This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CYBERNETICS 1

n-Dimensional Discrete Cat Map Generation

Using Laplace Expansions
Yue Wu, Member, IEEE, Zhongyun Hua, Student Member, IEEE, and Yicong Zhou, Senior Member, IEEE

Abstract—Different from existing methods that use matrix For example, the initial condition sensitivity and ergodicity
multiplications and have high computation complexity, this paper of a chaotic system are analogous to the diffusion and con-
proposes an efficient generation method of n-dimensional (nD) fusion properties [21] in cryptography, and the deterministic
Cat maps using Laplace expansions. New parameters are also
introduced to control the spatial configurations of the nD dynamics in a chaotic system corresponds to the deterministic
Cat matrix. Thus, the proposed method provides an efficient pseudorandomness in cryptography [22], [23].
way to mix dynamics of all dimensions at one time. To investi- Among various developed chaotic maps, the Cat map is
gate its implementations and applications, we further introduce a 2-D chaotic map named after Vlamimir Arnold [24], [25].
a fast implementation algorithm of the proposed method with Except for the common properties of chaotic systems, the Cap
time complexity O(n4 ) and a pseudorandom number genera-
tor using the Cat map generated by the proposed method. The map has many unique properties.
experimental results show that, compared with existing genera- 1) It is in the integer form and can be easily adapted to
tion methods, the proposed method has a larger parameter space an arbitrary finite precision [24], which benefits to the
and simpler algorithm complexity, generates nD Cat matrices difference between the precision infiniteness of chaotic
with a lower inner correlation, and thus yields more random systems and precision finiteness of cryptosystems [26].
and unpredictable outputs of nD Cat maps.
2) It is invertible and its inverse is also in the integer form
Index Terms—Arnold’s cat map, cryptography, Laplace expan- because its Cat matrix has determinant 1.
sion, n-dimensional (nD) Cat map. 3) It is area preserving [24], indicating that the Cat map
can be directly used as a permutation function, which is
I. I NTRODUCTION the fundamental building blocks for cryptography.
4) It is topologically transitive [27] and thus it is chaotic
S A TYPE of dynamical systems closely related to nat-
A ural processes, chaotic systems are quite sensitive to
their initial conditions. This indicates that their trajectories are
with high randomness.
5) It is an Anosov diffeomorphism and structurally sta-
ble [25], indicating that small perturbations in the system
exactly decided by their initial states. Any small difference in do not affect the qualitative behaviors of the map’s tra-
their initial states yields significantly different outcomes after jectory, and thus the Cat map itself can resist a certain
long time system evolution [1]–[3]. This is commonly referred level of noise.
as butterfly effect [4]. Examples of chaotic systems include the With these significant properties, the Cat map has been
logistic map [5], Chua’s circuit [6], and Chen–Lee system [7]. studied in both theory and practice and used in different
With significant properties of initial state sensitivity, ergod- subjects. Examples include the Cat map period distribu-
icity, and unpredictability, chaotic systems have been widely tions [28], [29], Cat map properties [30], Quantum Cat
reported in various subjects including mathematics, physics, maps [31]–[34], Cat maps in other domains [35], and Cat map-
computer sciences, biology, engineering, economics, robotics, based pseudorandom number generator (PRNG) [36], image
geology, neuron science, and chemistry [8]–[11]. Among all encryption [24], and watermarking [18].
chaos-based applications, cryptography is the most popular A Cat map usually contains a transformation matrix called
one [12]–[18]. This is because many properties of chaotic the Cat matrix. Several Cat map generation methods have been
systems can be found similar in cryptography [19], [20]. developed to construct different Cat matrices such that the
generated Cat maps have high randomness and large param-
Manuscript received July 9, 2015; accepted September 21, 2015. This work
was supported in part by the Macau Science and Technology Development eter spaces. According to the dimensions of the Cat matri-
Fund under Grant FDCT/017/2012/A1, and in part by the Research ces, these generation methods can be roughly classified into
Committee at the University of Macau under Grant MYRG2014-00003-FST, 2-D [24], [25], 3-D [24], [37]–[39], and n-dimensional (nD)
Grant MRG017/ZYC/2014/FST, Grant MYRG113(Y1-L3)-FST12-ZYC, and
Grant MRG001/ZYC/2013/FST. This paper was recommended by Associate Cat map generation methods [40]–[42]. A 2-D Cat matrix is
Editor J. Cao. (Corresponding author: Yicong Zhou.) a 2 × 2 matrix containing four elements, while a 3-D Cat
Y. Wu is with the Department of Electrical and Computer Engineering, matrix is a 3 × 3 matrix with nine elements. When being used
Tufts University, Medford, MA 02155 USA (e-mail: [email protected]).
Z. Hua and Y. Zhou are with the Department of Computer and Information in a cryptosystem, they commonly have a parameter space
Science, University of Macau, Macau, China (e-mail: [email protected]; less than the required size to resist brute-force attacks [43],
[email protected]). and thus they have to be used with additional components
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. to ensure security [24]. On the other hand, an nD Cat map
Digital Object Identifier 10.1109/TCYB.2015.2483621 with a parametric nD Cat matrix seems to have a quite large
2168-2267 c 2015 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.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CYBERNETICS

parameter space. However, because of using a large number Based on the way how to construct the nD Cat matrix CnD
∗ ,
of matrix multiplications and dependency relationships among there exist several Cat map generation methods.
the elements in the Cat matrix, existing nD Cat map genera-
tion methods produce nD Cat matrices with highly correlated A. Arnold’s Cat Map Generation Method
matrix elements, which may downgrade the security of cryp- The original Cat map called the Arnold’s Cat map is a 2-D
tography applications [40]. A detailed literature review on Cat map. Its Cat matrix C2-D
these methods is given in Section II. A is defined by [25]
 
In this paper, we propose a new nD Cat map generation 1 1
C2-D
A = . (2)
method using Laplace expansions to iteratively construct an 1 2
nD Cat matrix. Its properties are discussed and theoretical
analysis is provided. The main contributions of this paper can In practice, the Arnold’s Cat matrix C2-DA is commonly
be summarized as follows. extended into a parametric form for cryptography applica-
1) To overcome the high computation complexity of exist- tions [24] and defined as follows:
 
ing methods that are rooted in matrix multiplications, we 1 a
propose a new nD Cat matrix generation method that is Cpara. =
2-D
(3)
b ab + 1
the first time to use Laplace expansions to efficiently
 
generate nD Cat matrices.  
para.  = 1 and Cpara.
where a and b are integers. Obviously, C2-D 2-D
2) We further propose new parameters to control the spatial
is indeed a Cat matrix with area preserving, invertibility, and
configurations of an nD Cat matrix. This provides an
other properties.
efficient way to mix dynamics of all dimensions at one
time.
3) We introduce a new fast algorithm to implement the pro- B. 3-D Cat Map Generation Methods
posed method. It can reduce the time complexity from To enhance the security level of the Cat-map-based cryptog-
O(n5 ) to O(n4 ). raphy and other applications, several 3-D Cat map generation
4) We provide theoretical analysis and comprehensive per- methods were developed.
formance evaluations of the proposed method with Lian et al. [37] extended C2-D para. into the x − z and
respect to the parameter space, algorithm complex- y − z planes, and then proposed a parametric 3-D Cat
ity, matrix element correlation, Shannon entropy, and matrix C3-D
L , as
Kolmogorov entropy. The comparison results show that ⎡ ⎤
1 0 a
the proposed method have a larger parameter space, less ⎢ ⎥
computation complexity, more independent, and random L =⎣
C3-D bc 1 abc + c ⎦ (4)
matrix elements and can generate outputs with better bcd + b d abcd + ab + cd + 1
randomness and unpredictability than existing methods.
where a, b, c, and d are integer parameters.
5) To investigate the real-world application of the proposed
Chen et al. [24] constructed a parametric 3-D Cat matrix by
method, we introduce a new PRNG using the Cat maps
firstly extending C2-D
para. with respect to the x, y, and z direc-
generated by the proposed method. The randomness
tions and then mixing all coupled Cat matrices using matrix
quality of the proposed PRNG is evaluated using two
multiplications. Consequently, their 3-D Cat matrix C3-D
C is the
test standards.
product of three extended Cat matrices of pairwise-coupled
The rest of this paper is organized as follows. Section II
dynamics as
briefly reviews the existing Cat map generation methods,
zy
C = CC CC CC
C3-D x
Section III introduces our nD Cat map generation method
using Laplace expansions, Section IV analyzes the proposed ⎡
ax az by + 1 az
method, Section V compares the proposed method with exist- = ⎣ bz + ax by (az bz + 1) az bz + 1
ing methods, Section VI proposes a new PRNG and evaluates by (ax bx + 1) bx
its performance, and Section VII concludes this paper. ⎤
ay + ax az (ay by + 1)
ay bz + ax (ay by + 1)(az bz + 1) ⎦ (5)
II. E XISTING C AT M AP G ENERATION M ETHODS (ax bx + 1)(ay by + 1)
This section briefly reviews the parametric Arnold’s, 3-D, where ax , bx , ay , by , az , and bz are six integer control
and nD Cat map generation methods, and discusses their parameters.
properties. Also based on C2-Dpara. , Liu et al. [38] introduced a para-
All discrete Cat maps can be represented as a general form
metric 3-D Cat matrix C3-D
U by mixing dynamics on all three
Y(t + 1) = CnD
∗ Y(t) mod N (1) dimensions into the extended dimension as
⎡ ⎤
1 a 0
where CnD∗ is an nD Cat matrix generated by the method ⎣ b ab + 1 0 ⎦
∗ and |CnD U =
C3-D (6)
∗ | = 1, Y(t) = [y1 (t), y2 (t), . . . , yn (t)] , and
T
c d 1
Y(t + 1) = [y1 (t + 1), y2 (t + 1), . . . , yn (t + 1)] are the input
T

and output of the nD Cat map. N is the number of finite states. where parameters a, b, c, and d are integers.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WU et al.: nD DISCRETE CAT MAP GENERATION USING LAPLACE EXPANSIONS 3

(n−1)D (n−1)D
Recently, Pan and Li [39] also introduced their parametric where 0 denotes the zero matrix and CN , CN ,
I II
3-D Cat matrix adding two control parameters c and d, but (n−1)D (n−1)D (n−1)D
CN , and CN are four submatrices of CN par-
III IV
in a formulation different from Lian et al.’s [37] in (4). In titioned with respect to the pth row and pth column, that is
particular, Pan’s 3-D Cat matrix C3-D
P is defined as follows:
⎡ ⎤
(n−1)D (n−1)D
⎡ ⎤ (n−1)D
CNI C NII
1 a c CN = ⎣ (n−1)D (n−1)D
⎦.
⎣ b ab + 1 CN CN
P =
C3-D bc ⎦ (7) III IV
d abcd cd + 1 Since all |Bp | = 1, then |CnD
N | = 1 and CnD
N is indeed an nD
Cat matrix.
where a, b, c, and d are integer control parameters.
D. Discussion
C. nD Cat Map Generation Methods When given the number of finite states, the trajectory of the
Besides the efforts on generating 3-D Cat matrices, high- Arnold’s Cat map can be predicted. The parameter space of the
dimensional Cat map generation methods have also been parametric Arnold’s Cat matrix C2-D para. is obviously too small
explored. for cryptography applications. The 3-D Cat matrices gener-
Falcioni et al. [40] proposed their nD Cat matrix CnD F by ated by above-mentioned methods have the largest parameter
replacing parameters in C2-D
para. with matrix ones as defined space of M 6 (M is the number of possible values of each
  matrix parameter) attained by C3-D C . Taking M = 256 as an
Im×m Pa example, then (256) = 2  2 , which is far smaller than
6 48 128
CnD = (8)
F Pb Pb Pa + Im×m the required size to resist brute-force attacks [43]. This implies
where I is the mth order identity matrix with 1 ≤ m ≤ n, and that applications directly using these parametric 3-D Cat matri-
Pa and Pb are arbitrary parameter matrices of size m × (n − m) ces are insecure. Consequently, these 3-D Cat matrices are
and (n − m) × m, respectively. It is easy to verify that commonly used with additional components [24].
  The parametric nD Cat matrices have many matrix elements.
 nD 
C  = |I| · (Pb Pa + I) − Pb I−1 Pa  However, because of many matrix multiplications involved in
F
their generation methods, their computation complexity is high
= |I| · |I| and their matrix elements are strongly correlated with each
= 1. (9) other. Correlated elements are difficultly coprime even with
coprime parameters. For a Cat map system, coprime elements
When m = 1 and n = 2, this nD Cat matrix becomes the often result in a long period of outputs [28], which is desired
parametric 2-D Cat matrix C2-D para. in (3). for many applications [44]. Also, when used in cryptography
Tang and Tang [41] proposed an nD Cat map gen- applications, correlated elements might reduce the actual key
eration method by coupling 2-D dynamics developed by space of the security system and thus it is vulnerable for the
Chen et al. [24]. It firstly couples dynamics in any two of the brute-force attacks [40]. This is because with correlated ele-
nDs, and then mixes all these coupled matrices together as ments, the actual number of distinctive Cat matrices that can
be generated could be less than the parameter space where
T = CT1,2 CT1,3 · · · CT1,n CT2,3 · · · CTn−1,n
CnD nD nD nD nD nD
(10)
parameters are commonly used as keys.
where each coupled Cat matrix CnD Tp,q (q > p) is an identical III. P ROPOSED M ETHOD
matrix except for elements CTp,q (p, q) = apq , CnD
nD
Tp,q (q, p) = This section presents the proposed nD Cat map generation
bpq and CnD (q, q) = a b
pq pq + 1. In total, CnD is a method in detail. First, a theorem for the nD Cat matrix is
Tp,q
n  T
mixed matrix of 2 nD pairwise-coupled Cat matrices. Since introduced. Based on this theorem, we then give a lemma
Tp,q | = 1, then |CT | = 1.
|CnD nD
of extending a Cat matrix from (n − 1)D to nD. Finally, we
Starting with C2-D N = Cpara. , Nance [42] recursively con-
2-D propose our method to generate nD Cat matrices.
structed his nD Cat matrix CnD N using matrix union, which is
defines as A. Theorem for nD Cat Matrix
An nD Cat matrix is an n × n square matrix with the

n
constraint that its determinant is 1. First of all, we intro-
N = B1 B2 · · · Bn =
CnD Bp (11)
p=1 duce Theorem 1 that allows us to construct an nD Cat matrix
conditionally.
where the pth basis Cat matrix Bp (1 ≤ p ≤ n) is obtained by Theorem 1: For an n × n square matrix X, if an element Xij
(n−1)D
expanding CN with respect to the pth diagonal elements as with a cofactor AXij = 1 satisfies the relation
⎡ (n−1)D ⎤ ⎛ ⎞
CN 0p×1 CN
(n−1)D n
⎢ I II
⎥ Xij = (−1)i+j ⎝1 − Xkj (−1)k+j AXkj ⎠ (12)
⎢ 01×(n−p−1) ⎥
Bp = ⎢ 01×p 1 ⎥ k=1,k=i
⎣ ⎦
(n−1)D (n−1)D for all other elements and their cofactors, then |X| = 1 and X
CN 0(n−p−1)×1 CN
III IV is an nD Cat matrix.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CYBERNETICS

Appendix A provides a detailed proof of Theorem 1. From

Theorem 1, if an element in an n × n square matrix has a
cofactor of value 1 and the relation (12) is satisfied, the deter-
minant of this square matrix is constant 1 and thus it is an nD
Cat matrix.

B. Extending (n − 1)D Cat Matrix to nD

From Theorem 1, it can be known that when some depen-
dency relationships are satisfied, an nD Cat matrix can be
constructed. Lemma 1 gives a definition of extending an
(n − 1)D Cat matrix into an nD one.
Lemma 1: If an n × n matrix X extended from an (n − 1)D
Cat matrix C(n−1)D as the submatrix associated with the ele-
ment Xij , and its elements and submatrices satisfy the relation
in (12), the n × n matrix X is an nD Cat matrix.
Proof: Since C(n−1)D is the submatrix associated with
element Xij , then
 
 
AXij = C(n−1)D  = 1.

Because (12) is also satisfied, X is then an nD Cat matrix as Fig. 1. Extending an (n − 1)D Cat matrix to nD. (a) (n − 1)D Cat
matrix C(n−1)D . (b) Extending matrix with respect to the element located
Theorem 1 states. nD .
at (i, j). (c) Adding new elements on the ith row or the jth column except Ci,j
Lemma 1 indicates that an nD Cat matrix CnD can be gen- (d) Computing the element Ci,j nD using (12) and generating new nD
erated in a parametric way, where the parameters consist of a Cat matrix CnD .
given (n − 1)D Cat matrix C(n−1)D , the location of the ele-
ment Xij associated with C(n−1)D , and 2(n − 1) additional
In the rest of this paper, when generating an nD Cat matrix
matrix elements. Fig. 1 illustrates the procedures of extending
using our proposed method, CnD ours∗ indicates that all parame-
an (n − 1)D Cat matrix to nD. As can be seen, two ran-
ters in I and J are restricted to be constant 2 (Fig. 2) and all
dom spatial parameters (i, j) divide the (n − 1)D Cat matrix
(n−1)D (n−1)D (n−1)D parameters in P and Q are randomly determined, while CnD ours
[Fig. 1(a)] into four parts: 1) CUL ; 2) CUR ; 3) CLL ;
(n−1)D denotes that all parameters in I, J, P, and Q are randomly
and 4) CLR , and then add new elements on the ith row and
determined.
jth column expect CnD i,j , whose value is computed using (12).
The generated nD Cat matrix is shown in Fig. 1(d).
IV. M ETHOD D ISCUSSION
Because most parameters to generate an nD Cat matrix
C. Proposed nD Cat Map Generation Method using Laplace expansions are highly independent, users have
Suppose 1-D Cat matrix C1-D = [1]1 , any Cat matrix with great flexibility to select different parameters to generate a
a desired dimension can be constructed using Lemma 1 and large number of nD Cat matrices. This section discusses
C1-D . The proposed nD Cat matrix generation method using the parameter settings and proposes a computation efficiency
Laplace expansions is defined as follows: improvement method for generating nD Cat matrices with
Laplace expansions.
Ours = L(I, J, P, Q)
CnD (13)

where I = {i1 , i2 , . . . , in−1 } and J = { j1 , j2 , . . . , jn−1 } are A. Parameter Settings

two integer lists and ik , jk ∈ [1, k + 1] for all k ∈ [1, n − 1]; In the proposed nD Cat map generation method in
P = {P1 , P2 , . . . , Pn−1 } and Q = {Q1 , Q2 , . . . , Qn−1 } are two Algorithm 1, the parameters in I and J control the locations of
vector lists and Pk and Qk are two k-length vectors for all the dependent elements and they are called the spatial config-
k ∈ [1, n − 1]. Algorithm 1 describes the detailed procedures uration parameters (SCPs), and the parameters in P and Q are
of our proposed nD Cat map generation method. Fig. 2 shows newly added independent matrix elements in each iteration of
an example of generating a 6-D Cat matrix. All dependent the matrix expansion and they are called matrix entries param-
elements in each Cat matrix can be determined via (12). For eters (MEPs). Two SCPs are needed to generate an nD Cat
example, C2-D
2,2 in Fig. 2(b) can be calculated by (14) and C2,2
3-D matrix CnD from C(n−1)D . Each SCP is an integer within range
in Fig. 2(c) can be calculate by (15) of [1, n] and thus has n possible choices [Fig. 1(b)]. Therefore,
when generating CnD from C1-D , the total number of SCPs
2-D
C2,2 = ab + 1 (14) NSCP and different choices CSCP can be calculated as
3-D
C2,2 = abcd + cd + ef − bce − adf + 1. (15) NSCP = 2(n − 1)
1 We just use C1-D as a dummy case for generating high-dimensional Cat
CSCP = n2 × (n − 1)2 × · · · × 22
matrices, because it involves no dynamics mixing. = (n!)2 .
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WU et al.: nD DISCRETE CAT MAP GENERATION USING LAPLACE EXPANSIONS 5

Fig. 2. Example of generating a 6-D Cat matrix. (a) 1-D Cat matrix C1-D (dummy). (b) 2-D Cat matrix C2-D . (c) 3-D Cat matrix C3-D . (d) 4-D Cat
matrix C4-D . (e) 5-D Cat matrix C5-D . (f) 6-D Cat matrix C6-D .

Ours = L(I, J, P, Q)
Algorithm 1 Proposed nD Cat Map Generation Method CnD
Input: I = {i1 , i2 , . . . , in−1 } is an vector with ik ∈ [1, k + 1] for all k ∈ [1, n − 1]
Input: J = { j1 , j2 , . . . , jn−1 } is an vector with jk ∈ [1, k + 1] for all k ∈ [1, n − 1]
Input: P = {P1 , P2 , . . . , Pn−1 } is a vector list and Pk is a k-length vector for all k ∈ [1, n − 1]
Input: Q = {Q1 , Q2 , . . . , Qn−1 } is a vector list and Qk is a k-length vector for all k ∈ [1, n − 1]
1: C1D = [1];
2: for k = 1 to n − 1 do
3: Extend CkD into a (k + 1) × (k + 1) matrix C(k+1)D by placing its partitions CkD kD kD kD
UL , CUR , CLL and CLR with respect to
the associated element location (ik , jk );
4: Set elements in the ik -th row and jk -th column in the extended matrix with respect to the parameter vectors Pk and Qk
as follows  
(k+1)D Pkm if m < jk (k+1)D Qkm if m < ik
Cik ,m = and C m,jk = ;
Pm−1 if m > jk
k Qm−1 if m > ik
k

(k+1)D
5: Calculate element Cik ,jk using (12).
6:end for  
Ensure: nD Cat matrix CnD with CnD  = 1

TABLE I
When extending an (n − 1)D Cat matrix C(n−1)D to CnD , PARAMETER S PACES OF nD C AT M APS G ENERATED
2(n − 1) independent matrix elements are needed [Fig. 1(c)]. U SING L APLACE E XPANSIONS
Thus, the total number of MEPs can be computed as
NMEP = 2(n − 1) + 2(n − 2) + · · · + 2
= n2 − n.
If we assume that all MEPs are selected from a symbol set with
M possible values, then the total number of different choices
to form (n2 − n) MEPs is CMEP = M n −n . Finally, due to
2

the independency of SCPs and MEPs, the whole parameter (Q is an orthogonal matrix and R is an upper triangular
space is simply the production of their individual parameter matrix) [45]. Thus, computing the unknown element by (12)
spaces, namely CS = CSCP CMEP , which can be seen in costs the computation complexity of order O(n4 ). Extending
Table I. Thus users have great flexility to choose different an (n − 1)D Cat matrix to nD also requires the computation
parameter settings to generate a large number of different nD complexity of order O(n4 ) because it needs to compute only
Cat matrices. the unknown element in (12). Therefore, generating an nD Cat
matrix from the 1-D Cat matrix C1-D requires the computa-
B. Computation Efficiency Improvement tion complexity of order O(n5 ). In order to further improve
Computing the unknown element in (12) requires to per- the computation efficiency, we introduce Theorem 2.
form n multiplications and n additions and to evaluate the Theorem 2: In the context of (12), Xij can be alternatively
determinants of (n − 1) matrices with size of (n − 1) × computed as
(n − 1). Calculating the determinant of an n × n matrix
Xij = (−1)i+j (1 − |G|) (16)
requires the computation complexity of order O(n3 ) when
using some methods like the Gaussian elimination, lower- with the aid of an auxiliary matrix G, whose element Gij = 0
upper (LU) matrix decomposition, or QR decomposition and all other elements are identical to those in X.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CYBERNETICS

Appendix B provides the mathematical proof of Theorem 2.

As a result, computing the unknown element in an extended
Cat matrix using (16) instead of (12) only requires to compute
two additions and the determinant of an n × n matrix. Its com-
putation complexity is of order O(n3 ). Thus, generating an nD
Cat matrix needs the computation complexity of order O(n4 ).
In this way, we can successfully reduce the computation
complexity of the proposed method for one magnitude.

C. Chaotic Behavior Analysis

The “chaos” phenomenon is difficult to exactly define and
qualitatively measure since everyone has a different viewpoint
about what the chaos is. The Lyapunov exponent (LE) [46] is
to describe the exponential divergence of two close trajecto- Fig. 3. Mean lgtLE values of nD Cat maps generated by the proposed
method.
ries. It can be used to measure whether a dynamical system
has chaotic behaviors. With a positive LE value, the difference TABLE II
of two close trajectories of a dynamical system will exponen- PARAMETER S PACES OF D IFFERENT nD
tially increase, and ultimately make the two trajectories totally C AT M AP G ENERATION M ETHODS
different. Thus, a dynamical system with a positive LE value
can be considered as chaotic. For high dimension dynamical
systems, they have more than one LE value and the largest
LE (lgtLE) value is an indictor of its chaotic behaviors.
An nD discrete dynamical system can be defined by
⎛ 1⎞
A
⎜ A2 ⎟
X(t + 1) = F(X(t)) = ⎜ ⎟
⎝ · · · ⎠X(t) (17)
A n

where X(t) = [x1 (t), x2 (t), . . . , xn (t)]T and X(t + 1) =

[x1 (t +1), x2 (t +1), . . . , xn (t +1)]T are the input and output. If
proposed method can generate nD Cat maps with complex
Ai (i = 1, 2, . . . , n) are differentiable functions, the Jacobian
chaotic behaviors.
matrix J of the system F(·) can be defined as follows:
⎛∂ 1 ∂ 1 ∂A 1 ⎞
A A
···
⎜ ∂x 1 ∂ ∂x n ⎟ V. P ERFORMANCE E VALUATIONS AND C OMPARISONS
⎜ ∂ 2 ∂ x22 ∂ ⎟
⎜ A A
··· A2 ⎟ In this section, the performance of the proposed nD Cat map
⎜ ⎟
J = ⎜ ∂x 1 ∂x 2 ∂xn ⎟. (18) generation method is compared with several other methods.
⎜ ⎟
⎜ ··· ··· ··· ··· ⎟
⎝ ∂ A n ∂A n ∂A n ⎠
··· A. Parameter Space
∂x 1 ∂x 2 ∂x n
Therefore, in each iteration, J(t) corresponds to the itera- The computation precision of implementing a chaotic map
tion output X(t). If J(t) has m eigenvalues, given by μi (t), in software or hardware is finite, implying that the chaotic
the n LE values of the system F(·) are defined as follows: complexity of a chaotic map in real numbers will inevitably
 k  degrade [47]. Consequently, the parameter space of an nD
1 Cat map generation method is always limited. Moreover, the
λj = lim ln μj (t) (19) parameter space is also closely related to the total number of
k→∞ k
t=1 distinctive Cat matrices, which is known as the security key
where j = 1, 2, . . . , n. Chaotic behaviors of the system F(·) space in cryptography applications [24], [37]. Therefore, an nD
is demonstrated by the lgtLE value among λj . Cat map generation method with sufficiently large parameter
To analyze the chaotic behaviors of the nD Cat maps gen- space is desired to against the brute-force attacks.
erated by the proposed method, we perform the following Table II shows the parameter spaces of different nD Cat
experiment: for our nD Cat map generation method, we set map generation methods. It is clear that the proposed method
M = 2 (MPEs are restricted to {0, 1}) and randomly generate has the largest parameter space in all listed methods. This
5×2n nD Cat maps for n ∈ {3, 4, . . . , 10}. The lgtLE values of is because, compared with other methods that use only the
these nD Cat maps are calculated. Fig. 3 plots the mean lgtLE matrix entries as the parameters, the proposed method intro-
values for different dimensions n. As can be seen, the mean duces parameters to control the spatial configurations of an nD
lgtLE values of nD Cat maps are all positive and they become Cat matrix. These spatial parameters can significantly increase
bigger when the dimension n increases. This means that the the parameter space of an nD Cat matrix.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WU et al.: nD DISCRETE CAT MAP GENERATION USING LAPLACE EXPANSIONS 7

TABLE III
A LGORITHM C OMPLEXITY OF D IFFERENT nD C AT M AP G ENERATION M ETHODS

Fig. 4. Matrix element correlation analysis of different nD Cat map generation methods.

∗1 , C∗2 , . . . , C∗1000 for M = 4 (MEPs are restrict to

CnD
B. Algorithm Complexity nD nD

Because of the high complexity of generating a multidi- {0, 1, 2, 3}). Suppose p and q are two different indexes
mensional Cat matrix, the computation and memory costs (1 ≤ p, q ≤ n2 ), the correlation coefficient (CE) between pth
are quite important factors to evaluate whether an nD Cat and qth matrix elements in CnD ∗ is defined as follows:
map generation method is suitable for the applications.   
CnD
E Sp − μSp Sq − μSq
Table III lists the algorithm complexity of different nD Cat CE (p, q) =
∗ (20)
map generation methods. As can be seen, for generating σSp σSq
an nD Cat matrix, the proposed method requires n2 mem- where μ and σ are the mean value and standard deviation and
ory cells and its computation complexity is O(n4 ), which Sp and Sq are sequences of matrix elements in the form of
is one magnitude less than the Tang and Tang’s [41] and  
Nance’s [42] methods. Sp = CnD∗1 [p], . . . , C∗1000 [p]
nD
 
Sq = CnD∗1 [q], . . . , C∗1000 [q] .
nD
(21)

C. Matrix Element Correlation They contain the pth and qth elements in these 1000 ran-
nD
In the process of the nD Cat matrix generation, the set- domly generated matrices, respectively. Thus, CEC∗ (p, q) is
tings of matrix elements are totally decided by the generation the correlation coefficient between the sequences Sp and Sq .
method. The Cat matrices generated by different methods have The CE value closing to 0 means low correlation between
different degrees of inner correlations between their matrix Sp and Sq while closing to 1 means high correlation.
elements. As discussed in Section II-D, an nD Cat matrix Fig. 4 plots the magnitudes of these correlation coeffi-
with low inner correlation is commonly preferred for many cients for 3-D, 4-D, 5-D, and 6-D Cat matrices generated
applications. by various methods. As can be seen for each color matrix,
To analyze the correlations of matrix elements in nD Cat the values in the pth row denote the correlation coefficients
matrices generated by different methods, we perform the between the pth matrix element and all the n2 matrix ele-
nD nD nD
following correlation analysis: for each nD Cat map gen- ments, namely CEC∗ (p, 1), CEC∗ (p, 2), . . . , CEC∗ (p, n2 ),
nD nD nD
eration method ∗, we randomly generate 1000 Cat matrices because CEC∗ (p, q) = CEC∗ (q, p) and CEC∗ (p, p) = 1,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CYBERNETICS

TABLE IV
S TATISTICS OF nD C AT M ATRIX E LEMENTS T HAT C OULD BE
C ONTROLLED BY D IFFERENT N UMBERS OF MEP S IN
D IFFERENT M ETHODS

(a) (b)

all magnitudes are on the main diagonal symmetry and the Fig. 5. Shannon entropy values of nD Cat maps for N = 2. (a) Mean of
magnitudes on the main diagonal line are 1, marked as red. Shannon entropy values of nD Cat maps versus map dimension n. (b) Sorted
Shannon entropy scores of 10-D Cat maps for each generation method.
It is easy to see that CnD ours∗ generated with specific spatial
configurations have patterns in their element correlation plots,
while CnDours generated with randomly determined spatial con- Hmin = 0 implies that the system is completely predictable
figurations have CE values close to 0. Besides, it is noticeable and Hmax = n log2 N when if and only if all observations
that multidimensional Cat matrices generated by the proposed absolutely uniformly distribute over the N n states.
method have smaller CE values, and thus have less correlated Two experiment groups of Shannon entropy analysis are
matrix elements than those of other methods. performed to evaluate the randomness of various nD Cat
All these methods have their own characteristics in gen- map generation methods. The first group of experiments is
erating multidimensional Cat matrices. For example, if the to analyze the Shannon entropy values against the Cat map
generated multidimensional Cat matrices always contain a con- dimension n by fixing N = 2. For each Cat map generation
stant element in a fixed position, the correlation between this method, we randomly generate 5N n nD Cat maps and com-
element and any other elements is always 0. On the other pute the Shannon entropy values of these nD Cat maps for
hand, the more elements controlled by a single MEP, the n ∈ {3, 4, . . . , 10}. The observation time series is of length
less correlation between two Cat matrix elements. For the 2n+1 with respect to the initial state Y(0) = 1n×1 .
multidimensional Cat matrices generated by various methods, The mean Shannon entropy results of different methods are
Table IV lists the statistics of their elements that could be con- given in Fig. 5(a). As can be seen, the values of all gener-
trolled by different numbers of MEPs. It is clear that CnD ours∗ ation methods increase as the dimension n increases except
generated by the proposed method with fixed spatial param- for the Cat maps generated by the Nance’s [42] method.
eters has the largest number of elements controlled by one The low randomness of the Nance’s method is caused by
MEP, implying that most of matrix elements in the nD Cat its small parameter space and high matrix element correla-
matrices are independent. Considering spatial parameters, no tions, which can be seen in Table II and Fig. 4, respectively.
element in CnDours is a constant and all of them are determined The nD Cat maps generated by the proposed methods achieve
in a more complicated way involving both SCPs and MEPs. high Shannon entropy values for all test dimensions. Fig. 5(b)
Consequently, for the Cat matrices generated by the proposed shows the Shannon entropy values of 10-D Cat maps gen-
method, any two of their elements are uncorrelated as shown erated by different methods in the ascending order. Our
in Fig. 4. methods CnD nD
ours∗ and Cours have bigger Shannon entropy val-
ues than other methods in the first 4000 10-D Cat maps.
D. Shannon Entropy Tang and Tang’s [41] method can achieve the biggest Shanon
Due to the finite precision of digital devices, the nD Cat entropy values when the number of 10-D Cat maps is over
maps are commonly used in discrete forms in real-world 4000. At last, our proposed and Tang and Tang’s [41] meth-
applications as shown in (1). In (1), Y(t) and Y(t + 1) are ods can achieve the same biggest Shannon entropy values.
observations of the outputs of the nD Cat map with respect to Fig. 5(b) also shows that the outputs of 10-D Cat maps
time t and t + 1. Each observation has n channels and each generated by our methods CnD nD
ours∗ and Cours have similar ran-
channel has N states. Thus observation Y(t) has the change to domness. This is because their MEPs are randomly selected
attain a finite number N n of states. from the same data set. Table V lists the maximum Shannon
The randomness of the discrete nD Cat map can be entropy values for nD Cat maps generated by different meth-
measured by the Shannon entropy [17] defined as follows: ods with various dimensions. It is noticeable that the proposed
n
and Tang and Tang’s [41] methods are able to generate nD

N
Cat maps with high Shannon entropy values that are quite
H=− Pr(k) log2 Pr(k) (22) close to the theoretical maximum Shannon entropy values
k=1
Hmax = n log2 N = n.
where Pr(k) denotes the probability of seeing an observation The second group of experiments is to analyze the Shannon
Y(t) in the kth state. Then we have 0 ≤ H ≤ n log2 N and entropy values against the state number N by fixing the dimen-
a bigger Shannon entropy value indicates better randomness. sion n = 3. The state number N = {2, 3, . . . , 10} and for each
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WU et al.: nD DISCRETE CAT MAP GENERATION USING LAPLACE EXPANSIONS 9

TABLE V
M AXIMUM S HANNON E NTROPY VALUES OF nD C AT M APS
G ENERATED BY D IFFERENT M ETHODS W ITH N = 2

Fig. 7. Mean Kolmogorov entropy values of nD Cat maps generated by

different methods.

numbers N. As can be seen, the proposed method outper-

forms all other listed generation methods. Its generated 3-D
Cat maps can achieve the highest Shannon entropy values for
N = {2, 3, . . . , 10} that are extremely close to the theoreti-
cal maximum scores. These further verify that the Cat maps
generated by the proposed method are more stable to generate
(a) (b) random-like outputs.

Fig. 6. Shannon entropy of 3-D Cat maps. (a) Mean of Shannon entropy E. Kolmogorov Entropy
values of 3-D Cat maps versus the number of channel states N. (b) Sorted
Shannon entropy scores of 3-D Cat maps with N = 10 for each generation The Kolmogorov entropy [49] can measure whether extra
method. information is needed to predict the trajectory of a dynamical
system, which is defined as follows:
TABLE VI
M AXIMUM S HANNON E NTROPY VALUES OF 3-D C AT M APS KE = lim τ −1 lim lim Km,τ (ε) (23)
G ENERATED BY D IFFERENT M ETHODS τ →0 ε→0 m→∞
where m is the embedding dimension, Km,τ (ε) is defined by

Km,τ (ε) = − p(i1 , . . . , im ) log p(i1 , . . . , im ). (24)
i1 ,...,im ≤n(ε)

Divide the phase plane into m nonoverlapping partitions

φi1 , . . . , φim , then p(i1 , . . . , im ) is the probability of correctly
predicted trajectory in partition φi1 at time τ , in partition φi2 at
time 2τ, . . . , in partition φim at time mτ . Positive Kolmogorov
entropy value indicates chaotic behaviors of the dynamical
system and a bigger value means better unpredictability.
For different nD Cat map generation methods, we inves-
tigate their Kolmogorov entropy values against the Cat map
dimension n. When generating nD Cat matrices, the number
3-D Cat map generation method, randomly generating 5N 3 of possible values is set as M = 4 (MPEs are restricted to
3-D Cat maps and computing their Shannon entropy values {0, 1, 2, 3}). For each generation method with different dimen-
using an observed time series of length N 4 with respect to sions n, we randomly generate 10n Cat maps and then obtain
initial state Y(0) = 13×1 . The mean Shannon entropy values trajectories of these nD Cat maps with the number of finite
of different methods are shown in Fig. 6(a). It is clear that on states N = 4 and initial value Y(0) = 1n×1 . For each iteration
average the 3-D Cat maps generated by the proposed method output Y(t) = [y1 (t), y2 (t), . . . , yn (t)]T , we convert it into float
have higher mean Shannon entropy values than other meth- number within the range of [0, 1] by
ods. Fig. 6(b) shows the Shannon entropy values of 3-D Cat n
maps generated by different methods in the ascending order 4i−1 yi (t)
yt = i=1 n . (25)
with N = 10. Compared with other listed generation methods, 4
the proposed method has the fastest rate to achieve its highest The generated trajectories are with length of 12 000 and
Shannon entropy values. we use the method proposed in [50] to calculate their
Table VI lists the maximum Shannon entropy values of 3-D Kolmogorov entropy values. Fig. 7 plots the mean values of
Cat maps generated by different methods with different state Kolmogorov entropy results of different generation methods.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON CYBERNETICS

TABLE VII
M AXIMUM KOLMOGOROV E NTROPY VALUES OF nD C AT M APS
G ENERATED BY D IFFERENT M ETHODS

As can be seen, the nD Cat maps generated by the proposed Fig. 8. Diehard statistical test results.
methods can achieve positive mean Kolmogorov entropy val-
ues and the values become bigger as the dimension n increases. TABLE VIII
For other methods, their randomly generated nD Cat maps T EST U01 R ESULTS FOR B INARY S EQUENCES
W ITH D IFFERENT L ENGTHS
achieve quite small mean Kolmogorov entropy values that are
close to 0. This is because the trajectories of the discrete Cat
map are periodic [29] and its period is small when the ele-
ments of its Cat matrix are highly correlated. Table VII lists
the maximum Kolmogorov entropy values of nD Cat maps
generated by different methods with various dimensions n.
It is obvious that the proposed method is able to generate nD
Cat maps with bigger Kolmogorov entropy values than other
TABLE IX
methods. P ERFORMANCE C OMPARISONS OF D IFFERENT nD
C AT M AP G ENERATION M ETHODS
VI. P ROPOSED P SEUDORANDOM N UMBER G ENERATOR
To show the effectiveness of the nD Cat maps generated by
our proposed method, as an example, this section introduces
a PRNG using the 5-D Cat maps generated by the proposed
method.
Here, we set N = 1 and the initial value Y(0) as a float
number in (1). Given I, J, P, and Q, we use (13) to produce
a 5-D Cat matrix C5-Dours . Suppose {Y(t)|t = 1, 2, . . .} is the
output of C5-D
ours with the initial value Y(0). Y(t) is with size The sorted 234 p-values are all distributed on or nearby the
of 5 × 1. The proposed PRNG is defined as follows: uniform distribution line (the red line in Fig. 8). This means
 
T(t) = Bin Y(t) (26) that the generated random number sequence can pass the
31:32 Diehard statistical test.
where Bin[α]31:32 is a function to convert the float number α
into a 52-bit binary stream using the IEEE 754 standard [51]
and then fetch its 31st and 32nd digital numbers. B. TestU01 Test
TestU01 is a random number test package that provides con-
A. Diehard Statistical Test vinced empirical description for the quality of the PRNG [53].
Diehard statistical test suit [52] is a battery of statistical It contains six test batteries and three of them (Rabbits,
tests that is widely used for measuring the quality of a PRNG. Alphabit, and BlockAlphabit) are used to test the random-
It contains 15 subtests and shows good performance for data ness of binary sequences. Rabbits applies 38 subtests while
sequences with a large size. These 15 empirical subtests are Alphabit and BlockAlphabit apply 17 subtests. Each subtest
designed to find the nonrandomness areas in a large-size data will generate a p-value and the random number sequence is
sequence and total 234 p-values are generated. For a ran- considered to pass the subtest if the generated p-value is within
dom number sequence with high quality, these p-values are the range of [0.001, 0.999].
expected to randomly distribute to pass the Diehard statistical In our experiment, binary sequences with different lengths
test. If six or more p-values with 0 or 1 are obtained, a data are generated by the proposed PRNG and tested by Rabbits,
sequence is considered to fail the test. The size of the test Alphabit, and BlockAlphabit. The test results are shown in
sequence is suggested as 11 468 800 byte. Table VIII. As can be seen, binary sequences with different
Fig. 8 plots the Diehard statistical test results of a ran- lengths can pass almost all the subtests except for one subtest
dom number sequence generated by the proposed PRNG. in Rabbits.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WU et al.: nD DISCRETE CAT MAP GENERATION USING LAPLACE EXPANSIONS 11

VII. C ONCLUSION Proof: The determinant of matrix G can be expanded as

In this paper, we have introduced an efficient method of 
n
generating nD Cat maps using Laplace expansions. It can |G| = Gpq (−1)p+q AG
ij
iteratively generate any dimension Cat matrix from a 1-D p=1
Cat matrix. Unlike existing nD Cat map generation methods, 
n
which only consider MEPs to control the matrix elements, the = Gij (−1)i+j AG
ij + Gkj (−1)k+j AG
kj .
proposed method also introduces SCPs to control the spatial k=1,k=i
configurations.
Theoretical analysis and experiment evaluations have been Because Gij = 0 and for arbitrary elements Gpq in auxiliary
done to show that the proposed nD Cat map generation method matrix G with either p = i or q = j, we have Gpq = Xpq ,
can combine the dynamics of different dimensions and gen- the above equation can be rewrote as
erate an nD Cat matrix with (n2 − n) independent matrix 
n
elements. The performance of the proposed method has been |G| = 0(−1)i+j AG
ij + Xkj (−1)k+j AXkj
compared with other nD Cat map generation methods and the k=1,k=i
comparison results are shown in Table IX. As can be seen, 
n
the proposed method can achieve the best performance in the = Xkj (−1)k+j AXkj .
parameter space, matrix element correlation, Shannon entropy, k=1,k=i
Kolmogorov entropy, and the second-best performance in algo-
Substitute this equation to (12), we then obtain (16).
rithm complexity. To investigate its real-world applications, we
have proposed a new PRNG using the 5-D Cat maps gener-
ated by the proposed method. Our future work will explore its R EFERENCES
applications in secure communication and data encryption. [1] M. R. Frank, L. Mitchell, P. S. Dodds, and C. M. Danforth, “Standing
swells surveyed showing surprisingly stable solutions for the Lorenz’96
model,” Int. J. Bifurcat. Chaos, vol. 24, no. 10, pp. 1430027–14,
ACKNOWLEDGMENT 2014.
[2] Y. Zhou, Z. Hua, C. M. Pun, and C. L. P. Chen, “Cascade chaotic system
The authors would like to thank the anonymous reviewers with applications,” IEEE Trans. Cybern., vol. 45, no. 9, pp. 2001–2012,
for their valued comments and suggestions, which greatly help Sep. 2015.
us to improve the quality of this paper. [3] S. Y. Li, L. M. Tam, S. E. Tsai, and Z. M. Ge, “Novel fuzzy modeling
and synchronization of chaotic systems with multinonlinear terms by
advanced Ge-Li fuzzy model,” IEEE Trans. Cybern., to be published,
A PPENDIX A 2015
[4] H.-G. Chou, C.-F. Chuang, W.-J. Wang, and J.-C. Lin, “A fuzzy-
P ROOF OF T HEOREM 1 model-based chaotic synchronization and its implementation on a secure
communication system,” IEEE Trans. Inf. Forensic Security, vol. 8,
Here proves Theorem 1 in Section III-A. no. 12, pp. 2177–2185, Dec. 2013.
Proof: Compute the determinant of an n × n matrix X using [5] R. Law, D. J. Murrell, and U. Dieckmann, “Population growth in
Laplace expansions space and time: Spatial logistic equations,” Ecology, vol. 84, no. 1,
pp. 252–262, 2003.

n [6] N. Boccara, “Van der pol oscillator,” Essentials of Mathematica:
|X| = Xpq (−1)p+q AXpq With Applications to Mathematics and Physics. New York, NY,
USA: Springer, 2007, pp. 505–508.
p=1 [7] H.-K. Chen and C.-I. Lee, “Anti-control of chaos in rigid body motion,”

n Chaos Solitons Fractals, vol. 21, no. 4, pp. 957–965, 2004.
= Xij (−1)i+j AXij + Xkj (−1)k+j AXkj . [8] Z.-P. Wang and H.-N. Wu, “On fuzzy sampled-data control of
chaotic systems via a time-dependent Lyapunov functional approach,”
k=1,k=i IEEE Trans. Cybern., vol. 45, no. 4, pp. 819–829, Apr. 2015.
[9] S. H. Strogatz, Nonlinear Dynamics and Chaos: With Applications
Because AXij = 1 and X satisfies to Physics, Biology, Chemistry, and Engineering. Cambridge, MA,
⎛ ⎞ USA: Perseus, 2001.
[10] H.-J. Stöckmann, Quantum Chaos: An Introduction. New York, NY,
n
Xij = (−1)i+j ⎝1 − Xkj (−1)k+j AXkj ⎠. USA: Cambridge Univ. Press, 2006.
[11] M. Shen, W.-N. Chen, J. Zhang, H. S.-H. Chung, and O. Kaynak,
k=1,k=i “Optimal selection of parameters for nonuniform embedding of chaotic
time series using ant colony optimization,” IEEE Trans. Cybern., vol. 43,
Then no. 2, pp. 790–802, Apr. 2013.
[12] Y. Wu, Y. Zhou, and L. Bao, “Discrete wheel-switching chaotic system
|X| = Xij (−1)i+j AXij + 1 − Xij (−1)i+j and applications,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 61,
no. 12, pp. 3469–3477, Dec. 2014.
= Xij (−1)i+j 1 + 1 − Xij (−1)i+j = 1. [13] F. Pareschi, G. Setti, and R. Rovatti, “Implementation and testing of
high-speed CMOS true random number generators based on chaotic
systems,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 57, no. 12,
Therefore, X is an nD Cat matrix. pp. 3124–3137, Dec. 2010.
[14] Y. Zhou, L. Bao, and C. L. P. Chen, “Image encryption using a new
parametric switching chaotic system,” Signal Process., vol. 93, no. 11,
A PPENDIX B pp. 3039–3052, 2013.
P ROOF OF T HEOREM 2 [15] S.-L. Chen, T. Hwang, and W.-W. Lin, “Randomness enhancement using
digitalized modified logistic map,” IEEE Trans. Circuits Syst. II, Exp.
Here proves Theorem 2 in Section IV-B. Briefs, vol. 57, no. 12, pp. 996–1000, Dec. 2010.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON CYBERNETICS

[16] Z. Li, K. Li, C. Wen, and Y. C. Soh, “A new chaotic secure communi- [44] N. J. Corron and S. D. Pethel, “Control of long-period orbits and
cation system,” IEEE Trans. Commun., vol. 51, no. 8, pp. 1306–1312, arbitrary trajectories in chaotic systems using dynamic limiting,”
Aug. 2003. Chaos Interdiscipl. J. Nonlin. Sci., vol. 12, no. 1, pp. 1–7, 2002.
[17] Z. Hua, Y. Zhou, C.-M. Pun, and C. L. P. Chen, “2D sine logistic [45] J. A. Storer, An Introduction to Data Structures and Algorithms. Basel,
modulation map for image encryption,” Inf. Sci., vol. 297, pp. 80–94, Switzerland: Springer, 2001.
Mar. 2015. [46] A. Wolf, J. B. Swift, H. L. Swinney, and J. A. Vastano, “Determining
[18] S. Chen and H. Leung, “Ergodic chaotic parameter modulation with Lyapunov exponents from a time series,” Physica, vol. 16, no. 3,
application to digital image watermarking,” IEEE Trans. Image Process., pp. 285–317, 1985.
vol. 14, no. 10, pp. 1590–1602, Oct. 2005. [47] K. J. Persohn and R. J. Povinelli, “Analyzing logistic map pseudo-
[19] T. Habutsu, Y. Nishio, I. Sasase, and S. Mori, A Secret Key random number generators for periodicity induced by finite precision
Cryptosystem by Iterating a Chaotic Map. Berlin, Germany: Springer, floating-point representation,” Chaos Solitons Fractals, vol. 45, no. 3,
1991, pp. 127–140. pp. 238–245, 2012.
[20] J. Fridrich, “Symmetric ciphers based on two-dimensional chaotic [48] A. M. Davie and A. J. Stothers, “Improved bound for complexity of
maps,” Int. J. Bifurcat. Chaos, vol. 8, no. 6, pp. 1259–1284, 1998. matrix multiplication,” Proc. Roy. Soc. Edin. Sect. A Math., vol. 143,
[21] D. Xiao, X. F. Liao, and S. J. Deng, “One-way hash function con- no. 2, pp. 351–369, 2013.
struction based on the chaotic map with changeable-parameter,” Chaos [49] P. Faure and H. Korn, “A new method to estimate the Kolmogorov
Solitons Fractals, vol. 24, no. 1, pp. 65–71, 2005. entropy from recurrence plots: Its application to neuronal signals,”
[22] V. Patidar, K. K. Sud, and N. K. Pareek, “A pseudo random bit Phys. D Nonlin. Phenom., vol. 122, no. 1, pp. 265–279, 1998.
generator based on chaotic logistic map and its statistical testing,” [50] J. A. C. Gallas, “Structure of the parameter space of the Hénon map,”
Inform. Slovenia, vol. 33, no. 4, pp. 441–452, 2009. Phys. Rev. lett., vol. 70, no. 18, p. 2714, 1993.
[23] Y. Zhou, L. Bao, and C. L. P. Chen, “A new 1D chaotic system for [51] IEEE Standard for Floating-Point Arithmetic, IEEE Standard 754-2008,
image encryption,” Signal Process., vol. 97, pp. 172–182, Apr. 2014. 2008.
[24] G. Chen, Y. Mao, and C. K. Chui, “A symmetric image encryption [52] G. Marsaglia. Diehard Statistical Tests. [Online]. Available:
scheme based on 3D chaotic cat maps,” Chaos Solitons Fractals, vol. 21, http://www.stat.fsu.edu/pub/diehard/, accessed Sep. 22, 2015.
no. 3, pp. 749–761, 2004. [53] P. L’Ecuyer and R. Simard, “Testu01: A C library for empirical testing
[25] J. Ford, G. Mantica, and G. Ristow, “The Arnol’d cat: Failure of the of random number generators,” ACM Trans. Math. Softw., vol. 33, no. 4,
correspondence principle,” Phys. D Nonlin. Phenom., vol. 50, no. 3, p. 22, 2007.
pp. 493–520, 1991.
[26] L. Kocarev and S. Lian, Chaos-Based Cryptography: Theory, Algorithms
and Applications, vol. 354. Berlin, Germany: Springer, 2011.
[27] X. Tong and M. Cui, “Image encryption with compound chaotic Yue Wu (S’08–M’12) received the B.E. degree in
sequence cipher shifting dynamically,” Image Vis. Comput., vol. 26, telecommunication engineering from the Huazhong
no. 6, pp. 843–850, 2008. University of Sciences and Technologies, Hubei,
[28] G. Gaspari, “The Arnold cat map on prime lattices,” Phys. D Nonlin. China, in 2001, the M.Sc. degree in applied mathe-
Phenom., vol. 73, no. 4, pp. 352–372, 1994. matics from the University of Toledo, Toledo, OH,
[29] F. Chen, K.-W. Wong, X. Liao, and T. Xiang, “Period distribution of USA, in 2008, and the Ph.D. degree in electrical
the generalized discrete Arnold cat map for N = 2e ,” IEEE Trans. Inf. engineering from Tufts University, Medford, MA,
Theory, vol. 59, no. 5, pp. 3249–3255, May 2013. USA, in 2012.
[30] I. Antoniou, B. Qiao, and Z. Suchanecki, “Generalized spectral He was a Scientist with the Raytheon BBN
decomposition and intrinsic irreversibility of the Arnold Cat Map,” Technologies, Cambridge, U.K. He is currently a
Chaos Solitons Fractals, vol. 8, no. 1, pp. 77–90, 1997. Computer Scientist with the Information Sciences
[31] S. Weigert, “Quantum chaos in the configurational quantum cat map,” Institute, University of Southern California Viterbi School of Engineering.
Phys. Rev. A, vol. 48, no. 3, pp. 1780–1798, 1993. His current research interests include information security, image processing,
[32] A. Lakshminarayan and N. Balazs, “On the quantum cat and sawtooth pattern recognition, and their real applications.
maps—Return to generic behaviour,” Chaos Solitons Fractals, vol. 5,
no. 7, pp. 1169–1179, 1995.
[33] S. P. Kuznetsov, “Disheveled Arnold’s cat and the problem of quantum- Zhongyun Hua (S’14) received the B.S. degree
classic correspondence,” Phys. D Nonlin. Phenom., vol. 137, no. 3, in software engineering from Chongqing University,
pp. 205–227, 2000. Chongqing, China, in 2011, and the M.S. degree in
[34] S. V. Neshveyev, “On the K-property of quantized Arnold cat maps,” software engineering from the University of Macau,
J. Math. Phys., vol. 41, no. 4, pp. 1961–1965, 2000. Macau, China, in 2013, where he is currently pur-
[35] A. M. F. Rivas, M. Saraceno, and A. M. Ozorio de Almeida, suing the Ph.D. degree from the Department of
“Quantization of multidimensional cat maps,” Nonlinearity, vol. 13, Computer and Information Science.
no. 2, p. 341, 2000. His current research interests include chaos-based
[36] L. Barash and L. N. Shchur, “Periodic orbits of the ensemble applications, multimedia security, and signal/image
of Sinai-Arnold cat maps and pseudorandom number generation,” processing.
Phys. Rev. E, vol. 73, no. 3, 2006, Art. ID 036701.
[37] S. Lian, Y. Mao, and Z. Wang, “3D extensions of some 2D chaotic
maps and their usage in data encryption,” in Proc. 4th Int. Conf. Control
Autom., Montreal, QC, Canada, 2003, pp. 819–823. Yicong Zhou (M’07–SM’14) received the B.S.
[38] H. Liu, Z. Zhu, H. Jiang, and B. Wang, “A novel image encryption degree from Hunan University, Changsha, China,
algorithm based on improved 3D chaotic cat map,” in Proc. 9th Int. in 1992, and the M.S. and Ph.D. degrees from Tufts
Conf. Young Comput. Sci., Hunan, China, 2008, pp. 3016–3021. University, Medford, MA, USA, in 2008 and 2010,
[39] T. G. Pan and D. Y. Li, “A new algorithm of image encryption based on respectively, all in electrical engineering.
3D Arnold Cat,” in Adv. Eng. Forum, vol. 1, pp. 183–187, Sep. 2011. He is currently an Assistant Professor with
[40] M. Falcioni, L. Palatella, S. Pigolotti, and A. Vulpiani, “Properties the Department of Computer and Information
making a chaotic system a good pseudo random number generator,” Science, University of Macau, Macau, China. He
Phys. Rev. E, vol. 72, no. 1, 2005, Art. ID 016220. has authored/co-authored over 90 papers, including
[41] K. W. Tang and W. K. S. Tang, “A chaos-based secure voice commu- 14 IEEE T RANSACTION papers, six most down-
nication system,” in Proc. IEEE Int. Conf. Ind. Technol., Hong Kong, loaded/popular papers in corresponded journals, and
2005, pp. 571–576. one highly cited paper within the top 1% of published papers in the Institute
[42] J. Nance, “Periods of the discretized Arnold Cat Map for Scientific Information (ISI) database up to 2015. His current research
and its extension to n dimensions,” [Online]. Available: interests include chaotic systems, multimedia security, image processing and
http://arxiv.org/pdf/1111.2984v1.pdf, 2013. understanding, and machine learning.
[43] G. Alvarez and S. Li, “Some basic cryptographic requirements for Dr. Zhou was a recipient of the third price of the Macau Natural Science
chaos-based cryptosystems,” Int. J. Bifurcat. Chaos, vol. 16, no. 8, Award in 2014. He is a member of the International Society for Optical
pp. 2129–2151, 2006. Engineers and the Association for Computing Machinery.