You are on page 1of 10

Applied Mathematical Modelling 36 (2012) 4502–4511

Contents lists available at SciVerse ScienceDirect

Applied Mathematical Modelling


journal homepage: www.elsevier.com/locate/apm

Different Zhang functions leading to different Zhang-dynamics models


illustrated via time-varying reciprocal solving
Yunong Zhang ⇑, Fen Li, Yiwen Yang, Zhan Li
School of Information Science and Technology, Sun Yat-sen University, Guangzhou 510006, PR China

a r t i c l e i n f o a b s t r a c t

Article history: Along with neural dynamics (based on analog solvers) widely arising in scientific compu-
Received 8 April 2011 tation and optimization fields in recent decades which attracts extensive interest and
Received in revised form 18 November 2011 investigation of researchers, a novel type of neural dynamics, called Zhang dynamics
Accepted 27 November 2011
(ZD), has been formally proposed by Zhang et al. for the online solution of time-varying
Available online 3 December 2011
problems. By following Zhang et al.’s neural-dynamics design method, the ZD model, which
is based on an indefinite Zhang function (ZF), can guarantee the exponential convergence
Keywords:
performance for the online time-varying problems solving. In this paper, different indefi-
Zhang functions (ZFs)
Zhang dynamics (ZD) models
nite Zhang functions, which can lead to different ZD models, are proposed and developed
Exponential convergence as the error-monitoring functions for the time-varying reciprocal problem solving. Addi-
Time-varying reciprocal solving tionally, for the goal of developing the floating-point processors or coprocessors for the
MATLAB Simulink modeling future generation of computers, the MATLAB Simulink modeling and simulative verifica-
tions of such different ZD models are further presented for online time-varying reciprocal
solving. The modeling results substantiate the efficacy of such different ZD models for
time-varying reciprocal solving.
Ó 2011 Elsevier Inc. All rights reserved.

1. Introduction

In various scientific and engineering fields, many fundamental mathematical operations are encountered, such as linear/
nonlinear equations solving, linear/nonlinear inequalities solving, quadratic program and matrix inversion. In view of the
fundamental roles of these mathematical operations, a number of algorithms have been proposed and investigated to solve
these problems. However, most of these reported algorithms are theoretically/intrinsically for static (or to say, time-
invariant, constant) problems solving [1–7]. In [1], Kadalbajoo and Kumar devoted a numerical study of the boundary value
problems for nonlinear singularly perturbed differential-difference equations with small delay. Salahi [4] considered the
least squares solution of linear inequalities with uncertainty in the data, and proved the robust counterpart of the least
squares solution of linear inequalities equivalent to a second order conic linear program. Besides the mentioned algorithms,
there has also been a method of using dynamic systems to solve linear or nonlinear equations [8–11].
Generally speaking, for solving the aforementioned problems, there are two general types of solutions. One type is the
numerical algorithms performed on digital computers, which are usually of serial-processing nature and may not be efficient
enough for large-scale online or real-time applications due to the high computational complexity. As another important type
of the solutions, many parallel-processing computational methods have been proposed, analyzed, and implemented on spe-
cific architectures, e.g., the analog and neural-dynamic solvers which are closely related to the dynamic systems method.

⇑ Corresponding author.
E-mail addresses: zhynong@mail.sysu.edu.cn, ynzhang@ieee.org (Y. Zhang).

0307-904X/$ - see front matter Ó 2011 Elsevier Inc. All rights reserved.
doi:10.1016/j.apm.2011.11.081
Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511 4503

In recent decades, neural dynamics (based on analog solvers) has been widely arising in scientific computation and opti-
mization fields, attracting extensive interest and investigation of researchers. Due to the in-depth research, numerous solv-
ers in the form of neural dynamics have been proposed and investigated to solve a variety of problems. Moreover, such a
neural-dynamics approach is now regarded as a powerful alternative for online computation and optimization, owing to
its potential parallel-processing distributed nature and convenience of hardware implementations.
However, many of the neural-dynamic approaches may be efficient for static/time-invariant applications, and may be less
feasible for most of the time-varying applications which require a faster convergence to the time-varying solutions. Since
2001, different from the conventional neural-dynamics approach, a novel type of neural dynamics, called Zhang dynamics
(ZD), has been formally proposed by Zhang et al. [12–16] for the online solution of time-varying problems, e.g., time-varying
Sylvester equation solving [12], time-varying nonlinear equations solving [13], time-varying convex quadratic program [14],
and time-varying matrix inversion [15,16]. According to the Zhang et al.’s neural-dynamics design method, the Zhang
dynamics (ZD) is designed based on an indefinite Zhang function (ZF) as the error-monitoring function. Here, the word
‘‘indefinite’’ means that such an error-monitoring function can be positive, zero, negative or even lower-unbounded. This
differs from the usual situation involved in the design of traditional algorithms; for example, a norm-based positive-definite
energy-function is usually used in the design of gradient-based dynamics [14,15]. Thus, by making good use of the ZF as well
as the time-derivative information of the time-varying coefficients involved in the time-varying problem, the resultant ZD
models can methodologically avoid the lagging errors generated by the conventional algorithms. The ZD models can guar-
antee much better convergence performance to the time-varying theoretical solution of the time-varying problem in an
error-free manner.
In this paper, we propose, develop and investigate different continuous-time Zhang dynamics (CTZD) models for solving
online the time-varying reciprocal problem [which can be viewed as the fundamental and important operation in a floating-
point divider/processor in the form of f(x) = ax  1] by defining different ZFs as the error-monitoring functions. In addition to
the theoretical analysis and verification of the convergence characteristics of the proposed CTZD models, the MATLAB Sim-
ulink modeling and simulative examples are investigated with the goal of developing the floating-point processors or copro-
cessors for the future generation of computers. From the modeling results, the efficacy of the proposed CTZD models based
on different ZFs for online time-varying reciprocal solving is verified evidently. Note that, for better understanding, we give
the following definition of Zhang function (ZF), and that we use the term ‘‘error function’’ throughout the rest of this paper
more frequently than ‘‘Zhang function (ZF)’’.

Definition 1. The Zhang function (ZF), defined in this study, is the design basis of Zhang dynamics (ZD), which differs
evidently from the usual error functions in the study of conventional algorithms as follows. Compared with the norm-based
scalar-valued positive or lower-bounded energy function used in the design of gradient-based dynamics, (1) ZF is indefinite,
meaning that it can be positive, zero, negative or even lower-unbounded; and (2) ZF can be matrix- or vector-valued (when
solving a time-varying matrix- or vector-valued problem), so as to fully monitor and control the time-varying problem
solving process.

To the best of the authors’ knowledge, there is no literature handling such a time-varying reciprocal problem at the pres-
ent stage, and almost all reported algorithms are just for the static/time-invariant reciprocal solving [17–19]. Therefore, the
main contributions of this paper are very clear and listed in the following facts.

– Different CTZD models are proposed by defining different error functions for solving online the time-varying reciprocal
problem for the first time. This is also the main motivation of this work.
– Together with the theoretical analysis, this paper presents the convergence results of the proposed CTZD models, which
substantiate the efficacy of the proposed CTZD models for solving online the time-varying reciprocal problem.
– The MATLAB Simulink modeling technique is exploited in this paper for verifying the CTZD models, which drastically
reduces the programming efforts, and is good for the hardware/circuit implementations.
– The modeling results are presented, where different CTZD models are exploited to solve online the time-varying recipro-
cal problem, and the efficacy of the proposed CTZD models is thus verified evidently.

2. Preliminaries and ZD method

In this section, we introduce different error functions and propose the resultant CTZD models for solving online the time-
varying reciprocal problem. Meanwhile, some propositions and the related theoretical analysis are given.

2.1. Preliminaries

Without loss of generality, let us consider the time-varying reciprocal problem which is expressed in the following form:
f ðxðtÞ; tÞ ¼ aðtÞxðtÞ  1 ¼ 0 2 R; t 2 ½0; þ1Þ: ð1Þ
_
In Eq. (1), a(t) – 0 2 R denotes a smoothly time-varying scalar, and aðtÞ 2 R denotes the time derivative of a(t). In this work,
we aim at finding the unknown x(t) 2 R in real time t such that Eq. (1) holds true.
4504 Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511

2.2. CTZD models based on different error functions

The error function is the design basis for deriving a CTZD model. For presentation convenience, we denote the error func-
tion by e(t) with ė(t) as the time derivative of e(t) in this paper.
By following Zhang et al.’s neural-dynamics design method [12], we firstly define an indefinite error function as the error-
monitoring function to monitor the process of time-varying problems solving. Secondly, we choose its time-derivative ė(t)
via the following ZD design formula to force e(t) to converge to zero:
deðtÞ
_
eðtÞ :¼ ¼ c/ðeðtÞÞ; ð2Þ
dt
where the design-parameter c > 0 2 R corresponds to the reciprocal of a capacitance parameter, which should be set as large
as the hardware would permit, or selected appropriately for experimental and/or simulative purposes. In addition, different
values of c may affect the convergence performance of the resultant CTZD model. On the other hand, /() : R ? R denotes a
general monotonically-increasing odd activation function. There are many kinds of activation functions, such as the linear
activation function, the sigmoid activation function, the power-sigmoid activation function, the hyperbolic sine activation
function and the power-sum activation function. In this paper, we choose the linear activation function /(e) = e. As a result,
the ZD design formula (2) reduces to
deðtÞ
_
eðtÞ :¼ ¼ ceðtÞ: ð3Þ
dt

Remark 1. As we probably know, many conventional algorithms, such as gradient-based dynamics, are designed for static/
time-invariant problems solving, but may not be efficient enough for time-varying cases. In addition, the static problem can
be viewed as a special case of the time-varying problem. Moreover, by calculus, the analytical solution of (2) can be given,
which is different corresponding to different activation function /() used, and globally (exponentially) converges to zero
[16]. For example, (1) using the linear activation function /(e) = e, we can obtain e(t) = exp(ct)e(0) ? 0 as time t tends to
infinity, where exponential convergence rate c > 0, and e(0) denotes the initial error; and (2) using the bipolar sigmoid
activation function /(e) = (1  exp(ne))/(1 + exp(ne)) with n P 1, we can obtain e(t) = sgn(e(0)) ln (1 + z(t))/n ? 0 as time
t tends to infinity, where
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
g expðnctÞ g expðnctÞ 2
zðtÞ ¼  1þ 1
2 2
and g = (exp(ne(0))  1)2/exp(ne(0)). Evidently, the analytical solutions of (2) corresponding to using different activation
functions (especially, using the linear activation function) achieve the global (exponential) convergence [16], and thus the ZD
method and models can avoid the lagging errors (generated by conventional algorithms).
Furthermore, for the excellent property of global exponential convergence of ZD design formula (3), we have the follow-
ing theorem.

Theorem 1. As for ZD design formula (3) which is also a dynamic system, starting from any initial error e(0), the error function
e(t) globally and exponentially converges to zero with rate c > 0.

Proof. For (3), by calculus, we obtain its analytical solution as e(t) = exp(ct)e(0). Based on the definition of global exponen-
tial convergence, we can draw the conclusion that, starting from any e(0), the error function e(t) globally and exponentially
converges to zero with rate c > 0, as time t tends to infinity. The simple proof is thus completed. h

Specifically, for the online solution of time-varying reciprocal problem (1), in this paper, we can define different error
functions as follows:
1
eðtÞ :¼ xðtÞ  ; ð4Þ
aðtÞ
1
eðtÞ :¼ aðtÞ  ; ð5Þ
xðtÞ
eðtÞ :¼ aðtÞxðtÞ  1; ð6Þ
1
eðtÞ :¼  1: ð7Þ
aðtÞxðtÞ
According to the ZD design formula (3), different error functions can lead to different CTZD models. Therefore,

(1) let us consider ZD design formula (3) and error function (4). Then, we have
 
1 1
_ þ
xðtÞ _
aðtÞ ¼ c xðtÞ  ;
a2 ðtÞ aðtÞ
a2 ðtÞxðtÞ
_ _  cða2 ðtÞxðtÞ  aðtÞÞ:
¼ aðtÞ ð8Þ
Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511 4505

Thus, we obtain CTZD model (8) for time-varying reciprocal solving. For further illustration and investigation, we can also
exploit the other three error functions (i.e. error functions (5)–(7)) to build other types of CTZD models.
(2) Considering ZD design formula (3) and error function (5), we have
 
1 1
_ þ
aðtÞ _
xðtÞ ¼  c aðtÞ 
x2 ðtÞ xðtÞ
which can be further written as

2
_
xðtÞ _
¼ aðtÞx ðtÞ  cðaðtÞx2 ðtÞ  xðtÞÞ: ð9Þ
Therefore, CTZD model (9) for time-varying reciprocal solving is obtained.
(3) By combining ZD design formula (3) and error function (6), we have

_
aðtÞxðtÞ _
þ aðtÞxðtÞ ¼ cðaðtÞxðtÞ  1Þ;
and then
_
aðtÞxðtÞ _
¼ aðtÞxðtÞ  cðaðtÞxðtÞ  1Þ: ð10Þ
CTZD model (10) for time-varying reciprocal solving is thus obtained.
(4) With ZD design formula (3) and error function (7) combined, we have
 
1 1
 _
ðaðtÞxðtÞ _
þ aðtÞxðtÞÞ ¼ c 1 ;
a2 ðtÞx2 ðtÞ aðtÞxðtÞ
which can be further written as

_
aðtÞxðtÞ _
¼ aðtÞxðtÞ þ cðaðtÞxðtÞ  a2 ðtÞx2 ðtÞÞ: ð11Þ
Therefore, we come to CTZD model (11) for time-varying reciprocal solving.

As a result, we have obtained four different types of CTZD models (i.e., CTZD models (8)–(11)) for time-varying reciprocal
solving, which correspond to four different types of error functions (i.e., error functions (4)–(7)). For readers’ convenience
and also for comparison, the four different CTZD models are listed below:

a2 ðtÞxðtÞ
_ _  cða2 ðtÞxðtÞ  aðtÞÞ;
¼ aðtÞ
2
_
xðtÞ _
¼ aðtÞx ðtÞ  cðaðtÞx2 ðtÞ  xðtÞÞ;

_
aðtÞxðtÞ _
¼ aðtÞxðtÞ  cðaðtÞxðtÞ  1Þ;

_
aðtÞxðtÞ _
¼ aðtÞxðtÞ þ cðaðtÞxðtÞ  a2 ðtÞx2 ðtÞÞ;

corresponding to error functions (4)–(7), respectively. In the ensuing subsection, we will have some propositions about the
CTZD models with an illustrative theoretical analysis shown.

2.3. Propositions and theoretical analysis

Before presenting more theoretical results, it is worth pointing out that the ensuing propositions in this subsection can be
viewed as the special cases of the above presented Theorem 1. Moreover, for the completeness of this paper, the proof of
Proposition 1 is given in detail as an exemplar analysis.

Proposition 1. Consider a smoothly time-varying scalar a(t) – 0 2 R involved in time-varying reciprocal problem (1). Starting
from randomly-generated initial state x(0) – 0 2 R which has the same sign as a(0), the neural state x(t) of CTZD model (8) derived
from error function (4) exponentially converges to the time-varying theoretical reciprocal a1(t) of a(t).

Proof. We use the well-known Lyapunov method to prove the exponential convergence of CTZD model (8).
Firstly, starting with error function (4), we define a Lyapunov candidate
 2
1 1
VðxðtÞ; tÞ ¼ xðtÞ  P 0;
2 aðtÞ
where V(x(t),t) = 0 for any x(t) = a1(t), and V(x(t),t) > 0 for any x(t) – a1(t). Then, we derive its time derivative as
4506 Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511

     2
_ dVðxðtÞ; tÞ 1 1 1
VðxðtÞ; tÞ :¼ ¼ xðtÞ  _ þ 2 aðtÞ
 xðtÞ _ ¼ c xðtÞ  ¼ 2cVðxðtÞ; tÞ: ð12Þ
dt aðtÞ a ðtÞ aðtÞ

_
Since V(x(t), t) P 0, then VðxðtÞ; _
tÞ ¼ 2cVðxðtÞ; tÞ 6 0, which guarantees the (final) negative-definiteness of VðxðtÞ; tÞ.
Furthermore, from (12), we have

VðxðtÞ; tÞ ¼ Vðxð0Þ; 0Þe2ct :


That is,
 2  2
1 1 1 1
xðtÞ  ¼ xð0Þ  e2ct :
2 aðtÞ 2 að0Þ
Thus, we have
   
   
xðtÞ  1  ¼ xð0Þ  1 ect : ð13Þ
 aðtÞ   að0Þ
With a :¼ jx(0)  1/a(0)j, Eq. (13) is rewritten as
 
 
xðtÞ  1  ¼ aect : ð14Þ
 aðtÞ
Eq. (14) means that x(t) exponentially converges to a1(t) with the convergence rate c > 0. That is, starting from randomly-
generated initial state x(0) – 0 2 R which has the same sign as a(0), the neural state x(t) of CTZD model (8) exponentially
converges to the time-varying theoretical reciprocal a1(t) of a(t) involved in time-varying Eq. (1). h

Proposition 2. Consider a smoothly time-varying scalar a(t) – 0 2 R involved in time-varying reciprocal problem (1). Starting
from randomly-generated initial state x(0) – 0 2 R which has the same sign as a(0), the neural state x(t) of CTZD model (9) derived
from error function (5) converges to the time-varying theoretical reciprocal a1(t) of a(t), with the error defined in error function
(5) exponentially convergent to 0.

Fig. 1. Block diagrams of CTZD models for time-varying reciprocal solving.


Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511 4507

Proposition 3. Consider a smoothly time-varying scalar a(t) – 0 2 R involved in time-varying reciprocal problem (1). Starting
from randomly-generated initial state x(0) – 0 2 R which has the same sign as a(0), the neural state x(t) of CTZD model (10)
derived from error function (6) converges to the time-varying theoretical reciprocal a1(t) of a(t), with the error defined in error
function (6) exponentially convergent to 0.

Proposition 4. Consider a smoothly time-varying scalar a(t) – 0 2 R involved in time-varying reciprocal problem (1). Starting
from randomly-generated initial state x(0) – 0 2 R which has the same sign as a(0), the neural state x(t) of CTZD model (11)
derived from error function (7) converges to the time-varying theoretical reciprocal a1(t) of a(t), with the error defined in error
function (7) exponentially convergent to 0.

3. MATLAB Simulink modeling

For possible hardware implementation based on digital circuits and also for the goal of developing the floating-point pro-
cessors or coprocessors for the future generation of computers, the MATLAB Simulink modeling of the proposed CTZD models
(i.e., CTZD models (8)–(11)) is investigated and presented in this paper. Before doing this, we need to transform some of the
proposed CTZD models into the following explicit forms, with the corresponding block diagrams depicted in Fig. 1.

 For CTZD model (8),


_
xðtÞ ¼ ð1  a2 ðtÞÞxðtÞ _  cða2 ðtÞxðtÞ  aðtÞÞ:
_  aðtÞ

 For CTZD model (9), it is already in the explicit form, and does not need to be transformed.
 For CTZD model (10),

Fig. 2. Overall Simulink models of CTZD models (8) and (9) for online time-varying reciprocal solving (i.e., time-varying Eq. (1) solving).
4508 Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511

_
xðtÞ _
_  aðtÞxðtÞ
¼ ð1  aðtÞÞxðtÞ  cðaðtÞxðtÞ  1Þ:
 For CTZD model (11),
_
xðtÞ _
_  aðtÞxðtÞ
¼ ð1  aðtÞÞxðtÞ þ cðaðtÞxðtÞ  a2 ðtÞx2 ðtÞÞ:

Therefore, the overall Simulink models of the proposed CTZD models are shown in Figs. 2 and 3, in which a(t) is generated
by employing the ‘‘MATLAB Function’’ block using the ‘‘Clock’’ block as its input.

4. Simulative examples

In the previous sections, we have proposed the CTZD models based on different error functions for time-varying recipro-
cal solving, together with corresponding propositions and theoretical analysis. Based on the above-mentioned overall Sim-
ulink models depicted in Figs. 2 and 3, the ensuing illustrative examples are shown to verify the efficacy of the proposed
CTZD models.

Example 1. Let us consider the following time-varying reciprocal problem (in which a(t) = sin(3t) + cos(sin(2t)) + 2 is
involved):

f ðxðtÞ; tÞ ¼ ðsinð3tÞ þ cosðsinð2tÞÞ þ 2ÞxðtÞ  1 ¼ 0: ð15Þ


The proposed CTZD models, i.e., CTZD models (8)–(11), are exploited to solve this problem (15). It is easy to see that the
theoretical initial reciprocal value is x⁄(0) = 1/a(0)  0.333. For the convenience of observation, the initial state (or to say,
starting value) x(0) is randomly generated within [0.2, 0.4], i.e., x(0) 2 [0.2, 0.4]. The simulative results based on the Simulink
models are presented in Fig. 4. As shown in the figure, with design-parameter c = 10, the states x(t) of the proposed CTZD
models all converge to the time-varying theoretical reciprocal a1(t), i.e., the time-varying theoretical solution of (15), in

Fig. 3. Overall Simulink models of CTZD models (10) and (11) for online time-varying reciprocal solving (i.e., time-varying Eq. (1) solving).
Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511 4509

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

Fig. 4. State trajectories of CTZD models (8)–(11) with design-parameter c = 10 for solving the time-varying reciprocal problem (15), where dash-dotted
curves correspond to the time-varying theoretical reciprocal a1(t), i.e., the theoretical solution of (15).

a rather short time (i.e., in less than 1 s). Through this example, we have primarily verified the efficacy of the proposed CTZD
models for solving online the time-varying reciprocal problem.

Example 2. In this example, we are considering a more complicated situation of the time-varying reciprocal problem, i.e.,
f ðxðtÞ; tÞ ¼ ðsinðcosð4tÞÞ þ expð sinð3tÞÞ þ cosðtÞ þ 2ÞxðtÞ  1 ¼ 0; ð16Þ
so as to investigate the general applicability of the proposed CTZD models. In other words, a(t) = sin(cos(4t)) +
exp(sin(3t)) + cos(t) + 2.
Similar to the way of Example 1, the proposed CTZD models are exploited to solve this problem (16). As shown in Fig. 5,
with the design-parameter c = 10 and the initial state x(0) randomly generated within [0.1, 0.4], the state trajectories of the
proposed CTZD models all fit well with the time-varying theoretical reciprocal a1(t), i.e., the time-varying theoretical
solution of (16), rapidly (i.e., in less than 1 s).
From the above two examples, we can draw a conclusion that, either for a simple problem (15) or a more complicated
problem (16), the proposed CTZD models based on different error functions all can solve the time-varying reciprocal problem
efficiently; i.e., with appropriate values of design-parameter c and initial state x(0), the states of the proposed CTZD models
all converge to the time-varying theoretical solution of the time-varying reciprocal problem rapidly and accurately. Thus, the
efficacy of the proposed CTZD models is substantiated well.

Example 3. As mentioned in the previous sections, the value of design-parameter c may affect the convergence performance
of the proposed CTZD models. Let us consider the following time-varying reciprocal problem (in which
a(t) = sin(t)cos(2t) + 2):
4510 Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

Fig. 5. State trajectories of CTZD models (8)–(11) with design-parameter c = 10 for solving the time-varying reciprocal problem (16), where dash–dotted
curves correspond to the time-varying theoretical reciprocal a1(t), i.e., the theoretical solution of (16).

0.1

−0.1
0 2 4 6 8 10
0.1

−0.1
0 0.2 0.4 0.6 0.8 1
0.1

−0.1
0 0.02 0.04 0.06 0.08 0.1

Fig. 6. Computational errors e(t) :¼ x(t)  1/(sin(t) cos(2t) + 2) of CTZD model (8) for solving the time-varying reciprocal problem (17) with different values
of c.
Y. Zhang et al. / Applied Mathematical Modelling 36 (2012) 4502–4511 4511

f ðxðtÞ; tÞ ¼ ðsinðtÞ cosð2tÞ þ 2ÞxðtÞ  1 ¼ 0: ð17Þ


In this example, CTZD model (8) is exploited to solve (17) with the initial state x(0) randomly generated within [0.4, 0.6]. The
computational errors e(t) :¼ x(t)  1/(sin(t)cos(2t) + 2) with respect to different values of c are displayed in Fig. 6. As seen
from the figure, the convergence time of the computational error e(t) to zero is becoming much shorter (i.e., from about
0.6 s to about 0.006 s) when the value of c increases from 10 to 103. This simulative result indicates that the design-param-
eter c plays an important role in the proposed CTZD model (8), and should be set as large as the hardware would permit, or
selected appropriately large for experimental and/or simulative purposes. Furthermore, the exponential-convergence char-
acteristics of CTZD model (8) can also be seen comparatively from this figure.

5. Conclusions

In this paper, by following Zhang et al.’s neural-dynamics design method and based on the Zhang function (ZF) as the
error-monitoring function, a novel type of neural dynamics, called Zhang dynamics (ZD), has been presented and investi-
gated for online solution of time-varying reciprocal problem, which is in the time-varying form of f(x(t), t) = a(t)x(t)  1 = 0.
Based on different error functions, different CTZD models have been proposed, developed and investigated. Moreover, the-
oretical analysis and results are given to substantiate the exponential convergence of the proposed CTZD models. For pos-
sible hardware implementations based on digital circuits and for the goal of developing the floating-point processors or
coprocessors for the future generation of computers, the MATLAB Simulink modeling of the proposed CTZD models has been
presented and investigated in this paper. Through illustrative simulation examples, the efficacy of the proposed CTZD models
has been further demonstrated.

Acknowledgements

This work is supported by the National Natural Science Foundation of Chinaunder Grants 61075121 and 60935001, and
also by the Fundamental Research Funds for the Central Universities of China.

References

[1] M.K. Kadalbajoo, D. Kumar, A computational method for singularly perturbed nonlinear differential-difference equations with small shift, Appl. Math.
Model. 34 (2010) 2584–2596.
[2] M. Dehghan, M. Hajarian, An iterative method for solving the generalized coupled Sylvester matrix equations over generalized bisymmetric matrices,
Appl. Math. Model. 34 (2010) 639–654.
[3] W. Wang, Y. Zhang, S. Li, Stability of continuous Runge–Kutta-type methods for nonlinear neutral delay-differential equations, Appl. Math. Model. 33
(2009) 3319–3329.
[4] M. Salahi, Robust least squares solution of linear inequalities, Appl. Math. Lett. 23 (2010) 605–608.
[5] A. Dax, A hybrid algorithm for solving linear inequalities in a least squares sense, Numer. Algorithms 50 (2009) 97–114.
[6] Y. Xia, G. Feng, J. Wang, A recurrent neural network with exponential convergence for solving convex quadratic program and related linear piecewise
equations, Neural Netw. 17 (2004) 1003–1015.
[7] R.H. Sturges Jr., Analog matrix inversion (robot kinematics), IEEE J. Robotic. Autom. 4 (1988) 157–162.
[8] O. Wolkenhauer, M. Ullah, P. Wellstead, K.H. Cho, The dynamic systems approach to control and regulation of intracellular networks, FEBS Lett. 579
(2005) 1846–1853.
[9] G.P. Cai, J.Z. Hong, S.X. Yang, Dynamic analysis of a flexible hub-beam system with tip mass, Mech. Res. Commun. 32 (2005) 173–190.
[10] B. Deviren, M. Keskin, Dynamic phase transitions and compensation temperatures in a mixed spin-3/2 and spin-5/2 Ising system, J. Statist. Phys. 140
(2010) 934–947.
[11] L. Borland, Microscopic dynamics of the nonlinear Fokker–Planck equation: a phenomenological model, Phys. Rev. E 57 (1998) 6634–6642.
[12] Y. Zhang, D. Jiang, J. Wang, A recurrent neural network for solving Sylvester equation with time-varying coefficients, IEEE Trans. Neural Netw. 13
(2002) 1053–1063.
[13] Y. Zhang, C. Yi, W. Ma, Comparison on gradient-based neural dynamics and Zhang neural dynamics for online solution of nonlinear equations, Lect.
Notes Comput. Sci. 5370 (2008) 269–279.
[14] Y. Zhang, Z. Li, Zhang neural network for online solution of time-varying convex quadratic program subject to time-varying linear-equality constrains,
Phys. Lett. A 373 (2009) 1639–1643.
[15] Y. Zhang, W. Ma, B. Cai, From Zhang neural network to Newton iteration for matrix inversion, IEEE Trans. Circ. Syst-I 56 (2009) 1405–1415.
[16] Y. Zhang, S.S. Ge, Design and analysis of a general recurrent neural network model for time-varying matrix inversion, IEEE Trans. Neural Netw. 16
(2005) 1477–1490.
[17] G. Hanrot, J. Rivat, G. Tenenbaum, P. Zimmermann, Density results on floating-point invertible numbers, Theor. Comput. Sci. 291 (2003) 135–141.
[18] E. Croot, R. Li, H.J. Zhu, The abc conjecture and correctly rounded reciprocal square roots, Theor. Comput. Sci. 315 (2004) 405–417.
[19] G. Agrawal, A. Khandelwal, E.E. Swartzlander Jr., An improved reciprocal approximation algorithm for a Newton Raphson divider, in: Proceedings of
Advanced Signal Processing Algorithms, Architectures, and Implementations XVII, San Diego, CA, USA, August 2007, pp. 1–12.

You might also like