A survey on observability of Boolean control networks

Observability is a fundamental property of a partially observed dynamical system, which means whether one can use an input sequence and the corresponding output sequence to determine the initial state. Observability provides bases for many related problems, such as state estimation, identification, disturbance decoupling, controller synthesis, etc. Until now, fundamental improvement has been obtained in observability of Boolean control networks (BCNs) mainly based on two methods—Edward F. Moore’s partition and our observability graph (or their equivalent representations found later based on the semitensor product (STP) of matrices (where the STP was proposed by Daizhan Cheng)), including necessary and sufficient conditions for different types of observability, extensions to probabilistic Boolean networks (PBNs) and singular BCNs, even to nondeterministic finite-transition systems (NFTSs); and the development (with the help of the STP of matrices) in related topics, such as computation of smallest invariant dual subspaces of BNs containing a set of Boolean functions, multiple-experiment observability verification/decomposition in BCNs, disturbance decoupling in BCNs, etc. This paper provides a thorough survey for these topics. The contents of the paper are guided by the above two methods. First, we show that Moore’s partition-based method closely relates the following problems: computation of smallest invariant dual subspaces of BNs, multiple-experiment observability verification/decomposition in BCNs, and disturbance decoupling in BCNs. However, this method does not apply to other types of observability or nondeterministic systems. Second, we show that based on our observability graph, four different types of observability have been verified in BCNs, verification results have also been extended to PBNs, singular BCNs, and NFTSs. In addition, Moore’s partition also shows similarities between BCNs and linear time-invariant (LTI) control systems, e.g., smallest invariant dual subspaces of BNs containing a set of Boolean functions in BCNs vs unobservable subspaces of LTI control systems, the forms of quotient systems based on observability decomposition in both types of systems. However, there are essential differences between the two types of systems, e.g., “all plausible definitions of observability in LTI control systems turn out to be equivalent” (by Walter M. Wonham 1985), but there exist nonequivalent definitions of observability in BCNs; the quotient system based on observability decomposition always exists in an LTI control system, while a quotient system based on multiple-experiment observability decomposition does not always exist in a BCN.


Background
In the background of control, it is well known and has been widely accepted that controllability and observability are the most fundamental properties in control systems in the state-space form [1] since 1960. A control system (usually a physical system, represented by a differential equation) is controllable if for every two states x and x , there is an input B Kuize Zhang kuize.zhang@surrey.ac.uk 1 Department of Computer Science, University of Surrey, Guildford GU2 7XH, UK sequence that drives x to x starting at some initial time. Usually, the state of a control system is partially observed, which is represented by an output function. A control system is observable if the initial state can be determined by an input sequence and the corresponding output sequence.
In his seminal 1963 paper [2], Kalman proved that a linear control system is an irreducible realization of an impulse-response matrix if and only if the system is completely controllable and completely observable. In addition, he also gave a canonical form of a linear control system in terms of four mutually exclusive parts: (A) completely controllable but unobservable, (B) completely controllable and completely observable, (C) uncontrollable and unobservable, (D) uncontrollable but completely observable. The form was later called Kalman decomposition. Following these seminal papers, the two notions have been extended to many different kinds of control systems using/developing new mathematical tools. The control systems cover linear systems [2,3], nonlinear systems [4][5][6], switched systems [7], networked systems [8,9], etc.
It might not be well known to the control community that controllability and observability are also two fundamental properties in the computer science community. They are in different terms, and appeared even earlier. On [2, pp. 169], Kalman compared his irreducible realization with E.F. Moore's minimization of machines (in his seminal 1956 paper [10]): "The class of all machines which are indistinguishable from a given strongly connected machine S by any single experiment has a unique (up to isomorphism) member with a minimal number of states. This unique machine, called the reduced form of S, is strongly connected and has the property that any two of its states are distinguishable". Moore's seminal paper stimulated many branches in computer science which are currently called model-based testing, see the very early book [11] and recent monograph [12], as well as references therein, where [11] seems to be the first book on topics that received attention in both control science and computer science. The sequential machines studied in [10] were later called Moore machines, which consist of finitely many states, inputs, and outputs. The state update/transition function of a Moore machine is a function that maps every pair of a state and an input to a state, the output function is a function that maps a state to an output. Moore machines are actually among the earliest models of computer science. They are deterministic, in almost the same form compared with control systems, but discrete in both time and space. By recasting controllable into a Moore machine, we obtain strongly connected, because the following two conditions are equivalent: (1) for every two states, there is an input sequence that drives the one to the other, (2) the state-transition graph 1 of a Moore machine is strongly connected. Hence, the terminology "strongly connected" in [10] means "completely controllable" in [2] (see [2, pp. 170]). That is, this notion performs complete correspondence in linear control systems and Moore machines. On the other hand, indistinguishability of two states in Moore machines means that under each input sequence, the two output sequences generated by the two states are the same. Hence, the "existence of two indistinguishable states" seems to correspond to unobservability of linear control systems. However, it is not true, because Moore machines are essentially nonlinear. This incorrespondence will be directly reflected by the fact that there are nonequivalent notions of observability for Moore machines, but "all plausible definitions of observability in linear time-invariant control systems turn out to be equivalent" [3], because the difference of two solutions of a linear time-invariant control system is still its solution, resulting in that different definitions of observability do not depend on inputs. The closedness of difference of solutions in linear time-invariant control systems also results in that controllability is dual to observability in this type of systems: a linear time-invariant control system is controllable if and only if its dual system is observable.
To sum up, there is essential difference between linear control systems and Moore machines due to various notions of observability, while there is an extreme similarity between irreducible realization of linear control systems and minimization of Moore machines. Except for this similarity, there is another extreme similarity-verification of multiple-experiment observability 2 (see Definition 12) of Moore machines [10, Theorem 6] and unobservable subspace computation in linear time-invariant control systems [3]. This similarity is due to the finiteness of the number of inputs and states in Moore machines and the linearity feature and finiteness of dimensions of linear systems. This similarity plays a central role in the solvability of several control problems in Boolean control networks (BCNs) (they were proposed in [14,15]). In hindsight, the mathematical essence in the multiple-experiment observability decomposition [16] in BCNs, the mathematical essence in the disturbance decoupling problem of Boolean networks (BNs) [17], the invariant dual subspaces (actually unobservable subspaces) of BNs [18], and the observability verification method [19] for BNs, turn out to coincide with Moore's partition-based verification method for multiple-experiment observability of Moore machines [10, Theorem 6]. See Sect. 3 for details. In [20] and [13,Remark 4.1], Moore's partition-based method was briefly restated. The observability verification method used in [20] is, however, essentially different from Moore's method. The method used in [20] is an extension of the weightedpair-graph method (actually an inverse method used for verifying the negation of observability) proposed in [21]. Different from Moore's method, the weighted-pair-graph method applies to more notions of observability, and can be extended to nondeterministic finite-transition systems [20]. The weighted-pair-graph method (with its slight variants) has been widely used in observability verification until now, e.g., in BCNs [22][23][24][25][26], in probabilistic BNs [27,28], in singular BCNs [29], as well as in observability disturbance analysis of BNs [30]. See Sect. 4 for details.
Now we can introduce the model studied in the current paper-Boolean control networks (BCNs). What are the relations between BCNs and the two models-linear control systems and Moore machines, as discussed above? BCNs are special Moore machines for which the numbers of states, inputs, and outputs are all powers of 2. BNs are BCNs whose inputs are constant. Now that BCNs are special Moore machines, is it true that one only needs to study Moore machines, but does not need to study BCNs separately? The answer is No, because when mentioning a Moore machine, its size is the numbers of states, inputs, outputs, and transitions (between states driven by inputs); while mentioning a BCN, its size is the numbers of state nodes, input nodes, output nodes, and the lengths of its updating functions and output functions which are all Boolean functions. In other words, the focus is mainly on states, inputs, and outputs in Moore machines, while the focus is mainly on state nodes, input nodes, and output nodes in BCNs. For example, if an algorithm for verifying observability of a BCN runs in time exponential in the size of the BCN, then it runs in time polynomial in the size of the Moore machine represented by the BCN. As a matter of fact, not all Moore machines can be represented by BCNs. Hereinafter, when mentioning an algorithm for a Moore machine, its time complexity and space complexity are in the size of the Moore machine; when mentioning an algorithm for a BCN, its time complexity and space complexity are in the size of the BCN. Another difference between BCNs and Moore machines lies in that each node of a BCN can be regarded as a variable, then it makes sense to endow a BCN with a coordinate, which does not apply to a Moore machine. Hence, it makes sense to study coordinate-based problems in BCNs such as the disturbance decoupling problem [17,31] and the observability decomposition problem [16]. It is worthy mentioning that the disturbance decoupling problem can also be defined in a coordinate-independent form, which was called the "original disturbance decoupling problem" in [16]. The semitensor product of matrices first proposed by Cheng [32] in 2001 is a convenient and appropriate tool used for representing coordinates and transformations between different coordinates in BCNs [33,34]. Here we want to point out that the area of the control-theoretic problems of BCNs and the area of modelbased testing in computer science are quite far from each other, and many results obtained in the latter area were not known to most researchers in the former area. Of course, the latter area is much more developed. Later when introducing specific topics, we will introduce relevant overlaps.
As mentioned above, the controllability of a Moore machine is the same as the strong connectedness of its statetransition graph. Hence, it appears that it is not worthy devoting much space to the discussion on controllability of Moore machines, because it can be verified in linear time by Tarjan's algorithm. Hence, controllability of a BCN can be verified in exponential time also by Tarjan's algorithm. On the other hand, verifying controllability of BCNs is NP-hard [35]. Hence, the interesting topics on controllability of BCNs should be on developing fast algorithms for special types of BCNs [36]. Also as mentioned before, there exist nonequivalent notions of observability in Moore machines, so does for BCNs [37], even for controllable BCNs [38]. Therefore, it is interesting to discuss the relations between different notions of BCNs.
In the following, we start to show the main parts of the survey. Different from many other survey papers in which the principal line is based on problems/definitions, we choose the principle line based on methods. Our choice is more fundamental, because in this area, there are quite a lot of problems/definitions, but there are only a very small number of methods (up to equivalence).

Notation
• ∅: empty set • ⊂ and : subset and strict subset relations, respectively • 2 X : power set of set X • Z (Z + ): set of (positive) integers • N: set of natural numbers (including 0) • R n : set of n-length real column vectors • R m×n : set of m × n real matrices • D: set {0, 1} • ∧, ∨, ¬: logical connectives conjunction, disjunction, negation • δ i n : i-th column of the identity matrix I n

Equivalence relations and partitions
Let X be a set. A partition of X is a set of pairwise disjoint nonempty subsets of X whose union is equal to X , where these subsets are called the parts (or cells) of the partition. A partition of X is finite if it consists of finitely many subsets of X , or is discrete if all cells are singletons. For two partitions α, β of X , we say α refines β, or α is finer than β, or β is coarser than α, or α is a refinement of β, denoted by α β or β α, if for every M ∈ α, there exists N ∈ β such that M ⊂ N . When α strictly refines β, we denote α β or β ≺ α.
On a given set X , a relation ∼ on X is a subset of X ×X . We say two states x, x ∈ X have relation ∼ (or x has relation ∼ with x ) if (x, x ) ∈∼, which is denoted by x ∼ x . A relation ∼⊂ X × X is an equivalence relation if it is reflexive, symmetric, and transitive. That is, for all x, y, z ∈ X , Let relation ∼⊂ X × X be symmetric and transitive. Define [x] ∼ := {x} ∪ {y ∈ X |y = x ∧ y ∼ x}. Not that {y ∈ X |y = x ∧ y ∼ x} may be equal to ∅. One easily sees that for all x, y ∈ X , either Let ∼⊂ X × X be an equivalence relation. For x ∈ X , [x] ∼ (= {y ∈ X |y ∼ x}) is called the equivalence class generated by x, i.e., the set of all elements of X that have the relation ∼ with x. For all x, y ∈ X , [x] ∼ = [y] ∼ if and only if x ∼ y.
For two equivalence relations α, β on a set X , we say α refines β if the partition of X induced by α refines the partition of X induced by β, equivalently, for all x, y in X , if xα y then xβ y.
Let f : X → Y be a function. f defines an equivalence relation on X : for all x, x ∈ X , x and x have this relation if and only if f (x) = f (x ). The partition induced by this relation is called the partition induced by f and is denoted by P( f ). Conversely, given a set X and a partition α of X , α induces (at least) one function f α on X : for all x, x ∈ X , xαx if and only if f α (x) = f α (x ). Clearly, f α induces α.

The notion of Moore machine
where • X is a finite set of states, • U is a finite set of inputs, • f : X × U → X is the update/transition function, • Y is a finite set of outputs, and • h : X → Y is the output function.
Another widely studied class of systems called Mealy machines generalize Moore machines in the sense that h also depends on U , i.e., h is a function from X × U to Y [39].

The notion of Boolean (control) network
Definition 2 A Boolean network (BN) is formulated as the following logical form: where t ∈ N denotes any discrete time step, x i (t), y k (t) ∈ D denote the value of state node x i and the value of output node y j at time step t, f i , h k : D n → D are Boolean functions, i ∈ 1, n , k ∈ 1, q . In this survey, we mainly consider observability, so we need to consider the output function in (1). A BN (1) can be briefly denoted as the compact form where t ∈ N; x(t) ∈ D n and y(t) ∈ D q stand for the state and the output at time step t, respectively; f : D n → D n and h : D n → D q are logical mappings.
The dependency graph of a BN (1) is a directed graph The state-transition graph of a BN (2) is a directed graph For a BN, the size of its state-transition graph is exponential in the size of its dependency graph.

Example 1 Consider the following simple BN:
where t ∈ N; A(t), B(t), C(t) ∈ D. Its dependency graph and state-transition graph are shown in Fig. 1a and 1b, respectively.
Definition 3 A Boolean control network (BCN) is defined by the following logical form: . .
where t ∈ N denotes any discrete time step; x i (t), u j (t), and y k (t) ∈ D denote the values of state node x i , input node u j , and output node y k at time step t, respectively, f i : D m+n → D and h k : D n → D are Boolean functions, i ∈ 1, n , j ∈ 1, m , k ∈ 1, q . A BCN (4) is represented in the compact form where t ∈ N; x(t) ∈ D n , u(t) ∈ D m , and y(t) ∈ D q stand for the state, input, and output of (5) at time step t; f : D m+n → D n and h : D n → D q are logical mappings. By definition, a BCN (5) is a Moore machine, but not vice versa.
The dependency graph of a BCN (4) is a directed graph The state-transition graph of a BCN (5) is a weighted directed graph (V, E, W), where V = D n ; E ⊂ V × V is as follows: for all x, x ∈ D n , (x, x ) ∈ E if and only if there exists u ∈ D m such that x = f (u, x); and for every edge By definition, for a BCN, the size of its state-transition graph is exponential in the size of its dependency graph. In addition, more than one BCN may share the same dependency graph.
Example 2 Consider the following simple BCN: where t ∈ N, A(t), B(t), u(t), y(t) ∈ D. Its dependency graph and state-transition graph are shown in Fig. 2a and 2b, respectively.

The notion of finite automaton
An alphabet Σ is a nonempty finite set such that each finite sequence of elements of Σ has a unique decomposition of elements of Σ. Elements of Σ are called letters, finite sequences 120 K. Zhang where state q 0 is initial (with an input arrow from nowhere), q 2 is final (represented by double circles), symbols beside arrows denote letters. Transition q 0 a − → q 1 means that when A 1 is in state q 0 and reads letter a, A 1 transitions to state q 1 . The other transitions can be explained similarly of letters are called words. Σ * denotes the set of all words over Σ, including the empty word . Σ + := Σ * \ { }. Σ ω denotes the set of infinite sequences over Σ. Deterministic finite automata will be used to verify four definitions of observability of BCNs (see Sect. 4.2) as well as observability of nondeterministic finite-transition systems (NFTSs, see Sect. 4.5). Nondeterministic finite automata will be used to verify the observability of NFTSs as well.

Definition 4 A deterministic finite automaton is a quintuple
• q 0 ∈ Q is the initial state (aka start state), and • F ⊂ Q is the set of final states (aka accept states).

Definition 5
Let Σ be an alphabet and A = (Q, Σ, δ, q 0 , F) be a DFA. A word w ∈ Σ * is called accepted by DFA A if δ(q 0 , w) ∈ F. The formal language recognized by DFA A is the set of all words accepted by A, i.e., {w ∈ Σ * |δ(q 0 , w) ∈ F}. A formal language recognized by some DFA is called regular.
In A 1 , all transitions ending with the unique final state q 2 are q 1 a − → q 2 and q 2 a − → q 2 , and the unique transition ending with q 1 is q 0 a − → q 1 . Hence, all transition sequences of length 2 ending with q 2 are q 0 Hence, all words accepted by A 1 end with aa. On the other hand, for every word waa, where w ∈ Σ * , no matter δ(q 0 , w) is equal to q 0 , q 1 , or q 2 , A 1 accepts waa. Hence, A 1 exactly accepts all words ending with aa.
• q 0 ∈ Q is the initial state, and • F ⊂ Q is the set of final states.
There exists a transition q a − → q if and only if q ∈ δ(q, a). When an NFA A is in a state q and reads a letter a, then A can transition to any state of δ(q, a). The transition function δ is also recursively extended to Q × Σ * → 2 Q as usual.

Definition 7 Let Σ be an alphabet and
The formal language recognized by A is the set of all words accepted by A, i.e., {w ∈ Σ * |δ(q 0 , w) ∩ F = ∅}.
In other words, for an NFA A = (Q, Σ, δ, q 0 , F) over an alphabet Σ and a word w = w 1 . . . w n ∈ Σ * , where w 1 , . . . , w n ∈ Σ, A accepts w if and only if there exist states q 1 , . . . , q n such that q i ∈ δ(q i−1 , w i ) for all 1 ≤ i ≤ n and q n ∈ F.
From now on, for an automaton A (either deterministic or nondeterministic), we use Acc(A) to denote its accessible part, i.e., the part of A that is obtained by removing all states that are not reachable from the initial state and their ingoing and outgoing transitions.

The algebraic form of Boolean control networks based on the semitensor product of matrices
Definition 8 [34] Let A ∈ R m×n , B ∈ R p×q , and α = lcm(n, p) be the least common multiple of n and p. The semitensor product (STP) of A and B is defined as where ⊗ denotes the Kronecker product.
From this definition, it is easy to see that the conventional product of matrices is a particular case of the STP, since if n = p then A B = AB. Since the STP keeps most properties of the conventional product, e.g., the associative law [13] 3 , the distributive law, inverse-order laws (e.g., (A B) Tr = B Tr A Tr ), etc. [34], we usually omit the symbol " " hereinafter.
It is very convenient to represent coordinates and coordinate transformations in BCNs using STP. When verifying properties of BCNs, STP is not necessary. As usual, STPbased verification is not very efficient compared with the verification based on state-transition graphs, because the former provides an enumeration of the states of a BCN, while in the latter one only needs to visit necessary states.

Definition 9
Let A ∈ R m×n , B ∈ R p×n . The Khatri-Rao product of A and B is defined by Proposition 1 [34] For a Boolean function f : D n → D, there exists a unique logical matrix F ∈ L 2×2 n such that Proposition 2 [34] For two logical functions F A 1 · · · A n and G A 1 · · · A n , where A i ∈ Δ, i ∈ 1, n , F ∈ L 2 m 1 ×2 n , G ∈ L 2 m 2 ×2 n , one has where H = F * G. Conversely, one also has Proposition 3 [34](Pseudocommutative law) Let A ∈ R m×n and z ∈ R t . Then Definition 10 [34] The matrix M k r = δ 1 k ⊕ · · · ⊕ δ k k is called the power-reducing matrix. Particularly, we denote M 2 r := M r .

Proposition 4 [34]
For power-reducing matrix M k r , we have A widely used property on the Kronecker product is as follows.
The algorithms for transforming the algebraic form of a BN/BCN back to their logical forms can be found in [33,34].

Different definitions of observability of Boolean control networks
Although in this survey we do not consider controllability, we still state its definition for ease of statement.
Clearly, a BCN (5) is controllable if and only if its statetransition graph is strongly connected.
As usual, a verification algorithm for a property is a necessary and sufficient condition for the property that is verifiable in finitely many steps.
Next we state the four definitions of observability BCNs considered in this survey. The terms "***-experiment observability" were chosen from [13], which were rephrased from the term "Gedanken experiment" used in [10]. An experiment is a process for which an input sequence is fed into a BCN to generate an output sequence, to distinguish between the real initial state and the candidates of the initial state. According to the number of experiments needed, the notions of observability are called "single-experiment" or "multiple-experiment". In this survey, we only consider preset experiments, where all input sequences in experiments are completely determined in advance. In addition, there are another kind of experiments called adaptive experiments, where in an input sequence, every input is determined by the (past) generated output sequence. See [11,39] for details. For example, the existence of adaptive distinguishing sequences (that can determine the initial state) was verified in exponential time in [11], and verified in polynomial time in [39], for Mealy machines. In [26], the existence of adaptive distinguishing sequences was studied for BCNs, but the necessary and sufficient condition for the existence given in [26] is not verifiable.
Definition 12 [10,41] A BCN (5) is called multiple-experiment observable if for every two different initial states x(0), x(0) ∈ D n , there is an input sequence such that the output sequences corresponding to x(0) and x(0) are different. Such an input sequence is called a distinguishing input sequence of x(0) and x(0) .
If a BCN is multiple-experiment observable, then every two different initial states have a distinguishing input sequence, and then the initial state can be determined by doing a finite number of experiments. In the first experiment, assume two different initial states as candidates, and feed one of their distinguishing input sequences into the BCN, then at least one of the obtained output sequences differs from the real output sequence (generated by the real initial state and the input sequence), hence at least one of the two candidates is not the real initial state, and mark each candidate that is not real. In the second experiment, repeat the first experiment by assuming two different initial states as candidates from the unmarked initial states. Repeat the experiment until there is only one unmarked initial state left. Then the only state is the real initial state.
In [10], the notion of observability was not explicitly mentioned, but a polynomial-time algorithm was given to verify multiple-experiment observability of Moore machines (see the proof of Theorem 6 therein), i.e., the algorithm runs in exponential time for BCNs. The algorithm is based on an equivalence relation ∼ k : two states x and x have this relation if and only if x and x have no distinguishing input sequence of length k. See Sect. 3 for details. The notion of "indistinguishability" was formally defined and verified in [10] as well. In [41], a sufficient but not necessary condition was given for multiple-experiment observability of BCNs.
Definition 13 [33,39] A BCN (5) is called strongly multipleexperiment observable if for every initial state x(0) ∈ D n , there exists an input sequence such that for each initial state x(0) ∈ D n different from x(0), the output sequences corresponding to x(0) and x(0) are different. Such an input sequence is called a distinguishing input sequence of x(0).
If a BCN is strongly multiple-experiment observable, then every initial state has a distinguishing input sequence, and then the initial state can be determined also by doing a finite number of experiments (sometimes only one experiment is enough). In the first experiment, assume one initial state x 0 as a candidate, and feed one of its distinguishing input sequences into the BCN, then x 0 is the real initial state x 0 if and only if the output sequence generated by x 0 and U is the same as the real output sequence (generated by x 0 and U ). With any luck, after the first experiment, the real initial state can be found. With no luck, after repeating this experiment for finitely many times for different candidates, the real initial state will be determined. Definition 13 is strictly stronger than Definition 12 [37,38].
In [39], the distinguishing input sequence of x(0) was called a unique input output sequence of x(0), and its existence problem was proven to be PSPACE-complete for Mealy machines. Hence, the algorithm runs in doubly exponential time for BCNs. The verification algorithm for the existence of a distinguishing input sequence of x(0) is as follows: apply an input sequence to the pair (x(0), {x ∈ D n |x = x(0), h(x) = h(x(0))}) and see whether some pair whose right component is empty can be reached, where in each reachable pair, the right component must be maximal and the states in the right component must produce the same output as the left component. The method proposed in [39] is not equivalent to the partition-based method proposed in [10] (see Sect. 3) or the inverse method proposed in [21] (see Sect. 4). In [33], a necessary and sufficient condition for strong multiple experiment observability of controllable BCNs was given, which is sufficient but not necessary for general BCNs.
Definition 14 [39,42] A BCN (5) is called single-experiment observable if there exists an input sequence such that for every two different initial states x(0), x(0) ∈ D n , the output sequences corresponding to x(0) and x(0) are different. Such an input sequence is called a distinguishing input sequence of (5).
If a BCN is single-experiment observable, then the initial state can be determined by every distinguishing input sequence. Hence, in order to determine the initial state, one only needs to do one experiment. Definition 14 is strictly stronger than Definition 13 [37,38].
In [39], the single-experiment observability verification problem was proven to be PSPACE-complete for Mealy machines, where the distinguishing input sequence was called preset distinguishing sequence. Hence, the verification algorithm runs in doubly exponential time for BCNs. The verification algorithm is based on the notion of initial state uncertainty: the initial state uncertainty of an input sequence u(0) . . . u(n) is the partition induced by the equivalence relation for which two states x(0) and x(0) have this relation if and only if u(0) . . . u(n) is not a distinguishing input sequence of x(0) and x(0) . Clearly, u(0) . . . u(n) is a distinguishing input sequence of BCN (5) if and only if the partition is discrete. The method proposed in [39] is not equivalent to the partition-based method proposed in [10] or the inverse method proposed in [21]. In [42], a necessary and sufficient condition for single-experiment observability of controllable BCNs was given, which is sufficient but not necessary for general BCNs. (1) . . . , the corresponding output sequences y(0)y(1) . . . and y (0)y (1) . . . are different.
If a BCN is arbitrary-experiment observable, then the initial state can be determined by every sufficiently long input sequence. Hence, to determine the initial state, one also only needs to do one experiment. Definition 15 is strictly stronger than Definition 14 [37,38].
In [43], an exponential-time algorithm for arbitraryexperiment observability of BCNs was given based on the equivalence relation ∼ k defined in [10] (see Sect. 3): a BCN is arbitrary-experiment observable if and only if for every pair of different periodic (state, input)-trajectories of the same minimal period k and the same input trajectory, the corresponding output trajectories are also different and periodic of minimal period k. It is not very convenient to apply this verification algorithm, because one needs to first find an upper bound on k and then check all sequences of length k. Apparently, the method is not equivalent to the inverse method proposed in [21]. The inverse method is easy-to-use.

The method
The partition-based method given by Moore [10] in 1956 is based on a sequence of equivalence relations. We restate the relations in terms of BCNs instead of Moore machines because BCNs are the considered model. Consider a BCN (5). For each k ∈ N, define an equivalence relation ∼ k on its state set D n : for every two states x and x in D n , x ∼ k x if and only if x and x have no k-length distinguishing input sequence. The partition induced by ∼ k is denoted by P(∼ k ). The result in [10, Theorem 6] is as follows.

all its cells are singletons).
Note that the term "observability" did not explicitly appear in [10]. For two states x and x , if they have no k-length distinguishing input sequence, then they have no distinguishing input sequence of length k − 1. Then for all k ∈ N, Moreover, one observes that satisfies that the partitions first strictly refine, and then remain the same, where ind = min{l ∈ N|P(∼ l ) = P(∼ l+1 )}.
Because the cardinality of D n is 2 n , one has ind ≤ 2 n − 1 and P(∼ 2 n −1 ) = P(∼ k ) for all k ≥ 2 n . That is, Theorem 6 holds. Given a partition P(∼ 0 ), one can compute P(∼ ind ) based on the observation (A) as follows. (1) P tmp := P(∼ 0 ). (2) Arbitrarily choose a cell C of P tmp , partition C as follows: two different states x, x of C remain in the same cell if and only if for all u, f (u, x) and f (u, x ) belong to the same cell of P tmp . (3) Repeat (2) until no cell can be divided, and then we obtain P(∼ ind ). Note that the final result P(∼ ind ) is unique and irrespective of the order of dividing cells.

Example 4 Consider BCN
where t ∈ N, x(t) ∈ D 2 , u(t), y(t) ∈ D, f and h are as follows: (a) f .
x 00 01 10 11 for all k > 2. In P(∼ 0 ), 00, 01, 10 form a cell because h(00) = h(01) = h(10) = 0, 11 forms the other cell because h(11) = 1 = h(00). By applying inputs 0 and 1 on states 01 and 10, they still stay in the same cell of P(∼ 0 ), then 01 and 10 are in the same cell of P(∼ 1 ). By applying input 0 on states 00 and 01, they transition into different cells of P(∼ 0 ), then they belong to different cells of P(∼ 1 ). The same argument indicates that states 00 and 10 belong to different cells of P(∼ 1 ). By applying input 0 on states 01 and 10, they transition into different cells of P(∼ 1 ), then they belong to different cells of P(∼ 2 ). Therefore, P(∼ 2 ) is discrete. By Theorem 6, BCN (12) is multiple-experiment observable. By the process of computing P(∼ 0 ), P(∼ 1 ), and P(∼ 2 ), one can find distinguishing input sequences of every pair of states. The empty word is a distinguishing input sequence of 00 and 11, 01 and 11, and 10 and 11. Input 0 is a distinguishing input sequence of 00 and 01, 00 and 10. Input sequence 00 is a distinguishing input sequence of 01 and 10.

An equivalent semitensor-product representation for Theorem 6
Hereinafter, regard a logical matrix L ∈ L R×T as a function from Consider the algebraic form (10) of BCN (5). By definition, the corresponding partition P(∼ 0 ) as defined in (11) is induced by H =: G 0 . Furthermore, the correspond- Hence Theorem 6 can be equivalently represented by the following result.

Proposition 7 A BCN (10) is multiple-experiment observable if and only if there exists k
In the special case that m = 0 (i.e., when the input is constant), Theorem 7 degenerates to the main result obtained in [19].

Relations with invariant dual subspaces and observability of Boolean networks
In [18], a notion of smallest invariant subspace of a BN (8a) containing a set of Boolean functions was studied and an algorithm was designed to compute such a subspace.
, the set of all Boolean functions of z 1 , . . . , z r . Let the structure matrix of z i = f i (x 1 , . . . , x n ) be F i ∈ L 2×2 n , i ∈ 1, r , F = F 1 * · · · * F r ∈ L 2 r ×2 n is called the structure matrix of F {z 1 , . . . , z r }. Next we give a partition-based description for F {z 1 , . . . , z r }.
It is not difficult to prove that for all Boolean functions z 1 , . . . , z r : D n → D, Then the following proposition follows.
• furthermore, That is, the subspace generated by a set of Boolean functions is determined by the partition induced by these Boolean functions. Hence, different sets of Boolean functions may generate the same subspace. A similar feature also appears in linear algebra: different bases may generate the same linear subspace. In addition, the above P defines a bijection from X * to Π , where X * denotes the set of all subspaces generated by sets of Boolean functions on D n , Π denotes the set of all partitions of D n . A subspace F {z 1 , . . . , z r } is called L-invariant [18] if there exists G ∈ L 2 r ×2 r such that F L = G F, where F is the structure matrix of the subspace, and the BN z(t + 1) = Gz(t) is called the dual BN of (8a) with respect to F {z 1 , . . . , z r }. Since F {z 1 , . . . , z n } is generated by a set of functions, in the current paper we call it dual subspace which follows the traditional terms of functional analysis; and then, we call an L-invariant subspace (this term was used in [18]) an L-invariant dual subspace. A number of properties on Linvariant dual subspaces of (8a) containing a set of Boolean functions were characterized in [18].
We give a new explanation for L-invariant dual subspaces from a partition point of view. A dual subspace i.e., the partition induced by F L is coarser than the partition induced by F, i.e., (from a dynamical-system point of view) the BN (8a) coarsens the partition induced by F, i.e., for every two states The structure matrix of the smallest L-invariant dual subspace of (8a) containing F {z 1 , . . . , z r } is computed as as an output function of BN (8a), then corresponding to the BN (L, F), the partition induced by O ind is exactly Moore's partition P(∼ ind ) as defined in (11). This is an essential relation between Moore's partition and the smallest L-invariant dual subspace of (8a) containing F {z 1 , . . . , z r } defined in [18]. By Theorem 6, this smallest L-invariant dual subspace containing F decides the observability of BN (L, F).

Relations with unobservable subspaces of linear time-invariant control systems
Consider the following linear control system: . It is well known that (13) is observable if and only if n−1 i=0 O i = {0}. One directly sees that BCNs (or Moore machines) are discrete in both time and space, but linear control systems are continuous in both time and space. Therefore, they look totally different. Is there any similarity between these two types of systems? It is hard to give a confirmatory answer. But, now there does exist one! The similarity comes from observability. One can see that the chain satisfies that the subspaces first strictly shrink, and then remain the same.
Comparing (14) with (11), one can see that the two chains are quite similar. In (11), the partitions strictly refine and then remain the same. In (14), the subspaces strictly shrink and then also remain the same. The former means that the pairs of indistinguishable initial states become fewer and fewer, the latter means unobservable initial states become fewer and fewer. The two descending chains are quite similar to each other and they represent the same physical phenomenon if their mathematical meanings are neglected. This similarity can also be regarded as one between discrete mathematics and continuous mathematics.

An application to disturbance decoupling of Boolean control networks
In this subsection, we study the disturbance decoupling problem of BCNs with disturbance. After finding solutions to the problem, one will see the solutions are just of Moore's partition style! This part contains relatively intricate technical details. The reader who is not interested in the technical details can jump over them and directly read the theorems and their explanations. The material of this part was chosen from the lecture notes of the course "Finite-State Dynamical Systems" given by the writer in the winter semester of 2021/22 at Technical University of Berlin, Germany. The coordinate-independent disturbance decoupling problem for BNs with disturbance has been completely solved in [17], and was called the original disturbance decoupling problem therein. The method of finding a solution used in this subsection starts from the definition of the problem and is technically different from the method used in [17]. However, when the input is constant, the solution found in this subsection is consistent with the one found in [17] (by comparing Theorem 10 with [17, Theorem 2]). The technique (based on Proposition 11) in the current paper used for checking the conditions in Theorem 10 is technically easier than the one used in [17].
We now formulate the disturbance decoupling problem for a BCN with disturbance: where The original disturbance decoupling problem means that, whether some state-feedback controller and some coordinate transformation can make the output of (15) not be affected by where G ∈ L 2 m ×2 n . A coordinate transformation is a bijection from Δ 2 n to Δ 2 n , equivalently a mapping z = T x, where z, x ∈ Δ 2 n , T ∈ L 2 n ×2 n is a nonsingular matrix. Since T is also a permutation matrix, the inverse transformation is x = T −1 z = T Tr z. For brevity, denote N := 2 n , M := 2 m , P := 2 p , Q := 2 q .

Problem 1 Consider a BCN (15) with disturbance ξ . Its original disturbance decoupling problem is to check whether there exists a state-feedback controller
and a coordinate transformation where G ∈ L M×N , T ∈ L N ×N , such that after feeding them into (15), the obtained closed-loop BCN satisfies that ξ does not affect y.
We next formulate a special case of the original disturbance decoupling problem which is coordinate-dependent. (15) with disturbance ξ . Its coordinate-dependent disturbance decoupling problem is to check whether there exists a state-feedback controller u(t) = Gx(t) (16) and a coordinate transformation z(t) = T x(t) (17), where G ∈ L M×N , T ∈ L N ×N , such that after feeding them into (15), one obtains the following decomposed BCN (in which ξ does not affect y):

Problem 2 [31] Consider a BCN
Sometimes, we write "a state-feedback controller G ∈ L M×N " for short, and write "a coordinate transformation T ∈ L N ×N " for short, where T is nonsingular.
The procedure of finding solutions to Problem 1 is almost the same as that of finding solutions to Problem 2. The latter is a little more complicated, because the decomposed form (19) is more fragmented. We next show how to find solutions to Problem 2.
To solve Problem 2, we first put a state-feedback controller candidate (16) and a coordinate transformation candidate (17) into (15), and obtain the following:

by Propositions 3 and 4)
where To make Problem (2) solvable, one needs to assume (20) coincides with (19), one then has where (21a) and (21b) hold by Proposition 2 (the right-hand side of (21a) is the structure matrix of the first s Boolean functions of (20f), the right-hand side of (21b) is the structure matrix of the last n − s Boolean functions of (20f)), and because 1 Tr S ⊗L 2 ⊗ 1 Tr P is obtained from (19b) as follows: Moreover, one has where By (21b), (22a), and (23), where S = 2 s for some s ∈ 1, n ,S = N /S,H ∈ L Q×S , Until now, the relationship between the solutions to Problem 2 and Moore's partition is not obvious to see, because in (26), the term "partition" does not explicitly appear.
Because there are finitely many state-feedback controllers and finitely many coordinate transformations, one can verify whether Problem 2 has a solution using Theorem 9.
Next, based on Theorem 9, we reveal the relationship between the solutions to Problem 2 and Moore's partition. The closed-loop BCN in Theorem 9 is where t = 0, 1, 2, . . . . By (26), one sees that for each i ∈ 1, N , all columns of (1 Tr Then one can construct a partition of the state space Δ N as S ] ∼ , then one can construct the following quotient BCN: where for any i ∈ 1, S and k ∈ 1, P . Because the functionality of a coordinate transformation T is renaming the vertices in the state-transition graph of a BCN and the quotient BCN (29) is equivalently obtained from a BCN (15) whose disturbance decoupling problem has a solution, we next give a new necessary and sufficient condition for Problem 2 to have a solution without using T . After putting a state-feedback controller (16) into (15), we obtain the following closed-loop BCN where t = 0, 1, 2, . . . , G i and L i are the same as those in (20),  (30) in the same cell of P ddp , for every k, k ∈ 1, P ,Lxδ k P andLx δ k P are also in the same cell of P ddp .
Proof "if": By (i) and (ii), write the partition P ddp as with T ∈ L N ×N as follows: where i ∈ 1,S , j ∈ 1, S . Then by (ii), for all j, j ∈ 1, S and i ∈ 1,S , That is, i.e., (26a) is satisfied. By (iii), for all i ∈ 1,S , j, j ∈ 1, S , and k, k ∈ 1, P , and are in the same cell of P ddp . Then we can denote the righthand side of (35c) and the right-hand side of (36c) by δ l t N and δ l t N , respectively, where l ∈ 1,S , t, t ∈ 1, S . Then Furthermore, That is, the value of (1 Tr S ⊗ IS)TL(T Tr ⊗ I P )δ j S δ iS δ k P does not depend on j, k. Hence for someL 2 ∈ LS ×S , i.e., (26b) is satisfied. By Theorem 9, Problem 2 has a solution as state-feedback controller G and coordinate transformation (31).
"only if": By Theorem 9, we construct the partition of Δ N which satisfies (i). By (26a), for all i ∈ 1,S and j, j ∈ 1, S , which satisfies (ii). By (26b), δ k P are in the same cell of P ddp as in (37), which satisfies (iii).

Remark 1
The main result obtained in [17] is as follows: for a BCN (15) with constant input, i.e., with m = 0, Problem 2 has a solution if and only if the state-transition graph of the reduced (15) has an equal, concolorous, perfect partition of its vertex set, where "equal", "concolorous", and "perfect" coincide with (i), (ii), and (iii) of Theorem 10, respectively. The method of checking the existence of such a partition designed in [17] is technically more complex than the method designed later based on Proposition 11. The reader who is interested in this topic could try to compare their efficiency.
By Theorem 10, to check whether Problem 2 has a solution, one needs to check whether there is a state-feedback controller G ∈ L M×N such that for the closed-loop BCN (30) obtained by feeding G into BCN (15), there is a partition P ddp of state space Δ N satisfying (i), (ii), and (iii). One can see that P ddp is very similar to and finer than 4 Moore's partition P(∼ ind ) as in (11) if one regards ξ as input, hence P ddp can be regarded as a Moore-style partition. By comparing P ddp with P(∼ ind ), one sees only minor differences between them: (a) in P(∼ ind ), the cardinalities of all cells need not be the same, (b) in P(∼ ind ), each input must drive the states of one cell into one cell, but different inputs may drive the states of one cell into different cells, (c) P(∼ ind ) is unique with respect to BCN (5), but P ddp may not be unique with respect to (30).
We define a relation ∼ ddp ∈ Δ N × Δ N with respect to the closed-loop BCN (30) as follows: two states x, x have relation ∼ ddp if and only if for every two disturbance sequences ξ 1 . . . ξ r and ξ 1 . . . ξ r , where r ∈ N (r = 0 means ξ 1 . . . ξ r = ξ 1 . . . ξ r = ), the output sequences generated by x and ξ 1 . . . ξ r and generated by x and ξ 1 . . . ξ r are the same. Relation ∼ ddp is symmetric and transitive, but not necessarily reflexive. As introduced in Sect. 2.2, ∼ ddp can induce a partition P(∼ ddp ) of Δ N . By definition, the equivalence relation that induces a given partition P ddp satisfying (ii) and (iii) is finer than ∼ ddp . Furthermore, we have the following result.

Proposition 11 A closed-loop BCN (30) admits a partition P ddp satisfying (ii) and (iii) of Theorem 10 if and only if the relation ∼ ddp corresponding to (30) is reflexive.
Proof "if": Assume the relation ∼ ddp corresponding to (30) is reflexive. That is, now ∼ ddp is an equivalence relation. For every two states x, x having relation ∼ ddp , choosing the empty input sequence , by definition we have H x = H x . That is, ∼ ddp satisfies (ii). Let x, x be two states having relation ∼ ddp and let ξξ 1 . . . ξ r , ξ ξ 1 . . . ξ r be two disturbance sequences, by definition the output sequence generated by x and ξξ 1 . . . ξ r is the same as the output sequence generated by x and ξ ξ 1 . . . ξ r . Then the output sequence generated bŷ Lxξ and ξ 1 . . . ξ r is the same as the output sequence generated byLx ξ and ξ 1 . . . ξ r . Due to the arbitrariness of ξ 1 . . . ξ r and ξ 1 . . . ξ r , we haveLxξ andLx ξ have relation ∼ ddp . That is, ∼ ddp satisfies (iii). Note that a nonreflexive relation ∼ ddp may not satisfy both (ii) and (iii), because the above x and x may not necessarily be different. "only if": Assume (30) admits a partition P ddp satisfying (ii) and (iii). By definition, every two states x, x in the same cell of P ddp satisfy that for all disturbance sequences ξ 1 . . . ξ r and ξ 1 . . . ξ r , the output sequences generated by x and ξ 1 . . . ξ r and generated by x and ξ 1 . . . ξ r are the same. Hence the equivalence relation inducing P ddp is finer than ∼ ddp . Compressing P ddp by repeating the following action: merging two cells C 1 , C 2 such that the states of C 1 and C 2 produce the same output and transition into the same cell, until no two such cells exist. Then the modified P ddp is the same as ∼ ddp .

Corollary 1
If the relation ∼ ddp corresponding to a closedloop BCN (30) is reflexive, then it is the coarsest among the partitions P ddp satisfying (ii) and (iii) of Theorem 10.
By Proposition 11, the existence of partition P ddp satisfying (ii) and (iii) of Theorem 10 can be checked by computing relation ∼ ddp as follows (since ∼ ddp is symmetric and transitive, it induces a partition P(∼ ddp ) of Δ N as introduced in Sect. 2.2, the following procedure works). Now we know how to check whether a closed-loop BCN (30) admits a partition satisfying (ii) and (iii) of Theorem 10. If the obtained P(∼ ddp ) has all cells with the same cardinality, then P(∼ ddp ) also satisfies (i); otherwise, one needs to check if there is a refinement of P(∼ ddp ) that satisfies (i), (ii), and (iii) simultaneously.
Remark 2 Problem 1 is coordinate-independent, so when looking for a partition P ddp as in Theorem 10, (i) plays no role. We obtain a reformulation of Theorem 10: Problem 1 has a solution if and only if there exists a state-feedback controller G ∈ L M×N and a partition P ddp of the state space Δ N such that (ii) and (iii) of Theorem 10 are satisfied. The existence of such P ddp can be checked by computing P(∼ ddp ) and then check whether ∼ ddp is reflexive by Proposition 11.

An application to observability decomposition of Boolean control networks
In this subsection, we introduce the observability decomposition problem of BCNs with respect to multiple-experiment observability (Definition 12). As mentioned in Theorem 6 (Sect. 3.1), Moore's partition directly provides an algorithm for verifying multiple-experiment observability of BCNs. In this subsection, we will show that Moore's partition also directly provides an algorithm for checking the existence of the observability decomposition of BCNs with respect to this type of observability, as well as computing the decomposition if one exists. When characterizing the observability decomposition, one will see that its form is quite similar to the disturbance decoupling problem studied in Sect. 3.5. That is to say, Moore's partition closely relate the observability decomposition problem and the disturbance decoupling problem in BCNs. Before formally introducing the observability decomposition problem, we show a motivating example.
Using argument similar to that used in Sect. 3.5, one can see that the existence of a coordinate transformation as in Problem 3 is equivalent to the condition that the cells of the corresponding Moore's partition have the same cardinality equal to 2 s for some s ≥ 1. (11) have the same cardinality equal to 2 s for some s ∈ 1, n .

Remark 3 In [3]
, it was shown that for a linear time-invariant control system (13), its quotient system over the quotient space R n / n−1 i=0 O i is observable, where recall that n−1 i=0 O i is the unobservable subspace of (13). Comparing this quotient system with the pair ((50b), (50c)), one can regard the latter as a counterpart of the former in a BCN (10). In addition, with respect to the disturbance decoupling problem, the quotient system of a BCN (15) with disturbance was derived in (29). An essential difference lies in that, the quotient system of (13) with respect to observability always exists, but the above quotient BCN ((50b), (50c)) of BCN (10) with respect to multiple-experiment observability and the quotient BCN (29) of BCN (15) with disturbance do not, which also reveals the essence of nonlinearity of BCNs.
Next we restate a related result obtained in [44] which is consistent with Theorem 12, but s was given and the results were represented by sub-BCNs B j which are obtained from BCN (10) by setting u(t) ≡ δ j M , j ∈ 1, M . By comparing these two results, one can see that it is easier to use Theorem 12 to check the existence of an observability decomposition for a BCN (10).
One can equivalently represent (50) as the following based on an appropriate coordinate transformation T ∈ L N ×N (if such a T exists): where j ∈ 1, M . Then the following result was obtained in [44].

Theorem 13 Let s be in 1, n ands = n−s. Then Problem 3 has a solution if and only if there is a partition P obsd of Δ N such that
(i) all the cells of P obsd have the same cardinality S = 2 s , (ii) for every two states x, x in the same cell of P obsd , it holds that H x = H x , (iii) for all j ∈ 1, M , for every two states x, x in the same cell of P obsd , L j x and L j x are also in the same cell of P obsd , (iv) P obsd is the coarsest among the partitions of Δ N that satisfy (ii) and (iii).
The conditions (i), (ii), and (iii) together imply (51). Then the condition (iv) implies that the pair ((50b), (50c)) is multiple-experiment observable. The latter was guaranteed by a proposition proven in [44]: Let P be a partition of Δ N that satisfies (ii) and (iii), then P is finer than the corresponding Moore's partition P(∼ ind ). If (iv) holds then corresponding to the quotient BCN obtained by identifying all states in the same cell of P obsd , P(∼ ind ) is discrete. A short conclusion In this section, we introduced several control problems of BCNs, including computation of the smallest invariant dual subspace of a BN containing a given set of Boolean functions, the multiple-experiment observability verification (decomposition) problem, and the disturbance decoupling problem. It is surprising that Moore's partition plays a central role in all these fundamental problems and hence closely relates them. However, Moore's partition has its own drawbacks, e.g., when using Moore's partition to verify arbitrary-experiment observability (Definition 15) of BCNs [43], the procedure and the obtained verification algorithm are very complex; Moore's partition does not apply to the other notions of observability of BCNs; in addition, Moore's partition cannot be extended to nondeterministic finite-transition systems in [20] and [13,Remark 4.1]. In the next section, we introduce a different method which applies to the four definitions of observability of BCNs as introduced in Sect. 2.7. Since this new method was found by the writer when trying to verify the negations of these definitions of observability of BCNs [21,37], it was called an inverse method in the current paper.

Introductory subsection
During the early stages of the study of observability problems of BCNs (between 2008 and 2016), the model-based testing area (including [10,39], [12] and references therein) in computer science should not be known to the researchers in the control area of BCNs. The earliest works on the verification of different types of observability of BCNs are under the controllability assumption (see a necessary and sufficient condition for strong multiple-experiment observability [33] and a necessary and sufficient condition for single-experiment observability [42], both based on the controllability assumption). Although an STP representation for Moore's partition was rediscovered in [33,42], Moore's partition does not apply to these two types of observability (later we will show the reason), hence necessary and sufficient conditions for these types of observability were not obtained in [33,42]. On the other hand, although Moore's partition applies to multiple-experiment observability (shown in Theorem 6), it was not used to verify this type of observability in [41]. Only a sufficient but not necessary condition was given for this type of observability in [41]. The first necessary and sufficient condition for observability of BCNs was given in [43] based on Moore's partition, where only arbitrary-experiment observability was studied. Although a necessary and sufficient condition for arbitrary-experiment observability was given in [43], the procedure of deriving the condition is rather complex, so is the condition itself: a BCN (10) is arbitrary-experiment observable if and only if for every pair of different periodic (state, input)-trajectories of the same minimal period k and the same input trajectory, the corresponding output trajectories are also different and periodic of minimal period k. The first necessary and sufficient condition for single-experiment observability of BCNs was given in [45]. The method used is to give a length upper bound on input sequences such that if no input sequence of that length is a distinguishing input sequence of (10) then no distinguishing input sequence exists. Hence until 2013, no necessary and sufficient condition for multiple-experiment observability or strong multipleexperiment observability of BCNs had been found in the control area of BCNs. The methods used in [43,45] cannot be used to deal with the other types of observability.
From 2012 to 2013, the writer was visiting Department of Mathematics and Statistics of University of Turku, Finland as a joint Ph.D. candidate. During that period, he was mainly studying cellular automata (discrete-time dynamical systems over the Cantor space, a natural generalization of deterministic Turning machines), and his main research interests gradually moved from control theory to theoretical computer science. Upon leaving, he discussed the decidability problem of one type of observability of BCNs with his colleges therein, including Jarkko Kari, Charalampos Zinoviadis, Valle Salo, and Illka Törmä. The idea of using finite automata and formal languages to verify observability of BCNs was born during the discussion but did not take shape during that period.
The idea took shape after the writer left Finland and just arrived in Singapore in around October, 2013. This time he considered the negation of observability of BCNs, and proposed the concept of weighted pair graph [21,37] (see Definition 16, later renamed observability graph in [13,23]), which aggregates every pair of trajectories of a BCN that produce the same output sequence. Note that this kind of thinking is essentially different from the ideas used in the previous papers on observability of BCNs, including [33,[41][42][43]45], because in these papers, the authors all considered directly verifying different definitions of observability themselves instead of their negations. After considering verifying the negations of various definitions of observability, the writer proposed a unified method to verify the four definitions of observability shown in Sect. 2.7 (see [21,37]). The overall idea is as follows: first, given a BCN (5), compute its weighted pair graph (i.e., observability graph); second, com-

BCN
Observability graph DFA Verify observability Fig. 4 Sketch of using the notions of observability graph and deterministic finite automaton to verify observability of BCNs [21,37] pute deterministic finite automata (DFAs) from the graph adapted to the four definitions of observability; third, use each DFA to verify the corresponding type of observability. See Fig. 4 for a sketch. Next, we show the implementation of the sketch.

Observability verification
Now we introduce the notion of weighted pair graph (i.e., observability graph). , v ))|, i.e., the number of inputs appearing in the edges starting from v. The diagonal subgraph of an observability graph is defined by all diagonal vertices and all edges between them, and denoted by . Similarly, the non-diagonal subgraph is defined by all nondiagonal vertices and all edges between them.

Definition 16 [21,37] Consider a BCN (5). A triple
For a BCN (5), the size of its observability graph G o is O(2 2n+m−1 ) = O (2 2n+m ). Because at each diagonal vertex of G o , there exist exactly 2 m outgoing edges and all these edges go to diagonal vertices, we can consider the diagonal subgraph as a single vertex. Therefore, the number of vertices can be equivalently bounded by 1+2 n (2 n −1)/2 from above. When drawing G o , the diagonal subgraph will be drawn as a single vertex . called complete in the current paper. To conveniently verify observability of BCNs, incomplete DFAs are also considered, in which δ is not defined at at least one pair in Q × Σ. The words w accepted by an incomplete DFA A are also defined as those satisfying δ(q 0 , w) ∈ F, and the language recognized by an incomplete DFA A is also defined as the set of all words accepted by A. The following direct proposition will be useful in verifying observability of BCNs.

Proposition 14 [21] Consider a (complete or incomplete) DFA A = (Q, Σ, δ, q 0 , Q) (all states are final). Assume each state is reachable from q 0 . Then A recognizes language Σ * if and only if A is complete.
Since we verify the negations of different definitions of observability of BCNs, we characterize their negations in the following four propositions. These propositions can be directly obtained by definition.

Proposition 15 A BCN (5) is not multiple-experiment observable (Definition 12) if and only if there exist two different
initial states x 0 , x 0 ∈ D n such that for each p ∈ N and each input sequence U ∈ (D m ) p , the corresponding output sequences are the same.

Proposition 16 A BCN (5) is not strongly multiple-experiment observable (Definition 13) if and only if there exists an ini-
tial state x 0 ∈ D n such that for each p ∈ N and each input sequence U ∈ (D m ) p , there exists an initial state x 0 ∈ D n different from x 0 such that the corresponding output sequences are the same.

Proposition 17 A BCN (5) is not single-experiment observable (Definition14) if and only if for each p ∈ N and each
input sequence U ∈ (D m ) p , there exist two different initial states x 0 , x 0 ∈ D n such that the corresponding output sequences are the same.

Proposition 18 A BCN (5) is not arbitrary-experiment observable (Definition 15) if and only if there exist two different
initial states x 0 , x 0 ∈ D n and an input sequence U ∈ (D m ) p with p sufficiently large such that the corresponding output sequences are the same.
One sees both input sequence 0 and input sequence 1 distinguish initial states 01 and 00, both input sequence 0 and input sequence 1 distinguish initial states 00 and 10, input sequence 01 distinguishes initial states 01 and 10. For example, an illustration of using input sequence 01 to distinguish states 01 and 10 is shown in Table 1.

Remark 4
When using Theorem 19 to verify multipleexperiment observability of a BCN (5), one needs to compute DFA G v oD for at most 2 n (2 n − 1)/2 times. Hence the overall time complexity of verifying this type of observability is O(2 4n+m−2 ) = O(2 4n+m ).

Verifying Definition 13
Verifying strong multiple-experiment observability (Definition 13) is more involved than verifying multiple-experiment observability. By Proposition 16, to verify Definition 13 for a BCN (5), one needs a powerset construction from the observability graph G o of (5). From now on we call a DFA with at least one initial state a deterministic finite-state automaton. Such an automaton is denoted by (Q, Σ, δ, Q 0 , F), where Q 0 ⊂ Q denotes the set of initial states.
For every initial state x 0 ∈ D n of (5), we consider the observability graph G o = (V, E, W) of (5) as a deterministic finite-state automaton where , v )).
Over G x 0 o , we construct a (complete or incomplete) DFA Recall that Acc(G x 0 oD ) is the accessible part of G x 0 oD . Then all words that are not accepted by Acc(G x 0 oD ) are distinguishing input sequences of x 0 . Theorem 20 [21,37] A BCN (5) is not strongly multipleexperiment observable if and only if there exists an initial state x 0 ∈ D n such that the corresponding DFA Acc(G x 0 oD ) (54) recognizes language (D m ) * , i.e., Acc(G x 0 oD ) is complete by Proposition 14.
From DFA Acc(G 10 oD ) (resp., Acc(G 01 oD )), one sees that for each positive-length input sequence U , 0U distinguishes initial state 10 (resp., 01) from any other initial state, but 1U cannot do that.
Taking U = 1 for example, one has results shown in Table 2.
From DFA Acc(G 00 oD ), one sees that each positive-length input sequence distinguishes initial state 00 from any other initial state.   Hence BCN (12) is strongly multiple-experiment observable by Theorem 20.

Remark 5
When using Theorem 20 to verify strong multipleexperiment observability of a BCN (5), one needs to compute DFA G x 0 oD for at most 2 n times (for all x 0 ∈ D n ). Hence the overall time complexity of verifying this type of observability is O(2 2 2n−1 −2 n−1 +1+m+n ) = O(2 2 2n−1 +m ).

Verifying Definition 14
To verify single-experiment observability (Definition 14) of a BCN (5), by Proposition 17, one also needs a powerset construction from the observability graph G o of (5), but this construction is different from G x 0 oD (54) used for verifying strong multiple-experiment observability.
We also consider G o = (V, E, W) as a deterministic finitestate automaton where V nd = {{x, x }|h(x) = h(x ), x = x } ⊂ V is the set of initial states, i.e., the initial states are exactly the non- ((v, v )). Over G V nd o , we construct a DFA where 2 V \ {∅} is the state set, V nd is the initial state, δ is defined in the same way as the one in G x 0 oD (54). The size of G V nd oD is O(2 2 2n−1 −2 n−1 +1+m ) = O(2 2 2n−1 +m ). All words that are not accepted by Acc(G V nd oD ) are distinguishing input sequences of (5).

Theorem 21 [21,37] A BCN (5) is not single-experiment observable if and only if the corresponding DFA
Example 10 Recall BCN (12) in Example 4. The corresponding DFA Acc(G V nd oD ) is shown in Fig. 8. This DFA is not complete, then by Theorem 21, (12) is single-experiment observable, and for every positive-length input sequence U , input sequence 0U distinguishes every pair of different initial states.
Taking U = 1 for example, one also has results shown in Table 2.

Verifying Definition 15
Different from verifying the previous three definitions of observability, one can directly use the observability graph G o of a BCN (5) to verify its arbitrary-experiment observability (Definition 15). By Proposition 18, the following result holds.

Theorem 22 [37] A BCN (5) is not arbitrary-experiment observable if and only if in its observability graph G o , there exist a non-diagonal vertex v, a cycle C, and a path from v to C.
Due to the finiteness of the number of states of (5) and the pigeonhole principle, a sufficiently long input sequence that is not a distinguishing input sequence of two different initial states x 0 , x 0 leads vertex {x 0 , x 0 } to a cycle in G o . The converse is also true. Hence Theorem 22 holds.

Example 11
Reconsider BCN (12) in Example 4. In its observability graph (Fig. 5), there is a path from non-diagonal vertex {01, 10} to its diagonal subgraph, then there is a path from {01, 10} to some cycle in the diagonal subgraph, (12) is not arbitrary-experiment observable by Theorem 22.

Remark 7
The time complexity of using Theorem 22 to verify arbitrary-experiment observability of a BCN (5) is

Observer design based on arbitrary-experiment observability
Consider a BCN (5) that is arbitrary-experiment observable. One can design an observer as a partial function obs : (5), there are at most 2 2n−1 − 2 n−1 non-diagonal vertices, and all input sequences of length 2 2n−1 −2 n−1 are distinguishing input sequences of every pair of different initial states. The construction of obs is as follows. Choose an initial state x 0 ∈ D n and an input sequence U ∈ (D m ) 2 2n−1 −2 n−1 , and feed them into (5) to obtain the output sequence Y ∈ (D q ) 2 2n−1 −2 n−1 +1 , then obs(U , Y ) = x 0 , because when the input sequence is U and the corresponding output sequence is Y , the initial state can only be x 0 .

Reconstructibility verification
Reconstructibility (also called detectability) is a property on whether one can use an input sequence and the corresponding output sequence to determine the current (i.e., final) state. This property is weaker than observability. When the initial state is not crucial but one wants to do quantitative analysis for a BCN (5) after some time, the property of reconstructibility applies. We characterize the counterparts of single-experiment observability (Definition 14) and arbitrary-experiment observability (Definition 15) of a BCN (5), for reconstructibility, which are called single-experiment reconstructibility and arbitrary-experiment reconstructibility. A homing input sequence u 1 . . . u n of (5), where n ∈ N, is such that for every two initial states x 0 , x 0 , if the corresponding current states x n and x n are different, then the output sequences generated by x 0 and u 1 . . . u n and generated by x 0 and u 1 . . . u n are different. Hence a homing input sequence can determine the current state.
Definition 17 [12,46] A BCN (5) is called single-experiment reconstructible if it has a homing input sequence. Definition 18 [43] A BCN (5) is called arbitrary-experiment reconstructible if all input sequences of length greater than some number in N are homing.
Definition 18 is strictly stronger than Definition 17 [46]. To verify reconstructibility, we define a notion of reconstructibility graph.
Definition 19 [46] Consider a BCN (5) and its observability graph G o (Definition 16). Its reconstructibility graph G r is defined by the non-diagonal subgraph of G o .

Example 12
Reconsider the BCN (12) in Example 4. Its reconstructibility graph is shown in Fig. 9. This figure is the non-diagonal subgraph of Fig. 5 (the observability graph of (12)).

Verifying Definition 18
We first verify Definition 18 and then Definition 17, because it is much easier to verify Definition 18. We directly use the reconstructibility graph of a BCN (5) to verify its arbitraryexperiment reconstructibility. The following result is rather intuitive.

Theorem 23 [46] Consider a BCN (5) and its reconstructibility graph G r . The BCN is not arbitrary-experiment reconstructible if and only if G r contains a cycle.
Remark 8 Compared with the necessary and sufficient condition for arbitrary-experiment reconstructibility of (5) given in [43], one can see that the expression of our condition in Theorem 23 is much simpler and much easier to check. Recall that the condition given in [43] was derived from Moore's partition (defined in (11)).

Example 13
Reconsider the BCN (12) in Example 4. Since in its reconstructibility graph (Fig. 9) there is no cycle, then by Theorem 23, (12) is arbitrary-experiment reconstructible. By Fig. 9 one can see that all input sequences starting from 0 of length at least 2 and all input sequences starting from 1 are homing. The procedure of using homing input sequence 01 to determine the current state is shown in Table 2.

Remark 9
The time complexity of using Theorem 23 to verify arbitrary-experiment reconstructibility of a BCN (5) is

Verifying Definition 17
Similar to verifying single-experiment observability (Sect. 4.2.3), to verify single-experiment reconstructibility (Definition 17) of a BCN (5), one also needs a powerset construction from its reconstructibility graph G r .
We also consider G r = (V, E, W) as a deterministic finitestate automaton where all states are initial, δ is a partial function v )).
Over G r , we construct a DFA One can see that the words that are not accepted by DFA Acc(G r D ) are homing input sequences of (5). Then the following result holds. (5)  Further discussion on verification of single-experiment reconstructibility It has been shown that the time complexity of using Theorem 24 to verify single-experiment reconstructibility of a BCN (5) is doubly exponential. Is it possible to reduce the complexity? The size of DFA G r D (58) is exponential in the size of the reconstructibility graph G r of (5). One possible way of reducing time complexity is to see whether one can directly use G r to verify this type of reconstructibility. Luckily, this kind of reduction works, because the condition in Theorem 24 can be equivalently represented by some kinds of "limit behavior" of G r . In detail, if one regards G r as a dynamical system, then the DFA G r D is complete implies after a sufficiently long time, all vertices of G r that are reachable have outdegree 2 m . This holds because G r is deterministic and all vertices are initial states of G r as a dynamical system. The converse implication also holds. The limit set of G r (i.e., the set of all vertices that can be reached at any time) generates a complete subgraph (all vertices have outdegree 2 m ). The conclusion is formulated as follows:

Theorem 24 [46] A BCN
Proposition 25 [46] Consider a BCN (5), its reconstructibility graph G r = (V, E, W), and the DFA G r D defined in (58). DFA G r D is complete if and only if G r has a complete subgraph.
By Theorem 24 and Proposition 25, the following result holds.

Theorem 26 [46] A BCN (5) is not single-experiment reconstructible if and only if its reconstructibility graph G r has a complete subgraph.
To check the condition in Theorem 26, one can remove the set V <2 m of all vertices of G r with outdegree less than 2 m and all vertices that will reach V <2 m . Then the rest of G r is a complete subgraph if it is nonempty. Hence this check can be done in time linear in the size of G r .

Example 15
Consider BCN where t ∈ N, x(t) ∈ D 2 , u(t), y(t) ∈ D, f and h are as follows: (a) f . The reconstructibility graph of (59) is shown in Fig. 11. In Fig. 11, there is a complete subgraph (the self-loop on vertex {01, 10}), then by Theorem 26, (59) is not singleexperiment observable. The corresponding DFA Acc(G r D ) (defined in (58)) is shown in Fig. 12. DFA Acc(G r D ) is complete, which matches Proposition 25.

Remark 11
The time complexity of using Theorem 26 to verify single-experiment reconstructibility of a BCN (5) is O(2 2n+m ). When applying the classical verification algorithm (based on a notion of current-state uncertainty, which is similar to the initial-state uncertainty on pp. 10) for the existence of homing sequences of Mealy machines [12, Chapter 1] to verify single-experiment reconstructibility of (5), it runs in time O(2 3n + 2 2n+m ), and homing input sequences can also be computed in this time.

Further applications of the observability graph to observability of Boolean control networks
As shown in Sect. 4.3.2, we first used DFA Acc(G r D ) (58) to verify single-experiment reconstructibility of a BCN (5) (Theorem 24), and then directly used the reconstructibility graph G r of (5) to verify this type of reconstructibility (Theorem 26). This modification leads to tremendous decrease of time complexity (from doubly exponential to exponential!). Then one may consider whether such tremendous decrease of time complexity could be obtained in verifying observability. For arbitrary-experiment of observability, we had already directly used the observability graph G o of (5) to do verification (Theorem 22 Proposition 27 [22,23] Consider a BCN (5). Two different initial states x 0 , x 0 ∈ D n with h(x 0 ) = h(x 0 ) has as one of their distinguishing input sequences. Two different initial states x 0 , x 0 ∈ D n with h(x 0 ) = h(x 0 ) has a distinguishing input sequence if and only if in the observability graph G o of (5), either outdeg({x 0 , By Proposition 27, a necessary and sufficient condition for multiple-experiment observability of (5) can be obtained.
Theorem 28 [22,23] A BCN (5) is multiple-experiment observable if and only if in its observability graph G o , for every non-diagonal vertex v with outdegree equal to 2 m , there is a path from v to some non-diagonal vertex with outdegree less than 2 m .  (5) is not multiple-experiment observable.

Remark 12
The time complexity of using Theorem 28 to verify multiple-experiment observability of a BCN (5) is O(2 2n+m ), which is lower than that of using Theorem 19 with complexity O(2 4n+m ) (see Remark 4).

Example 17
Consider BCN where t ∈ N, x(t) ∈ D 2 , u(t), y(t) ∈ D, f and h are as follows: (a) f .
To check the condition in Theorem 28, we compute as V <2 = {{00, 10}, {01, 00}}. The rest of G o contains a vertex {01, 10} whose outdegree is equal to 2. Therefore, (60) is not multiple-experiment observable by Theorem 28. By  Fig. 13, one easily sees that initial states 01 and 10 have no distinguishing input sequence.

Extensions of the observability graph to nondeterministic finite-transition systems
After writing the papers [21,37,46] and establishing the basic framework of using the notion of observability graph to verify the four kinds of observability of BCNs, the writer tried to extend the notion to nondeterministic finite-transition systems (NFTSs). Recall that Moore machines and Mealy machines are deterministic finite-transition systems. Luckily, the notion can be naturally extended, and hence the previously studied different notions of observability have been verified in NFTSs using the extended notion of observability graph [20]. However, in NFTSs, Moore's equivalence relation (see (11)) is not necessarily transitive [20], hence Moore's partition-based method does not apply to NFTSs. An NFTS is a sextuple where • X is a finite set of states, • X 0 ⊂ X a set of initial states, • U an (finite) alphabet of inputs, • Y an alphabet of outputs, and • h : X → Y the output map.
In [20], the NFTSs considered are a bit more general in the sense that f :

Different definitions of observability of nondeterministic finite-transition systems
Definition 20 [20] An NFTS S = (X , X 0 , U , f , Y , h) is called multiple-experiment observable if for every two different initial states x 0 , x 0 ∈ X 0 , there is an input sequence α ∈ U * such that every output sequence β of length |α| + 1 generated by x 0 and α is different from every output sequence β of length |α|+1 generated by x 0 and α. Such an α is called a distinguishing input sequence of x 0 and x 0 .
Definition 21 [20] An NFTS S = (X , is called single-experiment observable if there is an input sequence α ∈ U * such that for every two different initial states x 0 , x 0 ∈ X 0 , every output sequence β of length |α| + 1 generated by x 0 and α, and every output sequence β of length |α| + 1 generated by x 0 and α, it holds that β = β . Such an α is called a distinguishing input sequence of S. Definition 22 [20] An NFTS S = (X , X 0 , U , f , Y , h) is called arbitrary-experiment observable if for every two different initial states x 0 , x 0 ∈ X 0 , for every input sequence α ∈ U ω , every output sequence β ∈ Y ω generated by x 0 and α is different from every output sequence β ∈ Y ω generated by x 0 and α.

Observability graphs of nondeterministic finite-transition systems
In [20], the observability graph G o of a BCN (5) (Definition 16) was extended to an NFTS S as follows (in the form of its adjacency matrix (5) of [20]).
and With the notion of observability graph of an NFTS S, the procedures shown in Sect. 4 (52) and (56) were extended to S), and verification algorithms for the three types of observability were derived for NFTSs. The same as arbitrary-experiment observability of a BCN (5), this type of observability was also verified directly using the observability graph G o of S.

Verifying Definition 20
To verify multiple-experiment observability of S, we consider a non-diagonal vertex v := {x, x } ∈ V of its observability G o , where x, x ∈ X 0 , and regard G o as a nondeterministic finite automaton (NFA) where V is the state set, v is the initial state, δ : v ))}. Over G v o we construct the following DFA (the counterpart of DFA G v oD (52) of BCN (5)) where 2 V \ {∅} is the state set, {v} is the initial state, δ is defined in the same way as the one in G x 0 oD (54). to use Theorem 29 to verify multiple-experiment observability of an NFTS S.

Verifying Definition 21
To verify single-experiment observability of an NFTS S, we consider the observability G o of S as a nondeterministic finite-state automaton where is the same as the one defined in (61). Furthermore, we compute the following DFA from (63):

Verifying Definition 22
As mentioned above, one can directly use the observability graph G o of an NFTS S to verify its arbitrary-experiment observability.
Theorem 31 [20] An NFTS S is not arbitrary-experiment observable if and only if in its observability graph G o , there exists a non-diagonal vertex v, a cycle C, and a path from v to C.

Remark 15
The time complexity of using Theorem 31 to verify arbitrary-experiment observability of an NFTS S is O(|X | 4 |U |).

Observer design based on arbitrary-experiment observability
Similar to the observer design for a BCN (5) base on arbitrary-experiment observability as shown in Sect. 4.2.5, next we design an observer for an arbitrary-experiment observable NFTS S. Assume an NFTS S that is arbitraryexperiment observable and assume X 0 = X , one can design an observer as a partial function obs : U |X |(|X |−1)/2 × Y |X |(|X |−1)/2+1 → X 0 , because in the observability graph G o of S, there are at most |X |(|X | − 1)/2 non-diagonal vertices, and all input sequences of length |X |(|X | − 1)/2 are distinguishing input sequences of every pair of different initial states. The construction of obs is as follows. Choose an initial state x 0 ∈ X 0 and an input sequence α ∈ U |X |(|X |−1)/2 , and feed them into S and choose an arbitrary generated output sequence β ∈ Y |X |(|X |−1)/2+1 , then obs(α, β) = x 0 , because when the input sequence is α and the corresponding output sequence is β, the initial state can only be x 0 .
The writer wants to point out that when writing papers [21,37,46], he had not known the research area of model-based testing, including [10,12,39]. However, when writing paper [20], he found this theoretical computer science area, and was aware of that several notions of observability have been studied in this area, but the notions were not called observability. In [10,39], Moore machines and Mealy machines were studied, multiple-experiment observability was characterized in [10] (already introduced in Sect. 3.1), single-experiment observability and strong multiple-experiment observability were studied in [39] and their verification problems were proven to be PSPACE-complete in Mealy machines. However, none of the verification methods found [10,39] is equivalent to the observability-graph method (already shown on Page 9 in Sect. 2.7). Later on, when the writer started to look for observability results in NFTSs in the literature, he found paper [47], in which a structure equivalent to DFA Acc(G v oD ) (62) was directly used to verify multipleexperiment observability of NFTSs. Note that in [47], the counterpart of Acc(G v oD ) was only used to verify multipleexperiment observability. Note also that it is far from trivial to recover the observability graph G o of S from the counterpart, because the size of the counterpart is exponential in the size of G o .

Applications of the observability graph to observability of singular Boolean control networks and probabilistic Boolean networks
After the papers [21,37,46] were published, the basic idea of observability graph has been widely used in the control area of BCNs, not only in [22,23]. For example, in [25,26], the problem of set controllability/reachability of BCNs were verified; moreover, after computing the product (called parallel extension in [26], which is actually the adjacency matrix of the observability graph proposed in [21,37] when being applied to verify observability) of two copies of a BCN (5), multiple-experiment observability of (5) was verified by verifying set controllability for the product. Note that the technical details of verifying multiple-experiment observability of (5) by verifying set controllability of the parallel extension of (5) are exactly the same as those in [21,37], but in an equivalent matrix representation (that only looks different from the graph and automaton representation used in [21,37]). In [26], two other types of observability of (5) were also verified over the parallel extension of (5) (see Table 3).
The observability graph has also been used to study observability perturbation for multivalued logical networks (a deterministic extension of BNs) in [30]. Note that [ A variant of the reconstructibility graph (Definition 19) was used to study reconstructibility of singular BCNs in [29], where such networks are a subclass of NFTSs studied in [20]. A summary can be found in Table 3.
Observability results have also been extended to probabilistic Boolean networks (PBNs) [27,28,49,50], where in [27,28], the observability graph was used. In the PBNs studied in these papers, the stochastic switching signals are independent and identically distributed processes, hence the PBNs are actually discrete-time finite-state time-homogeneous Markov chains. Moreover, if all probabilities in such a PBN are removed then it becomes an NFTS in which the input is constant. That is, the systems considered in [20] are more general than the systems considered in [27,28,49,50].
The PBNs studied in [27,28,49,50] are as follows: where σ : N → 1, s with s ∈ Z + is an independent and identically distributed process and at each time t ∈ N, [ p 1 , . . . , p s ] =: p is the probability distribution of σ ; L 1 , . . . , L s ∈ L 2 n ×2 n are the system matrices; H ∈ L 2 q ×2 n is the output matrix; x(t) ∈ Δ 2 n ; y(t) ∈ Δ 2 q . 6 In [27,28,49,50], the authors considered p i ≥ 0. Here we consider p i > 0 with loss of generality because all structure matrices L i with p i = 0 could be removed equivalently.
Definition 24 [27,50] Let θ be in N. A PBN (65) is called observable with probability one on 0, θ if for every two different initial states x 0 , x 0 ∈ Δ 2 n , It is trivial to see that a PBN (65) is observable with probability one on 0, θ if and only if the corresponding BCN (66) is arbitrary-experiment observable (Definition 15) and all input sequences of length θ are distinguishing input sequences of (66). Hence, observability with probability one on 0, θ can be regarded as a slightly stronger version of arbitrary-experiment observability.
By definition, a PBN (65) is finite-time observable in probability if and only if the corresponding BCN (66) is multiple-experiment observable (Definition 12). Furthermore, (65) is observable in probability on 0, θ if and only if the corresponding (66) is multiple-experiment observable and all pairs of different initial states x 0 , x 0 have a distinguishing input sequence of fixed length θ . Hence observability in probability on 0, θ and finite-time observability in probability can be regarded as slight variants of multiple-experiment observability. [27,28] A PBN (65) is called asymptotically observable in distribution if for every two different initial Table 3 Complexity upper bounds for verifying four definitions of observability in BCNs, where the algorithms corresponding to the bounds with the same type of underlines are derived from mathematically equivalent methods, n and m denote the numbers of state nodes and input nodes. The method in [48] is function-based, all the other methods are state-based. So, the method in [48]  states x 0 , x 0 ∈ Δ 2 n , it holds that lim θ→∞ Prob{ y(θ ; σ, x 0 ) = y(θ ; σ, x 0 )} = 1.
Conversely, assume that (i) (66) satisfies Definition 12 and (ii) in the observability graph G o := (V, E, W) of (66), there is no path from any non-diagonal vertex to any diagonal vertex. We endow the edges of G o with probabilities according to the probability distribution p = [p 1 , . . . , p s ] as in (65): v)) p i . Denote the set of diagonal vertices and the set of non-diagonal vertices of G o by V d and V nd , respectively. Then by (ii), for all v ∈ V d and v ∈ V nd , p v,v = 0. Denote the adjacency matrix of the non-diagonal subgraph of G o by M V nd = ( p v,v ) v,v ∈V nd . Then also by (ii), the sum of the v-th column of (M V nd ) θ is equal to Prob{ y(θ ; σ, x 0 ) = y(θ ; σ, x 0 )}, where {x 0 , x 0 } = v. By (i) (i.e., for every {x 0 , x 0 } in V nd , x 0 and x 0 have a distinguishing input sequence in G o ), there existsθ ∈ Z + such that in (M V nd )θ , the sum of each column is less than 1. Then the spectral radius of (M V nd )θ is less than 1, so is M V nd . Hence lim θ→∞ (M V nd ) θ has all entries equal to 0. That is, for all x 0 , x 0 with (x 0 , x 0 ) ∈ V nd , lim θ→∞ Prob{ y(θ ; σ, x 0 ) = y(θ ; σ, x 0 )} = 0, lim θ→∞ Prob{ y(θ ; σ, x 0 ) = y(θ ; σ, x 0 )} = 1. For all x 0 , x 0 with H x 0 = H x 0 , Prob{ y(θ ; σ, x 0 ) = y(θ ; σ, x 0 )} = 1 for any θ ∈ N. Then (65) is asymptotically observable in distribution.
Based on the above discussion, a PBN (65) is asymptotically observable in distribution if and only if the corresponding BCN (66) satisfies Definition 12 and in the observability graph of (66), there is no path from any nondiagonal vertex to any diagonal vertex, which is exactly [28,Proposition 1]. Hence asymptotic observability in distribution of PBNs can be regarded as a slightly stronger version of multiple-experiment observability of BCNs (Definition 12).

Remark 16
The above three conclusions show that the three definitions of observability in probability on 0, θ with θ ∈ N, finite-time observability in probability, and asymptotic observability in distribution are rather close to each other and all can be regarded as slight variants of multiple-experiment observability of BCNs (the first and the third are slightly stronger than multiple-experiment observability, the second is exactly multiple-experiment observability). In addition, the three definitions do not depend on probability distributions of the stochastic switching signal σ . Formally, given two probability distributions p i , i = 1, 2, for a PBN (65), one has (65) with p = p 1 is observable if and only if PBN (65) with p = p 2 is observable with respect to any one of the above three definitions. [28,Lemma 3] is exactly [10,Theorem 6]. Note that this result was earlier restated in [20] and [13,Remark 4.1]. In [28], observability of switched BNs was mentioned. As one can easily see, observability of switched BCNs can be equivalently transformed to observability of BCNs if inputs and switching signals of switched BCNs are with the same quantifier (either ∃ or ∀). For a given switched BCN, one could regard the Cartesian product of the set of inputs and the codomain of the switching signal as the new set of inputs so that a new equivalent BCN is obtained. The results in [51] showed that Definition 14 extended to switched BCNs for which the switching signals are with the ∃ quantifier is actually Definition 14 of BCNs. As a sequence, controllability of switched BCNs studied in [52] (in which inputs and switching signals are both with ∃ quantifier) is actually controllability of BCNs.

Concluding remarks
This paper provided a thorough survey for the results related to observability of general Boolean control networks (BCNs). The contents of the paper were guided by two methods-Edward F. Moore's partition and our observability graph. First, Moore's partition closely relates the following problems: computation of smallest invariant dual subspaces of Boolean networks (BNs) containing a set of Boolean functions, multiple-experiment observability verification/decomposition in BCNs, and disturbance decoupling in BCNs. However, this method does not apply to other types of observability or nondeterministic systems. Second, based on our observability graph, four different types of observability were verified in BCNs, verification results were also extended to probabilistic BNs, singular BCNs, and nondeterministic finite-transition systems (NFTSs).
All the methods introduced in the paper are state-based, i.e., with respect to BCNs whose dependency graphs are with different degrees of sparsity, the methods show similar efficiencies. Hence, such methods generally cannot be used to verify observability of large-scale BCNs (i.e., those with more than approximately 30 nodes) in a reasonable amount of time. It is known that verifying observability of BNs is NP-hard [53], then it is also NP-hard to verify all different types of observability of BCNs, because different types of BCNs coincide with each other when the inputs are constant (see [53] and [13,Section 4.4]). Hence, it is very difficult to efficiently verify observability for large-scale BCNs. One way of overcoming the high complexity of verifying a large-scale BCN is to first partition the BCN into smallscale subnetworks (with not more than 20 nodes), and then verify these small subnetworks. Two classes of such partitions (called node aggregations) were given in [24]. In the first class, the subnetworks are BCNs, and the aggregated networks are acyclic; in the second class, the subnetworks are NFTSs, and the aggregated networks are not necessarily acyclic. The partition method was successfully applied to analyze arbitrary-experiment observability of a Boolean Tcell receptor kinetics model built in [54] with 37 state nodes and 3 input nodes. The result on the T-cell model obtained in [24] is the unique minimal set of 16 state nodes needed to be directly measured to make the overall model observable, as well as 5 of the 16 state nodes needed to be directly measured to make the model arbitrary-experiment reconstructible.
Efficient verification can be done in several special types of BCNs, e.g., a polynomial-time observability verification algorithm for conjunctive BNs (which only contain logical operator ∧) was given in [55].
All the results collected in the paper are state-based. As mentioned above, for different types of BCNs, they show similar efficiencies. In [48], function-based necessary and sufficient conditions for controllability and multipleexperiment observability of BCNs were given. The method is based on an algebraic form of a BCN, i.e., regarding each node as a variable in the finite field F 2 and regarding Boolean functions as polynomial functions on the vector space F n 2 . Then controllability and observability were verified by computing two affine varieties and checking their equality by comparing their Gröbner bases (that are computable). The verification algorithms depend on the lengths of Boolean functions of a BCN. Hence, when a BCN is sparse, the lengths of its Boolean functions are relatively small, the algorithms are efficient. Based on the algorithms, in [48], controllability and multiple-experiment observability were verified directly over the above mentioned Boolean T-cell model built in [54]. Different from the results obtained in [24], the unique minimal set of states nodes needed to be directly measured to make the overall model multiple-experiment observable is 14.
Recently, the study on the observability synthesis problem in BCNs based on state feedback started [56], in which only arbitrary-experiment observability was studied. The results were obtained by combining the notion of observability graph and the semitensor product of matrices. The synthesis algorithm designed therein is preliminary, a lot of work could be done to improve its efficiency and performance.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, 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 licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence 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. To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/.