Haar wavelet method for solving lumped and

distributed-parameter systems

C.F.Chen
C.H.Hsiao

Indexing terms: Haar wavelet, Matrix of integration, Dynamic systems

All previously mentioned orthogonal functions, how-

Abstract: An operational matrix of integration ever, are supported on the whole interval a I t 5 b.
based on Haar wavelets is established, and a This kind of global support is evidently a drawback for
procedure for applying the matrix to analyse certain analysis work, particularly systems involving
lumped and distributed-parameters dynamic abrupt variations or a local function vanishing outside
systems is formulated. The technique can be a short interval of time or space. Owing to cancella-
interpreted from the incremental and tions of many terms being required in order to obtain a
multiresolution viewpoint. Crude as well as reasonable accuracy, the recently developed technique,
accurate solutions can be obtained by changing wavelet analysis, could be a possible tool for solving
the parameter m; in the mean time, the main the difficulty in physics, communication and image
features of the solution are preserved. Several processing. This paper uses the simplest wavelet func-
nontrivial examples are included for tion, or Haar wavelet, to explore the new direction in
demonstrating the fast, flexible and convenient system analysis; i.e. we establish an operational matrix
capabilities of the new method. for integration via Haar wavelets. By using this new
matrix, the drawbacks caused by the whole range sup-
port situation are eliminated. We will apply this opera-
tional matrix to lumped-parameter systems first, and
1 Introduction then to distributed-parameter systems, to demonstrate
the simplicity of the approach.
Using orthogonal functions to construct operational The orthogonal set of Haar functions is defined as
matrices for solving identification and optimisation shown in Figs. 1-8 [8-101. That is a group of square
problems of dynamic systems, was initially established waves with magnitude of 21 in some intervals and
in 1975 when the Walsh-type operational matrix was zeros elsewhere. Just these zeros make the Haar
constructed by the present authors [l]. Since then, transform faster than other square functions such as
many operational matrices based on various orthogo- Walsh’s. The first curve of Fig. 1 is that h,(t) = 1
nal functions, such as block pulse [2], Laguerre [3, 41 during the whole interval 0 5 t 5 1. It is called the
Legendre [5], Chebyshev [6], and Fourier [7], have been scaling function. The second curve hl(t) is the
developed. The main characteristic of this technique fundamental square wave, or the mother wavelet which
is to convert a differential equation into an algebraic also spans the whole interval (0, I). All the other
equation, and hence the solution, identification and subsequent curves are generated from h l ( t ) with two
optimisation procedures are either greatly reduced or operations: translation and dilation. h2(t) is obtained
much simplified accordingly. from hl(t) with dilation, i.e. h,(t) is compressed from
The basic idea starts from the integral of a basic the whole interval (0, 1) to the half interval (0, 112) to
vector # ( t ) . We will give the following approximation
~
generate h2(t). h3(t) is the same as h2(t) but shifted
first. (translated) to the right by 1/2. Similarly, h2(t) is

Jo’ 4(T)d7. P4(t) (1)

compressed from a half interval to a quarter interval to
generate h4(t). The function h4(t) is translated to the
right by 1/4, 2/4, 3/4 to generate h,(t), h6(t), h7(t),
where # ( t ) = [#,(t), #l(t), ..., #mm-l(t)]l,in which the respectively. In general,
elements #,(t), ..., #m-l(t) are the basic functions,
h,(t) = h I ( 2 3 t - k / 2 3 ) (2a)
orthogonal on certain interval [a, b], where a and b
could be 0 and 1, respectively, and the matrix P can be n=2J+k, j > O , O<k<2’ (2b)
uniquely determined based on the particular orthogo- h,(t) is also included to make this set complete. We
nal function. have the following:
0IEE, 1996
IEE Proceedings online no. 19960702
Paper received 19th December 1995
Professor C.F. Chen is with the Department of Electrical Engineering,
National Cheng Kung University, Taiwan, Republic of China
C.H. Hsiao is with the Department of Electrical Engineering, Wu Feng
Institute of Technology and Commerce, Cha Yi, Taiwan, Republic of
China Fig. 1 First Huar function and corresponding integral

IEE Proc.-Control Theory Appl.. Vol. 144, No. 1, January 1997 87

 Usually, the series expansion of eqn. 4 contains infinite
terms for approximating smooth y(t). If y ( t ) is piece-
wise constant by itself, or may be approximated as
piecewise constant during each subinterval, then eqn. 4
will be terminated at finite terms.
Fig.2 Second Haar function and corresponding integral We use row vectors to denote time functions, and
column vectors x(t), u(t), y ( t ) to denote the state, the
input, and the output vectors, respectively. The first
four Haar functions can be expressed as follows:
hop) = [ 1 1 1 11
U h l ( t ) = [l 1 -1 -I]
Fig.3 Third Haar function and corresponding integral
h2(t) = [ l -1 0 01
h3(t) = [ O 0 1 -11 (6)
For instance, if y ( t ) = [9 1 2 01 is piecewise constant,
then
U

Fig.4 Fourth H a m function and corresponding integral

+ + +
y ( t ) = 3ho(t) 2hl(t) 4h:,(t) h3(t) = c t H ( t ) ( 7 )
where H(t) is H4(t)and

Lhg(t)l to 0 1 -11
U The Haar coefficient ci can be obtained by applying
Fig.5 Fgth Haar function and corresponding integral eqn. 5 directly; however, it is more convenient to evalu-
ate it by matrix inversion.
Ct = y(t)HL1 = [ 3 2 4 11 (9)
1 2
U
Fig. 6 Sixth H a m function and corresponding integral
L1 -1 0 -21
Eqn. 9 is called the forward transform, which turns the
time function y ( t ) into the coefficient vector ct, and
eqn. 7 is known as the inverse transform, which recov-
ers y(t) from et. Since H and N-l contain many zeros,
U this phenomenon makes the Haar transform much
Fig.7 Seventh Haar function and corresponding integral faster than the Fourier transform, and it is even faster
than the Walsh transform.
rt
2 integration of Haar wavelets
In studying differential equation models of dynamic
U
systems, it is frequently needed to perform integrations
Fig.8 Eighth Haar function and corresponding integral in order to get the dynamic problem solved. The new
approach we are developing is also using the integra-
tion technique.
This orthogonal basis is not a recent invention [SI. It Let us consider a 4th-order system. The integrals of
is a reminiscent of the Walsh basis [l], in which each the first four Haar wavelets can be expressed as:
Walsh function contains many wavelets to fill the inter-
val (0, 1) complelely, and to form a global basis. While l h n ( t ) d t = t , o < t < l = -1 [ l 3 5 71 (11)
each Haar function contains just one wavelet during 8
some subinterval of time, and remains zero elsewhere; 1
the Haar set forms a local basis.
We have noticed that all the Haar wavelets are
1 t
hl(t)dt=
t, o < t < -

1-t,
1
-<t<l
]+,l 3 3 11
orthogonal to each other: 2-
(12)

Therefore, they form a very good transform basis. Any

function y(t), which is square integrable in the interval
(0, l), namely Jdy2(t)dt is finite, can be expanded into
Haar wavelets [lo, 111:
+ +
y(t) = coho(t) Clhl(t) c2h2(t) ' . + (4) '

ci = 2 j 1 1
y(t)h$)dt (5)

88 IEE Proc -Control Theory Appl., Vol. 144, No. 1, January 1997
Writing eqns. 11 to 14 together, we obtain All these matrices P2, P4, P8 satisfy eqn. 19. Of course,
rl 3 5 71 we should have a general and more rigorous proof in
addition to the numerical check. Let us partition P,
into four submatrices:
0 1 11 Lo (23)
Let the integrals be expanded into Haar series. Thus,
Eqn. 19 consists of the following four equations:
H4(t)dt = I 3 4 H 4 ( t ) (16) pma = pm/2 (24a)
1
r 4 -2 -1 -11 Pmb = -- Hm/2
2m
1
P,, = -H-l
2m
Pmd = null matrix
Following the same procedures,-the integration matrix
We now offer the general proof.
P8 can be obtained as follows:
Proof 2.1: Choose m = 4 for illustration. As shown in
eqn. 15, J,j H4 (z)dz contains four integrals. By defini-
tions of eqns. 21 and 23, Pda is the transform of the
first two integrals into the first two Haar functions
ho(t) and hl(t). On the other hand, P2 of eqn. 22 is also
the transform of these two integrals into ho(t) and h,(t).
Therefore they should be equal and eqn. 24a should
hold, although eqn. 17 is evaluated with four subinter-
In general for an mth-order system with m = 2j, j is a vals while eqn. 22 is evaluated with two subintervals.
positive integer, the P, is given as That makes no difference.
Proof 2.2: By definition, P4b is the transform of the
integrals ho(t) and hl(t) into series of h2(t)and h3(t).
To prove matrix eqn. 19 we verify it numerically first.
For an mth-order system the Haar matrix H, is
defined with m Haar functions in row vector form as which shows that the integral of ho(t) is a ramp
for matrix eqn. 8. In other words, function, and the integral of hl(t) is a triangular
function consisting of a rising ramp and a falling ramp.
It is noted that the absolute value of the slopes of these
ramps is the same.
The transform is carried out with H c l . As shown in
eqn. 10, the third column of Hcl has 2 and -2 in its (1,
3), (2, 3) entries. When eqn. 25 is postmultiplied with
this column, the second column of eqn. 25 is sub-
stracted from its first column, and the difference is
divided by 2. That means we are calculating the nega-
tive slope of J; ho(z)dzand Jot hl(z)dz.Namely, we are
The integrals of Haar functions are triangular func- taking the negative derivatives of these integrals. The
tions, as explained earlier. The operation matrix P, is same situation occurs when eqn. 25 is postmultiplied
the Haar transform coefficient matrix of these integrals with the fourth column of H r l . Because integration
as defined by eqn. 16: and differentiation cancel each other, we have P4b
equals -(1/2m)H2.

L U -1A
r8 -4 -2 -2
1
p4 = 16 1 1 0 0
Proof2.3: When the integrals of h2(t)and h3(t)are mul-
.1 -1 0 0
tiplied by the first two columns of Hcl, this results in
32 -16 -8 -8 -4 -4 -4 -4 P4c.That means P4cis the transform of the integrals of
-16 0 -8 8 -4 -4 4 4 h2(t) and h3(t)into series of ho(t)and h,(t). These inte-
4 4 0 0 - 4 4 0 0 grals are triangular functions which span the first and
the second half intervals. In numerical expression, the
1 4 - 4 0 0 0 0 - 4 4 averaged values are taken to represent these triangular
Pa = -
64 1 1 2 0 0 0 0 0 functions.
1 1 - 2 0 0 0 0 0
1 - 1 0 2 0 0 0 0
- 1 -1 0 -2 0 0 0 0
IEE Proc.-Control Theory Appl., Vol. 144, No. 1, January 1997 89
Comparing the first two columns of H c l and H y l , we The Kronecker product A 0 Pt is defined as [131
find that the former is just a dilation of the latter. This -j P l l A P2d . . . Pm1A1
means that the second and the fourth rows of the first
two columns of H L ~are the same as the first and the
third rows there. Therefore,
A8Pt= I . p2:A :* 'mZA 1 (36)
rl 1i LPIiA P z ~ A PmiAj ..:
After F is obtained, it can be used in eqn. 32 to find
x(t), which can be entered into eqn. 29 to calculate y ( t )

=&
1 1 01] [: 3=p;' 1
with only a few matrix operations.
The above derivations illustrate that solving the
dynamic equation via the Haar wavelets approach is
remarkably feasible. In practical calculation, one
Proof 2.4: By definition, Pdd is the transform of the should study the property of P and H closely to choose
integrals of h,(t) and h3(t).As shown in eqn. 27 its each the most effective calculating subroutines. Zeros
row contains two equal positive entries, while each col-
involved in H and P will greatly simplify the solution
umn of the last half of Hcl contains two entries of 22.
procedures. Alternatively, eqn. 34 can be rewritten as
Therefore these rows and columns are orthogonal to
each other, and Pddis a null matrix. A-'F - F P = A-'GL (37)
To make the above analyses and proofs clear and for nonsingular A , this is a Lyapunov matrix equation
compact, m = 4 has been used. In general, the above [14], which arises in several areas of control including
proofs can be applied for any m. stability theory, and the study of the RMS behaviour
This completes the proof of eqn. 19.
of systems. The subroutine lyup (A-l, -P, A-'GI) of
3 Analysis of lumped-parameter systems via MATLAB, or the subroutine srhw of the Appendix
Haar wavelets may be applied directly to obtain F with A and G1
being given.
A lumped-parameter linear system with n states x(t), p
inputs u(t), and 4 outputs y ( t ) is usually described by low-pass PI motor and
the state equation and the output equation with prop- filter controller bad

erly dimensioned A , B, C, D matrices. 1 K S-' 250 S-l 25 S-' 18 S-' I

~ ( t=) Ax(t)+ B u ( t ) , ~(0)

= xo (28)
+
y ( t ) = C x ( t ) DU(t) (29)
-5.125

Assume that u(t) is square integrable in the interval 0 5

t < 1. Its Haar series expansion can be expressed as
Fig. 9 Signal-$ow graph of a position contvol system
~ ( t=)G H ( t ) (30)
where G is a p x m matrix, which can be obtained with
the method mentioned in Section 1. Example I : The signal flow graph of a practical system
To avoid dealing with impulse functions, we usually [15] is shown in Fig. 2. Its dynamic equation is
expand the variable k(t) instead of x(t) itself, into a expressed as
Haar series. r 0 18 0 0 1 r 01
X(t)= F H ( t ) (31)
X ( t )=
Integrating yields 0 250
x(t) =
i' X ( T ) +XO =F
i' H ( T ) ~ T + x=
o F P H ( t )+ X O
(32)
L-k -92.25 0 -2501 Lkl
Two sets of solutions are obtained via the Haar
approach: for k = 168 (damping ratio 5 = 0.5) and k =
Substituting eqns. 30, 31 and 32 into eqn. 28, we obtain 390 (< = 0.2) with m = 16 and m = 64, respectively. In
+ +
F H ( t ) = A F P H ( t ) Ax0 B G H ( t ) Figs. 10 and 11 solid and dotted curves express the
details for the two cases.
+
[ F - A F P ] H ( t )= {[Axo, O , O , . . . ,O] B G } H ( t ) (33)
The number of zero columns involved in eqn. 33 is m
- 1, or
n
F - AFP [Axo,O , O , . . . 01 + BG = GI
)

[fo)fi,...,fm-l]
-A[f~,f~,-.-,fm-,]P
-
[go)g l , . . . gm-11
- ) (34)
whereA and g, are the (i + 1)st columns of F and GI,
respectively. Since the unknown matrix F in eqn. 34 is
premultiplied by A and postmultiplied by P , it should
be solved with Kronecker product as 0.2 -

fo 0 00 0.01 0.02 0.03 0.04 0.05

. 0.06 0.07 0.08 0.09 0.1
4
fl t1me.s
= [I- A 8 (35) Fig. 10 Huar solutions of a 4th-order system
-: x l ( t ) with 6= 0.5, m = 16
fm-1 .. . ....... : x l ( t ) with = 0.2, m = 64

90 IEE Proc.-Control Theory Appl., Vol. 144, No. I, January 1997

 4r Various transform methods, such as the Laplace
-- transform, the Fourier transform, the Walsh transform
31 --- ---- and the block pulse transform etc., have been available
I -
for a long time, as already mentioned in Section 1. A
new approach to solve for the state variables in distrib-
uted-parameter systems via Haar wavelets is proposed
in this Section. Compared with the classical methods,
the new approach is much more elegant in theory,
more convenient in numerical calculation. and most of
all, it is much faster in data processing [16].
-21
---- To illustrate the new idea clearly and concretely, let
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 us consider an initially relaxed infinite transmission line
time,s with R resistance, L inductance, G conductance, and C
Fig. 11 Haar solutions of a 4th-order system capacitance per unit length, as shown in Fig. 13. When
-: x2(t) with 5 = 0.5, m = 16 Ax approaches zero, the voltage and the current in the
...........: x2(t) with 5 = 0.2, m = 64
line are governed by the well known telegraphist's
equation [17-191.
To emphasise the piecewise constant properties of ai
Ri+L-z--
av
Haar solutions, 256 points have been drawn in these at ax
curves. The time interval of 0.1 s has been normalised av ai
to be 1 in the Haar expansion. For m = 64, each subin- Gv+C-=-- (39)
at dX
terval of 0.1164s contains four points with the same
values. Similarly, for m = 16, each subinterval of
0.1/16s contains 16 points. If 64 points of Haar solu-
a2v
-= L
8x2
a2v
C S + (RC + LG)-av
at
+ RGu (40)
tions are used directly via the conventional plot sub-
routine with each Haar solution as the midpoint value
a2i
-- - LC-
a2i + (RC + LG)-ai + RGi
8x2 at2 at
(41)
of each subinterval, it results in very smooth curves as
shown in Fig. 12. In principle, we apply the Haar transform in the time
domain, and solve the space domain problem with the
conventional method. Let us illustrate this idea with
the following examples.
Example 2: Suppose a unit step voltage is applied to
the initially relaxed infinite line at x = 0. The driving
point current i(0, t ) must be found.
Solution: Let us apply the Haar transform in the time
domain and expand a2vldt2 into a Haar series:
'\, x2(1) ,/' d2v/dt2 = at(x)H(t) (42)
-I - Y where a(x) is an m-vector function of x,x denotes the
-2
\
.--' ,

I
space distance. The x in a(x) will be dropped to
simplify the notation.

4 Analysis of distributed-parameter systems via

Haar wavelets where dv(x, O)/& and y(x, 0) have been set to zero for
the initially relaxed system. Eqns. 42, 43 and 44 can be
It is well known that the governing principles of dis- applied to eqn. 40
tributed-parameter systems are described by a set of at - atM2 = 0 (45)
partial differential equations. These are usually rather n
too complicated to be investigated theoretically and a = d2a/dx2 (46)
solved numerically, since such equations have at least
two independent variables and two sets of boundary M e [LCI+(RC+LG)P+RGP2]$P-l (47)
conditions to be satisfied. Solving eqn. 45, we have
at = a: exp(-Mz) + ai exp(Mz) (48)

-
I (x.t) RAx/2 LAx12 I(x +Ax, t )
I 1 1 - I
where al and a2 are two constant vectors to satisfy the
boundary conditions. The voltage y(x, t ) must be finite
when x + CO. It is necessary that a2 = 0,
v(x,t ) = a4 exp(-Mz)P2H(t) (49)
Lax12 RAx12
When a unit step voltage is applied at x = 0, it means
Y ( 0 , t ) = 1:
Ax
Fig. 13 Elementary length of a transmission line v(0,t ) = at,P2H(t)= [I,1,... ,I]
IEE Proc.-Control Theory Appl., Vol. 144, No. 1, January 1997 91
 a For leakage-free noninductive cable G = 0, L = 0,
a:P2 = [ I , o , .. . ,o]= ai (50) eqns. 38 and 40 are reduced to the following forms
at = [I,0 , . . . , I ] P - ~
= aiPp2 (50a) Ri = - ~ L J / ~ x (58)
Applying the Haar transform to eqn. 38 at the driving
point, namely x = 0 d2v/dx2= RCdv/dt (59)
Eqn. 59 is known as a diffusion equation. It is well
di(0,t)/dt = btH(t) (51) known that the diffusion of a liquid in a homogeneous
medium, the transmission of heat through a slab of
i(0,t) = dt + i(0,O) = b t P H ( t ) (52) uniform thickness and the distribution of alternating
current in a homogeneous plane conductor can all be
Substituting eqns. 49, 51 and 52 into eqn. 38 with x = described with the same form of partial differential
0, we have equations.
+
b t [ R P L I ] = a4MP2 Haar wavelets are introduced to solve these partial
differential equations.
bt = a i M ( R P +LI)-lP2 Solution: In the Haar domain, let us assume a d a t can
= m a : [ ( G / C ) P+ 1 ] 9 [ ( R / L ) +
P I]-3P (53) be expanded in a Haar series as
where the cummutative property has been applied. av/dt = at ( x ) H ( t ) (60)
Finally, we have Integrating eqn. 60 and applying the integration matrix
P of Section 2, we have
z(0,t) = m a h [ ( G / C ) P + I]i[(R/L)P
+Ilp+H(t)
where a. is defined by eqn. 50.
(54) Y(Z,t ) = at
si H ( t ) d t = atPPH(t)
Entering eqns. 60 and 61 into eqn. 59 yields,
(61)

The analytic solution can be found in [19], or

a t P H ( t )= RCatH(t)
at = RCatP-'
n
a = d2a/dx2 (62)
Now, the conventional method is applied to solve the
ordinary differential equation of eqn. 62. All the eigen-
values of eqn. 62 with positive real parts must be
dropped. Therefore,
(57)
at = ah exp(-zdXZP-1/2) (63)
where 1, is the modified Bessel function of the first
kind of the zeroth order. v(z,t ) = ai exp(-z&&?P-1/2)PH(t) (64)
Let us take some numerical values to compare these
results. Suppose R = 0.002Q/m, C = 5 x 10-l2F/m, L =
z(x,t ) = exp(-zJRCF-1/2)P1/2H(t) (65)
5 x 10M6H/m,G = 0 to cancel the second term in a; = [0,. . . ,o, 2m,. . . ,am] (66)
eqn. 55. The Haar solution of eqn. 54 is obtained via Note that
the inverse Haar transform with the known coefficient
vector, and displayed in Fig. 14. m = 16 has been aiP = [I,0,.. . ,o] (67)
applied. The attenuated Bessel solution is evaluated by Eqns. 66 and 67 are set up to satisfy the boundary con-
applying subroutine Bessel (alpha, X) in MATLAB, dition with a unit step voltage applied at x = 0. That
and drawn in Fig. 14. The difference between these means,
curves is very small. This confirms that the Haar u ( ~ , t=) ahPH(t) = ho(t) = 1 (68)
approach is applicable to distributed-parameter sys- For any fixed distance x, say d [ R q = 1142,
tems.
at = ah exp(-P-'/2//z/z) (69)
l
T\
can be evaluated by applying MATLABs function
2 09 -'\ expm (.). Then v(x, t ) and i(x, t ) are obtained simply by
matrix multiplication.
E 00-
$ 07- The analytic solutions have been described in [19] as
.- 06-
U

6 05-
Q
0 04-
c
? 03-
L

U 02- --
0 11 where erfc (.) denotes the complementary error func-
0 5 10 15 20 25 tion.
time,ms The Haar solution and the Fourier solution have
Fig. 14 Huur and Bessel solutions of transmission line current been drawn together in Fig. 15 for comparison. It
shows that the Haar approach provides excellent
results in numerical evaluation, even if a small number
Example 3: Consider a unit step voltage is applied to a m is used. In both examples, we let m = 16. The evalu-
leakage-free noninductive cable. The voltage v(x, t ) and ated curves will be closer and closer to the real values,
the current i(x, t ) are required to be found. if a larger m is assigned.
92 IEE Proc -Control Theory A p p l , Vol 144, No I , January 1997
 9 DAUBECHIES, I.: ‘The wavelet transform, time-frequency local-
ization and signal analysis’, ZEEE Trans. Znfor. Theory, 1990, 36,
0.61 pp. 961-1005
10 STRANG, G.: ‘Wavelet transforms versus Fourier transforms’.
Research notes, MIT, 1992, pp. 1-18
11 STRANG, G.: ‘Wavelet and dilation equations a brief introduc-

“1y
tion’, S I A M Rev.,1989, 31, pp. 614-627
12 PENTHLAND, A.P.: ‘Interpolation using wavelet bases’, IEEE
Trans. Pattern Analysis & Machine Intelligence, 1994, 16, (4), pp.
410414
13 REGALIA, P.A., and MITRA, S.K.: ‘Kronecker products, uni-
, , , , tary matrices and signal processing applications’, S I A M Rev.,
0.1
0’
0 0.1 0.2 0.3 0.4 0.5
2
0.6 0.7 0.8 0.9
- 1.0
1989, 31, (4), pp. 586-613
14 BARTLS, R.H., and STEWART, G.W.: ‘Solution of the matrix
AX + XB = C’, Commun. A C M , 1972, 15,
15 HSIAO, C.H.: ‘The design and manufacturing of a secant correc-
tor for some tracking equipments’. Third conference on Science
time,lt/x RC is normalised t o 1 and technology of national defense, Chung Cheng Institute of Sci-
Fig. 15 Haar and erfc solutions of transmission line problem ence and Technology, Tao-Yuan, Taiwan, ROC, 15 October
1994, pp. 951-957
16 VIADYANATHAN, P.P.: ‘Multirate system and filter banks’
5 Conclusions (Prentice-Hall, Englewood cliffs, NJ, 1993)
17 CHANG, F.Y.: ‘Waveform relaxation analysis of RLCG trans-
mission lines’, IEEE Trans. Circuits Syst., 1990, 37, pp, 1 3 9 4
The Haar wavelet orthogonal functions and their inte- 1415
gration matrices have been introduced to solve the 18 CHANG, F.Y.: ‘Transient analysis of lossy transmission lines
lumped and the distributed-parameter systems. For the with arbitrary initial potential and current distributions’, , 1992,
39. DV. 180-198
lumped case, the Kronecker product with some existing 19 CHENG, D.K.: ‘Analysis of linear systems’ (Addison-Wesley,
subroutines may be applied to solve for F in eqn. 34. New York, 1961)
For the distributed case, the involved equations are
usually irrational and solved by the transform method 8 Appendix: Algorithms to solve eqn. 34
followed by the classic approach. The universal tool is Eqn. 34 is given as
the operational matrix P for the integrals of the Haar
wavelets, which is positive definite for any order of m. F - A F P = GI (72)
Therefore, its inverse and square root exist. where the unknown n x m matrix F is required to be
The theoretical elegance of the Haar approach can be solved. As we explained before, either the Kronecker
appreciated from the simple mathematical relations and product or the subroutine lyap (A-’, --P, A-’G1) may be
their compact derivations and proofs. The accurate and applied to evaluate F. Studying the subroutine Zyap in
fast capabilities of the method have been demonstrated MATLAB, we find it is also based on the Kronecker
with numerical examples. Among the well known product. As shown in eqn. 35, A 0 Pt is dimensioned
wavelets, the Haar wavelet is the simplest one. This as nm x nm square matrix. When m or n, or the prod-
new approach, formulated for the Haar wavelet, can uct nm is large, the calculation of the inverse of an nm-
also be extended to other wavelets. Compared with the square matrix is a big task even for the modern com-
Fourier method, the Haar wavelet approach apparently puters.
has two advantages: (i) it is particularly suitable for In this Appendix, we are trying to take advantage of
analysing systems involving abruptly varying functions, using the particular property of P to avoid the inver-
and (ii) it is a computer-oriented method, because no sion of a big nm-square matrix directly.
imaginary numbers are involved in the calculation. As shown in eqn. 24, the m-square matrix P, may be
divided in four submatrices.
6 Acknowledgment

This research project is partially supported by Varitro-

nix Ltd (Hong Kong) and partially by the National Pam = Pm/2 Pbm = - &Hm/2
Science Council (Taiwan, Republic of China). 1 H-1 (73)
Pcm = G m / 2 Pdm = null matrix
7 References Accordingly, F and GI may be divided into two parts
with the following definitions:
1 CHEN, C.F., and HSIAO, C.H.: ‘A state-space approach to n a
Walsh series solution of linear systems’, Int. J. System Sci., 1965, F = [Pam Fbm] G I = [Gam Gbm] (74)
6, (9), pp. 833-858
2 CHEN, C.F., TSAY, Y.T., and WU, T.T.: ‘Walsh operational Entering eqns. 73 and 74 into eqn. 72, we have
matrices for fractional calculus and their application to distrib-
uted systems’, J. Franklin Inst., 1977, 303, pp. 267-284 Fam - A[FamPam + Fbrnpcm] = Gam (75)
3 HWANG, C., and SHIH, Y.P.: ‘Laguerre operational matrices
for fractional calculus and applications’, Znt. J. Control, 1981, 34,
pp. 557-584
Fbm = Gbm AFamPbm + (76)
4 KING, R.E., and PARASKEVOPOULOS, P.N.: ‘Parameter Substituting eqn. 76 into eqn. 75 yields
identification of discrete time SISO systems’, Int. J. Control, 1979,
30, pp. 1023-1029 Fam - AFamPm/, - A[Gbm + AFamPbmlPcm = Gam
5 CHANG, R.Y., and WANG, M.L.: ‘Legendre polynomials
approximation to dynamical linear state space equations with ini- Fa, - [I + (A/2m)2]-1AFamPm/2
tial and boundary value conditions’, Znt. J. Control, 1984, 40, pp.
215-232 + +
= [I ( A / a ~ ~ ) ~ l - ~ ( GAGbmPcm)
am (77)
6 PARASKEVOPOULOS, P.N.: ‘Chebyshev series approach to This completes the first step of elimination, since
system identification, analysis and optimal control’, J. Franklin
Inst., 1983, 316, pp. 135-157 eqn. 77 has the same form as eqn. 7 2 but the dimen-
7 PARASKEVOPOULOS, P.N., SPARCIS, P.D., and MON- sion has been reduced by half. Following the same pro-
ROURSOS, S.G.: ‘The Fourier series operational matrix of inte- cedure, we have the following:
gration’, Int. J. System Sci., 1985, 16, pp. 171-176
8 HAAR, A.: ‘Zur Theorie der orthogonalen functionen Systeme’, a
Math. Ann., 1910, 69, pp. 331-371 1 Fbm/2 1 (78)
IEE Proc.-Control Theory A p p l , Vol. 144, No. I , January 1557 93
where
A1 = A, A2 = [I + ( A / ~ v I ) ~ ] - ~ A(81)
Similarly,
Fam/2 - A2[Fam/2Pam/2 + Fbm/ZPcm/2] = Gam/?, (82)
Fbm/Z = G6m/2 f A2Fum/2Pbm/2 (83)
-1G
Fum/2 - A3Fam/2Pm/4 =[I + ( A Z / m ) 1
2
am12

+ A3Gbm/2Pcm/2 (S4)
A3 = [I +(A~/vI)~]-'A~ (85)
The above steps can be repeated with the following
recursive formulae:
Fb(2m/2c) = G b ( 2 m / z 2 )f A z F ~ ( 2 m / Z ~ ) P b ( 2 m / 2 ~ )
2 = 1 , z . . . , ( J - 1) (86)

94 IEE Proc.-Control Theory Appl., Vol. 144, No. I , January 1997