GGAME BRUSH NUMBER
WILLIAM B. KINNERSLEY AND PAWE(cid:32)L PRA(cid:32)LAT
Abstract.
We study a two-person game based on the well-studied brushing process ongraphs. Players Min and Max alternately place brushes on the vertices of a graph. Whena vertex accumulates at least as many brushes as its degree, it sends one brush to eachneighbor and is removed from the graph; this may in turn induce the removal of othervertices. The game ends once all vertices have been removed. Min seeks to minimize thenumber of brushes played during the game, while Max seeks to maximize it. When bothplayers play optimally, the length of the game is the game brush number of the graph G ,denoted b g ( G ).By considering strategies for both players and modelling the evolution of the game withdifferential equations, we provide an asymptotic value for the game brush number of thecomplete graph; namely, we show that b g ( K n ) = (1+ o (1)) n /e. Using a fractional version ofthe game, we couple the game brush numbers of complete graphs and the binomial randomgraph G ( n, p ). It is shown that for pn (cid:29) ln n asymptotically almost surely b g ( G ( n, p )) =(1 + o (1)) pb g ( K n ) = (1 + o (1)) pn /e . Finally, we study the relationship between the gamebrush number and the (original) brush number. Introduction
Imagine a network of pipes that must be periodically cleaned of a regenerating contami-nant, say algae. In cleaning such a network, there is an initial configuration of brushes onvertices, and every vertex and edge is initially regarded as dirty . A vertex is ready to becleaned if it has at least as many brushes as incident dirty edges. When a vertex is cleaned,it sends one brush along each incident dirty edge; these edges are now said to be clean . (Nobrush ever traverses a clean edge.) The vertex is also deemed clean. Excess brushes remainon the clean vertex and take no further part in the process. (In fact, for our purposes inthis paper, we may think about clean vertices as if they were removed from the graph.) Thegoal is to clean all vertices (and hence also all edges) of the graph using as few brushes aspossible. The minimum number of brushes needed to clean a graph G is the brush number of G , denoted b ( G ).Figure 1 illustrates the cleaning process for a graph G where there are initially 2 brushesat vertex a . The solid edges indicate dirty edges while the dotted edges indicate clean edges.For example, the process starts with vertex a being cleaned, sending a brush to each ofvertices b and c .This model, which was introduced in [15], is tightly connected to the concept of minimumtotal imbalance of a graph, which is used in graph drawing theory. The cleaning process hasbeen well studied, especially on random graphs [1, 18]. (See also [13] for algorithmic aspects,[16, 17] for a related model of cleaning with brooms, [6] for a variant with no edge capacityrestrictions, [4] for a variant in which vertices can send out no more than k brushes, and [11]for a combinatorial game.) Owing to inspiration from chip-firing processes [2, 14], brushesdisperse from an individual vertex in unison, provided that their vertex meets the criteria a r X i v : . [ m a t h . C O ] J a n igure 1. An example of the cleaning process for graph G .to be cleaned. Models in which multiple vertices may be cleaned simultaneously are called parallel cleaning models ; see [10] for more details. In contrast, sequential cleaning models mandate that vertices get cleaned one at a time. The variant considered in [15] and the onewe consider in this paper are sequential in nature.The brushing game that we introduce in this paper is a two-player game played on a graph G . Initially, every vertex and edge is dirty and there are no brushes on any vertices. Theplayers, Max and Min, alternate turns; on each turn, a player adds one brush to a vertexof his or her choosing. When a vertex accumulates at least as many brushes as it has dirtyneighbors, it fires , sending one brush to each dirty neighbor. All edges incident to this vertexbecome clean, and the vertex itself becomes clean. This may in turn make other verticesready to fire, so the process continues until we obtain a stable configuration. (It is knownthat the sequence in which vertices fire does not affect the distribution of brushes on dirtyvertices—see below for more details.) The game ends when all vertices (and so all edges aswell) are clean. Max aims to maximize the number of brushes played before this point, whileMin aims to minimize it. When Min starts and both players play optimally, the length ofthe game on G is the game brush number of G , denoted b g ( G ); we use (cid:98) b g ( G ) to denote thevariant of the game in which Max plays first.The game brush number follows in the same spirit as the game matching number [7], gamechromatic number [8], game domination number [5], toppling number [3], etc., in whichplayers with conflicting objectives together make choices that produce a feasible solutionto some optimization problem. The general area can be called competitive optimization .Competitive optimization processes can also be viewed as on-line problems in which onewants to build a solution to some optimization problem, despite not having complete controlover its construction. In this context, the game models adversarial analysis of an algorithmfor solving the problem: one player represents the algorithm itself, the other player representsthe hypothetical adversary, and the outcome of the game represents the algorithm’s worst-case performance.Throughout this paper, we consider only finite, simple, undirected graphs. For backgroundon graph theory, the reader is directed to [19].1.1. Main results.
Let us start with the following convenient bound proved in Section 2.
Theorem 1.
Always (cid:12)(cid:12)(cid:12) b g ( G ) − (cid:98) b g ( G ) (cid:12)(cid:12)(cid:12) ≤ . Theorem 1 is best possible. For the star K , k − , we have b g ( K , k − ) = 2 k − (cid:98) b g ( K , k − ) = 2 k ; in particular, we have b g ( P ) = 1 and (cid:98) b g ( P ) = 2. On the other hand, for ≥ b g ( C n ) = 3 and (cid:98) b g ( C n ) = 2. (We remark without proof that the graph G k obtained by taking 5 k triangles and identifying a single vertex of each has b g ( G k ) = 8 k + 1and (cid:98) b g ( G k ) = 8 k ; this yields another family—with unbounded brush number—witnessingsharpness of the bound. We leave the details to the reader.)For the remainder of the paper, we will be concerned primarily with the asymptotics of b g over various families of graphs. Hence the difference between b g and (cid:98) b g is unimportant; weuse whichever is most convenient (but prefer b g in general).Because the game produces a feasible solution to the original problem, the value of thegame parameter is bounded by that of the original optimization parameter. Proposition 2.
Always b ( G ) ≤ b g ( G ) ≤ b ( G ) − and b ( G ) ≤ (cid:98) b g ( G ) ≤ b ( G ) . In Section 2, we show that these bounds, though elementary, are best possible in a strongsense:
Theorem 3.
For every rational number r in [1 , , there exists a graph G such that b g ( G ) b ( G ) = r. We next turn our attention to complete graphs. Our next main result (proved in Section 3)provides the asymptotic behavior of the game brush number for K n . Theorem 4. b g ( K n ) = (1 + o (1)) n /e Finally, we move to random graphs. The random graph G ( n, p ) consists of the probabilityspace (Ω , F , Pr), where Ω is the set of all graphs with vertex set { , , . . . , n } , F is the familyof all subsets of Ω, and for every G ∈ Ω,Pr( G ) = p | E ( G ) | (1 − p )( n ) −| E ( G ) | . This space may be viewed as the set of outcomes of (cid:0) n (cid:1) independent coin flips, one for eachpair ( u, v ) of vertices, where the probability of success (that is, adding edge uv ) is p. Notethat p = p ( n ) may (and usually does) tend to zero as n tends to infinity. All asymptoticsthroughout are as n → ∞ (we emphasize that the notations o ( · ) and O ( · ) refer to functions of n , not necessarily positive, whose growth is bounded). We say that an event in a probabilityspace holds asymptotically almost surely (or a.a.s. ) if the probability that it holds tends to1 as n goes to infinity.Our main result here is the following: Theorem 5.
For p = p ( n ) (cid:29) ln n/n and G ∈ G ( n, p ) , a.a.s. b g ( G ) = (1 + o (1)) pb g ( K n ) = (1 + o (1)) pn /e. The result is proved in Section 4.2.
Preliminaries and relation to the brush number
We begin with a formal definition of the brushing game, along with some terminologyused throughout the paper. To facilitate reasoning about the game, we define it in greatergenerality than was mentioned in the introduction. A configuration of a graph G is anassignment of some nonnegative integer number of brushes to each vertex of G ; we representa configuration by a map f : V ( G ) → N ∪ { } . efinition 6. The brushing game on a graph G with initial configuration f has players Max and
Min . Initially, all vertices of G are deemed dirty , and each vertex v contains f ( v )brushes.The players alternate turns. At the beginning of each turn, the player whose turn it isadds one brush to any dirty vertex. After this, if some vertex v has at least as many brushesas dirty neighbors, then v fires : one brush is added to each dirty neighbor of v , and v itselfis marked clean . (The process of firing v is also referred to as cleaning v .) Vertices continueto fire, sequentially, until no more vertices may fire, at which point the turn ends. (If two ormore vertices are simultaneously able to fire, then the order of firings is chosen arbitrarily—although we will see in Observation 7 that the order does not matter.) If at this point allvertices of G are clean, then the game ends.When both players play optimally, the number of brushes placed during the game is the game brush number of G with initial configuration f , denoted b g ( G ; f ) if Min takes the firstturn and by (cid:98) b g ( G ; f ) when Max does. When f is identically zero, we write b g ( G ) in place of b g ( G ; f ) and refer to it as the game brush number of G . When Min (resp. Max) starts thegame, we sometimes refer to the process as the Min-start (resp.
Max-start ) game.In the Min-start (resp. Max-start) game, a round of the game consists of one turn by Min(resp. Max) and the subsequent turn by Max (resp. Min).We say that G can be cleaned by the configuration f when some list of vertex cleanings,starting from f , results in every vertex of G being cleaned. Given configurations f and g ofa graph G , we say that f dominates g provided that f ( v ) ≥ g ( v ) for all v ∈ V ( G ).The following well-known facts will be of use: Observation 7 ([15]) . For any graph G ,(1) Let f and g be configurations of G such that f dominates g . If G can be cleaned by g , then it can also be cleaned by f .(2) If G can be cleaned by f , then every maximal list of vertex firings (starting from f )cleans all vertices of G . Fact (2) above implies that in the brushing game, it is the multiset of brushes placed thatdetermines whether the game has ended; the order of moves is irrelevant.We also need the following result from [15].
Theorem 8 ([15], Theorem 3.1) . The brush number of any graph is at least half the numberof vertices of odd degree.
Now we are ready to prove Theorem 1. We obtain the theorem as a simple consequenceof a more general lemma.
Lemma 9.
Let f and g be configurations of a graph G . If f dominates g , then b g ( G ; f ) ≤ b g ( G ; g ) and (cid:98) b g ( G ; f ) ≤ (cid:98) b g ( G ; g ) . Moreover, b g ( G ; g ) − b g ( G ; f ) ≤ (cid:80) v ( f ( v ) − g ( v )) and (cid:98) b g ( G ; g ) − (cid:98) b g ( G ; f ) ≤ (cid:80) v ( f ( v ) − g ( v )) .Proof. We prove b g ( G ; f ) ≤ b g ( G ; g ) and b g ( G ; g ) − b g ( G ; f ) ≤ (cid:80) v ( f ( v ) − g ( v )); the proofsof the other inequalities are nearly identical. Consider two instances of the brushing game on G : the f -game , in which we have initial configuration f , and the g -game , in which we haveinitial configuration g . Intuitively, the additional brushes initially present in the f -game annot lengthen the game. However, for a formal proof, more care is needed: these extrabrushes could cause some vertices to fire earlier in the f -game, which could invalidate someof Min’s desired moves (since she cannot add brushes to clean vertices).We give a strategy for Min in the f -game. Min plays the f -game and g -game simul-taneously; the f -game is the “real” game in which both players play, while the g -game is“imagined” by Min to guide her play in the f -game. At all times, we denote by f ∗ the currentconfiguration of the f -game and by g ∗ the current configuration of the g -game. Likewise, forall v ∈ V ( G ), at all times c f ( v ) and c g ( v ) denote the number of clean neighbors of v in the f -game and in the g -game, respectively.Min aims to ensure that the f -game finishes no later than the g -game. She maintains twoinvariants: after each turn, (1): f ∗ ( v ) − c f ( v ) ≥ g ∗ ( v ) − c g ( v ) for each vertex v that is dirty in both games, and (2): every clean vertex in the g -game is also clean in the f -game.Note that firing the neighbor of some dirty vertex v increases f ∗ ( v ) by 1 and decreases c f ( v )by 1, so invariant (1) is maintained under firing a vertex (in both games). We claim thatinvariant (2) follows from invariant (1). We prove this through induction on the number ofturns played. Clearly, invariant (2) holds at the beginning of the game. Fix some nonnegativeinteger t and suppose both invariants hold after t turns; we show that invariant (2) mustalso hold after t + 1 turns. Consider the state of both games at the beginning of turn t + 1, before any vertices have fired. A clean vertex v may fire in the f -game if and only if f ∗ ( v ) ≥ deg( v ) − c f ( v ) or, equivalently, f ∗ ( v ) − c f ( v ) ≥ deg( v ) − c f ( v ). Likewise, v may firein the g -game if and only if g ∗ ( v ) ≥ deg( v ) − c g ( v ). Suppose now that vertices x , x , . . . , x k fire in the g -game, in that order. Consider each x i in turn. If x i is already clean in the f -game, then after it fires in the g -game, both invariants still hold. On the other hand, if x i is dirty in the g -game, then f ∗ ( x ) − c f ( x ) ≥ g ∗ ( x ) − c g ( x ) ≥ deg( x ) − c g ( x ) ≥ deg( x ) − c f ( x ) , where the first inequality follows from invariant (1), the second from the assumption that x i can fire in the g -game, and the last from invariant (2). Thus x i may fire in the f -gameas well, and again both invariants hold. It follows that invariant (2) holds after all of the x i have fired and, hence, after turn t + 1. Thus we need only explicitly verify that invariant (1)holds throughout the game, since this implies that invariant (2) holds as well.Since f dominates g , invariant (1) holds initially. Subsequently, on Max’s turns, Maxplays in the f -game, after which Min imagines the same move in the g -game. (Note thatevery valid move in the f -game is valid also in the g -game, since every dirty vertex in the f -game is also dirty in the g -game.) This clearly maintains invariant (1). On Min’s turns,Min chooses some optimal move in the g -game; suppose she adds a brush to vertex v . If v is clean in the f -game, then Min plays the same move there, which maintains invariant(1). Otherwise, Min plays any valid move in the f -game. (If there are no valid moves inthe f -game, then the f -game has ended no later than the g -game, as desired.) In this case, v must have been clean in the f -game but dirty in the g -game, so v itself has no bearingon the invariant; indeed, since we consider only those vertices that are dirty in both games,invariant (1) is maintained.Min follows an optimal strategy for the g -game, so that game lasts for at most b g ( G ; g )turns. Throughout the game, every clean vertex in the g -game is also clean in the f -game; ence the f -game finishes no later than the g -game. Consequently b g ( G ; f ) ≤ b g ( G ; g ), asclaimed.Finally, to prove b g ( G ; g ) − b g ( G ; f ) ≤ (cid:80) v ( f ( v ) − g ( v )), we observe that Min couldapply the following strategy: starting from configuration g , for each vertex v , iterativelyadd up to f ( v ) − g ( v ) brushes to v (as long as v remains dirty). This requires at most (cid:80) v ( f ( v ) − g ( v )) turns by Min (and hence an equal number of turns by Max). The resultingconfiguration dominates f , so from this point, Min can end the game in at most b g ( G ; f )additional turns. (cid:3) Lemma 9 formalizes the intuition that extra brushes cannot hinder Min nor help Max. Asa consequence, when seeking an upper bound on b g ( G ) for some graph G , we may “ignore”brushes that have been placed, pretending that they simply do not exist; this cannot shortenthe remainder of the game. This can be quite useful when we want to give a strategy forMin but do not want to worry about some or all of Max’s moves. Likewise, when we seeka lower bound on b g ( G ), we may pretend that the graph contains extra brushes that werenever actually played. Conversely, ignoring a brush can increase the length of the game byat most 2, and adding a brush can decrease the length of the game by at most 2.Lemma 9 also yields a quick proof of Theorem 1: Proof of Theorem 1.
Let Max play an optimal first move in the Max-start game, and denotethe resulting configuration by f . The remainder of the game can be viewed as a Min-startgame with initial configuration f . Thus by Lemma 9, (cid:98) b g ( G ) = b g ( G ; f ) + 1 ≤ b g ( G ) + 1.Similarly, let Min play an optimal first move in the Min-start game, and denote theresulting configuration by g . This time, Lemma 9 yields b g ( G ) = (cid:98) b g ( G ; g ) + 1 ≤ (cid:98) b g ( G ) + 1.This completes the proof. (cid:3) Recall from Proposition 2 that always b ( G ) ≤ b g ( G ) ≤ b ( G ) − b ( G ) ≤ (cid:98) b g ( G ) ≤ b ( G ), because Min can use her first b ( G ) turns to play brushes in some configuration realizing b ( G ). Thus always 1 ≤ b g ( G ) /b ( G ) <
2. It is natural to ask whether the set of all such ratios,over all connected graphs G , is dense in [1 , n -comb B n and n -sunlet S n are the graphs obtained from P n and from C n , respectively, byattaching one pendant leaf to each vertex. Lemma 10.
Fix n , n , . . . , n m , all at least 2. Let G be the disjoint union of B n , B n , . . . , B n m .Suppose furthermore that, in each component of G , the two vertices of degree 2 contain onebrush each, while all other vertices are empty. If we play the brushing game from this initialconfiguration, then the number of turns needed to clean G is (cid:80) i n i − m , regardless of whichplayer moves first.Proof. Let n = (cid:80) i n i and let f denote the specified initial configuration of G . We provethat b g ( G ; f ) and (cid:98) b g ( G ; f ) are bounded above and below by n − m .We begin with the lower bound. The graph G can be obtained from S n + m by placing onebrush on each of m appropriately-chosen pendant leaves and their neighbors, then allowingthese vertices to fire. (Here we are supposing for convenience that clean vertices are deletedfrom the graph; we may do so because clean vertices no longer affect the game.) Since S n + m contains 2( n + m ) vertices of odd degree, Theorem 8 yields b ( S n + m ) ≥ n + m . Since 2 m rushes have already been placed, the number of additional brushes needed to clean G is atleast n − m ; this establishes the desired bound.For the upper bound, we give a strategy for Min. We use induction on n . When n = 2 theclaim is clear, so suppose n ≥
3. If Min plays first, then she plays on any vertex of degree 2.If the component in which she played was isomorphic to B , then the entire component getscleaned, and only m − n − − ( m − n − m −
1, and thedesired bound follows. If instead the component was isomorphic to B k for k ≥
3, then onlythe vertex at which Min played and its pendant leaf fire, leaving a component isomorphic to B k − with the desired initial configuration. The induction hypothesis again shows that thegame lasts at most n − m − v be the vertex at which Max plays. If v has degree 3, then Min playsat its pendant leaf; if v is itself a pendant leaf, then Min plays at its neighbor. In eithercase, v and its neighbor both fire, leaving a new graph H of the specified form; say H has m (cid:48) components and 2 n (cid:48) vertices. Of the two vertices cleaned, let x be the one with degree 3.If x has two neighbors of degree 2, then x belongs to a copy of B in G , all vertices of whichfire. Hence n (cid:48) = n − m (cid:48) = m −
1, so the number of turns needed to clean H is at most n (cid:48) − m (cid:48) , which simplifies to n − m −
2. If instead x has one neighbor of degree 2, then x ,its neighbor, and their pendant leaves all fire. Thus n (cid:48) = n − m (cid:48) = m , so again thenumber of turns needed to clean H is at most n − m −
2. Finally, if x has no neighbors ofdegree 2, then x and its neighbor fire, leaving two new components in H . Now n (cid:48) = n − m (cid:48) = m + 1, so once again the number of turns needed to clean H is at most n − m − G is at most 2 + ( n − m − n − m , as claimed. (cid:3) Before proceeding, we need one more technical lemma.
Lemma 11.
Let f and g be configurations of a graph G , and let v be a vertex of degree 2 in G . Suppose that f ( v ) = 0 , g ( v ) = 1 , and f ( u ) = g ( u ) for all u ∈ V ( G ) − { v } . If G can becleaned by g , then it can be cleaned by f .Proof. When some neighbor of v fires, v loses a dirty neighbor and gains a brush. Hence,whether we start from f or g , vertex v may fire when and only when one of its neighborshas fired. In both cases, when v fires, it sends one brush to each remaining dirty neighbor.Thus, after v fires, the number of brushes on each remaining dirty vertex is independent ofwhich configuration we started from. It follows that any list of vertex firings that cleans G starting from g also cleans G starting from f . (cid:3) We are now ready to prove Theorem 3.
Proof of Theorem 3.
For positive integers k and n , with k ≤ n , let G n,k be obtained from S n by choosing any k consecutive pendant edges and subdividing each one 2 n times. Weshow that by choosing n and k appropriately, we can make the ratio b g ( G n,k ) /b ( G n,k ) takeon any rational value in [1 , G n,k has 2 n vertices of odd degree, Theorem 8 yields b ( G n,k ) ≥ n . Conversely, it is easy to see that n brushes suffice to clean G n,k : place one rush on each of n − b ( G n,k ) = n . We claim that b g ( G n,k ) = n + k −
1, from which it would follow that whenever q ≤ p < q , we have b g ( G q,p − q +1 ) = ( q + ( p − q )) /q = p/q .To show that b g ( G n,k ) ≥ n + k −
1, we give a strategy for Max. Call the pendant pathsof length n threads , and call the vertices of degree 2 thread vertices . For as many turns ashe can, Max places a brush on any thread vertex that does not already have one. Since b ( G n,k ) = n , we may suppose that the game lasts no more than 2 n − n thread vertices, Max can adhere to this strategy until all of the threads havebeen cleaned; in particular, he never plays more than one brush on any thread vertex. Wemay also suppose without loss of generality that Min played no brushes on thread vertices,since she could have achieved the same result by playing a brush that thread’s leaf. Thus,by Lemma 11, we may “ignore” the brushes on thread vertices, in the sense that they haveno impact on the duration of the game. Once all the threads have been cleaned, Max playsarbitrarily. Consider the state of the game just after the last thread is cleaned, and let G (cid:48) be the graph induced by the clean edges of G . Each thread contains at least two verticesof odd degree in G (cid:48) : the leaf necessarily has degree 1, and the other endpoint has eitherdegree 1 or degree 3, depending on whether it has fired. Since G (cid:48) has 2 k vertices of odddegree, and since Max’s brushes did not contribute to its cleaning, Min must have played atleast k brushes so far. Consequently, Max plays at least k − b ( G n,k ) = n , and since Max plays at least k − n + k − k times. By Lemma 9, since we seek an upper bound on the length of thegame, we may ignore the brushes placed by Max during this time. At this point, the graphconsists of an ( n − k )-sunlet in which one edge has been subdivided k times, and in whicheach vertex of degree 2 contains one brush. If it is Min’s turn, then she plays on a vertex ofdegree 2. This causes all vertices of degree 2 to fire, and produces a copy of B n − k having theinitial configuration described in Lemma 10. Since at most 2 k brushes have already beenplayed, and Lemma 10 states that the remainder of the game lasts exactly n − k − k + ( n − k −
1) = n + k −
1, as desired. Likewise, weobtain the same bound if it is Max’s turn, and he chooses to play on a vertex of degree 2.If it is Max’s turn and he chooses to play on a leaf, then Min plays on its neighbor; if Maxchooses to play on a vertex of degree 3, then Min plays on its pendant leaf. In either case,we have a graph of the form specified in Lemma 10, except that one component containsthe subdivided edge. We claim that, despite these extra vertices, the game still lasts for atmost n − k − n − k − k + ( n − k −
1) = n + k − (cid:3) . Complete graphs
We next turn our attention to complete graphs. To simplify the presentation of our mainresult, we first present both players’ optimal strategies, we next analyze the performance ofthese strategies, and we save the proof of optimality for last. We give two strategies, one forMax and one for Min. • The balanced strategy for Max: on each turn, Max adds a brush to the dirty vertexhaving the fewest brushes. • The greedy strategy for Min: Min selects the dirty vertex having the most brushes,adds brushes to that vertex until it gets cleaned, and repeats.As we will show later, these strategies yield a Nash equilibrium—that is, the balancedstrategy is an optimal response to the greedy strategy and vice-versa. Hence the game brushnumber of K n is precisely the length of the game when the strategies are employed againsteach other. While there is no simple formula for b g ( K n ), we can use differential equations todetermine the asymptotics. Lemma 12.
Consider the brushing game on K n . When Max uses the balanced strategy andMin uses the greedy strategy, the game ends in (1 + o (1)) n /e turns.Proof. At all times throughout the game, let x denote the number of vertices cleaned so farand y the number of brushes placed. We define the k th phase of the game to be the sequenceof turns during which there are exactly k clean vertices, starting with the turn after that onwhich the k th vertex fires and ending on the turn on which the ( k + 1)st vertex fires.Consider the state of the game at the beginning of phase x (and suppose the game has notyet finished). We analyze the length of this phase. With x vertices cleaned, each remainingvertex must accumulate n − x − n − x remaining dirty vertices has the same number of brushes (up to a difference of 1). Henceeach dirty vertex has yn − x + O (1) brushes. Min must therefore place another n − x − yn − x + O (1)brushes before the next vertex fires; since Max also plays, 2( n − x − yn − x ) + O (1) brushes areintroduced in this phase.To investigate the asymptotic behavior of the game, we simplify the analysis by modelingthe brushing game (a discrete process) using differential equations (a continuous model).The argument above yields y ( x + 1) − y ( x ) = 2 (cid:18) n − x − yn − x (cid:19) + O (1) . We next normalize both parameters by letting t = x/n and f ( t ) = y/n . We thus arrive atthe differential equation f (cid:48) ( t ) = 2 (cid:18) − t − f ( t )1 − t (cid:19) , f (0) = 0 . It is easily verified that the solution to this differential equation is f ( t ) = − − t ) ln(1 − t ).The game ends during phase (1 + o (1)) t n , where t is such that f (cid:48) ( t ) = 0. Solving for t yields t = 1 − e − / and f ( t ) = 1 /e . Hence we expect that the number of brushes playedthroughout the game is (1 + o (1)) f ( t ) n = (1 + o (1)) n /e . or a formal proof, we now return to the original (discrete) model of the brushing game.Fix a nonnegative integer x , and suppose that the game does not end before phase x . Asargued above,2 (cid:18) n − x − y ( x ) n − x (cid:19) − C ≤ y ( x + 1) − y ( x ) ≤ (cid:18) n − x − y ( x ) n − x (cid:19) + C for some constant C . Intuitively, since the “true” length of this phase is close to its valueunder the differential equation model, the “true” length of the game should likewise be closeto the value suggested by the solution of the differential equation.We can prove this formally through induction on x . For all nonnegative integers x suchthat the game does not end before phase x , we claim that y ( x ) = 2( n − x ) ln (cid:18) − x/n (cid:19) + O ( x ) . More precisely, we prove that y ( x ) ≤ n − x ) ln (cid:16) − x/n (cid:17) + C (cid:48) x for some constant C (cid:48) ; theproof of the analogous lower bound is similar. The desired bound trivially holds when x = 0.For the inductive step, fix x ≥
0, suppose that the game does not end before phase x , andassume that y ( x ) ≤ n − x ) ln (cid:16) − x/n (cid:17) + C (cid:48) x . Note that the game must end before phase n/
2: once n/ n/ n/ − x ≤ n/
2. Now y ( x + 1) = y ( x ) + (cid:16) y ( x + 1) − y ( x ) (cid:17) ≤ n − x ) ln (cid:18) − x/n (cid:19) + C (cid:48) x + 2 (cid:18) n − x − y ( x ) n − x (cid:19) + C ≤ n − x ) (cid:18) ( n − x ) ln (cid:18) − x/n (cid:19) + 1 − (cid:18) − x/n (cid:19)(cid:19) − C (cid:48) xn − x + C (cid:48) x + C = 2( n − x ) (cid:18) ( n − x −
2) ln (cid:18) − x/n (cid:19) + 1 (cid:19) + C (cid:48) x (cid:18) − n − x (cid:19) + C ≤ n − x ) (cid:18) ( n − x −
2) ln (cid:18) − ( x + 1) /n (cid:19) + O (cid:18) n − x (cid:19)(cid:19) + C (cid:48) x + C = 2( n − x )( n − x −
2) ln (cid:18) − ( x + 1) /n (cid:19) + C (cid:48) x + O (1)= 2 (cid:16) ( n − x − + 1 (cid:17) ln (cid:18) − ( x + 1) /n (cid:19) + C (cid:48) x + O (1)= 2( n − ( x + 1)) ln (cid:18) − ( x + 1) /n (cid:19) + C (cid:48) x + O (1) , where the first and second inequalities follow from the induction hypothesis and the last linefollows from the assumption that x ≤ n/ / (1 − ( x +1) /n )) = O (1)). If C (cid:48) is takento be sufficiently large, then C (cid:48) x + O (1) ≤ C (cid:48) ( x + 1), so the claimed bound holds. A similarargument establishes the analogous lower bound on y ( x ). As with the differential equationmodel, the game ends immediately following phase t , where t is the largest nonnegativeinteger such that y ( t + 1) − y ( t ) >
0. Solving for t yields t = (1 + o (1))(1 − e − / ) n , from hich it follows that that y ( t + 1) = (1 + o (1)) n /e . Since exactly y ( t + 1) brushes areplayed throughout the game, this completes the proof. (cid:3) We now show that the strategies analyzed in Lemma 12 are optimal for both players. Inthis analysis, we focus on the number of brushes placed directly on each vertex (as opposedto those obtained from clean neighbors). Moreover, we “sort” the vertices by the numbersof brushes played, in non-increasing order: always v denotes the vertex having received themost brushes, v the vertex having received the next most, and so on. By the symmetry of K n , we may suppose that moves are always played so as to avoid re-indexing of the vertices.For example, if v i and v i +1 have received the same number of brushes, then placing a brushon v i +1 would force re-indexing: the old v i becomes the new v i +1 , and the old v i +1 becomesthe new v i . However, we may just as well suppose that the brush was placed directly on v i ,since this produces an isomorphic configuration. This assumption is equivalent to forbiddingeither player from adding a k th brush to v i until k brushes have been played on v i − . Sincewe track only the number of vertices played directly on each vertex, firing a vertex has noeffect on the indexing.Using these conventions, in the course of the game on K n , we clean v first, then v ,and so on. For i < n/
2, vertex v i receives i − n − i brushes to fire. Hence v i fires when and only when the following two conditions havebeen met: first, vertices v , v , . . . , v i − are already clean; second, at least n − i + 1 brusheshave been played on v i . Thus the game ends when and only when n − i + 1 brushes havebeen played on v i for all i in { , , . . . , (cid:98) n/ (cid:99)} . Viewing this condition graphically yields a“triangle” of brushes that must be placed before the game ends (see Figure 2). We refer tothis triangle as the critical triangle , to brushes placed within the triangle as in-brushes , andto brushes placed outside as out-brushes . (Formally, the k th brush played on vertex v i is anin-brush if k ≤ n − i +1 and an out-brush otherwise.) Since the number of in-brushes playedthroughout the game is fixed, Max aims to force many out-brushes to be played, while Minaims to prevent this. v v v v v v v v Figure 2.
An illustration of the critical triangle for K . (In-brushes are shaded.) ow we are ready to prove Theorem 4. Since the upper bound and lower bound usesubstantially different arguments, we split the proof into two parts. We begin with the lowerbound. Lemma 13. b g ( K n ) ≥ (1 + o (1)) n /e Proof.
We show that when Max uses the balanced strategy, Min’s optimal response is toplay the greedy strategy. Consequently, by Lemma 12, Max can enforce the claimed lowerbound by using the balanced strategy.Fix n and consider the brushing game on K n (where Max always uses the balanced strat-egy). Call a strategy for Min optimal if no strategy ends the brushing game on K n in fewerturns. Since Max’s strategy is fixed, a strategy for Min can be viewed as the sequence ofmoves she plays throughout the game. Call a move by Min greedy if it places a brush onthe least-indexed dirty vertex and non-greedy otherwise. It suffices to prove that the greedystrategy is optimal.Fix an optimal strategy A for Min. If in fact A is the greedy strategy, then there is nothingto prove. Suppose instead that A contains at least one non-greedy move, and suppose thatthe last such move is played on Min’s k th turn. We show how to transform A into a newoptimal strategy, B , that either has fewer non-greedy moves than A or has the same numberof non-greedy moves but with the last such move happening after Min’s k th turn. Iteratingthis transformation must eventually yield an optimal strategy with fewer non-greedy movesthan A , and continuing to iterate must eventually yield the greedy strategy (that is, theunique strategy having no non-greedy moves). Since the transformation always produces anoptimal strategy, the greedy strategy must be optimal, as claimed.Before giving the construction of strategy B , we observe that Min’s last move of the gamemust be greedy. If Min’s last move concludes the game, then with that move Min plays thefinal in-brush. Consequently, on Min’s last turn, only one in-brush remains to be played.Under the greedy strategy, Min always plays an in-brush; since Min’s final move places thelast in-brush, this move must in fact be greedy. Suppose instead that the game ends on oneof Max’s turns. In this case, Max plays the final in-brush; say he places this brush on vertex v i . By optimality of A , Min did not an out-brush on her last turn, since otherwise she couldhave ended the game by playing the final in-brush. Hence Min played an in-brush, say onvertex v j . Note that after Min’s last turn, exactly n − i brushes have been played on v i ,exactly n − j + 1 have been played on v j , and at most n − j have been played on v j +1 (wemust have j + 1 ≤ n since no in-brushes can ever be played on v n ). It follows that j < i :otherwise, on his ensuing turn, Max would play another brush on v j +1 rather than on v i .Thus, out of the two in-brushes remaining on Min’s final turn, she played the one on thelower-indexed vertex, so her move was in fact greedy.We now explain how to construct strategy B . Define the A -game (resp. B -game ) to bethe instance of the game when Min follows strategy A (resp. strategy B ). In the B -game,Min plays her first k − A -game. Since Min’s last turn in the A -gameis greedy, she must play at least k + 1 moves in that game. Let us consider Min’s k th and( k + 1)st turns. Say Min plays on vertex v i on her k th turn the A -game, and say that Maxresponds by playing on vertex v j . For her k th turn in the B -game, Min plays a greedy move,say on vertex v i (cid:48) ; say that Max subsequently plays on vertex v j (cid:48) . By assumption, i (cid:54) = i (cid:48) .Suppose first that j = j (cid:48) . On Min’s k th turn in the A -game, the greedy strategy dictatesplaying on v i (cid:48) , hence v i (cid:48) is the least-indexed dirty vertex. Neither Min’s nor Max’s k th ove adds any brushes to v i (cid:48) , so it remains the least-indexed dirty vertex on Min’s ( k + 1)stturn. By assumption, in the A -game, Min plays greedily after turn k ; thus, Min plays her( k + 1)st turn on v i (cid:48) . In the B -game, Min plays on v i for her ( k + 1)st turn and plays greedilythereafter. After Min’s ( k + 1)st turn the two games have identical configurations, proceedidentically, and finish simultaneously. Thus, strategy B must be optimal. Moreover, eitherMin plays fewer non-greedy moves than under strategy A (if v i turns out to be a greedymove in the B -game) or she plays the same number of non-greedy moves but with the lastone occurring on her ( k + 1)st turn (if v i is non-greedy).Now suppose j (cid:54) = j (cid:48) . Since Max uses the balanced strategy in both games, j (cid:54) = j (cid:48) implies i = j (cid:48) . In other words, the balanced strategy tells Max to play his k th turn on v j (cid:48) unless Minplays there first. Moreover, since Max uses the balanced strategy, if he ever causes a vertexto fire, subsequently all remaining vertices fire and the game ends. If j = i (cid:48) , then afterMax’s k th turn, both games have the same configuration. Min henceforth plays greedilyin the B -game and, as above, it follows that B is an optimal strategy of the desired form.Thus, suppose j (cid:54) = i (cid:48) . Since we have assumed that Min plays at least k + 1 turns under theoptimal strategy A , Max’s k th turn must not cause any vertices to fire. Hence v i (cid:48) remains thegreedy move on Min’s ( k + 1)st turn in the A -game, so by assumption she plays there. In the B -game, Min plays her ( k + 1)st move on v j and plays greedily thereafter: once again, bothgames have the same configuration, and the claim follows. This completes the proof. (cid:3) Our proof of the upper bound of Theorem 4 is more technical. For this proof, we considera generalization of the original game. In our abstraction of the brushing game on K n , playersalternately place brushes on a game board, with the game ending once the “critical triangle”is full. In particular, the board we have been considering has n columns, and the criticaltriangle has height n −
1. To prove the upper bound, we consider the family of games BG ( w, h, t ). Each such game is similar to the original, except that the game board has“width” w , the critical triangle has height h , and the game begins with t consecutive turnsby Max. Definition 14.
The game BG ( w, h, t ) has two players, Max and
Min , and is played on arectangular game board divided into square cells . The board consists of w columns and h rows of cells. Initially, each cell of the board is deemed empty , and each column is deemed dirty .The players take turns filling cells of the game board. On each turn, a player may fill thecell in row i , column j provided that: • column j is still dirty, • either i = 0 or the cell in row i −
1, column j has already been filled, and • either j = 1 or the cell in row i , column j − plays in that column. Aftereach turn, each dirty column k is deemed clean if that column has at least h − k + 2 filledcells, and either k = 1 or column k − fires . Columns continue to fire until no more columnsare able to fire. Once all columns have fired, the game ends.At the beginning of the game, Max takes t consecutive turns. After this, the playersalternate turns, with Min taking the ( t + 1)st turn, Max taking the ( t + 2)nd, and so on.Play continues in this manner until the game ends. The r th round of the game consists of urns t + 2 r − t + 2 r – that is, Min’s r th turn and Max’s subsequent turn. The length of the game is the number of turns taken throughout the course of the game. Min aims tominimize the length of the game, while Max aims to maximize it.Note that BG ( n, n − ,
0) is equivalent to the original brushing game on K n . To upper-bound b g ( K n ), we use the following strategy. As in the proof of the lower bound, we playtwo instances of the brushing game: the “real game” and the “ideal game”. We play somenumber of turns in both games and view the remainders as as “sub-games” isomorphic toinstances of BG ( w, h, t ) for appropriate choices of w , h , and t . We then use induction tobound the lengths of the sub-games. In what follows, we write “ b g ∗ ( w, h, t )” to refer to thelength of BG ( w, h, t ) when Min uses the greedy strategy and Max plays optimally. Lemma 15. b g ( K n ) ≤ (1 + o (1)) n /e Proof.
It suffices to show that when Min uses the greedy strategy, Max’s optimal responseis to play the balanced strategy; consequently, by Lemma 12, Min can enforce the claimedupper bound by using the greedy strategy. In fact, we prove a stronger claim: for any naturalnumbers w, h , and t with w > h , the balanced strategy is optimal for Max in BG ( w, h, t ),given that Min uses the greedy strategy. The original claim then follows by taking w = n , h = n −
1, and t = 0. We prove this stronger claim through induction on w . When w ≤ BG ( w, h, t ). As before, we compare two instances of the game. In the real game ,Max uses any fixed strategy, while in the ideal game , he uses the balanced strategy. In bothgames, Min uses the greedy strategy. We claim that the ideal game finishes no sooner thanthe real game, which would establish optimality of Max’s balanced strategy.If Max is forced to fill the critical triangle within his initial t turns, then the balancedstrategy clearly maximizes the length of the game, so we may suppose the game continuespast Max’s initial turns. Suppose column 1 fires in round r of the ideal game, and considerthe state of both games after r rounds. At this point, column 1 must have fired in the realgame as well, since the balanced strategy minimizes the number of cells filled in column 1by Max—in fact, when Min uses the greedy strategy, Max fills no cells in that column afterhis initial t turns. Let t (cid:48) = t + 2 r − h ; in both games, exactly t (cid:48) cells have been filled incolumns 2 , , . . . , w . All t (cid:48) of these cells must have been filled by Max in accordance withthe balanced strategy, since Min’s greedy strategy ensures that she played her first r movesin column 1. If the real game has finished after round r , then the claim holds. If instead thereal game has not yet finished, then the ideal game also cannot have finished and, moreover,column 2 cannot have fired in that game.Suppose first that column 2 has not yet fired in the real game. Since t (cid:48) cells have beenfilled in columns 2 and higher, we may view the remainder of the real game as an instanceof BG ( w − , h − , t (cid:48) ) in which Min uses the greedy strategy. In the real game, some of thecells in columns 2 and higher were filled by Min; in this new instance, we pretend that thesecells were all filled by Max during his initial t (cid:48) turns. The length of this game, and hencethe number of turns remaining in the real game, is at most b g ∗ ( w − , h − , t (cid:48) ) − t (cid:48) . Wemay likewise view the remainder of the ideal game as an instance of BG ( w − , h − , t (cid:48) ) inwhich Min again uses the greedy strategy. However, this time, all cells already filled actuallywere filled by Max and, moreover, they were filled in accordance with the balanced strategy. y the induction hypothesis, the number of turns remaining in the ideal game is exactly b g ∗ ( w − , h − , t (cid:48) ) − t (cid:48) , so the ideal game lasts at least as long as the real game, as claimed.Now suppose instead that after r rounds, exactly k columns have fired in the real gamefor some k ≥
2. If k = w , then the real game is complete and the claim follows, so supposeotherwise. This time, we cannot necessarily view the remainder of the real game as aninstance of BG ( w − , h − , t (cid:48) ), since it might be that more than h − k + 1 , k + 2 , . . . , w . Letting x denote the total number of moves played in columns 2 , , . . . , k ,the number played in columns k + 1 , k + 2 , . . . , w is t (cid:48) − x . Since column k + 1 has not yetfired, it has fewer than h − k filled cells. Hence we may view the remainder of the real gameas an instance of BG ( w − k, h − k, t (cid:48) − x ) in which Min uses the greedy strategy.Let T r and T i denote the number of turns remaining in the real game and the ideal game,respectively. We aim to show that T r ≤ T i , from which the claim would follow. We beginby upper-bounding T r . Just as before, we may pretend that Max played all t (cid:48) − x moves inthis new instance. By the induction hypothesis, we may suppose (since we seek an upperbound on T r ) that these moves were all played in accordance with the balanced strategy.This means that Max has filled y r rows and z r cells in the next row, where y r = (cid:4) t (cid:48) − xw − k (cid:5) let z r = t (cid:48) − x − y r ( w − k ). Thus T r ≤ b g ∗ ( w − k, h r , z r ) − z r , where h r = h − k − y r .Let us now analyze T i . In the ideal game, exactly t (cid:48) moves have been played in columns2 , , . . . , w , all played by Max in accordance with the balanced strategy. Let y i = (cid:4) t (cid:48) w − (cid:5) and let z i = t (cid:48) − y i ( w − w − h − − y i , we see that T i = b g ∗ ( w − , h i , z i ) − z i , where h i = h − − y i .For the sake of comparing the two games, we claim that y i ≤ y r + x − y r ( k − w − + 1. There are x − y r ( k −
1) filled cells in rows y r + 1 and higher of columns 2 , , . . . , k in the real game;if these moves had instead been played according to the balanced strategy, then they wouldhave completed an additional (cid:106) x − y r ( k − w − (cid:107) full rows, and might also have completed a rowthat was already partially full. Thus y i ≤ y r + x − y r ( k − w − ≤ y r + ( h − − y r )( k − w − ≤ y r + ( h − k − w − < y r + k, where the second inequality uses the observation that no more than h − x ≤ ( h − k − h < w . It now follows that h i = h − − y i > h − − ( y r + k ) = h − (2 + k ) − y r ≥ h − k − y r = h r . As argued above, T i = b g ∗ ( w − , h i , z i ) − z i and T r ≤ b g ∗ ( w − k, h r , z r ) − z r . Recall that weaim to show that T i ≥ T i . Toward this end, first note that adding k − b g ∗ ( w − k, h r , z r ) − z r ≤ b g ∗ ( w − , h r , z r ) − z r . Next, weadjust the height of this game. By the induction hypothesis, the greedy strategy is optimalfor Max in b g ( w − , h i , t ∗ ) for all t ∗ , hence b g ∗ ( w − , h r , z r ) − z r = b g ∗ ( w − , h i , z r + ( w − h i − h r )) − ( z r + ( w − h i − h r )) , ince Max’s first ( w − h i − h r ) moves in the latter game fill the first h i − h r rows, whichyields a configuration equivalent to the former game. Letting z ∗ r = z r + ( w − h i − h r ) nowyields T r ≤ b g ∗ ( w − k, h r , z r ) − z r ≤ b g ∗ ( w − , h r , z r ) − z r = b g ∗ ( w − , h i , z ∗ r ) − z ∗ r . Finally, it is easily seen that b g ∗ ( w − , h i , z i ) − z i ≤ b g ∗ ( w − , h i , z ∗ r ) − z ∗ r : the left side of theequation counts the number of turns remaining in the game after an initial z i greedy movesby Max, while the right side counts the number of turns remaining after z ∗ r greedy moves,and the presence of the additional z ∗ r − z i chips cannot lengthen the remainder of the game.Hence T i ≥ T r , which completes the proof. (cid:3) Theorem 4 now follows from Lemmas 13 and 15.4.
Random Graphs
We next seek to establish a connection between the brushing game on the complete graphand the game on the random graph. As an intermediate step, we introduce a fractionalvariant of the brushing game.The fractional brushing game behaves very similarly to the ordinary brushing game, butwith two important changes. First, vertices can contain non-integral numbers of brushes.Second, a vertex v fires when the number of brushes on v is at least p times the number ofdirty neighbors, at which time it sends p brushes to each dirty neighbor (where p is a fixedconstant between 0 and 1). Definition 16.
Given p ∈ (0 , fractional brushing game on a graph G has players Max and
Min . Throughout the game, each vertex of G contains some nonnegative real numberof brushes . Initially, all vertices of G are deemed dirty , and no vertex contains any brushes.The players alternate turns. At the beginning of each turn, if the number of brushes onsome vertex v is at least p times the number of dirty neighbors, then v fires : p brushes areadded to each dirty neighbor of v , and v itself is marked clean . (The process of firing v isalso referred to as cleaning v .) Vertices continue to fire, sequentially, until no more verticesmay fire. If at this point all vertices of G are clean, then the game ends. Otherwise, theplayer whose turn it is adds one brush to any dirty vertex, and the turn ends.Min aims to minimize, and Max aims to maximize, the number of turns taken beforethe game ends. When both players play optimally, the total number of turns taken is the fractional game brush number of G with parameter p , denoted b gp ( G ) if Min takes the firstturn and by (cid:98) b gp ( G ) when Max does. When Min (resp. Max) starts the game, we sometimesrefer to the process as the Min-start (resp.
Max-start ) game.In the Min-start (resp. Max-start) game, a round of the game consists of one turn by Min(resp. Max) and the subsequent turn by Max (resp. Min).We next show that the ordinary and fractional game brush numbers are closely connected.
Theorem 17.
For every graph G and for every positive p = p ( n ) , b gp ( G ) = (1 + o (1)) pb g ( G ) + O ( n ln n ) . Proof.
Initially, we assume that 1 /p is always integral. After proving the theorem under thisadditional assumption, we briefly explain the modifications needed to extend the argumentto all values of p . e bound b gp ( G ) both above and below in terms of b g ( G ). For the lower bound wegive a strategy for Max, and for the upper bound we give a strategy for Min. Since thesestrategies are quite similar, we present them simultaneously. Denote the players by “A”and “B” (where A and B could represent either one of Min or Max). Player A imaginesan instance of the ordinary game on G and uses an optimal strategy in that game to guidehis or her play in the fractional game. The games proceed simultaneously, except that forevery round played in the fractional game, player A simulates 1 /p rounds in the ordinarygame. To simplify the analysis, we introduce a third entity, the Oracle . At various points,the Oracle will “pause” the games and add extra brushes to various vertices, in an attemptto “synchronize” the games by ensuring that all clean vertices in the fractional game arealso clean in the ordinary game and vice-versa. Note that each brush introduced by theOracle can decrease the length of the corresponding game by at most 2. (This follows fromLemma 9; while the lemma only directly applies to the ordinary game, the same argumentsuffices for the fractional game with only cosmetic changes.) Moreover, note that no dirtyvertex ever contains brushes added by the Oracle.Before presenting the strategies, we introduce some additional notation. For a vertex v ,let x A ( v ) (resp. y A ( v )) denote the number of brushes played on v by A in the ordinarygame (resp. fractional game); define x B ( v ) and y B ( v ) similarly. The discrepancy for A ata vertex v is defined by d A ( v ) = px A ( v ) − y A ( v ), and the discrepancy for B is defined by d B ( v ) = px B ( v ) − y B ( v ).Player A plays as follows. By Theorem 1, we may assume without affecting the asymptoticsthat B plays first (in both games), so each round of the fractional game consists of a moveby B and the subsequent move by A. After B plays in the fractional game, A simulates 1 /p moves by B in the ordinary game; for each move, the imagined B plays on some dirty vertex v minimizing d B ( v ). Player A responds to each imagined move according to some optimalstrategy for the ordinary game. Now A plays, in the fractional game, on some dirty vertex v maximizing d A ( v ). Finally, if any vertices are clean in the fractional game but not in theordinary game, then the Oracle adds brushes to these vertices in the ordinary game in orderto clean them. (Note that to clean these vertices, it suffices to add at most p max {− d A ( v ) − d B ( v ) , } brushes to every such vertex v .) Likewise, if any vertices are clean in the ordinarygame but not in the fractional game, then the Oracle adds brushes in the fractional gameto clean them. (This time, the Oracle must add at most max { d A ( v ) + d B ( v ) , } brushes toeach such vertex v .) The Oracle repeats these steps until every vertex that is clean in onegame is also clean in the other.Each brush added by the Oracle changes the length of the corresponding game by at mosta constant number of turns. We want to show that the Oracle need not add many brushes.To this end, we claim that for every vertex v , at all points during the game, d A ( v ) = O (ln n ).Let D = (cid:80) w max { d A ( w ) , } , where the summation is over all vertices w that are dirty inthe ordinary game. In any given round, A’s plays in the ordinary game may increase D by as much as 1. If at this point D exceeds n , then d A ( v ) ≥ v , so A’ssubsequent move in the fractional game will reduce D by 1. Consequently, we have D < n at the beginning of each round, so D < n + 1 at all points during the game. It follows that,at all times, the number of vertices v for which d A ( v ) ≥ . n +12 . , which inturn is less than n/ n . ow let D = (cid:80) w (cid:0) max { d A ( w ) , . } − . (cid:1) , where again the sum is over all vertices w that are dirty in the ordinary game. As above, we must have D < m + 1 at all times,where m denotes the number of vertices with discrepancy at least 2 .
1. The argument aboveshows that m < n , so always D < n +22 . Thus, the number of vertices with discrepancyat least 4 . . D – is always atmost n +22 · . , which is less than n/ n . Iterating this argument shows thatfor all positive integers k , at any point in the game, at most n/ k vertices have discrepancyat least 2 . k . In particular, no vertex can ever have discrepancy at least 2 . (cid:100) log n (cid:101) + 1),hence d A ( v ) = O (ln n ) for every vertex v at all times. A similar argument shows that − d B ( v ) = O (ln n ) for all v .We now bound the length of the fractional game. Let (cid:96) o and (cid:96) f denote the lengths ofthe ordinary and fractional games, respectively, and let m o and m f denote the numbers ofbrushes added by the Oracle to the ordinary and fractional games, respectively. As arguedabove, when the Oracle adds brushes to a vertex v in the ordinary game, it adds at most p max {− d A ( v ) − d B ( v ) , } . Furthermore, once the Oracle adds brushes to a vertex v , thatvertex becomes clean in both games, so d A ( v ) and d B ( v ) cease to change; hence it sufficesto analyze the discrepancies at the end of the game. We now have m o ≤ (cid:88) v ∈ V ( G ) p max {− d A ( v ) − d B ( v ) , }≤ p (cid:88) v ∈ V ( G ) (cid:12)(cid:12) d A ( v ) (cid:12)(cid:12) + (cid:88) v ∈ V ( G ) (cid:12)(cid:12) d B ( v ) (cid:12)(cid:12) ≤ p (cid:88) v : d A ( v ) > d A ( v ) + 2 (cid:88) v : d B ( v ) < − d B ( v ) = O ( n ln n/p ) , where the third inequality uses the fact that (cid:80) v ∈ V ( G ) d A ( v ) = (cid:80) v ∈ V ( G ) d B ( v ) = 0. ByLemma 9, the Oracle’s intervention changes the length of the ordinary game by only O ( n ln n/p )turns. Likewise, m o = O ( n ln n ), so the Oracle’s interference changes the length of the frac-tional game by only O ( n ln n ) turns.To conclude the proof, suppose A is actually Min. Since Min follows an optimal strategy forthe ordinary game (putting aside the Oracle’s interference), we have (cid:96) o ≤ b g ( G )+ O ( n ln n/p ).Likewise, since Max follows an optimal strategy for the fractional game, we have (cid:96) f ≥ b gp ( G ) − O ( n ln n ). The Oracle ensures that the two games finish within one round of eachother, hence b gp ( G ) − O ( n ln n ) ≤ (cid:96) f = p(cid:96) o + O (1) ≤ p ( b g ( G ) + O ( n ln n/p )) + O (1) = pb g ( G ) + O ( n ln n ) , hence b gp ( G ) ≤ pb g ( G ) + O ( n ln n ), which establishes the claimed upper bound on b gp ( G ).Finally, suppose A is Max. This time, Max follows an optimal strategy for the ordinarygame, so (cid:96) o ≥ b g ( G ) − O ( n ln n/p ). Similarly, (cid:96) f ≤ b gp ( G ) + O ( n ln n ). Combining theseobservations, we have b gp ( G ) − O ( n ln n ) ≥ (cid:96) f = p(cid:96) o + O (1) ≥ p ( b g ( G ) + O ( n ln n/p )) + O (1) = pb g ( G ) + O ( n ln n ) , ence b gp ( G ) ≥ pb g ( G ) + O ( n ln n ). This establishes the lower bound, and completes theproof under the assumption that 1 /p is integral.Suppose now that 1 /p is not integral. In the argument above, after every round in thefractional game, player A simulated 1 /p rounds in the ordinary game; when 1 /p is not aninteger, he or she cannot do this. Instead, after round i in the fractional game, A simulates (cid:100) i/p (cid:101) − (cid:100) ( i − /p (cid:101) rounds in the ordinary game. The remainder of the argument proceedsjust as before. (cid:3) Theorems 4 and 17 yield the following important corollary.
Corollary 18. If p (cid:29) ln n/n , then b gp ( K n ) = (1 + o (1)) pb g ( K n ) = (1 + o (1)) pn /e . The final tool we need is the following well-known result known as the Chernoff Bound:
Theorem 19 ([12]) . Let X be a random variable that can be expressed as a sum X = (cid:80) ni =1 X i of independent random indicator variables X i , where X i is a Bernoulli random variable with Pr [ X i = 1] = p i (the p i need not be equal). For t ≥ , Pr [ X ≥ E [ X ] + t ] ≤ exp (cid:18) − t E [ X ] + t/ (cid:19) and Pr [ X ≤ E [ X ] − t ] ≤ exp (cid:18) − t E [ X ] (cid:19) . In particular, if ε ≤ / , then Pr [ | X − E [ X ] | ≥ ε E [ X ]] ≤ (cid:18) − ε E [ X ]3 (cid:19) . We are now ready to prove the main result of this section.
Proof of Theorem 5.
Corollary 18 establishes the second equality in the theorem statement,so it suffices to prove the first. Let G be any graph on n vertices. To obtain an upper boundfor b g ( G ), we provide a strategy for Min that mimics her optimal strategy for the fractionalgame on K n , which yields an upper bound on b g ( G ) in terms of b gp ( K n ). Max plays hismoves in the real game on G , but Min interprets them as moves in the imaginary fractionalgame on K n . Min plays according to an optimal strategy in the imaginary game, then makesthe same move in the real game. Our hope is that the two games behave very similarly for G ∈ G ( n, p ).As in the proof of Theorem 17, we introduce an Oracle to enforce synchronization betweenthe games. The Oracle can, at any time, clean a vertex in either game by adding extrabrushes. Suppose some vertex fires in the real game but not in the imaginary one. Whenthis happens, the Oracle cleans this vertex in the imaginary game; by Lemma 9, this cannotincrease the length of the imaginary game. Now suppose instead that some vertex v fires inthe imaginary game but not in the real one. In this case, the Oracle adds brushes to v inthe real game until it fires. By Lemma 9, each brush placed by the oracle can decrease thelength of the real game by at most 2.To obtain an upper bound on the length of the real game, we must bound the numberof brushes added by the Oracle in the real game. Let π = ( v , v , . . . , v n ) be the cleaningsequence produced during the game, that is, the order in which the vertices fire. Considerthe state of the game when only vertices v , v , . . . , v i − are clean. In the imaginary game, v i eceived ( i − p brushes from earlier neighbors and needs a total of ( n − i ) p to fire. In the realgame, it received deg − π ( v i ) = | N ( v i ) ∩ { v , v , . . . , v i − }| brushes and needs deg( v i ) − deg − π ( v i ).If the Oracle adds brushes to clean v i in the real game, then it must be that, compared tothe imaginary game, v i either received fewer brushes or needs more to fire (or both). TheOracle must add at most D π ( v i ) brushes, where D π ( v i ) = max (cid:110)(cid:16) ( i − p − ( n − i ) p (cid:17) − (cid:16) deg − π ( v i ) − (deg( v i ) − deg − π ( v i )) (cid:17) , (cid:111) = max (cid:110)(cid:16) i − p − ( n − p (cid:17) − (cid:16) − π ( v i ) − deg( v i ) (cid:17) , (cid:111) ≤ (cid:8) ( i − p − deg − π ( v i ) , (cid:9) + max { deg( v i ) − ( n − p, } . Hence the length of the game is at most b gp ( K n ) + 2 D π ( G ), where D π ( G ) = 2 n (cid:88) i =1 max (cid:8) ( i − p − deg − π ( v i ) , (cid:9) + n (cid:88) i =1 max { deg( v i ) − ( n − p, } . Of course we do not know, in advance, the cleaning sequence π . However, we still obtain thefollowing upper bound: b g ( G ) ≤ b gp ( K n ) + 2 max π D π ( G ) . A similar argument yields a lower bound on b g ( G ). We still provide a strategy for Min,but this time, she uses an optimal strategy for the real game on G to guide her play in theimaginary (fractional) game on K n . If a vertex fires in the imaginary game but not in thereal one, then the Oracle cleans that vertex by providing extra brushes in the real game;this cannot increase the length of the real game. If a vertex v fires in the real game butnot in the imaginary one, then the Oracle adds cleans v by adding brushes in the imaginarygame; each brush added can decrease the length of the imaginary game by at most 2. Whenthe Oracle adds brushes in the imaginary game, it is because more brushes were received orfewer were needed (or both) in the real game than in the imaginary game. Using symmetryand the notation introduced above, we obtain b gp ( K n ) ≤ b g ( G ) + 2 D ¯ π ( G ), where ¯ π is thereverse of π . Consequently, we get b gp ( K n ) ≤ b g ( G ) + 2 max π D ¯ π ( G ) = b g ( G ) + 2 max π D π ( G ) . It follows that | b g ( G ) − b gp ( K n ) | ≤ π D π ( G ) . Now let G ∈ G ( n, p ), let d = d ( n ) = p ( n − ω = ω ( n ) = d/ ln n (note that ω tendsto infinity as n → ∞ ). To complete the proof, it suffices to show that a.a.s. D π ( G ) = o ( pn ).We can easily bound the second summation in the definition of D π ( G ) using the ChernoffBound.Fix a vertex v i . Clearly E [deg( v i )] = p ( n −
1) = d , so it follows immediately from theChernoff Bound (Theorem 19) thatPr [ | deg( v i ) − d | ≥ εd ] ≤ (cid:18) − ε d (cid:19) = 2 n − , or ε = (cid:112) n/d = (cid:112) /ω = o (1). Hence by the Union Bound, a.a.s. for all i ∈ { , , . . . , n } we have deg( v i ) = (1 + o (1)) d . As a consequence, a.a.s. n (cid:88) i =1 max { deg( v i ) − ( n − p, } = n (cid:88) i =1 max { deg( v i ) − d, } ≤ n · o ( d ) = o ( pn ) . Estimating the first summation in the definition of D π ( G ) is slightly more complicated,since we must consider all possible cleaning sequences π . First, observe that the partialsum containing only the first n/ω / terms can be bounded (deterministically, for any π ) asfollows: n/ω / (cid:88) i =1 max (cid:8) ( i − p − deg − π ( v i ) , (cid:9) ≤ p n/ω / (cid:88) i =1 ( i −
1) = O ( pn /ω / ) = o ( pn ) . Now fix some cleaning sequence π and some i exceeding n/ω / . Clearly, E (cid:2) deg − π ( v i ) (cid:3) = p ( i − ε = ε ( n ) = √ /ω / = o (1). We call vertex v i bad if E (cid:2) deg − π ( v i ) (cid:3) − deg − π ( v i ) >ε E (cid:2) deg − π ( v i ) (cid:3) . Applying the Chernoff Bound again, the probability that v i is bad is at mostexp (cid:18) − ε p ( i − (cid:19) ≤ exp (cid:0) − ω − / p ( nω − / ) (cid:1) ≤ exp (cid:0) − ω / ln n (cid:1) =: q. The important observation is that the events { v i is bad } are mutually independent, so theprobability that there are at least n/ω / bad vertices is at most2 n q n/ω / = exp (cid:0) O ( n ) − ω / n ln n (cid:1) = exp (cid:0) − (1 + o (1)) ω / n ln n (cid:1) = o (1 /n n ) = o (1 /n !) . Hence, by the Union Bound, a.a.s. for all possible cleaning sequences π there are at most n/ω / bad vertices. It follows that a.a.s. the first summation in the definition of D π ( G ) canbe bounded as follows: n (cid:88) i =1 max (cid:8) ( i − p − deg − π ( v i ) , (cid:9) = o ( pn ) + (cid:88) i>n/ω / max (cid:8) ( i − p − deg − π ( v i ) , (cid:9) ≤ o ( pn ) + n ( εd ) + ( n/ω / )( np ) = o ( pn ) . This completes the proof. (cid:3)
References [1] N. Alon, P. Pra(cid:32)lat, and N. Wormald, Cleaning regular graphs with brushes,
SIAM Journal on DiscreteMathematics (2008), 233–250.[2] A. Bjorner, L. Lovasz, and P.W. Shor, Chip-firing games on graphs, European Journal of Combinatorics , (1991), 283–291.[3] A. Bonato, W.B. Kinnersley, and P. Pra(cid:32)lat, Toppling numbers of complete and random graphs, DiscreteMathematics and Theoretical Computer Science (2014), 229–252.[4] P. Borowiecki, D. Dereniowski, and P. Pra(cid:32)lat, Brushing with additional cleaning restrictions,
TheoreticalComputer Science (2014), 76–86.[5] B. Brˇesar, S. Klavˇzar, and D.F. Rall, Domination game and an imagination strategy,
SIAM Journal onDiscrete Mathematics (2010), 979–991.[6] D. Bryant, N. Francetic, P. Gordinowicz, D. Pike, and P. Pra(cid:32)lat, Brushing without capacity restrictions, Discrete Applied Mathematics (2014) 33–45.[7] D.W. Cranston, W.B. Kinnersley, S. O, and D.B. West, Game matching number of graphs,
DiscreteApplied Mathematics (2013), 1828–1836.
8] T. Dinski and X. Zhu, Game chromatic number of graphs,
Discrete Mathematics (1999), 109–115.[9] S. Gaspers, M.-E. Messinger, R.J. Nowakowski, and P. Pra(cid:32)lat, Clean the graph before you draw it!,
Information Processing Letters (2009), 463–467.[10] S. Gaspers, M.-E. Messinger, R.J. Nowakowski, and P. Pra(cid:32)lat, Parallel cleaning of a network withbrushes,
Discrete Applied Mathematics (2009) 467–478.[11] P. Gordinowicz, R. Nowakowski, and P. Pra(cid:32)lat, polish —Let us play the cleaning game,
TheoreticalComputer Science (2012), 123–132.[12] S. Janson, T. (cid:32)Luczak, A. Ruci´nski,
Random Graphs , Wiley, New York, 2000.[13] D. Lokshtanov, N. Misra, and S. Saurabh, Imbalance is Fixed Parameter Tractable,
Proceedings ofInternational Computing and Combinatorics Conference (COCOON 2010) .[14] C. Merino, The chip-firing game,
Discrete Mathematics (2005) 188–210.[15] M.-E. Messinger, R.J. Nowakowski, and P. Pra(cid:32)lat, Cleaning a network with brushes,
Theoretical Com-puter Science (2008), 191–205.[16] M.-E. Messinger, R.J. Nowakowski, and P. Pra(cid:32)lat, Cleaning with Brooms,
Graphs and Combinatorics (2011), 251–267.[17] P. Pra(cid:32)lat, Cleaning random d –regular graphs with Brooms, Graphs and Combinatorics (2011), 567–584.[18] P. Pra(cid:32)lat, Cleaning random graphs with brushes, Australasian Journal of Combinatorics (2009),237–251.[19] D.B. West, Introduction to Graph Theory, 2nd edition , Prentice Hall, 2001.
Department of Mathematics, University of Rhode Island, Kingston, RI, USA, 02881
E-mail address : [email protected] Department of Mathematics, Ryerson University, Toronto, ON, Canada, M5B 2K3
E-mail address : [email protected]@ryerson.ca