On the Price of Independence for Vertex Cover, Feedback Vertex Set and Odd Cycle Transversal

Let $vc(G)$, $fvs(G)$ and $oct(G)$, respectively, denote the size of a minimum vertex cover, minimum feedback vertex set and minimum odd cycle transversal in a graph $G$. One can ask, when looking for these sets in a graph, how much bigger might they be if we require that they are independent; that is, what is the price of independence? If $G$ has a vertex cover, feedback vertex set or odd cycle transversal that is an independent set, then we let $ivc(G)$, $ifvs(G)$ or $ioct(G)$, respectively, denote the minimum size of such a set. Similar to a recent study on the price of connectivity (Hartinger et al. EuJC 2016), we investigate for which graphs $H$ the values of $ivc(G)$, $ifvs(G)$ and $ioct(G)$ are bounded in terms of $vc(G)$, $fvs(G)$ and $oct(G)$, respectively, when the graph $G$ belongs to the class of $H$-free graphs. We find complete classifications for vertex cover and feedback vertex set and an almost complete classification for odd cycle transversal (subject to three non-equivalent open cases). We also investigate for which graphs $H$ the values of $ivc(G)$, $ifvs(G)$ and $ioct(G)$ are equal to $vc(G)$, $fvs(G)$ and $oct(G)$, respectively, when the graph $G$ belongs to the class of $H$-free graphs. We find a complete classification for vertex cover and almost complete classifications for feedback vertex set (subject to one open case) and odd cycle transversal (subject to three open cases).


Introduction
We define a number of transversals of the vertex set of a graph G. A set S ⊆ V (G) is a vertex cover if for every edge uv ∈ E(G), at least one of u and v is in S, or, equivalently, if the graph G − S contains no edges. A set S ⊆ V (G) is a feedback vertex set if for every cycle in G, at least one vertex of the cycle is in S, or, equivalently, if the graph G − S is a forest. A set S ⊆ V (G) is an odd cycle transversal if for every cycle in G containing an odd number of vertices, at least one vertex of the cycle is in S, or, equivalently, if the graph G − S is bipartite. For each of these transversals, one usually wishes to investigate how small they can be and there is a vast research literature on this topic. One can add an additional constraint: require the transversal to be an independent set, that is, a set of vertices that are pairwise non-adjacent. It might be possible that no such transversal exists under this constraint. For example, a graph G has an independent vertex cover if and only if G is bipartite. We are interested in the following research question:

How is the minimum size of a transversal in a graph affected by adding the requirement that the transversal is independent?
Of course, this question can be interpreted in many ways; for example, one might ask about the computational complexity of finding the transversals. In this paper, we focus on the following: for the three transversals introduced above, is the size of a smallest possible independent transversal (assuming one exists) bounded in terms of the minimum size of a transversal? That is, one might say, what is the price of independence?
To the best of our knowledge, the term price of independence was first used by Camby [4] in a recent unpublished manuscript. She considered dominating sets of graphs (sets of vertices such that every vertex outside the set has a neighbour in the set). As she acknowledged, though first to coin the term, she was building on past work. In fact, Camby and her co-author Plein had given a forbidden induced subgraph characterization of those graphs G for which, for every induced subgraph of G, there are minimum size dominating sets that are already independent [6], and there are a number of further papers on the topic of the price of independence for dominating sets (see the discussion in [4]).
We observe that this incipient work on the price of independence is a natural companion to recent work on the price of connectivity, investigating the relationship between minimum size transversals and minimum size connected transversals (which, in contrast to independent transversals, will always exist for the transversals we consider, assuming the input graph is connected). This work began with the work of Cardinal and Levy in their 2010 paper [8] and has since been taken in several directions; see, for example, [1,5,7,9,11,12].
In this paper, as we broaden the study of the price of independence by investigating the three further transversals defined above, we will concentrate on classes of graphs defined by a single forbidden induced subgraph H, just as was done for the price of connectivity [1,12]. That is, for a graph H, we ask what, for a given type of transversal, is the price of independence in the class of H-free graphs? The ultimate aim in each case is to find a dichotomy that allows us to say, given H, whether or not the size of a minimum size independent transversal can be bounded in terms of the size of a minimum transversal. We briefly give some necessary definitions and notation before presenting our results.
A colouring of a graph G is an assignment of positive integers (called colours) to the vertices of G such that if two vertices are adjacent, then they are assigned different colours. A graph is k-colourable if there is a colouring that only uses colours from the set {1, . . . , k}. Equivalently, a graph is k-colourable if we can partition its vertex set into k (possibly empty) independent sets (called colour classes or partition classes). For s, t ≥ 0, let K s,t denote the complete bipartite graph with partition classes of size s and t, respectively (note that K s,t is edgeless if s = 0 or t = 0). For r ≥ 0, the graph K 1,r is also called the (r + 1)-vertex star; if r ≥ 2 we say that the vertex in the partition class of size 1 is the central vertex of this star. For n ≥ 1, let P n and K n denote the path and complete graph on n vertices, respectively. For n ≥ 3, let C n denote the cycle on n vertices. For r ≥ 1, let K + 1,r denote the graph obtained from K 1,r by subdividing one edge. The disjoint union G + H of two vertex-disjoint graphs G and H is the graph with vertex set V (G) ∪ V (H) and edge set E(G) ∪ E(H). We denote the disjoint union of r copies of a graph G by rG.
The Price of Independence for Vertex Cover. As mentioned above, a graph has an independent vertex cover if and only if it is bipartite. For a bipartite graph G, let vc(G) denote the size of a minimum vertex cover, and let ivc(G) denote the size of a minimum independent vertex cover. Given a class X of bipartite graphs, we say that X is ivc-bounded if there is a function f : Z ≥0 → Z ≥0 such that ivc(G) ≤ f (vc(G)) for every G ∈ X and we say that X is ivc-unbounded if no such function exists, that is, if there is a k such that for every s ≥ 0 there is a graph G in X with vc(G) ≤ k, but ivc(G) ≥ s.
In our first main result, proven in Section 2, we determine for every graph H, whether or not the class of H-free bipartite graphs is ivc-bounded.

Theorem 1. Let H be a graph. The class of H-free bipartite graphs is ivc-bounded if and
only if H is an induced subgraph of K 1,r + rP 1 or K + 1,r for some r ≥ 1.
The Price of Independence for Feedback Vertex Set. A graph has an independent feedback vertex set if and only if its vertex set can be partitioned into an independent set and a set of vertices that induces a forest; graphs that have such a partition are said to be near-bipartite. In fact, minimum size independent feedback vertex sets have been the subject of much research from a computational perspective: to find such a set is, in general, NP-hard, but there are fixed-parameter tractable algorithms and polynomial-time algorithms for certain graph classes; we refer to [2] for further details. For a near-bipartite graph G, let fvs(G) denote the size of a minimum feedback vertex set, and let ifvs(G) denote the size of a minimum independent feedback vertex set. Given a class X of near-bipartite graphs, we say that X is ifvs-bounded if there is a function f : Z ≥0 → Z ≥0 such that ifvs(G) ≤ f (fvs(G)) for every G ∈ X and ifvs-unbounded otherwise. In our second main result, proven in Section 3, we determine for every graph H, whether or not the class of H-free near-bipartite graphs is ifvs-bounded.

Theorem 2. Let H be a graph. The class of H-free near-bipartite graphs is ifvs-bounded
if and only if H is isomorphic to P 1 + P 2 , a star or an edgeless graph.
The Price of Independence for Odd Cycle Transversal. A graph has an independent odd cycle transversal S if and only if it has a 3-colouring, since, by definition, we are requesting that S is an independent set of G such that G − S has a 2-colouring. For a 3-colourable graph G, let oct(G) denote the size of a minimum odd cycle transversal, and let ioct(G) denote the size of a minimum independent odd cycle transversal. Given a class X of 3colourable graphs, we say that X is ioct-bounded if there is a function f : Z ≥0 → Z ≥0 such that ioct(G) ≤ f (oct(G)) for every G ∈ X and ioct-unbounded otherwise.
In our third main result, proven in Section 4, we address the question of whether or not, for a graph H, the class of H-free 3-colourable graphs is ifvs-bounded. Here, we do not have a complete dichotomy due to three missing cases, as we will discuss at the end of Section 4.

63:4
On the Price of Independence Theorem

Let H be a graph. The class of H-free 3-colourable graphs is ioct-bounded:
if H is an induced subgraph of P 4 or K 1,3 + sP 1 for some s ≥ 0 and only if H is an induced subgraph of K + 1,4 or K 1,4 + sP 1 for some s ≥ 0.

Vertex Cover
We start with a useful lemma.

Lemma 4.
Let r, s ≥ 1. If G is a (K 1,r + sP 1 )-free bipartite graph with bipartition (X, Y ) such that |X|, |Y | ≥ rs + r − 1, then either: every vertex of G has degree less than r or fewer than s vertices of X have more than s − 1 non-neighbours in Y and fewer than s vertices of Y have more than s − 1 non-neighbours in X.
Proof. Let G be a (K 1,r + sP 1 )-free bipartite graph with bipartition (X, Y ) such that |X|, |Y | ≥ rs + r − 1. No vertex in X can have both r neighbours and s non-neighbours in Y , otherwise G would contain an induced K 1,r + sP 1 . Therefore every vertex in X has degree either at most r − 1 or at least |Y | − (s − 1) ≥ rs + r − s. By symmetry, we may assume that there is a vertex x ∈ X of degree at least r. Suppose, for contradiction, that there is a set X ⊆ X of s vertices, each of which has more than s − 1 non-neighbours in Y . Then every vertex of X has degree at most r − 1.
is a K 1,r + sP 1 , a contradiction. It follows that fewer than s vertices in X have more than , there must be a vertex y ∈ Y that is complete to X , and therefore has deg(y) ≥ r. Repeating the above argument, it follows that fewer than s vertices of Y have more than s − 1 non-neighbours in X. This completes the proof.
Recall that a graph has an independent vertex cover if and only if it is bipartite.
Proof. Let G be a (K 1,r + sP 1 )-free bipartite graph. Fix a bipartition (X, Y ) of G. Let S be a minimum vertex cover of G, so |S| = vc(G). We may assume that vc(G) ≥ 2, otherwise ivc(G) = vc(G), in which case we are done. We may also assume that |X|, |Y | > vc(G)r+rs > rs + r − 1, otherwise X or Y is an independent vertex cover of the required size, and we are done. If every vertex of G has degree at most r − 1, then S = (S ∩ Y ) ∪ (N (S ∩ X)) is an independent vertex cover in G of size at most vc(G)(r − 1), and we are done. By Lemma 4, we may therefore assume that fewer than s vertices of X have more than s − 1 non-neighbours in Y . We will show that this leads to a contradiction. Since |X|, |Y | ≥ vc(G) + s, there must be a set S of vc(G) + 1 vertices in X that have at least vc(G) + 1 neighbours in Y . If a vertex x ∈ V (G) has degree at least vc(G) + 1, then |N (x)| > |S|, so x ∈ S. Therefore every vertex of S must be in S, contradicting the fact that |S | = vc(G) + 1 > vc(G) = |S|.
Proof. Clearly it is sufficient to prove the lemma for connected graphs G. Let G be a connected K + 1,r -free bipartite graph. Fix a bipartition (X, Y ) of G. Let S be a minimum vertex cover of G, so |S| = vc(G). We may assume that vc(G) ≥ 2, otherwise ivc(G) = vc(G) and we are done. We may also assume that |X|, |Y | > (vc(G)) 2 (r − 1), otherwise X or Y is an independent vertex cover of the required size.
If there are two vertices x, y ∈ X with dist(x, y) = 2 and deg(x) ≥ deg(y) + (r − 1), then x, y, a common neighbour of x and y, and . Note that if a vertex x ∈ V (G) has degree at least vc(G) + 1, then |N (x)| > |S| and so x ∈ S.
We are now ready to prove the main result of this section.

Theorem 1 (restated). Let H be a graph. The class of H-free bipartite graphs is ivc-bounded if and only if
H is an induced subgraph of K 1,r + rP 1 or K + 1,r for some r ≥ 1.

Proof.
If H is an induced subgraph of K 1,r + rP 1 or K + 1,r for some r, then Lemma 5 or 6, respectively, implies that the class of H-free bipartite graphs is ivc-bounded. Now let H be a graph and suppose that there is a function f : Z ≥0 → Z ≥0 such that ivc(G) ≤ f (vc(G)) for all H-free bipartite graphs G. We will show that H is an induced subgraph of K 1,r + rP 1 or K + 1,r for some r. For r ≥ 1, s ≥ 2, let D r s denote the graph formed from 2K 1,s and P 2r by identifying the two end-vertices of the P 2r with the central vertices of the respective K 1,s 's (see also Figure 1; a graph of the form D 1 s is said to be a double-star). It is easy to verify that vc(D r s ) = r + 1 and ivc(D r s ) = r + s. Therefore, for every r ≥ 1, D r f (r+1) cannot be H-free by definition of f . Note that for r, s, t ≥ 1, if s ≤ t then D r s is an induced subgraph of D r t . Therefore, for each r ≥ 1, there must be an s such that D r s is not H-free. In other words, for each r ≥ 1, H must be an induced subgraph of D r s for some s. In particular, this means that we may assume that H is an induced subgraph of D 1 t for some t ≥ 1. If H contains at most one of the central vertices of the stars that form the D 1 t , then H is an induced subgraph of K 1,t + tP 1 and we are done, so we may assume H contains both central vertices. If one of these central vertices has at most one neighbour that is not a central vertex, then H is an induced subgraph of K + 1,t+1 , and we are done. We may therefore assume that H contains an induced D 1 2 . However, for every s ≥ 1, D 2 s is D 1 2 -free and therefore H-free. This contradiction completes the proof.

Feedback Vertex Set
Recall that a graph has an independent feedback vertex set if and only if it is near-bipartite.
Proof. Let G be a (P 1 + P 2 )-free near-bipartite graph. Note that G is a P 3 -free graph, so G is a disjoint union of cliques. It follows that G is a complete multi-partite graph, say with a partition of its vertex sets into k non-empty independent sets V 1 , . . . , V k . We may assume that k ≥ 2, otherwise G is an edgeless graph, in which case ifvs(G) = fvs(G) = 0 and we are done. Since G is near-bipartite, it contains an independent set I such that G − I is a forest. Note that I ⊆ V i for some i ∈ {1, . . . , k}. Since near-bipartite graphs are 3-colourable, it follows that k ≤ 3. Furthermore, if k = 3, then |V j | = 1 for some j ∈ {1, 2, 3} \ {i}, otherwise G − I would contain an induced C 4 , a contradiction. In other words G is either a complete bipartite graph or the graph formed from a complete bipartite graph by adding a dominating vertex. First suppose that k = 2, so G is a complete bipartite graph. Without loss of generality assume that |V 1 | ≥ |V 2 | ≥ 1. Let S be a feedback vertex set of G. If there are two vertices in V 1 \ S and two vertices in V 2 \ S, then these vertices would induce a C 4 in G − S, a contradiction. Therefore S must contain all but at most one vertex of V 1 or all but at most one vertex of V 2 , so fvs(G) ≥ min(|V 1 | − 1, |V 2 | − 1) = |V 2 | − 1. Let I be a set consisting of |V 2 | − 1 vertices of V 2 . Then I is independent and G − I is a star, so I is an independent feedback vertex set. It follows that ifvs(G) ≤ |V 2 | − 1. Since fvs(G) ≤ ifvs(G), we conclude that ifvs(G) = fvs(G) in this case. Now suppose that k = 3, so G is obtained from a complete bipartite graph by adding a dominating vertex. Without loss of generality assume that |V 1 | ≥ |V 2 | ≥ |V 3 | = 1. Let S be a feedback vertex set of G. By the same argument as in the k = 2 case, S must contain all but at most one vertex of V 1 or all but at most one vertex of V 2 . If there is a vertex in V i \ S for all i ∈ {1, 2, 3}, then these three vertices would induce a C 3 in G − S, a contradiction. Therefore S must contain every vertex in V i for some i ∈ {1, 2, 3}. Since |V 1 | ≥ |V 2 | ≥ |V 3 | = 1, it follows that |S| ≥ min(|V 2 | − 1 + |V 3 |, |V 2 |) = |V 2 |. Therefore fvs(G) ≥ |V 2 |. Now V 2 is an independent set and G − V 2 is a star, so V 2 is an independent feedback vertex set. It follows that ifvs(G) ≤ |V 2 |. Since fvs(G) ≤ ifvs(G), we conclude that ifvs(G) = fvs(G).
Proof. Fix integers k ≥ 0 and r ≥ 1. Suppose G is a K 1,r -free near-bipartite graph with a feedback vertex set S such that |S| = k. Since G is near-bipartite, V (G) can be partitioned into an independent set V 1 and a set V (G) \ V 1 that induces a forest in G. Since forests are bipartite, we can partition V (G) \ V 1 into two independent sets V 2 and V 3 .
To prove the lemma, we will iteratively modify S until we obtain an independent feedback vertex set S of G with |S | ≤ (2r 2 − 5r + 3)|S|. Every vertex u ∈ S has at most 2r − 2 neighbours in F . Consider two neighbours v, w of u in F . As F is a forest, there is at most one induced path in F from v to w, so there is at most one induced cycle in G[F ∪ {u}] that contains all of u, v and w. Therefore G[F ∪ {u}] contains at most 2r−2 2 = 1 2 (2r − 2)(2r − 2 − 1) = 2r 2 − 5r + 3 induced cycles. Note that every cycle in G contains at least one vertex of V 1 . Therefore, if s ∈ S ∩(V 2 ∪V 3 ), then we can find a set X of at most 2r 2 −5r+3 vertices in V 1 \S such that if we replace s in S by the vertices of X, then we again obtain a feedback vertex set. Repeating this process iteratively, for each vertex we remove from S ∩(V 2 ∪V 3 ), we add at most 2r 2 −5r+3 vertices to S ∩ V 1 . We stop the procedure once S ∩ (V 2 ∪ V 3 ) becomes empty, at which point we have produced a feedback vertex set S with |S | ≤ (2r 2 − 5r + 3)|S|. Furthermore, at this point S ⊆ V 1 , so S is independent. It follows that ifvs(G) ≤ (2r 2 − 5r + 3) fvs(G).
Note that all near-bipartite graphs are 3-colourable (use one colour for the independent set and the two other colours for the forest). We prove the following lemma. Lemma 9. Let k ≥ 3. The class of C k -free near-bipartite graphs is ifvs-unbounded and ioct-unbounded.
Proof. For r, s ≥ 2, let S r s denote the graph constructed as follows (see also Figure 2). Start with the graph that is the disjoint union of 2s copies of P 2r , and label these copies U 1 , . . . , U s , V 1 , . . . , V s . Add a vertex u adjacent to both endpoints of every U i and a vertex v adjacent to both endpoints of every V i . Finally, add an edge between u and v.
Every induced cycle in S r s is isomorphic to C 2r+1 , which is an odd cycle. Thus a set S ⊆ V (S r s ) is a feedback vertex set for S r s if and only if it is an odd cycle transversal for S r s . It follows that fvs(S r s ) = oct(S r s ) and ifvs(S r s ) = ioct(S r s ). Now {u, v} is a minimum feedback vertex set of S r s , so fvs(S r s ) = oct(S r s ) = 2. However, any independent feedback vertex set S contains at most one vertex of u and v; say it does not contain u. Then it must contain at least one vertex of each U i . It follows that ifvs(S r s ) = ioct(S r s ) ≥ s + 1. Since for every s ≥ 2, k ≥ 3, the graph S k s is C k -free. This completes the proof.

63:8
On the Price of Independence Figure 3 The graphs T5 and T 5 . The edge x2x3 is present in T5, but not in T 5 .

Theorem 2 (restated). Let H be a graph. The class of H-free near-bipartite graphs is ifvs-bounded if and only if
H is isomorphic to P 1 + P 2 , a star or an edgeless graph.
Proof. If H = P 1 + P 2 , then the theorem holds by Lemma 7. If H is isomorphic to a star or an edgeless graph, then H is an induced subgraph of K 1,r for some r ≥ 1. In this case the theorem holds by Lemma 8. Now suppose that the class of H-free near-bipartite graphs is ifvs-bounded. By Lemma 9, H must be a forest. We will show that H is isomorphic to P 1 + P 2 , a star or an edgeless graph.
We start by showing that H must be (P 1 +P 3 , 2P 1 +P 2 , 2P 2 )-free. Let vertices x 1 , x 2 , x 3 , x 4 , in that order, form a path on four vertices. For s ≥ 3, let T s be the graph obtained from this path by adding an independent set I on s vertices (see also Figure 3) that is complete to the path and note that T s is near-bipartite. Then {x 1 , x 2 , x 3 } is a minimum feedback vertex set in T s . However, if S is an independent feedback vertex set, then S contains at most two vertices in {x 1 , . . . , x 4 }. Therefore S must contain at least s − 1 vertices of I, otherwise T s − S would contain an induced C 3 or C 4 . Therefore fvs(T s ) = 3 and ifvs(T s ) ≥ s − 1. Note that T s is (P 1 + P 3 , 2P 1 + P 2 , 2P 2 )-free (this is easy to see by casting to the complement and observing that T s is the disjoint union of a P 4 and a complete graph). Therefore H cannot contain P 1 + P 3 , 2P 1 + P 2 or 2P 2 as an induced subgraph, otherwise T s would be H-free, a contradiction.
Next, we show that H must be P 4 -free. For s ≥ 3 let T s be the graph obtained from T s by removing the edge x 2 x 3 (see also Figure 3). Then {x 1 , x 2 , x 3 } is a minimum feedback vertex set in T s , so fvs(T s ) = 3. By the same argument as for T s , we find that ifvs(T s ) ≥ s − 1 Now the complement T s is the disjoint union of a C 4 and a complete graph, so T s is P 4 -free. Therefore H cannot contain P 4 as an induced subgraph.
We may now assume that H is a (P 1 + P 3 , 2P 1 + P 2 , 2P 2 , P 4 )-free forest. If H is connected, then it is a P 4 -free tree, so it is a star, in which case we are done. We may therefore assume that H is disconnected. We may also assume that H contains at least one edge, otherwise we are done. Since H is (2P 1 + P 2 )-free, it cannot have more than two components. Since H is 2P 2 -free, one of its two components must be isomorphic to P 1 . Since H is a (P 1 + P 3 )-free forest, its other component must be isomorphic to P 2 . Hence H is isomorphic to P 1 + P 2 . This completes the proof.

Odd Cycle Transversal
Recall that a graph has an independent odd cycle transversal if and only if it is 3-colourable. We show the following two lemmas.
Proof. Let G be a P 4 -free 3-colourable graph. It suffices to prove the lemma component-wise, so we may assume that G is connected. Note that G cannot contain any induced odd cycles on more than three vertices, as it is P 4 -free. Let (V 1 , V 2 , V 3 ) be a partition of V (G) into independent sets. We may assume that G is not bipartite, otherwise ioct(G) = oct(G) = 0, in which case we are done. As G is connected, P 4 -free and contains more than one vertex, its complement G must be disconnected. Therefore we can partition the vertex set of G into two parts X 1 and X 2 such that X 1 is complete to X 2 . No independent set V i can have vertices in both X 1 and X 2 , so without loss of generality we may assume that X 1 = V 1 and ] is a P 4 -free bipartite graph, it is readily seen that it is a disjoint union of complete bipartite graphs. Note that G−X 1 is a bipartite graph, so X 1 is an odd cycle transversal of G. Furthermore, X 1 is independent. Now let S be a minimum vertex cover of G[X 2 ]. Observe that G − S is bipartite, so S is an odd cycle transversal of G. Since G[X 2 ] is the disjoint union of complete bipartite graphs, for every component C of G[X 2 ], S must contain one part of the bipartition of C, or the other; by minimality of S, it only contains vertices from one of the parts. It follows that S is independent.
We now claim that every minimum odd cycle transversal S of G contains either X 1 or a minimum vertex cover of G[X 2 ], both of which we have shown are independent odd cycle transversals; by the minimality of S, this will imply that S is equal to one of them. Indeed, suppose for contradiction that there is a vertex x ∈ X 1 \ S and two adjacent vertices y, z ∈ X 2 \ S. Then G[{x, y, z}] is a C 3 in G − S. This contradiction completes the proof.
Proof. Fix an integer k ≥ 0. Let G be a K 1,3 -free 3-colourable graph with an odd cycle transversal S such that |S| = k. Fix a partition of V (G) into three independent sets V 1 , V 2 , V 3 . Without loss of generality assume that 3 . Let S = S and note that G − S is bipartite by definition of odd cycle transversal. To prove the lemma, we will iteratively modify S until we obtain an independent odd cycle transversal S of G with |S | ≤ 3k. Suppose x ∈ V i for some i ∈ {1, 2, 3}. Then x has no neighbours in V i since V i is an independent set. For j ∈ {1, 2, 3} \ {i}, the vertex x can have at most two neighbours in V j , otherwise G would contain an induced K 1,3 . It follows that deg(x) ≤ 4 for all x ∈ V (G).
As G − S is a bipartite K 1,3 -free graph, it is a disjoint union of paths and even cycles. Every vertex u ∈ S has at most four neighbours in V (G) \ S . An induced odd cycle in G − (S \ {u}) consists of the vertex u and an induced path P in G − S between two neighbours v, w of u such that P ∩ N (u) does not contain any vertices apart from v and w. If u has q neighbours in some component C of G − S , then there can be at most q such paths P that lie in this component. It follows that there are at most four induced odd cycles in G − (S \ {u}). Note that every induced odd cycle in G contains at least one vertex in each V i . Therefore, if s ∈ S ∩ (V 2 ∪ V 3 ), then we can find a set X of at most four vertices in V 1 \ S such that if we replace s in S by the vertices of X, then we again obtain an odd cycle transversal. Repeating this process iteratively, for each vertex we remove from S ∩ (V 2 ∪ V 3 ), we add at most four vertices to S ∩ V 1 , so |S | increases by at most 3. We stop the procedure once S ∩ (V 2 ∪ V 3 ) becomes empty, at which point we have produced an odd cycle transversal S with |S | ≤ |S| + 3|S ∩ (V 2 ∪ V 3 )| ≤ k + 2 × 2k 3 = 3k. Furthermore, at this point S ⊆ V 1 , so S is independent. It follows that ioct(G) ≤ 3 oct(G).

63:10
On the Price of Independence Proof. Fix r, s ≥ 1 and k ≥ 0. Let G be a (K 1,r + sP 1 )-free 3-colourable graph with a minimum odd-cycle transversal T on k vertices. Fix a partition of V (G) into three independent sets V 1 , V 2 , V 3 . We may assume that oct(G) ≥ 2, otherwise ioct(G) = oct(G) and we are done. If |V i | ≤ max(oct(G)r + r 2 + 3rs − 2r, f (oct(G))) for some i ∈ {1, 2, 3}, then deleting V i from G yields a bipartite graph, so ioct(G) ≤ max(oct(G)r + r 2 + 3rs − 2r, f (oct(G))) and we are done. We may therefore assume that |V i | > max(oct(G)r + r 2 + 3rs − 2r, f (oct(G))) for all i ∈ {1, 2, 3}. If G is K 1,r -free, then ioct(G) ≤ f (oct(G)), so suppose that G contains an induced K 1,r , say with vertex set X. Note that |X| = r + 1, and each V i can contain at most r vertices of X, since every V i is an independent set.
For every i ∈ {1, 2, 3}, there cannot be a set of s vertices in V i \ X that are anticomplete to X, otherwise G would contain an induced K 1,r + sP 1 , a contradiction. For every i ∈ {1, 2, 3}, since |V i | > oct(G)r + r 2 + 3rs − 2r ≥ r 2 + 3rs, it follows that |V i \ X| ≥ |V i | − r > (s − 1) + (r + 1)(r − 1) = (s − 1) + |X|(r − 1). Hence for every i ∈ {1, 2, 3}, there must be a vertex x ∈ X that has at least r neighbours in V i . Applying this for each i in turn, we find that at least two of the graphs in } contain a vertex of degree at least r; without loss of generality assume that this is the case for Let V 2 and V 3 denote the set of vertices in V 2 and V 3 , respectively, that have more than s − 1 non-neighbours in V 1 . By Lemma 4, |V 2 |, |V 3 | ≤ s − 1.
Suppose a vertex x ∈ V 2 \V 2 is adjacent to a vertex y ∈ V 3 \V 3 . By definition of V 2 and V 3 , the vertices x and y each have at most s − 1 non-neighbours in V 1 . Since |V 1 | − 2(s − 1) ≥ oct(G)+1, it follows that |N (x)∩N (y)∩V 1 | ≥ oct(G)+1 so N (x)∩N (y)∩V 1 ⊆ T . We conclude that at least one of x or y must be in T . In other words, ] of size at most oct(G)+2(s−1). By Lemma 5, there is an independent vertex cover T of G[V 2 ∪V 3 ] of size at most (oct(G)+2(s−1))r+rs = oct(G)r+3rs−2r. Note that by definition of vertex cover, (V 2 ∪V 3 )\T is an independent set, and so G−T is bipartite. Therefore T is an independent odd cycle transversal for G of size at most oct(G)r + 3rs − 2r. This completes the proof.
The following result follows immediately from combining Lemmas 11 and 12.
Proof. Let s ≥ 2. We construct the graph Q s as follows (see also Figure 4). First, let A, B and C be disjoint independent sets of s vertices. Choose vertices a ∈ A, b ∈ B and c ∈ C. Add edges so that a is complete to B ∪ C, b is complete to A ∪ C and c is complete to A ∪ B. Let Q s be the resulting graph and note that it is 3-colourable with colour classes A, B and C.
Note that {a, b} is a minimum odd cycle transversal of Q s , so oct(Q s ) = 2. Let S be a minimum independent odd cycle transversal. Then S contains at most one vertex in {a, b, c}, say S contains neither b nor c.

Lemma 15.
Let H be a graph with more than one vertex of degree at least 3. Then the class of H-free 3-colourable graphs is ioct-unbounded.
Proof. Let s ≥ 1. We construct the graph Z s as follows (see also Figure 5). Start with the disjoint union of s copies of P 4 and label these copies U 1 , . . . , U s . Add an edge ab and make a and b adjacent to the endpoints of every U i . Let Z s be the resulting graph and note that Z s is 3-colourable (colour a and b with Colours 1 and 2, respectively, colour the endpoints of the U i s with Colour 3 and colour the remaining vertices of the U i s with Colours 1 and 2).
Note that Z s − {a, b} is bipartite, so {a, b} is a minimum odd cycle transversal and oct(Z s ) = 2. However, any independent odd cycle transversal S contains at most one vertex of a and b; say it does not contain a. For every i ∈ {1, . . . , s}, the graph Z s [U i ∪ {a}] is a C 5 . Therefore S must contain at least one vertex from each U i . It follows that ioct(Z s ) ≥ s.
Let H be a graph with more than one vertex of degree at least 3. By Lemma 9, we may assume that H is a forest. It remains to show that Z s is H-free. Suppose, for contradiction, that Z s contains H as an induced subgraph and let x and y be two vertices that have degree at least 3 in H. Since H is a forest, x and y must each have three pairwise non-adjacent neighbours in Z s . The endpoints of each U i have exactly three neighbours, but two of them (a and b) are adjacent. Without loss of generality we may therefore assume that x = a and y = b. Since x has degree at least 3 in H, the vertex x must have a neighbour z = y in H and so z must be the endpoint of a U i . Therefore x, y and z are pairwise adjacent, so H[{x, y, z}] is a C 3 , contradicting the fact that H is a forest. It follows that Z s is H-free. This completes the proof.

63:12
On the Price of Independence Proof. Let s ≥ 1. We construct the graph Y s as follows (see also Figure 6).  d 1 and y adjacent to a 1 , b 1 , c 1 and d 1 . Let Y s be the resulting graph.
First note that Y s is K 1,5 -free. The vertices y, a 1 and d 1 all have degree 5, but their neighbourhood is not independent, so they cannot be the central vertex of an induced K 1,5 . All the other vertices have degree at most 4, so they cannot be the central vertex of an induced K 1,5 either. Therefore no vertex in Y s is the central vertex of an induced K 1,5 , so Y s is K 1,5 -free.
The graph Y s − {x, y} is bipartite with bipartition classes: Furthermore, Y s is 3-colourable with colour classes: In fact, we will show that this 3-colouring is unique (up to permuting the colours). To see this, suppose that c : V (Y s ) → {1, 2, 3} is a 3-colouring of Y s . Since x and y are adjacent we may assume without loss of generality that c(x) = 1 and c(y) = 2. Since a 1 and d 1 are adjacent to both x and y, it follows that c(a 1 ) = c(d 1 ) = 3. Since b 1 is adjacent to y and a 1 , it follows that c(b 1 ) = 1. By symmetry c(c 1 ) = 1.
We prove by induction on i that for every i ∈ {1, . . . , 3s}, c(a i ) = c(d i ) ≡ i + 2 mod 3 and c(b i ) = c(c i ) ≡ i mod 3. We have shown that this is true for i = 1. Suppose that the claim holds for i − 1 for some i ∈ {2, . . . , 3s}. Then c(a i−1 ) = c(d i−1 ) ≡ (i − 1) + 2 mod 3 and c(b i−1 ) = c(c i−1 ) ≡ i − 1 mod 3. Since b i is adjacent to b i−1 and d i−1 , it follows that c(b i ) ≡ i mod 3. Since a i is adjacent to b i and a i−1 , it follows that c(a i ) ≡ i + 2 mod 3.
By symmetry c(c i ) ≡ i mod 3 and c(d i ) ≡ i + 2 mod 3. Therefore the claim holds for i. By induction, this completes the proof of the claim and therefore shows that the 3-colouring of Y s is indeed unique.
Furthermore, note that the colour classes in this colouring have sizes 4s + 1, 4s + 1 and 4s, respectively. A set S is an independent odd cycle transversal of a graph if and only if it is a colour class in some 3-colouring of this graph. It follows that ioct(Y s ) = 4s. This completes the proof.
We summarise the results of this section in the following theorem.

Theorem 3 (restated). Let H be a graph. The class of H-free 3-colourable graphs is ioct-bounded:
if H is an induced subgraph of P 4 or K 1,3 + sP 1 for some s ≥ 0 and only if H is an induced subgraph of K + 1,4 or K 1,4 + sP 1 for some s ≥ 0.
Proof. If H is an induced subgraph of P 4 or K 1,3 + sP 1 for some s ≥ 0, then the class of H-free 3-colourable graphs is ioct-bounded by Lemma 10 or Corollary 13, respectively. Now suppose that the class of H-free 3-colourable graphs is ioct-bounded. By Lemma 9, H must be a forest. By Lemma 16, H must be K 1,5 -free. Since H is a K 1,5 -free forest, it has maximum degree at most 4. By Lemma 14, H must be (P 1 + P 4 , 2P 2 )-free.
First suppose that H is P 4 -free, so every component of H is a P 4 -free tree. Hence every component of H is a star. In fact, as H has maximum degree at most 4, every component of H is an induced subgraph of K 1,4 . As H is 2P 2 -free, at most one component of H contains an edge. Therefore H is an induced subgraph of K 1,4 + sP 1 for some s ≥ 0 and we are done. Now suppose that H contains an induced P 4 , say on vertices x 1 , x 2 , x 3 , x 4 in that order and let X = {x 1 , . . . , x 4 }. Since H is a forest, every vertex v ∈ V (H) \ X has at most one neighbour in X. A vertex v ∈ V (H) \ X cannot be adjacent to x 1 or x 4 , since H is 2P 2 -free. By Lemma 15, the vertices x 2 and x 3 cannot both have neighbours outside X; without loss of generality assume that x 3 has no neighbours in V (H) \ X. Since H is (P 1 + P 4 )-free, every vertex v ∈ V (H) \ X must have at least one neighbour in X, so it must be adjacent to x 2 . As H has maximum degree at most 4, it follows that H is an induced subgraph of K + 1,4 . This completes the proof.

Conclusions
To develop an insight into the price of independence for classical concepts, we have investigated whether or not the size of a minimum independent vertex cover, feedback vertex set or odd cycle transversal is bounded in terms of the minimum size of the not-necessarilyindependent variant of each of these transversals for H-free graphs (that have such independent transversals). We have determined the answer for every graph H except for the three missing cases for odd cycle transversal listed in Open Problem 17. While we note that the bounds we give in some of our results (Lemmas 7 and 10) are tight, in this paper we were mainly concerned with obtaining the dichotomy results on whether there is a bound, rather than trying to find exact bounds.
As results for the price of connectivity implied algorithmic consequences [9,13], it is natural to ask if our results for the price of independence have similar consequences. The problems Independent Vertex Cover, Independent Feedback Vertex Set and Independent Odd Cycle Transversal ask to determine the minimum size of the corresponding independent transversal. The first problem is readily seen to be polynomialtime solvable. The other two problems are NP-hard for H-free graphs whenever H is not a linear forest [3], just like their classical counterparts Feedback Vertex Set [16,17] and Odd Cycle Transversal [10] (see also [14,15]). The complexity of these four problems restricted to H-free graphs is still poorly understood when H is a linear forest. Our results suggest that it is unlikely that we can obtain polynomial algorithms for the independent variants based on results for the original variants, as the difference between ifvs(G) and fvs(G) and between ioct(G) and oct(G) can become unbounded quickly.