No distributed quantum advantage for approximate graph coloring

. We give an almost complete characterization of the hardness of c -coloring χ -chromatic graphs with distributed algorithms, for a wide range of models of distributed computing. In particular, we show that these problems do not admit any distributed quantum advantage. To do that:

1 α ) rounds, with α = c−1 χ−1 .2. We prove that any distributed algorithm for this problem requires Ω(n 1 α ) rounds.Our upper bound holds in the classical, deterministic LOCAL model, while the near-matching lower bound holds in the non-signaling model.This model, introduced by Arfaoui and Fraigniaud in 2014, captures all models of distributed graph algorithms that obey physical causality; this includes not only classical deterministic LOCAL and randomized LOCAL but also quantum-LOCAL, even with a pre-shared quantum state.
We also show that similar arguments can be used to prove that, e.g., 3-coloring 2-dimensional grids or c-coloring trees remain hard problems even for the non-signaling model, and in particular do not admit any quantum advantage.Our lower-bound arguments are purely graph-theoretic at heart; no background on quantum information theory is needed to establish the proofs.

Introduction
In this work, we settle the distributed computational complexity of approximate graph coloring, for deterministic, randomized, and quantum versions of the LOCAL model of distributed computing.
In brief, the setting is this: We have an input graph G with n nodes.Each node is a computer and each edge represents a communication link.Computation proceeds in synchronous rounds: each node sends a message to each of its neighbors, receives a message from each of its neighbors, and updates its own state.After T rounds, each node has to stop and announce its own output, and the outputs have to form a proper c-coloring of the input graph G.If the chromatic number of G is χ, in this setting it is trivial to find a χ-coloring in T = O(n) rounds, as in O(n) rounds all nodes can learn the full topology of their own connected component and they can locally find an optimal coloring by brute force without any further communication.But the key questions are: How well can we color graphs in T ≪ n rounds?And how much does it help if we use quantum computers that can exchange quantum information, possibly with a pre-shared entangled state?

Main result
We show that for all constants c, χ, and α, it is possible to find a c-coloring of a χ-colorable graph in T = Õ(n 1/α ) communication rounds if and only if For example, if the graph is bipartite (χ = 2), this means that the complexity of 2-coloring is Θ(n) rounds, 3-coloring is Θ( √ n) rounds, and 4-coloring is Θ(n 1/3 ) rounds.Here we use Õ and Θ to hide polylogarithmic factors, that is, our results are tight up to polylogarithmic factors.
Perhaps the biggest surprise is that this result holds for a wide range of models of distributed computing: the answer is the same for deterministic, randomized, and quantum versions of the LOCAL model, and it holds even if the algorithm has access to shared randomness or pre-shared quantum state (as long as the quantum state is prepared before we reveal the structure of graph G).
In particular, we show that there is no distributed quantum advantage for approximate graph coloring in the context of the LOCAL model, at least up to polylogarithmic factors.

Significance and motivation
Our work is directly linked to two lines of research: understanding the quantum advantage in distributed settings, and the complexity of distributed graph coloring in classical settings.

Distributed quantum advantage.
There is a long line of work [4,21,27,42,43,46,54,65,67] on quantum advantage in the CONGEST model-this is a bandwidth-limited version of the LOCAL model.However, much less is known about quantum advantage in the LOCAL model.
Earlier work by Gavoille et al. [33] and Arfaoui and Fraigniaud [6] on quantum-LOCAL brought primarily bad news: they showed that many classical LOCAL model lower bounds still hold in the quantum-LOCAL model.The quantum advantage demonstrated by [33] was limited to constant factors or required pre-shared quantum resources.The major breakthrough was the recent work by Le Gall et al. [47] that demonstrated that there is a problem that can be solved in only 2 rounds using quantum communication, whereas solving it in the classical setting requires Ω(n) rounds.
However, the problem from Le Gall et al. [47] is very different from the classical problems commonly studied in the field of distributed graph algorithms, and most importantly, it is not a 1 locally checkable problem.Locally checkable problems are graph problems in which the task is to find a feasible solution subject to local constraints-perhaps the best-known example of such a problem is graph coloring.A lot of recent work on the classical LOCAL model has focused on locally checkable problems, and there is nowadays a solid understanding of the landscape of the distributed computational complexity of such problems for the classical models-see, e.g., [8,9,10,11,18,19,22,24,30,35,36,61].However, what is wide open is how quantum-LOCAL changes the picture.
A major open problem is whether there is any locally checkable graph problem that can be solved asymptotically faster in quantum-LOCAL in comparison with the classical randomized LOCAL model, and it has been conjectured that no such problem exists [63].In this work we provide more evidence in support of this conjecture: we show that various problems related to graph coloring do not admit any significant quantum advantage.
Hardness of distributed coloring.In a very recent work, Akbari et al. [2] studied the notion of locality in three different settings: distributed, dynamic, and online graph algorithms.They showed that for locally checkable problems in rooted regular trees the three notions of locality coincide, but more generally the notions are distinct.The prime example of a problem that separates the models is 3-coloring bipartite graphs: the distributed locality (i.e., round complexity) of the problem is Ω( √ n) [20], but the online locality is O(log n) [2].While this demonstrates that there is large gap between distributed and online settings, this also highlights a blind spot in our understanding of seemingly elementary questions in the classical LOCAL model: What, exactly, is the distributed complexity of 3-coloring bipartite graphs?Can we solve it in Õ( √ n) rounds?And, more generally, what is the distributed complexity of c-coloring χ-colorable graphs?
Given the prominent role graph coloring plays in distributed graph algorithms, the state of the art is highly unsatisfactory-the upper and lower bounds are far from each other, even if we consider the seemingly elementary question of coloring bipartite graphs: • As mentioned above, the complexity of 3-coloring bipartite graphs is known to be somewhere between Ω( √ n) and O(n).Brandt et al. [20] show that 3-coloring 2-dimensional grids requires Ω( √ n) rounds, and even though they study toroidal grids (which are not necessarily bipartite), the same result can be adapted to also show that 3-coloring bipartite graphs requires Ω( √ n) rounds.It is not known if this is tight; to the best of our knowledge, there is no upper bound other than the trivial O(n)-round algorithm.
• The complexity of 4-coloring bipartite graphs is only known to be somewhere between Ω(log n) and O(n).Linial's [49] lower bound for coloring trees applies, so we know that the complexity has to be at least Ω(log n), but beyond that very little is known.The lower bound construction from [20] cannot be used here since it is easy to 4-color grids.To come up with a nontrivial upper bound, it would be tempting to use network decompositions in the spirit of Barenboim [12], but we are lacking network decomposition algorithms with suitable parameters, and in any case this approach cannot produce 4-colorings or 5-colorings in o( √ n) rounds.
In this work we solve all these open questions, up to polylogarithmic factors, for the general task of c-coloring χ-colorable graphs.We show that there is plenty of room for improvement in both upper and lower bounds.For example, in the case of 4-coloring bipartite graphs, the right bound turns out to be Θ(n 1/3 ), which is far from what can be achieved with the state of the art outlined above.

Contributions in more detail
We will now describe all of our results and contributions in more detail; we refer to Section 2 for an overview of the proof ideas and to Sections 4 and 5 for the proofs.

Classical upper bound (Section 4)
Let us start with the upper bound.We design new distributed algorithms with the following properties: Theorem 1.1.There exists a det-LOCAL algorithm A det and a rand-LOCAL algorithm A rand that, given a parameter α ∈ N, find a proper vertex coloring with α(χ − 1) + 1 colors in any graph with chromatic number χ, as follows: • A det runs in O n 1/α log 3−1/α n • (log log n) O(1) rounds.
We note that the algorithms do not need to know χ; it is sufficient to know α and n.As a corollary, we can, e.g., 3-color bipartite graphs in Õ( √ n) rounds by setting α = 2.The number of colors c = α(χ − 1) + 1 may look like a rather unnatural expression, and there does not seem to be a priori any reason to expect that this would be tight-however, as we will see, this is indeed exactly the right number.

Non-signaling model
Our main goal is to show that the algorithms in Theorem 1.1 are optimal (up to polylogarithmic factors), not only in the classical models but also in, e.g., all reasonable variants of the quantum-LOCAL model.To this end, we work in the non-signaling model, as defined by Arfaoui and Fraigniaud [6]; this is essentially equivalent to the φ-LOCAL model defined earlier by Gavoille et al. [33].
The non-signaling model is a characterization of output distributions that do not violate the nosignaling from the future principle or, equivalently, the causality principle [25].To better understand this, suppose we have some classical rand-LOCAL algorithm A that runs in T rounds and outputs a vertex coloring.Let p(G) be the output distribution of A when run on a graph G.The key observation is that this distribution is not arbitrary-in particular, it must satisfy the following property:

Definition 1.2 (Non-signaling distribution, informal version). The output distribution p(G) of
A is non-signaling beyond distance T if the following holds: Let V be a set of nodes of a graph Put otherwise, changes more than distance T away from U cannot influence the output distribution of U .It is not hard to see that this holds for det-LOCAL and rand-LOCAL, even if the algorithm has access to shared randomness.But what makes this notion particularly useful is that it is satisfied also by the quantum-LOCAL model, even with a pre-shared quantum state [6,33].Informally, a system that violates the non-signaling property would violate causality and enable faster-than-light communication, which is something that quantum physics does not allow.
We use NS-LOCAL to refer to the non-signaling model.We say that A is an NS-LOCAL algorithm that runs in T rounds if it produces an output distribution that is non-signaling beyond distance T .We will then prove statements of the form "any NS-LOCAL algorithm for this problem requires at least T rounds."As a corollary, this gives a T -round lower bound for det-LOCAL, rand-LOCAL, and quantum-LOCAL, even if we have access to shared randomness and pre-shared quantum states.This also puts limits on the existence of so-called finitely dependent colorings [39,40].

Non-signaling lower bounds (Section 5)
The precise version of our lower bound result states that, for every large enough number of nodes n, there exists a χ-chromatic graph on n nodes that is hard to color in NS-LOCAL.
Suppose A is an NS-LOCAL algorithm for c-coloring graphs in the family F of χ-chromatic graphs of n nodes with success probability q > ε.Then the running time of A is at least A key observation is that, if the parameters χ, c, and ε in Theorem 1.3 are constants, then T = Ω(n 1 α ), which matches the upper bound in Theorem 1.1 up to polylogarithmic factors.In particular, there is at best polylogarithmic room for any distributed quantum advantage.
While Theorem 1.3 implies bounds for coloring bipartite graphs in general, we will also use our techniques to prove bounds for specific bipartite graphs.By prior work, it is known that 3-coloring 2-dimensional grids is hard in the det-LOCAL model [1,20].We show that this also holds for the non-signaling model: Suppose A is an NS-LOCAL algorithm that 3-colors n 1 × n 2 grids with probability q > ε.Then, the running time of A is at least This result is easiest to interpret in the case of a square grid, i.e., n 1 = n 2 .Then the lower bound (for constant ε) is simply Ω( √ n), where n = n 1 • n 2 , and this is trivially tight since the diameter of the grid is O( √ n); hence the problem can be solved in O( √ n) rounds with a det-LOCAL algorithm.In particular, there is no room for distributed quantum advantage (beyond possibly constant factors).
Finally, we revisit the classical result by Linial [49] about the hardness of coloring trees.We show that essentially the same lower bound holds in the non-signaling model: Theorem 1.5.Let c ≥ 2 be an integer, and ε ∈ (0, 1).Suppose A is an NS-LOCAL algorithm that c-colors trees of size n ∈ N with probability q > ε.Then, for infinitely many n, as long as ε > e −n , the running time of A is at least

Key new ideas and techniques
In this section, we will give an informal overview of the key new ideas and techniques that we use to prove Theorems 1.1, 1.3, 1.4 and 1.5.We refer to Sections 4 and 5 for the formal proofs.

Classical upper bound (Section 4)
Background and prior work.The only existing distributed algorithm for solving the approximate coloring problem in general graphs that we are aware of is a folklore algorithm based on network decompositions [7,50].For parameters α and d, an (α, d)-network decomposition is a partition of the nodes V of a graph G = (V, E) into clusters of (weak) diameter at most d together with a proper α-coloring of the cluster graph; recall that the weak diameter of a cluster C ⊆ V is the maximum distance in G between any two nodes in C.
Given such a decomposition, it is not hard to see how to color graphs of chromatic number χ with αχ colors in d rounds by using disjoint color palettes for clusters of different colors: For every i ∈ [α], the nodes in a cluster of color i use colors from the palette {i, α + i, 2α + i, . . .}.Each such cluster C is colored by having a leader node collect the entire topology of the cluster and then brute forcing an optimal coloring φ C of the cluster.Since the graph has chromatic number χ, coloring φ C uses at most χ colors.Finally, the leader broadcasts the coloring and assigns each node v in C the color α(φ C (v) − 1) + i.This results in a proper coloring with at most αχ colors.In addition, the nodes do not require knowledge of χ in advance.This algorithm has for example been used by Barenboim [12] to compute a non-trivial approximate coloring in a constant number of rounds.Our algorithm is based on two new ideas that are outlined below.
New ingredient 1: New network decomposition algorithms.Network decomposition algorithms mostly focus on optimizing the product of α and d (as most applications of network decompositions require time proportional to αd) or on minimizing the number of cluster colors for a given maximum cluster diameter (e.g., [7,12,34,50,61]).We are interested in network decompositions with a fixed number of cluster colors α that is beyond our control and where we wish to optimize the value of d.By using existing clustering techniques [23,34,57,61] with some minor adaptations, we give randomized and deterministic algorithms that, for any parameter ε > 0, compute a set of non-adjacent clusters such that the cluster diameter of each cluster is polylog(n)/ε and the total number of unclustered nodes is at most εn.For every integer α, this can in turn be used to compute an (α, d)-network decomposition with d = Õ(n 1/α ).
Let us illustrate the idea for the case α = 2. Setting ε = 1/ √ n, we compute a set of non-adjacent clusters of diameter Õ( √ n) such that at most O( √ n) nodes remain unclustered.The constructed clusters can all be colored with color 1 and the connected components of the unclustered nodes form the clusters of color 2. We thus obtain a (2, Õ( √ n))-network decomposition in time Õ( √ n) and one can therefore color graphs of chromatic number χ with 2χ colors in Õ( √ n) rounds.

New ingredient 2:
The hiding trick.In Section 4 we show how to reduce the number of colors used while keeping the round complexity the same.The main new idea is what we call the hiding trick: First we make sure that clusters of the same color are at least four hops apart; this can be achieved by running a network decomposition protocol on G 3 .For simplicity, assume that we have a (2, Õ( √ n))-network decomposition.Let C be a cluster of color 1.We first extend C to an extended cluster C 1 that includes all unclustered nodes that are adjacent to C. Next we find a proper χ-coloring of C 1 using colors {1, . . ., χ} by brute force.Finally, any boundary node of C 1 that has color χ is removed, thus yielding a cluster C 0 with C ⊆ C 0 ⊆ C 1 .Note that C 0 is colored using at most χ colors and that and all boundary nodes of C 0 have a color different from χ.We have effectively hidden the color χ inside the cluster.Now we continue with the rest of the process and can safely use colors {χ, . . ., 2χ − 1} to color the uncolored nodes in clusters of color 2. The end result is a proper (2χ − 1)-coloring, and the running time is simply a constant times the cluster With this strategy, for instance, we can compute a 3-coloring of bipartite graphs in Õ( √ n) rounds.Theorem 1.1 is in essence a formalization and generalization of the hiding trick.We can hide one color in each cluster and therefore reuse one of the colors for all of the α cluster colors.This results in a coloring with α(χ − 1) + 1 colors.In addition, if one chooses the color palettes carefully, it is not necessary for the nodes to know χ beforehand.At first this may seem like an ad-hoc trick that cannot possibly be optimal-after all, we are saving only one color in each step.However, our matching lower bound shows that the hiding trick is essentially the best that we can do in distributed coloring, even if we have access to quantum resources.

Non-signaling lower bounds (Section 5)
Prior work on classical lower bounds.As a warm-up, let us first see how one could prove a lower bound similar to Theorem 1.3 for classical models.For concreteness, let us focus on the case χ = 2 and c = 3 in the det-LOCAL model.Then α = 2, and we would like to show that 3-coloring bipartite graphs requires Ω( √ n) rounds.Here we can make use of existential graph-theoretic arguments similar to the one already used by Linial [49].Let A be a det-LOCAL algorithm that purportedly finds a 3-coloring in bipartite graphs in T (n) = o( √ n) rounds.Now suppose that we are able to construct a graph G on n nodes with the following properties: The graph G is not bipartite, but (since A operates in the LOCAL model) we can apply A to G anyway and observe what happens.As the chromatic number of G is at least 4, certainly A cannot 3-color G; hence there has to be at least one node v such that in the local neighborhood X of v the output of A is not a valid 3-coloring.However, by assumption, the local neighborhood of v up to distance Θ( √ n) is bipartite, so with some cutting and pasting we can construct another graph G ′ = (V, E ′ ) on the same set of nodes such that G ′ is bipartite and the graph induced by the radius-T (n) neighborhood of X is the same in G and G ′ .Hence the output of A in X is the same in both graphs, which means A produces an invalid coloring in G ′ (which is bipartite) in the neighborhood X.The key point in this argument is the existence of the cheating graph G that "fools" A as A cannot locally tell the difference between G and the valid input graph G ′ .
New ingredient 1: Bogdanov's construction.To apply the proof strategy outlined above, we need a suitable construction of a cheating graph.It turns out we can make direct use of Bogdanov's [16] graph-theoretic work-this is a 10-year-old result, but so far this seems to have been a blind spot in the research community, and we are not aware of any lower bounds in the context of distributed graph algorithms that make use of it.
Together with our new algorithm from Theorem 1.1, this then gives a near-complete characterization of the complexity of c-coloring χ-colorable graphs in det-LOCAL.A similar argument applies (with some adjustments) to the rand-LOCAL model as well; in particular, we can exploit the independence of the output distribution between well-separated subgraphs of the input graph to boost the failing probability of an algorithm (see Section 5.2 for the details).
New ingredient 2: Defining cheating graphs for NS-LOCAL.While proof techniques based on cheating graphs are commonly used in the context of det-LOCAL and rand-LOCAL, we stress the same line of reasoning does not hold in quantum-LOCAL or NS-LOCAL.In fact, in the context of NS-LOCAL new challenges emerge, which we discuss later in this section.Our new proof strategy overcomes these issues: it builds on the idea of cheating graphs, but it allows us to directly derive a lower bound for NS-LOCAL.To the best of our knowledge, this is the first work establishing that Linial's argument [49] can be adapted to more general non-signaling models.This is one of our main technical contributions.
In Section 5 we present a new definition of a cheating graph that is applicable in NS-LOCAL.Suppose we are interested in a locally checkable problem P in graph family F. Definition 2.1 (Cheating graph, informal version).Consider a sufficiently large N > 0. A graph G is a cheating graph for (P, F) if (1) problem P is not solvable on G; (2) for a suitable k, we can cover G with k subgraphs G (1) , . . ., G (k) such that P is solvable over each of the graphs induced by their radius-T (n) neighborhoods, where n = |V (G)| • N ; (3) we can take any N subgraphs G (x 1 ) , . . ., G (x N ) together with their radius-T (n) neighborhoods, possibly with replacement, form their disjoint union G, and find a graph H ∈ F of size n that contains an induced subgraph isomorphic to G.
See Definition 5.6 for the formal version.We show that the existence of graphs of arbitrarily large size with the above properties directly implies a lower bound equal to T to the problem P over the graph family F that holds also in NS-LOCAL-this is formalized in Theorem 5.7.
We note that the precise requirements for k and N depend on T : in Section 5.5 we will exploit the fact that when T is small we can afford a large k, while in Sections 5.3 and 5.4 we deal with a large T , and then it will be important to construct a cheating graph with a constant k.
Our definition of a cheating graph is somewhat technical, but through three examples we demonstrate that this is indeed an effective way of proving lower bounds.
New ingredient 3: New analysis of Bogdanov's construction.While in the classical models we could directly apply Bogdanov [16] as a black box, this is no longer the case in NS-LOCAL.Nevertheless, in Section 5.3 we show that the construction of [16] indeed gives a cheating graph for c-coloring χ-chromatic graph.
It is known that the graph constructed by [16] is locally χ-chromatic, but globally has chromatic number greater than c, implying property (1) in Definition 2.1.We further go through the details of the construction and prove that the graph also satisfies properties (2) and (3) from Definition 2.1these are properties outside the scope of [16].Then from Theorem 5.7 we obtain Theorem 1.3.
New ingredient 4: New analysis of quadrangulations of the Klein bottle.In Section 5.4 we make use of properties of quadrangulations of the Klein bottle [5,58,59] to construct a family of graphs that is locally grid-like but is not 3-colorable.We show that such quadrangulations give cheating graphs for 3-coloring grids, and then Theorem 5.7 implies Theorem 1.4.
New ingredient 5: New analysis of Ramanujan graphs.In Section 5.5 we revisit the construction of Ramanujan graphs [52], that is, high-girth and high-chromatic regular graphs, which Linial used in his lower-bound proof.Again, we show that it provides us with a cheating graph (Definition 2.1) for c-coloring trees, and Theorem 1.5 follows.
Discussion.While quadrangulations of the Klein bottle and Ramanujan graphs have been used in prior work to prove lower bounds for the classical models, by e.g.Aboulker et al. [1] and Linial [49], we remark that to our knowledge, this is the first time that the applicability of Bogdanov's graph-theoretic work [16] in the context of distributed computing and quantum computing lower bounds is recognized (in spite of it being a 10-year-old result).
We also note that the classical version of Theorem 1.4 by [20] uses fundamentally different proof techniques: the argument in [20] is primarily algorithmic, while our proof is primarily graphtheoretic.The algorithmic proof from the prior work seems to be fundamentally incompatible with the NS-LOCAL model, while the graph-theoretic proof also yields a lower bound for NS-LOCAL.This suggests a general blueprint for lifting prior lower bounds from det-LOCAL or rand-LOCAL to NS-LOCAL: (1) re-prove the previous result using existential graph-theoretic arguments, and (2) apply the cheating graph idea to lift it to NS-LOCAL.
While the proof technique developed in this work is applicable in many graph problems, there are also problems for which cheating graphs do not exist (e.g., sinkless orientation on 2-regular graphs).An open research direction is developing proof strategies that can be used to derive NS-LOCAL lower bounds for those cases.

Lower bound technique in more details
Fix a sufficiently large N > 0. Consider any locally checkable problem P over some graph family F. We want to show that, whenever a cheating graph (Definition 2.1) for the pair (P, F) exists, any T -round algorithm solving the problem has failing probability at least 1 − (1 − 1/k) N , where k is the size of the subgraph cover of the cheating graph.
Let G be the cheating graph.For simplicity, we can think of P as the 3-coloring problem, and F to be the family of bipartite graphs.Provided that F respects some natural properties, properties (1) and (2) from Definition 2.1 ensure that we can get a lower bound in rand-LOCAL.Indeed, assume there is a T -round randomized algorithm A that 3-colors bipartite graphs.Clearly, A fails to 3-color G with probability 1.Hence, there is an i ⋆ ∈ [k] such that the failing probability of A over G (i ⋆ ) is at least 1/k.Hence, A will fail on any bipartite graph of at most n = |V (G)| • N nodes containing an induced subgraph isomorphic to the radius-T (n) neighborhood of G (i ⋆ ) with probability 1/k.If k is not small enough (e.g., k = w(1)), the failing probability tends to 0. In rand-LOCAL we can amplify the failure probability as follows: Suppose F contains a graph H N of at most n = |V (G)| • N nodes that contains, as subgraphs, N disjoint copies of the radius-T (n) neighborhood of G (i ⋆ ) in G.This is always possible if F is the family of bipartite graphs.By independence, the failing probability of A over H N is at least 1 − (1 − 1/k) N (see Fig. 1a).Hence, such an algorithm cannot exist.The property that F contains H N is reasonable for many natural problems (e.g., c-coloring χ-chromatic graphs for all combinations of c and χ) where, given a graph for which the problem is solvable, one can connect disjoint copies of the graphs and obtain a solvable instance of the problem.
However, as we anticipated, in the NS-LOCAL model some issues arise: (i) If two graphs G and H have different sizes, then even if they share two identical subgraphs G ′ and H ′ with isomorphic radius-T neighborhoods, a non-signaling output distribution is not guaranteed to be identical over G ′ and H ′ ; this is due to the no-cloning principle [25,55,66].
(ii) If we look at the output distributions for two subsets of nodes X and Y , then even if X and Y are far from each other, we cannot assume that the outputs of these subsets are independent.
Issue (i) puts some limits on the choice of the graph used to "fool" the algorithm, while issue (ii) makes it necessary to deal with possible dependencies among different parts of the input graph.Such issues are the reason why we require the cheating graph to satisfy property (3) in Definition 2.1.
G (4) G (5)  G (6)   G (8)  G (9) G ( 7) (a) Construction for the rand-LOCAL model.For any T (n)-round algorithm A solving the problem, there is an Graph H x for x = (4, 3, . . ., 9) G1( 4) G (2) 1 (b) Construction for the NS-LOCAL model.We start with N copies G 1 , . . ., G N of G and consider their disjoint union.We prove that, in this specific graph, there is already a combination of indices To solve issue (i), we consider a graph that is the disjoint union of N copies G 1 , . . ., G n of the cheating graph G: such graph has |V (G)| • N vertices, exactly the same as the graph of property (3) from Definition 2.1.Consider now any NS-LOCAL algorithm A that 3-colors bipartite graphs with locality T , and apply it to the graph ⊔ i∈[N ] G i .Clearly, A will fail to solve the problem in each G j with probability 1.At this point, we cannot continue as before: while it is true that in each G j we can find an index i ⋆ such that the probability of A failing in G (i ⋆ ) j is at least 1/k, we cannot use independence to increase the failing probability.
Property (3) ensures that, for a sufficiently large N , and for any sequence of indices x = (x 1 , . . ., x N ) ∈ [k] N , there exists a graph H x of size |V (G)| • N that contains an induced subgraph isomorphic to the disjoint union of the radius-T neighborhoods of G (x 1 ) , . . ., G (x N ) .However, correlations among these subgraphs of H x might hold.To overcome this issue, we need to consider all possible sequences of subgraphs G 1b). Fix a total order for the elements in [k] N , and let its ordered elements be x 1 , . . ., , where z i is the i-th element of x j , for each i = 1, . . ., N .Furthermore, for each index x ∈ [k] N , let I x be the set of all indices y ∈ [k] N such that y and x share the same element at the i-th position, for some i, i.e., x(i) = y(i).Notice that, for x = (x 1 , . . ., x N ), ∪ y∈Ix F y describes the event that there is an i ∈ implying that the dependencies behave "well enough", hence solving issue (ii).We proceed by contradiction: we assume that, for all and the sum of their probability is not 1.To better deal with the math, we define E x 1 = F x 1 and, recursively, we define reaching a contradiction.Thus, there exists an , and . By the property of the NS-LOCAL model, we get that the failing probability of A on

Preliminaries
We consider the set N of natural numbers to start with 0. We also define N + = N \ {0}.For any positive integer n ∈ N + , we denote the set {1, . . ., n} by [n].
Graphs.All graphs in this paper are simple graphs without self-loops.Let G = (V, E) be a simple graph with n ∈ N nodes.If the set of nodes and the set of edges are not specified, we refer to them by V (G) and E(G), respectively.For any edge e = {u, v} ∈ E(G), we also write e = uv = vu.
If G is a subgraph of H, we write G ⊆ H.For any subset of nodes A ⊆ V , we denote by G[A] the subgraph induced by the nodes in A. For any nodes u, v ∈ V , dist G (u, v) denotes the distance between u and v in G (i.e., the number of edges of any shortest path between u and v in G); if u and v are disconnected, then dist The coloring φ is said to be proper if we have φ(u) ̸ = φ(v) for every uv ∈ E. If, for some χ ∈ N, there exists a proper χ-coloring for G and χ is minimal with this property, then G is said to be χ-chromatic; we also say that the chromatic number of G, denoted by X (G), is χ.In the c-coloring problem, the input is a graph G, and the task is to find a proper c-coloring of G.
The LOCAL model.The LOCAL model is a distributed system consisting of a set of |V | = n processors (or nodes) that operates in a sequence of synchronous rounds.In each round the processors may perform unbounded computations on their respective local state variables and subsequently exchange of messages of arbitrary size along the links given by the underlying input graph.Nodes identify their neighbors by using integer labels assigned successively to communication ports.(This assignment may be done adversarially.)Barring their degree, all nodes are identical and operate according to the same local computation procedures.Initially all local state variables have the same value for all processors; the sole exception is a distinguished local variable x(v) of each processor v that encodes input data.
Let c ≥ 1 be a constant, and let Σ in be a set of input labels.The input of a problem is defined in the form of a labeled graph (G, x) where G = (V, E) is the system graph, V is the set of processors (hence it is specified as part of the input), and ] and of an input label λ in (v) ∈ Σ in to each processor v.The output of the algorithm is given in the form of a vector of local output labels λ out : V → Σ out , and the algorithm is assumed to terminate once all labels λ out (v) are definitely fixed.We assume that nodes and their links are fault-free.The local computation procedures may be randomized by giving each processor access to its own set of random variables; in this case, we are in the randomized LOCAL (rand-LOCAL) model as opposed to deterministic LOCAL (det-LOCAL).
The running time of an algorithm is the number of synchronous rounds required by all nodes to produce output labels.If an algorithm running time is T , we also say that the algorithm has locality T .Notice that T can be a function of the size of the input graph.
We say that the c-coloring problem over some graph family F has complexity T in the det-LOCAL model if there exists a det-LOCAL algorithm solving the problem in time T for all input graphs G ∈ F, but no det-LOCAL algorithm solves the problem in time T − 1 (where T can be a function of the input graph size) for all input graphs G ∈ F. The complexity in the rand-LOCAL model is defined similarly.

New classical graph coloring algorithms
In this section we prove the existence of algorithms in det-LOCAL and rand-LOCAL that almost match the NS-LOCAL lower bound.We recall here the precise statement for the reader's convenience.

Theorem 1.1.
There exists a det-LOCAL algorithm A det and a rand-LOCAL algorithm A rand that, given a parameter α ∈ N, find a proper vertex coloring with α(χ − 1) + 1 colors in any graph with chromatic number χ, as follows: 1) rounds.
Notice that, if we plug in the same parameters (with c = α(χ − 1) + 1) in Theorem 1.3 with an appropriate choice of constant ε, we get that α = c−1 χ−1 , implying a lower bound of Ω n 1/α /χ 1+1/α for the problem.Therefore, for constant χ and α, our algorithms give a perfect trade-off between quality of approximation and time complexity up to logarithmic factors.
Fast coloring from fast network decomposition.Our algorithm follows an approach similar to that of Barenboim et al. [13], which in turn is based on network decompositions.
• For every i, the (weak • µ is an α-coloring of S. In addition, the coloring µ is presented explicitly to the nodes of G; that is, every node v ∈ C i knows the cluster color µ(C i ) of its respective cluster C i .
We recall the algorithm of Barenboim et al. [13].Suppose we are given an (α, d(n))-network decomposition D. We iterate sequentially through the α cluster colors of D. For each cluster color a and each cluster C having the cluster color µ(C) = a, collect the entire topology of C in some arbitrary node v ∈ C (chosen by, e.g., leader election).The node v then computes a perfect coloring π for the nodes of C that uses at most χ colors and then broadcasts to each node u ∈ C its color π(u).Clearly in the LOCAL model each iteration takes at most diameter of C i many rounds, so at most O(d(n)) rounds.With a clever implementation, this process can also be sped up by doing all the iterations in parallel.We will see this in the later sections.
In order for this strategy to work, we must ensure that neighboring clusters use distinct sets of colors; otherwise, the colorings of the nodes of neighboring clusters may not match.To deal with this, we have clusters of different cluster colors use completely different sets of colors for the nodes, which we will refer to as color palettes.More precisely, each cluster C of cluster color µ(C) = a may only use colors from the palette p a = {(a, b) | b ∈ N + }.Since there are α many colors for the clusters, (assuming each cluster is colored using at most χ colors) this then gives a coloring of G with αχ colors.As we color each cluster by gathering its entire structure in a single node, knowledge of χ is not needed in order to use the optimal number of colors for each cluster.
The "hiding trick".We show how to optimize the above strategy using what we call a "hiding trick".We add one special hidden color −1 in common to all color palettes p a and that is guaranteed to appear only in the "inside" of the clusters; that is, if a node v in some cluster C has a neighbor that is not in C, then v is guaranteed to not be colored −1.This ensures the colorings produced by two neighboring clusters are still compatible since the only color shared by their palettes is the hidden color −1, which is only present in the "insides" of the clusters.Since the palettes now share exactly one color, this allows us to save α − 1 colors in total.Surprisingly enough, this small modification is enough to attain the minimum number of colors possible, that is, α(χ − 1) + 1 colors (as per our lower bound from Theorem 1.3).
Fast decomposition from fast clustering.The specific complexity of the resulting algorithm depends on value d(n) of the underlying decomposition D. We show how to obtain a decomposition with To do so, we show how to efficiently turn any existing clustering algorithm into a network decomposition algorithm.The difference between the two is that the former only needs to group a subset of nodes in the graph, whereas the latter must partition the entire graph.
meeting the following conditions: • S 1 , . . ., S k are mutually non-adjacent; that is, the distance between any two nodes u ∈ S i and v ∈ S j where i ̸ = j is at least 2.
• For every i, the (weak • D contains at most λ|V (G)| vertices.
Our conversion from a clustering algorithm into a network decomposition one is by a bootstrapping procedure: if the clustering algorithm is guaranteed to cluster at least half of the nodes in G, then we can apply it again and again until only an ε fraction of nodes remains unclustered, where ε is a parameter of our choosing.For an appropriate choice of ε, the fraction of nodes that remains is sufficiently small that we can directly gather the remaining nodes into their own cluster (simply by grouping every remaining connected component of the graph).
Organization.In Section 4.1 we show how to obtain a fast coloring algorithm given the underlying network decomposition D. In Section 4.2 we then show how to obtain such a decomposition following the two-step approach described above: first we show our bootstrapping result for clustering algorithms in Section 4.2.1 and then how this implies a network decomposition algorithm in Section 4.2.2.Plugging in the state-of-the-art for clustering algorithms, we then obtain Theorem 1.1.

The hiding trick
The main result of this section is the following, which is a coloring algorithm based on our so-called hiding trick (see also Lemma 4.4 below).The algorithm presupposes the existence of a network decomposition algorithm, which we show how to obtain in Section 4.2.
Theorem 4.3.Fix some parameter α ∈ N and suppose there is an (α, d(n))-network decomposition algorithm B for det-LOCAL or rand-LOCAL that runs in time d(n).Then there is an algorithm A that (α(χ − 1) + 1)-colors any χ-chromatic graph G with n nodes in O(d(n)) time.Moreover, A works in det-LOCAL or rand-LOCAL, depending on which model B runs in.In addition, if B is randomized (i.e., it is a rand-LOCAL algorithm) and succeeds with probability 1 − 1/poly(n), then so does A succeed with probability 1 − 1/poly(n).
The core idea of our algorithm is the following constructive lemma.It shows that, in any graph G, it is always possible to color a subset A of nodes in a way that "hides" one designated hidden color −1.To color A in this manner, it may be necessary to fix the color of some nodes in N (A) as well.
Lemma 4.4 (Hiding Lemma).Let G = (V, E) be a graph, and let χ be the chromatic number of G.For any subset A ⊆ V , there exists A ⊆ A ′ ⊆ (A∪N (A)) and a proper coloring φ : Proof.Since G is χ-colorable, there exists a χ-coloring ψ of A ∪ N (A).We uncolor some of the nodes of N (A) such that none of the uncolored nodes has a neighbor with color −1.Formally, let and φ = ψ↾ A ′ (i.e., the restriction of ψ to A ′ ).Since we only uncolor nodes in N (A) and ψ is a proper χ-coloring of A ∪ N (A), φ is certainly a proper coloring of A ′ ⊆ (A ∪ N (A)).We now argue that v ∈ V \ A ′ has no neighbor that is colored −1 by φ.Since A ⊆ A ′ , we must consider only the following two cases: Note that Lemma 4.4 (in its current form) cannot be immediately applied to a decomposition of G to produce colorings for the clusters of G.The reason for that is the following: Consider two clusters C and C ′ that are assigned the same cluster color a by the network decomposition.(In particular, this means C and C ′ are not adjacent.)Recall that we will color the nodes of C and C ′ using the same color palette.However, if we color them such as in Lemma 4.4, then we are potentially also coloring nodes in N (C) and N (C ′ ) and, for all we know, the intersection N (C) ∩ N (C ′ ) may not be empty.Hence we cannot simply use the same color palette in both clusters, as this could potentially lead to an invalid coloring.
Therefore, we would like that not only C and C ′ but also N (C) and N (C ′ ) are not adjacent.Equivalently, we wish for the distance between C and C ′ to be at least 4. We can indirectly guarantee this by using an (α, d(n))-network decomposition D of G 3 instead of G.This is because being not adjacent in G 3 immediately implies a distance of at least 4 in G. Asymptotically speaking, this does not incur any additional cost (compared to computing a decomposition of G) since, for any k, we can simulate each round of communication in G k by using k rounds of communication in G.
We now present Algorithm 1, which, given a χ-chromatic graph G and an (α, d(n))-network decomposition D of G 3 , colors G using α(χ − 1) + 1 colors.By proving the correctness of Algorithm 1 we then obtain Theorem 4.3.
Let us give a brief high-level description of Algorithm 1.Each cluster C i acts independently and based on the cluster color a that is assigned to it by the decomposition D. First, the entire topology of C i ∪ N (C i ) is gathered in some leader node leader i .Next leader i brute-forces an optimal coloring φ i of its respective cluster C i using a color palette p a that depends on a. (Without

Algorithm 1 Coloring a decomposition
Elect a leader node leader i of C i 3: Collect the entire topology of a ← the cluster color assigned to C i by D 5: if a = α then 7: leader i broadcasts φ i to all nodes in C i 9: else 10:  end if 21: end for 22: Each node outputs its own color restriction we use only the smallest elements of p a (according to the natural ordering of p a ) in φ i .This enables the nodes to choose correct colors even without knowledge of χ.)If a = α, φ i is simply a χ-coloring of C i whose existence is guaranteed by the χ-chromaticness of G. Otherwise (i.e., if a < α), leader i instead computes a coloring φ i of C ′ i ⊆ C i ∪ N (C i ) according to Lemma 4.4.Each coloring φ i is broadcasted to all nodes that may have been assigned a color by φ i .The nodes that were assigned multiple colors (i.e., potentially those at the border of two distinct clusters) then simply choose one of them arbitrarily.Lemma 4.5.Algorithm 1 computes a valid coloring of G using no more than α(χ − 1) + 1 colors.
Proof.First we argue that no more than α(χ − 1) + 1 colors are used in total.For a cluster color To show that the color is proper, first observe that, based on Lemma 4.4, no node in N (C i ) is assigned −1 by φ i ; as a result, if a node v is assigned the color −1 by φ i , then necessarily v ∈ C i .Consider any two adjacent nodes v and v ′ and recall that both of them choose the largest color among any of the colors they were assigned.For the sake of contradiction, suppose that both nodes pick the same color x.Consider the following two cases: x ̸ = −1.Since the intersection of any two color palettes is {−1}, this implies that v and v ′ were assigned colors from the same color palette p a .Since the φ i are all valid colorings, the colors of v and v ′ come from different coloring functions φ i and φ i ′ , respectively.However, since φ i and φ i ′ use the same color palette, the respective clusters C i and C i ′ are assigned the same cluster color by D. This means that C i and C i ′ are not adjacent in G 3 and, in turn, the distance between nodes in N (C i ) and N (C i ′ ) is at least 2, which immediately contradicts v and v ′ being adjacent.
x = −1.This means that both of the nodes are assigned their color by the coloring of their respective cluster.Let i and i ′ be the numbers of the clusters of v and v ′ , respectively.If i = i ′ , then both v and v ′ pick their colors according to φ i = φ i ′ , which contradicts φ i being a proper coloring.Hence let i ̸ = i ′ .We have then that C i and C i ′ are adjacent clusters and, since they are distinct, without restriction we have i ̸ = α.Since v ′ ∈ N (C i ) and the color of both v and v ′ is −1, Lemma 4.4 implies that v ′ is in the domain of φ i (as otherwise it would be adjacent to v, which has the color −1).Now since φ i (v) = −1, we know that φ i (v ′ ) > −1, which means that the color of v ′ cannot be −1.
Note that, in the proof above, we simply showed that no more than c = α(χ − 1) + 1 colors are used, but there is still a minor technicality to be dealt with since the colors do not come from the set [c] as the definition of c-coloring demands (but rather either the color is −1 or a pair (a, b)).A straightforward way of resolving this is, for instance, remapping

Fast network decomposition
Given any clustering algorithm, we can obtain a network decomposition algorithm as follows.
Theorem 4.7.Let f, g : N → N be arbitrary functions and suppose there is an O(f (n))-round distributed (1/2, g(n))-clustering algorithm named cluster for det-LOCAL or rand-LOCAL.There is an algorithm A that, given a graph G = (V, E) and any α ∈ N + , computes an (α, O(n rounds.The algorithm A works in det-LOCAL or rand-LOCAL, depending on which model cluster itself is based on.In addition, if cluster is randomized and succeeds with probability 1 − 1/poly(n), then A also succeeds with probability 1 − 1/poly(n).
We mention two state-of-the-art clustering algorithms that can be plugged into Theorem 4.7, one for the det-LOCAL model and one for the rand-LOCAL model.For the det-LOCAL model we use the clustering algorithm from [34].In fact this algorithm actually works even in the more restricted CONGEST model, where each node can only send O(log n)-bit messages each round.

Theorem 4.8 ([34]
).There exists an algorithm that computes a Plugging in this algorithm in Theorem 4.7, we obtain the first item of Theorem 1.1.For the rand-LOCAL model (i.e., the second item of Theorem 1.1), we use the following.
There are two steps to the proof of Theorem 4.7.In Section 4.2.1 we show a bootstrapping result where we use the (1/2, g(n))-clustering algorithm cluster to obtain a (ε, O(g(n)/ε))-clustering algorithm for any ε of our choosing.Plugging in an adequate value for ε, in Section 4.2.2 we then obtain the network decomposition algorithm of Theorem 4.7.

Fast clustering
Next we show our bootstrapping procedure, with which we can reduce a 1/2 fraction of unclustered nodes to any ε of our choosing.The price to pay is only a multiplicative O(1/ε) factor in the diameter of the clusters and a multiplicative O ε −1 log(1/ε) factor in the running time.Formally, what we achieve is the following: Theorem 4.10.Let cluster be as in Theorem 4.7.For any 0 < ε ≤ 1, there is an algorithm ε-cluster that computes an The algorithm ε-cluster can be implemented in the same distributed model as cluster and its runtime is dominated by 2 log(1/ε) invocations of cluster.In addition, if cluster is randomized (i.e., it works in rand-LOCAL) and succeeds with probability 1 − 1/poly(n), then ε-cluster also succeeds with probability 1 − 1/poly(n).
We first describe the algorithm (Algorithm 2) and then prove it satisfies the properties of Theorem 4.10.Note that, in the description of Algorithm 2 in all three Lines 6,7 and 8 the neighborhoods are always taken with respect to G.

Algorithm 2 Bootstrapping cluster
Require: G = (V, E), 0 < ε ≤ 1 Require: Clustering algorithm cluster as in Theorem 4. Run cluster on G 2R+1 , producing a clustering C 5: for each cluster C ∈ C in parallel do 6: Mark all nodes in N j * (C) \ N j * −1 (C) for deletion 8: Add N j * −1 (C) as a cluster to C ′

9:
end for 10: Let U be the set of nodes not marked for deletion or that did not join a cluster of C ′

11:
G ← G[U ] 12: end for 13: Each node outputs whether it is part of a cluster of C ′ or not At a high level, Algorithm 2 works in multiple iterations, in each of which we first invoke cluster.For each new cluster C that is computed, we delete some boundary around this cluster and separate it from the rest of the graph.We then add C to our final clustering.Since cluster clusters at least half of the nodes, the size of our graph is halved in each iteration; hence after O(2 log 1/ε) iterations we are left with only an ε fraction of nodes.
The main challenge lies in the choice of the boundaries and preventing too many nodes from being deleted.To solve this we appeal to the computational power of the LOCAL model: For each cluster C, we inspect the (4/ε)-hop neighborhood of C and find j * such that the number of nodes at distance exactly j * from C (i.e., |N j * (C) \ N j * −1 (C)|) is minimized.As we will show, there is always a choice of j * such that the number of nodes we delete is not too large.
We now turn to showing that Algorithm 2 satisfies the properties in Theorem 4.10.To give an overview, we need to prove the following: 1.The clusters created have diameter O(g(n)/ε) and are non-adjacent (Lemma 4.11).
We address these claims now one by one.Proof.We start by analyzing a single iteration of the for loop on Line 3 and show that a cluster with diameter O(g(n)/ε) is created.Afterwards we prove that, once a good cluster is created, it is preserved by later iterations.Algorithm 2 runs the algorithm cluster of Theorem 4.8 on the power graph G 2R+1 .Let C i , C j be two clusters created by cluster.For any nodes u ∈ C i and v ∈ C j with i ̸ = j, we have dist G ′ (u, v) ≥ 2, which implies dist G (u, v) ≥ 2R + 2. Let S i ∈ S and S j ∈ S be the clusters that are fixed by Line 8 in the iterations of C i and C j , respectively.We observe that S i (resp., S j ) only contains nodes at distance at most R − 1 from C i (resp., C j ).Hence it follows that, for any u ′ ∈ S i and v ′ ∈ S j , we have dist In addition, cluster guarantees that the diameter of the clusters is at most When creating a fixed cluster in Line 8 we increase the diameter by at most 2j * ≤ 2R − 2 hops.As a result, the diameter of any fixed cluster is bounded above by Since the entire one-hop-neighborhood of every cluster is deleted, in the next iterations all of the new clusters will also have at least one deleted node between themselves and any other previously created cluster.Also no nodes of previously fixed clusters are ever considered for another cluster or deleted.Hence the subsequent iterations do not interfere with the previous ones.Lemma 4.12.Algorithm 2 deletes at most an ε fraction of nodes during its execution.All nodes that are not deleted eventually join a cluster.
Proof.Without loss of generality, we assume ε ≤ 1/2, as otherwise a single execution of cluster already gives the result.We upper-bound the number of deleted nodes using an inductive argument.
Let S be the set of all clusters following a single execution of the for loop on Line 3. In addition, for a cluster C ∈ S, let del(C) denote the set of nodes marked for deletion by C (during the execution of the for loop on Line 5).Observe that, for any two distinct clusters C 1 ̸ = C 2 , we have that del(C 1 ) ̸ = del(C 2 ).This is due to the fact that dist G (C 1 , C 2 ) ≥ 4R + 2 and that dist G (v, C i ) ≤ j * ≤ R − 1 for any v ∈ del(C i ).Also, by an averaging argument, for any cluster This means we can upper-bound the total number of deleted nodes by By Theorem 4.8, at least half of the remaining nodes in G are clustered in each iteration of the for loop on Line 3. Arguing as before, we get that the number of remaining nodes is halved with each execution of the loop.This means that, during the execution of the loop, the number of deleted nodes is at most In turn, due to the aforementioned progress guarantee, the 2 log(1/ε) repetitions of the loop ensure at most a 2 −2 log(1/ε) = ε 2 fraction of nodes are left standing at the end and are then deleted in the final step of Algorithm 2. As a result, using that ε ≤ 1/2, the total number of deleted nodes is at most Proof.We analyze the runtime of a single iteration of the for loop on Line 3. When ran on G 2R+1 , cluster takes O(f (n)) time, so for each round of cluster we spend O(R) = O(1/ε) rounds to simulate its execution on G. Hence we need O(f (n)/ε) rounds in total to run cluster on G.
Next we collect the entire R-neighborhood of a cluster C in some leader node, compute j * , and then broadcast j * to all nodes inside of N j * (C).This all requires O(g(n)/ε + R) = O(g(n)/ε) rounds.Hence a single iteration of the for loop on Line 3 costs O((f (n) + g(n))/ε) rounds in total.The final deletion procedure does not cost any additional rounds since each node knows at the end whether it has joined a cluster or not.Since we repeat the loop O(log(1/ε)) times, the claim follows.
Proof.Let c > 0 be such that cluster succeeds with probability at least 1 − 1/n c .In addition, let us assume ε > 1/n as otherwise the claim is trivial.The observation to make is that, if all executions of cluster by Algorithm 2 are correct, then the result of Algorithm 2 is also correct.Since there are 2 log(1/ε) executions of cluster in total, this means that, for any 0 < c < c (and, in particular, for any fixed choice of such a c) and large enough n, the probability that Algorithm 2 succeeds is at least This concludes the analysis of Algorithm 2, from which Theorem 4.10 follows.

Fast network decomposition from fast clustering
Finally we show how a clustering algorithm as in Theorem 4.10 implies a network decomposition algorithm, thereby giving a proof of Theorem 4.7.
The strategy followed by Algorithm 3 is very much straightforward: First apply the clustering algorithm α−1 times.The remaining graph contains then at most many nodes.At this point we can just put all remaining connected components into their own clusters, which will trivially have diameter at most O(g(n)/ε).

Algorithm 3 (α, O(n 1/α ))-network decomposition from clustering
Require: G = (V, E), α ∈ N + Require: Clustering algorithm ε-cluster from Theorem 4.10 Color every cluster C ∈ C i with the cluster color µ(C) = i Proof.Clearly Algorithm 3 produces clusters with the correct diameter: ε-cluster produces clusters with diameter O(g(n)/ε) = O n 1/α g(n) and there are only ε α−1 n = O(g(n)/ε) nodes to be clustered in Line 7. As for µ being a proper cluster coloring, notice that ε-cluster already guarantees the clusters formed in clustering C i are non-adjacent; this is also guaranteed in the clustering created in Line 7. Regarding the round complexity, we have α − 1 many invocations of ε-cluster and then at most ε α−1 n many nodes to cluster in Line 7. Hence using that the running time of each invocation of

10), we can upper-bound the round complexity by
.
5 New lower bounds in the non-signaling model

Framework
In this section we define the framework in which our technique is developed.We start with the notion of labeling problem.
Definition 5.1 (Labeling problem).Let Σ in and Σ out two sets of input and output labels, respectively.
A labeling problem P is a mapping (G, λ in ) → {λ (out,i) } i∈I , with I being a discrete set of indices, that assigns to every graph G with any input labeling λ in : V (G) → Σ in a set of permissible output vectors λ (out,i) : V (G) → Σ out that might depend on (G, λ in ).The mapping must be closed under graph isomorphism, i.e., if φ : ) is an isomorphism between G and G ′ , and A labeling problem can be thought as defined for any input graph of any amount of nodes.If the set of permissible output vectors is empty for some input (G, λ in ), we say that the problem is not solvable on the input (G, λ in ): accordingly, the problem is solvable on the input (G, λ in ) if One observation on the generality of definition of labeling problem follows: one can actually consider problems that require to output labels on edges.This variation of Definition 5.1 does not affect in any way the applicability and the generality of the result we present in Section 5, namely, our lower bound technique.
We actually focus on labeling problems where, for any input graph, an output vector λ out is permissible if and only if the restrictions of the problem on any local neighborhoods can be solved and there exist compatible local permissible output vectors whose combination provides λ out .This concept is grasped by the notion of locally verifiable labeling (LVL) problems, the generalization of locally checkable labeling (LCL) problems to unbounded degree graphs, first introduced by Naor and Stockmeyer [60].For any function f : A → B and any subset A ′ ⊆ A, let us denote the restriction of f to A ′ by f ↾ A ′ .Furthermore, we define a centered graph to be a pair (H, v H ) where H is a graph and v H ∈ V (H) is a vertex of H that we name the center of H.The radius of a centered graph is the maximum distance from v H to any other node in H. Definition 5.2 (Locally verifiable labeling problem).Let t ∈ N. Let Σ in and Σ out two sets of input and output labels, respectively, and P a labeling problem.P is locally verifiable with checking radius t if there exists a (possibly infinite) family S = {((H, v H ), λin , λout ) i } i∈I of tuples, where (H, v H ) is a centered graph of radius at most t, λin : V (H) → Σ in is an input labeling for H, λout : V (H) → Σ out is an output labeling for H (which can depend on λin ) with the following property • for any input (G, λ in ) to P, an output vector We remark that the notion of an (LVL) problem is a graph problem, and does not depend on the specific model of computation we consider (hence, the problem cannot depend on, e.g., node identifiers).Next definition introduces the concept of outcome of an algorithm.Definition 5.3 (Outcome).Let Σ in and Σ out be two sets of input and output labels, respectively.An outcome O is a mapping (G, x) → {(λ (out,i) , p i )} i∈I , with I being a discrete set of indices, assigning to every input graph G with any input data x = (id : 1. for all i ∈ I, p i > 0; 2. i∈I p i = 1; 3. p i represents the probability of obtaining λ (out,i) as the output vector of the distributed system.
Let T ≥ 0 be any integer.We say that an outcome O on some graph family F has locality T if there exists a distributed algorithm in the LOCAL model which, for any input (G, x) where G ∈ F, produces the same probability distribution over output vectors as O after T rounds of computation.Notice that T can be a function of the size of the input graph.
An algorithm can be thought of producing an output distribution on every input: whenever the computations of a node in a given round are defined, the algorithm proceeds normally; if at some round some computation is undefined for a node, the node outputs some "garbage label", say ⊥: we remark that we can assume Σ out always contains such a garbage label without loss of generalization.Hence, an outcome can be always thought of as being defined on the family of all graphs and all valid inputs: for this reason, we will omit specifying the graph family on which the outcome is defined.
We say that an outcome O over some graph family F solves problem P over F with probability p if, for every G ∈ F and any input data x = (id, λ in ), it holds that When p = 1, we will just say that O solves problem P over the graph family F.

We now define the complexity class L[T ].
A problem P over some graph family F belongs to the class L[T ] (respectively, L[T, p], for some p ∈ [0, 1]) if there exists a distributed algorithm in the LOCAL model which produces an outcome O with locality at most T which solves problem P on F (respectively, solves problem P with probability p).In such case we write (P, F) ∈ L[T ] (respectively, (P, F) ∈ L[T, p]).
The next computational model tries to capture the fundamental properties of any physical computational model (in which one can run either deterministic, random, or quantum algorithms) that respects causality.The defining property of such a model is that, for any two (labeled) graphs (G 1 , x 1 ) and (G 2 , x 2 ) that share some identical subgraph (H, y), every node u in H must exhibit identical behavior in G 1 and G 2 as long as its local view, that is, the set of nodes up to distance T away from u together with input data and port numbering, is fully contained in H.As the port numbering can be computed with one round of communication through a fixed procedure (e.g., assigning port numbers 1, 2, . . ., deg(v) based on neighbor identifiers in ascending order) and we care about asymptotic bounds, we will omit port numbering from the definition of local view.
The model we consider has been introduced by Arfaoui and Fraigniaud [6] and is equivalent to the φ-LOCAL model by Gavoille et al. [33]; however, as in [6], we explicitly require the outcome to be defined for every possible graph: in fact, as argued for distributed algorithms before, every physical procedure producing outcomes for graphs should produce some outcome on any input (possibly, by using some garbage label as before). 1n order to proceed, we first define the non-signaling property of an outcome.Let T ≥ 0 be an integer, and I a set of indices.For any set of nodes V , subset S ⊆ V , and for any input (G = (V, E), x), we define its T -local view as the set where dist G (u, v) is the distance in G. Furthermore, for any subset of nodes S ⊆ V and any output distribution {(λ (out,i) , p i )} i∈I , we define the marginal distribution of {(λ (out,i) , p i )} i∈I on set S as the unique output distribution {( λ(out,i) , pi )} i∈I acting on S which satisfies the condition pj = i : λ(out,j) =λ (out,i) [S] p i , where λ (out,i) [S] is the restriction of output λ (out,i) to the processes in S. Definition 5.4 (Non-signaling outcome).An outcome O : (G, x) → {(λ (out,i) , p i )} i∈I is nonsignaling beyond distance T if for all set of nodes V and all subsets S ⊆ V , for any pair of inputs ( As every (deterministic or randomized) algorithm running in time at most T in the LOCAL model produces an outcome which has locality T , we can provide lower bounds for the LOCAL model by proving them in the NS-LOCAL model.For the sake of readability, we assume that every outcome O that has locality T can be produced by a hypothetical non-signaling LOCAL algorithm A with running time T .This is just an artifact of the text and does not affect in any way the validity of our proofs.
We now present a lower bound technique that works for LVL problems in NS-LOCAL.

Lower bound technique
We first introduce the notion of subgraph cover.

Indistinguishability argument in the classical LOCAL model
Our technique is an extension of the indistinguishability argument already exploited to prove lower bounds in the LOCAL model [26,32,45].Let P be an LVL problem over some graph family F.
The indistinguishability argument basically says that the output of a node v running a T -round algorithm A cannot distinguish between inputs that differ only outside its T -view.Hence, A cannot solve the problem.However, in the aforementioned works, all the different inputs considered to "confuse" the nodes were solvable instances.Another approach, the one that we consider in this work, was introduced by Linial [49] to prove a lower bound for c-coloring trees: it uses graphs outside the input graph family (on which the problem is impossible to be solved) which is locally everywhere a solvable instance.Linial used a high-girth graph G (which locally looks like a tree but lies outside the input graph family) that has chromatic number bigger than c.This immediately yields a lower bound that is some constant fraction of girth(G).In this sense, the approach is purely existential graph-theoretic at heart.We generalize this latter method all the way up to the NS-LOCAL model, and we also present a technique to boost the failing probability of any outcome.As our argument presents some technicalities, we proceed step by step and present it first for the det-LOCAL model, then for rand-LOCAL, and finally for NS-LOCAL.
The argument for the det-LOCAL model goes roughly as follows.
Indistinguishability argument: det-LOCAL model.Suppose we have an LVL problem P, with checking radius t > 0, that is solvable over some graph family F, and assume we have a det-LOCAL algorithm A that solves P over F and has running time T (n) ≥ t > 0, n being the size of the input graph.We remark that T might also depend on other parameters of the input graph, such as the maximum degree.However, we omit such dependencies for the sake of readability.Let us fix the size n of the input graph.Suppose there exists a graph G n / ∈ F such that P is not solvable over G n , and let us run A for T (n) rounds over G n (G n does not have necessarily size n-we force outputs after time T (n) if the protocol did not produce any or if, at any time, at any node the computational procedure is not well-defined).As P is not solvable over G n , we know that there is a node ] are given exactly the same identifiers and input labels.As Hn and G n [N t (v)] look identical, A must have produced the same non-admissible output over Hn in time T (n), which is a contradiction.
This argument can be extended to the rand-LOCAL model with some care: while in the det-LOCAL model the local failure is deterministic and takes necessarily place in all graphs that locally look like the bad neighbor, this is not the case for rand-LOCAL.We now show how to deal with random outputs.
Indistinguishability argument: rand-LOCAL model.We keep the same hypothesis (except that now A is a rand-LOCAL algorithm) and, in addition, we ask that the graph G n / ∈ F, over which P is not solvable, admits a subgraph cover {G (i) n } I∈I with the following properties: (2) For each i ∈ I, there exists a graph Again, we run A on G n , and we know that G n will contain at least one t-neighborhood that has a non-admissible output vector.As {G n } I∈I is a subgraph cover, there exists i ⋆ ∈ I such that the probability of A failing over n )] are given the same identifiers and input labels, A fails on H (i ⋆ ) n with probability at least 1/|I|.We observe that property (2) is sufficient but not necessary to the technique: it is sufficient to ensure the existence of the graph H (i) Nevertheless, in many practical scenarios actually determining i ⋆ is hard, while it is easier to ensure (2) for many graph families.
This result is useful when |I| is not too large: however, in many cases, it is not possible to find subgraph covers with few elements; furthermore, one may want a failure probability that is higher than a constant value.Luckily, other properties of the rand-LOCAL model come to our aid and sometimes allow to boost the failing probability.Let N ∈ N + .The idea is to replicate N times the T (n)-neighborhood of the bad subgraph (making sure we still obtain a graph that belongs to the graph family under consideration) and exploit the independence of the outcome generated by A over subsets of the nodes that are "far enough".More formally, assume that N T (n) (G (i) n ) has size at most ⌊n/N ⌋ for each i ∈ I. Let us replace property (2) as follows: (2) For each choice of indices x N = (x 1 , . . ., x N ) ∈ [|I|] N , there exists a graph H x N ∈ F of size n which contains a subgraph Hx N such that Hx N is isomorphic to the disjoint union N j=1 G (x j ) n , and By independence, the probability that A solves P over H x N is now (1 − 1/|I|) N (assuming again the same identifiers and input labels are given to ).In both arguments for the det-LOCAL and the rand-LOCAL model, we denominate the graph G n as the cheating graph, because it allows us to "trick" the distributed algorithm, since nodes cannot distinguish between different inputs if they have the same local view.

Indistinguishability argument in the NS-LOCAL model
The argument outlined in Section 5.2.1 cannot be directly applied to the NS-LOCAL model for two reasons: 1. Independence is not guaranteed between far away subsets of nodes (e.g., there could be some shared resources).
2. We cannot consider an outcome over two graphs G, H of different sizes and require it to have the same output distribution over two subgraphs G ′ ⊆ G, H ′ ⊆ H that have the same local neighborhood due to the no-cloning principle [25,55,66] (in fact, the properties of a non-signaling outcome hold only for graphs of the same size; see Definition 5.4).
However, we overcome these issues and show that: 1.The dependencies actually go "in the right direction", i.e., the bound on the failing probability does not decrease w.r.t. the bound we showed in the rand-LOCAL model; 2. We can restrict ourselves to graphs of same sizes in many cases, as we show in the applications of our lower bound technique (Sections 5.3-5.5).
Some technicalities are required to obtain 1.This section is devoted to the formal proof of our lower bound technique in the NS-LOCAL model.The technique applies to LVL problems restricted to graph families meeting some specific properties.Definition 5.6 (Cheating graph).Let P be any LVL problem with checking radius t ∈ N that is solvable over some graph family F. Suppose that, for some integer n ∈ N, there exists a triple (k, N, T ) ∈ N 3 (that can depend on n and, possibly, other parameters defining the graph family F), with T ≥ t, and a graph G n of size at most ⌊n/N ⌋, such that the following properties are met: n , and Then we say that G n is an (n, k, N, T )-cheating graph for the pair (P, F) or, more generally, that F admits an (n, k, N, T )-cheating graph for P.
Remark 1. Being P an LVL problem, Definition 5.6.(ii).(b)implies that P is solvable on We now present our general lower bound theorem.
Theorem 5.7.Let P be an LVL problem with checking radius t, and F be a graph family that admits an (n, k, N, T )-cheating graph for P. Suppose O is an outcome over F in NS-LOCAL with locality T ≥ t.Then, there exists a graph H ∈ F on n vertices such that the probability of O solving P on H is at most (1 − 1/k) N .Furthermore, H can be chosen among the graphs in the family Definition 5.6

.(ii).(b).
Proof.Let G n be a (n, k, N, T )-cheating graph for (P, F).We know that G n has size at most ⌊n/N ⌋ and satisfies the properties listed in Definition 5.6.Now, consider a new graph that consists of N disjoint copies G n,1 , . . ., G n,N of G n , and n − |V (G n )| isolated nodes.
For each i = 1, . . ., N , consider the subgraph cover {G for G n,i given by Definition 5.6.Let O be any outcome having locality T and solving problem P over F.
As P is not solvable on G n,i , then the failing probability of O over G n,i is 1, for each i = 1, . . ., N .Consider one of the G n,i and notice that, if O produces a permissible vector output on for each j ∈ [k] at the same time, then, by definition of LVL problem, we have a global permissible vector output on G n,i (which does not exist).Hence, by Definition 5.6.(ii).(a),there must exist n,i ) and the output vector on N t (v) is not permissible: in such a case, we say that G (j) n,i contains a bad node.We now prove that there exists a sequence of indices n,j with probability at least 1 − (1 − 1/k) N .If we had independence between "far away" parts of the graphs (as in the rand-LOCAL model), this thesis would be trivial (see Section 5.2.1).However, in the NS-LOCAL model non-trivial dependencies are possible (e.g., pre-shared quantum state).
We here present a shorter proof by induction on N , as we already gave a somewhat "constructive" intuition in Section 2.2.Assume N = 1: as O fails on G n,1 with probability 1, and the latter is covered by {G contains a bad node is at least 1/k.Now, assume N > 1 and the claim to be true for n,i (we remark that P is solvable on n,i by Remark 1): the inductive hypothesis can be rewritten as Pr ∪ N i=1 E < 1 otherwise the thesis is trivial.Let us denote the complement of any event A by Ā.As O fails on G n,N with probability 1, we know that ) .
By the law of total probability, we get that ; by the union bound, it follows there exists Then, by the inclusion-exclusion principle and the law of total probability, proving the claim.By Definition 5.6.(ii).(b),there exists a graph H x N ∈ F over the same set of n nodes that contains a subgraph Hx N with Consider the same identifiers and input labels over n,i )]: by the definition of NS-LOCAL, the probability that O fails on Hx N ⊆ H x N is the same as that on N j=1 G (x j ) n,j , yielding the thesis.
As long as one can find a cheating graph for a pair (P, F), where P is an LVL problem and F an input graph family, the lower bound technique can be applied.In Sections 5.3-5.5, all the analysis that is carried out serves to show that there exists such a cheating graph for, respectively, c-coloring χ-chromatic graphs, 3-coloring grids, and c-coloring trees.

Lower bound for c-coloring χ-chromatic graphs
The goal of this section is to prove Theorem 1.3 by showing that the family of χ-chromatic graphs admits cheating graphs for the c-coloring problem (Definition 5.6).We restate the theorem for the sake of readability.
Suppose A is an NS-LOCAL algorithm for c-coloring graphs in the family F of χ-chromatic graphs of n nodes with success probability q > ε.Then the running time of A is at least We base our analysis on [16,Theorem 1.2], a result that has gone relatively unnoticed and lies at the intersection between graph theory, combinatorics, and topology, which ensures the existence of a graph with high chromatic number which, locally, is χ-chromatic.The first half of the sections aims at constructing such graph, and the second half is devoted to the proof that this graph admits a (small enough) subgraph cover that satisfies the properties of Definition 5.6.
Preliminaries.We first define some graph operations.For any two graphs G, H, we define the intersection graph G ∩ H as a graph whose vertex set is the set V (G) ∩ V (H), and whose edge set is the set E(G) ∩ E(H).Similarly, we define the union graph G ∪ H as a graph whose vertex set is the set V (G) ∪ V (H), and whose edge set is the set E(G) ∪ E(H).We define the difference graph G \ H as the subgraph of G induced by V (G) \ V (H).The T -local chromatic number of a graph G, denoted by LX T (G), is the minimum c ∈ N such that the graph induced by the T -neighborhood of any node is c-colorable.More formally Given two graphs G and H, a function f : Notice that the composition of homomorphisms is a homomorphism: hence, if G is homomorphic to H, then X (H) ≥ X (G).Furthermore, we define the tensor product of graphs G and H as a graph G × H whose vertex set is V (G) × V (H), and whose edge set is determined by the following: for any (g, h), (g Remark 2. [16, Theorem 1.2] has been stated for χ ≥ 3 since the result for χ = 2 was already known from a different construction provided by [62] (a proof translated in English was reproduced by [37]).Nevertheless, the proof of [16,Theorem 1.2] also holds for the case χ = 2.
Remark 3. Theorem 5.8 holds even for r ≥ 1, as explicitly mentioned at the end of [16,Section 4], slightly changing the proof.
We will discuss the tightness and the related works of this result already in Section 5.3.2.We proceed proving Theorem 1.3.The idea of the whole proof is to show that the graph from Theorem 5.8 provides a cheating graph for the c-coloring graphs problem and the family of χ-chromatic graphs.
We now construct the graph from Theorem 5.8.
The r-join of graphs.Given two graphs G and H, we aim to define the r-join operation G ⋆ r H, for any r ≥ 0. The vertex set Furthermore, let to v j+1 for any 0 ≤ j ≤ i − 1, and v i = u.Similarly, set w r+1 = h, w j = (g ′ , h ′ , j) for any odd j, and w j = (g, h, j) for any even j.We have that w j is connected to w j−1 for any i + 1 ≤ j ≤ r + 1, and w i = u.
We are ready to define the graph from Theorem 5.8 which we will prove to be a cheating graph for the family of χ-chromatic graphs.
See Fig. 3 for some examples (notice that r = 1 in the examples: as observed in Remark 3, the result still holds in this case).
Further discussion on G k is deferred to Section 5.3.1.In order to continue, we state the following lemma for an induced subgraph, whose proof is trivial.Lemma 5.12.Let T ∈ N. Let G be a connected graph, and For a graph G and any node v ∈ V (G), we define by dist G (v, H) = min u∈V (H) {dist G (u, v)} the distance between v and any subgraph H ⊆ G.We write dist(v, H) when the underlying graph G is clear from the context.We now prove some key-properties of G k which allow us to show that G k is a cheating graph.
Proof.We prove the thesis by induction on k.Remember that G 2 is obtained by the 2r-join of two disjoint copies χ is connected to the first copy of K (1) χ , and Consider the graphs G 2 ] and G (2) 2 ].The cover property and properties (i)-(iii) are straightforward.Clearly, G 2 = ∪ i∈ [2] 2 )]) = χ as the projections pr 2 )] are homomorphisms.Moreover, it is easily verifiable that G χ (by the inductive hypothesis) and K (2) χ are connected, Lemma 5.10 implies their join is connected.Then, Lemma 5.12 implies G (1) 2 is connected.The same applies for G χ )].As for property (iv), consider any two nodes u ∈ V (K 2 )] and has distance T from 2 )] and has distance T from G (2) 2 .Let k ≥ 3 and assume the thesis is true for G k−1 .We now construct the subgraph cover of Then, we define We now prove that the family {G k ) (by observing that the latter contains V (G k−1 ) and by using the inductive the hypothesis on the subgraph covering) or to χ ), and j ∈ [2r].By the inductive hypothesis, there exists i ∈ The path w 0 w 1 . . .w j defined by w 0 = u 1 , w k = (v 1 , v 2 , k) for 1 ≤ k ≤ j odd, w k = (u 1 , u 2 , k) for 2 ≤ k ≤ j even connects u 1 to v if j is odd.The path w 0 w 1 . . .w j defined by As u and v are connected, it means that uv 1 is an edge in G k−1 .By the inductive hypothesis on the subgraph covering, there exists i ∈ From the inductive hypothesis on the subgraph covering, there exists

Property (i). Consider G (i)
χ as a subgraph, χ colors are also necessary.

Similarly, consider G
k−1 (by the inductive hypothesis (ii)) and K (k) χ are connected, their join is connected.Then, Lemma 5.12 implies that G

Property (iii). Consider any node
Remark 5. We highlight that the subgraph cover from Lemma 5.13 is not unique: other families of graphs could be used obtaining a shorter and simpler proof at the expense of a higher number of graphs in the family (e.g., a number of subgraphs that is exponential in k).The latter would result in a worse bound on T in the next corollary.
We are now ready to show that the family of χ-chromatic graphs admits a cheating graph for the c-coloring graphs problem.
Proof.There exists a unique integer r ≥ 3 such that We claim that the graph G k+1 defined in Definition 5.11 by iterating 2r-join operations is an (n, k + 1, N, T )-cheating graph for (P, F).Clearly, P is not solvable on G k+1 , while property Definition 5.6.(ii).(a)follows by Lemma 5.13.(iii) (since coloring is an LVL problem with checking radius t = 1).
For any choice of indices x N = (x 1 , . . ., x N ) ∈ [k + 1] N , we now show that there exists a connected graph H x N ∈ F on n nodes that admits a subgraph Hx N such that (1) k+1,j+1 ) = T (such nodes exist because of Lemma 5.13.(iv)).All remaining nodes form a path of which one endpoint is connected to any node in G k+1,N [N T (G k+1,N .Property (1) follows by construction.As for property (2), we observe that the chromatic number of k+1,j ) is χ-chromatic by Lemma 5.13.(i), and χ ≥ 2. Furthermore, H x N is connected by Lemma 5.13.(ii) combined with Lemma 5.12 and observing that connected disjoint connected components through paths.
The proof of our main lower bound now follows easily.
Proof of Theorem 1.3.Let k = α and . By Corollary 5.14, F admits an (n, k + 1, N, T )-cheating graph for the c-coloring graphs problem for any n ≥ ((6χ + 1) k+1 − 1)N/6 and for some Theorem 5.7 implies that there is a connected graph H ∈ F on n nodes such that the probability that any outcome O with locality T is c-coloring H is at most (1 − 1/(k + 1)) N .We remind the reader that the graph can be chosen as in Definition 5.6.(ii).(b),and Corollary 5.14 implies that such graph is connected.By definition of k and N , this probability is at most ε.By observing that log , we get the thesis.

Notes on the graph G k
This section is dedicated to the reader that is interested in the topological elements underlying the construction of the graph G k in [16].Familiarity with the notion of topological space, join of topological spaces, homotopy equivalence, abstract simplicial complex, and geometric realization of an abstract simplicial complex is required.Every graph G can be associated with an abstract simplicial complex N C(G) called the neighborhood complex, defined as follows: that is, N C(G) consists in all subsets A ⊆ V of nodes that have a common neighbor.Let us denote the geometric realization of N C(G) by ∥N C(G)∥ (the geometric realization is unique up to homeomorphisms).We say that a non-empty topological space X is m-connected if each continuous map π : S i−1 → X extends to a continuous map π : D i → X for each i = 1, . . ., m, where -dimensional sphere, and Lovász [51] proved the following theorem.
Theorem 5.15 ([51]).Let G be any graph such Lovász's result provides a clear and effective tool to bound from below the chromatic number of a graph.Bogdanov [16] linked the r-join operation between graphs to the join operation of topological spaces, proving the following lemma.
By applying recursively the above procedure, one can understand why the graph G k from Definition 5.11 has chromatic number bounded from below by k(χ − 1) + 1.
As for the properties of the local chromatic number, Bogdanov [16] also showed the following simple result, which can be proved by using the projections pr G and pr H , that immediately gives the desired result.Lemma 5.17 ([16]).Let G, H be two graphs.For every r ∈ N + it holds that LX r (G ⋆ 2r H) = min{LX r (G), LX r (H)}.

On highly chromatic graphs with small local chromatic number
In this section we present previous works studying the relation between the chromatic number and the local chromatic number of a graph.We remind to the reader that the local chromatic number LX r (G) of radius r of a graph G is the maximum number of colors required to color any r-ball in G.We remark that, in this paper, we depart from the standard definition of local chromatic number in the literature that started with [29], which is not useful for our purposes.For positive integers c, χ, r, define f χ (c, r) to be the maximal integer n such that every graph G of n nodes with LX r (G) ≤ χ is c-colorable.The first bounds on f χ (c, r) held for specific values of χ, especially χ ∈ {2, 3} [14,28,62] (for a summary of the progress of such results we defer the reader to [3,17]).We only describe the results that consider a generic χ, as they are more directly related to our work.Kierstead et al. [44] proved that, for each k ∈ N, i.e., any graph G having LX r (G) ≤ χ and |V | ≤ ⌊r/(2k)⌋ k has chromatic number at most k(χ−1)+1.Thirty years later, Bogdanov [16] proved that this result is basically tight when k, χ are fixed: it showed that, for each k ≥ 2, i.e., there exists a graph : such result is built upon a lemma by Lovász [51].The two above results show an interesting phenomenon.For constant positive integers χ and c, the minimum number of vertices of a graph G with LX r (G) ≤ χ and χ(G) = c is roughly r , that is, it jumps to the powers of r where the exponents are the values c congruent to 1 modulo χ − 1.
The estimate by Kierstead et al. [44] breaks when k ≳ r.Bogdanov [17] and Alon and Ben-Eliezer [3] investigated and provided better lower bounds to f χ (c, r) when c ≳ (χ − 1)r; Alon and Ben-Eliezer [3] estimated also upper bounds to f χ (c, r) which are roughly tight for fixed r.We do not discuss such results in details as they are not useful for our purposes: we are interested in the asymptotic relation between the locality radius r and the number of nodes n given the local and global chromatic number.Such case is covered by Eq. ( 2): in particular, we used the example graph that was built in [16] as our baseline for the lower bound proof.
We remark that our result for the specific case χ = 2 could also be achieved through other graphs that were studied in the literature, e.g., the generalized Mycielski graph (studied by Stiebitz [62], see [37] for an English version of the proof).

No quantum advantage for 3-coloring grids
In this section we prove that 3-coloring n 1 × n 2 grids of requires time Ω(min(n 1 , n 2 )) rounds in the NS-LOCAL model, by using the same graph-theoretical lower bound argument of Section 5.
For any two integers a ≤ b, we denote the set {a, a + 1, . . ., b} by [a : b].

Quadrangulation of the Klein bottle.
To establish the truth of our claim, we make use of results on the chromatic number of quadrangulations of surfaces [5,58,59].Following the preliminaries of [59], by surface we mean a compact connected 2 manifold without boundary.A quadrangulation of a surface is a graph without self-loops on that surface with all faces being quadrilaterals.For our purpose, we can restrict to the class of simple graphs.Let G be a quadrangulation of a surface S.An orientation of a face of G is a closed walk along its boundaries.Given two oriented faces sharing at least one boundary edge e, we say that the consistency of the orientation is not broken at e if we traverse edge in opposite directions when considering the two orientations of the faces.Given an arbitrary orientation of all faces we can count the edges that break consistency of the orientation of the surface.Note that reversing the orientation of a face changes the status of its four edges, and thus the parity of the number of edges breaking consistency does not change.We say that G is even or odd depending on this parity.By the previous remark, the parity of the quadrangulation is determined by any orientation of all faces and is invariant, i.e., it depends only on the graph G.
The following theorem was independently proved by [5,58], but we report its formulation by [59].
Proof.Notice that G n 2 ,n 1 is isomorphic to G n 1 ,n 2 , hence we focus on the latter.It is easy to see that G n 1 ,n 2 is a quadrangulation of the Klein-bottle.Indeed, H n 1 ,n 2 is a quadrangulation of the square [0, 1] × [0, 1]: it suffices to map the nodes (i, j) into (i/n 1 , j/n 2 ), for all i, j.As G n 1 ,n 2 is obtained by node identification through the relation ∼ V , which acts on the border of the square, we have a quadrangulation of the Klein-bottle.Let's orient the faces of G n 1 ,n 2 as in Fig. 4.Even though in H n 1 ,n 2 the orientation is even as consistency is never broken, through simple observations, one can verify that G n 1 ,n 2 is an odd quadrangulation if and only if n is odd, as the orientation consistency is broken only at edges belonging to the set {{(i, 0), (i + 1, 0)} | 0 ≤ i ≤ n 1 − 1}.Then Theorem 5.19 gives the thesis.
It suffices to construct a suitable subgraph cover for G n 1 ,n 2 which shows that the family of grid graphs admits cheating graphs for the 3-coloring problem.First, let us denote by Q n 1 ,n 2 any subgraph of the infinite two-dimensional lattice that is isomorphic to an (n 1 + 5)/2 × (n 2 + 5)/2 grid, and by Q T n 1 ,n 2 the graph induced by its T -neighborhood in the lattice.Lemma 5.21.Let 5 ≤ n 1 , n 2 be odd integers, and let G n 1 ,n 2 be the graph defined in Definition 5. 18.Let T = ⌊ min(n 1 ,n 2 )−5 4 ⌋.There exists a subgraph cover {G (i) [4] of G n 1 ,n 2 such that the following statements hold: (iii) For each v ∈ V (G n 1 ,n 2 ), there exists i ∈ [4] such that N 1 (v) ⊆ V (G (i) n 1 ,n 2 ).Proof.Observe that claim (ii) implies claim (i): hence, we prove claim (ii).Consider the graph H n 1 ,n 2 used to construct the KB-gadget in Definition 5.18.Let V 1 be the set of nodes (i, j) of H n 1 ,n 2 respecting the following property Similarly, we define V 2 , V 3 , V 4 by the following properties : n 1 , j ∈ [0 : 1] or i ∈ 0 : For an example of V 1 , V 2 see Fig. 5.Then, define S i to be subset of V (G n 1 ,n 2 ) induced by V i after applying the equivalence relation ∼ V in V (H n 1 ,n 2 ), and n 1 ,n 2 is isomorphic to Q n 1 ,n 2 for each i ∈ [4].Furthermore, the T -view of G (i) n 1 ,n 2 identifies in H n 1 ,n 2 four graphs that do not intersect, implying claim (ii).Claim (iii) is trivial.Remark 6.The strange shape of the subgraph cover is needed for property (iii).However, the analysis carries on by considering "simpler" subgraph covers which don't meet property (iii) as a valid coloring can be checked by looking at single edges, not necessarily entire neighborhoods.Nevertheless, we choose to use the general result for LVLs (Theorem 5.7).Lemma 5.21 implies that the family of grid graphs admits cheating graphs for the 3-coloring graphs problem.    1.There exists a constant λ such that, for each n ≥ 1, there is a prime p ∈ [n + 1 : (1 + λ)n] such that p ≡ 1 (mod 4); 2. If p is a prime such that p ≡ 1 (mod 4), there exist infinitely many primes q such that q ≡ 1 (mod 4) and p (mod q) is a non-zero perfect square; Proof.For the first result, we follow the same approach of [64].Let us focus on the second claim: we use the same approach as in [53].By the law of quadratic reciprocity [38,Theorem 98], p (mod q) is a perfect square if and only if q (mod p) is a perfect square.Consider the arithmetic progression 1 + m(4p) for m ≥ 1.By Lemma 5.24,there exist infinitely many m such that 1 + 4pm is a prime.Set q = 1 + 4pm.Now, q ≡ 1 (mod 4) and q ≡ 1 (mod p), which implies that p ≡ 1 (mod q).Through Lemma 5.23 and Corollary 5.25, we can show that the family of trees admits a cheating graph for the c-coloring problem.
3) of Definition 2.1 ensures that we can construct an admissible instance H x as shown in the figure, with |V (H x )| = n.By the properties of the NS-LOCAL model, since H x and i∈[N ] G i share the same local view around j∈[N ] G (xj ) j, A fails on H x too with at least the same probability.

Figure 1 :
Figure 1: Illustration of the lower-bound argument based on the cheating graph G.For n = |V (G)|•N , the problem is solvable in each T (n)-radius neighborhood of G (i) , i ∈ [9], but not on G.
Hence, since ψ is a proper coloring, every node in N (v) colored by ψ has a color that is different from −1.Since φ is a restriction of ψ, the same holds for any node in N (v) colored by φ. v / ∈ A ∪ N (A).Then v is only adjacent to nodes in N (A).Since no node in N (A) is colored −1 by φ by definition, clearly v has no neighbor colored −1 by φ.

11 : 19 :
leader i broadcasts φ i to all nodes in C i ∪ N (C i ) 12: end if 13: end for 14: for each node u in parallel do 15: Φ u ← the set of all colors assigned to u by the φ i 16: if Φ u = {−1} then 17: Color u with the color −1 18: else Color u with an arbitrary color from Φ u \ {−1} 20: a be the set of colors actually used by the nodes to color clusters with the color a.By the χchromaticness of G, every minimal coloring of any induced subgraph of G uses at most χ colors, so we have |p ′ a | ≤ χ.Since the intersection of two palettes p a and p b is exactly {−1} for a ̸ = b, we use α a=1 p ′ a = |{−1}| + α a=1 p ′ a \ {−1} ≤ α(χ − 1) + 1 colors in total, as claimed.

Lemma 4 . 6 .
−1 to 1 and every pair (a, b) ∈ p a to α(b − 1) + a + 1. (Note this gives a bijection between {−1} ∪ a∈[α] p a and N + .)Since we use only the smallest elements of the palette p a in each respective coloring φ i , we know that any (a, b) ∈ p ′ a must be such that b ≤ χ − 1. Hence the largest color used is α(χ − 2) + α + 1 = c (corresponding to (α, χ − 1) ∈ p α ).Given the (α, d(n))-network decomposition D, Algorithm 1 can be run distributedly in O(d(n)) rounds in the det-LOCAL model.Proof.The only lines in the algorithm that require communication are Lines 2, 3 and 11.Since our clusters have diameter at most d(n), Line 2 requires only O(d(n)) rounds of communication.The other two trivially take only d(n) + 1 rounds since the message size in the det-LOCAL model is unbounded and also any node in N (C i ) has distance at most d(n) + 1 to its respective leader i .All other steps in the algorithm are local computations that do not incur any cost in the det-LOCAL model.Together, Lemmas 4.5 and 4.6 prove Algorithm 1 satisfies the requirements of Theorem 4.3, thus concluding its proof.

5 :U 7 : 9 :
← U \ C∈C i C 6: end for Form a clustering C α of G[U ] by having each connected component form its own cluster 8: Color every cluster C ∈ C α with the cluster color µ(C) = α Each node v ∈ C outputs its cluster color µ(C) the output distributions corresponding to these inputs have identical marginal distributions on the set S. Notice that T can depend on the input labeled graph.Definition 5.4 is also the more general definition for the locality of an outcome: an outcome O has locality T if it is non-signaling beyond distance T .The NS-LOCAL model.The non-signaling LOCAL (NS-LOCAL) model is a computational model that produces non-signaling outcomes.Let p ∈ [0, 1].The complexity class N S[T, p] is defined by all pairs (P, F) where P is a problem and F is a graph family such that there exists an outcome O that is non-signaling beyond distance T which solves P over F with probability at least p.If p = 1, we just say that (P, F) ∈ N S[T ].

Lemma 5 . 13 .
Let χ ≥ 2, r ≥ 3, and k ≥ 2 be integers.Define T = ⌊ 2r 3 ⌋.Let G k be the graph defined in Definition 5.11 built with copies of K χ .There exists a subgraph cover {G(i) k } i∈[k]  of G k such that the following statements hold:

Lemma 5 .
16 ([16]).Let G, H be any two graphs, and r∈ N + .Then, ∥N C(G ⋆ r H)∥ ≃ ∥N C(G)∥ ⋆ ∥N C(H)∥, where ≃ means homotopy equivalent and the latter ⋆ operator represents the join operation between topological spaces.Note that, for any three topological spaces A, A ′ , and B, ifA ≃ A ′ , then A ⋆ B ≃ A ′ ⋆ B[56].Theorem 5.15 and Lemma 5.16 can be combined to construct graphs of high chromatic number.E.g., if G = K(1)

Figure 4 :
Figure 4: Representations of graphs G n 1 ,n 2 and H n 1 ,n 2 with n 1 = n 2 = 3. Nodes on the borders are identified in G n 1 ,n 2 as indicated by the same colors.Each face of the graph is oriented the same as the oriented circle inside it.

Figure 5 :
Figure 5: Two elements of the subgraph cover of the KB-gadget for n 1 = n 2 = 7. Nodes with identical colors are identified.V 3 and V 4 can be obtained by symmetries.

Lemma 5 . 26 .
Let c ≥ 2 be an integer.For every N ≥ 1 and infinitely many n ∈ N, there exists a value T = Θ log c n N such that the family F of all trees of size n of height at most 2T + 3 admits an (n, k = n/N, N, T )cheating graph for the c-coloring problem.
Then, Pr[A fails on H N ] ≥ 1 − (1 − 1/9) N by independence, where H N is an admissible instance.As long as |V (H N )| ≤ n, this gives the lower bound.
Now, clearly, G n 1 ,n 2 is everywhere locally grid-like; however, we claim that X (G) ≥ 4 if n 1 is odd.