x\} V 1 = { w , x } , and V 2 = { y , z } V_2 = \{y,z\} V 2 = { y , z } . Here D 0 = 3 D_0 = 3 D 0 = 3 , D 1 = D 2 = 4 D_1 = D_2 = 4 D 1 = D 2 = 4 , and K = 26 K = 26 K = 26 . A thick blue line between P ( u , a ) P(u,a) P ( u , a ) and P ( v , b ) P(v,b) P ( v , b ) represents D i D_i D i distinct edges, where u , v ∈ V i u,v \in V_i u , v ∈ V i . Internal clique edges are omitted.
Lemma 6.2. There is a set X ⊆ V ( H ) X \subseteq V(H) X ⊆ V ( H ) with ∣ δ H ( X ) ∣ ≥ t |\delta_H(X)| \ge t ∣ δ H ( X ) ∣ ≥ t if and only if there is a set Y ⊆ V ( G ) Y \subseteq V(G) Y ⊆ V ( G ) with ∣ Y ∣ = ∣ V ( G ) ∖ Y ∣ |Y| = |V(G) \setminus Y| ∣ Y ∣ = ∣ V ( G ) ∖ Y ∣ and ∣ δ G ( Y ) ∣ ≤ K |\delta_G(Y)| \le K ∣ δ G ( Y ) ∣ ≤ K .
Proof. Suppose first that X ⊆ V ( H ) X \subseteq V(H) X ⊆ V ( H ) satisfies ∣ δ H ( X ) ∣ ≥ t |\delta_H(X)| \ge t ∣ δ H ( X ) ∣ ≥ t , and set
Y : = ⋃ v ∈ X P ( v , 0 ) ∪ ⋃ v ∈ V ( H ) ∖ X P ( v , 1 ) . Y := \bigcup_{v\in X} P(v,0) \cup\bigcup_{v\in V(H)\setminus X} P(v,1). Y := v ∈ X ⋃ P ( v , 0 ) ∪ v ∈ V ( H ) ∖ X ⋃ P ( v , 1 ) .
For each v ∈ V ( H ) v \in V(H) v ∈ V ( H ) , the set Y Y Y contains exactly one of the two equally sized cliques P ( v , 0 ) P(v,0) P ( v , 0 ) and P ( v , 1 ) P(v,1) P ( v , 1 ) . Hence ∣ Y ∣ = ∣ V ( G ) ∖ Y ∣ |Y| = |V(G) \setminus Y| ∣ Y ∣ = ∣ V ( G ) ∖ Y ∣ . We evaluate ∣ δ G ( Y ) ∣ |\delta_G(Y)| ∣ δ G ( Y ) ∣ . The edges introduced inside group V i V_i V i contribute to ∣ δ G ( Y ) ∣ |\delta_G(Y)| ∣ δ G ( Y ) ∣ by D i ∣ V i ∣ ( ∣ V i ∣ − 1 ) D_i|V_i|(|V_i|-1) D i ∣ V i ∣ ( ∣ V i ∣ − 1 ) . The two edges representing u v ∈ E ( H ) uv \in E(H) uv ∈ E ( H ) contribute zero when u v ∈ δ H ( X ) uv \in\delta_H(X) uv ∈ δ H ( X ) and two otherwise. Therefore,
∣ δ G ( Y ) ∣ = ∑ i = 0 ⌈ n / q ⌉ − 1 ( D i ⋅ ∣ V i ∣ ⋅ ( ∣ V i ∣ − 1 ) ) + 2 ( m − ∣ δ H ( X ) ∣ ) ≤ K . |\delta_G(Y)| = \sum_{i=0}^{\lceil n/q\rceil-1} \left(D_i \cdot|V_i| \cdot(|V_i|-1)\right) + 2(m-|\delta_H(X)|) \le K. ∣ δ G ( Y ) ∣ = i = 0 ∑ ⌈ n / q ⌉ − 1 ( D i ⋅ ∣ V i ∣ ⋅ ( ∣ V i ∣ − 1 ) ) + 2 ( m − ∣ δ H ( X ) ∣ ) ≤ K .
Conversely, let Y ⊆ V ( G ) Y \subseteq V(G) Y ⊆ V ( G ) minimize ∣ δ G ( Y ) ∣ |\delta_G(Y)| ∣ δ G ( Y ) ∣ among all sets satisfying ∣ Y ∣ = ∣ V ( G ) ∖ Y ∣ |Y| = |V(G) \setminus Y| ∣ Y ∣ = ∣ V ( G ) ∖ Y ∣ . Assume ∣ δ G ( Y ) ∣ ≤ K |\delta_G(Y)| \le K ∣ δ G ( Y ) ∣ ≤ K . Every clique P ( v , a ) P(v,a) P ( v , a ) lies entirely in Y Y Y or in V ( G ) ∖ Y V(G) \setminus Y V ( G ) ∖ Y : a nontrivial partition of a clique of size at least K + n + 2 K+n+2 K + n + 2 cuts more than K K K edges. Let a i a_i a i be the number of cliques from group V i V_i V i contained in Y Y Y . Since each such clique has size ( K + n + 2 ) ( q + 1 ) i (K+n+2)(q+1)^i ( K + n + 2 ) ( q + 1 ) i , we have
0 = ∣ Y ∣ − ∣ V ( G ) ∖ Y ∣ = 2 ( K + n + 2 ) ∑ i = 0 ⌈ n / q ⌉ − 1 ( a i − ∣ V i ∣ ) ⋅ ( q + 1 ) i . 0 = |Y| - |V(G)\setminus Y| = 2(K+n+2)\sum_{i=0}^{\lceil n/q\rceil-1} (a_i-|V_i|)\cdot(q+1)^i. 0 = ∣ Y ∣ − ∣ V ( G ) ∖ Y ∣ = 2 ( K + n + 2 ) i = 0 ∑ ⌈ n / q ⌉ − 1 ( a i − ∣ V i ∣ ) ⋅ ( q + 1 ) i .
We have − q ≤ a i − ∣ V i ∣ ≤ q -q \le a_i - |V_i| \le q − q ≤ a i − ∣ V i ∣ ≤ q , and the only multiple of q + 1 q+1 q + 1 in this interval is zero. Reducing the balance equation modulo q + 1 q+1 q + 1 therefore gives a 0 = ∣ V 0 ∣ a_0 = |V_0| a 0 = ∣ V 0 ∣ . Divide the equation by q + 1 q+1 q + 1 and repeat the argument to obtain a i = ∣ V i ∣ a_i = |V_i| a i = ∣ V i ∣ for every i i i .
We claim that, for every v ∈ V ( H ) v \in V(H) v ∈ V ( H ) , exactly one of P ( v , 0 ) P(v,0) P ( v , 0 ) and P ( v , 1 ) P(v,1) P ( v , 1 ) is contained in Y Y Y . Suppose otherwise, and fix a group V i V_i V i in which this fails. Because exactly ∣ V i ∣ |V_i| ∣ V i ∣ of the 2 ∣ V i ∣ 2|V_i| 2∣ V i ∣ cliques P ( x , a ) P(x,a) P ( x , a ) are contained in Y Y Y , there exist u , v ∈ V i u,v \in V_i u , v ∈ V i such that P ( u , 0 ) ∪ P ( u , 1 ) ⊆ Y P(u,0) \cup P(u,1) \subseteq Y P ( u , 0 ) ∪ P ( u , 1 ) ⊆ Y and P ( v , 0 ) ∪ P ( v , 1 ) ⊆ V ( G ) ∖ Y P(v,0) \cup P(v,1) \subseteq V(G) \setminus Y P ( v , 0 ) ∪ P ( v , 1 ) ⊆ V ( G ) ∖ Y . Define Y ′ : = ( Y ∖ P ( u , 0 ) ) ∪ P ( v , 0 ) Y' := (Y \setminus P(u,0)) \cup P(v,0) Y ′ := ( Y ∖ P ( u , 0 )) ∪ P ( v , 0 ) . Since ∣ P ( u , 0 ) ∣ = ∣ P ( v , 0 ) ∣ |P(u,0)| = |P(v,0)| ∣ P ( u , 0 ) ∣ = ∣ P ( v , 0 ) ∣ , we have ∣ Y ′ ∣ = ∣ Y ∣ |Y'| = |Y| ∣ Y ′ ∣ = ∣ Y ∣ .
Under the first rule, for every w ∈ V i ∖ { u , v } w \in V_i \setminus\{u,v\} w ∈ V i ∖ { u , v } and a ∈ { 0 , 1 } a \in\{0,1\} a ∈ { 0 , 1 } , the number of cut edges between P ( w , a ) P(w,a) P ( w , a ) and P ( u , 0 ) ∪ P ( v , 0 ) P(u,0) \cup P(v,0) P ( u , 0 ) ∪ P ( v , 0 ) is the same for Y Y Y and Y ′ Y' Y ′ , while the number of cut edges between P ( u , a ) P(u,a) P ( u , a ) and P ( v , b ) P(v,b) P ( v , b ) over a , b ∈ { 0 , 1 } a,b \in\{0,1\} a , b ∈ { 0 , 1 } decreases from 4 D i 4D_i 4 D i to 2 D i 2D_i 2 D i . Every clique other than P ( u , 0 ) P(u,0) P ( u , 0 ) and P ( v , 0 ) P(v,0) P ( v , 0 ) lies on the same side of the cut for Y Y Y and Y ′ Y' Y ′ . Therefore, a second-rule edge can change its membership in the cut only if it has an endpoint in P ( u , 0 ) ∪ P ( v , 0 ) P(u,0) \cup P(v,0) P ( u , 0 ) ∪ P ( v , 0 ) . There are at most deg H ( u ) + deg H ( v ) \deg_H(u) + \deg_H(v) deg H ( u ) + deg H ( v ) such edges, so
∣ δ G ( Y ) ∣ − ∣ δ G ( Y ′ ) ∣ ≥ 2 D i − deg H ( u ) − deg H ( v ) ≥ 2. |\delta_G(Y)| - |\delta_G(Y')| \ge2D_i - \deg_H(u) - \deg_H(v) \ge2. ∣ δ G ( Y ) ∣ − ∣ δ G ( Y ′ ) ∣ ≥ 2 D i − deg H ( u ) − deg H ( v ) ≥ 2.
This contradicts the choice of Y Y Y and proves the claim.
We now recover a cut of H H H from this bisection. Define X : = { v ∈ V ( H ) : P ( v , 0 ) ⊆ Y } X := \{v \in V(H) : P(v,0) \subseteq Y\} X := { v ∈ V ( H ) : P ( v , 0 ) ⊆ Y } . For each group V i V_i V i , exactly D i ∣ V i ∣ ( ∣ V i ∣ − 1 ) D_i|V_i|(|V_i| - 1) D i ∣ V i ∣ ( ∣ V i ∣ − 1 ) edges added by the first rule belong to δ G ( Y ) \delta_G(Y) δ G ( Y ) . For each edge of E ( H ) ∖ δ H ( X ) E(H) \setminus\delta_H(X) E ( H ) ∖ δ H ( X ) , both edges added by the second rule belong to δ G ( Y ) \delta_G(Y) δ G ( Y ) . Thus,
∣ δ G ( Y ) ∣ = ∑ i = 0 ⌈ n / q ⌉ − 1 ( D i ⋅ ∣ V i ∣ ⋅ ( ∣ V i ∣ − 1 ) ) + 2 ( m − ∣ δ H ( X ) ∣ ) ≤ K . |\delta_G(Y)| = \sum_{i=0}^{\lceil n/q\rceil-1} \left(D_i \cdot|V_i| \cdot(|V_i|-1)\right) + 2(m-|\delta_H(X)|) \le K. ∣ δ G ( Y ) ∣ = i = 0 ∑ ⌈ n / q ⌉ − 1 ( D i ⋅ ∣ V i ∣ ⋅ ( ∣ V i ∣ − 1 ) ) + 2 ( m − ∣ δ H ( X ) ∣ ) ≤ K .
The definition of K K K therefore gives ∣ δ H ( X ) ∣ ≥ t |\delta_H(X)| \ge t ∣ δ H ( X ) ∣ ≥ t , so X X X is the required cut of H H H .
We have
∣ V ( G ) ∣ = 2 ( K + n + 2 ) ∑ i = 0 ⌈ n / q ⌉ − 1 ( ∣ V i ∣ ⋅ ( q + 1 ) i ) |V(G)| = 2(K+n+2)\sum_{i=0}^{\lceil n/q\rceil-1}\left(|V_i|\cdot(q+1)^i\right) ∣ V ( G ) ∣ = 2 ( K + n + 2 ) i = 0 ∑ ⌈ n / q ⌉ − 1 ( ∣ V i ∣ ⋅ ( q + 1 ) i )
≤ 2 ( K + n + 2 ) ∑ i = 0 ⌈ n / q ⌉ − 1 ( q ⋅ ( q + 1 ) i ) \le2(K+n+2)\sum_{i=0}^{\lceil n/q\rceil-1}\left(q\cdot(q+1)^i\right) ≤ 2 ( K + n + 2 ) i = 0 ∑ ⌈ n / q ⌉ − 1 ( q ⋅ ( q + 1 ) i )
≤ 8 q 2 ( n + m ) ( q + 1 ) n / q + 1 = 2 ( n / q + 1 ) log 2 ( q + 1 ) + log 2 ( 8 q 2 ( n + m ) ) . \le8q^2(n+m)(q+1)^{n/q+1}=2^{(n/q+1)\log_2(q+1)+\log_2(8q^2(n+m))}. ≤ 8 q 2 ( n + m ) ( q + 1 ) n / q + 1 = 2 ( n / q + 1 ) l o g 2 ( q + 1 ) + l o g 2 ( 8 q 2 ( n + m )) .
Since G G G is simple, it has fewer than ∣ V ( G ) ∣ 2 |V(G)|^2 ∣ V ( G ) ∣ 2 edges. Thus, the construction takes time polynomial in ∣ V ( G ) ∣ |V(G)| ∣ V ( G ) ∣ .
Theorem 1.4. Let g : N → R ≥ 0 g:\mathbb{N}\to\mathbb{R}_{\ge0} g : N → R ≥ 0 be unbounded and nondecreasing, satisfy g ( x ) = O ( x ) g(x)=O(x) g ( x ) = O ( x ) , and satisfy g ( λ x ) = O ( g ( x ) ) g(\lambda x)=O(g(x)) g ( λ x ) = O ( g ( x )) for every fixed constant λ > 0 \lambda>0 λ > 0 . Suppose that Minimum Bisection can be solved in O ∗ ( 2 o ( g ( k ) ) ) O^*(2^{o(g(k))}) O ∗ ( 2 o ( g ( k )) ) time. Then 3-SAT can be solved in 2 o ( g ( q s ) ) + O ( s log q / q + log s ) 2^{o(g(qs))+O(s\log q/q+\log s)} 2 o ( g ( q s )) + O ( s l o g q / q + l o g s ) time for every integer q q q with 2 ≤ q ≤ s 2\le q\le s 2 ≤ q ≤ s , where s s s is the total number of variables and clauses in the input 3-CNF formula.8
In particular, the theorem yields three consequences.
For every fixed 0 < α < 1 0<\alpha<1 0 < α < 1 , if 3-SAT has no O ∗ ( 2 o ( s α ) ) O^*(2^{o(s^\alpha)}) O ∗ ( 2 o ( s α ) ) -time algorithm, then Minimum Bisection has no O ∗ ( 2 o ( ( k / log k ) α / ( 2 − α ) ) ) O^*(2^{o((k/\log k)^{\alpha/(2-\alpha)})}) O ∗ ( 2 o (( k / l o g k ) α / ( 2 − α ) ) ) -time algorithm.
For every fixed c > 0 c>0 c > 0 , if 3-SAT has no O ∗ ( 2 o ( s / ( log s ) c ) ) O^*(2^{o(s/(\log s)^c)}) O ∗ ( 2 o ( s / ( l o g s ) c ) ) -time algorithm, then Minimum Bisection has no O ∗ ( 2 o ( k / ( ( log k ) 2 c log log k ) ) ) O^*(2^{o(k/((\log k)^{2c}\log\log k))}) O ∗ ( 2 o ( k / (( l o g k ) 2 c l o g l o g k )) ) -time algorithm.
If 3-SAT has no O ∗ ( 2 o ( s ) ) O^*(2^{o(s)}) O ∗ ( 2 o ( s ) ) -time algorithm, equivalently, if ETH holds, then Minimum Bisection has no O ∗ ( 2 o ( k ) ) O^*(2^{o(k)}) O ∗ ( 2 o ( k ) ) -time algorithm.
Proof. Suppose that Minimum Bisection can be solved in O ∗ ( 2 h ( k ) ) O^*(2^{h(k)}) O ∗ ( 2 h ( k ) ) time, where h ( k ) = o ( g ( k ) ) h(k)=o(g(k)) h ( k ) = o ( g ( k )) . We may replace h ( x ) h(x) h ( x ) by max 0 ≤ j ≤ x h ( j ) \max_{0\le j\le x}h(j) max 0 ≤ j ≤ x h ( j ) and assume that h h h is nondecreasing. The redefined h h h is still o ( g ) o(g) o ( g ) because g g g is nondecreasing and unbounded. Let φ \varphi φ be a 3-CNF formula of size s s s , and fix an integer q q q with 2 ≤ q ≤ s 2\le q\le s 2 ≤ q ≤ s . Apply Lemma 6.1 to obtain ( H , t ) (H,t) ( H , t ) , and then apply the construction above to obtain ( G , K ) (G,K) ( G , K ) . Write n : = ∣ V ( H ) ∣ n:=|V(H)| n := ∣ V ( H ) ∣ and m : = ∣ E ( H ) ∣ m:=|E(H)| m := ∣ E ( H ) ∣ . The two reductions are correct by Lemmas 6.1 and 6.2 .
The maximum-degree bound in Lemma 6.1 gives D i = O ( 1 ) D_i=O(1) D i = O ( 1 ) . Since ∣ V i ∣ ≤ q |V_i|\le q ∣ V i ∣ ≤ q and ∑ i ∣ V i ∣ = n \sum_i|V_i|=n ∑ i ∣ V i ∣ = n , the definition of K K K gives K = O ( q n + m ) = O ( q s ) K=O(qn+m)=O(qs) K = O ( q n + m ) = O ( q s ) . Hence there is an absolute constant C C C such that K ≤ C q s K\le Cqs K ≤ C q s . Since g g g is nondecreasing and g ( C x ) = O ( g ( x ) ) g(Cx)=O(g(x)) g ( C x ) = O ( g ( x )) , we have h ( K ) ≤ h ( C q s ) = o ( g ( C q s ) ) = o ( g ( q s ) ) h(K)\le h(Cqs)=o(g(Cqs))=o(g(qs)) h ( K ) ≤ h ( C q s ) = o ( g ( C q s )) = o ( g ( q s )) . Moreover,
∣ V ( G ) ∣ ≤ 8 q 2 ( n + m ) ( q + 1 ) n / q + 1 = 2 O ( s log q / q + log s ) . |V(G)|\le8q^2(n+m)(q+1)^{n/q+1}=2^{O(s\log q/q+\log s)}. ∣ V ( G ) ∣ ≤ 8 q 2 ( n + m ) ( q + 1 ) n / q + 1 = 2 O ( s l o g q / q + l o g s ) .
The hypothetical algorithm applied to ( G , K ) (G, K) ( G , K ) thus decides whether φ \varphi φ is satisfiable in 2 o ( g ( q s ) ) ⋅ 2 O ( s log q / q + log s ) 2^{o(g(qs))} \cdot2^{O(s \log q/q+\log s)} 2 o ( g ( q s )) ⋅ 2 O ( s l o g q / q + l o g s ) time, as claimed.
We finally verify the three consequences in the statement. We choose q q q separately for each consequence. Let r ( s ) r(s) r ( s ) tend to infinity sufficiently slowly. For the first consequence, put β : = α / ( 2 − α ) \beta:= \alpha/(2-\alpha) β := α / ( 2 − α ) , use g ( k ) : = ( k / log k ) β g(k) := (k/\log k)^{\beta} g ( k ) := ( k / log k ) β , and choose q : = ⌈ s 1 − α log s ⋅ r ( s ) ⌉ q := \lceil s^{1-\alpha}\log s\cdot r(s)\rceil q := ⌈ s 1 − α log s ⋅ r ( s )⌉ . For the second consequence, use g ( k ) : = k / ( ( log k ) 2 c log log k ) g(k) := k/((\log k)^{2c}\log\log k) g ( k ) := k / (( log k ) 2 c log log k ) and choose q : = ⌈ ( log s ) c log log s ⋅ r ( s ) ⌉ q := \lceil(\log s)^{c}\log\log s\cdot r(s)\rceil q := ⌈( log s ) c log log s ⋅ r ( s )⌉ . For the third consequence, use g ( k ) : = k g(k) := k g ( k ) := k and choose q : = ⌈ r ( s ) ⌉ q := \lceil r(s)\rceil q := ⌈ r ( s )⌉ . In the three cases, r r r can be chosen so slowly that the exponent o ( g ( q s ) ) + O ( s log q / q + log s ) o(g(qs))+O(s\log q/q+\log s) o ( g ( q s )) + O ( s log q / q + log s ) is, respectively, o ( s α ) o(s^{\alpha}) o ( s α ) , o ( s / ( log s ) c ) o(s/(\log s)^{c}) o ( s / ( log s ) c ) , and o ( s ) o(s) o ( s ) . These three bounds contradict the three hypotheses on 3-SAT in the theorem statement.
Acknowledgments
The author thanks Tomohiro Koana for useful discussions and for improving the presentation of the paper.
Declaration of generative AI use
ChatGPT 6 Pro generated the proofs of Theorems 1.1 to 1.3. GPT-6 Astra generated the gadget construction for Theorem 1.4 as part of the ETH-tightness proof, before either OpenAI or Anthropic publicly announced a refutation of ETH. GPT-5.6 Sol and GPT-6 Astra were used in drafting and editing the manuscript. The author verified and revised the proofs and the manuscript and takes full responsibility for the paper.
References