Hedetniemi's conjecture and adjoint functors in thin categories

We survey results on Hedetniemi's conjecture which are connected to adjoint functors in the"thin"category of graphs, and expose the obstacles to extending these results.


Introduction and terminology
Hedetniemi's conjecture states that the chromatic number of the product of two graphs is equal to the minimum of the chromatic numbers of the factors. This conjecture has caught the attention of some category theorists, who recognize their turf in its setting of morphisms and products in Cartesian closed categories. Yet, the input from category theorists has not gone much beyond reformulations of the problem and rediscovery of basic results.
In our opinion there is potential for deeper input. However, we will argue that the consideration of the graph-theoretic product in terms of its categorial definition is perhaps the wrong approach. The most relevant property of the product seems to be that it is preserved by right adjoint functors. Indeed, significant results have been proved using this fact.
The purpose of this paper is to expose this point of view in greater detail. However for this purpose, we will need to escape the usual category of graphs and homomorphisms and work instead in thin categories.
1.1. Graphs, digraphs, homomorphisms and chromatic numbers. It will be useful to consider both finite graphs and finite digraphs. A digraph G consists of the following data: 1. a finite set V (G) of vertices; 2. a finite set A(G) of arcs; 3. two maps t G , h G : A(G) → V (G) that determine the tail and the head of each arc. For a ∈ A(G), we say that it is an arc from t G (a) to h G (a) and write a : t G (a) → h G (a).
A digraph is simple if for any u, v ∈ V (G) there is at most one arc from u to v. For simple digraphs it is common for an arc u → v to be called (u, v), and so A(G) is usually viewed as a binary relation on V (G). When A(G) is symmetric, G is called a graph; an edge of G is a pair {u, v} with (u, v), (v, u) ∈ A(G).
A homomorphism f of a digraph G to a digraph H is a pair of maps (f 0 , f 1 ), f 0 : V (G) → V (H), f 1 : A(G) → A(H), such that both of the following diagrams commute: Homomorphisms are composed component-wise. Where H is simple, f is usually identified with f 0 because f 1 is uniquely determined by it. The product G × H of two digraphs is given by This construction, along with the obvious projections, is the product in the category of digraphs and their homomorphisms. Note that symmetry is preserved by the product, that is, the product of two graphs is a graph.
The complete graph K n on n vertices is the (simple) graph defined by We will also be interested in the characterization of so-called "multiplicative graphs". A graph K is called multiplicative if the existence of a homomorphism of a product G×H to K implies the existence of a homomorphism of a factor G or H to K. Hence Hedetniemi's conjecture is equivalent to the statement that the complete graphs are multiplicative. However, few graphs have yet been shown to be multiplicative.
1.3. Structure of the paper. The paper is structured as follows. In Section 2 we present the thin categories of graphs and digraphs, and functors relevant to our discussion. In Section 3 we give an exposition of some of the strange results on Conjecture 1.2 obtained using such functors. In Section 4 we see how attempts to improve on the previous results have led to proving the multiplicativity of some "circular complete graphs"; we then discuss apparent limitations of the method of adjoint functors towards Hedetniemi's conjecture. We provide proof of the results which we can prove using properties of adjoint functors, and we propose open problems throughout the paper.
2. Adjoint functors in the thin categories of graphs and digraphs 2.1. Thin categories. In the context of colourings and Hedetniemi's conjecture, the existence or non-existence of a homomorphism between two graphs or digraphs seems to be more relevant than the structure of homomorphisms between them. Hence we focus our study on thin categories. A thin category is a category in which there is at most one morphism between any two objects. The thin category of digraphs (resp. graphs) is the category whose objects are digraphs (resp. graphs) and where there is a unique morphism from G to H if there exists a homomorphism of G to H. If a homomorphism of G to H exists, we write G → H; we write G → H if it does not. Thus → is a binary relation on the class D of finite digraphs or on its subclass G of finite graphs. More precisely, → is a preorder on G and on D, and Hedetniemi's conjecture can be naturally stated in terms of the quotient order G/↔. The relation ↔ is naturally defined by G ↔ H if G → H and H → G; G and H are then called homomorphically equivalent. Note that G and H are homomorphically equivalent if and only if they are isomorphic in the thin category.
Both G/↔ and D/↔ are distributive lattices with (G/↔) ∧ (H/↔) = (G × H)/↔. (In fact, both G/↔ and D/↔ are Heyting algebras because the categories of graphs and digraphs are Cartesian closed.) An element K of G (resp. D) is multiplicative if and only if K/↔ is meet-irreducible in G/↔ (resp. D/↔). Thus Hedetniemi's conjecture is equivalent to the statement that for every complete graph K n , K n /↔ is meet-irreducible in G/↔. This reformulation does not simplify matters since the structure of G/↔ is complex, and in particular, dense above K 2 (see [17]).
In thin categories, adjunction comes down to an equivalence between two existential statements. Specifically, two functors L, R between thin categories (G or D) are respectively left and right adjoints of each other if the following property holds: This condition is weaker than adjunction in the usual categories with all homomorphisms. Thus all adjoint pairs in these categories (which we present next) are also adjoint pairs in the thin categories, but there are other thin adjoints as we shall see in 2.5.

2.2.
Pultr template and functors. The functors relevant to our discussion are connected to those introduced by Pultr in [22]. Pultr worked, among others, in the usual categories of graphs and digraphs (as defined in 1.1). He characterized the adjoint functors in these categories by means of the following construction.
(i) A Pultr template is a quadruple T = (P, Q, ǫ 1 , ǫ 2 ) where P , Q are digraphs and ǫ 1 , ǫ 2 homomorphisms of P to Q. (ii) Given a Pultr template T = (P, Q, ǫ 1 , ǫ 2 ), the left Pultr functor Λ T is the following construction: For a digraph G, Λ T (G) contains one copy P u of P for every vertex u of G, and for every arc a : u → v of G, Λ T (G) contains a copy Q a of Q with ǫ 1 (P ) identified with P u and ǫ 2 (P ) identified with P v . (iii) Given a Pultr template T = (P, Q, ǫ 1 , ǫ 2 ) the central Pultr functor Γ T is the following construction: For a digraph H, the vertices of Γ T (H) are the homomorphisms f : P → H, and the edges of Γ T (H) are the homomorphisms g : Q → H; such g is an arc from f 1 = g • ǫ 1 to f 2 = g • ǫ 2 . For any Pultr template T , Λ T : D → D and Γ T : D → D are respectively left and right adjoints of each other (see [22,7]). We use the adjective "central" rather than "right" for Γ T because in some important cases, Γ T itself admits a right adjoint in the thin category.
The right adjoints into G need an additional property of the Pultr template: both P and Q must be graphs, and moreover, Q must admit an automorphism q with q • ǫ 1 = ǫ 2 and q • ǫ 2 = ǫ 1 . The existence of such an automorphism makes the conditions in (ii) and (iii) symmetric, so that Λ T (G) is well defined, and Γ T (H) is a graph rather than a digraph. (For the adjunction to work in the "non-thin" category of graphs and homomorphisms, we need to allow graphs with multiple edges.) We now give a simple example of a Pultr template, which will be useful later on. We let T (3) = (K 1 , P 3 , ǫ 1 , ǫ 2 ), where P 3 is the path with three edges and ǫ 1 , ǫ 2 map K 1 to the endpoints of P 3 . Then Λ T (3) (G) is obtained from G by replacing each edge by a path with three edges, and Γ T (3) (H) is obtained from H by putting edges between vertices joined by a walk of length 3 in H. In particular, if H is the cycle with five vertices, then Γ T (3) (H) is the complete graph K 5 . The adjunction between Λ T (3) and Γ T (3) then reads as follows.
For any graph G, G is 5-colourable if and only if Λ T (3) (G) admits a homomorphism to the five-cycle. This is a "complexity reduction" of the problem of determining whether a graph is 5-colourable to the problem of determining whether a graph admits a homomorphism to the five-cycle. Since the former problem is NP-complete, so is the latter. Using an arsenal of similar reductions, Hell and Nešetřil [16] eventually proved that for any non-bipartite graph H without loops, the problem of determining whether a graph admits a homomorphism to H is NP-complete.

"Graph products" and variants on Hedetniemi's conjecture.
Other examples of Pultr templates and functors are connected to the various "products" used in graph theory. Indeed there are many useful ways to define an edge set on the Cartesian product V (G) × V (H) of the vertex sets of simple graphs G and H. We consider the following examples from [18].
• The direct product G × H is the product we introduced earlier: it For a fixed graph H and ⋆ ∈ {×, , •}, let T (⋆, H) be the Pultr template (H ⋆ K 1 , H ⋆ K 2 , ǫ 1 , ǫ 2 ) where ǫ 1 (u) = (u, 0) and ǫ 2 (u) = (u, 1). Then for any graph G, Λ T (⋆,H) (G) = G ⋆ H. Therefore any product by a fixed graph is a left adjoint, and the equivalence holds. In particular, Γ T (×,H) (K) is the exponential graph usually denoted K H , which is the exponential object in the category G. Exponential graphs connect to Hedetniemi's conjecture as follows: Thus, Hedetniemi's conjecture is equivalent to the statement that Similar observations hold in the case of the other graph products: This is the Sabidussi identity, which is sometimes viewed as a companion formula to Hedetniemi's conjecture.
There is no corresponding formula for the lexicographic product. In general, we have [9]). Thus χ(G • H) does not depend on the structure of H, just on its chromatic number. However, the structure of G is relevant. We have G , and the latter is homomorphically equivalent to the Kneser graph K(n, m) defined by The name of these graphs is derived from Kneser's conjecture which states that χ(K(n, m)) = n−2m+2. Kneser's conjecture was proved by Lovász [19] in a famous paper which introduced the field of "topological bounds" on the chromatic number.

Products, functors and Hedetniemi's conjecture.
Our examples so far have situated Hedetniemi's conjecture within the more general problem of determining the chromatic number of graphs of the type R(K n ), where R is some right adjoint in the thin category of graphs. These questions can be easy or hard, but there seems to be nothing special about the fact that × is the "correct" product in the category of graphs. The special property of × that will be relevant to our discussion is the fact that thin right adjoints commute with it, up to homomorphic equivalence: Here, × cannot be replaced by the various "products" of graph theory. It is a property that can be used to say something meaningful about Hedetniemi's conjecture, when suitable functors are found, as shown in Section 3.

2.5.
A right adjoint of a central Pultr functor. Our last example of this section introduces a functor that we will use in Section 4. It is our first example of a thin right adjoint that is not a central Pultr functor. It can be motivated by the following question.
Does there exist, for every integer n, a graph G n such that χ(G n ) = n and G n admits a n-colouring f : G n → K n where the neighbourhood of each colour class is an independent set?
In [10], Gyárfás, Jensen and Stiebitz present the question as a strengthening of a question of Harvey and Murty. Note that if two adjacent vertices u 1 and u 2 are respective neighbours of two identically coloured vertices u 0 and u 3 , then u 0 , u 1 , u 2 , u 3 is a walk of length three between two identically coloured vertices. Therefore the above question can be reformulated in terms of the functor Γ T (3) introduced in the first example of this section.
Does there exist, for every integer n, a graph G n such that Once again, the question reduces to determining the chromatic number of some thin right adjoint of K n . The question is answered in the affirmative in [10].
2.6. Excursion: Topological bounds on chromatic numbers. As an aside note, we mention that the method for finding a lower bound on χ(Ω T (3) (K n )) in [10] is the topological method devised by Lovász [19] for Kneser graphs. Thus the "topological bounds" are tight for chromatic numbers of Ω T (3) (K n ) and Γ T (•,Km) (K n ). They are also trivially tight for Γ T ( ,H) (K n ), and Hedetniemi's conjecture would imply that they are also tight for Γ T (×,H) (K n ). Since it is easy to devise graphs for which such topological bounds are not tight, it is intriguing to see it work well in the case of many right adjoints.
Is it a coincidence? In modern terms, the bounds can be defined via functors which associate to a graph a "hom-complex" and eventually an object in a category of topological spaces. This point of view is developed in Section 7 of [4]. Perhaps the effectiveness of the topological bounds can be traced functorially in some way. As far as we know, there is no unifying theory as to when topological bounds on the chromatic number are tight.
3. The arc graph construction 3.1. Multiplicative complete graphs in G and D, and the Poljak-Rödl function. The complete graphs K 1 and K 2 can be shown to be multiplicative both in G and D with relatively straightforward arguments. The first nontrivial case of Hedetniemi's conjecture was established by El-Zahar and Sauer in the aptly named paper [5].
However, a few years before, Poljak and Rödl [21] had proved that for every n ≥ 3, there are digraphs G n and H n both with n + 1 vertices and chromatic number n + 1, such that χ(G n × H n ) = n. Thus K n fails to be multiplicative in D for any n ≥ 3. This justifies some skepticism towards Hedetniemi's conjecture.
Let φ : N → N be defined by Hedetniemi's Conjecture 1.1 is that φ(n) = n for all n ∈ N, but so far this has been verified only for n ≤ 4. The weak Hedetniemi Conjecture 1.2 is that φ is unbounded. Poljak and Rödl proved the following result. This result and its developments, along with their relationship with adjoint functors, are the topic of this section.
3.2. The arc graph construction and its chromatic properties. Let P i be the directed path with i arcs, that is, the digraph defined by V ( P i ) = {0, . . . , i} and A( P i ) = {(0, 1), . . . , (i − 1, i)}. Consider the Pultr template T = ( P 1 , P 2 , ǫ 1 , ǫ 2 ), where ǫ 1 and ǫ 2 map the arc of P 1 respectively to the first and the second arc of P 2 . Then Γ T : D → D is the arc graph construction usually denoted δ: The vertices of δ(G) = Γ T (G) are the arcs of G, and its arcs correspond to pairs of consecutive arcs of G. We will use the following property of the arc graph construction.
Proof. We present a proof using adjoint functors. It turns out that δ has a thin right adjoint, which we will call δ R . The vertices of δ R (G) are the ordered pairs (U, V ) of sets of vertices of G such that (u, v) ∈ A(G) for all u ∈ U and v ∈ V . The arcs of δ R (G) are the ordered pairs ((U, V ), (W, X)) such that V ∩ W = ∅. Note that the vertices of δ R (K n ) are the ordered pairs (U, V ) of disjoint sets of vertices of K n . The map f : is an endomorphism, with image of size 2 n . Therefore χ(δ R (K n )) ≤ 2 n . Using adjunction we then get Also, δ R (K n ) contains a copy of K ( n ⌊n/2⌋ ) induced by the sets (U, U ) such that |U | = ⌊n/2⌋. Hence Therefore we have χ(δ(G)) ≃ log(χ(G)) and similarly χ(δ n (G)) ≃ log n (χ(G)) (where the exponents represent composition). These are essentially the only known examples of right adjoints with non-trivial predictable effects on the chromatic number of general graphs. Chromatic numbers of graphs obtained by applying arc-graph-like functors to transitive tournaments have recently been studied in [2,1]. However, the following is not known.  (i) If ψ is bounded, then its least upper bound is 3.
(ii) If φ is bounded, then its least upper bound is at most 9.
The proof method is simply a finer analysis of the chromatic properties of δ; in particular this accounts for the value 9 = 3 2 in the second statement. Thus the case of undirected graphs is never dealt with directly. Proof methods have always used directed graphs.
Problem 3.7. Is there a functor R : G → G that allows to provide tighter upper bounds on φ, if the case that it is bounded?
A later result gives a further link between φ and ψ:

Theorem 3.8 ([24]). φ is bounded if and only if ψ is bounded.
Thus the weak Hedetniemi Conjecture 1.2 is equivalent for graphs and for digraphs. It would be interesting to connect the case of digraphs more directly to Hedetniemi's conjecture. Problem 3.9. Is there a way to prove that if ψ grows too slowly, then Hedetniemi's conjecture is false? (Note that if ψ ′ is sub-quadratic, then Hedetniemi's conjecture is false.) 3.4. Multiplicative complete graphs. The last tentative application of δ to Hedetniemi's conjecture is an appealing argument that seems to have been first noticed by Roman Bašic. Proposition 3.10. If K n is multiplicative, then K ( n ⌊n/2⌋ ) is multiplicative. Proof. The proof uses a strengthening of Lemma 3.3 (ii) in the case of undirected graphs: Sperner's theorem implies that if G is an undirected graph, then χ(δ(G)) is the largest integer n such that χ(G) ≤ n ⌊n/2⌋ . In other words, for an undirected graph G we have δ(G) → K n ⇔ G → K ( n ⌊n/2⌋ ) . Now suppose that G × H → K ( n ⌊n/2⌋ ) . Then δ(G × H) → K n , that is, δ(G) × δ(H) → K n . If K n is multiplicative, this implies δ(G) → K n or δ(H) → K n , whence G → K ( n ⌊n/2⌋ ) or H → K ( n ⌊n/2⌋ ) . Thus if K n is multiplicative, then K ( n ⌊n/2⌋ ) is multiplicative. The attentive reader may have noticed that the above argument switches between G and D. It correctly uses the multiplicativity of K n in D to prove the multiplicativity of K ( n ⌊n/2⌋ ) in G. However, multiplicativity is category sensitive. The precise formulation of Proposition 3.10 is the following.
Proposition 3.11 (The precise formulation of Proposition 3.10). If K n is multiplicative in D, then K ( n ⌊n/2⌋ ) is multiplicative in G. Thus the result turns out to be trivial, since the only complete graphs that are multiplicative in D are K 0 , K 1 and K 2 . Attempts have been made to fix the argument, or to find similar arguments, using other functors. (The symmetrisation of δ does not seem to be the correct functor, either.) In Section 4, we examine the results obtained using the functors related to the Pultr template T (3) of Section 2, and its generalizations.
3.5. Excursion: odd girth of shift graphs. Let K n denote the transitive tournament on n vertices, that is, The shift graph S(n, k) is the graph δ k−1 ( K n ). The shift graphs (or their symmetrisations) are folklore "topology free" examples of graphs with large chromatic number and no short odd cycles. We show how these properties connect to adjoint functors. Proposition 3.12. S(n, k) has no odd cycle with fewer than 2k + 1 vertices, and for log k−1 2 (n) > m we have χ(S(n, k)) ≥ m (where the exponent represents composition).
Proof. We have χ( K n ) = n, hence the iterated logarithmic lower bounds on χ(S(n, k)) follow from Proposition 3.3, that is, from the right adjoint of δ. We now show that the absence of short odd cycles follows from the left adjoint of δ. Suppose that some orientation C of an odd cycle admits a homomorphism to S(n, k) = δ(S(n, k − 1)), for some k > 1. Then δ L (C) admits a homomorphism to S(n, k − 1), where δ L is the left adjoint of δ. By construction, the number of arcs in δ L (C) is equal to the number of vertices in C. Moreover, since C admits a homomorphism to S(n, k) which does not contain an oriented cycle, C must contain a source and a sink. The copies of P 1 corresponding to sources and sinks of C are pendant; therefore any odd cycle of δ L (C) has fewer arcs than C. Since we have C → δ(K 2 ) ≃ K 2 , we get δ L (C) → K 2 whence δ L (C) indeed has an odd cycle. The smallest odd cycle in K n has three vertices; iteratively, this implies that the smallest odd cycle in S(n, k) has at least 2k + 1 vertices.

Path functors and circular graphs
The natural generalization of Proposition 3.11 is the following. : G → G of Section 2 admits the right adjoint Ω T (3) : G → G. Therefore, using the fact that K 3 is multiplicative in G, we first get that Ω T (3) (K 3 ) is also multiplicative in G. Then, recursively, the graphs Ω n T (3) (K 3 ) are all multiplicative in G. It turns out that Ω n T (3) (K 3 ) is homomorphically equivalent to the odd cycle C 3 n+1 with 3 n+1 vertices. So this sequence of derivations yields particular cases of the following result of Häggkvist, Hell, Miller and Neumann Lara. The proof given in [11] is an adaptation of the proof of Theorem 3.1 given in [5]. However it is possible to derive the multiplicativity of all odd cycles from that of K 3 using the following adjoint functors that generalize Γ T (3) and Ω T (3) : Let T (2k + 1) = (K 1 , P 2k+1 , ǫ 1 , ǫ 2 ), where P 2k+1 is the path with 2k + 1 edges and ǫ 1 , ǫ 2 map K 1 to the endpoints of P 2k+1 . The right adjoint of Γ T (2k+1) is Ω T (2k+1) is characterized in [12]. It is defined as follows.
• The vertices of Ω T (2k+1) (G) are the (k + 1)-tuples (A 0 , . . . , A k ) such that each A i ⊆ V (G); A 0 is a singleton; every vertex of A i−1 is adjacent to every vertex of A i for i = 1, . . . , k; and . . , k, and every vertex of A k is adjacent to every vertex of B k . To prove Theorem 4.2 with these functors, we use the following. Theorem 4.2 can then be proved as follows. We note that Ω T (2k+1) (K 3 ) and Ω T (3) (C 2k+1 ) are both isomorphic to the odd cycle C 3(2k+1) . Therefore by Proposition 4.1, the multiplicativity of C 3(2k+1) follows from the multiplicativity of K 3 , and the multiplicativity of C 2k+1 follows from the multiplicativity of C 3(2k+1) .

Circular complete graphs and the circular chromatic number.
We now extend Theorem 4.2 to some "circular complete graphs" defined as follows. Let Z s = {0, 1, . . . , s − 1} denote the cyclic group with s elements. For integers r, s such that 1 ≤ r ≤ s/2, the circular complete graph K s/r is the graph with vertex set Z s and with edges {x, y} such that y − x ∈ {r, r + 1, . . . , s − r}. In particular, K s/1 is the complete graph K s , and K (2r+1)/r is the odd cycle C 2r+1 .
The conventional notation K s/r is slightly ambiguous, since it confuses two integer parameters s and r with a single rational parameter s/r. We will write s/r for the parameter pair and s r for the corresponding fraction. The notation is motivated by the following result, which alleviates some of its ambiguity.

Lemma 4.4 ([3]).
( The circular chromatic number χ c (G) of a graph G is defined as the infimum of the values s r such that G → K s/r . By Lemma 4.4, the infimum is attained. We have χ c (G) ≤ χ(G) < χ c (G) + 1, hence the circular chromatic number is a refinement of the chromatic number.
As in the case of the chromatic number, the inequality χ c (G × H) ≤ min{χ c (G), χ c (H)} follows immediately from definitions. Zhu [27] conjectured that equality always holds. This is a strengthening of Hedetniemi's conjecture. It is equivalent to the statement that the circular complete graphs are multiplicative. In this direction, the following is known. The proof uses the functors Γ T (3) and Ω T (3) as follows.
(i) When s r < 3, we have Γ T (3) (K s/r ) ↔ K s/(3r−s) . (ii) When s r < 12 5 , we have Ω T (3) (Γ T (3) (K s/r )) ↔ K s/r . (iii) Therefore by Proposition 4.1, when s r < 12 5 , K s/r is multiplicative if and only if K s/(3r−s) is multiplicative. (iv) Starting with the odd cycles C 2n+1 = K (2n+1)/n which are multiplicative by Proposition 4.2, we can use item (iii) above to infer the multiplicativity of many circular complete graphs K s/r = Γ m T (3) C 2n+1 . The set of rationals s r for which K s/r is proved to be multiplicative in this fashion is a dense subset of the interval (2,4).
(v) The remaining graphs K s/r , 2 < s r < 4 are proved to be multiplicative by a density argument: Suppose that It turns out that the multiplicativity of K 3 implies that of all the graphs K s/r : 2 < s r < 4. Up to homomorphic equivalence, these are the only graphs known to be multiplicative (other than K 1 and K 2 ). The proof method exposed here does not extend to other circular complete graphs. Property (i) breaks down at s r = 3, since Γ T (3) (K s/r ) contains loops whenever s r ≥ 3. Property (ii) breaks down at 12 5 , we have Γ T (3) (K 12/5 ) ↔ K 4 , and Ω T (3) (K 4 ) is 4-chromatic by the result of [10] discussed in Section 2.

Compositions and chains of functors.
For odd integers m, n, we consider the functor L m n = Γ T (m) • Λ T (n) and its right adjoint R n m = Γ T (n) • Ω T (m) . In particular, the circular complete graphs K s/r with s an odd multiple of 3 are the graphs R m n (K 3 ) (≃ L m n (K 3 )) such that m < 3n. More generally, for any odd cycle C k with k ≥ m n , we have L m n (C 2k+1 ) ↔ R m n (C 2k+1 ) ↔ K s/r , where s = nk and r = nk−m 2 . As noted in [13], the circular chromatic number of G can therefore be expressed in terms of the infimum of the values m n such that G → R m n (K 3 ). Since R m n is the thin right adjoint of L n m , we can look instead at the values m n such that L n m (G) admits a homomorphism to K 3 . In this characterization, the "base graph" K 3 could be replaced by any of the graphs K s/r with 2 < s r < 4, with an appropriate change of parameters. However, replacing K 3 by K 4 or other graphs yields new graph invariants, defined by homomorphisms into chains of graphs which mimic the circular complete graphs.
At the functorial level, we can order thin functors in an obvious way, Proof. The implications m n ≤ m ′ n ′ ⇒ L m n ≤ L m ′ n ′ and R m n ≤ R m ′ n ′ are proved in [13]. The converse implications are witnessed by suitably chosen circular complete graphs.
Up to equivalence, we can assimilate L m n and R m n to functors L m/n and R m/n corresponding to the rational parameter m n . These are two chains of functors isomorphic to the positive rationals with odd numerator and denominator. This order is in fact isomorphic to the rationals, but our specific labelling leaves holes. In particular, the adjoint functors L 1/2 and R 2/1 are not defined. We next see that this specific pair of holes can be filled by Pultr functors. 4.4. A limit functor. Consider the Pultr template T (2) = (P 1 , P 1 P 2 , ǫ 1 , ǫ 2 ) where P 1 and P 2 are path with vertex-sets {0, 1} and {0, 1, 2} respectively, is the Cartesian product of section 2, and (ǫ 1 (0), ǫ 1 (1), ǫ 2 (0), ǫ 2 (1)) = ((0, 0), (1, 0), (1, 2), (0, 2)). Having P 1 rather than P 1 as first coordinate of the template makes a difference only for graphs with no edges. In all other cases, we could replace P 1 by P 1 and get an equivalent functor. Proposition 4.8. For all G ∈ G, we have L m n (G) → Λ T (2) (G) when m n < 1 2 , and Λ T (2) (G) → L m n (G) when m n > 1 2 ; as well as R n m (G) → Γ T (2) (G) when n m < 2, and Γ T (2) (G) → R n m (G) when n m > 2. We omit the proof, which is similar to many proofs involving the functors associated with the templates T (2k + 1). The main point is that Λ T (2) and Γ T (2) fulfill the role of L 1/2 and R 2/1 . It is not clear which of the other rational holes in the chains {L m/n } m,n odd and {R n/m } n,m odd can be filled, and how. It can be shown that the irrational holes cannot be filled by finite constructions. Thus, a workable theory of convergence for sequences of thin functors remains to be developed.
One interesting aspect of the template T (2) is that Γ T (2) has a partial right adjoint. Let Ω T (2) (G) be the graph defined as follows.
We again omit the standard proof for the sake of briefness. We note that the result allows the derivation of the multiplicativity of Ω T (2) (K 3 ) = K 12/5 directly from that of K 3 , instead of through density arguments as in the proof of Proposition 4.5. The range of graphs on which Ω T (2) acts as a right adjoint of Γ T (2) is not precisely known, but it extends to the odd cycles and allows to reprove the multiplicativity of some circular complete graphs.
In short, adjoint functors allow the extension of the proof of multiplicativity of K 3 to the circular complete graphs K s/r , 2 < s r < 4, sometimes in many ways, using compositions, limits and partial right adjoints. However, the multiplicativity of K 4 , the "next case of Hedetniemi's conjecture" remains open. We now present a difficulty in using the adjoint functors method to derive the multiplicativity of other complete graphs from that of K 3 .

4.5.
Obstacle: A stronger form of multiplicativity. El-Zahar and Sauer actually proved a stronger form of Theorem 3.1: They conjectured that a similar strong form of multiplicativity would hold for all complete graphs. However this conjecture was refuted in [25]. The counterexamples are of interest to us. Let G m,n be the graph obtained by identifying one end of P n to a vertex of K 7/2 and the other end to a vertex of K m . For m > 4, G m,n × G m,n ↔ G m,n → K 4 . Let L m,n be the subgraph of G m,n × G m,n induced by V (G m,n × K 7/2 ) ∪ V (K 7/2 × G m,n ). Now suppose that for some T , Γ T admits a right adjoint Ω T such that Ω T (K 3 ) = K 4 . Then for n ≥ 3, m ≥ 5, we have G m,n → K 4 whence Γ T (G m,n ) → K 3 . However from L m,n → K 4 = Ω T (K 3 ), we get Γ T (L m,n ) → K 3 . However Γ T (L m,n ) contains a copy of Γ T (G m,n )×Γ T (K 7/2 )∪Γ T (K 7/2 )× Γ T (G m,n ). Hence by Theorem 4.10, since Γ T (G m,n ) → K 3 , we must have that Γ T (K 7/2 ) is bipartite. It can be shown that this is incompatible with the fact that Γ T admits a right adjoint.
In short, the right adjoints we know seem to transfer a property that is stronger than multiplicativity, and this seems to be an obstacle in deriving the multiplicativity of other complete graphs from that of K 3 .
We note that Wrochna [26] has recently proved that all square-free graphs are multiplicative. The functors detailed here can therefore be used to derive from this the multiplicativity of a larger class of graphs. However, Wrochna states that his method also yields the stronger form of multiplicativity for square-free graphs. Hence the multiplicativity of larger complete graphs still seems out of reach.

Concluding comments
We have shown how some important facts concerning the weak Hedetniemi conjecture and the characterization of multiplicative graphs are connected to adjoint functors in the thin category of graphs. In addition, the multi-factor version of the weak Hedetniemi conjecture can also be tackled through adjoint functors, as shown in [6]. The characterization of central Pultr functors which admit right adjoints is initiated in [8]. However, as we have seen, further developments may depend on functors that escape the mould of Pultr functors. In short, the tools devised so far in this line of study are categorial in nature, and seem to have reached their limit in their present form. Perhaps the contribution of category theory to Hedetniemi's conjecture would be to see the way to sharpen these tools.