Abstract
Coverability in Petri nets finds applications in verification of safety properties of reactive systems. We study coverability in the equivalent model: Vector Addition Systems with States (VASS).
A kVASS can be seen as k counters and a finite automaton whose transitions are labelled with k integers. Counter values are updated by adding the respective transition labels. A configuration in this system consists of a state and k counter values. Importantly, the counters are never allowed to take negative values. The coverability problem asks whether one can traverse the kVASS from the initial configuration to a configuration with at least the counter values of the target.
In a wellestablished line of work on kVASS, coverability in 2VASS is already PSPACEhard when the integer updates are encoded in binary. This lower bound limits the practicality of applications, so it is natural to focus on restrictions. In this paper we initiate the study of 2VASS with one unary counter. Here, one counter receives binary encoded updates and the other receives unary encoded updates. Our main result is that coverability in 2VASS with one unary counter is in NP. This improves upon the inherited stateoftheart PSPACE upper bound. Our main technical contribution is that one only needs to consider runs in a certain compressed linear form.
Filip Mazowiecki is supported by the ERC grant INFSYS, agreement no. 950398. Henry SinclairBanks is supported by EPSRC Standard Research Studentship (DTP), grant EP/T5179X/1. Karol Węgrzycki is supported by the ERC grant TIPEA agreement no. 850979.
You have full access to this open access chapter, Download conference paper PDF
Similar content being viewed by others
Keywords
1 Introduction
Vector Addition Systems with States (VASS) are a wellstudied class of infinitestate systems (see the survey [37]). These are finite automata with counters that can be updated, but are never allowed to take negative values. Thus, a configuration consists of a state and a vector over the natural numbers. The central decision problems are the reachability and coverability problems. The reachability problem asks whether from a given start configuration one can reach the target configuration. The coverability problem is the same except that the target configuration need not be reached exactly, counter values are allowed to be greater. Both problems are not only mathematically elegant, but they have interesting theoretical applications [7] and implementations [6]. Coverability is provably a simpler problem that is better suited for applications; reachability tools are mostly applied to coverability benchmarks [14]. Yet coverability has applications in the verification of safety conditions in reactive systems [17, 21]. Such systems may require additional data structures to be accurately represented, like counters for example. Safety conditions often boil down to whether a particular state can be reached as opposed to a particular configuration [8].
Coverability and reachability have been studied for decades. The equivalent model of Petri nets was introduced already in the sixties [34]. For general VASS, Lipton proved in 1976 an EXPSPACE lower bound that applies to both coverability and reachability [31]. Two years later, Rackoff proved a matching EXPSPACE upper bound for coverability [35]. Later in 1981, Mayr proved that reachability is decidable [32] without providing an upper bound for the algorithm. The construction was simplified by Kosaraju [24] and Lambert [25], and a recent series of papers by Leroux and Schmitz ended in 2019 by proving an Ackermann upper bound [27]. A matching Ackermann lower bound was published in 2021 by two independent groups [12, 26].
Plenty of attention has been given to VASS with fixed dimension, that is when the number of counters k is invariable, denoted kVASS. For fixed dimension VASS it matters much whether the counter updates are encoded in unary or binary. Already, Rackoff gives NL and PSPACE upper bounds for coverability in unary encoded and binary encoded kVASS, respectively [35]. The coverability problem where there are no counters is just directed graph reachability that is NLcomplete [3]. Thus, coverability in unary encoded kVASS is NLcomplete, for every fixed k. Coverability in binary encoded 1VASS is in NC\({}^2\) [2], it can therefore be decided in deterministic polynomial time. If there are two or more binary counters, coverability is PSPACEhard [5] via a reduction from reachability in bounded onecounter automata that is PSPACEcomplete [18]. Therefore, coverability in binary encoded kVASS is PSPACEcomplete for every \(k \ge 2\). See Figure 1 for the complexities of coverability in VASS with a fixed number of unary and binary encoded counters. This is all in striking contrast to the reachability problem in fixed dimension VASS, since reachability in 8VASS is already known to be nonelementary [13].
There is a prominent line of work on 2VASS with various encodings. The seminal paper in 1979 of Hopcroft and Pansiot [23] shows reachability in 2VASS is decidable, proving that the reachability set is effectively semilinear. Moreover, in the same paper the authors show, by an example, that the 3VASS reachability set need not be semilinear. Later, this was improved as it was shown that for 2VASS the reachability relation is effectively semilinear [28]. This proof shows that every 2VASS can be characterised by a flat model, i.e. where the underlying finite automaton does not contain nested cycles. A more careful analysis of that paper, resulted in a PSPACE upper bound result for reachability in binary encoded 2VASS [5]. Since coverability in binary encoded 2VASS is PSPACEhard [5], the authors were able to conclude that both coverability and reachability are PSPACEcomplete. Just as coverability demonstrated the difference encoding makes to complexity, so does reachability; later it was proved that reachability in unary encoded 2VASS is NLcomplete [16].
Our Results and Techniques. We consider the coverability problem for 2VASS with one unary counter. Here, updates of one counter are encoded in binary and the updates of the other are encoded in unary, see Figure 2 for an example. Notice that the unary counter need not be limited to polynomially bounded values. Otherwise, the value of the unary counter could be encoded into the states for an instance of coverability in binary encoded 1VASS. Furthermore, we do not impose any restrictions on the initial and the target configurations, i.e. both coordinates of these vectors are encoded in binary. Our main result is that coverability in 2VASS with one unary counter is in NP.
Coverability in binary encoded kVASS is PSPACEcomplete, for \(k \ge 2\). The lower bound limits the practicality of applications. Therefore, it is sensible to consider restricted variations and quantify their complexity. We remark that coverability in fixed dimension VASS had widelyopen complexity if there was exactly one binary counter and at least one unary counter. See Figure 1 for a summary of the known results.
The natural starting point is the characterisation of runs via linear path schemes [4]. Intuitively, the authors prove that if coverability or reachability holds then there is a witnessing path of a specific shape. Namely, all paths can be characterised by a bounded language defined by a regular expression of the form \(\tau _0 \gamma _1^* \tau _1 \ldots \tau _{k1} \gamma _k^* \tau _k\). Here \(\tau _0, \dots , \tau _k\) are paths that connect disjoint cycles \(\gamma _1, \dots , \gamma _k\). Since the language is bounded, checking if there is a path for a given expression essentially amounts to an instance of integer linear programming. In particular, the authors argue that both k and \(\tau _0  + \gamma _1  + \tau _1  + \ldots + \tau _{k1}  + \gamma _k  + \tau _k \) are pseudopolynomially bounded [4]. However, a polynomial bound would immediately yield an NP upper bound as such a regular expression can be guessed. Given that coverability in 2VASS with two binary counters is PSPACEhard [5], we cannot simply directly apply the known results when dealing with 2VASS with one binary and one unary counter. In Section 3, we provide a detailed discussion and a difficult yet motivating example in Figure 3.
To overcome this problem, we show that coverability can be witnessed by paths in compressed linear form. We relax the condition of the bounded language, by allowing to nest linear forms, provided that the exponents are fixed. Intuitively, an expression of the form \((\tau \gamma ^* \tau ')^*\) is still forbidden, but we allow for \((\tau \gamma ^e \tau ')^*\), where e is fixed but can be exponentially large (encoded using polynomially many bits). Such a form easily provides an NP upper bound.
We rely on two crucial observations to prove that we can focus on paths in compressed linear form. First, notice that the \(*\) operation in a linear path scheme corresponds to iterating some cycle in the VASS. Since \(\gamma _1, \ldots , \gamma _k\) need to be short, one naturally focuses on short cycles. The issue is that there are exponentially many cycles of polynomial size. In Section 4 we prove that for coverability there are only polynomially many ‘optimal’ cycles. In Section 5 we deal with the problem when some cycle \(\gamma \) occurs many times in a linear path scheme witnessing coverability, resulting in a polynomial bound on k, the width of the linear path scheme. Then we prove that, either we can merge some \(\gamma _i\) and \(\gamma _j\) thus reducing the width, or that there is a cycle that has positive effect on one counter and nonnegative effect on the other counter. Intuitively, in the latter case, we can reduce the problem to coverability in 1VASS by pumping such a cycle that forces one counter to take an arbitrarily large value. Moreover, such a cycle is witnessed by a linear path scheme. Since we need to pump this cycle, we require compressed linear forms to describe the repetitions of the cycle.
We highlight that both our crucial observations rely on that we work with coverability, not reachability. We further highlight that we address these crucial observations through our technical contributions that often depend on the fact there is one unary counter.
Further Related Work. Asymmetric treatment of the counters has been already considered for VASS. Recall that Minsky machines can be seen as VASS with the additional ability of zerotesting. For this model coverability is undecidable [33], even with two counters. This raised natural questions of what happens where only one of the counters is able to be reset or tested for zero. This, and more generally, reachability in VASS with hierarchical zerotests are known to be decidable [36]. There is a further investigation into VASS with one zerotest [20]. Recently, work has appeared containing detailed analysis about 2VASS where counters have different powers [19, 29]. Finally, one of the most famous open problems in the community is whether reachability is decidable for 1VASS with a pushdown stack. For these systems, coverability is known to be decidable [30]. The best known lower bound is that coverability, thus reachability also, is PSPACEhard [15]. Our model, 2VASS with one unary counter, can be seen as 1VASS with a singleton alphabet pushdown stack.
The complexity of reachability in binary encoded 3VASS remains an intriguing open problem. It is PSPACEhard, like in dimension two, and the only known upper bound is primitive recursive, but not even elementary [27]. Recent works on reachability in fixed dimension VASS [9, 11, 13] provide new examples and a better understanding of the VASS model. Interestingly, many techniques applied to fixed dimension VASS are very closely related to recent progress on the nonelementary and Ackermann lower bounds for general VASS [10, 12, 26]. We finally and additionally motivate coverability in VASS with one binary counter and (at least) one unary counter as an avenue for finding new techniques to approach VASS problems with.
2 Preliminaries
Given an integer \(z \in \mathbb {Z}\) we denote \(\textrm{bitsize}(z) = \log _2(z  + 1) + 1\). For a vector \(\textbf{v} \,{:}{=}\, (v_1,v_2)\) we use \((\textbf{v})_1 \,{:}{=}\, v_1\) and \((\textbf{v})_2 \,{:}{=}\, v_2\) to be the projections to the first and second coordinates, respectively. We use \(\textbf{v}_{\max } \,{:}{=}\, \max \{v_1,v_2\} + 1\) to denote the size of vector \(\textbf{v}\). We write \(\textbf{v} \le \textbf{w}\) if the inequalities hold on each coordinate. We write \(\textbf{v} < \textbf{w}\) if at least one of the inequalities is strict.
A 2VASS with one unary counter \(V = (Q, T)\) consists of a finite set of control states Q and a set of transitions \(T \subseteq Q \times \mathbb {Z}\times \{1, 0, 1\} \times Q\). We shall refer to the first counter as the binary counter and the second counter as the unary counter. The size of V is \(V  = Q  + \sum _{(p,b,u,q) \in T} \textrm{bitsize}(b)\). With \(V _{\max } \,{:}{=}\, Q  + T \cdot T _{\max }\) we denote the total ‘pseudopolynomial size’ of the automaton, where \(T _{\max }\) denotes the maximum absolute value that occurs in the transitions. Note that in a standard 2VASS both counters are in binary, i.e. the domain of updates for the second counter is also \(\mathbb {Z}\).
A path \(\pi \) in V is a, possibly empty, sequence of transitions \(\pi = (t_i)_{i=1}^m\) such that \(t_i = (q_{i1}, b_i, u_i, q_i) \in T\). A path is simple if \(q_0, \ldots , q_m\) are distinct. A path is a cycle if \(q_0 = q_m\) and \(m > 0\) (thus empty cycles are forbidden). We call it a \(q_0\)cycle to emphasise the first and last state of the cycle. A cycle is simple if \(q_1, \ldots , q_m\) are distinct. A cycle is short if \(m \le Q \). The length of a path is the number of transitions in the path, denoted \(\textrm{len}(\pi ) = m\). We write \(\pi [i..j]\) to denote the path that is the subsequence of transitions \((t_i, \ldots , t_j)\) in \(\pi \).
A configuration \((p, \textbf{u}) \in Q \times \mathbb {N}^2\), denoted \(p(\textbf{u})\), is a state paired with the current binary and unary counter values. A run is a sequence of configurations \((q_i(\textbf{v}_i))_{i=0}^m\) such that \((q_{i1}, (\textbf{v}_i)_1  (\textbf{v}_{i1})_1, (\textbf{v}_i)_2  (\textbf{v}_{i1})_2, q_i) \in T\). A run can equivalently be defined by the sequence of configurations induced by following a path \(\pi \) starting from an initial configuration \(q_0(\textbf{v}_0)\). We denote this run \(q_0(\textbf{v}_0)\xrightarrow {\pi }q_m(\textbf{v}_m)\). We also write \(q_0(\textbf{v}_0)\xrightarrow {*}q_m(\textbf{v}_m)\) to indicate the existence of a run between two configurations.
In this paper we study the coverability problem for VASS.
Do note that the initial configuration \(p(\textbf{u})\) and the target configuration \(q(\textbf{v})\) have both the binary and unary components encoded as binary integers. The reachability problem for VASS—which we will not study in this paper—requires \(\textbf{v}' = \textbf{v}\).
Consider a path \(\pi = (t_i)_{i=1}^m\), where \(t_i = (q_{i1}, b_i, u_i, q_i)\). The effect of \(\pi \) is the sum of the counter updates, i.e. the vector \(\textrm{eff}(\pi ) \,{:}{=}\, \sum _{i=1}^m (b_i,u_i)\). We often focus on the two projections: the binary effect \(\textrm{eff}_b(\pi ) \,{:}{=}\, \sum _{i=1}^m b_i\), and the unary effect \(\textrm{eff}_u(\pi ) \,{:}{=}\, \sum _{i=1}^m u_i\).
We say that a cycle \(\gamma \) is monotone if \(\textrm{eff}(\gamma ) \ge \textbf{0}\) or \(\textrm{eff}(\gamma ) \le \textbf{0}\). Otherwise, we say that \(\gamma \) is nonmonotone. Note the two variants of a nonmonotone cycle: a positivenegative cycle \(\textrm{eff}_b(\gamma ) > 0\) and \(\textrm{eff}_u(\gamma ) < 0\), and a negativepositive cycle \(\textrm{eff}_b(\gamma ) < 0\) and \(\textrm{eff}_u(\gamma ) > 0\).
Let \(\gamma \) be a cycle. Given \(e \in \mathbb {N}\) we write \(\gamma ^e\) for the path obtained by e repetitions of \(\gamma \). We refer to e as the exponent. A linear path scheme is a regular expression of the form \(\tau _0 \gamma _1^* \tau _1 \cdots \tau _{k1} \gamma _k^* \tau _k\), where the paths \(\tau _0, \tau _1, \ldots , \tau _k\) connect disjoint cycles \(\gamma _1, \ldots , \gamma _k\). Note that a collection of cycles is disjoint if no two cycles have a common state. Given \(\ell = (\tau _0, \gamma _1, \tau _1, \dots , \tau _{k1}, \gamma _k, \tau _k)\), we say the a path \(\pi \) is in linear form \(\ell \) if \(\pi = \pi _\ell = \tau _0 \gamma _1^{e_1} \tau _1 \cdots \tau _{k1} \gamma _k^{e_k} \tau _k\) for some exponents \(e_1, \ldots , e_k\). Note that in this definition every path has a linear form, e.g. \(\tau _0 = \pi \) is valid. To leverage the definition, we will ask whether paths are in a linear form of certain size. The size of a linear form \(\ell \) is \(\sum _{i=0}^{k} \textrm{len}(\tau _i) + \sum _{i = 1}^k \textrm{len}(\gamma _i)\). The size of \(\pi _\ell \) is \( \sum _{i=0}^{k} \textrm{len}(\tau _i) + \sum _{i = 1}^k \textrm{len}(\gamma _i) + \sum _{i = 1}^k \textrm{bitsize}(e_i)\), i.e. includes the exponents. We refer to k as the width of the linear form.
3 Coverability in 2VASS with One Unary Counter
In this section we briefly discuss why the stateoftheart techniques are not enough to prove that coverability in 2VASS with one unary counter is in NP. Blondin et al. [4] show that for a given 2VASS V there exists a set of linear path schemes S such that if \(p(\textbf{u})\xrightarrow {*}q(\textbf{v})\) in V, then there exists a path \(\pi \) in a linear path scheme \(\rho \in S\) such that \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{v})\). For every linear path scheme \(\rho \in S\) the width of \(\rho \), and therefore the width of every path, is bounded above by \(\textsf {poly}(Q,T_{\max })\) [4, Theorem 3.1]. Such a path \(\pi \) is not necessarily a polynomial size witness, as the width depends on \(T _{\max }\) polynomially. We provide an example of a 2VASS with one unary counter where the width of every linear form \(\ell \) for a path is exponential in the input size. This demonstrates that the combinatorial structure of linear path schemes is not selfsufficient to show that there always exists a polynomial size witness of coverability.
Paths in Compressed Linear Form. Nevertheless, there is a natural way to succinctly describe the path presented in Figure 3. Let \(\sigma = \lambda \rho \), and note that
All paths and cycles are ‘small’, and the bitsize of N and \(N^2\) are polynomial in n, so \(\sigma \) itself is a path in linear form. We introduce the following generalisation of linear form paths that encapsulates the idea behind paths of this kind of arrangement.
Definition 1 (Compressed linear form path)
A path \(\pi \) is in compressed linear form if \(\pi = \rho _0 \sigma _1^{f_1} \rho _1 \cdots \rho _{k1} \sigma _k^{f_k} \rho _k\) for some connected paths in linear form \(\rho _0, \rho _1, \ldots , \rho _k\); cycles in linear form \(\sigma _1, \ldots , \sigma _k\); and exponents \(f_1, \ldots , f_k\). The size of a compressed linear form path is the sum of the sizes of all \(\rho _i\) and \(\sigma _i\) (including the bitsize of their exponents) plus the bitsize of the exponents \(f_i\) (Fig. 4).
The following theorem is our main contribution.
Theorem 1
Let V be a 2VASS with one unary counter and fix two configurations \(p(\textbf{u})\) and \(q(\textbf{v})\). If \(p(\textbf{u})\xrightarrow {*}q(\textbf{v})\), then there exists a path in compressed linear form \(\pi \) such that \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{v}')\) and \(\textbf{v}' \ge \textbf{v}\). The size of the compressed linear form path is polynomial in \(V  + \textrm{bitsize}(\textbf{u}) + \textrm{bitsize}(\textbf{v})\).
Corollary 1
Coverability in 2VASS with one unary counter is in NP.
Proof
By Theorem 1 it suffices to consider paths in compressed linear form of polynomial size, that can be guessed in NP. It suffices to observe that a coverability instance on a given compressed linear form amounts to an instance of integer linear programming. Intuitively, this is because the nested cycles are fixed. Thus to check whether a run drops below zero it suffices to check before applying a cycle and after applying it for the last time (see e.g. [5, Section V, Lemma 14]). \(\square \)
We highlight that it is rather unexpected that only one extra ‘level’ of linear form paths is enough to obtain polynomial size witnesses of coverability in a 2VASS with one unary counter, since the problem is PSPACEcomplete for general 2VASS. Roughly speaking, the example given in Figure 3 observes the most complex behaviour possible and this instance of coverability is witnessed by a compressed linear form path. More specifically, compressed linear form paths containing only one linear form cycle suffice as witnesses for coverability in 2VASS with one unary counter. Therefore, all witnesses can be represented by a compressed linear form path \(\rho \sigma ^N \tau \) where \(\rho \) and \(\tau \) are linear form paths to and from the single linear form cycle \(\sigma \) which is iterated N times.
The rest of the paper is dedicated to proving Theorem 1. We heavily exploit both distinguishing features of the problem: the fact that one counter receives unary encoded updates (as opposed to both counters in binary) and the fact that we aim to assert coverability (as opposed to reachability). Our approach is as presented in the introduction. In 4 we observe that we can polynomially bound the total number of distinct short cycles. We formalise this and show that there are only polynomially many ‘irreplaceable’ short cycles. In 5 we provide a ‘reshuffling procedure’. If some short cycle \(\gamma \) repeats exponentially many times we aim to modify the path \(\pi \) by moving the cycles \(\gamma \) close to each other. Then either every short cycle \(\gamma \) will appear only in polynomially many ‘bundles’ \(\gamma ^e\), or we find a cycle \(\sigma \) such that \(\textrm{eff}(\sigma ) > \textbf{0}\). In the latter case, by pumping \(\sigma \) we are essentially left with one counter. Finally, in Section 6 we conclude the proof of Theorem 1.
4 Replacing Short Cycles
In this section, we show that there are only polynomially many short cycles that need occur in a run witnessing coverability. Fix a path \(\pi = (q_{i1}, b_i, u_i, q_i)_{i=1}^k\). Let \(0 \le i_b, i_u \le k\) be the first indices such that \(g_b = \sum _{i=1}^{i_b} b_i\) and \(g_u = \sum _{i=1}^{i_u}u_i\) are at their lowest, respectively. Note that \(g_b, g_u \le 0\) since by convention if we consider \(i_b, i_u = 0\) then the sum evaluates to 0. We call and denote these two numbers the binary guard \(\textrm{grd}_b(\pi ) = g_b\) and the unary guard \(\textrm{grd}_u(\pi ) = g_u\). The following claim immediately follows from these definitions.
Claim 1
Both \(\textrm{grd}_b(\pi [i_b+1 .. k]) = 0\) and \(\textrm{grd}_u(\pi [i_u+1 .. k]) = 0\).
Much like the nadir of a cycle in a onecounter net, defined in [1], we define the binarynadir state as \(q_{i_b}\), i.e. the first state in which the binary counter first attains the lowest value when executing \(\pi \). We call the binarynadir decomposition \(\pi = \pi _1^b \pi _2^b\), for \(\pi _1^b = \pi [1 .. i_b]\) and \(\pi _2^b = \pi [i_b+1 .. k]\), as intimated in Claim 1. Notice that this decomposition necessitates the binary guard of the path \(\pi \) is equal to the binary effect of the prefix \(\pi _1^b\), \(\textrm{grd}_b(\pi ) = \textrm{eff}_b(\pi _1^b) = \textrm{grd}_b(\pi _1^b)\). Furthermore, the suffix of the binarynadir decomposition has zero binary guard \(\textrm{grd}_b(\pi _2^b) = 0\). We primarily utilise binarynadir states and binarynadir decompositions, hence the omission of matching unarynadir states and unarynadirdecompositions.
Definition 2 (Replaceable cycles)
Let \(\gamma \) be a qcycle and let p be the binarynadir state of \(\gamma \). We say that \(\gamma \) is replaceable if there exists a qcycle \(\gamma '\) with the same binarynadir state p, such that

(a)
\(\textrm{eff}_b(\gamma ') \ge \textrm{eff}_b(\gamma )\) and \(\textrm{eff}_u(\gamma ') \ge \textrm{eff}_u(\gamma )\),

(b)
\(\textrm{grd}_b(\gamma ') \ge \textrm{grd}_b(\gamma )\) and \(\textrm{grd}_u(\gamma ') \ge \textrm{grd}_u(\gamma )\), and

(c)
\(\textrm{len}(\gamma ') \le \textrm{len}(\gamma )\).
Additionally, at least one inequality is strict and we write \(\gamma \prec \gamma '\).
We say a cycle is irreplaceable if it is not replaceable. We also say that an irreplaceable qcycle \(\gamma \) with the binarynadir state p is characterised by the five values: \(\textrm{eff}_b(\gamma )\), \(\textrm{eff}_u(\gamma )\), \(\textrm{grd}_b(\gamma )\), \(\textrm{grd}_u(\gamma )\), and \(\textrm{len}(\gamma )\).
Lemma 1 (Replacing cycles)
Let \(\pi = \pi _1 \gamma \pi _2\), where \(\gamma \) is a qcycle. Suppose \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{v})\) then the following hold.

If \(\gamma \) is replaceable, then there exists an irreplaceable qcycle \(\gamma \prec \gamma '\) such that \(p(\textbf{u})\xrightarrow {\pi _1 \gamma ' \pi _2}q(\textbf{v}')\).

If \(\gamma \) is irreplaceable, then for every irreplaceable qcycle \(\gamma '\) that has the same characterisation as \(\gamma \), \(p(\textbf{u})\xrightarrow {\pi _1 \gamma ' \pi _2}q(\textbf{v}')\).
In both cases \(\mathbf{v'} \ge \textbf{v}\) and \(\textrm{len}(\pi ) \ge \textrm{len}(\pi _1 \gamma ' \pi _2)\).
For convenience, we define the polynomial \(R(Q ) \,{:}{=}\, Q ^4(Q +1)(2Q +1)^2\).
Lemma 2
There exists at most \(R(Q )\) many irreplaceable short cycles with different characterisations.
Proof
We fix two states q and p and consider only qcycles \(\gamma \) with the binarynadir state p. Thus in the final argument one must multiply everything by \(Q ^2\). Since we consider short cycles, the unary effect and the unary guard are small, i.e. \(Q  \le \textrm{eff}_u(\gamma ) \le Q \) and \(Q  \le \textrm{grd}_u(\gamma ) \le 0\).
Towards a contradiction, suppose there exists more than \(Q ^2(Q +1)(2Q +1)^2\) many such irreplaceable qcycles with different characterisations. By the pigeonhole principle there must exist two cycles, denoted in binarynadir decomposition \(\gamma = \gamma _1\gamma _2\) and \(\gamma ' = \gamma '_1\gamma '_2\), that have the same values \(\textrm{eff}_u(\gamma _1) = \textrm{eff}_u(\gamma '_1)\), \(\textrm{eff}_u(\gamma _2) = \textrm{eff}_u(\gamma '_2)\), \(\textrm{grd}_u(\gamma ) = \textrm{grd}_u(\gamma ')\), \(\textrm{len}(\gamma _1) = \textrm{len}(\gamma '_1)\), and \(\textrm{len}(\gamma _2) = \textrm{len}(\gamma '_2)\).
We know that the irreplaceable qcycles \(\gamma \) and \(\gamma '\) have different characterisations, so it must be the case that their binary effects differ \(\textrm{eff}_b(\gamma ) \ne \textrm{eff}_b(\gamma ')\). Otherwise, the cycle with the lesser binary guard is replaceable, because the unary effect, unary guard, and length do not differ. Without loss of generality, suppose \(\textrm{eff}_b(\gamma ) > \textrm{eff}_b(\gamma ')\), then \(\textrm{grd}_b(\gamma ) < \textrm{grd}_b(\gamma ')\). Otherwise, \(\gamma '\) would be replaceable as \(\gamma \prec \gamma '\).
Now consider the qcycle \(\sigma = \gamma _1'\gamma _2\), also with the binarynadir state p. We will show that \(\gamma \prec \sigma \) contradicting the fact that \(\gamma \) is an irreplaceable qcycle. First, observe that \(\sigma \) has greater binary effect than \(\gamma \) as
where the inequality holds because \(\textrm{grd}_b(\gamma ) < \textrm{grd}_b(\gamma ')\). Second, \(\sigma \) and \(\gamma \) have equal unary effect because \(\textrm{eff}_u(\gamma _1') = \textrm{eff}_u(\gamma _1)\). Third, we show that \(\sigma \) has a greater binary guard than \(\gamma \). Since \(\gamma _2\) is the suffix of the binarynadir decomposition of \(\gamma \), it must be true that \(\textrm{grd}_b(\gamma _2) = 0\). By Claim 1 \(\textrm{grd}_b(\sigma ) = \textrm{grd}_b(\gamma _1')\). Combining these facts, \(\textrm{grd}_b(\sigma ) = \textrm{grd}_b(\gamma ') > \textrm{grd}_b(\gamma )\). Fourth, \(\sigma \) has at least the unary guard of \(\gamma \) because, in particular, the unary guard of the prefix of a path is at most the unary guard of the entire path.
Fifth and finally, \(\sigma \) and \(\gamma \) have equal length because \(\textrm{len}(\gamma '_1) = \textrm{len}(\gamma _1)\). We have at least one strict inequality. Thus, we have reached the desired contradiction. \(\square \)
5 Reshuffling Linear Form Paths
5.1 Reshuffling Procedure
There can be many linear forms for a path \(\pi \). We will try to find an ‘optimal’ one, so we introduce a cost function to quantify linear forms. Recall that a linear form \(\ell \) is a sequence of paths \(\tau _0, \tau _1, \ldots , \tau _k\) and a sequence of cycles \(\gamma _1, \ldots , \gamma _k\). If \(\pi \) is in the linear form \(\ell = (\tau _0, \gamma _1, \tau _1, \dots , \tau _{k1}, \gamma _k, \tau _k)\) then we write \(\pi _\ell = \tau _0 \gamma _1^{e_1} \tau _1 \cdots \tau _{k1} \gamma _k^{e_k} \tau _{k}\), where \(\pi = \pi _\ell \) (the index is here to stress the exact linear form). For this section, we will consider linear forms only containing short cycles \(\gamma \), they will play a key role in the following arguments.
We define a cost function that assigns, to a linear form \(\ell \), the following pair of naturals \(C(\ell ) \,{:}{=}\, \left( \sum _{i=0}^k \textrm{len}(\tau _i), k \right) \). For convenience, we define the polynomial \(P(Q ) \,{:}{=}\, 2(Q ^2 + 1)(Q ^2 + 2) \cdot R(Q )\), where R is the polynomial defined for Lemma 2. We say that a linear form \(\ell \) is narrow if \(C(\ell ) \le (Q (P(Q )+1), P(Q ))\), otherwise we say that \(\ell \) is wide. We say that the triple \((\pi ', \sigma , \pi '')\) is a monotone cycle decomposition of a path \(\pi \) if \(\sigma \) is a monotone cycle, \(\pi = \pi ' \sigma \pi ''\), and \(\textrm{len}(\sigma ) < \textrm{len}(\pi )\).
Lemma 3 (Reshuffling)
Let \(\pi \) be a path such that \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{v})\). Then there exists a path \(\rho \) such that \(p(\textbf{u})\xrightarrow {\rho }q(\textbf{w})\) where \(\textbf{w} \ge \textbf{v}\), \(\textrm{len}(\rho ) \le \textrm{len}(\pi )\), and either

(i)
there exists a narrow linear form for \(\rho \), or

(ii)
there exists a monotone cycle decomposition of \(\rho \).
Proof
We start with a series of preparations. In the early part of this proof, we provide simple observations to ascertain some auspicious properties of our path. In the later part of this proof, we present the ‘reshuffling procedure’ and conclude with one of the cases in the statement of this lemma. In this proof we will compare linear forms using the lexicographic order \(\prec _{lex}\), that is known to be a linearorder and a wellorder. Formally,
We start with a path \(\pi '\) such that \(p(\textbf{u})\xrightarrow {\pi '}q(\textbf{v}')\) where \(\textbf{v}' \ge \textbf{v}\), \(\textrm{len}(\pi ') \le \textrm{len}(\pi )\), and \(\pi '\) has a linear form \(\ell '\) that has the least cost among all linear forms for all likepaths. That means there does not exist another path \(\pi ''\) such that \(p(\textbf{u})\xrightarrow {\pi ''}q(\textbf{v}'')\) where \(\textbf{v}'' \ge \textbf{v}\), \(\textrm{len}(\pi '') \le \textrm{len}(\pi )\), and \(\pi ''\) has a linear form \(\ell ''\) such that \(C(\ell '') \prec _{lex}C(\ell ')\).
For the first observation, suppose there exists \(0 \le i \le k\) such that \(\textrm{len}(\tau _i) > Q \). Then the path \(\tau _i\) can be written as \(\tau _i = \tau ' \gamma \tau ''\), where \(\gamma \) is a short cycle. We can define the linear form \(\ell ''\) by modifying \(\ell '\) where \(\tau _i\) is swapped for \(\tau ' \gamma \tau ''\). Although this increments the number of cycles k, we decrease the total length of the paths as \(\textrm{len}(\tau ') + \textrm{len}(\tau '') < \textrm{len}(\tau _i)\) (recall that empty cycles are forbidden). Thus \(C(\ell '') \prec _{lex}C(\ell ')\) contradicting the assumption that \(\ell \) has minimum cost. Therefore, we assume that \(\textrm{len}(\tau _i) \le Q \) for all \(0 \le i \le k\).
For the second observation, we define \(U \,{:}{=}\, \{0 \le i \le m: (\textbf{v}_i)_2 < Q \}\) to be the set of indices of configurations in the run that have unary counter value less than \(Q \). Observe that if \(U  > Q ^2 + 1\) then there are two indices \(0< i < j \le m\) such that the two corresponding configurations in the run have matching states \(q_{i} = q_{j}\) and equal unary counter values \((\textbf{v}_{i})_2 = (\textbf{v}_{j})_2\). Then, regardless of sign of its binary effect, \(\pi '[i..j]\) is a monotone cycle. Here, case (ii) immediately holds by decomposing \(\pi '\) itself using the monotone cycle \(\pi '[i .. j]\), given that \(i > 0\) and \(j \le m\) implies \(\textrm{len}(\pi '[i .. j]) = j  i < m = \textrm{len}(\pi ')\). Therefore, we assume \(U  \le Q ^2 + 1\). We continue with the aim of satisfying the conditions of case (ii) by finding a monotone cycle decomposition.
Let \(d = \{\gamma _1, \ldots , \gamma _k\} \) be the number of distinct cycles in the linear form \(\ell '\). By Lemma 1 and Lemma 2, we can assume that \(d \le R(Q )\). Otherwise, we can exchange replaceable qcycles for irreplaceable qcycles using the first point in Lemma 1. It is possible that for a particular characterisation, we can observe more than one irreplaceable qcycle. Then using the second point in Lemma 1, we can arbitrarily select one of these irreplaceable qcycles with equal characterisations to exchange all others with. By applying these cycle replacements to \(\pi '\), we obtain a different path \(\rho \). Definition 2 ensures that we do so without decreasing the effect (a), without allowing the counters to take a negative value (b), and without increasing the length of the path (c). Therefore \(p(\textbf{u})\xrightarrow {\rho }q(\textbf{w})\) and \(\textbf{w} \ge \textbf{v}' \ge \textbf{v}\), and \(\textrm{len}(\rho ) \le \textrm{len}(\pi ') \le \textrm{len}(\pi )\). We remark since cycles have been exchanged oneforone, then \(\rho \) takes a linear form \(\ell \) with the same path segments as \(\ell '\). Therefore, it is clear that neither the number of cycles k, nor the sum of the lengths of the paths between cycles, have changed. We also know that \(\ell \) is a linear form for \(\rho \) with minimum cost \(C(\ell ) = C(\ell ')\), as per the initialisation in this proof.
Suppose \(\rho = \rho _\ell = \tau _0 \gamma _1^{e_1} \tau _1 \cdots \tau _{k1} \gamma _k^{e_k} \tau _k\). Let \((q_j(\textbf{v}_j))_{j=0}^m\) be the run obtained by following the path \(\rho _\ell \) from the initial configuration \(q_0(\textbf{v}_0) = p(\textbf{u})\) to the final configuration \(q_m(\textbf{v}_m) = q(\textbf{w})\). We may assume that \(\ell \) is wide. Otherwise, case (i) is immediately satisfied. We also know that \(\textrm{len}(\rho _\ell ) \ge \max \{(C(\ell ))_1, (C(\ell ))_2\} > P(Q )\). We may also assume that each cycle \(\gamma _1, \ldots , \gamma _k\) is nonmonotone, i.e. it is positivenegative or negativepositive. Otherwise, case (ii) immediately holds by decomposing \(\rho \) itself using some monotone cycle \(\gamma _i\), given that \(\textrm{len}(\gamma _i) \le Q < P(Q ) < \textrm{len}(\rho _\ell )\). Notice this is valid since each \(e_i > 0\) by the minimality of \(C(\ell )\), otherwise you can write \(\cdots \tau _{i1} \gamma _i^0 \tau _i\cdots \) with one less cycle, decreasing \((C(\ell ))_2\).
From the first observation, we get \(\sum _{i=0}^k \textrm{len}(\tau _i) \le (k+1) Q \). Given that \(\ell \) is wide, either \(Q (P(Q )+1) < (C(\ell '))_1 = \sum _{i=0}^k \textrm{len}(\tau _i) \le (k+1) Q \) that implies \(P(Q ) < k\), or \(P(Q ) < (C(\ell '))_2 = k\). Regardless, \(P(Q ) < k\) holds. Recall that \(U  \le Q ^2 + 1\) from the second observation. Since there are relatively ‘few’ configurations indexed by U, there must exist a relatively ‘distant’ pair of consecutive configurations indexed by U. More formally, there are i and j such that \(0 \le i < j \le k\) and \(j  i \ge 2(Q ^2+2)R(Q )\) and all configurations that occur in the run over the path segment \(\tau _{i} \gamma _{i+1}^{e_{i+1}} \cdots \gamma _{j}^{e_j} \tau _{j}\) have unary counter value at least \(Q \). Notice that \(j  i\) is the number of cycles in this path segment. Since \(j  i \ge 2(Q ^2 + 2)R(Q )\) and by pigeonhole principle on the number of irreplaceable cycles, there is a common irreplaceable cycle \(\gamma \) repeated at least \(x = 2(Q ^2 + 2)\) many times. We will focus on the first x such occurrences of this cycle. Let \(s_1, \ldots , s_x\) be the indices of this cycle \(\gamma \), i.e. \(\gamma = \gamma _{s_1} = \ldots = \gamma _{s_x}\). To highlight these cycles, we decompose this path segment into
where \(f_j \,{:}{=}\, e_{s_j}\) and \(\varLambda _j\) are the concatenated paths (and cycles) in between iterations of \(\gamma \), see Figure 5. To reiterate, we know that all configurations that occur in the run over this path segment have at least \(Q \) unary counter value and \(\gamma \) is a short cycle.
Reshuffling Procedure. In the rest of the proof we will modify the path segment (above) of the path \(\rho _\ell \) with a procedure that we call reshuffling. At the end of this procedure we will find a monotone cycle and satisfy case (ii) of this lemma. We either find this cycle directly, or we obtain a linear form \(\ell ''\) such that \(C(\ell '') \prec _{lex}C(\ell )\) contradicting the assumption that \(\ell \) has minimal cost.
Note that \(x = 2(Q ^2 + 2)\) is even, and for every pair of consecutive cycles \(\gamma _{2j1}\) and \(\gamma _{2j}\) (for \(1 < 2j \le x\)), consider the subsegment \(\gamma ^{f_{2j1}} \varLambda _{2j1} \gamma ^{f_{2j}}\). There are two scenarios depending on the variant of the nonmonotone cycle \(\gamma \). In the scenario where \(\gamma \) is positivenegative, we move an iteration of \(\gamma \) from right to left obtaining \(\gamma ^{f_{2j1}+1} \varLambda _{2j1} \gamma ^{f_{2j}1}\). In the scenario where \(\gamma \) is negativepositive, we move an iteration of \(\gamma \) in the opposite direction obtaining \(\gamma ^{f_{2j1}1} \varLambda _{2j1} \gamma ^{f_{2j}+1}\).
We repeat this procedure until one of two conditions are met. The first is when there are no iterations of \(\gamma \) on one side, so either \(f_{2j1}\) or \(f_{2j}\) becomes 0. The second is when there appears a configuration, in the run over the path subsegment after reshuffling, with unary counter value less than \(Q \). See Figure 6 for a pictorial presentation of reshuffling in the scenario where \(\gamma \) is positivenegative.
We claim that after each reshuffling step, the corresponding run remains executable, so we must check that both counters remain nonnegative. Notice that by only moving a cycle, the total effect of the path subsegment remains the same. Therefore, if the run was executable before reshuffling, we can safely assume that the prefix before the path subsegment and the suffix after the path subsegment are still executable. For that reason, consider the counter values of configurations occurring in the run over the reshuffled path subsegment. We focus on a single step of the reshuffling procedure that concerns the subsegment \(\gamma ^{f_{2j1}} \varLambda _{2j1} \gamma ^{f_{2j}}\).
Suppose \(\gamma \) is a positivenegative cycle. Then the reshuffling procedure moves \(\gamma \) from right to left. We claim that since \(f_{2j1} > 0\) and \(\varLambda _0 \gamma ^{f_1} \varLambda _1 \cdots \varLambda _{2j1} \gamma ^{f_{2j1}}\) is executable, the subsegment \(\varLambda _0 \gamma ^{f_1} \varLambda _1 \cdots \varLambda _{2j1} \gamma ^{f_{2j1}+1}\) is executable from the initial configuration. This is because one prerequisite of the reshuffling procedure is that all configurations occurring in the run over the path subsegment have at least \(Q \) unary counter value. Moreover, the cycle \(\gamma \) has length at most \(Q \) so \(\textrm{grd}_u(\gamma ) \ge Q \) means the unary counter value remains nonnegative. As for the binary counter value, since a single execution of \(\gamma \) increases the binary counter and an iteration of \(\gamma \) was already executed before reshuffling, \(\varLambda _0 \gamma ^{f_1} \varLambda _1 \cdots \varLambda _{2j1} \gamma ^{f_{2j1}+1}\) is executable. In the same way, from the initial configuration, \(\varLambda _0 \gamma ^{f_1} \varLambda _1 \cdots \varLambda _{2j1} \gamma ^{f_{2j1}+1} \varLambda _{2j} \gamma _{2j}^{f_{2j}1}\) is executable. This is because \(\textrm{eff}_u(\gamma ) \ge Q \), and again, all configurations occurring in the run over the path subsegment have at least \(Q \) unary counter value, and also because of the monotonicity on the binary counter.
The argument when \(\gamma \) is a negativepositive cycle is analogous. This concludes the correctness analysis of the reshuffling procedure.
Finishing Reshuffling. We analyse what happens when reshuffling is finished. Suppose that there exists a pair \(2j1\) and 2j such that the reshuffling finishes under the first condition where all iterations of \(\gamma \) have been moved to one side of \(\varLambda _{2j1}\). In this case we obtain a new linear form \(\ell ''\) for \(\rho \), where one collection of the cycle \(\gamma \) has been removed (decrementing k). So \((C(\ell ''))_2 = k  1 < (C(\ell ))_2\) and the two adjacent path segments can be combined without changing the summed length of paths so \((C(\ell ''))_1 = (C(\ell ))_1\). Therefore, \(C(\ell '') \prec _{lex}C(\ell )\) contradicting the assumption \(\ell \) has the minimal cost.
Otherwise, for every \(1 \le j \le x/2\) the reshuffling of pair \(2j1\) and 2j finishes under condition the second condition. So there is a configuration with unary counter value less than \(Q \) in the run induced from the path \(\rho \) for each pair \(2j1\) and 2j (see Figure 7). Recall that \(\frac{x}{2} = Q ^2 + 2\), that is the number of pairs. Akin to the first observation (in the beginning of this proof), we use the pigeonhole principle on the number of such configurations to obtain two configurations with matching states and equal unary counter values. The path segment inducing the part of the run between these two configurations is a monotone cycle, regardless of the binary effect. Again, it must be true that the length of this cycle is less than the length of the whole path, so we obtain a monotone cycle decomposition of \(\rho \). Thus case (ii) of the lemma holds. \(\square \)
5.2 Applying Reshuffling
Lemma 3 does not necessarily return a narrow linear form for a path \(\pi \) witnessing coverability. Instead it may return a monotone cycle decomposition \((\rho , \sigma , \tau )\) of \(\pi \). Our next goal is to show that there exists polynomial size certificates for \(\rho \) and \(\sigma \) (Lemma 4), and then to show that there exists a polynomial size certificate for \(\tau \) (Lemma 5). Like linear forms, there can be many monotone cycle decompositions for a path. Following, we will use the cost function assigning monotone cycle decompositions to pairs of natural numbers \(D((\rho , \sigma , \tau )) \,{:}{=}\, (\textrm{len}(\rho \sigma ), \textrm{len}(\sigma ))\). Note that we can compare two decompositions using their cost, even if they are for two different paths.
Lemma 4
Suppose \(p(\textbf{u})\xrightarrow {*}q(\textbf{v})\) yet there is no narrow linear form \(\ell \) for any path \(\pi \) such that \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{w})\) and \(\textbf{w} \ge \textbf{v}\), then there exists a path \(\pi '\) such that

(a)
\(p(\textbf{u})\xrightarrow {\pi '}q(\textbf{w}')\) where \(\textbf{w}' \ge \textbf{v}\),

(b)
there is a monotone cycle decomposition \((\rho , \sigma , \tau )\) of \(\pi '\) where \(\textrm{eff}(\sigma ) > \textbf{0}\), and

(c)
there are narrow linear forms for both \(\rho \) and \(\sigma \).
Proof
We will again use the lexicographical order \(\prec _{lex}\) to compare the cost of monotone cycle decompositions. Let \(\pi \) be a path of minimum length such that \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{w})\) where \(\textbf{w} \ge \textbf{v}\). Let \(c = (\rho , \sigma , \tau )\) be the monotone cycle decomposition of \(\pi \) that minimizes the cost D(c) under the \(\prec _{lex}\) order. Such a decomposition must exist, otherwise applying Lemma 3 would return a narrow linear form \(\ell '\) for \(\rho \) such that \(p(\textbf{u})\xrightarrow {\rho }q(\textbf{w}')\) and \(\textbf{w}' \ge \textbf{w} \ge \textbf{v}\), contradicting an assumption of this lemma. Observe that \(\textrm{eff}(\sigma ) > \textbf{0}\), otherwise one can remove \(\sigma \) and consider the shorter path \(\rho \tau \), contradicting the minimal length of \(\pi \). Next, we argue that \(\rho \) and \(\sigma \) do not have monotone cycle decompositions, we then leverage Lemma 3 to obtain the narrow linear forms required.
Path \(\rho \) cannot be decomposed further. Towards a contradiction, assume that there is a monotone cycle decomposition \(c' = (\rho ', \sigma ', \tau ')\) of \(\rho \). Observe that the following monotone cycle decomposition \(c' = (\rho ', \sigma ', \tau ' \sigma \tau )\) of \(\pi \) has lower cost \(D(c') \prec _{lex}D(c)\) as \((D(c'))_1 = \textrm{len}(\rho ') + \textrm{len}(\sigma ') < \textrm{len}(\rho ) + \textrm{len}(\sigma ) = (D(c))_1\). This contradicts the assumption that \((\rho , \sigma , \tau )\) has minimum cost.
Suppose \(p(\textbf{u})\xrightarrow {\rho }p'(\textbf{x})\). Since there is no monotone cycle decomposition, applying Lemma 3 to \(\rho \) returns a path \(\rho '\) with a narrow linear form such that \(p(\textbf{u})\xrightarrow {\rho '}p'(\textbf{x}')\) where \(\textbf{x}' \ge \textbf{x}\) and \(\textrm{len}(\rho ') \le \textrm{len}(\rho )\).
Cycle \(\sigma \) cannot be decomposed further. Towards a contradiction, assume that there is a monotone cycle decomposition \((\rho ', \sigma ', \tau ')\) of \(\sigma \). Observe that the following monotone cycle decomposition \(c' = (\rho \rho ', \sigma ', \tau ' \tau )\) of \(\pi \) has lower cost \(D(c') \prec _{lex}D(c)\) as \((D(c'))_1 = \textrm{len}(\rho ) + \textrm{len}(\rho ') + \textrm{len}(\sigma ') \le \textrm{len}(\rho ) + \textrm{len}(\sigma ) = (D(c))_1\) and \((D(c'))_2 = \textrm{len}(\sigma ') < \textrm{len}(\sigma ) = (D(c))_2\). This contradicts the assumption that \((\rho , \sigma , \tau )\) has minimum cost.
Suppose \(p'(\textbf{x})\xrightarrow {\sigma }p'(\textbf{y})\). Since there is no monotone cycle decomposition, applying Lemma 3 to \(\sigma \) returns a path \(\sigma '\) with a narrow linear form such that \(p'(\textbf{x})\xrightarrow {\sigma '}p'(\textbf{y}')\) where \(\textbf{y}' \ge \textbf{y}\) and \(\textrm{len}(\sigma ') \le \textrm{len}(\sigma )\). In particular, it is also true that \(\textrm{eff}(\sigma ') \ge \textrm{eff}(\sigma ) > \textbf{0}\).
Replacing \(\rho '\) for \(\rho \) and \(\sigma '\) for \(\sigma \) in \(\pi \) yields a path \(\pi '\). Clearly if \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{w})\) where \(\textbf{w} \ge \textbf{v}\), then \(p(\textbf{u})\xrightarrow {\pi '}q(\mathbf{w'})\) where \(\textbf{w}' \ge \textbf{w} \ge \textbf{v}\). Finally, \((\rho ', \sigma ', \tau )\) is monotone cycle decomposition of \(\pi '\) such that \(\textrm{eff}(\sigma ') > \textbf{0}\) and \(\rho '\) and \(\sigma '\) have narrow linear forms, as required. \(\square \)
We now aim to obtain a narrow linear form for \(\tau \). Note that Lemma 4 gives us a monotone cycle \(\sigma \) with positive effect on at least one counter, i.e. \(\textrm{eff}(\sigma ) > \textbf{0}\). By pumping \(\sigma \) we can force one of the counters to take an arbitrarily large value (following, the vector \(\textbf{x}\) reflects this large value for Lemma 5). Then, loosely speaking, the problem reduces to coverability in 1VASS. However, proving the existence of a polynomial size compressed linear form path in Theorem 1 requires more care. Note that Lemma 5 is stated for 2VASS (not necessarily with one unary counter). First we need to recall the following bound on counter values observed throughout runs. Recall that \(V _{\max } \,{:}{=}\, Q  + T  \cdot T _{\max }\) is the pseudopolynomial size of the input.
Theorem 2
(Corollary from Theorem 3.2 in [4]). Consider a 2VASS (with both counters in binary) \(V = (Q,T)\) and let \(p(\textbf{u})\xrightarrow {*}q(\textbf{v})\), then there exists a run \(p(\textbf{u}) = q_0(\textbf{v}_0)\), \(q_1(\textbf{v}_{1}), \ldots , q_m(\textbf{v}_m) = q(\textbf{v})\) such that \(\textbf{v}_0 _{\max }, \textbf{v}_1 _{\max }, \ldots ,\) \(\textbf{v}_m _{\max } \le (V _{\max } + \textbf{u} _{\max } + \textbf{v} _{\max })^{\mathcal {O}(1)}\).
In the following lemma, that is proved in the appendix, given a 2VASS V, the initial configuration \(p(\textbf{u})\), and target configuration \(q(\textbf{v})\), we write B in place of \((V _{\max } + \textbf{u} _{\max } + \textbf{v} _{\max })^{\mathcal {O}(1)}\) from Theorem 2 and we fix \(\textbf{x} = (4BQ ^2V _{\max }^2, 0)\).
Lemma 5
Consider a 2VASS (with both counters in binary) \(V = (Q,T)\) and let \(p(\textbf{u})\xrightarrow {*}q(\textbf{v})\), then there exists a narrow linear form path \(\pi '\) such that \(p(\mathbf{u + x})\xrightarrow {\pi '}q(\textbf{v}')\) for some \(\textbf{v}' \ge \textbf{v}\).
6 Proof of Theorem 1
Before proving Theorem 1, we employ the fact that for a general 2VASS, not necessarily with one unary counter, the exponents of cycles in linear forms can be pseudopolynomially bounded.
Lemma 6
(Corollary from Lemma 18 in [5]). Let \(\pi \) be path in a 2VASS with a linear form \(\pi = \tau _0 \gamma _1^{f_1} \tau _1 \ldots \gamma _k^{f_k} \tau _k\) such that \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{v})\). Then there exist a path \(\pi ' = \tau _0 \gamma _1^{e_1} \tau _1 \cdots \tau _{k1} \gamma _k^{e_k} \tau _k\) such that \(p(\textbf{u})\xrightarrow {\pi '}q(\textbf{v}')\) where \(\textbf{v}' \ge \textbf{v}\) and \(\textrm{bitsize}(e_1), \ldots , \textrm{bitsize}(e_k)\) are all bounded by a polynomial in \(V  + \textrm{bitsize}(\textbf{u}) + \textrm{bitsize}(\textbf{v})\).
Proof
of Theorem 1. Let \(p(\textbf{u})\xrightarrow {\pi }q(\textbf{v})\) for some path \(\pi \). If there is a narrow linear form \(\ell \) for \(\pi \) then by Lemma 6 we obtain \(\pi ' = \tau _0 \gamma _1^{e_1} \tau _1 \cdots \tau _{k1} \gamma _k^{e_k} \tau _k\) such that \(p(\textbf{u})\xrightarrow {\pi '}q(\textbf{v}')\) where \(\textbf{v}' \ge \textbf{v}\) and \(\textrm{bitsize}(e_1), \ldots , \textrm{bitsize}(e_k)\) are all bounded above by a polynomial in \(V  + \textrm{bitsize}(\textbf{u}) + \textrm{bitsize}(\textbf{v})\). Since \(\ell \) is a narrow linear form, we know that \(k \le P(Q )\) so \(\sum _{i = 1}^k \textrm{len}(\gamma _i) \le kQ  \le Q P(Q )\) and we also know that \(\sum _{i=0}^{k} \textrm{len}(\tau _i) \le Q (P(Q )+1)\). Together, this implies the linear form path \(\pi '\) is of polynomial size.
It remains to consider the case when there is no narrow linear form \(\ell \) for \(\pi \). By Lemma 4 (via Lemma 3) there exists a path \(\pi '\) such that \(p(\textbf{u})\xrightarrow {\pi '}q(\textbf{v}')\) and \(\textbf{v}' \ge \textbf{v}\). Moreover, there is a monotone cycle decomposition \((\rho , \sigma , \tau )\) of \(\pi '\) such that \(\textrm{eff}(\sigma ) > \textbf{0}\) and there are narrow linear forms for both \(\rho \) and \(\sigma \).
Assume that \((\textrm{eff}(\sigma ))_1 > 0\). This is without loss of generality because if \((\textrm{eff}(\sigma ))_1 = 0\) then one can flip the coordinates in V, \(\textbf{u}\) and \(\textbf{v}\) (for the remainder of the proof it will not matter that one counter is unary). Let \(p'(\textbf{m})\) be the configuration such that \(p(\textbf{u})\xrightarrow {\;\rho \;}p'(\textbf{m})\xrightarrow {\sigma \tau }q(\textbf{v}')\). Observe that since \(\textrm{eff}(\sigma ) > \textbf{0}\) for every \(i \in \mathbb {N}\) the path \(\rho \sigma ^i\) induces the run \(p(\textbf{u})\xrightarrow {\rho \sigma ^i}p'(\textbf{m}+ i\cdot \textrm{eff}(\sigma ))\). Consider \(x = (\textbf{x})_1 = 4BQ ^2V _{\max }^2\) (for Lemma 5), clearly x is large enough so that \(p(\textbf{u})\xrightarrow {\rho \sigma ^x}p'(\textbf{m}')\) and \(\textbf{m}' \ge \textbf{m} + \textbf{x}\). By Lemma 5 there exists a narrow linear form for a path \(\tau '\) such that \(p'(\textbf{m}')\xrightarrow {\tau '}q(\textbf{v}'')\) and \(\textbf{v}'' \ge \textbf{v}'\).
We conclude by considering the compressed linear form path \(\rho \sigma ^x \tau '\) such that \(p(\textbf{u})\xrightarrow {\rho \sigma ^x \tau '}q(\textbf{v}'')\) and \(\textbf{v}'' \ge \textbf{v}' \ge \textbf{v}\). Since \(\rho \), \(\sigma \), and \(\tau '\) have narrow linear forms, we can also bound the exponents using Lemma 6 as in the beginning of this proof. Finally, \(\textrm{bitsize}(x)\) is polynomial in \(V  + \textrm{bitsize}(\textbf{u}) + \textrm{bitsize}(\textbf{v})\) much like the exponents of the cycles in the linear forms. Therefore, the size of the compressed linear form \(\rho \sigma ^x \tau '\) is polynomial in \(V  + \textrm{bitsize}(\textbf{u}) + \textrm{bitsize}(\textbf{v})\). \(\square \)
7 Conclusion and Future Work
In this paper we proved that coverability in 2VASS with one unary counter is in NP, a drop in complexity from PSPACE for general 2VASS. We achieve this by using our new techniques. Most notably, we polynomially bounded the number of short cycles that need to be used (Section 4). Then, we attempt to find a polynomial linear form path by replacing short cycles and reshuffling the path (Section 5).
A natural extension is to consider whether coverability in 3VASS with one binary counter and two unary counters is also in NP. More generally, there is the problem of determining the complexity of coverability in kVASS with one binary counter and \(k1\) unary counters. The technique for polynomially bounding the number of short cycles that need be used can easily be generalised to these higher dimension VASS with only one binary counter. However, it is not clear how to modify and use our reshuffling technique. Another open problem is whether reachability in 2VASS with one unary counter is also in NP. Note that completeness would immediately follow from the fact that reachability in binary encoded 1VASS is NPhard [22].
References
Shaull Almagor, Udi Boker, Piotr Hofman, and Patrick Totzke. Parametrized Universality Problems for OneCounter Nets. In Igor Konnov and Laura Kovács, editors, 31st International Conference on Concurrency Theory, CONCUR 2020, September 14, 2020, Vienna, Austria (Virtual Conference), volume 171 of LIPIcs, pages 47:1–47:16. Schloss Dagstuhl  LeibnizZentrum für Informatik, 2020. https://doi.org/10.4230/LIPIcs.CONCUR.2020.47
Shaull Almagor, Nathann Cohen, Guillermo A. Pérez, Mahsa Shirmohammadi, and James Worrell. Coverability in 1VASS with Disequality Tests. In Igor Konnov and Laura Kovács, editors, 31st International Conference on Concurrency Theory, CONCUR 2020, September 14, 2020, Vienna, Austria (Virtual Conference), volume 171 of LIPIcs, pages 38:1–38:20. Schloss Dagstuhl  LeibnizZentrum für Informatik, 2020. https://doi.org/10.4230/LIPIcs.CONCUR.2020.38
Sanjeev Arora and Boaz Barak. Computational complexity: a modern approach. Cambridge University Press, 2009
Michael Blondin, Matthias Englert, Alain Finkel, Stefan Göller, Christoph Haase, Ranko Lazić, Pierre McKenzie, and Patrick Totzke. The Reachability Problem for TwoDimensional Vector Addition Systems with States. J. ACM, 68(5):34:1–34:43, 2021. https://doi.org/10.1145/3464794
Michael Blondin, Alain Finkel, Stefan Göller, Christoph Haase, and Pierre McKenzie. Reachability in TwoDimensional Vector Addition Systems with States Is PSPACEComplete. In 30th Annual ACM/IEEE Symposium on Logic in Computer Science, LICS 2015, Kyoto, Japan, July 610, 2015, pages 32–43. IEEE Computer Society, 2015. https://doi.org/10.1109/LICS.2015.14
Michael Blondin, Christoph Haase, and Philip Offtermatt. Directed reachability for infinitestate systems. In Jan Friso Groote and Kim Guldstrand Larsen, editors, Tools and Algorithms for the Construction and Analysis of Systems  27th International Conference, TACAS 2021, Held as Part of the European Joint Conferences on Theory and Practice of Software, ETAPS 2021, Luxembourg City, Luxembourg, March 27  April 1, 2021, Proceedings, Part II, volume 12652 of Lecture Notes in Computer Science, pages 3–23. Springer, 2021. https://doi.org/10.1007/9783030720131_1
Mikołaj Bojańczyk, Claire David, Anca Muscholl, Thomas Schwentick, and Luc Segoufin. Twovariable logic on data words. ACM Trans. Comput. Log., 12(4):27:1–27:26, 2011. https://doi.org/10.1145/1970398.1970403
Hubert Comon and Yan Jurski. Multiple counters automata, safety analysis and presburger arithmetic. In Alan J. Hu and Moshe Y. Vardi, editors, Computer Aided Verification, 10th International Conference, CAV ’98, Vancouver, BC, Canada, June 28  July 2, 1998, Proceedings, volume 1427 of Lecture Notes in Computer Science, pages 268–279. Springer, 1998. https://doi.org/10.1007/BFb0028751
Wojciech Czerwiński, Slawomir Lasota, Ranko Lazic, Jérôme Leroux, and Filip Mazowiecki. Reachability in Fixed Dimension Vector Addition Systems with States. In Igor Konnov and Laura Kovács, editors, 31st International Conference on Concurrency Theory, CONCUR 2020, September 14, 2020, Vienna, Austria (Virtual Conference), volume 171 of LIPIcs, pages 48:1–48:21. Schloss Dagstuhl  LeibnizZentrum für Informatik, 2020. https://doi.org/10.4230/LIPIcs.CONCUR.2020.48
Wojciech Czerwiński, Sławomir Lasota, Ranko Lazic, Jérôme Leroux, and Filip Mazowiecki. The Reachability Problem for Petri Nets Is Not Elementary. J. ACM, 68(1):7:1–7:28, 2021. https://doi.org/10.1145/3422822
Wojciech Czerwiński, Sławomir Lasota, Christof Löding, and Radoslaw Piórkowski. New Pumping Technique for 2Dimensional VASS. In Peter Rossmanith, Pinar Heggernes, and JoostPieter Katoen, editors, 44th International Symposium on Mathematical Foundations of Computer Science, MFCS 2019, August 2630, 2019, Aachen, Germany, volume 138 of LIPIcs, pages 62:1–62:14. Schloss Dagstuhl  LeibnizZentrum für Informatik, 2019. https://doi.org/10.4230/LIPIcs.MFCS.2019.62
Wojciech Czerwiński and Łukasz Orlikowski. Reachability in Vector Addition Systems is Ackermanncomplete. In 62nd IEEE Annual Symposium on Foundations of Computer Science, FOCS 2021, Denver, CO, USA, February 710, 2022, pages 1229–1240. IEEE, 2021. https://doi.org/10.1109/FOCS52979.2021.00120
Wojciech Czerwiński and Łukasz Orlikowski. Lower Bounds for the Reachability Problem in Fixed Dimensional VASSes. In Christel Baier and Dana Fisman, editors, LICS ’22: 37th Annual ACM/IEEE Symposium on Logic in Computer Science, Haifa, Israel, August 2  5, 2022, pages 40:1–40:12. ACM, 2022. https://doi.org/10.1145/3531130.3533357
Alex Dixon and Ranko Lazic. KReach: A Tool for Reachability in Petri Nets. In Armin Biere and David Parker, editors, Tools and Algorithms for the Construction and Analysis of Systems  26th International Conference, TACAS 2020, Held as Part of the European Joint Conferences on Theory and Practice of Software, ETAPS 2020, Dublin, Ireland, April 2530, 2020, Proceedings, Part I, volume 12078 of Lecture Notes in Computer Science, pages 405–412. Springer, 2020. https://doi.org/10.1007/9783030451905_22
Matthias Englert, Piotr Hofman, Sławomir Lasota, Ranko Lazić, Jérôme Leroux, and Juliusz Straszyński. A lower bound for the coverability problem in acyclic pushdown VAS. Inf. Process. Lett., 167:106079, 2021. https://doi.org/10.1016/j.ipl.2020.106079
Matthias Englert, Ranko Lazić, and Patrick Totzke. Reachability in TwoDimensional Unary Vector Addition Systems with States is NLcomplete. In Martin Grohe, Eric Koskinen, and Natarajan Shankar, editors, Proceedings of the 31st Annual ACM/IEEE Symposium on Logic in Computer Science, LICS ’16, New York, NY, USA, July 58, 2016, pages 477–484. ACM, 2016. https://doi.org/10.1145/2933575.2933577
Javier Esparza, Ruslán LedesmaGarza, Rupak Majumdar, Philipp J. Meyer, and Filip Niksic. An SMTBased Approach to Coverability Analysis. In Armin Biere and Roderick Bloem, editors, Computer Aided Verification  26th International Conference, CAV 2014, Held as Part of the Vienna Summer of Logic, VSL 2014, Vienna, Austria, July 1822, 2014. Proceedings, volume 8559 of Lecture Notes in Computer Science, pages 603–619. Springer, 2014. https://doi.org/10.1007/9783319088679_40
John Fearnley and Marcin Jurdzinski. Reachability in twoclock timed automata is pspacecomplete. In Fedor V. Fomin, Rusins Freivalds, Marta Z. Kwiatkowska, and David Peleg, editors, Automata, Languages, and Programming  40th International Colloquium, ICALP 2013, Riga, Latvia, July 812, 2013, Proceedings, Part II, volume 7966 of Lecture Notes in Computer Science, pages 212–223. Springer, 2013. https://doi.org/10.1007/9783642392122_21
Alain Finkel, Jérôme Leroux, and Grégoire Sutre. Reachability for TwoCounter Machines with One Test and One Reset. In Sumit Ganguly and Paritosh K. Pandya, editors, 38th IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science, FSTTCS 2018, December 1113, 2018, Ahmedabad, India, volume 122 of LIPIcs, pages 31:1–31:14. Schloss Dagstuhl  LeibnizZentrum für Informatik, 2018. https://doi.org/10.4230/LIPIcs.FSTTCS.2018.31
Alain Finkel and Arnaud Sangnier. Mixing coverability and reachability to analyze VASS with one zerotest. In Jan van Leeuwen, Anca Muscholl, David Peleg, Jaroslav Pokorný, and Bernhard Rumpe, editors, SOFSEM 2010: Theory and Practice of Computer Science, 36th Conference on Current Trends in Theory and Practice of Computer Science, Spindleruv Mlýn, Czech Republic, January 2329, 2010. Proceedings, volume 5901 of Lecture Notes in Computer Science, pages 394–406. Springer, 2010. https://doi.org/10.1007/9783642112669_33
Pierre Ganty and Rupak Majumdar. Algorithmic verification of asynchronous programs. ACM Trans. Program. Lang. Syst., 34(1):6:1–6:48, 2012. https://doi.org/10.1145/2160910.2160915
Christoph Haase, Stephan Kreutzer, Joël Ouaknine, and James Worrell. Reachability in Succinct and Parametric OneCounter Automata. In Mario Bravetti and Gianluigi Zavattaro, editors, CONCUR 2009  Concurrency Theory, 20th International Conference, CONCUR 2009, Bologna, Italy, September 14, 2009. Proceedings, volume 5710 of Lecture Notes in Computer Science, pages 369–383. Springer, 2009. https://doi.org/10.1007/9783642040818_25
John E. Hopcroft and JeanJacques Pansiot. On the Reachability Problem for 5Dimensional Vector Addition Systems. Theor. Comput. Sci., 8:135–159, 1979. https://doi.org/10.1016/03043975(79)900410
S. Rao Kosaraju. Decidability of Reachability in Vector Addition Systems (Preliminary Version). In Harry R. Lewis, Barbara B. Simons, Walter A. Burkhard, and Lawrence H. Landweber, editors, Proceedings of the 14th Annual ACM Symposium on Theory of Computing, May 57, 1982, San Francisco, California, USA, pages 267–281. ACM, 1982. https://doi.org/10.1145/800070.802201
JeanLuc Lambert. A Structure to Decide Reachability in Petri Nets. Theor. Comput. Sci., 99(1):79–104, 1992. https://doi.org/10.1016/03043975(92)90173D
Jérôme Leroux. The Reachability Problem for Petri Nets is Not Primitive Recursive. In 62nd IEEE Annual Symposium on Foundations of Computer Science, FOCS 2021, Denver, CO, USA, February 710, 2022, pages 1241–1252. IEEE, 2021. https://doi.org/10.1109/FOCS52979.2021.00121
Jérôme Leroux and Sylvain Schmitz. Reachability in Vector Addition Systems is PrimitiveRecursive in Fixed Dimension. In 34th Annual ACM/IEEE Symposium on Logic in Computer Science, LICS 2019, Vancouver, BC, Canada, June 2427, 2019, pages 1–13. IEEE, 2019. https://doi.org/10.1109/LICS.2019.8785796
Jérôme Leroux and Grégoire Sutre. On Flatness for 2Dimensional Vector Addition Systems with States. In Philippa Gardner and Nobuko Yoshida, editors, CONCUR 2004  Concurrency Theory, 15th International Conference, London, UK, August 31  September 3, 2004, Proceedings, volume 3170 of Lecture Notes in Computer Science, pages 402–416. Springer, 2004. https://doi.org/10.1007/9783540286448_26
Jérôme Leroux and Grégoire Sutre. Reachability in TwoDimensional Vector Addition Systems with States: One Test Is for Free. In Igor Konnov and Laura Kovács, editors, 31st International Conference on Concurrency Theory, CONCUR 2020, September 14, 2020, Vienna, Austria (Virtual Conference), volume 171 of LIPIcs, pages 37:1–37:17. Schloss Dagstuhl  LeibnizZentrum für Informatik, 2020. https://doi.org/10.4230/LIPIcs.CONCUR.2020.37
Jérôme Leroux, Grégoire Sutre, and Patrick Totzke. On the Coverability Problem for Pushdown Vector Addition Systems in One Dimension. In Magnús M. Halldórsson, Kazuo Iwama, Naoki Kobayashi, and Bettina Speckmann, editors, Automata, Languages, and Programming  42nd International Colloquium, ICALP 2015, Kyoto, Japan, July 610, 2015, Proceedings, Part II, volume 9135 of Lecture Notes in Computer Science, pages 324–336. Springer, 2015. https://doi.org/10.1007/9783662476666_26
Richard Lipton. The Reachability Problem Requires Exponential Space. Department of Computer Science. Yale University, 62, 1976
Ernst W. Mayr. An Algorithm for the General Petri Net Reachability Problem. SIAM J. Comput., 13(3):441–460, 1984. https://doi.org/10.1137/0213029
Marvin L. Minsky. Computation: Finite and Infinite Machines. PrenticeHall, Inc., 1967
C. Petri. Kommunikation mit Automaten, Ph. D. dissertation. University of Bonn, 1962
Charles Rackoff. The Covering and Boundedness Problems for Vector Addition Systems. Theor. Comput. Sci., 6:223–231, 1978. https://doi.org/10.1016/03043975(78)900361
Klaus Reinhardt. Reachability in Petri Nets with Inhibitor Arcs. Electron. Notes Theor. Comput. Sci., 223:239–264, 2008. https://doi.org/10.1016/j.entcs.2008.12.042
Sylvain Schmitz. The Complexity of Reachability in Vector Addition Systems. ACM SIGLOG News, 3(1):4–21, 2016. URL: https://dl.acm.org/citation.cfm?id=2893585
Leslie G. Valiant and Mike Paterson. Deterministic OneCounter Automata. J. Comput. Syst. Sci., 10(3):340–350, 1975. https://doi.org/10.1016/S00220000(75)800055
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
Copyright information
© 2023 The Author(s)
About this paper
Cite this paper
Mazowiecki, F., SinclairBanks, H., Węgrzycki, K. (2023). Coverability in 2VASS with One Unary Counter is in NP. In: Kupferman, O., Sobocinski, P. (eds) Foundations of Software Science and Computation Structures. FoSSaCS 2023. Lecture Notes in Computer Science, vol 13992. Springer, Cham. https://doi.org/10.1007/9783031308291_10
Download citation
DOI: https://doi.org/10.1007/9783031308291_10
Published:
Publisher Name: Springer, Cham
Print ISBN: 9783031308284
Online ISBN: 9783031308291
eBook Packages: Computer ScienceComputer Science (R0)