Professional Documents
Culture Documents
6, JUNE 2016
Abstract Driven stable nonlinear feedback shift which limits their application to those codes with short
registers (NFSRs) with inputs are not only able to limit constraint lengths [8]. Although technological advancements
error propagations in convolutional decoders, but also helpful in microelectronics have made possible the introductions
to analyze the period properties of sequences generated by a
cascade connection of NFSRs in stream ciphers. An NFSR is (or modified versions) of both algorithms in cellular
driven stable if and only if the reachable set is a subset of telephones, relatively high cost is required in general.
the basin. Due to lack of efficient algebraic tools, the driven Compared to the other three algorithms, the threshold
stability of NFSRs with inputs has been much less studied. algorithm is less efficient, but is easier to implement, which
This paper continues to address this research using a Boolean leads to its applications in telephony and high frequency
control network approach. Viewing an NFSR with input as
a Boolean control network, we first give its Boolean control radio, where a moderate amount of coding gain is desired at
network representation, which is characterized with a state a relatively low cost [5], [8].
transition matrix. Some properties of the state transition matrix During the process of decoding, a decoding error tends
are then provided. Based on these, explicit forms are given to induce indefinitely long decoding errors. At the expense
for the reachable set and the set of basin. Two algorithms for of encoding queues, some schemes such as periodic
obtaining both the sets are provided as well. Compared with the
exhaustive search and the existing state operator method, the re-synchronization were suggested to control this error
Boolean control network approach requires lower computational propagation. In 1964, Massey and Liu proposed an alternative
complexity for those NFSRs with their stages greater than 1. way to limit such an error propagation, that is, using a
Index Terms Shift register, stability, convolutional decoder, driven stable nonlinear feedback shift register (NFSR) as
computational complexity, state transition matrix. the main building block in a convolutional decoder [9].
In their NFSR-based decoder, the feedback function represents
I. I NTRODUCTION a decoding algorithm (e.g., threshold algorithm). They also
gave an example therein to highlight the application of the
C ONVOLUTIONAL codes have been used in
many communication applications, such as mobile
communications and satellite communications. They mainly
NFSR-based decoder.
Unlike the well-developed theory of linear feedback shift
have several decoding algorithms, including the sequential, registers (LFSRs), the theory of NFSRs has not yet been
threshold, Viterbi, and BCJR. The sequential algorithm is well-understood, due to its complexity and lack of efficient
a probabilistic decoding method proposed by Wozencraft tools. Nevertheless, besides the application in convolutional
in 1961 [1]. In 1963, Massy proposed an algebraic decoding coders/decoders, LFSRs and NFSRs have been widely used in
method, called the threshold algorithm [2]. The Viterbi stream ciphers. In particular, NFSRs are more used in recently
algorithm is an optimum decoding method proposed by developed stream ciphers, such as Grain [10], Trivium [11]
Viterbi in 1967 [3], performing maximum likelihood and Mickey [12], thanks to their increased resistance to
decoding [5][7]. However, this does not imply that the Viterbi cryptanalysis attacks. Since the introductions of Grain and
algorithm is the best for all applications, since there are severe Trivium in 2005, the stream ciphers in Grain (or Trivium)
constraints imposed by hardware complexity [8]. The BCJR family have been attracting much attention. However, their
algorithm is an algorithm for maximum a posteriori decoding, cryptographic properties including period properties have not
proposed by Bahl, Cocke, Jelinek, and Raviv in 1974 [4]. been well-understood. Note that both Grain and Trivium
For the Viterbi and BCJR algorithms, their decoder use cascade connections of NFSRs as their main building
complexity grows exponentially with constraint length, blocks. In a cascade connection of NFSRs, those controlling
NFSRs are cascaded with their controlled NFSRs through
Manuscript received November 5, 2015; revised March 31, 2016; accepted their outputs, which implies that their outputs are the inputs
April 13, 2016. Date of publication April 21, 2016; date of current
version June 14, 2016. This work was supported in part by the Strategic of their controlled NFSRs. The periods of NFSR sequences
Priority Research Program of CAS under Grant No. XDA06010701, by the must be considered in order to analyze the cryptographical
National 973 Program of China under Grant No. 2011CB302400, and by the security of NFSR-based stream ciphers. Driven stable NFSRs
National Natural Science Foundation of China under Grant Nos. 61379139
and 61104075. The associate editor coordinating the review of this paper and with inputs are helpful to analyze the period properties of the
approving it for publication was L. Dolecek. sequences generated by a cascade connection of NFSRs in
The authors are with the State Key Laboratory of Information Security, stream ciphers.
Institute of Information Engineering, Chinese Academy of Sciences,
Beijing 100093, China (e-mail: zhongjianghua@iie.ac.cn; ddlin@iie.ac.cn). Those NFSRs with inputs holding constantly at zero
Digital Object Identifier 10.1109/TCOMM.2016.2557330 are called autonomous NFSRs, while the others are
0090-6778 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHONG AND LIN: DRIVEN STABILITY OF NFSRs WITH INPUTS 2275
called non-autonomous NFSRs. Usually, we also regard an In this paper, we restrict ourselves to the NFSRs with single
autonomous NFSR as an NFSR without input. Numerous inputs. The extension to the NFSRs with multiple inputs is
efforts have been made on autonomous NFSRs over the straightforward.
past decades [13][17]. However, most of them focused on The contribution of this paper is: a novel approach,
their periods or feedback functions. There are several studies called Boolean control network approach, which is proposed
addressed their stability. A sufficient condition was given to facilitate the study of driven stability of NFSRs with
to determine the global stability [9]. The relations between inputs, and leads to lower computational complexity than
stable NFSRs were revealed by Mowle [19]. He also provided the exhaustive search and the existing state operator
an algorithm for generating stable NFSRs [20]. k-stable method.
autonomous NFSRs were studied in [21]. The remainder of this paper is organized as follows.
In contrast, much fewer efforts were made on Section II briefly reviews some related work on Boolean
non-autonomous NFSRs. Massey and Liu introduced control networks. Section III gives the Boolean control
the notion of driven stability of non-autonomous NFSRs network representations of NFSRs. Section IV is our main
in [9], but they did not give any further study therein. Mowle results on driven stability of NFSRs. The paper is concluded
proposed a programmable procedure to determine the driven in Section V.
stability of non-autonomous NFSRs, via a next-state operator
and a past-state operator [22]. However, the state operator II. B OOLEAN C ONTROL N ETWORK
method still leads to relatively high time complexity of In this section, we first briefly review the semi-tensor
computations. To the best knowledge of the authors, since product of matrices. We then recall the multi-linear form of
then the driven stability of non-autonomous NFSRs has not a Boolean function, using the semi-tensor product. Finally,
been further studied, due to lack of efficient tools. we revisit some related results on Boolean control networks.
An NFSR was viewed as a finite-state automaton in [23] and Before all these introductions, we first give some notations
as a finite-state machine in [13]. In particular, an autonomous used in this paper.
NFSR was viewed as a Boolean network in [24][27].
F2 : the binary Galois field.
A Boolean network is an autonomous system that evolves as
Fn2 : the set of all n-dimensional vectors over F2 .
a finite state automaton through Boolean functions. Boolean
N: the set of nonnegative integers.
network was first introduced by Kauffman in 1969 to
In : the identity matrix of dimension n.
model a genetic network whose describing variables take
ni : the i -th column of the identity matrix In .
only two values, on and off (or equivalently, 1 and 0,
n = {ni |i = 1, 2, , n}.
respectively) [28]. Of course, the behavior of a Boolean
m n : the set of all m-dimensional vectors over n .
network depends on its internal structure. However, if the
Lnm : the set of n m matrices, whose columns belong
network behavior also depends on an external force, called
to n . If L Lnm , then L = [ni1 ni2 nim ]. For
(control) input, then the concept of a Boolean network is
the sake of simplicity, we write L in a compact form,
naturally extended to that of a Boolean control network.
as L = n [i 1 i 2 i m ].
Over the last decades Boolean (control) networks have
Col j (A): the j -th column of a matrix A.
attracted much attention in many communities, ranging
Col(A): the set of all columns of the matrix A.
from biology [29][31] and physics [32][34] to system
+, and : the ordinary addition, subtraction and
science [35][40].
multiplication in the real field, respectively.
In the community of system science, Cheng and his
and : the addition and multiplication modulo 2
co-workers developed an algebraic framework for Boolean
over F2 , respectively.
(control) networks, using a semi-tensor product approach [41].
In their work, a Boolean (control) network can be equivalently
converted into an algebraic form. Thanks to their algebraic A. Semi-Tensor Product
set-up, the problems related to Boolean functions can be Semi-tensor product of matrices was introduced by
converted into algebraic problems, which is very helpful to Cheng [41]. It is a generalization of the conventional matrix
analyze NFSRs. product, while it retains all major properties of the latter, such
Similar to an autonomous NFSR, it is natural to view as the associative law and the distributive law. The semi-tensor
a non-autonomous NFSR as a Boolean control network. product is defined as follows.
Motivated by this, this paper addresses the problem of driven Definition 1 [41]: Let A and B be matrices of dimensions
stability of NFSRs with inputs using a Boolean control n m and p q, respectively, and let be the least common
network approach. An NFSR is driven stable if and only multiple of m and p. The (left) semi-tensor product of A and B
q
if the reachable set is a subset of the basin. Viewing an m p matrix, given by
is defined as an n
NFSR with input as a Boolean control network, we first
A B = (A I m )(B I p ), (1)
give its Boolean control network representation, which is
characterized with a state transition matrix. Some properties where denotes the Kronecker Product [42].
of the matrix are then provided. Based on these, explicit Clearly, in Definition 1 if m = p, then the semi-tensor
expressions are given for the reachable set and the set of basin. product of A and B is reduced to their conventional
Two algorithms for obtaining both sets are provided as well. matrix product. As in the convention matrix product,
2276 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 64, NO. 6, JUNE 2016
in the semi-tenor product the power of a matrix A is Lemma 4 [41]: Boolean control network (5) has an
defined as Ak+1 = A Ak for any positive integer k. equivalent algebraic representation
A toolbox written in Matlab is available at the website
http://lsc.amss.ac.cn/dcheng/stp/STP.zip for some related x(t + 1) = L u x(t)u(t), t N, (6)
computations of the semi-tensor product. For the sake of where x 2n is the state, u 2m is the input, and
simplicity, in the sequel the symbol is omitted between
L u L2n 2n+m is the state transition matrix, satisfying
matrices if there is no confusion.
Col j (L u ) = Col j (G u1 ) Col j (G un ) (7)
B. Multi-Linear Form of Boolean Function
for all j = 1, 2, . . . , 2n+m , with G ui being the structure matrix
A Boolean function f with n variables is a mapping from of the i -th component of the vectorial function gu in (5) for
Fn2 to F2 , that is, f : Fn2 F2 . It is a linear (resp. nonlinear) any i {1, 2, . . . , n}.
Boolean function if f is a linear (resp. nonlinear) mapping Similarly, for the Boolean network
with respect to its variables in F2 . Let i be the decimal
number corresponding to the binary (i 1 , i 2 , . . . , i n ) via the X(t + 1) = g(X(t)), t N, (8)
mapping i = i 1 2n1 + i 2 2n2 + + i n . Then i ranges
from 0 to 2n 1. For the sake of simplicity, we denote with the state X = [X 1 X 2 X n ]T Fn2 , and the vectorial
f (i ) = f (i 1 , i 2 , . . . , i n ). Then [ f (0), f (1), . . . , f (2n 1)] function g : Fn2 Fn2 , it also has an equivalent algebraic
is called the truth table of f , arranged in the alphabet order, representation:
and [ f (2n 1), f (2n 2), . . . , f (0)] is called the truth table
x(t + 1) = Lx(t), t N, (9)
of f , arranged in the reverse alphabet order.
Identify 1 and 0, respectively, as [1 , 0]T and [0 1]T . where x 2n is the state, and L L2n 2n is the state
Accordingly, identify a variable X F2 as [X X 1]T . We say transition matrix, satisfying
[1 0]T , [0 1]T and [X X 1]T are, respectively, the vector
forms of their scalar forms, 1, 0 and X. To distinguish the Col j (L) = Col j (G 1 ) Col j (G n ) (10)
scalar form and the vector form of a variable, in the sequel
for all j = 1, 2, . . . , 2n , with G i being the structure matrix of
we use the notation X for a variable in F2 , while we use the
the i -th component of g in (8) for any i {1, 2, . . . , n}.
notation x for its vector variable in 2 , that is, X F2 , but
x 2 . Using the above vector forms, a Boolean function f
is changed to f : n2 2 . D. Reachable Set of Boolean Control Network
Lemma 2 [40], [43]: Any Boolean function f (x 1 ,
Definition 5 [40]: Consider Boolean control network (6).
x 2 , . . . , x n ) with x 1 , x 2 , . . . , x n 2 can be expressed as a
Given an initial state x0 and a destination state xd , xd is said
multi-linear form:
to be reachable from x0 if there exists an input sequence u(t0 ),
f (x 1 , x 2 , . . . , x n ) = F x 1 x 2 x n , (2) u(t0 + 1), . . . , u(td ), such that x(t0 ) = x0 and x(td ) = xd .
Lemma 6 [40]: Let L u L2n 2n+1 . Then
where F is called the structure matrix of f , and is uniquely
expressed as
u 2 n = L u
L k+1 L ku 2i n I2
i
(11)
s1 s2 ... s2n
F= (3) for any positive integer k and any i {1, 2, . . . , 2n }.
1 s1 1 s2 . . . 1 s2n
Let R(x0 ) represent the set of all states that are reachable
with [s1 , s2 , . . . , s2n ] being the truth table of f , arranged in from x0 via some input sequences. Then R(x0 ) is usually
the reverse alphabet order. called the reachable set of x0 .
Lemma 3 [40]: Suppose Lemma 7 [40]: Consider the Boolean control network (6).
x = x1 x2 xn (4) Assume p is the minimum positive integer p satisfying
by a cascade connection of NFSRs in stream ciphers. Take a for all i = 1, 2, . . . , n. Therefore, we have p2n+1 = 2n .
simple example, a cascade connection of NFSR1 into NFSR2, In particular, if the input U (t) 0 for any t N, then
in which NFSR1 is autonomous and controls NFSR2 through the non-autonomous NSFR is reduced to an autonomous one,
2278 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 64, NO. 6, JUNE 2016
q
which can be viewed as a Boolean network. In this case, for the state transition matrix L be Col j (L) = 2nj . Then similar
the sake of simplicity, we denote to Theorem 8, we have
h(X 1 , X 2 , . . . , X n ) = f (X 1 , X 2 , . . . , X n , 0), (1) (2) (n1) (n)
(16) q j = 2n 2n1 j 2n2 j 2 j j
and we say h the feedback function of its corresponding (22)
autonomous NFSR. Since f has been assumed to satisfy
for each j {1, 2, . . . , 2n }.
f (0, 0 . . . , 0, 0) = 0, we have h(0, 0, . . . , 0) = 0 for the
For any k {1, 2, . . . , 2n+1 }, denote by
autonomous NFSR throughout the paper. [k] [k] [k]
[X 1 X n U ] the state [X 1 X n U ]T Fn+1
T
2
The autonomous NFSR is called a Galois NFSR, where
corresponding to the decimal number 2n+1 k. Then, we
the feedback is applied to each storage device. If the
have
feedback is only applied to the right-most storage device,
1, if k is odd,
that is, the feedback connection vector a = [0 0 1], U = [k]
(23)
then the autonomous NFSR is called a Fibonacci NFSR. 0, if k is even.
Reference [15] showed that Galois NFSRs are equivalent
Similarly, for any j {1, 2, . . . , 2n }, denote by
to Fibonacci NFSRs. Moreover, Fibonacci NFSRs are more [ j] [ j] [ j]
commonly used to generate pseudo-random sequences for [X 1 X n1 X n ]T
the state [X 1 X n1 X n ]T Fn2
cryptographical securities. Throughout the paper, we assume corresponding to the decimal number 2n j . Note that the
that the feedback connection vector is a = [0 0 1], and the state transition matrices gu in (12) and g in (17) satisfy g(X) =
[ j] [ j]
autonomous NFSRs always mean the Fibonacci NFSRs with gu (X, 0). Then, [X 1 X n1 0] = [X 1[k] X n1
[k]
U [k] ]
their feedback functions h satisfying h(0, 0, . . . , 0) = 0.
with k = 2 j . Thus, we deduce that the components sk(i) ,
The n-stage autonomous NFSR can be expressed as a
k = 1, 2, . . . , 2n+1 , i = 1, 2, . . . , n, of the truth table of gui
nonlinear system (i)
and the components j , j = 1, 2, . . . , 2n , i = 1, 2, . . . , n, of
X(t + 1) = g(X(t)), (17) the truth table of gi satisfy sk(i) = j(i) when k = 2 j . Hence,
according to Equations (15) and (22), we have q j = p2 j .
where g(X)
= gu (X, 0). For any positive integer N, let Therefore, the results follows.
g N+1 (X) = g g N (X) , which indicates that the state g N (X)
Recall that a Boolean function (X 1 , . . . , X n , U ) is
is shifted N times from X. For the sake of simplicity, in
degenerated with respect to the variable U if is independent
the sequel we say the NFSR represented by (17) is the
of U , that is, (X 1 , . . . , X n , 1) = (X 1 , . . . , X n , 0).
corresponding autonomous NFSR of the non-autonomous one
A vectorial function u = [u1 u2 un ]T is degenerated
represented by (12).
with respect to the variable U if each component ui ,
For the autonomous NFSR, Reference [26] reveals the
i {1, 2, . . . , n}, is degenerated with respect to the variable U .
simple relation between the state transition matrix and the
If the state transition function gu in (12) is degenerated with
true table of the feedback function h. Assume the truth table
respect to the variable U , then the non-autonomous NFSR is
of h in (16) to be [1 , 2 , . . . , 2n ], arranged in the reverse
reduced to an autonomous NFSR.
alphabet order. Then the autonomous NFSR has an algebraic
Proposition 10: The state transition function gu
representation [26]:
in (12) is not degenerated with respect to the variable U if
x(t + 1) = Lx(t), t N, (18) and only if the state transition matrix L u in (14) satisfies
transition matrix L u in (14) satisfies integer N such that the state transition function g in (17)
(or equivalently, the state transition matrix L in (18)) satisfies
Col2 j 1 (L u ) = Col2 j (L u ), j = 1, 2, . . . , 2n . (26) g N (X) = 0 (or equivalently, L N x = 22n ).
n
Theorem 19: For an n-stage NFSR with input represented show that Col(L u 88 ) = {83 , 88 }, Col(L 2u 88 ) = {82 , 83 , 85 , 88 },
by (14), if the state transition function gu in (12) is degenerated Col(L 3u 88 ) = {81 , 82 , 83 , 85 , 86 , 88 }, and Col(L 4u 88 ) = 8 .
with respect to the variable U , then the reachable set Thus, according to Theorem 20, we obtain the reachable set
n
n
n 4
R(22n ) = Col L u 22n = {22n }. (27) R(88 ) = Col L ku 88 = 8 .
k=1
Proof: From Corollary 11, we have Col2n+1 1 (L u ) = Corollary 22: If the state transition function gu in (12) of
Col2n+1 (L u ). Taking into account Theorem 8, we deduce that an n-stage non-autonomous NFSR is not degenerated with
n
Col2n+1 1 (L u ) = Col2n+1 (L u ) = 22n . In addition,
according
respect to the variable U , and the non-autonomous NFSR is
n
to Lemma 18, we have Col L 2u 22n = Col L u 22n =
n driven stable, then its corresponding autonomous NFSR must
be locally stable.
{Col2n+1 1 (L u ), Col2n+1 (L u )}. Therefore, the result follows Proof: Let L u be the state transition matrix of the
from Lemma 7. non-autonomous NFSR. Then, according to the proof of
Theorem 20: For an n-stage NFSR with input represented Theorem 20, the two distinct columns Col2n+1 1 (L u ) and
by (14), if the state transition function gu in (12) is not n
Col2n+1 (L u ) are in the reachable set R(22n ). Since the NFSR
degenerated with respect to the variable U , then the reachable n
is driven stable and Col2n+1 (L u ) = 2n , we deduce that the
2
set n
n1
2 state Col2n+1 1 (L u ) that is not equal to 22n eventually reaches
n
n
R(22n ) =
n
Col L ku 22n . (28) the equilibrium state 22n . Thus, the result follows.
k=1 Next, we strive to give an algorithm in order to obtain
n n
n n the reachable set R(22n ). Lemma 6 shows that L k+1 u 2 n =
2
Moreover, the cardinality |R(22n )| of the reachable set R(22n ) n n
n
satisfies |R(22n )| 2. L u (L ku 22n ) I2 for any positive integer k. Let Yk = L ku 22n .
n
Proof: Straightforward computations show that L ku 22n Then Yk+1 = L u (Yk I2 ) and Yk L2n+1 2k . To obtain the
n n
L2n 2k and Col(L ku 22n ) Col(L u ) for any positive integer k. reachable set R(22n ), at most 2n1 iterations are required to
Since the state transition function gu in (12) is not degenerated do for Yk . To lower the spacial complexity of computations,
with respect to the variable U , according to Proposition 10 we we can only store the positions of the entry 1s in the columns
have Col2 j 1 (L
u )
= Col2 j (L u ), j =1, 2, . . .
, 2n . Hence,
if of all vectors and matrices.
k0 +1 2n k0 n
Col L u 2n is not a subset of i=1 Col L u 2n i 2 for Suppose that the state transition matrix
some positive integer k0 , then from Lemmas 6 and 18, we L u = 2n [ p1 p2 . . . p2n+1 ] (29)
deduce that L ku0 +1 22n has two more distinct columns than
n
Algorithm 1 Reachable Set that has only one predecessor. On the other hand, an n-stage
1. Set = 2n and R(2n ) = . NFSR has totally 2n possible states. Hence, Ns + Nb 2n .
2. Compute (), and set = (). Lemma 12 shows that Ns = Nb . Then the result follows.
3. Compute ( ), and set = ( ). Proposition 25: If the feedback function h(X 1 ,
4. Remove the repeated elements of , and set to the X 2 , . . . , X n ) of an n-stage autonomous NFSR satisfies
resulting vector. h(0, 0 . . . , 0) = h(1, 0, . . . , 0) = 0, then there is at least one
5. If Col() R(2n ), then remove the s columns that starting state in the state diagram of the autonomous NFSR.
belong to R(2n ),and set to the resulting vector and set Proof: If the feedback function h satisfying
R(2n ) = R(2n ) Col(), and goto Step 2. Otherwise, h(0, 0 . . . , 0) = h(1, 0, . . . , 0) = 0, then the state [0 0 0]T
output R(2n ) and stop. has two predecessors, [1 0 0]T and itself, which implies
that [0 0 0]T is a branch state. Thus, the result follows
from Lemma 12.
those of previously obtained s. In Matlab programme, we Reference [26] gives a way to find all starting states of an
can use the operation unique for the removal in Step 4, and autonomous NFSR, shown in the following lemma.
use the operation setdiff for the removal in Step 5. Since L u Lemma 26 [26]: Let L be the transition matrix of
has totally 2n+1 columns and has at most 2n distinct columns, an n-stage autonomous NFSR represented by (18). Then
j
the time complexity of computations for the two removals is 2n , j {1, 2, . . . , 2n }, is a starting state if and only if
j
at most 2n+1 + 2n operations. Moreover, the time complexity 2 n / Col(L).
of computations for the mapping is at most 2n+1 operations, Theorem 27: For an n-stage autonomous NFSR
and the time complexity of computations for the mapping with a feedback function h(X 1 , X 2 , . . . , X n ) satisfying
is at most 2n . In total, if the matrix L u is known, the time h(0, 0 . . . , 0) = h(1, 0, . . . , 0) = 0, let L be its state transition
n
complexity of computing R(2n ) (or equivalently, the reachable matrix. Then the basin of the equilibrium state 22n is
n
set R(22n )) is at most 2n+2 + 2n+1 operations. n
B(22n ) = L k 2i n |1 k K i , K i is the smallest ki
n
satisfying L ki 2i n = 22n , and 2i n
/ Col (L) with (33)
C. Basin some positive integer i 2n } .
Theorem 23: If the feedback function h(X 1 , X 2 , . . . , X n ) Proof: The result follows from Proposition 25 and
of an n-stage autonomous NFSR satisfies h(0, 0, . . . , 0) = 0 Lemma 26.
n n
and h(1, 0, . . . , 0) = 1, then the basin B(22n ) = {22n }. Example 28: Consider a 3-stage autonomous NFSR with a
Moreover, if the state transition function gu in (12) of its feedback function h(X 1 , X 2 , X 3 ) = X 1 X 2 X 3 . Obviously,
corresponding non-autonomous NFSR is (not) degenerated h(0, 0, 0) = h(1, 0, 0) = 0. Using Equations (19) and (20),
with respect to the variable U , then the non-autonomous NFSR we can easily obtain the state transition matrix L =
is (not) driven stable.
n
8 [2 3 5 8 1 4 5 8]. Clearly, only 86 , 87 / Col(L). For the
Proof: The state [0 0 0]T (or equivalently, the state 22n ) state 86 , it is easy to see that L86 = Col6 (L) = 84 , and
of the autonomous NFSR has two possible predecessors, L 2 86 = L(L86 ) = L84 = Col4 (L) = 88 . Thus, according to
[1 0 0]T and itself. Since the feedback function h satisfies
Theorem 27, we have {84 , 86 , 88 } R(88 ). Similarly, for the
h(0, 0, . . . , 0) = 0 and h(1, 0, . . . , 0) = 1, we can deduce that
state 86 , we can directly compute that L87 = 85 , L 2 87 = 81 ,
[0 0 0]T has itself as the only one predecessor. It implies
that the state [0 0 0]T forms a unit cycle, and this cycle L 3 87 = 82 , L 4 87 = 83 , and L 5 87 = 85 . It implies that 87
does not connect with any other states. Therefore, we have reaches a cycle consisting of states 85 , 81 , 82 , 83 , and keeps
n n n
B(22n ) = {22n }, which yields the cardinality of B(22n ) is 1. staying on this cycle and therefore, never reaches the state 88 .
Moreover, if the state transition function gu of the Hence, R(88 ) = {84 , 86 , 88 }.
non-autonomous NFSR is not degenerated with respect to the Note that an equilibrium state is a particular attractor (i.e.,
variable U , then according to Theorem 20, the cardinality a unit cycle). The K i in (33) is actually the transient period
n
of the reachable set R(22n ) is at least 2, which implies that of the state 2i n . Recall that for a given state x0 2n
n n
R(22n ) is not a subset of B(22n ). Therefore, in this case the of a Boolean network, the transient period of x0 , denoted
non-autonomous NFSR is not driven stable. However, if the by K (x0 ), is the smallest k such that x(0) = x0 and x(k) ,
state transition function gu of the non-autonomous NFSR is where is an attractor (i.e., a cycle) of the Boolean network.
degenerated with respect to the variable U , then according to Accordingly, the transient period of the Boolean network
n n
Theorem 19, the reachable set R(22n ) = {22n }, which indicates is defined as the maximum number of all K (x0 )s for any
n n x 0 2 n .
R(22n ) = B(22n ). Hence, in this case the non-autonomous
Cheng et al. gave an explicit form of the basin of a cycle C
NFSR is driven stable.
for a Boolean network in [37]. Let L and K , respectively,
Proposition 24: The number of starting states of an n-stage
be the state transition matrix and the transient period of the
autonomous NFSR is at most 2n1 .
Boolean network. Denote L (C) = {x|L x C} for any
Proof: Let Ns be the number of starting states, and Nb be
positive integer . Then the explicit form of the basin of the
the number of branch states. For any state of an autonomous
cycle C given in [37] is
NFSR, it must be a starting state that has no predecessors,
or be a branch state that has two predecessors, or be a state B(C) = L 1 (C) L 2 (C) L K (C). (34)
2282 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 64, NO. 6, JUNE 2016
Clearly, in order to find the basin of a particular cycle, D. Computational Complexity Comparison
n
an equilibrium state C = 22n , their method is to find the In this subsection, we compare the computational
n
predecessors successively from the equilibrium state 22n , while complexity of determining the driven stability of NFSRs
our method given in Theorem 27 is to find the successors for three methods: the exhaustive search, the existing state
successively from starting states. operator method, and our proposed Boolean control network
Reference [37] has shown that the transient period K of approach.
the Boolean network is equal to the smallest satisfying Assume that the algebraic normal form of the feedback
L {L +1 , L +2 , . . . , L 2 }. Computing such a K results
2n
function of the non-autonomous NFSR is known, denoted
2n+1 1
in (22n 1)2n operations, since it is required to compute by f (X 1 , . . . , X n , U ) = i=0 ai X 1i1 X nin U in+1 , where
the positions of the entry 1s in the columns of the ai F2 and i is the decimal number corresponding to the
2n
matrices L 2 , L 3 , . . . , L 2 , and the positions of the entry 1s binary (i 1 , . . . , i n , i n+1 ). Therefore, the algebraic normal form
in the columns of L +1 are just a permutation of those of the feedback function h(X 1 , . . . , X n ) = f (X 1 , . . . , X n , 0),
in the columns of L . In addition, computing L (C) is known as well. Suppose du and d to be the algebraic degrees
requires the exhaustive search of the positions of the entry of f (X 1 , . . . , X n , U ) and h(X 1 , . . . , X n ), respectively.
1s in the columns of L for any positive integer K , 1) Comparison of Time Complexity of Computations to
which results in 2n operations. Assume the matrix L is Obtain the Reachable Set: Using the Boolean control network
known. Then the time complexity of computing B(C) is totally approach, the time complexity of computations to obtain the
(22n + K 1)2n operations. In contrast, since L +1 22n =
n
reachable set is totally at most
L(L 22n ) and there are totally 2n states of an n-stage
n
d
u
n
R n+2 n+1
NFSR, our method given in Theorem 27 results in at most Cbcn = (n + 1)2 +2 (i + 1) +1
2n 1 operations, which is much lower than that of their i
i=1
method. n 2
+ 2 (n + 3n 1) (37)
In the following, we strive to give an algorithm to
n
obtain the basin B(22n ). For an n-stage autonomous NFSR operations. The time complexity of computations for the state
with a feedback function h(X 1 , X 2 , . . . , X n ) satisfying operator method in [22] is
h(0, 0, . . . , 0) = h(1, 0, . . . , 0) = 0, let S be a set of d
u
n+1
starting states in 2n , and denote its cardinality as |S|. Then Cso = 2
R n+3
N +2 n+1
(i + 1) +n+ N +1
i
1 |S| 2n1 . Suppose that the starting state set S has been i=1
obtained in term of Lemma 26, and its elements are denoted
j
+ 2n (n 2 + n 2) (38)
by 2vn , v = 1, 2, . . . , |S|. Let
operations, while it is
J = { j1, j2 , . . . , j|S | }, (35)
du
n+1
2n
R
Ces =2 4n + (i + 1) (39)
i
which is a set of the positions of the entry 1s of all the elements i=1
in the starting state S. R is lower
operations for the exhaustive search. Clearly, Cbcn
Similar to R(2n ), we define by B(2n ) the set of elements than Ces if n > 1, and it is lower than Cso for any positive
R R
that represent the positions of the entry 1s of all elements integer n.
n
in B(22n ). For the sake of convenience, we also called B(2n ) 2) Comparison of Time Complexity of Computations to
n
the basin of the equilibrium state 22n . Assume that the state Obtain the Basin: Using the Boolean control network
transition matrix L = 2n [q1 q2 . . . q2n ] is known. Define a B to obtain the basin is at
approach, the time complexity Cbcn
mapping most
d
: (i ) = qi . (36) n
Cbcn = 2
B n+2
+2 n
(i + 1) +2 1 (40)
i
Then the mapping and the matrix L are a one-to-one i=1
correspondence. Moreover, k and L k are a one-to-one operations. The time complexity of computations is
correspondence as well for any positive integer k. Note that d
n
any starting state of an NFSR eventually reaches a cycle and Ces = 2
B 2n
(i + 1) + 2n (41)
n i
keeps staying on it. B(22n ) is constituted by the starting states i=1
n
that eventually reach the state 22n and the states that those operations for the exhaustive search, while it is
starting states go through. Moreover, if there is some branch d
n
state, then the states that the branch state goes through is only B
Cso = 2n+2 M + 2n (i + 1) +M +1
required to compute for either starting state that is associated i
i=1
with the branch state.
+ 2n1 (n 2 + 2n 2) (42)
Therefore, based on the mapping and the set J , the
following gives Algorithm 2 to obtain the basin B(2n ) for an operations for the state operator method in [22], where the
n-stage NFSR with a feedback function h(X 1 , X 2 , . . . , X n ) iteration M 2n . Obviously, Cbcn
B is lower than C B if n > 1,
es
satisfying h(0, 0, . . . , 0) = h(1, 0, . . . , 0) = 0. B
and is also lower than Cso for any positive integer n.
ZHONG AND LIN: DRIVEN STABILITY OF NFSRs WITH INPUTS 2283
[22] F. J. Mowle, Readily programmable procedures for the analysis of [41] D. Cheng, H. Qi, and Y. Zhao, An Introduction to Semi-Tensor Product
nonlinear feedback shift registers, IEEE Trans. Comput., vol. C-18, of Matrices and Its Applications. Singapore: World Scientific, 2012.
no. 9, pp. 824829, Sep. 1969. [42] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. Cambridge,
[23] C. Fontaine, Nonlinear feedback shift register, in Encyclopedia of U.K.: Cambridge Univ. Press, 1991.
Cryptography and Security. New York, NY, USA: Springer-Verlag, [43] H. Qi and D. Cheng, Logic and logic-based control, J. Control Theory
2011, pp. 846848. Appl., vol. 6, no. 1, pp. 2636, 2008.
[24] H. Qi, On shift register via semi-tensor product approach, in Proc.
32nd Chin. Control Conf., Xian, China, Jul. 2013, pp. 208212.
[25] D. Zhao, H. Peng, L. Li, S. Hui, and Y. Yang, Novel way to research
nonlinear feedback shift register, Sci. China Inf. Sci., vol. 57, no. 9,
pp. 114, 2014.
[26] J. Zhong and D. Lin, A new linearization method for nonlinear feedback Jianghua Zhong received the B.S. degree in
shift registers, J. Comput. Syst. Sci., vol. 81, no. 4, pp. 783796, 2015. mathematics from Nanchang University, Nanchang,
[27] J. Zhong and D. Lin, Stability of nonlinear feedback shift registers, China, in 2000, the M.S. degree in applied
Sci. China Inf. Sci., vol. 59, no. 1, pp. 112, 2016. mathematics from Fuzhou University, Fuzhou,
[28] S. A. Kauffman, Metabolic stability and epigenesis in randomly China, in 2003, and the Ph.D. degree in system
constructed genetic nets, J. Theor. Biol., vol. 22, no. 3, pp. 437467, theory from the Academy of Mathematics and
1969. Systems Science, Chinese Academy of Sciences,
[29] S. E. Harris, B. K. Sawhill, A. Wuensche, and S. Kauffman, A model Beijing, in 2007. From 2007 to 2009, she was a
of transcriptional regulatory networks based on biases in the observed Post-Doctoral Fellow with the Royal Institute of
regulation rules, Complexity, vol. 7, no. 4, pp. 2340, 2002. Technology, Stockholm, Sweden. After that, she was
[30] S. Huang and D. E. Ingber, Shape-dependent control of cell growth, with the Institute of Complexity Science, Qingdao
differentiation, and apoptosis: Switching between attractors in cell University, Qingdao, China, until 2015, first as an Assistant Professor and
regulatory networks, Experim. Cell Res., vol. 261, no. 1, pp. 91103, then an Associate Professor. Since 2013, she has been with the State Key
2000. Laboratory of Information Security, Institute of Information Engineering,
[31] I. Shmulevich, E. R. Dougherty, S. Kim, and W. Zhang, Probabilistic Chinese Academy of Sciences. Her current research interests include nonlinear
Boolean networks: A rule-based uncertainty model for gene regulatory control, and nonlinear analysis in coding and cryptography.
networks, Bioinformatics, vol. 18, no. 2, pp. 261274, 2002.
[32] R. Albert and A.-L. Barabsi, Dynamics of complex systems: Scaling
laws for the period of Boolean networks, Phys. Rev. Lett., vol. 84,
pp. 56605663, Jun. 2000.
[33] M. Aldana, Boolean dynamics of networks with scale-free topology,
Phys. D, Nonlinear Phenomena, vol. 185, no. 1, pp. 4566, 2003. Dongdai Lin received the M.S. and Ph.D. degrees
[34] B. Samuelsson and C. Troein, Superpolynomial growth in the number in fundamental mathematics from the Institute of
of attractors in Kauffman networks, Phys. Rev. Lett., vol. 90, no. 9, Systems Science, Chinese Academy of Sciences, in
p. 098701, 2003. 1987 and 1990, respectively. He was an Associate
[35] D. Cheng, Input-state approach to Boolean networks, IEEE Trans. Professor from 1993 to 1998, a Professor with the
Neural Netw., vol. 20, no. 3, pp. 512521, Mar. 2009. Institute of Systems Science from 1998 to 2001,
[36] D. Cheng, Disturbance decoupling of Boolean control networks, IEEE and a Professor with the Institute of Software,
Trans. Autom. Control, vol. 56, no. 1, pp. 210, Jan. 2011. Chinese Academy of Sciences, from 2001 to 2011.
[37] D. Cheng and H. Qi, A linear representation of dynamics of Boolean He is currently a Professor with the Institute
networks, IEEE Trans. Autom. Control, vol. 55, no. 10, pp. 22512258, of Information Engineering, Chinese Academy of
Oct. 2010. Sciences, and the Director of the State Key
[38] D. Cheng and H. Qi, Statespace analysis of Boolean networks, IEEE Laboratory of Information Security. He has authored over 200 research papers
Trans. Neural Netw., vol. 21, no. 4, pp. 584594, Apr. 2010. in journals and conference proceedings. He is also involved in multivariate
[39] D. Cheng and H. Qi, Controllability and observability of Boolean public key cryptography, sequences and stream cipher, zero knowledge
control networks, Automatica, vol. 45, no. 7, pp. 16591667, 2009. proof, and network-based cryptographic computation. His research interests
[40] D. Cheng, H. Qi, and Z. Li, Analysis and Control of Boolean Networks: include cryptology, security protocols, symbolic computation, and software
A Semi-Tensor Product Approach. London, U.K.: Springer-Verlag, 2011. development.