Ambiguity, Weakness, and Regularity in Probabilistic Büchi Automata
aa r X i v : . [ c s . F L ] A p r Ambiguity, Weakness, and Regularityin Probabilistic Büchi Automata
Christof Löding ( B ) and Anton Pirogov ( B ) ⋆ ⋆⋆ RWTH Aachen University, Templergraben 55, 52062 Aachen, Germany {loeding,pirogov}@cs.rwth-aachen.de
Abstract.
Probabilistic Büchi automata are a natural generalizationof PFA to infinite words, but have been studied in-depth only ratherrecently and many interesting questions are still open. PBA are knownto accept, in general, a class of languages that goes beyond the regularlanguages. In this work we extend the known classes of restricted PBAwhich are still regular, strongly relying on notions concerning ambiguityin classical ω -automata. Furthermore, we investigate the expressivity ofthe not yet considered but natural class of weak PBA, and we also showthat the regularity problem for weak PBA is undecidable. Keywords: probabilistic · Büchi · automata · ambiguity · weak Probabilistic finite automata (PFA) are defined similarly to nondeterministicfinite automata (NFA) with the difference that each transition is equipped witha probability (a value between 0 and 1), such that for each pair of state andletter, the probabilities of the corresponding outgoing transitions sum up to 1.PFA have been investigated already in the 1960ies in the seminal paper of Rabin[18]. But while the development of the theory of automata on infinite words alsostarted around the same time [7], the model of probabilistic automata on infinitewords has first been studied systematically in [3]. The central model in thistheory is the one of probabilistic Büchi automata (PBA), which are syntacticallythe same as PFA. The acceptance condition for runs is defined as for standardnondeterministic Büchi automata (NBA): a run on an infinite word is acceptingif it visits an accepting state infinitely often (see [23,24] for an introduction tothe theory of automata on infinite words). In general, for probabilistic automataone distinguishes different criteria of when a word is accepted. In the positivesemantics, it is required that the probability of the set of accepting runs is greaterthan 0, in the almost-sure semantics it has to be 1, and in the threshold semanticsit has to be greater than a given value λ between 0 and 1. It is easy to see that ⋆ This work is supported by the German research council (DFG) Research TrainingGroup 2236 UnRAVeL ⋆⋆ The final authenticated publication is available online at https://doi.org/10.1007/978-3-030-45231-5_27
C. Löding and A. Pirogov
PFA with positive or almost-sure semantics can only accept regular languages,because these conditions correspond to the fact that there is an accepting run orthat all runs are accepting. For infinite words the situation is different, becausesingle runs on infinite words can have probability 0. Therefore, the existence ofan accepting run is not the same as the set of accepting runs having probabilitygreater than 0 (similarly, almost-sure semantics is not equivalent to all runsbeing accepting). And in fact, it turns out that PBA with positive (or almost-sure) semantics can accept non-regular languages [3]. This naturally raises thequestion under which conditions a PBA accepts a regular language.In [3] a subclass of PBA that accept only regular languages (under positivesemantics) is introduced, called uniform PBA. The definition uses a semanticcondition on the acceptance probabilities in end components of the PBA. Asyntactic class of PBA that accepts only regular languages (under positive andalmost-sure semantics) are the hierarchical PBA (HPBA) introduced in [8]. Thestate space of HPBA is partitioned into a sequence of layers such that for eachpair of state and letter there is at most one transition that does not increase thelayer. Decidability and expressiveness questions for HPBA have been studied inmore detail in [11,10]. While HPBA accept only regular languages for positiveand almost-sure semantics, it is not very hard to come up with HPBA thataccept non-regular languages under the threshold semantics [8,11] (see also theexample in Figure 2(a) on page 10). Restricting HPBA further such that thereare only two layers and all accepting states are on the first layer leads to a classof PBA (called simple PBA, SPBA) that accept only regular languages evenunder threshold semantics [9].In this paper, we are also interested in the question under which condi-tions PBA accept only regular languages. We identify syntactical patterns inthe transition structure of PBA whose absence guarantees regularity of the ac-cepted language. These patterns have been used before for the classification ofthe degree of ambiguity of NFA and NBA [25,19,16]. The degree of ambiguity ofa nondeterministic automaton corresponds to the maximal number of acceptingruns that a single input word can have. For NBA, the ambiguity can (roughly)be uncountable, countable, or finite. For positive semantics, we show that PBAwhose transition structure corresponds to at most countably ambiguous NBA,accept only regular languages. For almost-sure semantics, we need a slightlystronger condition for ensuring regularity. But both classes that we identify areeasily seen to strictly subsume the class of HPBA. For the emptiness and uni-versality problems for these classes we obtain the same complexities as the onesfor HPBA. In the case of threshold semantics, we show that finite ambiguityis a sufficient condition for regularity of the accepted language, generalizing acorresponding result for PFA from [12]. The class of finitely ambiguous PBAstrictly subsumes the class of SPBA.Besides the relation between regularity and ambiguity in PBA, we also inves-tigate the class of weak PBA (abbreviated PWA). In weak Büchi automata, theset of accepting states is a union of strongly connected components of the au-tomaton. We show that PWA with almost-sure semantics define the same class mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 3 of languages as PBA with almost-sure semantics (which implies that with posi-tive semantics PWA define the same class as probabilistic co-Büchi automata).This is in correspondence to results for non-probabilistic automata: weak au-tomata with universal semantics (a word is accepted if all runs are accepting)define the same class as Büchi automata with universal semantics, and nonde-terministic weak automata correspond to nondeterministic co-Büchi automata(see, e.g., [17], where weak automata are called weak parity automata). Further-more, it is known that universal Büchi automata, respectively nondeterministicco-Büchi automata, can be transformed into equivalent deterministic automata(with the same acceptance condition). An analogue of deterministic automatain the probabilistic setting are the so-called 0/1 automata, in which each wordis either accepted with probability 0 or with probability 1. It is known thatalmost-sure PBA can be transformed into equivalent 0/1 PBA (see the proof ofTheorem 4.13 in [4]). Concerning weak automata, a language can be acceptedby a deterministic weak automaton (DWA) if, and only if, it can be accepted bya deterministic Büchi and by a deterministic co-Büchi automaton (this followsfrom results in [14], see [6] for a more direct construction). We show an analogousresult in the probabilistic setting: The class of languages defined by 0/1 PWAcorresponds to the intersection of the two classes defined by PWA with almost-sure semantics and with positive semantics, respectively. It turns out that thisclass contains only regular languages, that is, 0/1 PWA define the same class asDWA.We also show that the regularity problem for PBA is undecidable (the prob-lem of deciding for a given PBA whether its language is regular). For PBAwith positive semantics this is not surprising, as for those already the emptinessproblem is undecidable [4]. However, for PBA with almost-sure semantics theemptiness and universality problems are decidable [1,2,8]. We show that regular-ity is undecidable already for PWA with almost-sure or with positive semantics.The proof also yields that it is undecidable for a fixed regular language whethera given PWA accepts this language.This work is organized as follows. After introducing basic notations in Sec-tion 2 we first characterize various regular subclasses of PBA that we derivefrom ambiguity patterns in Section 3 and then we derive some related complex-ity results in Section 4. In Section 5 we present our results concerning weakprobabilistic automata and in Section 6 we conclude.
First we briefly review some basic definitions.If Σ is a finite alphabet, then Σ ∗ is the set of all finite and Σ ω is the set ofall infinite words w = w w . . . with w i ∈ Σ . For a word w we denote by w ( i ) the i -th symbol w i .Classical automata used in this work have usually the shape ( Q, Σ, ∆, Q , F ) ,where Q is a finite set of states, Σ a finite alphabet, ∆ ⊆ Q × Σ × Q is the tran-sition relation and Q , F ⊆ Q are the sets of initial and final states, respectively. C. Löding and A. Pirogov
We write ∆ ( p, a ) := { q ∈ Q | ( p, a, q ) ∈ ∆ } to denote the set of successors of p ∈ Q on symbol a ∈ Σ , and ∆ ( P, w ) for P ⊆ Q, w ∈ Σ ∗ with the usualmeaning, i.e., states reachable on word w from any state in P .A run of an automaton on a word w ∈ Σ ω is an infinite sequence of states q , q , . . . starting in some q ∈ Q such that ( q i , w ( i ) , q i +1 ) ∈ ∆ for all i ≥ .We say that a set of runs is separated (at time i ) when the prefixes of length i of those runs are pairwise different.As usual, an automaton is deterministic if | Q | = 1 and | ∆ ( p, a ) | ≤ for all p ∈ Q, a ∈ Σ , and nondeterministic otherwise. For deterministic automata wemay use a transition function δ : Q × Σ → Q instead of a relation.Probabilistic automata we consider have the shape ( Q, Σ, δ, µ , F ) , i.e., thetransition relation is replaced by a function δ : Q × Σ × Q → [0 , which foreach state and symbol assigns a probability distribution on successor states (i.e. P q ∈ Q δ ( p, a, q ) = 1 for all p ∈ Q, a ∈ Σ ), and µ : Q → [0 , with P q ∈ Q µ ( q ) =1 is the initial probability distribution on states. The support of a distribution µ is the set supp ( µ ) := { x | µ ( x ) > } . Similarly as above, we may write δ ( µ, w ) and mean the resulting probability distribution after reading w ∈ Σ ∗ , whenstarting with probability distribution µ .For a probabilistic automaton A the underlying automaton A ⊳ is given byrecovering the transition relation ∆ := { ( p, x, q ) | δ ( p, x, q ) > } of positivelyreachable states and the initial state set Q := supp ( µ ) .As usual, a run of an automaton for finite words is accepting if it ends ina final state. For automata on infinite words, run acceptance is determined bythe Büchi (run visits infinitely many final states) or Co-Büchi (run visits finitelymany final states) conditions.We write p x → q if there exists a path from p to q labelled by x ∈ Σ + and p → q if there exists some x such that p x → q . The strongly connected component(SCC) of p ∈ Q is scc ( p ) := { q ∈ Q | p = q or p → q and q → p } . The set SCCs ( A ) := { scc ( q ) | q ∈ Q } is the set of all SCCs and partitions Q . An SCCis accepting ( rejecting ) if all (no) runs that stay there forever are accepting.An SCC is useless if no accepting run can continue from there. An automatonis weak , if the set of final states is a union of its SCCs. In this case, Büchiand Co-Büchi acceptance are equivalent and we treat weak automata as Büchiautomata.A classical automaton is trim if it has no useless SCCs, whereas a probabilisticautomaton is trim if it has at most one useless SCC, which is a rejecting sinkthat we canonically call q rej . We assume w.l.o.g. that all considered automataare trim, which also means that in an underlying automaton the sink q rej isremoved.We call transitions of probabilistic automata that have probability 1 deter-ministic and otherwise branching . If there are transitions p a → q and p a → q ′ with q = q ′ , we call this pattern a fork . Every branching transition clearly has at leastone fork. We call a ( p, q, q ′ ) fork intra-SCC , if p, q, q ′ are all in the same SCC,otherwise it is an inter-SCC fork. A run of an automaton is deterministic if itnever goes through forks, and limit-deterministic if it goes only through finitely mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 5 many forks. We say that two deterministic runs merge when they reach the samestate simultaneously. For a finite run prefix ρ , we call all valid runs with thisprefix continuations of ρ .A classical automaton A accepts w ∈ Σ ω if there exists an accepting run on w , and the language L ( A ) recognized by A is the set of all accepted words. If P is a set of states of an automaton, we write L ( P ) for the language acceptedby this automaton with initial state set P . For sets consisting of one state q , wewrite L ( q ) instead of L ( { q } ) .For a probabilistic automaton A and an input word w (finite or infinite),the transition structure of A induces a probability space on the set of runs of A on w in the usual way. We do not provide the details here but rather referthe reader not familiar with these concepts to [4]. In general, we write Pr ( E ) forthe probability of a measurable event E in a probability space. For probabilisticautomata, we consider positive , almost-sure and threshold semantics, i.e., anautomaton accepts w if the probability of the set of accepting runs on w is > , =1 or >λ (for some fixed λ ∈ ]0 , ), respectively. For an automaton A theselanguages are denoted by L > ( A ) , L =1 ( A ) and L >λ ( A ) , respectively, whereas L ( A ) := L ( A ⊳ ) is the language of the underlying automaton. A probabilisticautomaton is 0/1 if all words are accepted with either probability 0 or 1 (in thiscase the languages with the different probabilistic semantics coincide).To denote the type of an automaton, we use abbreviations of the form XYA ( γ ) where the type of transition structure is denoted by X ∈ { D (det.), N (nondet.),P (prob.) } , the acceptance condition is specified by Y ∈ { F (finite word),B (Büchi), C (Co-Büchi), W (Weak) } , and for probabilistic transitions thesemantics for acceptance is given by γ ∈ { >0,=1,> λ, / } .By L ( γ ) ( XYA ) we denote the whole class of languages accepted by the cor-responding type of automaton. If L is a set of languages, then L denotes theset of all complement languages (similarly, for a language L , we denote by L itscomplement), and BCl ( L ) the set of all finite boolean combinations of languagesin L . We use the notion of regular language for finite words and for infinite words(the type of words is always clear from the context). Ambiguity of automata refers to the number of different accepting runs on aword or on all words. An automaton is finitely ambiguous (on w ) if there areat most k different accepting runs (on w ) for some fixed k ∈ N , and in case ofat most one accepting run it is called unambiguous . If on each word there areonly finitely many accepting runs, but no constant upper bound over all words,then it is polynomially ambiguous if the number of different run prefixes thatare possible for any word prefix of length n can be bounded by a polynomial in n , and otherwise exponentially ambiguous . Finally, if if there exist words thathave infinitely many runs, but no word on which there are uncountably manyaccepting runs, then it is countably ambiguous , and otherwise it is uncountablyambiguous . C. Löding and A. Pirogov
In [16] (see also [19]), a syntactic characterization of those classes is presentedfor NBA by simple patterns of states and transitions. We define those patternshere and refer to [16] for further details. An automaton A has an IDA pattern if there exist two states p = q and a word v ∈ Σ ∗ such that p v → p , p v → q and q v → q . If additionally q ∈ F , then this is also an IDA F pattern. Finally, A has an EDA pattern if there exists a state p and v ∈ Σ ∗ such that thereare two different paths p v → p , and if additionally p ∈ F , this is also an EDA F pattern. If a PBA has no EDA pattern, we call it flat , reflecting the namingof a similar concept in other kinds of transition systems (e.g. [15]). The namesIDA and EDA abbreviate “infinite/exponential degree of ambiguity”, which theyindicated in the original NFA setting, and we keep those names for consistency.By k - NBA , n k - NBA , n - NBA , ℵ - NBA we denote the subsets of at mostfinitely, polynomially, exponentially and countably ambiguous NBA (and sim-ilarly for other types of automata). When speaking about ambiguity of somePBA A , we mean the ambiguity of the trimmed underlying NBA A ⊳ .In [8], hierarchical PBA (HPBA) were identified as a syntactic restrictionon PBA which ensures regularity under positive and almost-sure semantics. APBA with a unique initial state is hierarchical, if it admits a ranking on thestates such that at most one successor on a symbol has the same rank, and nosuccessor has a smaller rank. A HPBA has k levels if it can be ranked with only k different values. Simple PBA (SPBA) were introduced in [9] and are restrictedHPBA with two levels such that all accepting states are on level 0. SPBA unamb. ¬ IDA fin. amb. ¬ EDA , ¬ IDA F poly. amb. ¬ IDA F exp. amb. ¬ EDA flat ¬ EDA F countably amb. HPBA L > ( ℵ - PBA ) regular L =1 ( flat PBA ∪ k - PBA ) regular L >λ ( k - PBA ) regular Fig. 1: Illustration of the automata classes with restricted ambiguity as presentedfor NBA in [16], which are characterized by the absence of the state patterns
IDA , IDA F , EDA , and EDA F and their relation to the restricted classes called“Hierarchical PBA” ( HPBA ) [8] and “Simple PBA” (
SPBA ) [9]. We identify classesin this hierarchy which can be seen as extensions “in spirit” of respectively SPBAand HPBA, subsuming them while also preserving their good properties, as e.g.definition by syntactic means, regularity under different semantics and severalcomplexity results. mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 7
First, we show how HPBA relate to the ambiguity hierarchy, which can eas-ily be derived by inspection of the definitions. A visual illustration is given inFigure 1.
Proposition 1 (Relation of HPBA and the ambiguity hierarchy). HPBA ⊂ flat PBA ⊂ ℵ - PBA .2. k - PBA HPBA and
HPBA k - PBA .3.
SPBA ⊂ unambiguous PBA ⊂ k - PBA . Starting from these observations, this work was motivated by the questionwhether the ambiguity restrictions, which were only implicit in HPBA andSPBA, can be used explicitly to get larger classes with good properties. In thefollowing we will positively answer this question.
First, we observe that probabilistic automata can recognize regular languageseven under severe ambiguity restrictions.
Proposition 2.
Let A be a DBA. Then there exists an unambiguous PBA B such that L > ( B ) = L =1 ( B ) = L ( A ) .Proof. As A is a (w.l.o.g. complete) DBA, there exists exactly one run on eachword and all transitions when seen as PBA must have probability 1. Clearly thisunique natural 0/1 PBA obtained from A accepts the same language under bothprobable and almost-sure semantics and it is trivially unambiguous. ⊓⊔ Limit-deterministic NBA (LDBA) are NBA which are deterministic in allnon-rejecting SCCs. The natural mapping of LDBA into PBA [4, Lemma 4.2]already trivially yields countably ambiguous automata (because the determinis-tic part of the LDBA cannot contain an EDA F pattern, which implies uncount-able ambiguity [16]). The following result shows that already unambiguous PBAunder positive semantics suffice for all regular languages. Theorem 1.
Let L ⊆ Σ ω be a regular language.Then there exists an unambiguous PBA B such that L > ( B ) = L .Proof (sketch). Let A = ( Q, Σ, δ, q , c ) be a deterministic parity automatonaccepting L , i.e., a finite automaton with priority function c : Q → { , . . . , m } such that w ∈ L ( A ) iff the smallest priority assigned to a state on the uniquerun of A on w which is seen infinitely often is even.We construct an unambiguous LDBA for L , which then easily yields a PBA > by assigning arbitrary probabilities ([4, Lemma 4.2]) without influencing theambiguity. If the parity automaton A has m priorities, the LDBA B can beobtained by taking m +1 copies, where m of them are responsible for one priorityeach, and one is modified to guess which priority i on the input word is the mostimportant one appearing infinitely often along the run of A , and correspondinglyswitch into the correct copy. This switching is done unambiguously for the firstposition after which no priority more important than i appears. ⊓⊔ C. Löding and A. Pirogov
First we establish a result for flat PBA, i.e. PBA that have no
EDA pattern.In automata without
EDA pattern there are no states which are part of twodifferent cycles labeled by the same finite word. Even though we defined flatPBA by using an ambiguity pattern, the set of flat PBA does not correspond toan ambiguity class, but it is useful for our purposes due to the following property:
Lemma 1. If A is a flat PBA and w ∈ Σ ω , then the probability of a run of A on w to be limit-deterministic is 1.Proof. Let
Runs ( A , w ) denote the set of all runs of A on w and nldRuns ( A , w ) denote the subset containing all such runs that are not limit-deterministic. As A is flat, it has no EDA and thus also no
EDA F pattern, hence A is at mostcountably ambiguous (by [16]). Moreover, there are not only at most countablymany accepting runs on any word, but also countably many rejecting runs (whichcan be seen by a simple generalization of [16, Lemma 4]). But as all runs aredisjoint events, each run ρ that uses infinitely many forks has probability 0, andthe total number of runs is countable, we can see that Pr ( Runs ( A , w ) \ nldRuns ( A , w )) = X ρ ∈ Runs ( A ,w ) Pr ( ρ ) − X ρ ∈ nldRuns ( A ,w ) Pr ( ρ ) = 1 − . ⊓⊔ The following lemma characterizes acceptance of PBA under extremal se-mantics with restricted ambiguity and is crucial for the constructions in thefollowing sections:
Lemma 2 (Characterizations for extremal semantics).
Let A be a PBA.1. If A is at most countably ambiguous, then w ∈ L > ( A ) ⇔ there exists an accepting run on w that is limit-deterministic.2. If there are finitely many accepting runs of A on w , then w ∈ L =1 ( A ) ⇔ all runs on w are accepting and limit-deterministic.3. If A is flat, then w ∈ L =1 ( A ) ⇔ there is no limit-deterministic rejecting run on w .Proof. (1 . ) : For contradiction, assume that every accepting run on w goesthrough forks infinitely often. But then the probability of every individual ac-cepting run on w is 0. Each run is a measurable event (it is a countable intersec-tion of finite prefixes) and clearly disjoint from other runs, as two different runsmust eventually differ after a finite prefix. But as the number of accepting runsis countable by assumption, by σ -additivity it follows that the probability of allaccepting runs is also 0, contradicting the fact that w ∈ L > ( A ) .For the other direction, pick a limit-deterministic accepting run ρ of A on w and let uv = w and q ∈ Q such that the state of ρ after reading u is q and thereare no forks visited on v . Clearly, the probability to be in q after u in a run of A is positive (because u is finite), and the probability that A continues like ρ from q on v is . Hence, the probability of ρ is positive. mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 9 (2 . ) : The ( ⇐ ) direction is obvious. We now proceed to show ( ⇒ ) . Takesome time t after which all accepting runs on w separated. Assume that someaccepting run ρ is not limit-deterministic. But then ρ goes through infinitelymany forks after t which with positive probability lead to a successor from whichthe probability to accept is 0, and the probability of following ρ is also 0. Asthe probability to follow ρ until time t is positive, but after that the probabilityto accept is 0, this implies that there is a positive probability that A rejects w . Therefore, all accepting runs on w must be limit-deterministic. Now assumethat some run ρ on w is rejecting. Following this run until the time at which ρ isseparated from all accepting runs has positive probability and all continuationsmust be also rejecting, so A must reject w . (3 . ) : Clearly ( ⇒ ) holds, because a limit-deterministic rejecting run has pos-itive probability, i.e., if such a run exists on w , then A cannot accept almostsurely. For ( ⇐ ) , observe that because A is flat, we know by Lemma 1 thatwith probability 1 runs are limit-deterministic. Hence, if there exists no limit-deterministic rejecting run on w (which would have positive probability), thenwith probability 1 runs are limit-deterministic and accepting. ⊓⊔ Using these characterizations, we can provide simple constructions from prob-abilistic to classical automata.
Theorem 2.
Let A be a PBA that is at most countably ambiguous.Then L > ( A ) is a regular language.Proof (sketch). An NBA construction taking two copies of the PBA, where inthe first copy no state is accepting and the second copy has no forks, with thepurpose of guessing a limit-deterministic accepting run. ⊓⊔ Corollary 1. If L > ( A ) is not regular, then it contains an EDA F pattern. Theorem 3.
Let A be a PBA that is at most exponentially ambiguous or flat.Then L =1 ( A ) is regular and recognizable by DBA.Proof (sketch). Both cases (exp. ambiguous or flat) shown using a deterministicbreakpoint construction resulting in a DBA. In one case it checks whether allruns are accepting, in the other it checks that there are no limit-deterministicrejecting runs. ⊓⊔ Corollary 2. If L =1 ( A ) is not regular,then A contains both an EDA and an
IDA F pattern. The corollaries above follow directly from the theorems and the syntacticcharacterization of ambiguity classes [16]. The following proposition states thatthese characterizations of regularity in terms of the ambiguity patterns are tight. q a q b q + q $ b : 1 , a : a : a, ba : 1 , b : $ $ $ (b) q q a : 1 - λa : λb a (c) q q q q f a : λba : (1 - λ ) aa : (1 - λ ) a : λb Σ Fig. 2: (a) Some PWA which accepts the non-regular language { w = ( a + b ) ∗ $ ω | a ( w ) > b ( w ) } with a threshold of , where x ( w ) denotes the number ofoccurrences of x ∈ Σ in w ∈ Σ ω . (b) A family of PBA P λ from [4] such that L > ( P λ ) is not regular for any λ ∈ R . (c) A family of PWA ˜ P λ (closely relatedto [4, Fig. 6]) such that L =1 ( ˜ P λ ) is not regular for any λ ∈ R . Proposition 3.
There exist PBA...1. ...with
EDA F pattern (i.e. uncountably ambiguous) that acceptnon-regular languages under positive semantics.2. ...with no EDA F pattern (i.e. countably ambiguous) that acceptnon-regular languages under almost-sure semantics.Proof. (1.) Note that this statement just means that there are PBA acceptingnon-regular languages, which is well known. For example, the automata familyfrom [4, Fig. 3], depicted in Figure 2(b), accepts non-regular languages underpositive semantics and clearly contains an EDA F pattern, e.g. there are twodifferent paths from p to p on the word aab .(2.) The automata family depicted in Figure 2(c) is a simple modificationof the PBA family depicted in [4, Fig. 6] and recognizes the same non-regularlanguages under almost-sure semantics. It does not contain an EDA F pattern,because the accepting state is a sink, but it does contain an IDA F and an EDA pattern (both e.g. on aab ), so it is countably ambiguous and not flat. ⊓⊔ This completes our classification of regular subclasses of PBA under extremalsemantics that are defined by ambiguity patterns, showing that going beyond therestricted classes presented above (by allowing more patterns) in general leadsto a loss of regularity.Notice that the presented constructions do not track exact probabilities, justwhether transitions have a probability > or = 1 . This is a noteworthy obser-vation, as in general, the probabilities do matter for PBA, as shown in [4, Thm.4.7, Thm. 4.11]. Proposition 4.
Let A be a PBA. The exact probabilities in A do not influ-ence L > ( A ) if A is at most countably ambiguous, and L =1 ( A ) if A is at mostexponentially ambiguous or flat. mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 11 In this section we consider PBA under threshold semantics and we will see thatin this setting, we lose regularity much earlier than in the case of extremalsemantics, but there is still the large and natural subclass of finitely ambiguousPBA that retains regularity. Before we can show this, we need to derive a suitablecharacterization of such languages.We derive it from the following simple observation, which was also usedmore implicitly in the proof that Simple HPBA with threshold semantics areequivalent to DBA in [9].
Lemma 3.
Let A be a PBA. Then for every threshold λ ∈ ]0 , , there exists afinite set of probability values V ≥ λ ⊂ [ λ, such that for every finite run prefixwith probability v in A we have v ≥ λ ⇒ v ∈ V ≥ λ .Proof. Observe that given a finite set of real numbers R ⊂ [0 , , the set R ≥ λ := { r | r = Q i r i ≥ λ, r i ∈ R } must be finite, as in any sequence p p . . . of p i ∈ R ,only at most m = ⌈ log λ (max R ) ⌉ values can be < and such that the product ofthe sequence remains ≥ λ . In our case, let R be the set of distinct probabilitiesassigned to edges (including the initial edges) in A . As every finite run prefixby definition has the probability given by the product of the edge probabilities,this implies the statement. ⊓⊔ If there is just one accepting run (i.e., the automaton is unambiguous), onecan easily construct a nondeterministic automaton that guesses an accepting runand tracks it along with its probability value, of which there are only finitelymany above the threshold. In the case that there are multiple accepting runs, foracceptance only the sum of their probabilities matters. As individual runs canin principle have arbitrarily small probability values, it is not obvious that thesame approach (tracking a set of runs) can work. Determining a suitable cut-offpoint is not as simple, because it is not apparent when a single run becomesso improbable that it does not matter among the others. However, we will nowshow that such a cut-off point must exist:
Lemma 4.
Let A be a PBA, λ ∈ ]0 , a threshold and k ∈ N . There exists ε k ∈ ]0 , λ ] such that for all sets R t = { ρ ti } ji =1 of at most j ≤ k different runprefixes in A of the same length t ∈ N , Pr ( R t ) = P ji =1 Pr ( ρ ti ) < λ implies that Pr ( R t ) < λ − ε k .Proof. We prove this by induction on the number of runs k . For k = 1 , i.e. asingle run prefix, let V ≥ λ be the finite (by Lemma 3) set of different probabilityvalues ≥ λ and let E be the set of distinct probabilities in the automaton A .Then clearly v max ,<λ := max { a · b | a · b < λ, a ∈ V ≥ λ , b ∈ E } is the largestprobability value < λ that can correspond to a finite run prefix in A . Hence, wecan just pick an ε < λ − v max ,<λ and immediately get that for any run prefixwith probability v < λ , we have that v ≤ v max ,<λ < λ − ε .Now assume the statement holds for all sets with at most k run prefixes.Let R t be a set of k + 1 of different run prefixes of the same length such that Pr ( R t ) < λ and let ε := ε k . Then we know that for every subset S of at most k runs of R t we have Pr ( S ) < λ − ε . Also, every single run prefix can by Lemma 3have one of only finitely many probability values in V ≥ ε that are ≥ ε and thereexists a value v max ,<ε denoting the largest possible probability value < ε that asingle run prefix can have.If there exists a run prefix ρ ∈ R t with probability value v < ε , then weknow that Pr ( R t ) = Pr ( R t \ { ρ } ) + v < ( λ − ε ) + v max ,<ε < λ . If every run in R t has a probability value ≥ ε , then every run prefix in R t has as probabilityone of the values in V ≥ ε . Consider all sums of k values from V ≥ ε , which arefinitely many, and pick the largest sum s which is < λ . Choose ε k +1 such that ε k +1 < min( ε − v max ,<ε , λ − s ) to account for both cases. ⊓⊔ From this we can derive the following characterization of languages acceptedby finitely ambiguous PBA under threshold semantics:
Lemma 5.
Let A be a k -ambiguous PBA and λ ∈ ]0 , a threshold. There existsan ε ∈ ]0 , λ ] such that for all w ∈ Σ ω : w ∈ L >λ ( A ) iff there exists a set R oflimit-deterministic accepting runs of A on w with Pr ( R ) > λ , Pr ( S ) ≤ λ for all S ⊂ R and at most one run ρ ∈ R with Pr ( ρ ) < ε .Proof. Clearly ( ⇐ ) holds, as then w is accepted with probability ≥ Pr ( R ) > λ .We now show ( ⇒ ) . In a finitely ambiguous PBA there are only finitely manydifferent accepting runs on each word. Furthermore, as after finite time all ac-cepting runs have separated and each accepting run that visits forks infinitelyoften has probability 0, accepting runs that visit forks infinitely often do not con-tribute positively to the acceptance probability and thus can be ignored. Hence,if w ∈ L >λ ( A ) , there is a number of accepting runs that eventually all becomedeterministic and each such run has a positive probability, which must in totalbe > λ .Let R be a set of different limit-deterministic accepting runs of A on w suchthat Pr ( R ) > λ and Pr ( S ) ≤ λ for all S ⊂ R . As there are only finitely manyaccepting runs, such a set R must exist. Furthermore, notice that each limit-deterministic run has a finite prefix which has the same probability as the wholerun, so there exists a time t such that the probability of the set of all differentprefixes of runs in R of length t is exactly Pr ( R ) , so that Lemma 4 applies.Now pick an ε := ε k given by Lemma 4. We claim that at most one run ρ ∈ R can have a probability less than ε . If there is no such run in R , we aredone. Otherwise let ρ be a run with Pr ( ρ ) =: p < ε and notice that by choiceof R , we have that Pr ( R \ { ρ } ) =: s ≤ λ . It cannot be the case that s < λ , asthen by Lemma 4 we have s < λ − ε , which implies that Pr ( R ) = s + p < λ ,which is a contradiction. Hence, now assume that s = λ . But then, if there isany ρ ′ = ρ ∈ R such that Pr ( ρ ′ ) =: p ′ < ε , by the same argument we get thecontradiction that s − p ′ < λ − ε and hence s < λ . Therefore, no other run in R can have a probability < ε . ⊓⊔ Now we can perform the intended automaton construction to show:
Theorem 4. L >λ ( A ) is regular for each k -ambiguous PBA A and λ ∈ ]0 , . mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 13 Proof (sketch).
We use the characterization of Lemma 5 to construct a gener-alized Büchi automaton accepting L >λ ( A ) . Intuitively, the new automaton justguesses at most k different runs of A and verifies that the guessed runs are limit-deterministic and accepting. The automaton additionally tracks the probabilityof the runs over time, to determine whether the individual runs and their sumhave enough “weight”. The automaton rejects when the total probability of theguessed runs is ≤ λ , one of the runs goes into the rejecting sink q rej or a rundoes not see accepting states infinitely often.By Lemma 5 we only need to consider sets of runs with at most one run thathas a probability < ε , where ε := ε k is given by Lemma 4. For this single runwe also do not need to track the exact probability value, as its only purpose isto witness that the acceptance probability is strictly greater than λ , whereas allother runs must have one of the finitely many different probabilities which are ≥ ε and must sum to λ . ⊓⊔ This generalizes the corresponding result for PFA [12, Theorem 3]. The proofin [12] uses similar concepts, though a rather different presentation. In the settingof infinite words we additionally have to deal with a single run that has arbitrarilylow probability, and we have to ensure that this probability remains positive.After seeing that finitely ambiguous PBA retain regularity, we show that thisis the best we can do under threshold semantics:
Corollary 3.
There are polynomially ambiguous PBA A , that is, with an IDA pattern and no
EDA , IDA F patterns, such that L >λ ( A ) is not regular even forrational thresholds λ ∈ ]0 , .Proof. Follows from the fact that the PWA A from Figure 2(a), which recognizesa non-regular language (and is used to show Proposition 6), has just an IDA pattern in the underlying NBA, but no
EDA or IDA F patterns. ⊓⊔ This completes our characterization of languages which are recognized byPBA that are restricted by forbidden ambiguity patterns, so that we can stateour main result of this section (see Figure 1 for a visualization):
Theorem 5.
The following results hold about PBA with restricted ambiguity: – L > ( k - PBA ) = L > ( ℵ - PBA ) = L ( NBA ) – L =1 ( k - PBA ) = L =1 (2 k - PBA ) = L =1 ( flat PBA ) = L ( DBA ) ⊂ L =1 ( ℵ - PBA ) – L >λ ( k - PBA ) = L ( NBA ) ⊂ L >λ ( n k - PBA ) Proof.
The statements follow from the following inclusion chains: L ( NBA ) (1 . ) ⊆ L > ( k - PBA ) def. ⊆ L > ( ℵ - PBA ) (2 . ) ⊆ L ( NBA ) L ( DBA ) (3 . ) ⊆ L =1 ( k - PBA ) def. ⊆ L =1 (2 k - PBA ∪ flat PBA ) (4 . ) ⊆ L ( DBA ) (5 . ) ⊂ L =1 ( ℵ - PBA ) L ( NBA ) (1 . ) ⊆ L > ( k - PBA ) (6 . ) ⊆ L >λ ( k - PBA ) (7 . ) ⊆ L ( NBA ) (8 . ) ⊂ L >λ ( n k - PBA ) Where the marked relationships hold due to: (1.) Theorem 1, (2.) Theorem 2,(3.) Proposition 2, (4.) Theorem 3, (5.) Proposition 3, (6.) Simple transformationby adding a new accepting sink q acc and modifying the initial distribution µ [4,Lemma 4.16], (7.) Theorem 4, (8.) Corollary 3, and (def.) by definition of theambiguity-restricted automata classes. ⊓⊔ In this section, we state some upper and lower bounds on the complexity fordeciding emptiness and universality for PBA with restricted ambiguity, derivedfrom the characterizations and constructions presented above.
Theorem 6.
1. the emptiness problem for ℵ -PBA > is in NL
2. the universality problem for ℵ -PBA > is in PSPACE
3. the universality problem for at most exp. ambiguous or flat PBA =1 is in NL Proof. (1 . + 2 . ) : By Theorem 2 the languages of ℵ -PBA > are regular. Theconstruction of an NBA just uses two copies of the given PBA. For emptiness,it thus suffices to guess an accepted ultimately periodic word and verify that itis accepted by the NBA, which can be done in NL. Since universality for NBAin in PSPACE [21], we also obtain (2.). (3 . ) : If the automaton is at most exponentially ambiguous, there are onlyfinitely many accepting runs on each word and as we know by Lemma 2 that w ∈ L =1 ( A ) iff all runs are accepting, it suffices to guess a rejecting run in A ⊳ , which implies that the ultimately periodic word w labelling that run cannot be in L =1 ( A ) . If the automaton is flat, then we know that for each rejectedword there must exist a limit-deterministic rejecting run in the underlying NBA,which we also can guess. ⊓⊔ Type regular? Emptiness Universality > > λ > > k - PBA ✓ ∈ NL ∈ PSPACE ∈ PSPACE ∈ NL n k - PBA ✗ n - PBA flat
PBA ∈ NL c. ∈ PSPACE c. ∈ PSPACE c. ∈ NL c. ℵ - PBA ∈ PSPACE
Table 1: Summary of main results from Theorems 5 and 6 concerning PBA withambiguity restrictions. The completeness results follow from the hardness resultsfor HPBA (which are subsumed by flat PBA) from [8, Section 5], the
PSPACE inclusion of universality for almost-sure ℵ - PBA follows from [8, Theorem 4.4].Observe that ℵ -PBA > subsume HPBA > and the union of flat PBA =1 andexp. ambiguous PBA =1 subsumes HPBA =1 , while preserving the same complex-ity of the emptiness and universality problems. A summary of the main resultsfrom Theorem 5 and Theorem 6 is presented in Table 1. mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 15 We conclude with an observation relevant to the question about feasibilityof PBA with restricted ambiguity for the purpose of application in e.g. model-checking or synthesis.
Proposition 5 (Relationship to classical formalisms).–
There is a doubly-exponential lower bound for translation from LTL formulato countably ambiguous PBA with positive semantics. – There is an exponential lower bound for conversion from NBA to countablyambiguous PBA with positive semantics.Proof.
It is known [20, Theorem 2] that there is a doubly-exponential lowerbound from LTL to LDBA. It is also known that LTL to NBA has an exponentiallower bound (e.g. [5, Theorem 5.42]), which implies an exponential lower boundfrom NBA to LDBA.By Theorem 2 there is a polynomial transformation from countably ambigu-ous PBA with positive semantics into LDBA, which together with the aforemen-tioned bounds implies the claimed lower bounds. ⊓⊔ In this section we investigate the class of probabilistic weak automata (PWA),establishing the relation between different classes defined by PWA as shown inFigure 3 (see also the description of our contribution in the introduction).As a first remark, notice that PWA can be “complemented” by invertingaccepting and rejecting states and switching between dual semantics, e.g., for aPWA A we have L > ( A ) = L =1 ( A ) , where A is just A with inverted acceptingstate set F ′ = Q \ F .Since the overarching theme of this paper is trying to find regular subclassesof PBA, we will next establish the following result, showing that there is no hopeto find a complete syntactical characterization of regularity in PBA: Theorem 7.
The regularity of PWA (and therefore of PBA) under positive,almost-sure and threshold semantics is an undecidable problem.Proof (sketch).
Since L >λ ( PWA ) ⊇ L > ( PWA ) (see Theorem 10), L > ( PWA ) = L =1 ( PWA ) , and the class of regular ω -languages is closed under complement, itsuffices to show the statement for PWA =1 . We do this by reduction from thevalue 1 problem for PFA, which is the question whether for each ε > there ex-ists a word accepted by the PFA with probability > − ε . This problem is knownto be undecidable [13]. We consider a slightly modified version of the problemby assuming that no word is accepted with probability 1 by the given PFA. Theproblem remains undecidable under this assumption, because one can check if aPFA accepts a finite word with probability 1 by a simple subset construction.Given some PFA A , we construct a PWA =1 B by taking a copy of A andextending it with a new symbol such that from accepting states of A theautomaton is “restarted” on , while from non-accepting states leads into a L >λ ( PBA ) L >λ ( PWA ) L = ( P B A ) L = ( P W A ) L = ( P B A ) L > ( P W A ) PWA / L =1 ( PBA ) ∩ L =1 ( PBA ) L > ( PBA ) ω − Reg L ( D B A ) L ( D B A ) L ( DBA ) ∩ L ( DBA ) L ( DWA ) Fig. 3: Illustration of relationships between the class of languages acceptedby weak probabilistic automata under various semantics with other alreadyknown classes. The overlapping patterns indicate intersection of classes, wheredots mark L > ( PBA ) , and different diagonal lines respectively L =1 ( PBA ) and L =1 ( PBA ) . The dashed line indicates intersections with different subclasses ofregular languages. The class L >λ ( PBA ) contains all the other depicted classes, L >λ ( PWA ) contains the area inside the thick line. The depicted fact that L > ( PWA ) = L >λ ( PWA ) ∩ L > ( PBA ) is a conjecture, one direction is shownin Theorem 10.new part which ensures that infinitely many are seen and contains the onlyaccepting state of B . We show that L =1 ( B ) = ( Σ ∗ ω \ R , where R = ∅ if A does not have value 1, and R is non-empty but does not contain an ultimatelyperiodic word, otherwise. This implies that L =1 ( B ) is regular iff A does not havevalue 1. ⊓⊔ We will now show that PWA with almost-sure semantics are as expressive asPBA, and with positive semantics as expressive as PCA.
Theorem 8. L > ( PWA ) = L > ( PCA ) and L =1 ( PWA ) = L =1 ( PBA ) .Proof (sketch). It suffices to show the first statement. The second then fol-lows by duality, i.e., we can interpret a PBA =1 A recognizing L as a PCA > recognizing L and just apply the construction to get a PWA > B for L , suchthat B (with inverted accepting and rejecting states) is a PWA =1 for L . Inthe first statement the ⊆ inclusion is trivial, hence we only need to show that L > ( PCA ) ⊆ L > ( PWA ) .We construct a PWA > consisting of two copies of the original PCA > , a guess copy and a verify copy. In the first copy, the automaton can guess thatno final states will be visited anymore and switch to the verify copy, which isaccepting, but where all transitions into final states are redirected to a rejectingsink. ⊓⊔ Next, we show that languages that can be accepted by both, a PWA withalmost-sure semantics, and by a PWA with positive semantics, are regular and mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 17 can be accepted by a DWA. For the proof, we rely on a characterization ofDWA languages in terms of the Myhill-Nerode equivalence relation from [22]. Sowe first define this equivalence, and show that languages defined by PBA withpositive semantics have only finitely many equivalence classes. Then we comeback to the result for PWA.For L ⊆ Σ ω , define the Myhill-Nerode equivalence relation ∼ L ⊆ Σ ∗ × Σ ∗ by u ∼ L v iff uw ∈ L ⇔ vw ∈ L for all w ∈ Σ ω . Then the following holds: Lemma 6 (Finitely many Myhill-Nerode classes).
Languages in L > ( PBA ) have finitely many Myhill-Nerode equivalence classes.Proof. Let A = ( Q, Σ, δ, µ , F ) be some PBA > and u ∈ Σ ∗ some word and let µ u := δ ∗ ( µ , u ) be the probability distribution on states of A after reading u .Pick any w ∈ Σ ω and notice that uw ∈ L = L > ( A ) iff there exists some state q such that µ u ( q ) > and the probability to accept w from q is also > , as theproduct of two positive numbers clearly still is positive. But then, for any two u, v ∈ Σ ∗ we have that whenever µ u ( q ) > ⇔ µ v ( q ) > for all q , then we have uw ∈ L ⇔ vw ∈ L for all w ∈ Σ ω by the reasoning above, as the exact valuedoes not matter for acceptance, and therefore u ∼ L v . But as there are only atmost | Q | different possibilities how values in a distribution µ over Q are eitherequal to or greater than , this is an upper bound on the number of differentequivalence classes. ⊓⊔ Theorem 9. L > ( PWA ) ∩ L =1 ( PWA ) = L ( DWA ) = L ( PWA / ) Proof.
The inclusions L ( DWA ) ⊆ L ( PWA / ) ⊆ L > ( PWA ) ∩ L =1 ( PWA ) aretrivial, hence it remains to show L > ( PWA ) ∩ L =1 ( PWA ) ⊆ L ( DWA ) .So let L be a language from L > ( PWA ) ∩ L =1 ( PWA ) . We want to show that L can be accepted by a DWA. We use the following characterization of DWAlanguages [22, Theorem 21]: The DWA languages are precisely the languages withfinitely many Myhill-Nerode classes in the class G δ ∩ F σ in the Borel hierarchy.The classes G δ and F σ of the Borel hierarchy are often also referred to as Π and Σ . We do not introduce the details of this hierarchy here, but rather referthe reader not familiar with these concepts to [22] and [8].We already know that L has finitely many Myhill-Nerode classes by Lemma 6(as PWA are special cases of PBA). It remains to show that L is in the class G δ ∩ F σ . It is known that PBA with almost-sure semantics define languagesin G δ [8, Lemma 3.2]. Hence L is in G δ . Since L is accepted by a PWA withpositive semantics, the complement of L is accepted by a PWA with almost-sure semantics (as noted at the beginning of this section). We obtain that thecomplement of L is also in G δ again by [8, Lemma 3.2]. This means that L is in F σ , which by definition consists of the complements of languages from G δ . ⊓⊔ Concluding this section, we show a result about weak automata with thresh-old semantics, which (not surprisingly) turn out to be even more expressive. Acareful analysis of the PWA A in Fig. 2(a) shows the following result: Proposition 6.
For all thresholds λ ∈ ]0 , there exists a PWA A such that L >λ ( A ) is not regular and not P BA > recognizable. Putting things together, we can say the following about threshold PWA,establishing the relation of L >λ ( PWA ) to the other classes in Figure 3: Theorem 10 (Expressive power of threshold PWA). L > ( PWA ) ⊆ L >λ ( PWA ) ∩ L > ( PBA ) .2. L >λ ( PWA ) and L > ( PBA ) are incomparable (wrt. set inclusion).3. L > ( PWA ) ⊂ L >λ ( PWA ) ⊂ L >λ ( PBA ) .Proof. (1.) L > ( PWA ) ⊆ L > ( PBA ) by definition and L > ( PWA ) ⊆ L >λ ( PWA ) ,as any PWA > can be modified to a PWA >λ recognizing the same language byjust adding an additional accepting sink and modifying the initial distribution,just as described in [4, Lemma 4.16] for general PBA.(2.) By Proposition 6, there are languages recognized by PWA >λ that cannotbe recognized with PBA > . To show that there are languages accepted by PBA > that cannot be accepted by PWA >λ we can give a topological characterizationof languages accepted by PWA by a simple adaptation of [8, Lemma 3.2] andcombine it with other results shown in [8] to show that there are PBA > thataccept languages that cannot be accepted by PWA >λ .(3.) The first inclusion was discussed in (1.), the strictness follows fromProposition 6 and the fact that L > ( PWA ) = L =1 ( PBA ) ⊂ BCl ( L =1 ( PBA )) = L > ( PBA ) , where the first equality is Theorem 8 and the second is shown in [8].The second inclusion of the statement follows from (2.) and the fact from [4]that L > ( PBA ) ⊂ L >λ ( PBA ) . ⊓⊔ For the dual class L ≥ λ ( PWA ) one can show symmetric results that correspondto statements (1.) and (2.) above, for statement (3.) however there is no proofyet for the strictness of the inclusions (especially the second one), whereas thestatement L =1 ( PWA ) ⊆ L ≥ λ ( PWA ) ⊆ L ≥ λ ( PBA ) is obvious. We leave this issueas an open question. Another interesting question is whether > λ is equivalentto < λ (or dually for ≥ / ≤ ). By using notions from ambiguity in classical Büchi automata, we were able toextend the set of easily (syntactically) checkable PBA which are regular undersome or all of the usual semantics. As a consequence, ambiguity appears tobe an even more interesting notion in the probabilistic setting, as here it infact has consequences for the expressive power of automata, whereas in theclassical setting there is no such effect. Our results also indicate that to getnon-regularity, one requires the use of certain structural patterns which at leastimply the existence of the ambiguity patterns that we used. It is an open questionwhether it is possible to identify more fine-grained syntactic characterizations,patterns or easily checkable properties which are just over-approximated by theambiguity patterns and are required for non-regularity. mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 19
References
1. Baier, C., Bertrand, N., Größer, M.: On decision problems for proba-bilistic büchi automata. In: Foundations of Software Science and Com-putational Structures, 11th International Conference, FOSSACS 2008. Lec-ture Notes in Computer Science, vol. 4962, pp. 287–301. Springer (2008),https://doi.org/10.1007/978-3-540-78499-92. Baier, C., Bertrand, N., Größer, M.: Probabilistic automata over infinite words:Expressiveness, efficiency, and decidability. In: Proceedings Eleventh InternationalWorkshop on Descriptional Complexity of Formal Systems, DCFS 2009. EPTCS,vol. 3, pp. 3–16 (2009), https://doi.org/10.4204/EPTCS.33. Baier, C., Größer, M.: Recognizing omega-regular languages with probabilistic au-tomata. In: 20th IEEE Symposium on Logic in Computer Science (LICS 2005),26-29 June 2005, Chicago, IL, USA, Proceedings. pp. 137–146 (2005)4. Baier, C., Größer, M., Bertrand, N.: Probabilistic ω -automata. Journal of the ACM(JACM) (1), 1 (2012)5. Baier, C., Katoen, J.: Principles of model checking. MIT Press (2008)6. Boigelot, B., Jodogne, S., Wolper, P.: An effective decision procedure for lineararithmetic over the integers and reals. ACM Trans. Comput. Log. (3), 614–633(2005), https://doi.org/10.1145/1071596.10716017. Büchi, J.R.: On a decision method in restricted second order arithmetic. In: Studiesin Logic and the Foundations of Mathematics, vol. 44, pp. 1–11. Elsevier (1966)8. Chadha, R., Sistla, A.P., Viswanathan, M.: Power of randomization in automataon infinite strings. Logical Methods in Computer Science (2011)9. Chadha, R., Sistla, A.P., Viswanathan, M.: Probabilistic Büchi automata withnon-extremal acceptance thresholds. In: International Workshop on Verification,Model Checking, and Abstract Interpretation. pp. 103–117. Springer (2011)10. Chadha, R., Sistla, A.P., Viswanathan, M.: Emptiness under isolation and thevalue problem for hierarchical probabilistic automata. In: FOSSACS 2017. LNCS,vol. 10203, pp. 231–247 (2017), https://doi.org/10.1007/978-3-662-54458-711. Chadha, R., Sistla, A.P., Viswanathan, M., Ben, Y.: Decidable and expressiveclasses of probabilistic automata. In: FoSSaCS 2015. LNCS, vol. 9034, pp. 200–214. Springer (2015), https://doi.org/10.1007/978-3-662-46678-012. Fijalkow, N., Riveros, C., Worrell, J.: Probabilistic automata of bounded ambigu-ity. In: 28th International Conference on Concurrency Theory (CONCUR 2017).Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik (2017)13. Gimbert, H., Oualhadj, Y.: Probabilistic automata on finite words: Decidable andundecidable problems. In: International Colloquium on Automata, Languages, andProgramming. pp. 527–538. Springer (2010)14. Landweber, L.H.: Decision problems for ω -automata. Mathematical Systems The-ory , 376–384 (1969)15. Leroux, J., Sutre, G.: On flatness for 2-dimensional vector addition systems withstates. In: International Conference on Concurrency Theory. pp. 402–416. Springer(2004)16. Löding, C., Pirogov, A.: On finitely ambiguous Büchi automata. In: Developmentsin Language Theory - 22nd International Conference, DLT 2018, Tokyo, Japan,September 10-14, 2018, Proceedings. pp. 503–515 (2018)17. Löding, C., Thomas, W.: Alternating automata and logics over infinite words.In: Proceedings of the IFIP International Conference on Theoretical ComputerScience, IFIP TCS2000. LNCS, vol. 1872, pp. 521–535. Springer (2000)0 C. Löding and A. Pirogov18. Rabin, M.O.: Probabilistic automata. Information and control (3), 230–245 (1963)19. Rabinovich, A.: Complementation of finitely ambiguous Büchi automata. In: Devel-opments in Language Theory - 22nd International Conference, DLT 2018, Tokyo,Japan, September 10-14, 2018, Proceedings. pp. 541–552 (2018)20. Sickert, S., Esparza, J., Jaax, S., Křetínský, J.: Limit-deterministic Büchi automatafor linear temporal logic. In: Chaudhuri, S., Farzan, A. (eds.) Computer AidedVerification. pp. 312–332. Springer International Publishing, Cham (2016)21. Sistla, A.P., Vardi, M.Y., Wolper, P.: The complementation problemfor Büchi automata with applications to temporal logic (extended ab-stract). In: ICALP 1985. LNCS, vol. 194, pp. 465–474. Springer (1985),https://doi.org/10.1007/BFb001572522. Staiger, L.: Finite-state ω -languages. Journal of Computer and System Sciences (3), 434–448 (1983)23. Thomas, W.: Automata on infinite objects. In: Handbook of Theoretical Com-puter Science, vol. B: Formal Models and Semantics, pp. 133–192. Elsevier SciencePublishers, Amsterdam (1990)24. Thomas, W.: Languages, automata, and logic. In: Rozenberg, G., Salomaa, A.(eds.) Handbook of Formal Language Theory, vol. III, pp. 389–455. Springer (1997)25. Weber, A., Seidl, H.: On the degree of ambiguity of finite automata. TheoreticalComputer Science (2), 325–349 (1991)mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 21 A Proofs for section on ambiguity in PBA
A.1 Proof for Proposition 1Proposition 1 (Relation of HPBA and the ambiguity hierarchy). HPBA ⊂ flat PBA ⊂ ℵ - PBA .2. k - PBA HPBA and
HPBA k - PBA .3.
SPBA ⊂ unambiguous PBA ⊂ k - PBA .Proof.
1. First, observe that all states in the same SCC of a HPBA must have thesame rank, as otherwise the SCC contains a path where the ranks of thestates strictly decrease. Existence of an EDA pattern implies that there is atleast one intra-SCC fork, which implies that two successors must have thesame rank, which is forbidden for HPBA.On the other hand, it is easy to construct an automaton that has no EDApattern, but is not a valid HPBA because it has an intra-SCC fork. The sec-ond inclusion follows trivially because there are automata with EDA patternbut no EDA F pattern, and thus are at most countably ambiguous (by [16]).2. Finitely ambiguous automata have no IDA (and thus no EDA) patterns (by[16]), but even unambiguous PBA may contain an intra-SCC fork, meaningthat it cannot be a HPBA. On the other hand, HPBA may even have anIDA F pattern, which then implies infinite ambiguity.3. Clearly, level 1 of SPBA can be thought of a rejecting sink, as no acceptingstates are reachable. As in the trimmed automaton there is just one (useful)SCC containing states on level 0 and there are no intra-SCC forks in HPBA,transitions within level 0 are deterministic. Hence SPBA are trivially unam-biguous, which by definition is a strict subset of finitely ambiguous PBA. ⊓⊔ A.2 Proof for Theorem 1Theorem 1.
Let L ⊆ Σ ω be a regular language.Then there exists an unambiguous PBA B such that L > ( B ) = L .Proof. Let A = ( Q, Σ, δ, q , c ) be a deterministic parity automaton accepting L , i.e., a finite automaton with priority function c : Q → { , . . . , m } such that w ∈ L ( A ) iff the smallest priority assigned to a state on the unique run of A on w which is seen infinitely often is even.We will construct an unambiguous LDBA A ′ from A which also accepts L ,from which we will easily obtain an unambiguous PBA B . For this, we take m +1 copies of A and create a Büchi automaton which guesses the smallest prioritythat is seen infinitely often along the run in A , and ensure that only one correctguess is possible for each word.Formally, let A ′ = ( Q ′ , Σ, ∆ ′ , Q ′ , F ′ ) be an NBA with Q ′ := ˜ Q ·∪ Q ·∪ . . . ·∪ Q m consisting of m +1 copies of each state in Q , where the copies of q ∈ Q are denotedby ˜ q, q , . . . , q m , respectively, initial states defined as Q ′ := { ˜ q , q , . . . , q m } andfinal states defined as F ′ := { q i | c ( q ) = i and i is even } . The transition relation ∆ ′ := ˜ ∆ ∪ S mi =1 ∆ m is given by – ˜ ∆ := { (˜ p, a, q ′ ) | δ ( p, a ) = q and ( q ′ = ˜ q or q ′ = q j s.t. c ( q ) ≥ j > c ( p )) } , – ∆ i := { ( p i , a, q i ) | δ ( p, a ) = q and c ( p ) , c ( q ) ≥ i } .As the transitions defined by the ∆ i sets are just copies of a subset of thedeterministic transitions given by δ , and as all accepting states are only in theserestricted deterministic copies of A , clearly A ′ is an LDBA. Now we will showthat it accepts the same language and is unambiguous.If w ∈ L ( A ′ ) , then there exists a run ρ which either reaches or starts in one ofthe m copies of A that contain accepting states, and visits those infinitely often.Notice that we can easily obtain the run of A on w by projecting the states of ρ onto the original states in A . Now w.l.o.g. assume that ρ eventually is in the i -thcopy, i.e., eventually using states in Q i . As ρ is accepting, we have by definitionof F ′ that i must be even and ρ visits states q i in A ′ such that c ( q ) = i in A infinitely often. Also, ρ eventually never visits states q i with c ( q ) < i in A , astransitions with such states are not defined in ∆ i . This implies that the run of A on w is accepting.If w ∈ L ( A ) , let ρ now be the run on w in A , k the minimal priority k whichis seen along ρ infinitely often, and t the time where a state with priority < k isvisited for the last time (or if ρ never visits such states, let t := − ).First consider the runs that start in some initial state q j , which all proceeddeterministically in the corresponding restricted j -th copy of A . If j > k , thenat some point when in ρ a state q with priority ≤ k is visited, there exists nomatching transition to q j , so the run from q j terminates. If j < k , then if ρ reaches a state with priority < j , the run also terminates, and otherwise at somepoint ρ does not see states with priority < k , so that by definition of F ′ the rundoes not see accepting states anymore, and hence the run is rejecting. If j = k ,then the run terminates if some state with priority < k is visited at some point mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 23 along the run ρ , and otherwise (the case with t = − ) it can continue forever.Furthermore, by choice of k , states q j with c ( q ) = k are visited infinitely often,so that by definition of F ′ the run is accepting.Now consider the runs which start in ˜ q and observe that the automaton caneither use the unique transitions between states in ˜ Q , or at any point nondeter-ministically decide to switch into one of the restricted copies discussed above,but from any state ˜ p only to a q j ∈ Q j in a copy of A where only copies of states q ∈ Q with priorities c ( q ) ≥ j > c ( p ) can be reached.If t = − , i.e., no state with priority < k is ever visited by ρ , the runs of A ′ which forever visit states in ˜ Q are all rejecting, whereas runs that eventuallyswitch into one of the other copies can only choose to go to a copy with states Q j with j > k and hence these runs must terminate whenever ρ visits a state q with c ( q ) = k , which happens infinitely often, so all runs from ˜ q are rejecting.For t ≥ , observe the following. If a run eventually switches from ˜ Q to somestate in Q j with j = k , then as discussed above the run will either terminate (dueto missing transitions in ∆ j ) or be rejecting (by definition of F ′ ). Furthermore,if it switches too early to a state in Q k , it will also terminate (as ρ will visit atleast one more state with priority < k ), and a run cannot switch to states in Q k strictly after t , because by definition of ˜ ∆ this is only possible from a state withpriority < k . Hence, the only possible accepting run is the one which stays in ˜ Q until time t and in the next transition switches to some state q k ∈ Q k , fromwhere it continues deterministically and accepts, as then no more states withpriority < k are visited by ρ and hence no transitions that are missing in ∆ k are used, and furthermore infinitely many states q k ∈ F ′ are visited, which arecopies of states q with c ( q ) = k .So in any case, for every accepting run ρ in A there exists exactly one ac-cepting run in A ′ : for t = − it is the run starting in q k , and for t ≥ it is therun starting in ˜ q and switching to a state in Q k in the transition from time t to t + 1 . Therefore A ′ is an unambiguous LDBA accepting L . As all accepting runsin A ′ are limit-deterministic, we can trivially obtain the claimed unambiguousPBA B which accepts L under positive semantics by equipping edges in A ′ witharbitrary probabilities that result in valid probability distributions, because inany case the unique limit-deterministic accepting runs in B will have positiveprobability. ⊓⊔ A.3 Proof for Theorem 2Theorem 2.
Let A be a PBA that is at most countably ambiguous.Then L > ( A ) is a regular language.Proof. Let A = ( Q, Σ, δ, µ , F ) be a PBA that is at most countably ambiguous.We construct an NBA B accepting L > ( A ) , which intuitively consists of twocopies of A ⊳ . The first copy has no accepting states and the second copy has noforks.Let B = ( Q ′ , Σ, ∆ ′ , Q ′ , F ′ ) be an NBA, where Q ′ = Q × { n, d } consists oftwo copies of each state in A , Q ′ = { ( q, n ) | µ ( q ) > } , F ′ = { ( q, d ) | q ∈ F } ,and transitions ∆ := ∆ n ·∪ ∆ d ·∪ ∆ nd defined by – ∆ n = { (( p, n ) , a, ( q, n )) | δ ( p, a, q ) > } , – ∆ nd = { (( p, n ) , a, ( q, d )) | δ ( p, a, q ) > } , and – ∆ d = { (( p, d ) , a, ( q, d )) | δ ( p, a, q ) = 1 } .It is easy to see that the automaton accepts exactly those words for whichthere exists a limit-deterministic accepting run, hence by Lemma 2 we have L > ( A ) = L ( B ) . ⊓⊔ A.4 Proof for Theorem 3Theorem 3.
Let A be a PBA that is at most exponentially ambiguous or flat.Then L =1 ( A ) is regular and recognizable by DBA.Proof. Let A = ( Q, Σ, δ, µ , F ) be a PBA. There are two cases to consider—when A is exponentially ambiguous and when A is flat.First, assume that A is at most exponentially ambiguous, which means thaton each word there are only finitely many accepting runs. We construct a DBA B accepting L =1 ( A ) . By Lemma 2, B should accept if every run of A acceptsand is limit-deterministic. Notice, that we do not even need to check that theruns on w are limit-deterministic, because if all runs accept, this already implies w ∈ L =1 ( A ) . Hence, we just need to check that all runs accept, using a simplebreakpoint construction.Formally, let B := ( Q ′ , Σ, δ ′ , q ′ , F ′ ) with Q ′ := 2 Q × Q , q ′ := ( ∅ , supp ( µ )) ,F ′ := { ( S, ∅ ) | S ⊆ Q } and transition function δ ′ defined by – δ ′ (( S, ∅ ) , a ) := ( ∅ , ∆ ( S, a )) , and – δ ′ (( S, T ) , a ) := ( S ′ , T ′ ) for T = ∅ with T ′ = ∆ ( T, a ) \ F and S ′ := ∆ ( S ∪ T, a ) \ T ′ .It is easy to see that B sees accepting states infinitely often if and only ifon every path in A an accepting state is visited infinitely often, and hence byLemma 2 we have L =1 ( A ) = L ( B ) .Now assume that A is flat. In this case, we construct a DBA B accepting L =1 ( A ) , that by Lemma 2 should accept w iff there exists no limit-deterministicrejecting run of A . This is checked using a construction almost as above, but now mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 25 it suffices for a state to be at some point reached only by branching transitionsto be moved into the left set.Formally, define B as above, but with different δ ′ defined by – δ ′ (( S, ∅ ) , a ) := ( ∅ , ∆ ( S, a )) , and – δ ′ (( S, T ) , a ) = ( S ′ , T ′ ) for T = ∅ with • T ′ := { q | q F and ∃ p ∈ T s.t. δ ( p, a, q ) = 1 } , and • S ′ := ∆ ( S ∪ T, a ) \ T ′ .Let w = w w . . . ∈ Σ ω . If w L =1 ( A ) , by Lemma 2 there exists a limit-deterministic rejecting run ρ = q , q , . . . on w , then from some time t on onlydeterministic transitions (i.e., with δ ( q i , w i , q i +1 ) = 1 ) will be taken and all states q i for i ≥ t are rejecting. Hence by construction the set in the right component ofthe macrostate will always contain the current state along the run and thus willnever become empty anymore, so no accepting states of B are visited anymoreand hence w L ( B ) .On the other hand, if w ∈ L =1 ( A ) , then there are no limit-deterministicrejecting runs, which means that every run either sees accepting states infinitelyoften (in which case it is accepting), or uses branching transitions infinitely often(in which case it is not limit-deterministic). But then by construction, infinitelyoften all successor states in the sets will reach the left set and the right set mustbecome empty, and therefore w ∈ L ( B ) . ⊓⊔ A.5 Omitted details for Proposition 3(2)Lemma 7.
The automata in Figure 2(c) accept non-regular languages for all λ ∈ ]0 , .Proof. The PWA presented in Figure 2(c) is based on the PBA depicted in [4,Fig. 6] and accepts for some λ ∈ ]0 , the following language, which is known tobe not regular: ˜ L λ = ( a k ba k b . . . | k , k , . . . ∈ N ≥ such that ∞ Y i =1 (cid:0) − (1 − λ ) k i (cid:1) = 0 ) Notice that a ω is not accepted, as then q f can never be reached. Also, ifthere are finitely many b ’s, i.e., the word has the shape w = a k b . . . a k n ba ω ,then there is positive probability to not reach q f after reading the last b andafter that q f cannot be reached anymore, hence with positive probability theautomaton rejects w . Hence it is easy to see that all accepted words must be ofthe form ( a + b ) ω .Once a run has reached q f , it becomes accepting and stays accepting forever.The probability to reach q f from q on a k b is (1 − λ ) k , whereas the probability toavoid q f and come back to q instead is − (1 − λ ) k . Hence, Q ∞ i =1 (1 − (1 − λ ) k i ) is the probability of runs that avoid q f forever and therefore is exactly theprobability of rejecting runs. Therefore, we have L =1 ( ˜ P λ ) = ˜ L λ , as claimed. ⊓⊔ A.6 Proof for Theorem 4Theorem 4. L >λ ( A ) is regular for each k -ambiguous PBA A and λ ∈ ]0 , .Proof. We use the characterization of Lemma 5 to construct a generalized Büchiautomaton B (i.e., a Büchi automaton with multiple acceptance sets, where fromeach set at least one state must be visited infinitely often) accepting L >λ ( A ) ,which can easily be translated into an NBA.Intuitively, the new automaton B just guesses at most k different runs of A and verifies that the guessed runs are limit-deterministic and accepting. Theautomaton additionally tracks the probability of the runs over time, to determinewhether the individual runs and their sum have enough “weight”. More precisely,it tracks the probabilities of the current prefixes, which in the limit yield theprobabilities of the runs. As the runs we are interested in are limit-deterministic,there exists a finite prefix which has the probability of the whole run, hencetracking the prefix probabilities is sufficient for our purpose.The automaton rejects when the total probability of the guessed runs is ≤ λ ,one of the runs goes into the rejecting sink q rej or a run does not see acceptingstates infinitely often. Furthermore, the automaton shall guess no runs which aredefinitely useless for acceptance. By Lemma 5 we only need to consider sets ofruns with at most one run that has a probability < ε , where ε := ε k is given byLemma 4. For this single run we also do not need to track the exact probabilityvalue, as its only purpose is to witness that the acceptance probability is strictlygreater than λ , whereas all other runs must have one of the finitely many differentprobabilities which are ≥ ε and must sum to λ .Formally, let ε be as in Lemma 5, and V := V ≥ ε ·∪ { ⋆ n , ⋆ d } , where V ≥ ε is thefinite (by Lemma 3) set of different probability values ≥ ε that a run prefix of A can have, and the values ⋆ n , ⋆ d are to be interpreted as arbitrarily small valuessuch that < ⋆ n , ⋆ d < ε and are introduced for convenience to cover the case oftracking a single low-probability run imprecisely.Then B := ( Q ′ , Σ, ∆ ′ , Q ′ , F , . . . , F k ) is defined with – Q ′ := S ki =1 ( Q × V ) i (tuples of at most k states with probabilities), – Q ′ := { (( q , v ) ... ( q n , v n )) | ≤ n ≤ k, q i pw. diff. and ∀ ( q i , v i ) , µ ( q i ) = v i } , – F i := S i − j =1 ( Q × V ) j ∪ { (( q , v ) ... ( q i , v i ) , ... ) ∈ Q ′ | q i ∈ F, v i = ⋆ n } ∀ i ∈ { ...k } ,and for S = (( p , u ) , . . . , ( p m , u m )) , T = (( q , v ) , . . . , ( q n , v n )) ∈ Q ′ and symbol a ∈ Σ , the transition ( S, a, T ) is defined in ∆ ′ if – m ≤ n , P ni =1 v i > λ and ∀ i ∈ { . . . n } , q i = q rej , – there exists at most one v i such that v i < ε , and – there exist indices j < . . . < j m ≤ n and j m +1 = n + 1 such that for all i ∈ { . . . m } : • the states q j i , . . . , q j i +1 − are pairwise different, and • for all l ∈ { j i , . . . , j i +1 − } , we have: ∗ v l = u i · δ ( p i , a, q l ) if u i · δ ( p i , a, q l ) ≥ ε , ∗ v l = ⋆ n if u i ≥ ε and < u i · δ ( p i , a, q l ) < ε , mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 27 ∗ v l ∈ { ⋆ n , ⋆ d } if u i = ⋆ n and δ ( p i , a, q l ) > (guess when run is det.), ∗ v l = ⋆ d if u i = ⋆ d and δ ( p i , a, q l ) = 1 (ensure that run det.).This means that the automaton starts in a subset of the possible initial states(with respective initial probabilities), listed in a tuple in arbitrary order, and thenmust pick for each state at least one successor that has positive probability. Foreach state in the tuple also multiple different successors may be taken, whichmeans that the automaton then tracks these as distinct runs, but the totalnumber of tracked runs can be at most k . In other words, the automaton picksin each transition at most k different edges in the run tree of A and adjuststhe probabilities according to the probability of the respective finite path prefix.Hence, by construction, the automaton tracks at most k runs which are alldifferent, all but at most one have a probability ≥ ε , no run ever goes into therejecting sink q rej of A , and the total probability of these runs is > λ .If w ∈ L ( B ) , then there exists an accepting run ρ such that after somefinite time t the tuple size stabilizes at some size n ≤ k (as it is monotonicallyincreasing) and the sum of probabilities in the tuple stabilizes at some value > λ (as they are monotonically decreasing and can only take finitely many values).Furthermore, as ρ is accepting, infinitely many states along ρ are in the sets F i for i ≤ n , which means that in each tuple component accepting states of A arevisited infinitely often. Notice that this implies that after t , every state in a tuplehas exactly one selected successor, because each state must have at least one,but having more then one implies that the tuple would grow. Also, this successormust have probability 1 according to the transition distributions of A , as eitherthe total tracked probability would decrease, or there would be no transitionfor the single run which must at some point have the value ⋆ d assigned. Hence,there exist n different limit-deterministic accepting runs in A that have in totala probability > λ , witnessing that w ∈ L >λ ( A ) .If w ∈ L >λ ( A ) , then we can choose a set R of accepting runs as in Lemma 5,i.e., with total probability > λ , at most one run with a probability < ε , and allsubsets of R have probability < λ .The automaton B can guess this set R of runs, increasing the size of the tuplewhenever runs in R separate after sharing a common prefix. After some finitetime then all those runs become deterministic, i.e., only have unique successorswith probability 1, which means that the tracked probabilities do not decreaseanymore. For runs that have a probability ≥ ε , this means that the trackedvalue stabilizes eventually. For the possible single run with probability < ε , theautomaton eventually replaces its probability by ⋆ n and finally by ⋆ d , after therun has also become deterministic. As by assumption the runs are accepting, inevery component of the tuple infinitely often an accepting state is visited, suchthat by definition, infinitely often a state in F i is visited for all ≤ i ≤ k , hence w ∈ L ( B ) . ⊓⊔ B Proofs for section on weak PBA
B.1 Proof for Theorem 7Theorem 7.
The regularity of PWA (and therefore of PBA) under positive,almost-sure and threshold semantics is an undecidable problem.Proof.
Since L >λ ( PWA ) ⊇ L > ( PWA ) (see Theorem 10), L > ( PWA ) = L =1 ( PWA ) (see remark above), and the class of regular ω -languages is closed under com-plement, it suffices to show the statement for PWA =1 . We do this by reductionfrom the value 1 problem for PFA, which is the question whether for each ε > there exists a word accepted by the PFA with probability > − ε . This prob-lem is known to be undecidable [13]. We consider a slightly modified version ofthe problem by assuming that no word is accepted with probability 1 by thegiven PFA. The problem remains undecidable under this assumption, becauseone can check if a PFA accepts a finite word with probability 1 by a simplesubset construction.Let A = ( Q, Σ, δ, µ , F ) be some PFA. We construct a PWA B by taking acopy of A and extending it with a new symbol such that from accepting statesof A the automaton is “restarted” on , while from non-accepting states leadsinto a new part which ensures that infinitely many are seen and contains theonly accepting state of B .Formally, we construct the PWA B = ( Q ′ , Σ ′ , δ ′ , µ , F ′ ) with Q ′ := Q ·∪{ q , q a } , Σ ′ := Σ ·∪ { } , F ′ := { q a } by extending δ to δ ′ as follows: – δ ′ ( p, x, q ) := δ ( p, x, q ) ∀ p, q ∈ Q, x ∈ Σ , – δ ′ ( p, , q ) := µ ( q ) if p ∈ F and δ ′ ( p, , q ) = 1 if p ∈ Q \ F , – δ ′ ( q a , , q a ) = δ ′ ( q a , x, q a ) = δ ′ ( q , x, q ) = 1 ∀ x ∈ Σ , and – δ ′ ( q , , q ) = δ ′ ( q , , q a ) = .First notice that whenever a run reaches q , its continuations will almostsurely reach q a (and hence be accepting) iff is read infinitely often.If A does not have value 1, then there exists some ε > such that every wordis accepted by A with probability ≤ − ε . But as q can only be avoided byreaching a state that is accepting in A before reading , for any infinite sequenceof words w i ∈ Σ ∗ for i ∈ N we have that the probability to never reach q onthe word w = w w . . . is Q i Pr Acc ( A , w i ) ≤ Q i − ε = 0 , which means thaton any such w almost surely the state q will be reached and hence w will beaccepted. For words not of this shape, i.e. containing only finitely many , arun will either never reach q or stay in it forever never reaching q a . Thereforewe have L =1 ( B ) = ( Σ ∗ ω , which is a regular language.For the case that A does have value 1, recall that we assumed that no wordis accepted with probability 1. But since there are words accepted with proba-bility arbitrarily close to 1, there exists an infinite sequence of words w i ∈ Σ ∗ such that Q i Pr Acc ( A , w i ) > , and therefore on w = w w . . . with posi-tive probability q can be avoided forever, i.e., w L =1 ( B ) . Notice that sucha word w cannot be ultimately periodic, as then w could be written as uv ω mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 29 where v = w j w j +1 . . . w k for some j, k ∈ N , j ≤ k . If p is the probabilityto avoid q on v in B , then the probability to avoid q on w is at most Q i p ,which is 0 for p < and we already excluded that p = 1 (this would requirethat at least one word is accepted by A with probability 1), so all ultimatelyperiodic words are accepted by B . But then the subset R ⊆ ( Σ ∗ ω of wordsof the shape w w . . . that are rejected by B does not contain an ultimatelyperiodic word, so R cannot be regular and therefore L =1 ( B ) = ( Σ ∗ ω \ R isalso not regular. ⊓⊔ B.2 Proof for Proposition 6 q a q b q + q $ b : 1 , a : a : a, ba : 1 , b : $ $ $ Automaton in Figure 2(a).
Proposition 6.
For all thresholds λ ∈ ]0 , there exists a PWA A such that L >λ ( A ) is not regular and not P BA > recognizable.Proof. We show the result for λ = (in which case the PWA even has only ra-tional coefficients). The general statement follows, because one can easily modifythe PBA to accept the same language with any threshold λ ∈ ]0 , by [4, Lemma4.15].Consider the PWA A in Figure 2(a). Clearly, it can only positively acceptwords of shape ( a + b ) ∗ $ ω . Let w = u $ ω with u ∈ { a, b } ∗ and let a ( u ) denotethe number of occurrences of a ∈ Σ in u . Notice that on each b , half of theremaining probability of currently being in q b goes into the (implicit) rejectingsink, and on each a , half the probability of currently being at q a goes to q + . Theonly runs which can continue on $ ω after reading u are in q b or in q + after u and the unique possible run continuation on $ ω goes to and forever stays in theaccepting state q $ . Hence, we have: Pr ( A accepts w ) = ·
12 b ( u ) z }| { Pr ( A in q b after u ) + P a ( u ) i =1 12 i = · (1 −
12 a ( u ) ) z }| { Pr ( A in q + after u ) This means, that Pr ( A accepts w ) = · (1 −
12 a ( u ) +
12 b ( u ) ) , which is greaterthan if and only if a ( u ) > b ( u ) , and therefore L > ( A ) = { ( a + b ) ∗ $ ω | a ( u ) > b ( u ) } .Now it is easy to see that there are infinitely many Myhill-Nerode equivalenceclasses for this language, and hence it cannot be regular (as the implication“regular ⇒ finitely many Myhill-Nerode classes” also holds for infinite words).Furthermore, by Lemma 6 languages accepted by PBA > have only finitely manyclasses. Hence, this language cannot be accepted by any PBA > . ⊓⊔ B.3 Proof for Theorem 8Theorem 8. L > ( PWA ) = L > ( PCA ) and L =1 ( PWA ) = L =1 ( PBA ) .Proof. We show the first statement. The second then follows by duality, i.e., wecan interpret a PBA =1 A recognizing L as a PCA > recognizing L and just applythe construction to get a PWA > B for L , such that B (with inverted acceptingand rejecting states) is a PWA =1 for L . In the first statement the ⊆ inclusion istrivial, hence we only need to show that L > ( PCA ) ⊆ L > ( PWA ) .Now let A = ( Q, Σ, δ, µ , F ) be a PCA > . We refer to the states in F as bad states (since they occur only finitely often in accepting runs). Intuitively,the PWA > B accepting the same language is constructed as follows. Take twocopies of A , a guess copy and a verify copy. Each transition in the guess copyis modified to go into the verify copy with probability and all transitions tocopies of bad states in the verify copy are redirected to a rejecting sink.Formally, let Q g , Q v be two copies of the states Q and let q g and q v denotethe respective copy of q ∈ Q . The PWA B = ( Q ′ , Σ, δ ′ , µ ′ , F ′ ) is defined with Q ′ := Q g ∪ Q v ∪{ q rej } , µ ′ ( q g ) := µ ( q ) for all q g ∈ Q g and otherwise, F ′ := Q v ,and δ ′ defined as: – δ ′ ( p g , x, q g ) = δ ′ ( p g , x, q v ) = · δ ( p, x, q ) – δ ′ ( p v , x, q v ) = δ ( p, x, q ) if q F – δ ′ ( p v , x, q rej ) = 1 − P q F δ ( p, x, q ) Notice that we can write the set of accepting runs
AccRuns ( A , w ) on someword w ∈ Σ ω as a countable union of disjoint sets S i ≥ goodFrom ( i ) , such that goodFrom ( i ) contains the accepting runs where i is the smallest time such thatno state in F is visited at times ≥ i .Assume that w ∈ L > ( A ) . By σ -additivity, this implies Pr ( AccRuns ( A , w )) = P i ≥ Pr ( goodFrom ( i )) > and hence there is an i with Pr ( goodFrom ( i )) > .Let Q i ⊆ Q be the set of states occupied by some run in goodFrom ( i ) at time i . Clearly Q i is reached at time i with positive probability and by definition theruns in goodFrom ( i ) never see bad states after i . But then by construction, withpositive probability some runs of B stay in the guess copy until time i − andreach the verify copy at time i and then they proceed in the verify copy exactlyas the runs goodFrom ( i ) proceed after i in A . Hence, they never will visit states q v which correspond to states q ∈ F and thus forever stay in the verify copy(where all states are accepting) and therefore w ∈ L > ( B ) .The other direction is similar—if w ∈ L > ( B ) , then there exists some time i such that runs of B reach the verify copy at i and then with positive probabilitystay there, i.e., there is a subset goodFrom ( i ) of those runs that has positiveprobability, such that the runs never visit the rejecting sink after reaching i . Byconstruction, clearly the probability for corresponding runs in A is at least aslarge and hence w ∈ L > ( A ) . ⊓⊔ mbiguity, Weakness, and Regularity in Probabilistic Büchi Automata 31 B.4 Proof details for Theorem 10(2)
In this section we show that L > ( PBA ) and L >λ ( PWA ) are incomparable, i.e.,neither contains the other one. One direction directly follows by Proposition 6,i.e., there are languages recognized by PWA >λ that cannot be recognized withPBA > .For the other direction, the following result characterizes the languages ac-cepted by weak automata under extremal semantics in the Borel hierarchy, fromwhich the claim will follow. We do not introduce the details of this hierarchyhere, but rather refer the reader not familiar with these concepts to [22] and [8].Notice that the sets we call Π and Σ (using modern naming) are called G δ and F σ there.The result easily follows from an adaptation of [8, Lemma 3.2]: Lemma 8 (Topological characterization). If A is a PWA and λ ∈ [0 , athreshold, then L ≥ λ ( A ) is a Π set and L >λ ( A ) is a Σ set.Proof. The first statement is implied by [8, Lemma 3.2], as L ≥ λ ( A ) is a Π setfor any (even not weak) PBA. The second statement can be obtained for weakautomata by a simple adaptation of this proof, by showing that the set of wordsrejected by some PWA with probability ≤ (1 − λ ) is a Π set. The decompositionof paths into countable unions and intersections performed in the proof can bedone in the same way, due to the fact that in weak automata a run is rejectingif it sees rejecting states infinitely often (which means that the run eventuallystays in a rejecting SCC). But then clearly the complement of this set is the setof words that are accepted by A with probability > λ , which is exactly L >λ ( A ) and by definition is a Σ set. ⊓⊔ From Lemma 8 and the facts shown in [8] that L > ( PBA ) =