This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

On Approximate Reconfigurability of Label Cover

Naoto Ohsaka CyberAgent, Inc., Tokyo, Japan. [email protected]; [email protected]
Abstract

Given a two-prover game GG and its two satisfying labelings Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}}, the Label Cover Reconfiguration problem asks whether Οˆπ—Œ\psi_{\mathsf{s}} can be transformed into Οˆπ—\psi_{\mathsf{t}} by repeatedly changing the value of a vertex while preserving any intermediate labeling satisfying GG. We consider an optimization variant of Label Cover Reconfiguration by relaxing the feasibility of labelings, referred to as Maxmin Label Cover Reconfiguration: we are allowed to transform by passing through any non-satisfying labelings, but required to maximize the minimum fraction of satisfied edges during transformation from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}}. Since the parallel repetition theorem of RazΒ (SIAMΒ J.Β Comput.,Β 1998)Β [Raz98], which implies NP-hardness of Label Cover within any constant factor, produces strong inapproximability results for many NP-hard problems, one may think of using Maxmin Label Cover Reconfiguration to derive inapproximability results for reconfiguration problems. We prove the following results on Maxmin Label Cover Reconfiguration, which display different trends from those of Label Cover and the parallel repetition theorem:

  • β€’

    Maxmin Label Cover Reconfiguration can be approximated within a factor of nearly 14\frac{1}{4} for restricted graph classes, including slightly dense graphs and balanced bipartite graphs.

  • β€’

    A naive parallel repetition of Maxmin Label Cover Reconfiguration does not decrease the optimal objective value.

  • β€’

    Label Cover Reconfiguration on projection games can be decided in polynomial time.

The above results suggest that a reconfiguration analogue of the parallel repetition theorem is unlikely.

1 Introduction

Background.

In reconfiguration problems [IDH+11], given a pair of feasible solutions for a combinatorial problem, we wish to find a step-by-step transformation from one to the other while the feasibility of every intermediate solution is maintained. Under the framework of reconfiguration [IDH+11], numerous reconfiguration problems have been derived from classical search problems; refer to the surveys by NishimuraΒ [Nis18] and van den HeuvelΒ [vdH13] for an overview of reconfiguration.

In this article, we consider reconfigurability of Label Cover [ABSS97] and its approximation. Given a two-prover game GG and its two satisfying labelings Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}}, the Label Cover Reconfiguration problem asks whether Οˆπ—Œ\psi_{\mathsf{s}} can be transformed into Οˆπ—\psi_{\mathsf{t}} by repeatedly changing the value of a vertex preserving any intermediate labeling satisfying GG. We can consider an optimization variant of this problem by relaxing the feasibility of labelings, referred to as Maxmin Label Cover Reconfiguration: we are allowed to transform by passing through any non-satisfying labelings, but required to maximize the minimum fraction of satisfied edges during transformation from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}}. Solving Maxmin Label Cover Reconfiguration approximately, we may find approximate reconfiguration for Label Cover Reconfiguration consisting of almost-satisfying labelings.

The recent work of the authorΒ [Ohs23] demonstrated that assuming PSPACE-hardness of approximation for Maxmin CSP Reconfiguration (which implies that for Maxmin Label Cover Reconfiguration), a bunch of reconfiguration problems are also PSPACE-hard to approximate, say, within a factor of (1βˆ’Ξ΅)(1-\varepsilon) for some Ρ∈(0,1)\varepsilon\in(0,1). One limitation of this approach is that the value of such Ξ΅\varepsilon might be too small to rule out a 0.999​⋯​90.999\cdots 9-approximation algorithm. In NP optimization problems, the parallel repetition theorem due to RazΒ [Raz98] serves as the β€œmother” of many strong inapproximability results [Fei98, HΓ₯s99, HΓ₯s01]. The crux of this theorem is to imply along with the PCP theoremΒ [AS98, ALM+98] that for every Ρ∈(0,1)\varepsilon\in(0,1), there is a finite alphabet Ξ£\Sigma such that Label Cover on Ξ£\Sigma is NP-hard to approximate within a factor of Ξ΅\varepsilon. One might thus think of a reconfiguration analogue of the parallel repetition theorem for Maxmin Label Cover Reconfiguration, which would help in improving PSPACE-hardness of approximation for reconfiguration problems. Our contribution is to give evidence that such hopes are probably dashed.

Our Results.

We present a few results on Maxmin Label Cover Reconfiguration, which display different trends from those for Label Cover and the parallel repetition theorem.

  • β€’

    In SectionΒ 2, we prove that Maxmin Label Cover Reconfiguration can be approximated within a factor of nearly 14\frac{1}{4} for restricted graph classes, including slightly dense graphs and balanced bipartite graphs,

  • β€’

    In SectionΒ 3, we further show that a naive parallel repetition of Maxmin Label Cover Reconfiguration does not decrease the optimal objective value, defined as the worst fraction of satisfied edges during transformation.

  • β€’

    In SectionΒ 4, we develop a polynomial-time algorithm for deciding Label Cover Reconfiguration on projection games, which is based on a simple characterization of the reconfigurability between a pair of satisfying labelings.

Our results suggest that a reconfiguration analogue of the parallel repetition theorem is unlikely; thus, we should resort to a different approach to derive an improved factor of inapproximability for reconfiguration problems.

Preliminaries.

For two integers m,nβˆˆβ„•m,n\in\mathbb{N} with mβ©½nm\leqslant n, let [n]β‰œ{1,2,…,n}[n]\triangleq\{1,2,\ldots,n\} and [m..n]β‰œ{m,m+1,…,nβˆ’1,n}[m\mathrel{..}\nobreak n]\triangleq\{m,m+1,\ldots,n-1,n\}. For a statement PP, \llbracket​P​\rrbracket\llbracket P\rrbracket is 11 if PP is true, and 0 otherwise. We formally define Label Cover Reconfiguration and its optimization variant.A constraint graph is defined as a tuple G=(V,E,Ξ£,Ξ )G=(V,E,\Sigma,\Pi), where

  • β€’

    (V,E)(V,E) is an undirected graph called the underlying graph of GG,

  • β€’

    Ξ£\Sigma is a finite set called the alphabet, and

  • β€’

    Ξ =(Ο€e)e∈E\Pi=(\pi_{e})_{e\in E} is a collection of binary constraints, where each Ο€eβŠ†Ξ£e\pi_{e}\subseteq\Sigma^{e} consists of pairs of admissible values that endpoints of ee can take.

If the underlying graph of GG is a bipartite graph, GG is called a two-prover game or simply game. We write G=(X,Y,E,Ξ£,Ξ )G=(X,Y,E,\Sigma,\Pi) to stress that the underlying graph of GG is a bipartite graph with bipartition (X,Y)(X,Y). Unless otherwise specified, we use GG to represent a game. Moreover, GG is said to be a projection game if every constraint Ο€(x,y)\pi_{(x,y)} for (x,y)∈E(x,y)\in E has a projection property; i.e., each value β∈Σ\beta\in\Sigma for yy has a unique value α∈Σ\alpha\in\Sigma for xx such that (Ξ±,Ξ²)βˆˆΟ€(x,y)(\alpha,\beta)\in\pi_{(x,y)}. Such Ξ±\alpha is denoted Ο€(x,y)​(Ξ²)\pi_{(x,y)}(\beta).

A labeling for a constraint graph GG is a mapping ψ:Vβ†’Ξ£\psi\colon V\to\Sigma that assigns value of Ξ£\Sigma to vertex of VV. We say that ψ\psi satisfies edge e=(v,w)∈Ee=(v,w)\in E (or constraint Ο€e\pi_{e}) if (Οˆβ€‹(v),Οˆβ€‹(w))βˆˆΟ€e(\psi(v),\psi(w))\in\pi_{e}, and ψ\psi satisfies GG if it satisfies all edges of GG. For two satisfying labelings Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}} for GG, a reconfiguration sequence from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}} is any sequence of labelings starting from Οˆπ—Œ\psi_{\mathsf{s}} and ending with Οˆπ—\psi_{\mathsf{t}} such that each labeling is obtained from the previous one by changing the value of a single vertex; i.e., they differ in exactly one vertex. The Label Cover Reconfiguration problem is defined as follows:

{itembox}

[l]Label Cover Reconfiguration Input: a satisfiable game GG and its two satisfying labelings Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}}. Question: is there a reconfiguration sequence of satisfying labelings from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}}?

Label Cover Reconfiguration is known to be PSPACE-complete, e.g., [GKMP09]. We then proceed to an optimization variant [IDH+11] of Label Cover Reconfiguration, in which we are allowed to touch non-satisfying labelings. For a game G=(V,E,Ξ£,Ξ )G=(V,E,\Sigma,\Pi) and a labeling ψ:Vβ†’Ξ£\psi\colon V\to\Sigma, let 𝗏𝖺𝗅G⁑(ψ)\operatorname{\mathsf{val}}_{G}(\psi) denote the fraction of edges satisfied by ψ\psi; namely,

𝗏𝖺𝗅G⁑(ψ)β‰œ1|E|β‹…|{e∈E|Οˆβ€‹Β satisfies ​e}|.\displaystyle\operatorname{\mathsf{val}}_{G}(\psi)\triangleq\frac{1}{|E|}\cdot\left|\left\{e\in E\Bigm{|}\psi\text{ satisfies }e\right\}\right|. (1.1)

For any reconfiguration sequence \uppsi=⟨ψ(0),…,ψ(β„“)⟩\mathbf{\bm{\uppsi}}=\langle\psi^{(0)},\ldots,\psi^{(\ell)}\rangle, let 𝗏𝖺𝗅G⁑(\uppsi)\operatorname{\mathsf{val}}_{G}(\mathbf{\bm{\uppsi}}) denote the minimum fraction of satisfied edges over all ψ(i)\psi^{(i)}’s in \uppsi\mathbf{\bm{\uppsi}}; namely,

𝗏𝖺𝗅G⁑(\uppsi)β‰œminψ(i)∈\uppsi⁑𝗏𝖺𝗅G⁑(ψ(i)).\displaystyle\operatorname{\mathsf{val}}_{G}(\mathbf{\bm{\uppsi}})\triangleq\min_{\psi^{(i)}\in\mathbf{\bm{\uppsi}}}\operatorname{\mathsf{val}}_{G}(\psi^{(i)}). (1.2)

Then, Maxmin Label Cover Reconfiguration is defined as the following optimization problem:

{itembox}

[l]Maxmin Label Cover Reconfiguration Input: a satisfiable game GG and its two satisfying labelings Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}}. Question: maximize 𝗏𝖺𝗅G⁑(\uppsi)\operatorname{\mathsf{val}}_{G}(\mathbf{\bm{\uppsi}}) subject to that \uppsi\mathbf{\bm{\uppsi}} is a reconfiguration sequence from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}}.

2 Nearly 14\frac{1}{4}-factor Approximation

We show that Maxmin Label Cover Reconfiguration is approximately reconfigurable within a factor of nearly 14\frac{1}{4} for some graph classes. For a game G=(V,E,Ξ£,Ξ )G=(V,E,\Sigma,\Pi) and its two labelings Οˆπ—Œ,Οˆπ—:Vβ†’Ξ£\psi_{\mathsf{s}},\psi_{\mathsf{t}}\colon V\to\Sigma, let 𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}}) denote the maximum value of 𝗏𝖺𝗅G⁑(\uppsi)\operatorname{\mathsf{val}}_{G}(\mathbf{\bm{\uppsi}}) over all possible reconfiguration sequences \uppsi\mathbf{\bm{\uppsi}} from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}}; namely,

𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)β‰œmax\uppsi=βŸ¨Οˆπ—Œ,…,Οˆπ—βŸ©β‘π—π–Ίπ—…G⁑(\uppsi).\displaystyle\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})\triangleq\max_{\mathbf{\bm{\uppsi}}=\langle\psi_{\mathsf{s}},\ldots,\psi_{\mathsf{t}}\rangle}\operatorname{\mathsf{val}}_{G}(\mathbf{\bm{\uppsi}}). (2.1)
Theorem 2.1.

For a game GG having mm edges and two of its satisfying labelings Οˆπ—Œ,Οˆπ—:Vβ†’Ξ£\psi_{\mathsf{s}},\psi_{\mathsf{t}}\colon V\to\Sigma, the following holds:

  • β€’

    If the average degree of GG is at least 6Ξ΅\frac{6}{\varepsilon} for Ρ∈(0,14)\varepsilon\in(0,\frac{1}{4}), then 𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)β©Ύ14βˆ’Ξ΅\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})\geqslant\frac{1}{4}-\varepsilon. Moreover, the same result holds even if GG is a general (i.e., non-bipartite) constraint graph.

  • β€’

    If GG is balanced (i.e., the two parts have the same size), then 𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)β©Ύ14​(1βˆ’1m)\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})\geqslant\frac{1}{4}\left(1-\frac{1}{\sqrt{m}}\right).

The proof of TheoremΒ 2.1 relies on the following claim.

Claim 2.2.

For a constraint graph G=(V,E,Ξ£,Ξ )G=(V,E,\Sigma,\Pi) and its two satisfying labelings Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}}, it holds that

𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)β©ΎmaxSβŠ‚V⁑min⁑{|E​[S]||E|,|E​[Vβˆ–S]||E|},\displaystyle\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})\geqslant\max_{S\subset V}\min\left\{\frac{|E[S]|}{|E|},\frac{|E[V\setminus S]|}{|E|}\right\}, (2.2)

where E​[S]E[S] is the edge set of a subgraph of GG induced by vertex set SS.

Proof.

For a vertex set SβŠ‚VS\subset V, we define a labeling ψvia:Vβ†’Ξ£\psi_{\text{via}}\colon V\to\Sigma as follows:

ψvia​(v)β‰œ{Οˆπ—Œβ€‹(v)if ​v∈S,Οˆπ—β€‹(v)if ​vβˆ‰S.\displaystyle\psi_{\text{via}}(v)\triangleq\begin{cases}\psi_{\mathsf{s}}(v)&\text{if }v\in S,\\ \psi_{\mathsf{t}}(v)&\text{if }v\notin S.\end{cases} (2.3)

Consider transforming Οˆπ—Œ\psi_{\mathsf{s}} into ψvia\psi_{\text{via}} by changing the value of vβˆ‰Sv\notin S from Οˆπ—Œβ€‹(v)\psi_{\mathsf{s}}(v) to Οˆπ—β€‹(v)\psi_{\mathsf{t}}(v) one by one. Since the value of vertices in SS has never been changed, any edge of G​[S]G[S] are always satisfied during this transformation; i.e., 𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆvia)β©Ύ|E​[S]||E|\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\text{via}})\geqslant\frac{|E[S]|}{|E|}. Similarly, it follows that 𝗏𝖺𝗅G⁑(ψviaβ†­Οˆπ—)β©Ύ|E​[Vβˆ–S]||E|\operatorname{\mathsf{val}}_{G}(\psi_{\text{via}}\leftrightsquigarrow\psi_{\mathsf{t}})\geqslant\frac{|E[V\setminus S]|}{|E|}. Consequently, we derive

𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)β©Ύmin⁑{𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆvia),𝗏𝖺𝗅G⁑(ψviaβ†­Οˆπ—)}β©Ύmin⁑{|E​[S]||E|,|E​[Vβˆ–S]||E|},\displaystyle\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})\geqslant\min\Bigl{\{}\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\text{via}}),\operatorname{\mathsf{val}}_{G}(\psi_{\text{via}}\leftrightsquigarrow\psi_{\mathsf{t}})\Bigr{\}}\geqslant\min\left\{\frac{|E[S]|}{|E|},\frac{|E[V\setminus S]|}{|E|}\right\}, (2.4)

as desired. ∎

We focus on the case that G=(X,Y,E,Ξ£,Ξ )G=(X,Y,E,\Sigma,\Pi) is balanced; i.e., |X|=|Y||X|=|Y|. We wish to partition each XX and YY in a particular well-balanced manner. To this end, we use the following auxiliary lemma.

Lemma 2.3.

For any nn positive integers x1,…,xnx_{1},\ldots,x_{n} in the range of [n][n], there exists a partition (S,T)(S,T) of [n][n] such that

min⁑{βˆ‘i∈Sxi,βˆ‘i∈Txi}β©Ύmβˆ’m2,Β where ​mβ‰œβˆ‘i∈[n]xi.\displaystyle\min\left\{\sum_{i\in S}x_{i},\sum_{i\in T}x_{i}\right\}\geqslant\frac{m-\sqrt{m}}{2},\text{ where }m\triangleq\sum_{i\in[n]}x_{i}. (2.5)

Moreover, such SS and TT can be found in polynomial time.

Proof.

The case of n=1n=1 is trivial because m=1m=1 and so mβˆ’m2=0\frac{m-\sqrt{m}}{2}=0; we thus assume that nβ©Ύ2n\geqslant 2. Denote 𝐱​(S)β‰œβˆ‘i∈Sxi\mathbf{\bm{x}}(S)\triangleq\sum_{i\in S}x_{i} for any SβŠ†[n]S\subseteq[n]. Consider the following naive greedy algorithm for Bin Packing:

{itembox}

[l]Greedy algorithm

1:sort xix_{i}’s in descending order.
2:define S(0)β‰œβˆ…S^{(0)}\triangleq\emptyset and T(0)β‰œβˆ…T^{(0)}\triangleq\emptyset.
3:forΒ i=1​to​ni=1\;\textbf{to}\;nΒ do
4:Β Β Β Β Β if 𝐱​(S(iβˆ’1))<𝐱​(T(iβˆ’1))\mathbf{\bm{x}}(S^{(i-1)})<\mathbf{\bm{x}}(T^{(i-1)})Β then
5:Β Β Β Β Β Β Β Β Β define S(i)β‰œS(iβˆ’1)βˆͺ{i}S^{(i)}\triangleq S^{(i-1)}\cup\{i\} and T(i)β‰œT(iβˆ’1)T^{(i)}\triangleq T^{(i-1)}.
6:Β Β Β Β Β else
7:Β Β Β Β Β Β Β Β Β define S(i)β‰œS(iβˆ’1)S^{(i)}\triangleq S^{(i-1)} and T(i)β‰œT(iβˆ’1)βˆͺ{i}T^{(i)}\triangleq T^{(i-1)}\cup\{i\}. Β Β Β Β Β 
8:return S(n)S^{(n)} and T(n)T^{(n)}.

We show

|𝐱​(S(n))βˆ’π±β€‹(T(n))|β©½m.\displaystyle\left|\mathbf{\bm{x}}(S^{(n)})-\mathbf{\bm{x}}(T^{(n)})\right|\leqslant\sqrt{m}. (2.6)

Define

dβ‰œmn​ andΒ β€‹Ξ΅β‰œdn=mn2.\displaystyle d\triangleq\frac{m}{n}\text{ and }\varepsilon\triangleq\sqrt{\frac{d}{n}}=\sqrt{\frac{m}{n^{2}}}. (2.7)

Since for every k∈[n]k\in[n], x1,…,xkx_{1},\ldots,x_{k} are at least xkx_{k} and xk+1,…,xnx_{k+1},\ldots,x_{n} are at least 11, we have

xkβ‹…k⏟contribution of ​x1,…,xk+1β‹…(nβˆ’k)⏟contribution of ​xk+1,…,xnβ©½m=n​d⟹k​(xkβˆ’1)β©½(dβˆ’1)​n⟹xkβ©½dβˆ’1k​n+1.\displaystyle\begin{aligned} &\underbrace{x_{k}\cdot k}_{\text{contribution of }x_{1},\ldots,x_{k}}+\underbrace{1\cdot(n-k)}_{\text{contribution of }x_{k+1},\ldots,x_{n}}\leqslant m=nd\\ &\implies k(x_{k}-1)\leqslant(d-1)n\\ &\implies x_{k}\leqslant\frac{d-1}{k}n+1.\end{aligned} (2.8)

Observing easily that

|𝐱​(S(βŒˆΞ΅β€‹nβŒ‰βˆ’1))βˆ’π±β€‹(T(βŒˆΞ΅β€‹nβŒ‰βˆ’1))|β©½n\displaystyle\left|\mathbf{\bm{x}}(S^{(\lceil\varepsilon n\rceil-1)})-\mathbf{\bm{x}}(T^{(\lceil\varepsilon n\rceil-1)})\right|\leqslant n (2.9)

due to the nature of the greedy algorithm, we consider the following two cases:

(Case 1)

If the absolute difference between 𝐱​(S(βŒˆΞ΅β€‹nβŒ‰βˆ’1))\mathbf{\bm{x}}(S^{(\lceil\varepsilon n\rceil-1)}) and 𝐱​(T(βŒˆΞ΅β€‹nβŒ‰βˆ’1))\mathbf{\bm{x}}(T^{(\lceil\varepsilon n\rceil-1)}) is larger than 𝐱​([βŒˆΞ΅β€‹nβŒ‰..n])\mathbf{\bm{x}}([\lceil\varepsilon n\rceil\mathrel{..}\nobreak n]), then we would have added all of [βŒˆΞ΅β€‹nβŒ‰..n][\lceil\varepsilon n\rceil\mathrel{..}\nobreak n] into either S(n)S^{(n)} or T(n)T^{(n)}. Thus, the absolute difference between 𝐱​(S(n))\mathbf{\bm{x}}(S^{(n)}) and 𝐱​(T(n))\mathbf{\bm{x}}(T^{(n)}) will be simply

|𝐱​(S(n))βˆ’π±β€‹(T(n))|β©½|𝐱​(S(βŒˆΞ΅β€‹nβŒ‰βˆ’1))βˆ’π±β€‹(T(βŒˆΞ΅β€‹nβŒ‰βˆ’1))|⏟⩽nβˆ’π±β€‹([βŒˆΞ΅β€‹nβŒ‰..n])⏟⩾nβˆ’βŒˆΞ΅β€‹nβŒ‰+1β©½βŒˆΞ΅β€‹nβŒ‰βˆ’1⩽Ρ​n.\displaystyle\left|\mathbf{\bm{x}}(S^{(n)})-\mathbf{\bm{x}}(T^{(n)})\right|\leqslant\underbrace{\left|\mathbf{\bm{x}}(S^{(\lceil\varepsilon n\rceil-1)})-\mathbf{\bm{x}}(T^{(\lceil\varepsilon n\rceil-1)})\right|}_{\leqslant n}-\underbrace{\mathbf{\bm{x}}([\lceil\varepsilon n\rceil\mathrel{..}\nobreak n])}_{\geqslant n-\lceil\varepsilon n\rceil+1}\leqslant\lceil\varepsilon n\rceil-1\leqslant\varepsilon n. (2.10)
(Case 2)

Otherwise, the absolute difference between 𝐱​(S(n))\mathbf{\bm{x}}(S^{(n)}) and 𝐱​(T(n))\mathbf{\bm{x}}(T^{(n)}) must be at most xβŒˆΞ΅β€‹nβŒ‰x_{\lceil\varepsilon n\rceil}. Using Eq.Β 2.8, we can derive

|𝐱​(S(n))βˆ’π±β€‹(T(n))|β©½xβŒˆΞ΅β€‹nβŒ‰β©½dβˆ’1βŒˆΞ΅β€‹nβŒ‰β€‹n+1β©½dβˆ’1Ξ΅+1β©½dΞ΅,\displaystyle\left|\mathbf{\bm{x}}(S^{(n)})-\mathbf{\bm{x}}(T^{(n)})\right|\leqslant x_{\lceil\varepsilon n\rceil}\leqslant\frac{d-1}{\lceil\varepsilon n\rceil}n+1\leqslant\frac{d-1}{\varepsilon}+1\leqslant\frac{d}{\varepsilon}, (2.11)

where the last inequality is due to the fact that Ρ∈(0,1]\varepsilon\in(0,1].

In either case, we obtain

|𝐱​(S(n))βˆ’π±β€‹(T(n))|β©½max⁑{Ρ​n,dn}=d​n=m.\displaystyle\left|\mathbf{\bm{x}}(S^{(n)})-\mathbf{\bm{x}}(T^{(n)})\right|\leqslant\max\left\{\varepsilon n,\frac{d}{n}\right\}=\sqrt{dn}=\sqrt{m}. (2.12)

Consequently, we derive

2β‹…min⁑{𝐱​(S(n)),𝐱​(T(n))}+mβ©Ύmin⁑{𝐱​(S(n)),𝐱​(T(n))}+max⁑{𝐱​(S(n)),𝐱​(T(n))}=m⟹min⁑{𝐱​(S(n)),𝐱​(T(n))}β©Ύmβˆ’m2.\displaystyle\begin{aligned} &2\cdot\min\Bigl{\{}\mathbf{\bm{x}}(S^{(n)}),\mathbf{\bm{x}}(T^{(n)})\Bigr{\}}+\sqrt{m}\geqslant\min\Bigl{\{}\mathbf{\bm{x}}(S^{(n)}),\mathbf{\bm{x}}(T^{(n)})\Bigr{\}}+\max\Bigl{\{}\mathbf{\bm{x}}(S^{(n)}),\mathbf{\bm{x}}(T^{(n)})\Bigr{\}}=m\\ \implies&\min\Bigl{\{}\mathbf{\bm{x}}(S^{(n)}),\mathbf{\bm{x}}(T^{(n)})\Bigr{\}}\geqslant\frac{m-\sqrt{m}}{2}.\end{aligned} (2.13)

completing the proof. ∎

We are now ready to prove TheoremΒ 2.1.

Proof of TheoremΒ 2.1.

By ClaimΒ 2.2, it suffices to show the existence of a desired partition of VV for each case. The statement for the case of large average degree directly follows from KΓΌhn and OsthusΒ [KO03, Corollary 18], stating that for every graph GG of mm edges and average degree 6Ξ΅\frac{6}{\varepsilon}, its vertex set can be partitioned into SS and TT such that both G​[S]G[S] and G​[T]G[T] contain at least m​(14βˆ’Ξ΅)m\left(\frac{1}{4}-\varepsilon\right) edges.

Suppose then that G=(X,Y,E,Ξ£,Ξ )G=(X,Y,E,\Sigma,\Pi) is a balanced game over mm edges such that |X|=|Y||X|=|Y|. Using LemmaΒ 2.3, we construct a pair of partitions (S,SΒ―)(S,\overline{S}) and (T,TΒ―)(T,\overline{T}) of XX and YY, respectively, such that

min⁑{βˆ‘x∈SdG​(x),βˆ‘x∈SΒ―dG​(x)}β©Ύmβˆ’m2​ and ​min⁑{βˆ‘x∈TdG​(x),βˆ‘x∈TΒ―dG​(x)}β©Ύmβˆ’m2,\displaystyle\min\left\{\sum_{x\in S}d_{G}(x),\sum_{x\in\overline{S}}d_{G}(x)\right\}\geqslant\frac{m-\sqrt{m}}{2}\text{ and }\min\left\{\sum_{x\in T}d_{G}(x),\sum_{x\in\overline{T}}d_{G}(x)\right\}\geqslant\frac{m-\sqrt{m}}{2}, (2.14)

where dG​(v)d_{G}(v) denotes the degree of vv in GG. Define ΞΈβ‰œmβˆ’m4\theta\triangleq\frac{m-\sqrt{m}}{4}. By case analysis, we show that either of (SβˆͺT,SΒ―βˆͺTΒ―)(S\cup T,\overline{S}\cup\overline{T}) or (SβˆͺTΒ―,SΒ―βˆͺT)(S\cup\overline{T},\overline{S}\cup T) gives us a desired partition.

(Case 1)

eG​(S,T)β©½ΞΈe_{G}(S,T)\leqslant\theta: eG​(S,TΒ―)e_{G}(S,\overline{T}) and eG​(SΒ―,T)e_{G}(\overline{S},T) can be bounded from below as follows:

eG​(S,TΒ―)\displaystyle e_{G}(S,\overline{T}) =eG​(S,TβˆͺTΒ―)βˆ’eG​(S,T)β©Ύβˆ‘x∈SdG​(x)βˆ’ΞΈβ©Ύmβˆ’m4,\displaystyle=e_{G}(S,T\cup\overline{T})-e_{G}(S,T)\geqslant\sum_{x\in S}d_{G}(x)-\theta\geqslant\frac{m-\sqrt{m}}{4}, (2.15)
eG​(SΒ―,T)\displaystyle e_{G}(\overline{S},T) =eG​(SβˆͺSΒ―,T)βˆ’eG​(SΒ―,TΒ―)β©Ύβˆ‘y∈TdG​(y)βˆ’ΞΈβ©Ύmβˆ’m4.\displaystyle=e_{G}(S\cup\overline{S},T)-e_{G}(\overline{S},\overline{T})\geqslant\sum_{y\in T}d_{G}(y)-\theta\geqslant\frac{m-\sqrt{m}}{4}. (2.16)
(Case 2)

eG​(SΒ―,TΒ―)β©½ΞΈe_{G}(\overline{S},\overline{T})\leqslant\theta: Similarly to (Case 1), we can bound

eG​(S,TΒ―)β©Ύmβˆ’m4​ and ​eG​(SΒ―,T)β©Ύmβˆ’m4.\displaystyle e_{G}(S,\overline{T})\geqslant\frac{m-\sqrt{m}}{4}\text{ and }e_{G}(\overline{S},T)\geqslant\frac{m-\sqrt{m}}{4}. (2.17)
(Case 3)

eG​(S,T)>ΞΈe_{G}(S,T)>\theta and eG​(SΒ―,TΒ―)>ΞΈe_{G}(\overline{S},\overline{T})>\theta: We are done since SβˆͺTS\cup T and SΒ―βˆͺTΒ―\overline{S}\cup\overline{T} have a desired property.

Consequently, in either case, we obtain

min⁑{eG​(S,T),eG​(SΒ―,TΒ―)}β©Ύmβˆ’m4​ or ​min⁑{eG​(S,TΒ―),eG​(SΒ―,T)}β©Ύmβˆ’m4,\displaystyle\min\Bigl{\{}e_{G}(S,T),e_{G}(\overline{S},\overline{T})\Bigr{\}}\geqslant\frac{m-\sqrt{m}}{4}\text{ or }\min\Bigl{\{}e_{G}(S,\overline{T}),e_{G}(\overline{S},T)\Bigr{\}}\geqslant\frac{m-\sqrt{m}}{4}, (2.18)

which completes the proof. ∎

3 Parallel Repetition Does Not Amplify Gap

Here, we show that a naive parallel repetition of Label Cover Reconfiguration does not amplify the gap unlike the parallel repetition theorem [Raz98]. We first define the product of two games.

Definition 3.1.

Let G1=(X1,Y1,E1,Ξ£1,Ξ 1=(Ο€1,e)e∈E1)G_{1}=(X_{1},Y_{1},E_{1},\Sigma_{1},\Pi_{1}=(\pi_{1,e})_{e\in E_{1}}) and G2=(X2,Y2,E2,Ξ£2,Ξ 2=(Ο€2,e)e∈E2)G_{2}=(X_{2},Y_{2},E_{2},\Sigma_{2},\Pi_{2}=(\pi_{2,e})_{e\in E_{2}}) be two games. Then, the product of G1G_{1} and G2G_{2}, denoted G1βŠ—G2G_{1}\otimes G_{2}, is defined as a new game (X1Γ—X2,Y1Γ—Y2,E,Ξ£1Γ—Ξ£2,Ξ =(Ο€e)e∈E)(X_{1}\times X_{2},Y_{1}\times Y_{2},E,\Sigma_{1}\times\Sigma_{2},\Pi=(\pi_{e})_{e\in E}), where

Eβ‰œ{((x1,x2),(y1,y2))|(x1,y1)∈E1,(x2,y2)∈E2},\displaystyle E\triangleq\left\{((x_{1},x_{2}),(y_{1},y_{2}))\Bigm{|}(x_{1},y_{1})\in E_{1},(x_{2},y_{2})\in E_{2}\right\}, (3.1)

and the constraint Ο€eβŠ†(Ξ£1Γ—Ξ£2)e\pi_{e}\subseteq(\Sigma_{1}\times\Sigma_{2})^{e} for each edge e=((x1,x2),(y1,y2))∈Ee=((x_{1},x_{2}),(y_{1},y_{2}))\in E is defined as

Ο€eβ‰œ{((Ξ±1,Ξ±2),(Ξ²1,Ξ²2))|(Ξ±1,Ξ²1)βˆˆΟ€1,(x1,y1)​ and ​(Ξ±2,Ξ²2)βˆˆΟ€2,(x2,y2)}.\displaystyle\pi_{e}\triangleq\left\{((\alpha_{1},\alpha_{2}),(\beta_{1},\beta_{2}))\Bigm{|}(\alpha_{1},\beta_{1})\in\pi_{1,(x_{1},y_{1})}\text{ and }(\alpha_{2},\beta_{2})\in\pi_{2,(x_{2},y_{2})}\right\}. (3.2)

The ρ\rho-fold parallel repetition of GG, denoted GβŠ—ΟG^{\otimes\rho}, for any positive integer Οβˆˆβ„•\rho\in\mathbb{N} is defined as

GβŠ—Οβ‰œGβŠ—β‹―βŠ—GβŸΟβ€‹Β times.\displaystyle G^{\otimes\rho}\triangleq\underbrace{G\otimes\cdots\otimes G}_{\rho\text{ times}}. (3.3)

The parallel repetition theorem due to RazΒ [Raz98] states that for every game GG with 𝗏𝖺𝗅⁑(G)β©½1βˆ’Ξ΅<1\operatorname{\mathsf{val}}(G)\leqslant 1-\varepsilon<1, it holds that 𝗏𝖺𝗅⁑(GβŠ—Ο)β©½(1βˆ’Ξ΅Β―)ρlog⁑|Ξ£|\operatorname{\mathsf{val}}(G^{\otimes\rho})\leqslant(1-\overline{\varepsilon})^{\frac{\rho}{\log|\Sigma|}}, where Ρ¯>0\overline{\varepsilon}>0 depends only on the value of Ξ΅\varepsilon.

We can think of a reconfiguration analogue for parallel repetition. More precisely, for two labelings ψ1:(X1βˆͺY1)β†’Ξ£1\psi_{1}\colon(X_{1}\cup Y_{1})\to\Sigma_{1} for G1=(X1,Y1,E1,Ξ£1,Ξ 1)G_{1}=(X_{1},Y_{1},E_{1},\Sigma_{1},\Pi_{1}) and ψ2:(X2βˆͺY2)β†’Ξ£2\psi_{2}\colon(X_{2}\cup Y_{2})\to\Sigma_{2} for G2=(X2,Y2,E2,Ξ£2,Ξ 2)G_{2}=(X_{2},Y_{2},E_{2},\Sigma_{2},\Pi_{2}), the product labeling of ψ1\psi_{1} and ψ2\psi_{2}, denoted ψ1βŠ—Οˆ2\psi_{1}\otimes\psi_{2}, is defined as a labeling ψ:(X1Γ—X2)βˆͺ(Y1Γ—Y2)β†’Ξ£1Γ—Ξ£2\psi\colon(X_{1}\times X_{2})\cup(Y_{1}\times Y_{2})\to\Sigma_{1}\times\Sigma_{2} such that

Οˆβ€‹(v1,v2)β‰œ(ψ1​(v1),ψ2​(v2))​ for all ​(v1,v2)∈(X1Γ—X2)βˆͺ(Y1Γ—Y2).\displaystyle\psi(v_{1},v_{2})\triangleq(\psi_{1}(v_{1}),\psi_{2}(v_{2}))\text{ for all }(v_{1},v_{2})\in(X_{1}\times X_{2})\cup(Y_{1}\times Y_{2}). (3.4)

We write ΟˆβŠ—Ο\psi^{\otimes\rho} for denoting ΟˆβŠ—β‹―βŠ—ΟˆβŸΟβ€‹Β times.\underbrace{\psi\otimes\cdots\otimes\psi}_{\rho\text{ times}}. It is easy to see that 𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)=1\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})=1 implies 𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—Οβ†­Οˆπ—βŠ—Ο)=1\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes\rho}\leftrightsquigarrow\psi_{\mathsf{t}}^{\otimes\rho})=1. So, given that 𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)<1\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})<1, does the value of the ρ\rho-fold parallel repetition of a game decreases as the increase of ρ\rho? Unfortunately, the answer is negative:

Observation 3.2.

Let GG be a satisfiable game, and Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}} be its two satisfying labelings. Then, for every positive integer Οβˆˆβ„•\rho\in\mathbb{N}, it holds that

𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—Οβ†­Οˆπ—βŠ—Ο)⩾𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—).\displaystyle\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes\rho}\leftrightsquigarrow\psi_{\mathsf{t}}^{\otimes\rho})\geqslant\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}}). (3.5)
Proof.

Let G=(X,Y,E,Ξ£,Ξ )G=(X,Y,E,\Sigma,\Pi) be a satisfiable game. Suppose that we are given a reconfiguration sequence \uppsi=⟨ψ(0)​…,ψ(β„“)⟩\mathbf{\bm{\uppsi}}=\langle\psi^{(0)}\ldots,\psi^{(\ell)}\rangle from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}} such that 𝗏𝖺𝗅G⁑(\uppsi)=𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)\operatorname{\mathsf{val}}_{G}(\mathbf{\bm{\uppsi}})=\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}}). We will show

𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—kβŠ—Οˆπ—βŠ—Οβˆ’kβ†­Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆπ—βŠ—Οβˆ’k+1)⩾𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—)​ for all ​k∈[ρ].\displaystyle\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes k}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k}\leftrightsquigarrow\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k+1})\geqslant\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}})\text{ for all }k\in[\rho]. (3.6)

To this end, for each i∈[β„“]i\in[\ell], we bound

𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(iβˆ’1)βŠ—Οˆπ—βŠ—Οβˆ’kβ†­Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(i)βŠ—Οˆπ—βŠ—Οβˆ’k).\displaystyle\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i-1)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k}\leftrightsquigarrow\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k}). (3.7)

Fix k∈[ρ]k\in[\rho] and i∈[β„“]i\in[\ell]. We assume that ψ(iβˆ’1)\psi^{(i-1)} and ψ(i)\psi^{(i)} differ in some xβ‹†βˆˆXx^{\star}\in X; the case that they differ in some yβ‹†βˆˆYy^{\star}\in Y can be shown similarly. Construct then a reconfiguration sequence \uppsi(k,i)\mathbf{\bm{\uppsi}}^{(k,i)} from Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(iβˆ’1)βŠ—Οˆπ—βŠ—Οβˆ’k\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i-1)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k} to Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(i)βŠ—Οˆπ—βŠ—Οβˆ’k\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k} by changing the value of a vertex of GβŠ—β„“G^{\otimes\ell} if its value is different between the two assignments. Obviously, it holds that

𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(iβˆ’1)βŠ—Οˆπ—βŠ—Οβˆ’kβ†­Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(i)βŠ—Οˆπ—βŠ—Οβˆ’k)⩾𝗏𝖺𝗅GβŠ—β„“β‘(\uppsi(k,i)).\displaystyle\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i-1)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k}\leftrightsquigarrow\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k})\geqslant\operatorname{\mathsf{val}}_{G^{\otimes\ell}}(\mathbf{\bm{\uppsi}}^{(k,i)}). (3.8)

Observe now that Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(iβˆ’1)βŠ—Οˆπ—βŠ—Οβˆ’k\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i-1)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k} and Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(i)βŠ—Οˆπ—βŠ—Οβˆ’k\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k} differ on vertex (x1,…,xρ)∈Xρ(x_{1},\ldots,x_{\rho})\in X^{\rho} if and only if xk=x⋆x_{k}=x^{\star} and agree with each other on any vertex (y1,…,yρ)∈Yρ(y_{1},\ldots,y_{\rho})\in Y^{\rho}. Thus, for any assignment ψ∈\uppsi(k,i)\psi\in\mathbf{\bm{\uppsi}}^{(k,i)},

Οˆβ€‹(x1,…,xρ)\displaystyle\psi(x_{1},\ldots,x_{\rho}) =(Οˆπ—Œβ€‹(x1),…,Οˆπ—Œβ€‹(xkβˆ’1),{ψ(iβˆ’1)​(xk)orψ(i)​(xk)},Οˆπ—β€‹(xk+1),…,Οˆπ—β€‹(xρ))\displaystyle=(\psi_{\mathsf{s}}(x_{1}),\ldots,\psi_{\mathsf{s}}(x_{k-1}),\left\{\begin{array}[]{c}\psi^{(i-1)}(x_{k})\\ \text{or}\\ \psi^{(i)}(x_{k})\end{array}\right\},\psi_{\mathsf{t}}(x_{k+1}),\ldots,\psi_{\mathsf{t}}(x_{\rho}))
Οˆβ€‹(y1,…,yρ)\displaystyle\psi(y_{1},\ldots,y_{\rho}) =(Οˆπ—Œβ€‹(y1),…,Οˆπ—Œβ€‹(ykβˆ’1),ψ(iβˆ’1)​(yk),Οˆπ—β€‹(yk+1),…,Οˆπ—β€‹(yρ))\displaystyle=(\psi_{\mathsf{s}}(y_{1}),\ldots,\psi_{\mathsf{s}}(y_{k-1}),\psi^{(i-1)}(y_{k}),\psi_{\mathsf{t}}(y_{k+1}),\ldots,\psi_{\mathsf{t}}(y_{\rho}))
=(Οˆπ—Œβ€‹(y1),…,Οˆπ—Œβ€‹(ykβˆ’1),ψ(i)​(yk),Οˆπ—β€‹(yk+1),…,Οˆπ—β€‹(yρ)).\displaystyle=(\psi_{\mathsf{s}}(y_{1}),\ldots,\psi_{\mathsf{s}}(y_{k-1}),\psi^{(i)}(y_{k}),\psi_{\mathsf{t}}(y_{k+1}),\ldots,\psi_{\mathsf{t}}(y_{\rho})).

for all (x1,…,xρ)∈Xρ(x_{1},\ldots,x_{\rho})\in X^{\rho} and (y1,…,yρ)∈Yρ(y_{1},\ldots,y_{\rho})\in Y^{\rho}. By abuse of notation, denoting Οˆβ€‹(x,y)=(Οˆβ€‹(x),Οˆβ€‹(y))\psi(x,y)=(\psi(x),\psi(y)) for assignment ψ\psi and edge (x,y)(x,y); we bound the value of \uppsi(k,i)\mathbf{\bm{\uppsi}}^{(k,i)} as follows:

𝗏𝖺𝗅GβŠ—β„“β‘(\uppsi(k,i))\displaystyle\operatorname{\mathsf{val}}_{G^{\otimes\ell}}(\mathbf{\bm{\uppsi}}^{(k,i)})
β©Ύ\displaystyle\geqslant{} 1|E​(GβŠ—Ο)|β€‹βˆ‘x1,…,xρ∈X\displaystyle\frac{1}{|E(G^{\otimes\rho})|}\sum_{x_{1},\ldots,x_{\rho}\in X}
min{βˆ‘y1,…,yρ∈Y:(xj,yj)∈Eβ€‹βˆ€j\llbracketβ‹€1β©½jβ©½kβˆ’1Οˆπ—Œ(xj,yj)βˆˆΟ€(xj,yj)\rrbracketβ‹…\llbracketψ(iβˆ’1)(xk,yk)βˆˆΟ€(xk,yk)\rrbracketβ‹…\llbracketβ‹€k+1β©½jβ©½β„“Οˆπ—(xj,yj)βˆˆΟ€(xj,yj)\rrbracket,\displaystyle\min\left\{\sum_{\begin{subarray}{c}y_{1},\ldots,y_{\rho}\in Y:\\ (x_{j},y_{j})\in E\forall j\end{subarray}}\Bigl{\llbracket}\bigwedge_{1\leqslant j\leqslant k-1}\psi_{\mathsf{s}}(x_{j},y_{j})\in\pi_{(x_{j},y_{j})}\Bigr{\rrbracket}\cdot\Bigl{\llbracket}\psi^{(i-1)}(x_{k},y_{k})\in\pi_{(x_{k},y_{k})}\Bigr{\rrbracket}\cdot\Bigl{\llbracket}\bigwedge_{k+1\leqslant j\leqslant\ell}\psi_{\mathsf{t}}(x_{j},y_{j})\in\pi_{(x_{j},y_{j})}\Bigr{\rrbracket},\right.
βˆ‘y1,…,yρ∈Y:(xj,yj)∈Eβ€‹βˆ€j\llbracketβ‹€1β©½jβ©½kβˆ’1Οˆπ—Œ(xj,yj)βˆˆΟ€(xj,yj)\rrbracketβ‹…\llbracketψ(i)(xk,yk)βˆˆΟ€(xk,yk)\rrbracketβ‹…\llbracketβ‹€k+1β©½jβ©½β„“Οˆπ—(xj,yj)βˆˆΟ€(xj,yj)\rrbracket}\displaystyle\left.\sum_{\begin{subarray}{c}y_{1},\ldots,y_{\rho}\in Y:\\ (x_{j},y_{j})\in E\forall j\end{subarray}}\Bigl{\llbracket}\bigwedge_{1\leqslant j\leqslant k-1}\psi_{\mathsf{s}}(x_{j},y_{j})\in\pi_{(x_{j},y_{j})}\Bigr{\rrbracket}\cdot\Bigl{\llbracket}\psi^{(i)}(x_{k},y_{k})\in\pi_{(x_{k},y_{k})}\Bigr{\rrbracket}\cdot\Bigl{\llbracket}\bigwedge_{k+1\leqslant j\leqslant\ell}\psi_{\mathsf{t}}(x_{j},y_{j})\in\pi_{(x_{j},y_{j})}\Bigr{\rrbracket}\right\}

Since Οˆπ—Œβ€‹(xj,yj)\psi_{\mathsf{s}}(x_{j},y_{j}) satisfies (xj,yj)(x_{j},y_{j}) if 1β©½jβ©½kβˆ’11\leqslant j\leqslant k-1 and Οˆπ—β€‹(xj,yj)\psi_{\mathsf{t}}(x_{j},y_{j}) satisfies (xj,yj)(x_{j},y_{j}) if k+1β©½jβ©½β„“k+1\leqslant j\leqslant\ell, we further have

𝗏𝖺𝗅GβŠ—β„“β‘(\uppsi(k,i))\displaystyle\operatorname{\mathsf{val}}_{G^{\otimes\ell}}(\mathbf{\bm{\uppsi}}^{(k,i)})
β©Ύ\displaystyle\geqslant{} 1|E​(GβŠ—Ο)|β€‹βˆ‘x1,…,xρmin⁑{βˆ‘y1,…,yρ\llbracketβ€‹Οˆ(iβˆ’1)​(xk,yk)βˆˆΟ€(xk,yk)​\rrbracket,βˆ‘y1,…,yρ\llbracketβ€‹Οˆ(i)​(xk,yk)βˆˆΟ€(xk,yk)​\rrbracket}\displaystyle\frac{1}{|E(G^{\otimes\rho})|}\sum_{x_{1},\ldots,x_{\rho}}\min\left\{\sum_{y_{1},\ldots,y_{\rho}}\left\llbracket\psi^{(i-1)}(x_{k},y_{k})\in\pi_{(x_{k},y_{k})}\right\rrbracket,\sum_{y_{1},\ldots,y_{\rho}}\left\llbracket\psi^{(i)}(x_{k},y_{k})\in\pi_{(x_{k},y_{k})}\right\rrbracket\right\}
=\displaystyle={} 1|E​(GβŠ—Ο)|β€‹βˆ‘e1,…,ekβˆ’1,ek+1,…,eρ∈E\displaystyle\frac{1}{|E(G^{\otimes\rho})|}\sum_{e_{1},\ldots,e_{k-1},e_{k+1},\ldots,e_{\rho}\in E}
βˆ‘xkmin⁑{βˆ‘yk:(xk,yk)∈E\llbracketβ€‹Οˆ(iβˆ’1)​(xk,yk)βˆˆΟ€(xk,yk)​\rrbracket,βˆ‘yk:(xk,yk)∈E\llbracketβ€‹Οˆ(i)​(xk,yk)βˆˆΟ€(xk,yk)​\rrbracket}\displaystyle\sum_{x_{k}}\min\left\{\sum_{y_{k}:(x_{k},y_{k})\in E}\left\llbracket\psi^{(i-1)}(x_{k},y_{k})\in\pi_{(x_{k},y_{k})}\right\rrbracket,\sum_{y_{k}:(x_{k},y_{k})\in E}\left\llbracket\psi^{(i)}(x_{k},y_{k})\in\pi_{(x_{k},y_{k})}\right\rrbracket\right\}
=\displaystyle={} 1|E|β€‹βˆ‘xmin⁑{βˆ‘y:(x,y)∈E\llbracketβ€‹Οˆ(iβˆ’1)​(x,y)βˆˆΟ€(x,y)​\rrbracket,βˆ‘y:(x,y)∈E\llbracketβ€‹Οˆ(i)​(x,y)βˆˆΟ€(x,y)​\rrbracket}\displaystyle\frac{1}{|E|}\sum_{x}\min\left\{\sum_{y:(x,y)\in E}\left\llbracket\psi^{(i-1)}(x,y)\in\pi_{(x,y)}\right\rrbracket,\sum_{y:(x,y)\in E}\left\llbracket\psi^{(i)}(x,y)\in\pi_{(x,y)}\right\rrbracket\right\}
=\displaystyle={} min⁑{𝗏𝖺𝗅G⁑(ψ(iβˆ’1)),𝗏𝖺𝗅G⁑(ψ(i))}.\displaystyle\min\Bigl{\{}\operatorname{\mathsf{val}}_{G}(\psi^{(i-1)}),\operatorname{\mathsf{val}}_{G}(\psi^{(i)})\Bigr{\}}.

Concatenating reconfiguration sequences \uppsi(k,i)\mathbf{\bm{\uppsi}}^{(k,i)} for all i∈[β„“]i\in[\ell], we have

𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—kβŠ—Οˆπ—βŠ—Οβˆ’kβ†­Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆπ—βŠ—Οβˆ’k+1)β©Ύmin1β©½i⩽ℓ⁑𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(iβˆ’1)βŠ—Οˆπ—βŠ—Οβˆ’kβ†­Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆ(i)βŠ—Οˆπ—βŠ—Οβˆ’k)β©Ύmin0β©½i⩽ℓ⁑𝗏𝖺𝗅G⁑(ψ(i))=𝗏𝖺𝗅G⁑(\uppsi)=𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—).\displaystyle\begin{aligned} &\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes k}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k}\leftrightsquigarrow\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k+1})\\ &\geqslant\min_{1\leqslant i\leqslant\ell}\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i-1)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k}\leftrightsquigarrow\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi^{(i)}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k})\\ &\geqslant\min_{0\leqslant i\leqslant\ell}\operatorname{\mathsf{val}}_{G}(\psi^{(i)})\\ &=\operatorname{\mathsf{val}}_{G}(\mathbf{\bm{\uppsi}})=\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}}).\end{aligned} (3.9)

Consequently, we obtain

𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—Οβ†­Οˆπ—βŠ—Ο)β©Ύmink∈[ρ]⁑𝗏𝖺𝗅GβŠ—Οβ‘(Οˆπ—ŒβŠ—kβŠ—Οˆπ—βŠ—Οβˆ’kβ†­Οˆπ—ŒβŠ—kβˆ’1βŠ—Οˆπ—βŠ—Οβˆ’k+1)⩾𝗏𝖺𝗅G⁑(Οˆπ—Œβ†­Οˆπ—),\displaystyle\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes\rho}\leftrightsquigarrow\psi_{\mathsf{t}}^{\otimes\rho})\geqslant\min_{k\in[\rho]}\operatorname{\mathsf{val}}_{G^{\otimes\rho}}(\psi_{\mathsf{s}}^{\otimes k}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k}\leftrightsquigarrow\psi_{\mathsf{s}}^{\otimes k-1}\otimes\psi_{\mathsf{t}}^{\otimes\rho-k+1})\geqslant\operatorname{\mathsf{val}}_{G}(\psi_{\mathsf{s}}\leftrightsquigarrow\psi_{\mathsf{t}}), (3.10)

completing the proof. ∎

4 Polynomial-time Reconfiguration for Projection Games

We finally present a simple polynomial-time algorithm that decides reconfigurability between a pair of satisfying labelings for a projection game.

Theorem 4.1.

For a satisfiable projection game G=(X,Y,E,Ξ£,Ξ )G=(X,Y,E,\Sigma,\Pi) and its two satisfying labelings Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}}, we can decide in polynomial time if there is a reconfiguration sequence from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}} of satisfying labelings for GG.

Proof.

Since each connected component of GG can be processed independently, we only consider the case that GG is connected. If GG consists of a single vertex (i.e., XX or YY is empty), the answer is always β€œreconfigurable” as it has no edges; thus we can safely assume that X,Yβ‰ βˆ…X,Y\neq\emptyset. We show that Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}} are reconfigurable to each other if and only if Οˆπ—Œ\psi_{\mathsf{s}} and Οˆπ—\psi_{\mathsf{t}} agree on XX.

We first show the only-if direction. Suppose that Οˆπ—Œβ€‹(x)β‰ Οˆπ—β€‹(x)\psi_{\mathsf{s}}(x)\neq\psi_{\mathsf{t}}(x) for some x∈Xx\in X. Starting from Οˆπ—Œ\psi_{\mathsf{s}}, changing the value of xx from Οˆπ—Œβ€‹(x)\psi_{\mathsf{s}}(x) to any other value must violate edges e=(x,y)e=(x,y) incident to xx since the value of yy should be mapped to Οˆπ—Œβ€‹(x)\psi_{\mathsf{s}}(x) via Ο€e\pi_{e}, as claimed.

We then show the if direction. Suppose that Οˆπ—Œβ€‹(x)=Οˆπ—β€‹(x)\psi_{\mathsf{s}}(x)=\psi_{\mathsf{t}}(x) for every x∈Xx\in X. Then, it holds that Ο€e​(Οˆπ—Œβ€‹(y))=Οˆπ—Œβ€‹(x)=Οˆπ—β€‹(x)=Ο€e​(Οˆπ—β€‹(y))\pi_{e}(\psi_{\mathsf{s}}(y))=\psi_{\mathsf{s}}(x)=\psi_{\mathsf{t}}(x)=\pi_{e}(\psi_{\mathsf{t}}(y)) for every edge e=(x,y)∈Ee=(x,y)\in E. We can thus obtain a transformation from Οˆπ—Œ\psi_{\mathsf{s}} to Οˆπ—\psi_{\mathsf{t}} without breaking any constraint, by changing the value of each vertex yy of YY from Οˆπ—Œβ€‹(y)\psi_{\mathsf{s}}(y) to Οˆπ—β€‹(y)\psi_{\mathsf{t}}(y) one by one, as desired. ∎

References

  • [ABSS97] Sanjeev Arora, LΓ‘szlΓ³ Babai, Jacques Stern, and Z.Β Sweedyk. The hardness of approximate optima in lattices, codes, and systems of linear equations. J. Comput. Syst. Sci., 54(2):317–331, 1997.
  • [ALM+98] Sanjeev Arora, Carsten Lund, Rajeev Motwani, Madhu Sudan, and Mario Szegedy. Proof verification and the hardness of approximation problems. J. ACM, 45(3):501–555, 1998.
  • [AS98] Sanjeev Arora and Shmuel Safra. Probabilistic checking of proofs: A new characterization of NP. J. ACM, 45(1):70–122, 1998.
  • [Fei98] Uriel Feige. A threshold of ln⁑n\ln n for approximating set cover. J. ACM, 45(4):634–652, 1998.
  • [GKMP09] Parikshit Gopalan, PhokionΒ G. Kolaitis, Elitza Maneva, and ChristosΒ H. Papadimitriou. The connectivity of Boolean satisfiability: Computational and structural dichotomies. SIAM J. Comput., 38(6):2330–2355, 2009.
  • [HΓ₯s99] Johan HΓ₯stad. Clique is hard to approximate within n1βˆ’Ξ΅n^{1-\varepsilon}. Acta Math., 182:105–142, 1999.
  • [HΓ₯s01] Johan HΓ₯stad. Some optimal inapproximability results. J. ACM, 48(4):798–859, 2001.
  • [IDH+11] Takehiro Ito, ErikΒ D. Demaine, Nicholas J.Β A. Harvey, ChristosΒ H. Papadimitriou, Martha Sideri, Ryuhei Uehara, and Yushi Uno. On the complexity of reconfiguration problems. Theor. Comput. Sci., 412(12-14):1054–1065, 2011.
  • [KO03] Daniela KΓΌhn and Deryk Osthus. Partitions of graphs with high minimum degree or connectivity. J. Comb. Theory. Ser. B, 88(1):29–43, 2003.
  • [Nis18] Naomi Nishimura. Introduction to reconfiguration. Algorithms, 11(4):52, 2018.
  • [Ohs23] Naoto Ohsaka. Gap preserving reductions between reconfiguration problems. In STACS, pages 49:1–49:18, 2023.
  • [Raz98] Ran Raz. A parallel repetition theorem. SIAM J. Comput., 27(3):763–803, 1998.
  • [vdH13] Jan vanΒ den Heuvel. The complexity of change. In Surveys in Combinatorics 2013, volume 409, pages 127–160. Cambridge University Press, 2013.