16 views

Uploaded by Marina Kamalova

- Ruler Folding problem Paper
- Games Puzzles and Computation
- Complex problem solving field definition
- 02_The Role of Algorithms in Computing
- problemset3sol[1]
- Operations Research
- The Computer Science Handbook
- Daa Notes Full
- caim
- tocmod5
- Guide 1
- GATE Theory of Computation Book
- boca
- In Tract Ability
- Randomized Algorithms for 3-Sat
- Belyaev
- Linux Cpu Scheduler
- Amortized Analysis
- sigmod16a
- DETERMINAN 1

You are on page 1of 33

Jarkko Kari , Mikhail Volkov

Department of Mathematics FI-20014 University of Turku Turku, Finland

Department of Mathematics and Mechanics 620083 Ural State University Ekaterinburg, Russia email: jkari@utu.fi,Mikhail.Volkov@usu.ru

3

May 7, 2011 2 h 44

chapterKV

4 5 6

2010 Mathematics Subject Classication: 68Q45 68R10 y Key words: Finite automata, Synchronizing automata, Reset words, Cern s conjecture, Road Coloring Problem

Contents

1 Synchronizing automata, their origins and importance 2 Algorithmic and complexity issues y 3 Around the Cern conjecture 4 The Road Coloring Problem 5 Related work References Index 1 5 10 19 27 27 32

8 9 10 11 12 13 14

15

A complete deterministic nite automaton (DFA) (here and below stands for the state set and for the input alphabet) is called synchronizing if there exists a word whose action resets , that is, leaves the automaton in one particular state no matter at which state in it is applied: for all . Any word with this property is said to be a reset word for the automaton.

c:motivation

19 20

18

17

16

J. Kari, M. Volkov

KV:fig:C4

23 24 25 26

30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51

and below we adopt the convention that edges bearing multiple labels represent bunches of parallel KV:fig:C4 in Figure 1 represents the two parallel edges and . edges. In particular, the edge 2 After [35], the name homing sequence has become standard for the notion.

@ CA 5 6 C A 5

6A @ C9BA8 75

1 Here

22

21

Figure 1 shows a synchronizing automaton with 4 states1 denoted by . The reader can easily verify that the word resets the automaton leaving it in the state 1. With somewhat more effort one can also check that is the shortest reset word KV:fig:C4 y for . The example in Figure 1 is due to Cern , a Slovak computer scientist, in whose pioneering paper [18] the notion of a synchronizing automaton explicitly appeared for y the rst time. (Cern called such automata directable. The word synchronizing in this context was probably introduced by Hennie [39].) Implicitly, however, this concept has been around since the earliest days of automata theory. The very rst synchronizing automaton that we were able to trace back in the literature appeared in Ashbys classic book [5, pp. 6061], see [81, Section 1] for a discussion. In [18] the notion of a synchronizing automaton arose within the classic framework of Moores Gedanken-experiments [49]. For Moore and his followers nite automata served as a mathematical model of devices working in discrete mode, such as computers or relay control systems. This leads to the following natural problem: how can we restore control over such a device if we do not know its current state but can observe outputs produced by the device under various actions? Moore [49] has shown that under certain conditions one can uniquely determine the state at which the automaton arrives after a suitable sequence of actions (called an experiment). Moores experiments were adaptive, that is, each next action was selected on the basis of the outputs caused by the previous actions. Ginsburg [36] considered more restricted experiments that he called uniform. A uniform experiment2 is just a xed sequence of actions, that is, a word over the input alphabet; thus, in Ginsburgs experiments outputs were only used for calculating the resulting state at the end of an experiment. From this, just one further step was needed to come to the setting in which outputs were not used at all. It should be noted that this setting is by no means articialthere exist many practical situations when it is technically impossible to observe output signals. (Think of a satellite which loops around the Moon and cannot be controlled from the Earth while behind the Moon.) The original Gedanken-experiments motivation for studying synchronizing automata is still of importance, and reset words are frequently applied in model-based testing of reactive systems. See [21, 13] as typical samples of technical contributions to the area and [74] for a recent survey.

10

KV:fig:C4

1 20

' ' (

)( ' (

( ' 10

52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75

Another strong motivation comes from the coding theory. We refer to [11, Chapters 3 and 10] for a detailed account of profound connections between codes and automata; here we restrict ourselves to a special (but still very important) case of maximal prex codes. Recall that a prex code over a nite alphabet is a set of words in such that no word of is a prex of another word of . A prex code is maximal if it is not contained in another prex code over the same alphabet. A maximal prex code over is synchronized if there is a word such that for any word , one has . Such a word is called a synchronizing word for . The advantage of synchronized codes is that they are able to recover after a loss of synchronization between the decoder and the coder caused by channel errors: in the case of such a loss, it sufces to transmit a synchronizing word and the following symbols will be decoded correctly. Moreover, since the probability that a word contains a xed factor tends to 1 as the length of increases, synchronized codes eventually resynchronize by themselves, after sufciently many symbols being sent. (As shown in [14], the latter property in fact characterizes synchronized codes.) The following simple example illustrates these ideas: let and . Then is a maximal prex code and one can easily check that each of the words 010, 011110, 011111110, . . . is a synchronizing word for . For instance, if the code word 000 has been sent but, due to a channel error, the word 100 has been received, the decoder interprets 10 as a code word, and thus, loses synchronization. However, with a high probability this synchronization loss only propagates for a short while; in particular, the decoder denitely resynchronizes as soon as it encounters one of the segments 010, 011110, 011111110, KV:fig:decoding . . . in the received stream of symbols. A few samples of such streams are shown in Figure 2 in which vertical lines show the partition of each stream into code words and the boldfaced code words indicate the position at which the decoder resynchronizes. Sent Received Sent Received Sent Received

fig:decoding

76 77 78 79

Figure 2. Restoring synchronization If is a nite prex code over an alphabet , then its decoding can be implemented by a deterministic automaton that is dened as follows. Let be the set of all proper prexes of the words in (including the empty word ). For and , dene if if is a proper prex of a word of

80 81 82 83

The resulting automaton is complete whenever the code is maximal and it is easy to see that is a synchronizing automaton if and only if is a synchronized code. Moreover, a word is synchronizing for if and only if is a reset word for KV:fig:huffman and sends all states in to the state . Figure 3 illustrates this construction for

UVTSR9GI QP H

F D

E

'

TqY2xw&Ebts%bb%reS qqc g hc S S S Qc Q Q Qc Q TqY2xw&vtb2bb%bbQ qqc g hc S S Sc Q Q Qc Q Q TqY2cxwh2yQedb%creb2bbavtb2c tsb%reS qq g hc S Q Q Q S Sc Q Q Qc S S S S S Q Qc Q TqqY2xgw2cre2cbQb%cvbsbauet2Etbs%bbQ qc h h QS Q Q S S Q Qc QS Sc S S S Qc Q Q q qc h h h gc QSc Q Q Qc Q YTq2p#iare2bb%reS q qc h h h gc QS Q Qc Q Q YTq2p#iafedb%GbbQ

E G D D F D

D F

H

D i

4

84

J. Kari, M. Volkov

:fig:huffman

85 86 87 88 89 90 91 92 93 94 95

Figure 3. A synchronized code (on the left) and its automaton (on the right) Thus, (to be continued and supplied by some historical references). An additional source of problems related to synchronizing automata has come from robotics or, more precisely, from part handling problems in industrial automation such as part feeding, xturing, loading, assembly and packing. Within this framework, the concept of a synchronizing automaton was again rediscovered in the mid-1980s by Natarajan [50, 51] who showed how synchronizing automata can be used to design sensor-free orienters for polygonal parts, see [81, Section 1] for a transparent example illustrating Natarajans approach in a nutshell. Since the 1990s synchronizing automata usage in the area of robotic manipulation has grown into a prolic research direction but it is fair to say that publications in this area deal mostly with implementation technicalities. However, amongst them there are papers of signicant theoretical importance such as [27, 37, 20]. Recently, it has been realized that a notion that arose in studying of substitution systems is also closely related to synchronizing automata. A substitution on a nite alphabet is a map ; the substitution is said to be of constant length if all words , , have the same length. One says that satises the coincidence condition if there exist positive integers and such that all words have the same letter in the -th position. For an example, consider the substitution on dened KV:fig:substituion by . Calculating the iterations of up to (see Figure 4), we observe that satises the coincidence condition (with , ).

:substituion

104 105 106 107 108 109

Figure 4. A substitution satisfying the coincidence condition The importance of the coincidence condition comes from the crucial fact (established by Dekking [24]) that it is this condition that completely characterizes the constant length substitutions which give rise to dynamical systems measure-theoretically isomorphic to a translation on a compact Abelian group, see [61, Chapter 7] for a survey. For us, however, the coincidence condition is primarily interesting as yet another incarnation of synchro-

wzzywzzyzzyyyw aaXVV9X2X9WVXVz zyzyyywzywzwzz 92VWXaRaaVy yywzwzzyywzwzz V&VWVV9WRao aVy s s &It t s { x xwv t k hq 2)zVTyR9u ljRr

&VVRaz |VRaz | Vz |z | zyzyyyw yyw w | y w z w z z | z z WyXVWaW9y waW9y |y |y z | w z z y w z | z y | p VVXaWz9y 9Xz9y yWy | s w w | z x z x y | az~i&9y}| ywVyV}yo w m k h nj lji

f ged

96 97

KV:fig:C3

118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137

substitutions satisfying the coincidence condition correspond precisely to synchronizing automata, and moreover, given a substitution, the number of iterations at which the coincidence rst occurs is equal to the minimum length of reset word for the corresponding automaton. We mention in passing a purely algebraic framework within which synchronizing au- If space tomata also appear in a natural way. One may treat DFAs as unary algebras since each permits!! letter of the input alphabet denes a unary operation on the state set. A term in the , where is a variable language of such unary algebras is an expression of the form and is a word over an alphabet . An identity is a formal equality between two terms. A DFA satises an identity , where the words involved in the terms and are over , if and take the same value under each interpretation of their variables in the set . Identities of unary algebras can be of the from either (homotypical identities) or with (heterotypical identities). It is easy to realize that a DFA is synchronizing if and only if it satises a heterotypical identity, and thus, studying synchronizing automata may be considered as a part of the equational logic of unary algebras. In particular, synchronizing automata over a xed alphabet form a pseudovariety of unary algebras. See [12] for a survey of numerous publications in this direction; it is fair to say, however, that so far this algebraic approach has not proved to be really useful for understanding the combinatorial nature of synchronizing automata.

138

It should be clear that not every DFA is synchronizing. Therefore, the very rst question that we should address is the following one: given an automaton , how to determine whether or not is synchronizing?

s&complexity

141

140

139

Q S S S S Vv&bWVQ

" vF

X'

2' X'

'

2'

H " Y" F

X'

111

D " 2' xF D D F F D D l Q S S S S WiRaiX9ei" VVerQ d q q " v W' W' ! YYTql X&' ' #! ! U YTqTqY G ` q P

0 1

110

nizability. Indeed, there is a straightforward bijection between DFAs and constant length substitutions. Each DFA with denes a length substitution on that maps every to the word . (For instance, the KV:fig:C4 in Figure 1 induces the substitution .) automaton Conversely, each substitution such that all words , , have the same length gives rise to a DFA for which serves as the state set and which has input letters , say, acting on as follows: is the symbol in the -th position of the word . (For instance, the substitution considered in the previous paragraph KV:fig:C3 denes the automaton shown in Figure 5.) It is clear that under the described bijection

0123

123

03

23

er automaton

142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158

This question is in fact quite easy, and the most straightforward solution to it can be achieved via the classic subset construction by Rabin and Scott [62]. Given a DFA , we dene its subset automaton on the set of the non-empty subsets of by setting for each non-empty subset of and each . (Since we start with a deterministic automaton, we do not need adding the empty set to KV:fig:power automaton the state set of .) Figure 6 presents the subset automaton for the DFA shown in KV:fig:C4 Figure 1. Now it is obvious that a word is a reset word for the DFA if and only if labels a pathKV:fig:power automaton ending at a singleton. (For instance, the bold in starting at and path in Figure 6 represents the shortest reset word of the automaton .) Thus, the question of whether or not a given DFA is synchronizing reduces to the following reachability question in the underlying graph3 of the subset automaton : is there a path from to a singleton? The latter question can be easily answered by breadth-rst search, see, e.g., [22, Section 22.2]. The described procedure is conceptually very simple but rather inefcient because the power automaton is exponentially larger than . However, the following criterion of synchronizability gives rise to a polynomial algorithm. Proposition 2.1 ([18, Theorem 2]). A DFA for every there exists a word is synchronizing if and only if . and , then

op:quadratic 159

160

such that

161 162

Proof. Of course, only sufciency needs a proof. For this, take two states consider a word such that . Then . If

a graph we mean a quadruple of sets and maps: the set of vertices , the set of edges , a map that maps every edge to its tail vertex, and a map that maps every edge to its head vertex. Notice that in a graph, there may be several edges with the same tail and head. (Thus, our graphs are in fact directed multigraphs but since no other graph species show up in this chapter, we use a short name.) We assume the readers acquaintance with basic notions of graph theory such as path, cycle, etc. The underlying graph of an automaton is the graph obtained from by forgetting edge labels.

C b

3 By

'

C "$ " S GV$ c " c 2c " c X%! d! c c %! ! "$ R%! " ! E u

1 &0

10

1 &0

'

'

' (

y

( ' (

01

12

02

13

'

'

012

013

( (

' '

'

'

'

J. Kari, M. Volkov

023

163 164 165 166

is a reset word and is synchronizing. If , take two states and consider a word such that . Then . If , then is a reset word; otherwise we repeat the process. Clearly, a steps. reset word for will be constructed in at most One can treat Proposition 2.1 as a reduction of the synchronizability problem to a reachability problem in the subautomaton of whose states are couples (2element subsets) and singletons of . Since the subautomaton has time. This complexity bound breadth-rst search solves this problem in assumes that no reset word is explicitly calculated. If one requires that, whenever turns out to be synchronizing, a reset word is produced, then the best of the known algorithms (which is basically due to Eppstein [27, Theorem 6], see also [74, Theorem 1.15]) has an implementation that consumes time and working space, not counting the space for the output which is . For a synchronizing automaton, the subset automaton can be used to construct shortest reset words as they correspond to shortest paths from the whole state set to a singleton. Of course, this requires exponential (of ) time in the worst case. Nevertheless, there were attempts to implement this approach, see, e.g., [64, 79]. One may hope that, may yield a as above, a suitable calculation in the polynomial subautomaton polynomial algorithm. However, it is not the case, and moreover, as we will see, it is very unlikely that any reasonable algorithm may exist for nding shortest reset words in general synchronizing automata. In the following discussion we assume the readers acquaintance with some basics of computational complexity (such as the denitions of the complexity classes NP and coNP) that can be found, e.g., in [32, 54]. Consider the following decision problem: S HORT-R ESET-W ORD : Given a synchronizing automaton it true that has a reset word of length ? and a positive integer , is

KV:prop:quadratic

167 168

169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194

states,

Clearly, S HORT-R ESET-W ORD belongs to the complexity class NP: one can nondeterministically guess a word of length and then check if is a reset word for in time . Several authors [68, 27, 38, 72, 73] have proved that S HORT-R ESETW ORD is NP-hard by a polynomial reduction from SAT (the satisability problem for a system of clauses, that is, disjunctions of literals). We reproduce here Eppsteins reduction from [27]. Given an arbitrary instance of SAT with variables and clauses , we construct a DFA with 2 input letters and as follows. The state set of consists of states , , , and a special state . The transitions are dened by if the literal occurs in otherwise

for for

, ,

; ;

occurs in

for

S c a c " c " " e$ " " c $ uc c 2c " S c e vS s( E " rES vS rES vS S ES vS ! F YTqY(q Tq F ' S c c 9Y `c c u c " `c c c r

r c c c 93 `c " u c " `c c c c c y c r3 `c c c c c

c c

Sc " c

J. Kari, M. Volkov

in

g:A2_example

KV:fig:A2_example

and

196 197 198 199 200 201 202 203 204 205 206 207 208

:complexity1 209

210 211 212 213

Proposition 2.2. The problem S HORT-R ESET-W ORD is NP-complete. In fact, as observed by Samotij [73], the above construction yields slightly more4 . Consider the following decision problem: S HORTEST-R ESET-W ORD : Given a synchronizing automaton and a positive integer , is it true that the minimum length of a reset word for is equal to ?

4 Actually,

the reduction in [73] is not correct but the result claimed can be easily recovered as shown below.

195

If at some state in Figure 7 there is no outgoing edge labelled , the edge is assumed (those edges are omitted to improve readability). The two instances differ only in the rst clause: in it contains the literal while in it does not. Correspondingly, the automata and differ only by the outgoing edge labelled at the state : in it leads to (and therefore, it is not shown) while in it leads to the state and is shown by the dashed line. Observe that is satisable for the truth assignment , while is not satisable. It is not hard to check that the word resets while is reset by no word of length 3 but by every word of length 4. In general, it is easy to see that is reset by every word of length and is reset by a word of length if and only if is satisable. Therefore assigning the instance of S HORT-R ESET-W ORD to an arbitrary -variable instance of SAT, one obtains a polynomial reduction of the latter problem to the former. Since SAT is NPcomplete and S HORT-R ESET-W ORD lies in NP, we obtain the following.

KV:fig:A2_example

1 1! 1 3 ! 1 !

( ' 3F

3 T1! 3 3 ! p ! !

' ' F ( (

214 215 216 217 218 219 220 221 222 223 224 225 226 227 228

Assigning the instance of S HORTEST-R ESET-W ORD to an arbitrary system of clauses on variables, one sees that the answer to the instance is Yes if and only if is not satisable. Thus, we have a polynomial reduction from the negation of SAT to S HORTEST-R ESET-W ORD whence the latter problem is coNP-hard. As a corollary, S HORTEST-R ESET-W ORD cannot belong to NP unless NP = coNP which is commonly considered to be very unlikely. In other words, even non-deterministic algorithms cannot decide the reset threshold of a given synchronizing automaton, (that is, the minimum length of its reset words) in polynomial time. The exact complexity of the problem S HORTEST-R ESET-W ORD has been recently determined by Gawrychowski [33] and, independently, by Olschewski and Ummels [53]. It turns out that the appropriate complexity class is DP (Difference Polynomial-Time) introduced by Papadimitriou and Yannakakis [55]; this class consists of languages of the form where is a language from NP and a is a language in coNP. A standard DP-complete problem is SAT-UNSAT whose instance is a pair of clause systems , say, and whose question is whether is satisable and is unsatisable. Proposition 2.3. The problem S HORTEST-R ESET-W ORD is DP-complete. Proposition 2.3 follows from mutual reductions between S HORTEST-R ESET-W ORD and SAT-UNSAT obtained in [33, 53]. consists of all problems solvable by a deterministic The complexity class polynomial-time Turing machine that has an access to an oracle for an NP-complete problem, with the number of queries being logarithmic in the size of the input. The class DP is contained in (in fact, for every problem in DP two oracle queries sufce) and the inclusion is believed to be strict. Olschewski and Ummels [53] have shown that the problem of computing the reset threshold (as opposed to deciding whether it is equal to a given integer) is complete for the functional analogue of the class (see [75] for a discussion of functional complexity classes). Hence, this problem appears to be even harder than deciding the reset threshold. Recently Berlinkov [9] has shown (assuming P NP) that no polynomial algorithm can approximate within a constant factor the reset threshold of a given synchronizing automaton with two input letters. The problem of nding a reset word of minimum length (as opposed to computing only the length without writing down the word itself) may be even more difcult. From the cited result of [53] it follows that the problem is -hard but its exact complexity is not known yet. The hardness results in [9, 53] areKV:prop:complexity1 obtained via suitable encodings of SAT in the avor of the above proof of Proposition 2.2. Gerbush and Heeringa [34] have observed that some other well-known hard problems such as SCS (S HORTEST C OMMON S UPERSE QUENCE ) or S ET C OVER admit a transparent reduction to the problem of nding a reset word of minimum length for a given synchronizing automaton. In particular, since SCS is known to have no approximation within a constant factor unless P = NP [40], they have deduced a similar conclusion for approximating the reset threshold butin contrast to the cited result of [9]without any bound on the size of the input alphabet. Moreover, using a recent result on S ET C OVER [3], they have concluded that the reset threshold of synchronizing automata with states and unbounded alphabet cannot be approximated within the factor for some constant unless P = NP. It is a challenging problem to

KV:prop:complexity2

:complexity2 229

230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257

%2r

a2 R

a2

S x ar

a2

` s

Wl

10

258 259 260 261 262 263 264 265

J. Kari, M. Volkov

study approximation of the reset threshold within a logarithmic factor for synchronizing automata with a xed alphabet size. We mention that Pixley, Jeong and Hachtel [60] suggested an heuristic polynomial algorithm for nding short reset words in synchronizing automata that was reported to perform rather satisfactory on a number of benchmarks from [83]; further polynomial algorithms yielding short (though not necessarily shortest) reset words have been implemented by Trahtman [79] and Roman [66, 65]. Some algorithms for nding reset words will be also discussed in the next section.

266

The Cern conjecture. A very natural question to ask is the following: given a posy itive integer , how long can be reset words for synchronizing automata with states? y Cern [18] found a lower bound by constructing, for each , a synchronizing automwith states and 2 input letters whose shortest reset word has length . aton We assume that the state set of is and the input letters are and , subject to the following action on : if if

y conjecture

267 268 269 270 271 272

273 274

Our rst example of synchronizing automaton (see Figure 1) is, in fact, KV:fig:cerny-n automaton is shown in Figure 8 on the left.

:fig:cerny-n

275 276

The series was rediscovered many times (see, e.g., [48, 28, 27, 30]). It is easy to see that the word of length resets . Proposition 3.1 ([18, Lemma 1]). Any reset word for

V:prop:cerny 277

278 279 280 281

There are several nice proofs for this result. Here we present a recent proof from [4]; y it is based on a transparent idea and reveals an interesting connection between Cern s automata KV:sec:rcp and an extremal series of graphs discovered in Wielandts classic paper [82] (see Section 4).

&0 S 9 &0 #V S S

('

(

TqYq v q S

(

'

10

KV:fig:C4

. A generic

' S

.

(

Q " S " XQ 4q & s( X S ' e US q SQP V 4TqYTq&)VYX9 &0 S '

)( '

X0

282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299

11

Proof of Proposition 3.1. Let be a reset word of minimum length for . Since the letter acts on as a cyclic permutation, the word cannot end with . (Otherwise removing the last letter gives a shorter reset word.) Thus, for some is precisely the set . such that the image of under the action of Since the letter xes each state in its image , every occurrence of in except the last one is followed by an occurrence of . (Otherwise occurs in as a factor and reducing this factor to just results in a shorter reset word.) Therefore, if we let , then the word can be rewritten into a word over the alphabet . The actions of and induce a new DFA on the state set ; we denote this induced DFA KV:fig:cerny-n (shown in Figure 8 on the right) by . Since and act on in the same way, the word is a reset word for and brings the automaton to the state 2. If , the word also is a reset word for and it also brings the automaton to 2. Hence, for every , there is a path of length in from any given state to 2. In particular, setting , we conclude that for every there is a cycle of length in . The underlying graph of has simple cycles only of two lengths: and . Each cycle of must consist of simple cycles of these two lengths whence must be expressible as a non-negative integer combination of and each number . Here we invoke the following well-known and elementary result from arithmetics: Lemma 3.2 ([63, Theorem 2.1.1]). If are relatively prime positive integers, then is the largest integer that is not expressible as a non-negative integer combination of and . . Suppose Lemma 3.2 implies that that . Then there should be a path of this length from the state 1 to the state 2. Every outgoing edge of 1 leads to 2, and thus, in the path it must be followed by KV:lemma:sylvester a cycle of length . No cycle of such length may exist by Lemma 3.2. Hence . Since the action of on any set of states cannot change the cardinality of and the action of can decrease the cardinality by 1 at most, the word must contain at least occurrences of . Hence the length of over is at least and contain at least occurrences of . Since each occurrence of in corresponds to an occurrence of the factor in , we conclude that the length of over is at least . Thus, . y If we dene the Cern function as the maximum length of shortest reset words for synchronizing automata with states, the above property of the series , , y yields the inequality . The Cern conjecture is the claim that the equality holds true. y y In the literature, one often refers to Cern s paper [18] as the source of the Cern y only conjecture. In fact, the conjecture was not yet formulated in that paper. There Cern observed that and concluded the paper with the following remark: The difference between the bounds increases rapidly and it is necessary to sharpen them. One can expect an improvement mainly for the upper bound. The conjecture in its present-day form was formulated a bit later, after the expectation in the above quotation was conrmed by [76]. (Namely, Starke improved the upper bound

KV:lemma:sylvester

ma:sylvester 300

301 302

303 304 305 306 307 308 309 310 311 312 313

314 315 316 317 318 319 320 321 322 323 324 325

S w S S 9 b S 9g TqYTqX&I U 0 P q S S c 9 c ' $ c c $ ' v dn U 9( ' P H $ ( H ( H n S H U %P ( c H c S I c H c S n I I 9 c H c S S S E c c n S n c H c i G c H H c ( aU %P H H $ ( ' H $ U %P ( ( ' ( ' ' ' US q S a 4TqYTqXVP $ USQ aYX9P ' $ U %9)( ' P $ (

X(0

KV:prop:cerny

12

326 327 328

J. Kari, M. Volkov

from [18] to , which was the rst polynomial upper bound for .) y in his talks in the second half of Cern explicitly stated the conjecture the 1960s; in print the conjecture rst appeared in [19]. y An upper bound. The best upper bound for the Cern function achieved so far guarantees that for every synchronizing automaton with states there exists a reset word of length . Such a reset word arises as the output of the following greedy algorithm. G REEDY C OMPRESSION

3: 4: 5: 6: 7: 8: 9: 10:

332 333 334 335 336 337 338 339

If , then clearly the main loop of Algorithm 1 is executed at most times. Finding the word in line 7 amounts to reading the labels along a shortest path between a couple contained in and a singleton in the automaton (see the discussion after KV:prop:quadratic KV:Greedy Proposition 2.1). Breadth-rst search does this in time. Thus, Algorithm 1 is polynomial in the size of . In order to evaluate the length of the output word , we estimate the length of each word produced by the main loop. Consider a generic step at which and let with , . Then each of the sets

KV:Greedy

340 341

contains exactly states. Furthermore, since , there exist two distinct states such that . Now dene couples , , such that , for .

onfiguration

342

' W' YTqq ' ' q V' $ W' YTqq ' $ ' $ q S TTqTqYV $ ' " $ ' " qS $ U P W' " W' " $ W' " c uc c c W' " TTqYt X' " qq 2' ' T""Y" ' H S c c H y c " c r

S

KV:Greedy

while do if for all then return Failure else take a word of minimum length with Updating the current word Updating the current set return

c G2c YI c c H"

H

c cVc TI c " S 2c c

1: 2:

KV:Greedy

S 9 d

` 2 22 S H

qS $ TTqTqYVG

H Y"I H

W c c

TTqTqYVG qS

343 344 345 346 347 348 349 350 351

13

Then the condition that is a word of minimum length with implies that for . Indeed, if for some , then already the word of length would satisfy contradicting the choice of . Thus, we arrive at a problem from combinatorics of nite sets that can be stated as follows. Let . A sequence of -element subsets of an -element set is called 2-renewing if each contains a couple such that for each . What is the maximum length of a 2-renewing sequence as a function of and ? The problem was solved by Frankl [29] who proved the following result5 . Proposition 3.3. The maximum length of a -renewing sequence of -element subsets in an -element set is equal to . Thus, if is the length of the word that Algorithm 1 appends to the current word after the iteration KV:prop:frankl step that the algorithm enters while the current set contains states, then Proposition 3.3 guarantees that . Summing up all these inequalities from to , one arrives at the aforementioned bound

KV:Greedy

:prop:frankl 352

353

358 359 360 361 362 363 364 365 366 367 368 369 370

In the literature the bound (3.1) is usually attributed to Pin who explained the above conKV:Greedy nection between Algorithm 1 and the combinatorial problem on the maximum length of 2-renewing sequences and conjectured the estimation for this length in his talk at the Colloquium on Graph Theory and Combinatorics held in Marseille in 1981. (Frankl learned this conjecture from Pinand proved itduring another colloquium on combiKV:eq:pin natorics held in Bielefeld in November 1981.) Accordingly, the usual reference for (3.1) is the paperKV:eq:pin on the talk. The full story is however more complicated. Actually, [59] based the bound (3.1) rst appeared in [28] where it was deduced from a combinatorial conjecKV:eq:pin ture equivalent to Pins one. The conjecture however remained unproved. The bound (3.1) then reoccurredKV:eq:pin but the argument justifying it in these papers was insufcient. in [46, 47] KV:prop:frankl In 1987 both (3.1) and Proposition 3.3 were independently rediscovered by Klyachko, Rystsov and Spivak [45] who were aware of [28, 46, 47] but neither [59] nor [29]. We If space include here a proof of Frankls result following [45]. permits!!

2

KV:eq:pin

Proof of Proposition 3.3. Let . First, we exhibit a 2-renewing sequence of -element subsets in of length . For this put , list all couples of in some order and let be the union of with the -th couple is 2-renewing. in the list. Clearly, the sequence Now we assign to each -element subset

KV:prop:frankl

of

5 Actually Frankl [29] considered and solved a more general problem concerning the maximum length of (analogously dened) -renewing sequences of -element subsets in an -element set for any xed .

6 8 @95

U 43TYTqTq & 20)('1 &$# " q 3 P 1 % q " " TTqYTq ! 2 U q SP 2 TqYTqWW U 4Yq Tq TqXWWG SP

q 3 w 2

354

" " " c c2c W' "TY" ' ' T"Y" X' c c c c % c H"

(3.1) KV:eq:pin

14

A

376

J. Kari, M. Volkov

377 378 379 380 381 382 383 384 385 386

387 388 389 390 391 392 393 394 395 396 397 398 399 400

401 402

qQ 3P RbU &# $ A b3 9v e S 3 v p q q q QTTYT Q 0 e p p p n Q G v B I FF ' B F ' F 3 9 3 YT"" 3 yS b3 FF " 3 F F F F F B 1 SF F 1B F q FF v " 3 9 YT"" 3 F F F C B C B F 3 " 3 3wS b3 F 3 9 3 "YT"" 3 wS b3 FF F x x UB F F UB F 3 F 3 9e 3 YT"" e 3 e wS e b3 FF Ue3P %fi# 3 " Q9u3 %fb3 h E $ 3 9 F 3 9 s YT"" F t F swrq F S e p p p e p Ue3 afiP g e f3 bc c " A E A q S & TqYTqV Ec " c 1 d 1 ' % 20)(&b# " A YTqYTq4 " A q A U 43TqYTq &P q 3 1 1 (D)(c# " TTqTqY " 1 ' % 1 ' % q 20)(&c# " A YTqTqY4 " A q 2 E A A TYTqY4 A $ qq F $ 3 3 FF F b A Y a F 3 FF F 3 Y V n`W A A q A A TYTqY4 A qq $ TYTqTqW B B n Q XF $ V UTSF R F 3 3 Q3W3R 3 U $ 43V &3 P A U $ Q3 W&P A TqYTqTq A A YTqTqY4 A q TTqYTq q PI2 ' B F 'B F F 3 9 3 T"T"" 3 F F A q FFF 1 B F 1 F 9 T"T"" B F F C B 33 9 33 T"T"" 33 F C B DF F

. . . . . . . . . .. . . . . . . . . . . -determinant , whence this value is not 0. Hence If If all . . . . . . . . . . . . .. . . . . . . . . . . Clearly, cients

Observe that for any 2-renewing sequence of -element subsets in , the polynomials are linearly independent. Indeed, if they were linearly dependent, then by a basic lemma of linear algebra, some polynomial should be expressible as a linear combination of the preceding polynomials . By the denition of a 2-renewing sequence, contains a couple such that . If we substitute , and for in each for all polynomial , then the polynomials vanish (since the two last columns in each of the resulting determinants become proportional) and so does any linear combination of the polynomials. The value of however is the determinant being the product of a Vandermonde -determinant with the cannot be equal

to a linear combination of . We see that the length of any 2-renewing sequence cannot exceed the dimension of the linear space over spanned by all polynomials of the form . In order to prove that the dimension is at most , it sufces to show that the space is spanned by the polynomials , where is the 2-renewing sequence constructed in the rst paragraph of the proof. For this, take an arbitrary -element subset of . We claim that the polynomial is a linear combination of . We induct on the cardinality of the set . , then is the union of with some couple from , whence for some . Thus, and our claim holds true. , there is . Let . There exists a polynomial over such that and for . Consider the determinant

as the rst column is the sum of the next columns with the coef. Thus, expanding by the rst column gives the identity

B '

nomial

in variables

of reals:

403

15

Since

404 405

whence mials

and by the inductive assumption, the polynoKV:eq:combination are linear combinations of . From (3.2) we as well.

KV:Greedy

409 410 411 412 413 414 415 416 417 418 419 420

y If one executes Algorithm 1 on the Cern automaton (Figure 6 is quite helpful here), one sees that the algorithm returns the word of length 10 which is not the shortest reset word for . This reveals one of the main intrinsic difculties of the synchronization problem: the standard optimality principle does not apply here since it is not true that the optimal solution behaves optimally also in all intermediate steps. In our KV:Greedy example, the optimal solution is the word but it cannot be found by Algorithm 1 rather than on the second execution of because the algorithm chooses the main loop. Actually, the gap between the reset threshold of a synchronizing automKV:Greedy aton and the length of the reset word that Algorithm 1 returns on the automaton may be y arbitrarily large6 : one can calculate that for the Cern automaton whose reset threshold KV:Greedy is , Algorithm 1 produces a reset word of length . The behaviour of KV:Greedy Algorithm 1 on average is not yet understood; practically it behaves rather well. The extension algorithm. While studying Algorithm 1 has provided the best currently y known upper bound for the Cern function in the general case, the most impressive partial y results proving the Cern conjecture for some special classes of automata have been obtained via analysis a different algorithm. This algorithm also operates in a greedy manner but builds a reset word in the opposite direction. For a DFA , a subset and a word , we denote by the full pre-image of under the action of , that is, . In what follows, we denote the same a singleton set and its single elementKV:Extension to lighten notation. KV:Greedy In contrast to Algorithm 1, it is not clear whether Algorithm 2 admits a polynomialtime implementation. Moreover, in general we know no non-trivial bound on the length KV:Extension of the words that the main loop of Algorithm 2 appends to the current word. However, one can isolate some cases in which rather strong bounds on do exist. The following denition is convenient for subsequent discussion. Given a number , a DFA is said to be -extensible if for each proper non-singleton subset , there exists a word of length at most such that . The following observation explains the importance of this property.

KV:sec:algorithms&complexity from the non-approximation results discussed in Section 2 KV:Greedy because Algorithm 1 is not really deterministic. Indeed, in general there may be several words satisfying the conditions in line 7 of the algorithm and it has not been specied which one of the words should be taken.

6 We observe that this does not immediately follow

KV:fig:power automaton

KV:Greedy

421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436

c cuwc c d QGIp H cc

3P U &#r

" c U !

43' H ' ' ' ' H ( ( 43( 43( 1 &0 ' 1&430 ( ' ( ' ( ' A 1 1 ' (D)(%c# " A TYTqqY4 " A q 3P $ %U & i A S Vc c Rc U iE $ c 3P bU iP V rae&P g Gb%U &P $ %U id $ 3 g U 3 3 g 3P @ 3 3 U e iPb 3 Ib9u3 3 Q 4%U &P# $ A b3 9Sv b A 3 %fi# i Gt9fu3 Ue3P $ S e

and , the identity rewrites as for all . For each such

(3.2) KV:eq:combination

W 0

!P G

c c wp

`G

E p

S 9

`r

16

J. Kari, M. Volkov

439

times and Proof. If we run Algorithm 2 on , the main loop is executed at most each word that it appends to the current word has length at most . Hence the length of the reset word returned by the algorithm does not exceed . If , then y which complies with the Cern conjecture. we get the bound y The approach to the Cern conjecture via extensibility traces back to Pins paper [58] such that is prime and some letter of 1978. Pin observed that every DFA acts as a cyclic permutation of is 1-extensible provided some other letter acts on as a non-permutation. Thus, such is synchronizing and its reset threshold does not exceed . 20 years later Dubuc [25] generalized Pins result by showing that every synchronizing automata in which some letter acts as a cyclic permutation of the state set is 1-extensible. Kari [44] proved 1-extensibility of Eulerian7 synchronizing automata. In all these papers 1-extensibility is obtained via linear-algebraic arguments; we include here a proof from [44] as quite a representative example of these linearization techniques. Theorem 3.5 ([44, Theorem 2]). If a synchronizing automaton then it has a reset word of length at most , where is Eulerian, .

KV:Extension

thm:eulerian 453

454

Proof. For every vertex in an Eulerian graph, its in-degree and its out-degree are equal. In the underlying graph of a DFA the out-degree of every vertex is equal to the cardinality of the input alphabet. Hence, if , then each vertex in the underlying graph of

7 A graph is strongly connected if for every pair of its vertices, there exists a path from one to the other. A graph is Eulerian if it is strongly connected and each of its vertices serves as the tail and as the head for the same number of edges. A DFA is said to be Eulerian if so is its underlying graph. More generally, we freely transfer graph notions (such strong connectivity) from graphs to automata they underlie.

445

c c S S r9 W `

c p c

444

Sp V swS p p

`G S 7 V S

ly c c

438

xtensibility 437

Proposition 3.4. If is an -extensible automaton with states, then is synchronizing y and the reset threshold of is at most . In particular, the Cern conjecture holds true for -extensible automata.

`G

G REEDY E XTENSTION 1: if for all and then 2: return Failure 3: else 4: such that Initializing the current word 5: such that Initializing the current set 6: while do 7: if for all then 8: return Failure 9: else 10: take a word of minimum length with 11: Updating the current word 12: Updating the current set KV:Extension 13: return Algorithm 2. Extension algorithm calculating a reset word for

c G2c E c c H

p W7 S

'

p

S c 9G `c

458

17

has in-degree

460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478

Thus, a word satises if and only if the vector lies beyond the subspace of all vectors orthogonal to . We aim to bound the minimum length of such word but rst we explain why words sending beyond exist. Since the automaton There is a is synchronizing and strongly connected (as it is Eulerian), there exists a word bug in [44] such that one can rst synchronize to a state and then move into by in this place applying a word that labels a path from to a state in . Then

Now consider the chain of subspaces , where is spanned by all vectors of form with . Clearly, if for some then for all whence for every . Let be the least number such that for some word of length , that is, the smallest such that . Then in

m m m F ge h m m ' m m e m m n Xc c F f e m TqY m m qq e f f f f q Q Rb ge ge ge F ge c c sSb c c b c c ! ! ! " E G m F m F h e c c c c q $ 4 F h ce d F i F h be $ F F h be c c c c $ 4 c c F c c F h ce d i h ce h e Q $ F Q F b lk j g F Ed g gbe $ D i i Q h xd 4 hbe c c c c f t f e f r u c ge d D E ge c d F F E YTqTq S q E E U 4YTqTqXWWP q S c c c c S q c c c c H c Xc ( c c ( c c ' c c ' ' c ' c c c c ( c 2c ( c c c c vc ' c

E ~ E F ( H c &c c # '

459

holds true since the left-hand side of (3.3) is the number of edges in the underlying graph KV:eq:eulerian of with ends in . The equality (3.3) readily implies that for each , one of the following alternatives takes place: either for all letters or for some . Now assume that a subset and a word are such that and is a word of minimum length with this property. We write for some and and let . Then by the choice of and whence . Thus, must fall into the second of the above alternatives and so for some . The word has the same length as and has the property that . Having this in mind, we now aim to prove that for every proper subset , there exists a word of length at most such that . It is here where linear algebra comes into the play. We may assume that . Assign to each subset its characteristic vector in the linear space of -dimensional row vectors over as follows: -th entry of is 1 if , otherwise it is equal to 0. For instance, is the all ones row vector and the vectors form the standard basis of . Observe that for any vector , the inner product is equal to the sum of all entries of . In particular, for each subset , we have . Further, assign to each word the linear operator on dened by for each . It is then clear that for each . The inequality that we look for can be rewritten as or . Let . Then as and . Since for every word , we have . Hence

2 c c Vc ' c

KV:eq:eulerian

18

489

J. Kari, M. Volkov

the chain

496 497 498 499 500 501 502 503 504 505

We mention in passing that the upper bound provided by Theorem 3.5 is far from being tight. So far experiments have discovered no Eulerian synchronizing automaton with states whose reset threshold would exceed . y Return to our discussion of extensibility. Even though the approach to the Cern conjecture via 1-extensibility has proved to be productive in several special cases, it cannot resolve the general case because there exist synchronizing automata that are not 1discovered by Kari [42], extensible. KV:fig:Kari The rst example here was the 6-state automaton see Figure 10. This automaton is synchronizing with reset threshold 25, the shortest reset word being . Kari found as a counter example to a gen y eralized form of the Cern conjecture proposed in Pins thesis [57] but the automaton is remarkable in several other respects. In particular, one can verify that no word of

KV:fig:Kari

506 507 508 509 510 511 512 513 514 515 516

length 6 or 7 is such that the full pre-image of the set under the action of has more than 4 elements. Recently Berlinkov [10] has constructed a series of synchronizing automata that for each contains an automaton that is not -extensible. The question of whether or not all synchronizing automata are 2-extensible remains open. 2-extensibility (and thusby KV:prop:extensibility Proposition 3.4a quadratic in the state number upper bound for the reset threshold) has been established for several classes of synchronizing automata by Rystsov [69, 70, 71]. Recently a slightly relaxed version of 2-extensibility has been veried by B al, Berline kov and Perrin [8, 6] for the important class of the so-called one-cluster automata. A DFA is called one-cluster if there exist a letter that labels only one simple

U 24y r R`XP

v S w t u 1

'

'

'

'

'

'

'

r s

p q

491

p q

490

and, in particular,

. But by our choice of we have whence . Since is the orthogonal complement of a 1-dimensional subspace, , and we conclude that . As shown in the rst paragraph of the proof, the above implies that for every proper , there exists a word of length at most such that . subset KV:Extension Then Algorithm 2 run on returns a reset word of length at most .

KV:thm:eulerian

all inclusions are strict whence

KV:fig:cerny-n

517 518 519 520 521

19

cycle. (For instance, the automata KV:fig:Kari and shown in Figure 8 are one-cluster while Karis automaton shown in Figure 10, is not. A mass example of one-cluster automata KV:sec:motivation is provided by the decoders of nite maximal prex codes discussed in Section 1.) If KV:Extension , and one can modify Algorithm 2 is this cycle, then it is easy to see that as follows. R ELATIVE E XTENSTION

1: 2: 3: 4: 5: 6: 7: 8: 9: 10:

of minimum length with Updating the current word Updating the current set

Algorithm 3. Modied extension algorithm for a one-cluster automaton with being a unique simple cycle labelled

530

A graph in which each vertex has the same out-degree (say, ) is called a graph of constant outdegree and the number is referred to as the outdegree of . If we take an alphabet whose of size is equal to the outdegree of , then we can label the edges of by letters of such that the resulting automaton will be complete and deterministic. Any DFA obtained this way is referred to as a coloring . Given a graph, it is reasonable to ask under which conditions it admits a coloring satisfying some good properties. In this section we analyze the so-called Road Coloring Problem that is certainly the most famous question within this framework. The Road Coloring Problem asks under which conditions graphs of constant outdegree admit a synchronizing coloring. The problem was explicitly stated by Adler, Goodwyn and Weiss [1] in 1977; in an implicit form it was present already in an earlier memoir by Adler and Weiss [2]. Adler,

533

532

531

KV:sec:rcp

c c

c c R

522

In [8, 6] it KV:One-clusterthat the length of each word appended by the main loop has been shown of Algorithm 3 does not exceed , and this clearly implies a quadratic in upper bound on the reset threshold for one-cluster synchronizing automata. A similar result has been obtained by Carpi and DAlessandro [16]. Steinberg [77, 78] has generalized the above approach and slightly improved the upper bound. Namely, Steinberg has proved that a one-cluster synchronizing automaton with states has a reset word of length at most y . He also has veried the Cern conjecture for one-cluster synchronizing automata with -cycles of prime cardinality.

`rG

:One-cluster

return

z

H c Xc z E c c

for all

then

'

where

i i~ i} | i{ ' " n X0

' z r

' S

20

543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559

J. Kari, M. Volkov

Goodwyn and Weiss considered only strongly connected graphs; as we shall see below this is quite a natural assumption since the general case easily reduces to the case of strongly connected graphs. The name of the problem suggested in [1] comes from the fol, lowing interpretation. In every strongly connectedsynchronizing automaton one can assign to state an instruction (a reset word) such that following one will surely arrive at from any initial state. (Indeed, for this one should rst follow an arbitrary reset word leading to some state , say, and then follow a word that labels a path connecting and such a path exists because of strong connectivity.) Thus, in order to help a traveler lost on a given strongly connected graph of constant outdegree to nd his/her way from wherever he/she could be, we should if possible color (that is, label) the edges of such that becomes a synchronizing automaton and then tell the traveler the magic sequence of colors representing a reset word leading to the travelers destination. The original motivation in [2, 1] came from symbolic dynamics. However, the Road Coloring Problem is quite natural also from the viewpoint of the reverse engineering of synchronizing automata: we aim to relate geometric properties of graphs to combinatorial properties of automata built on those graphs. The following necessary condition was found in [1]: Proposition 4.1. If a strongly connected graph admits a synchronizing coloring, then the g.c.d. of lengths of all cycles in is equal to . Proof. Arguing by contradiction, let be a common divisor of lengths of the cycles in . Let denote the vertex set of . Take a vertex and, for , let there exists a path from to of length

562 563 564 565 566 567 568 569 570 571 572 573

Clearly,

Let where . This means that in there are two paths from to : of length and of length . Since is strongly connected, there exists also a path from to of length , say. Combining it with each of the two paths above we get a cycle of length and a cycle of length . Since divides the length of any cycle in , we have and , whence , a contradiction. Thus, is a disjoint union of , and by the denition each edge in . Then denitely cannot be converted into a synchronizing leads from to automaton by any coloring of its edges: no paths of the same length originated in and can terminate in the same vertex because they end in and in respectively. Graphs satisfying the conclusion of Proposition 4.1 are called primitive8 . Adler, Goodwyn and Weiss [1] conjectured that primitivity is not only necessary for a graph to have a synchronizing coloring but also sufcient. In other word, they suggested the follow8 In the literature such graphs are sometimes called aperiodic. The term primitive comes from the notion of a primitive matrix in the Perron-Frobenius theory of non-negative matrices: it is known (and easy to see) that a graph is primitive if and only if so is its incidence matrix.

. We claim that

if

KV:prop:primitivity

561

:primitivity 560

!

577 578 579 580 581

21

ing Road Coloring Conjecture: every strongly connected primitive graph with constant out-degree admits a synchronizing coloring. The Road Coloring Conjecture has attracted much attention. There were several interesting partial results (see, e.g., [52, 31, 56, 41, 15, 43, 44]), and nally the problem was solved (in the afrmative) in August 2007 by Trahtman [80]. Trahtmans proof heavily depends on a neat idea of stability which is due to Culik, Karhum ki and Kari [23]. Let a be a DFA. We dene the stability relation on as follows:

582 583

Any pair such that and is called stable. The key observation by Culik, Karhum ki and Kari [23] was the following: a Proposition 4.2. If every strongly connected primitive graph with constant out-degree and more than one vertex has a coloring with a stable pair of vertices, then the Road Coloring Conjecture is true. Proof. Let be a strongly connected primitive graph with constant out-degree. We show that has a synchronizing coloring by induction on the number of vertices in . If has only one vertex, there is nothing to prove. If has more than one vertex, then it admits a coloring with a stable pair of states by the letters of some alphabet . Let be the automaton resulting from this coloring. It is easy to check that the stability relation is a congruence of . Since the relation is non-trivial, the quotient automaton has fewer vertices. It is clear that is strongly connected, moreover, since each cycle in induces a cycle of the same length in , the underlying graph of the latter automaton is primitive as well. Therefore, the graph admits a synchronizing coloring by the induction assumption. We lift this coloring to a coloring of in the following natural way. Every transition in the automaton induces the transition in (here and stand for the -classes of the vertices and respectively ). Now, if the transition is being recolored to for some , then the transition becomes . A crucial feature of this recoloring procedure is that it is consistent and are with the stability relation in the following sense. Suppose two transitions with the same label in such that and . Then , and the two transitions induce the same transition in . If it is being recolored to for some , then the two transitions are being changed and still have a in the same way such that the resulting transitions common label. Let be the automaton resulting from the described recoloring; we want to show that is synchronizing. Take a reset word for the synchronizing coloring of that we started with. If we apply to the states of the automaton , it will lead them all into a set that is contained in a single class of the relation . We induct on . If , then is a reset word for . If , take two states . Since they form a stable pair in , there exists a word such that . (Here and below subscripts indicate the automaton in which paths are being considered.) As discussed above, since , the paths started at and and labelled in have a common label , say,

KV:prop:ckk 584

585 586

587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614

22

J. Kari, M. Volkov

in as well. Thus, . Consider the set of the end points of all paths in that originate in and are labelled . Observe that and, since , the set is still contained in a single class of the relation . Therefore the induction assumption applies. Proposition 4.2 localizes the initial task: while synchronization is a global property in which all vertices are involved, the proposition shows that we may look at some pair of vertices. We need a further localization that allows us to concentrate on the action of a single letter. For this, we need some auxiliary notions and results. be a DFA. A pair of distinct vertices is compressible if Let for some ; otherwise it is incompressible. A subset is said to be compressible if contains a compressible pair and to be incompressible if every pair of distinct vertices in is incompressible. Clearly, if is incompressible, then for every word , the set also is incompressible and . Lemma 4.3. Let be a DFA and let maximum size in . Suppose that there exists a word in . Then has a stable pair.

KV:prop:ckk

compressible 628

629 630

631 632 633 634 635 636 637 638 639 640

be such that but for all . Take Proof. Let an arbitrary word ; we aim to show that for a suitable word . Clearly, we may assume that . Since the set is incompressible, the state forms an incompressible pair with every state in . Similarly, since the set is incompressible, the state also forms an incompressible pair with every state in , and of course every pair of distinct states in is incompressible too. Now has more than elements so it must be compressible, and the above analysis shows that the only pair in which may be compressible is the pair . Thus, there is a word such that , and the pair is stable. Suppose that is a DFA. Fix a letter and remove all edges of except those labelled . The remaining graph is called the underlying graph of or simply the -graph. Thus, in the -graph every vertex is the tail of exactly one edge. From every state , one can start a path in the -graph:

Since the set is nite, states in this path eventually begin repeating, that is, for some non-negative integer and some integer we have . In other words, KV:fig:a-orbit each path in the -graph eventually arrives at a cycle, see Fig. 11. The least non-negative integer such that for some is called the -level of the state and the state is called the root of . The cycles of the -graph are referred to as -cycles. Lemma 4.4. Let be a strongly connected DFA. Suppose that there is a letter such that all states of maximal -level have the same root. Then has a stable pair.

647 648

Q G ' ' ` ! " ! ' ! ' " ' ' ' " ! ' E! " ' ' " ! ' E! q q " q " " rTqYq ' #! TqYq ' ! ' #! ! ' !' ' ' ' ' ff $ %! ! H $ H "`%! "! H $ " `"%! ! $ " g " U "%! %v! P %$v "$ "$ c c U 9X%! 9" ! P 4&avY$g" v &$v " "t! " "4v "$v $ ! 4"%! " $ " " $ ! ! " H $ " T"%! H! E H ! v $ ! U P $ " ! "! %! E `r} f U )c Y" `PG " " y! `G

"

" c cVc c

$ 2H

!

H 2 " $H "

:fig:a-orbit

660

662

664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679

9 This

(

'

'

'

'

Q !

( "

663

Proof. If has just one vertex, it is nothing to prove. Thus, we assume that has more than one vertex and prove that it admits a coloring with a stable pair of statesthe result KV:prop:ckk will then follow from Proposition 4.2. Fix an arbitrary coloring of by letters from an alphabet and take an arbitrary letter . We induct on the number of states that do not lie on any -cycle in the chosen coloring. We say that a vertex of is ramied if it serves as the tail for some edges with different heads. Suppose that . This means that all states lie on the -cycles. If we suppose that no vertex in is ramied, then there is just one -cycle (since is strongly connected) and all cycles in have the same length. This contradicts the assumption that is primitive9. Thus, let be a vertex which is ramied. Then there exists a letter such that the states and are not equal. We exchange the labels of the edges and KV:fig:rcp-basis , see Fig. 12. It is clear that in the new coloring there isroot one state of maximal only KV:lemma:common -level, namely, the state . Thus, Lemma 4.4 applies and the induction basis is veried. Now suppose that . We denote by the maximum -level of the states in the chosen coloring. Observe that implies .

Theorem 4.5 ([80]). Every strongly connected primitive graph with constant out-degree admits a synchronizing coloring.

"

652

'

651

$ i $ ' ' " ' " ' " ' " ' ' ' ' V" ' q i $ b $ $ 9 b " $ E

650

649

Proof. Let be the set of all states of -level . Then for all whence no pair of vertices from is incompressible. Thus, any incompressible set in has at most one common state with . Take an incompressible set of maximum size in and choose any state . Since the automaton is strongly connected, there is a path from to a state in . If is the word that labels this path, then is an incompressible set of maximum size and it has exactly one common state with (namely, ). Then is an incompressible set of maximum size that has all its states except one (namely, ) in some -cyclesthe latter conclusion is ensured by our choice of . If is the l.c.m. of the lengths of all simple -cycles, then xesKV:lemma:maximal -cycle but all states in every incompressible . We see that Lemma 4.3 applies (with in the role and in the role of ).

Pu n (Qn Xu&4n ' ' & n &4n ' ' ' ' u n u Qn

23

%! ! $

'

'

4n

'

'

24

J. Kari, M. Volkov

ig:rcp-basis

' ' "R$ $ ' (

ig:rcp-case1

692 693

702

such that

'

'"

'

'

Sz

695

694

Case 2: is on . Let be the least integer such that . The state is also on . Let be the least integer such that . Then the length of is . KV:fig:rcp-subcase21 . Again, we swap the labels of and , see Fig. 14. Subcase 2.1: If , then the swapping creates an -cycle of length increasing the number of states on the -cycles. If , then the -level of becomes whence all states of maximal -level in the new automaton are -ascendants of and thus have the same root.

'

691

' )"$ $ S

'

'

690

689

We swap the labels of and , see Fig. 13. If was on the -path from to , then the swapping creates a new -cycle increasing the number of states on the -cycles. If was not on the -path from to , then the -level of becomes whence all states of maximal -level in the new automaton are -ascendants of and thus have as the common root.

'

fig:rcp-case1

'

'

( $

'

' $

'

' ' $

688

Case 1:

is not on

'

682

'

681

' $ (

'

680

Let be a state of level . Since is strongly connected, there is an edge with , and by the choice of , the label of this edge is some letter . Let . One has . Let be the root of and let be the -cycle on which lies. The following considerations split in several cases. In each case except one we can recolor by swapping the labels of two edges so that the new coloring either satises the KV:lemma:common root premise of Lemma 4.4 (all states of maximal -level have the same root) or has more states on the -cycles (and the induction assumption applies). In the remaining case nding a stable pair will be easy.

'

'

' '

'

'

' "

' (

'

' '

25

cp-subcase21

703 704

and

is ramied. Since

C

cp-subcase22

705 706 707 708

We swap the labels of and , see Fig. 14. If still lies on an -cycle, then the length of the -cycle is at least and the number of states on the -cycles increases. Otherwise, the -level of becomes at least whence all states of maximal -level in the new automaton are -ascendants of and have a common root.

KV:fig:rcp-subcase21

711

cp-subcase23

712 713

Figure 16. Recoloring reducing Subcase 2.3 to Subcase 2.2 If we swap the labels of and , then we nd ourselves in the conditions KV:fig:rcp-subcase23 of Subcase 2.2 (with and playing the roles of and respectively), see Fig. 16. Subcase 2.4: and neither nor is ramied. In this subcase it is clear that and form a stable pair whichever coloring of KV:fig:rcp-subcase24 chosen, see Fig. 17. This completes the proof.

C

'

! 3 $

$ %! $

$ %!

710

709

. Let be the state on the -path from to such that Subcase 2.3: and is ramied. Since is ramied, there is a letter that .

$3 '

'

'

1

'

" ' ! E S

C

1 ' ' ( $

'

( ' $

3

1 ' $ %! ' ! ' C ' " ! ! ' ' ' ' ' $ 3 3 3

1

'

'

C

'

such

such

is

26

J. Kari, M. Volkov

cp-subcase24

KV:thm:road coloring

The above proof of Theorem 4.5 is constructive and can be unfolded to an algorithm that, given a strongly connected primitive graph with constant out-degree, nds a synchronizing coloring of ; moreover, this can be done in time quadratic in the number of vertices in , see [7]. If one drops the primitivity condition, one can prove (basically by the same method) the following generalization of the Road Coloring Theorem, see [7]: Theorem 4.6. Suppose that is the g.c.d. of the lengths of cycles in a strongly connected graph with constant out-degree. Then admits a coloring for which there is a word such that . Finally, we discuss a general version of the Road Coloring Problem in which graphs are not assumed to be strongly connected. Given an arbitrary graph , a vertex is said to be reachable from a vertex if there is a path from to . Clearly, the reachability relation is transitive, and the mutual reachability relation is an equivalence on the vertex set of . The subgraphs induced on the classes of the mutual reachability relation are strongly connected and are called the strongly connected components of the graph . The reachability relation induces a partial order on the set of the strongly connected components: a component precedes a component in this order if some vertex of is reachable from some vertex of . The following result shows that the general case of the Road KV:thm:road Coloring Problem easily reduces to its strongly connected case (solved by Theorem 4.5): Corollary 4.7. A graph with constant out-degree admits a synchronizing coloring if and only if has the least strongly connected component and this component is primitive. An interesting issue related to the Road Coloring Problem is the choice of the optimal synchronizing coloring for a given graph. Clearly, graphs admitting a synchronizing coloring may have many colorings and reset thresholds of the resulting synchronizing y automata may drastically differ. For instance, it is easy to see that the Cern automaton whose reset threshold is believed to be maximum possible for an -state automaton admits a recoloring with reset threshold is as low as (and moreover, every strongly connected graph with constant out-degree that has a loop admits a synchronizing coloring whose reset threshold is less that the number of vertices of .) Nevertheless, there exist graphs whose synchronizing coloring are slowly synchronizing KV:fig:Wielandt automata. As an example, consider the Wielandt graph shown in Figure 18. It has vertices , say, and edges: two edges from to for each

724 725

726 727 728 729 730 731 732 733 734 735

:rcp-general 736

737

738 739 740 741 742 743 744 745 746 747 748

S &

S 9

S 9G

" Vc c Q

-imprimitive 723

' '

3

coloring

S q S TqYTqVYXQ

X0

27

fig:Wielandt

749 750 751 752 753 754 755 756 757 758 759 760 761 762

, and the edges from to and . The graph (more precisely, its incidence matrix) rst appeared in Wielandts seminal paper [82] where Wielandt stated that for every primitive non-negative -matrix , the matrix is positive. The incidence matrix of was used to show that this bound is tight (that is, its -th power still has some 0 entries); later it was observed to be the only (up to a simultaneous permutation of rows and columns) matrix with this property, see [26]. It is easy to realize KV:fig:cerny-n of the graph that every coloring is isomorphic to the automaton shown in Figure 8 on the right. Since is strongly connected and primitive, the Road Coloring Theorem implies that is synchronizing (of course, this can also be veried directly). In [4] it isKV:prop:cerny reset threshold of shown that the is , see the proof of Proposition 3.1 above. The aforementioned extremal property of the Wielandt graphs gives some evidence for conjecturing that this series of graphs may yield the extremal value also for the reset threshold of synchronizing colorings of -vertex y graph. In other words, we suggest a conjecture that is in a sense parallel to the Cern one. Conjecture 4.8. Every strongly connected primitive graph with constant outdegree and vertices admits a synchronizing coloring that can be reset by a word of length . y We observe that while there is a clear analogy between Conjecture 4.8 and the Cern conjecture, the validity of none of them immediately KV:conj:hybrid of the other. implies the validity Some rst partial results related to Conjecture 4.8 can be found in [78, 17]. Roman [67] has shown that the problem of nding the optimal synchronizing coloring for a given graph is computationally hard. Namely, the following decision problem is NPcomplete: B OUNDED -S YNCHRONIZING -C OLORING : Given a strongly connected primitive graph with constant outdegree, is it true that has a synchronizing coloring with a reset word of length ?

KV:conj:hybrid

:conj:hybrid 763

764

774

5 Related work

:sec:related

r y

G n

S 9

1 2

g

q S YTqq E S Q n

Q

3

S q qS I YTqYV n

28

775

J. Kari, M. Volkov

References

[1] R. L. Adler, L. W. Goodwyn, and B. Weiss. Equivalence of topological Markov shifts. Israel J. Math, 27(1):4963, 1977. 19, 20 [2] R. L. Adler and B. Weiss. Similarity of automorphisms of the torus. Memoirs Amer. Math. Soc., 98, 1970. 19, 20 [3] N. Alon, D. Moshkovitz, and S. Safra. Algorithmic construction of sets for k-restrictions. ACM Trans. Algorithms, 2(2):153177, 2006. 9 [4] D. Ananichev, V. Gusev, and M. Volkov. Slowly synchronizing automata and digraphs. In P. Hlin n and A. Ku era, editors, Mathematical Foundations of Computer Science, volume e y c 6281 of Lecture Notes in Comput. Sci., pages 5564. Springer-Verlag, 2010. 10, 27 [5] W. R. Ashby. An introduction to cybernetics. Chapman & Hall, 1956. 2 [6] M.-P. B al, M. Berlinkov, and D. Perrin. A quadratic upper bound on the size of a synchronize ing word in one-cluster automata. Int. J. Foundations Comp. Sci., 22(2):277288, 2011. 18, 19 [7] M.-P. B al and D. Perrin. A quadratic algorithm for road coloring. Technical report, Universit e e Paris-Est, 2008. 26 [8] M.-P. B al and D. Perrin. A quadratic upper bound on the size of a synchronizing word in onee cluster automata. In V. Diekert and D. Nowotka, editors, Developments in Language Theory, Lecture Notes in Comput. Sci., pages 8190. Springer-Verlag, 2009. 18, 19 [9] M. Berlinkov. Approximating the minimum length of synchronizing words is hard. In F. Ablayev and E. W. Mayr, editors, Computer Science in Russia, volume 6072 of Lecture Notes in Comput. Sci., pages 3747. Springer-Verlag, 2010. 9 [10] M. Berlinkov. On a conjecture by Carpi and DAlessandro. In Y. Gao, H. Lu, S. Seki, and S. Yu, editors, Developments in Language Theory, volume 6224 of Lecture Notes in Comput. Sci., pages 6675. Springer-Verlag, 2010. 18 [11] J. Berstel, D. Perrin, and C. Reutenauer. Codes and automata. Number 129 in Encyclopedia of Mathematics and its Applications. Cambridge University Press, 2009. 3 c [12] S. Bogdanovi , B. Imreh, M. Ciri , and T. Petkovi . Directable automata and their generalc c izations: a survey. Novi Sad J. Math., 29(2):2969, 1999. 5 [13] V. Boppana, S. Rajan, K. Takayama, and M. Fujita. Model checking based on sequential ATPG. In Computer Aided Verication, volume 1622 of Lecture Notes in Comput. Sci., pages 418430. Springer-Verlag, 1999. 2 [14] R. M. Capocelli, L. Gargano, and U. Vaccaro. On the characterization of statistically synchronizable variable-length codes. IEEE Transactions on Information Theory, 34(4):817825, 1988. 3 [15] A. Carbone. Cycles of relatively prime length and the road coloring problem. Israel J. Math., 123:303316, 2001. 21 [16] A. Carpi and F. DAlessandro. The synchronization problem for locally strongly transitive automata. In R. Kr lovic and D. Niwinski, editors, Mathematical Foundations of Computer a Science, volume 5734 of Lecture Notes in Computer Science, pages 211222. Springer-Verlag, 2009. 19 y [17] A. Carpi and F. DAlessandro. On the hybrid Cern -Road Coloring Problem and Hamiltonian paths. In Y. Gao, H. Lu, S. Seki, and S. Yu, editors, Developments in Language Theory, volume 6224 of Lecture Notes in Computer Science, pages 124135. Springer-Verlag, 2010. 27

776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819

820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863

29

y [18] J. Cern . Pozn mka k homog nnym eksperimentom s kone n mi automatami. Matematickoa e c y fyzikalny Casopis Slovenskej Akad mie Vied, 14(3):208216, 1964. (in Slovak). 2, 6, 10, 11, e 12 y [19] J. Cern , A. Pirick , and B. Rosenauerov . On directable automata. Kybernetika, 7(4):289 a a 298, 1971. 12 [20] Y.-B. Chen and D. J. Ierardi. The complexity of oblivious plans for orienting and distinguishing polygonal parts. Algoritmica, 14:367397, 1995. 4 [21] H. Cho, S.-W. Jeong, F. Somenzi, and C. Pixley. Synchronizing sequences and symbolic traversal techniques in test generation. J. Electronic Testing, 4:1931, 1993. 2 [22] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to algorithms. MIT Press and McGraw-Hill, 2001. 6 [23] K. Culik II, J. Karhum ki, and J. Kari. A note on synchronized automata and Road Coloring a Problem. Int. J. Found. Comput. Sci., 13:459471, 2002. 21 [24] F. M. Dekking. The spectrum of dynamical systems arising from substitutions of constant length. Z. Wahrsch. Verw. Gebiete, 41:221239, 1978. 4 y [25] L. Dubuc. Sur le automates circulaires et la conjecture de Cern . RAIRO Inform. Th or. App., e 32:2134, 1998. (in French). 16 [26] A. L. Dulmage and N. S. Mendelsohn. Gaps in the exponent set of primitive matrices. Ill. J. Math., 8:642656, 1964. 27 [27] D. Eppstein. Reset sequences for monotonic automata. SIAM J. Comput., 19:500510, 1990. 4, 7, 10 [28] M. A. Fischler and M. Tannenbaum. Synchronizing and representation problems for sequential machines with masked outputs. In Proc. 11th Annual Symp. Foundations Comput. Sci, pages 97103. IEEE Press, 1970. 10, 13 [29] P. Frankl. An extremal problem for two families of sets. European J. Combinatorics, 3:125 127, 1982. 13 [30] D. Frettl h and B. Sing. Computing modular coincidences for substitution tilings and point o sets. Discrete Comput. Geom., 37:381407, 2007. 10 [31] J. Friedman. On the road coloring problem. Proc. Amer. Math. Soc., 110:11331135, 1990. 21 [32] M. R. Garey and D. S. Johnson. Computers and intractability: a guide to the theory of NPcompleteness. Freeman, 1979. 7 [33] P. Gawrychowski. Complexity of shortest synchronizing word. Private communiction, 2008. 9 [34] M. Gerbush and B. Heeringa. Approximating minimum reset sequences. In M. Domaratzki and K. Salomaa, editors, Implementation and Application of Automata, volume 6482 of Lecture Notes in Comput. Sci., pages 154162. Springer-Verlag, 2011. 9 [35] A. Gill. State-identication experiments in nite automata. Inform. Control, 4(2-3):132154, 1961. 2 [36] S. Ginsburg. On the length of the smallest uniform experiment which distinguishes the terminal states of a machine. J. Assoc. Comput. Mach., 5:266280, 1958. 2 [37] K. Goldberg. Orienting polygonal parts without sensors. Algorithmica, 10:201225, 1993. 4 [38] P. Goral ik and V. Koubek. Rank problems for composite transformations. Internat. J. Algebra c Comput., 5:309316, 1995. 7

30

864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908

J. Kari, M. Volkov

[39] F. C. Hennie. Fault detecting experiments for sequential circuits. In Switching Circuit Theory and Logical Design, pages 95110. IEEE Press, 1964. 2 [40] T. Jiang and M. Li. On the approximation of shortest common supersequences and longest common subsequences. SIAM J. Comput., 24(5):11221139, 1995. 9 [41] N. Jonoska and S. Suen. Monocyclic decomposition of graphs and the road coloring problem. Congressum numerantium, 110:201209, 1995. 21 [42] J. Kari. A counter example to a conjecture concerning synchronizing words in nite automata. Bull. European Assoc. Theor. Comput. Sci., 73:146, 2001. 18 [43] J. Kari. Synchronization and stability of nite automata. J. Universal Comp. Sci., 2:270277, 2002. 21 [44] J. Kari. Synchronizing nite automata on Eulerian digraphs. Theoret. Comput. Sci., 295:223 232, 2003. 16, 17, 21 [45] A. A. Klyachko, I. K. Rystsov, and M. A. Spivak. An extremal combinatorial problem associated with the bound of the length of a synchronizing word in an automaton. Cybernetics and System Analysis, 23(2):165171, 1987. translated from Kibernetika, No. 2, 1987, pp. 1620, 25. 13 [46] Z. Kohavi and J. Winograd. Bounds on the length of synchronizing sequences and the order of information losslessness. In Z. Kohavi and A. Paz, editors, Theory of Machines and Computations, pages 197206. Academic Press, 1971. 13 [47] Z. Kohavi and J. Winograd. Establishing certain bounds concerning nite automata. J. Comput. System Sci., 7(3):288299, 1973. 13 [48] A. E. Laemmel and B. Rudner. Study of the application of coding theory. Technical Report PIBEP-69-034, Dept. Electrophysics, Polytechnic Inst. Brooklyn, Farmingdale, N.Y., 1969. 10 [49] E. F. Moore. Gedanken experiments on sequential machines. In C. E. Shannon and J. McCarthy, editors, Automata Studies, pages 129153. Princeton Universty Press, 1956. 2 [50] B. K. Natarajan. An algorithmic approach to the automated design of parts orienters. In Proc. 27th Annual Symp. Foundations Comput. Sci., pages 132142. IEEE Press, 1986. 4 [51] B. K. Natarajan. Some paradigms for the automated design of parts feeders. Internat. J. Robotics Research, 8(6):89109, 1989. 4 [52] G. L. OBrien. The road coloring problem. Israel J. of Math., 39:145154, 1981. 21 [53] J. Olschewski and M. Ummels. The complexity of nding reset words in nite automata. In P. Hlin n and A. Ku era, editors, Mathematical Foundations of Computer Science, number e y c 6281 in Lecture Notes in Comput. Sci., pages 568579. Springer-Verlag, 2010. 9 [54] C. H. Papadimitriou. Computational complexity. Addison-Wesley, 1994. 7 [55] C. H. Papadimitriou and M. Yannakakis. The complexity of facets (and some facets of complexity). J. Comput. System Sci., 28(2):244259, 1984. 9 [56] D. Perrin and M. P. Sch tzenberger. Synchronizing prex codes and automata and the road u coloring problem. In Symbolic dynamics and its applications, volume 135 of Contemporary Mathematics, pages 295318. Amer. Math. Soc., 1992. 21 y [57] J.-E. Pin. Le probl` me de la synchronisation et la conjecture de Cern . Th` se de 3` me cycle, e e e Universit Paris VI, 1978. 18 e y [58] J.-E. Pin. Sur un cas particulier de la conjecture de Cern . In Proc. 5th Colloq. on Automata, Languages, and Programming (ICALP), volume 62 of Lecture Notes in Comput. Sci., pages 345352. Springer-Verlag, 1978. (in French). 16

909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952

31

[59] J.-E. Pin. On two combinatorial problems arising from automata theory. Ann. Disc. Math., 17:535548, 1983. 13 [60] C. Pixley, S.-W. Jeong, and G. D. Hachtel. Exact calculation of synchronization sequences based on binary decision diagrams. In Proc. 29th Design Automation Conf., pages 620623. IEEE Press, 1992. 10 [61] N. Pytheas Fogg. Substitutions in dynamics, arithmetics and combinatorics, volume 1794 of Lecture Notes in Mathematics. Springer-Verlag, 2002. Edited by V. Berth , S. Ferenczi, C. e Mauduit and A. Siegel. 4 [62] M. O. Rabin and D. Scott. Finite automata and their decision problems. IBM J. Res. Develop., 3(2):114125, 1959. 6 [63] J. L. Ramrez Alfonsn. The diophantine Frobenius problem. Oxford University Press, 2005. 11 [64] J.-K. Rho, F. Somenzi, and C. Pixley. Minimum length synchronizing sequences of nite state machine. In Proc. 30th Design Automation Conf., pages 463468. ACM, 1993. 7 [65] A. Roman. Genetic algorithm for synchronization. In A. Dediu, A. Ionescu, and C. Martn Vide, editors, Language and Automata Theory and Applications, volume 5457 of Lecture Notes in Comput. Sci., pages 684695. Springer-Verlag, 2009. 10 [66] A. Roman. Synchronizing nite automata with short reset words. Applied Mathematics and Computation, 209(1):125136, 2009. 10 [67] A. Roman. The NP-completeness of the Road Coloring Problem. 111(7):342347, 2011. 27 Inf. Process. Lett.,

[68] I. K. Rystsov. On minimizing length of synchronizing words for nite automata. In Theory of Designing of Computing Systems, pages 7582. Institute of Cybernetics of Ukrainian Acad. Sci., 1980. (in Russian). 7 [69] I. K. Rystsov. Almost optimal bound of reccurent word length for regular automata. Cybernetics and System Analysis, 31:669674, 1995. translated from Kibernetika i Sistemnyj Analiz, No. 5, 1995, pp. 4048. 18 [70] I. K. Rystsov. Quasioptimal bound for the length of reset words for regular automata. Acta Cybernetica, 12:145152, 1995. 18 [71] I. K. Rystsov. Reset words for automata with simple idempotents. Cybernetics and System Analysis, 36:339344, 2000. translated from Kibernetika i Sistemnyj Analiz, No. 3, 2000, pp. 3239. 18 [72] A. Salomaa. Composition sequences for functions over a nite domain. Theoret. Comput. Sci., 292:263281, 2003. 7 [73] W. Samotij. A note on the complexity of the problem of nding shortest synchronizing words. In Proc. AutoMathA 2007, Automata: from Mathematics to Applications. Univ. Palermo, 2007. (CD). 7, 8 [74] S. Sandberg. Homing and synchronizing sequences. In M. Broy, B. Jonsson, J.-P. Katoen, M. Leucker, and A. Pretschner, editors, Model-Based Testing of Reactive Systems, volume 3472 of Lecture Notes in Comput. Sci., pages 533. Springer-Verlag, 2005. 2, 7 [75] A. L. Selman. A taxonomy of complexity classes of functions. J. Comput. System Sci., 42(1):357381, 1994. 9 [76] P. H. Starke. Eine Bemerkung uber homogene Experimente. Elektronische Informationverar beitung und Kybernetik, 2:257259, 1966. (in German). 11

32

953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969

Index

y [77] B. Steinberg. The averaging trick and the Cern conjecture. In Y. Gao, H. Lu, S. Seki, and S. Yu, editors, Developments in Language Theory, volume 6224 of Lecture Notes in Comput. Sci., pages 423431. Springer-Verlag, 2010. 19 y [78] B. Steinberg. The Cern conjecture for one-cluster automata with prime length cycle. CoRR, abs/1005.1835, 2010. 19, 27 [79] A. Trahtman. An efcient algorithm nds noticeable trends and examples concerning the y Cern conjecture. In R. Kr lovi and P. Urzyczyn, editors, 31st Int. Symp. Math. Foundations a c of Comput. Sci., volume 4162 of Lecture Notes in Comput. Sci., pages 789800. SpringerVerlag, 2006. 7, 10 [80] A. Trahtman. The Road Coloring Problem. Israel J. Math., 172(1):5160, 2009. 21, 23 y [81] M. Volkov. Synchronizing automata and the Cern conjecture. In C. Martn-Vide, F. Otto, and H. Fernau, editors, Language and Automata Theory and Applications, volume 5196 of Lecture Notes in Comput. Sci., pages 1127. Springer-Verlag, 2008. 2, 4 [82] H. Wielandt. Unzerlegbare, nicht negative Matrizen. Math. Z., 52:642648, 1950. (in German). 10, 27 [83] S. Yang. Logic synthesis and optimization benchmarks. Technical Report User Guide Version 3.0, Microelectronics Center of North Carolina, Research Triangle Park, NC, 1991. 10

Index

970

2-renewing sequence, 13 automaton -extensible, 15 y Cern , 10 Kari, 18 one-cluster, 18 synchronizing, 1 B OUNDED -S YNCHRONIZING C OLORING, 27 y Cern conjecture, 11 y Cern function, 11 characteristic vector, 17 coincidence condition, 4 coloring (of a graph), 19 compressible pair, 22 compressible set, 22 couple, 7 graph, 6 of a letter, 22 of constant outdegree, 19 primitive, 20 Wielandt, 26 greedy algorithm compression, 12 extension, 15 identity of unary algebras, 5 heterotypical, 5 homotypical, 5 incompressible pair, 22 incompressible set, 22 prex code, 3 maximal, 3 synchronized, 3 reachability relation, 26 reset threshold, 9

1005 1006

1007

reset word, 1 Road Coloring Conjecture, 21 Road Coloring Problem, 19 S HORT-R ESET-W ORD, 7 S HORTEST-R ESET-W ORD, 8 stability relation, 21 stable pair, 21 strongly connected component, 26 subset automaton, 6 substitution, 4 of nite length, 4 synchronizing word of a code, 3 unary term, 5 underlying graph (of an automaton), 6 vertex ramied, 23

977 978

1016

1017 1018

1019 1020

1003 1004

33

- Ruler Folding problem PaperUploaded byVishnuprakash Puthiyakovilakath
- Games Puzzles and ComputationUploaded byIndi Fernandes
- Complex problem solving field definitionUploaded byAntonio Moreno Fleitas
- 02_The Role of Algorithms in ComputingUploaded byRohit Chaudhary
- problemset3sol[1]Uploaded byGobara Dhan
- Operations ResearchUploaded bymiclausbestia
- The Computer Science HandbookUploaded byAnonymous n6EXrLLvL
- Daa Notes FullUploaded byAarif Muhammed
- caimUploaded byPayal
- tocmod5Uploaded byJibu James
- Guide 1Uploaded bysiva2203
- GATE Theory of Computation BookUploaded byMims12
- bocaUploaded byohudalharthi
- In Tract AbilityUploaded byPenélope Molina Sierra
- Randomized Algorithms for 3-SatUploaded bySari Barhum
- BelyaevUploaded bymanu_1608
- Linux Cpu SchedulerUploaded byRashminMewada
- Amortized AnalysisUploaded byKaran Desai
- sigmod16aUploaded byNorbert Feron
- DETERMINAN 1Uploaded byGus Edi
- Matrix Det EXAMPLE.pdfUploaded byVijaybalaji
- Lecture Notes on Lie AlgebrasUploaded byalin444444
- Multivariate Notes r1Uploaded byAurea De Jesus Romo
- Determinants 2008Uploaded byKeval Velani
- Assignment (Ch Determinants)Uploaded byAtul Verma
- Automata Theory NotesUploaded byshailaja
- L.NO.04 MATHS 311Uploaded bycskaditya
- Data Structure 2Uploaded byWaqas Ahmed
- Optimal ROC CurveUploaded byCesar Costa
- CD QuestionsUploaded bySri Harsha Sasanapuri

- 11_Topic11__MeasurementDataProc.10.16.17docx (1)Uploaded bykat
- Eced NotesUploaded byaishuvc1822
- US Federal Trade Commission: divorcementUploaded byftc
- Control 123Uploaded byhahaha
- welcome to 8th grade mathUploaded byapi-324368265
- HW4Uploaded byKassandra Gianinoto
- LinearEquationsUploaded byTitser Laarni
- Deepstar - Multiphase FlowUploaded byfebrian
- Structural Analysis of Disc Brake Rotor for Different MaterialsUploaded byAnonymous kw8Yrp0R5r
- Calculation of the Theoretical Weight of SteelsUploaded byhorascanman
- Relativistic_Heat_Conduction.pdfUploaded byJose Daniel Hoyos
- 3_1Uploaded byapi-3717843
- FInal Exam SummitUploaded byHedly KS
- I'm from the street man I feel like sesame, uhUploaded by;(
- f-16 nnUploaded bykkkraja
- Vibration Theory Ch3Uploaded byrmartinson8
- International Boiling Point Project RubricUploaded bygjohnston9077
- An Exact Mapping from Navier-Stokes Equation to Schrodinger Equation via Riccati EquationUploaded byRyanElias
- Applied Mathematics in Engineering and ReliabilityUploaded byG Delis
- Mathematica SolutionUploaded byShakeel Ahmad Kasuri
- ETF2100_Assignment_2_S1_2018.pdfUploaded byHw Solution
- Engineering. Electronics. .Electronics.and.Circuit.analysisUploaded bydehgan
- Application Note - PMC Axis ControlUploaded byAbhishek Chaubey
- Class-7_46Uploaded byNabarunSarma
- Articolo-3Uploaded bysoniafranco
- SSRN-id809107Uploaded byapi-3715003
- 08 Answers to All ProblemsUploaded byteetie
- Generation Expansion Planning in Wind-Thermal Power SystemsUploaded byNan Zhang
- Random ProcessUploaded byChaithanya Suresh
- Chapter18GroupTechnologyandCADPlanningUploaded bymohanmech2006886