A better lower bound on the average degree of online k-list-critical graphs

We improve the best known bounds on the average degree of online k-list-critical graphs for k > 6. Specifically, for k > 7 we show that every non-complete online k-list-critical graph has average degree at least k−1+ (k−3) 2(2k−3) k4−2k3−11k2+28k−14 and every non-complete online 6-list-critical graph has average degree at least 5 + 93 766 . The same bounds hold for offline k-list-critical graphs.


Introduction
Imagine a classroom with n children and one teacher.Each child wishes to go to lunch with a group of her friends, but the teacher is dreadfully strict and will only allow certain groups of students to leave for lunch at a time.The teacher starts by giving each child k lunch tokens.Then she selects a group of students S and the children must decide on a group of friends F from S to go to lunch.The group of friends F leave for lunch and the remaining students in S each give the teacher one of their lunch tokens.The teacher keeps picking groups of students and allowing the children to decide on a subgroup of friends to go to lunch like this until either all of the students have gone to lunch with a lunch token in hand; or some student has no lunch tokens left.The teacher wants to be strict, but fair, so she desires to find the smallest number k of tokens to give each student so that there is always a way for the children to play that ends up with all the children going to lunch with a token and a group of friends.To her chagrin, she cannot find an easy way to determine the smallest k for a given classroom of children; it seems she has constructed a nearly impossible problem for herself in her perverse authoritarian zeal.We do not really want to help her solve her problem so she can inflict this game upon the children, but it is an interesting problem nonetheless.So, let us think about the problem and just not tell her of any progress we make.First off, we can abstract things a little and consider a graph G with vertex set consisting of all the children such that vertices x and y (representing children) are adjacent just in case x and y are not friends.Then a group of friends is just an independent set in G.We can think of giving a number of the electronic journal of combinatorics 25(1) (2018), #P1.51 lunch tokens to each student as a function from V (G) to N where f (v) is the number of tokens child v gets.For such a function f , say that G is online f -list-colorable just in case V (G) is empty, or for every S ⊆ V (G) there is an independent set F ⊆ S such that G − F is online f -list-colorable where f is the function from V (G) \ F to N given by f (v) = f (v) − 1 for v ∈ S \ F and f (v) = f (v) for v ∈ V (G) \ S. For k ∈ N, we say that G is online k-list-colorable just in case G is online f -list-colorable where f (v) = k for all v ∈ V (G).The smallest k for which G is online k-list-colorable is the online list-chromatic number of G.The teacher's problem then becomes the following.
Problem.Given a graph G that represents a classroom of children and their friend relationships, how can we best find the online list-chromatic number of G?
This problem is hard (supposing other problems we think are hard are in fact hard [6]), we will not solve it here.Surely, more friendships amongst the children should make it easier for the teacher to use fewer tokens per child.This intuition is most succinctly captured by the average degree d(G) of the vertices in the non-friend graph G defined above.Specifically, a low average degree should make it easier for the teacher to use fewer tokens per child.We improve the best known quantitative bounds for this problem.A graph G is online k-list-critical if G is not online (k − 1)-list-colorable, but every proper subgraph of G is.For further graph theory definitions and notation, we refer the reader to Diestel's textbook [3].
Main Theorem.For k 7, every non-complete online k-list-critical graph has average degree at least Every non-complete online 6-list-critical graph has average degree at least 5 + 93 766 .The theorem also holds, with a nearly identical proof, for offline list coloring.The proof is similar to the 4-list-critical case in [12], but now we incorporate reducibility lemmas from Kierstead and Rabern [7].Basically, we show that the average degree of the subgraph induced on vertices of degree k − 1 is small, which implies that the number of edges incident to the vertices of degree at least k must be large, and hence the number of vertices of degree at least k must be large; that is, the graph must have high average degree.A tight bound on the average degree of the subgraph induced on vertices of degree k − 1 in an online k-list-critical graph was proved by Gallai [5].The connected graphs in which each block is a complete graph or an odd cycle are called Gallai trees.Gallai [5] proved that in a k-critical graph, the vertices of degree k − 1 induce a disjoint union of Gallai trees.The same is true for online k-list-critical graphs [1,4,13].Since Gallai's bound is tight, it may appear that there is no hope of improvement using the above method.While it is true that the upper bound on average degree of Gallai trees cannot be improved in general, it can be improved in the absence of certain bad properties.Let G be an online k-list-critical graph and let L be the subgraph of G induced on vertices of degree k − 1.If the presence of bad properties in L could be shown to lead to reducible configurations in G, we would have a pathway to improvement.Kostochka and Stiebitz [9] made the first progress along these lines.Further improvements in [7], [2] and [12] follow the same general outline.As in [2] and [12], it is convenient to have a measure of how bad L is.So, if b is a function measuring badness, this could be realized as an upper bound of the form: Of course, we can measure badness along multiple axes.In our proof we use two badness measures β(L) and q(L), so the upper bound looks like: High β(L) badness leads to reducible configurations by kernel-perfect orientations and high q(L) badness leads to reducible configurations by Alon-Tarsi orientations.That means the same proof shows that Main Theorem holds for offline k-list-critical graphs as well (in fact, for the larger class of OC-irreducible graphs with δ(G) = k − 1 defined in section 5).Let c * k (L) be the number of components of L containing a copy of K k−1 .Let q k (L) be the number of non-cut vertices in L that appear in copies of K k−1 .Let β k (L) be the independence number of the subgraph of L induced on the vertices of degree k − 1.When k is defined in context, we just write c * (L), q(L) and β(L).We need upper bounds on our badness parameters q(L) and β(L).More general versions of the following two lemmas are proved in Section 2 and Section 3, respectively.The first lemma allows us a measure of control over K k−1 blocks in L. The second lemma allows us some control over the number of vertices in L with no neighbors in H. , then In Section 4, we use these two lemmas and a little careful counting to prove general lower bounds on the average degree of online k-list-critical graphs in terms of parameters p, h, z, f that are functions of k and satisfy bounds of the form for every Gallai tree T with ∆(T ) k − 1 (and a few further caveats).So, basically Section 4 is about turning upper bounds on the average degree of Gallai trees into lower bounds on the average degree of online k-list-critical graphs.
Finally, Section 5 proves upper bounds on the average degree of Gallai trees in terms of the parameters p, h, z, f .

Bounding q(L)
This section is devoted to extracting the reusable Lemma 2.1 from the proof of Kierstead and Rabern [7].All of the hard work was already done in [7].
Observation.The hypotheses of Lemma 2.1 are satisfied by non-complete k-critical, klist-critical, online k-list-critical and k-AT-critical graphs.
The proof of Lemma 2.1 requires the following four lemmas from [7].
Lemma 2.2.Let G be a graph and f : the electronic journal of combinatorics 25(1) (2018), #P1.51 Proof.Suppose not and choose a counterexample G minimizing |G|.Then |G| 3 and we have Let T k be the Gallai trees with maximum degree at most k − 1, excepting K k .For a graph G, let W k (G) be the set of vertices of G that are contained in some K k−1 in G.
Lemma 2.3.Let k 5 and let G be a graph with x ∈ V (G) such that: . ., H t , and all are in T k ; and We also have the following version with asymmetric degree condition on B. The point here is that this works for k 5.The consequence is that we trade a bit in our bound for the proof to go through with k ∈ {5, 6}.
Proof of Lemma 2.1.Let H be the subgraph of G induced on k + -vertices and let D be the components of L containing a copy of K k−1 .Put W := W k (L) and L := V (L) \ W . Define an auxiliary bipartite graph F with parts A and B where: Define . Now the second inequality in the lemma follows as before.

Bounding β(L)
This section is devoted to extracting the reusable Lemma 3.1 from the proof of Rabern [12].
If G is not OC-reducible to any nonempty induced subgraph, then it is OC-irreducible.
Observation.The hypotheses of Lemma 3.1 are satisfied by k-critical, k-list-critical and online k-list-critical graphs.
The proof of Lemma 3.1 requires the following lemma from Kierstead and Rabern [8] that generalizes a kernel technique of Kostochka and Yancey [10].
Definition.The maximum independent cover number of a graph G is the maximum mic(G) of I, V (G) \ I over all independent sets I of G.
Summing the bound in Claim 1 over all components of H and plugging into (1) gives Applying Lemma 3.2 using (2) and solving for β(L) proves the claim.

General lower bounds on average degree
This is the counting portion of the proof, which is simpler and more general than the counting in [7] and [2].
Definition 3. A quadruple (p, h, z, f ) of functions from N to R is r-Gallai if for every k r and Gallai tree T = K k with ∆(T ) k − 1, the following hold: , where L is the subgraph of G induced on (k − 1)-vertices.
Proof.Let H − the subgraph of G induced on k-vertices, H the subgraph of G induced on k + -vertices, H + the subgraph of G induced on (k + 1) + -vertices and D the components of L containing K k−1 .Plainly, the following bounds hold.
Now factoring out k − 1 gives the desired bound.
A nearly identical argument, using the other inequality in Lemma 1.1, proves a bound that holds for k 5.
, where L is the subgraph of G induced on (k − 1)-vertices.
When k = 4, we cannot apply Lemma 1.1, but using h(k) = 0 and running through the same argument proves the following bound for k 4.
When z(k) < 2, using Lemma 1.2 worsens the lower bound, so we may as well use z(k) = 0; that is, drop the β(L) term entirely.Doing so in the above argument shows that Theorems 4.1, 4.2, 4.3 hold for z(k) = 0 if we replace k + 1 in the denominator with k + 2. This gives the bounds proved by discharging in Cranston and Rabern [2].

Gallai quadruples
All known proofs of lower bounds for average degree of list-critical graphs are essentially a counting argument combined with the fact that some quadruple is Gallai.
We give a a list of inequalities that provide a sufficient condition for (p, h, z, f ) to be 5-Gallai.These inequalities take a form quite similar to the inequalities in Cranston and Rabern [2], but now they involve z(k) as well.The sufficiency proof is a small modification of the proof in [2].To use a Gallai quadruple in Lemma 4.1, we want 2h(k) + f (k) 0 to get rid of the term involving c * (L).Similarly, for Lemma 4.2, we want h(k) + f (k) 0. Finding the p, h, z, f that give the largest average degree subject to these constraints is a fractional linear program that can be converted to a linear program and solved for each k.This is useful for verification of bounds, but we want a formula in terms of k.For k 7, we use the following quadruple.
Lemma 5.5.The tuple 3k−7 For k = 6, we use the following quadruple.For k = 5, the quadruple in Lemma 5.4 is the optimal choice of p, h, z, f .
Lemma 5.6.The tuple 3k−5 Now on to the sufficiency proof.We prove two technical lemmas, the proof of Lemma 5.5 and Lemma 5.6 are straightforward computations from these.For an endblock B of a Gallai tree T , let x B be the cutvertex contained in B.
Lemma 5.7.Let z : N → R such that z(k) = 0 or z(k) 2 for all k ∈ N.For all k 5 and Gallai trees T with ∆(T ) k − 1 and K k−1 ⊆ T , we have Proof.Suppose the lemma is false and choose a counterexample T minimizing |T |.Claim 1. T has at least two blocks.
If T has only one block, then 2 T Suppose T has an endblock B that is not K k−2 .Then removing V (B) \ {x B } from T to get T and applying minimality of Claim 4. T does not exist.
By the previous claims, we know that every endblock T is a K k−2 that shares a vertex with an odd cycle.Pick and endblock B that is the end of a longest path in the block-tree of T .Let C be the odd cycle sharing x B with B. Since B is the end of a longest path in the block-tree, there is a neighbor y of x B on C such that d T (y) = 2 or y is contained in another endblock A (which must be a K k−2 ).First, suppose d T (y) = 2. Removing V (B) ∪ {y} from T to get T and applying minimality of |T | gives the contradiction (since the electronic journal of combinatorics 25(1) (2018), #P1.51 Hence y is contained in another K k−2 endblock A. Removing V (B) ∪ V (A) from T to get T * and applying minimality of |T | gives the contradiction (since β(T * ) < β(T )) whenever p, f , h and z satisfy all of the following conditions: ( (2) p(k) Proof.Suppose the lemma is false and choose a counterexample T minimizing |T |.Claim 1. T has at least two blocks.
Suppose T has an endblock B that is not Suppose B = K t for 4 t k − 3. Then we have the contradiction Finally, suppose B is an odd cycle of length .Then, we have .
Claim 5. T does not exist.
By Claims 2 and 3, all but at most one endblock of T is K k−2 with a cutvertex that is also in an odd cycle.Pick and endblock B that is the end of a longest path in the blocktree of T .Let C be the odd cycle sharing x B with B. Since B is the end of a longest path in the block-tree, there is a neighbor y of x B on C such that d T (y) = 2 or y is contained in another endblock A (which must be a K k−2 ).First, suppose d T (y) = 2. Removing V (B) ∪ {y} from T to get T and applying minimality of |T | gives (since q(T ) = q(T ) and β(T ) < β(T )) That is all we need to prove our lower bounds on average degree.If a good upper bound on c * (L) is known, it may be better to allow 2h(k) + f (k) > 0. In that case, one could use the following.

Lemma 1 . 1 .
Let G be a non-complete online k-list-critical graph where k 5. Let L be the subgraph of G induced on (k − 1)-vertices, H − the subgraph of G induced on k-vertices and H + the subgraph of G induced on (k + 1) + -vertices.Then q(L) c * (L) + 4 H − + H + , L , and if k 7, then q(L) 2c * (L) + 3 H − + H + , L .

Lemma 1 . 2 .
Let G be an online k-list-critical graph.Let L be the subgraph of G induced on (k − 1)-vertices and H the subgraph of G induced on k + -vertices.If 2 λ 6(k−1) k For a graph G, {X, Y } a partition of V (G) and k 4, let B k (X, Y ) be the bipartite graph with one part Y and the other part the components of G[X].Put an edge between y ∈ Y and a component T of G[X] iff N (y) ∩ W k (T ) = ∅.The next lemma tells us that we have a reducible configuration if this bipartite graph has minimum degree at least three.Lemma 2.4.Let k 7 and let G be a graph with Y ⊆ V (G) such that: 1.K k ⊆ G; and 2. the components of G − Y are in T k ; and 3

Lemma 2 . 5 .
Let k 5 and let G be a graph with Y ⊆ V (G) such that: 1.K k ⊆ G; and 2. the components of G − Y are in T k ; and the electronic journal of combinatorics 25(1) (2018), #P1.51

3 .
and A is the disjoint union of the following sets A 1 , A 2 and A 3 ,2.A 1 = D and each T ∈ D is adjacent to all y ∈ B where N (y) ∩ W k (T ) = ∅, For each v ∈ L , let A 2 (v) be a set of |N (v) ∩ B| vertices connected to N (v) ∩ B by a matching in F .Let A 2 be the disjoint union of the A 2 (v) for v ∈ L , 4.For each y ∈ B, let A 3 (y) be a set of d H (y) vertices which are all joined to y in F .Let A 3 be the disjoint union of the A 3 (y) for y ∈ B.

Lemma 3 . 1 .
Let G be an OC-irreducible graph with δ(G) = k − 1.Let L be the subgraph of G induced on (k − 1)-vertices and H the subgraph of G induced on k + -vertices.If 2 λ 6(k−1) k , then

P 3 or K 3
which all satisfy the desired bound.the electronic journal of combinatorics 25(1) (2018), #P1.51 Proof of Lemma 3.1.Fix λ with 2 λ 6(k−1) k .Let M be the maximum of I, V (G) \ I over all independent sets I of G with I ⊆ H. Since the vertices in L with k − 1 neighbors in L have no neighbors in H, mic(G) M + (k − 1)β(L).
This is a contradiction unless k = 5 and B = K 3 , but then B = K k−2 , a contradiction.Claim 3. If B is an endblock of T , then d T (x B ) = k − 1. Suppose B is an endblock of T with d T (x B ) < k − 1.Then B = K k−2 by Claim 2 and hence d T (x B ) = k − 2. Removing V (B) from T to get T * and applying minimality of |T | gives the contradiction

Table 1 :
Lower bounds on average degree d