Walsh 0 erational Matrices for Fractiona Calculus and Their f Application to Distributed Systems

by c. F. CHEN Department and

and Y. T. TSAY
Engineering, University of Houston, Houston, Texas

of Electrical

T. T. wu

Engineering Massachusetts

and Applied

Physics

Department,

Harvard

University,

Cambridge,

ABSTRACT:

The Walsh operational matrix for performing integration and solving state equations is generalized to fractional calculus for investigating distributed systems. A new set of orthogonal functions is derived from Walsh functions. By using the new functions, the generalized Walsh operational matrices corresponding to J/s, J(s+ l), e- and eMJ etc. are established. Several distributed parameter problems are solved by the new approach.

I. Introduction

Since Corrington (1) constructed Walsh tables for solving high-order differential equations and Chen and Hsiao (2) developed the Walsh operational matrix for solving state equations, the Walsh operational method has been successfully applied to various problems such as time domain synthesis (3) in system theory; piecewise gain determination in optimal control (4) and the direct method of solution of variational problems (5). The unique feature of the operational matrix of Chen and Hsiao is that it performs like an integrator in the time domain or l/s in the Laplace domain. We naturally ask: can this operational matrix be generalized to fractional calculus? In other words, are there Walsh operational matrices corresponding to irrational functions such as Js, J(s+ 1) or transcendental functions such as e --s e+ etc.? , This paper will investigate this problem and develop several Walsh matrices for fractional calculus, then apply these matrices for solving distributed parameter systems.

II. Review of Operational

Matrix for Znteger Integration

It is known that Walsh functions 4i(t) are a set of square waves which are orthonormal. Figure 1 shows the functions from & to &5 in the dyadic order. For an arbitrary function f(t), we can expand it into Walsh series, if the 267

C. F. Chen,

Y. T. Tsay and T. T. Wu

FIG. 1. Walsh functions. function is absolutely integrable in [0, 1). We write: (I)

f(t) = EOG&(0 where G = are determined I0 &(t)f(r) dr

(2)

such that the following integral square error is minimized, i.e. E= [f(t) - nfO GbW12 dt. (3)

For example, if we expand f(r) = t, the result is f(t) = #Jo(t) -ad*(t) -Q&(f) -+#Q(f) + . . . (4)

Partial sums of (4) form stairwise waves that approximate the original function t, as shown in Fig. 2. If we perform integration on the Walsh functions, various triangular waves are obtained. Evaluating the Walsh coefficients for these triangular waves, 268
Journal of The Franklin Institute

Walsh Operational Matrices for Fractional Calculus

lzc1/2~o-1/4~, 1120, -lIG#, -1189, I!!z!c-

Fro. 2. Decomposing a ramp into Walsh series.

gives the following formula: (Typically, with no loss of generality, only the first eight Walsh waves are considered.)
f _;i 1 -%

-4
0 0

0
0

0 0
0

$ Q
=

0 0

-Q
0

4
0

--ii

0 0 0 4 0 Q 0 0 ____________________-. ,__________________, 0 0 0 0 ii 0 0 0 0 0 0

(5)

A
0 0

0 0

0 0 0

0 0 0

0 0

0
0

l iz
0

or

The subscript means the dimension taken. It is preferable to take 2*, where s1 is an integer, as a dimension number. Note that the ith row of the square matrix shows the Walsh series coefficients of the function J4i-1 dt, i = 1, 2, . . . . For example, the first row of (5) gives the Walsh coefficients of Jc$~dt. The eight coefficients may be found in the first eight terms of (4), of course. The square matrix G performs as an integrator and is called the Walsh operational matrix for integer integration. It is interesting to note that if G is partitioned into four equal parts as shown; the upper left corner of GCBxsjis identical to GC4X4j;and the upper left part of Gc4x4j is Gc2x2j. Therefore, the regularity of the structure of the G matrix enables us to enlarge the G matrix.
Vol. 303, No. 3. March 1977

269

C. F. Chen, Y. T. Tsay and T. T. Wu We write the general operational matrix for integration as follows:

1:
2:

i
i : : (61

___+___: : 1I -- 2 i -; (3 : . . . . : mIm I ____:____-::.____________~

: -- 5 i O(B) / = m G(?nXrn) 5 m : 2m (4) __________________________:__ __________:

f_I (9) ; O(T) i m __~~~~~~~~~~~~~~~~~~~~~~~~____________.________~~~~~ &I@) : %

The operational matrix (6) perform integration for integer calculus nicely. Therefore, it can be used to solve linear differential equations effectively. Chen and Hsiao demonstrate its power in solving state equations, in synthesizing transfer functions in the time domain, and in finding time-varying gains for optimal control. We will generalize this matrix for fractional integration or non-integer calculus. III. A Set of New Functions Since Gc,,,,,,,, corresponds to l/s, if we want to study 4s for example, what we have to effect is an extension of the operational matrix idea: (1) to invert the operational matrix in order to find a matrix corresponding to the differentiator; (2) to take the square root of the differentiator matrix thus obtained for finding a new matrix equivalent to Js. However, difficulties arise: The inversion of a large matrix is not a trivial matter if m is large! Secondly, for taking the square root of a matrix, we naturally think of diagonalization of the matrix in question. It is easy to find that all the eigenvalues of the operational matrix are clustered together. There are m repeated roots and the value is 1/2m. Therefore, diagonalization of this problem is not a routine matter. For overcoming these difficulties, we develop a new set of orthogonal functions. There is no loss of generality if we use GcaX4)for deriving the new set. Consider:

(7)

Or, in general,

[I 1
270

&,,, dt = G(,~,,&+

(8)

Journal

of The Franklin

Institute

Walsh Operational Matrices for Fractional Calculus

We define a set of new functions which relates the set of Walsh functions by the following formula:

or, in general,

&I, = W(?nXl?&(?n,

(9a)

where the $i are called the block pulse functions with unity height and l/m width. They are shown in Fig. 3 for the first four functions. The square matrix is in (9), denoted by WC4X4j, called the Walsh matrix. The relation between the matrix and Walsh wave forms is shown in Fig. 4. From the figure we see that it is very easy to construct the Walsh matrix because the elements of the matrix correspond directly to the magnitude of the wave forms. One of the properties of the W matrix is that W?LXrnj= mL, or W;nXrn) =1_ WCrnXrn). m Substituting (9a) into (8), yields
,

(10)

(11)

I0

WC mxm,kn,

dt

&w,,,W~mxm,~w.

(12)

FIG. 3. Block pulse functions. 271

Vol.

303. No. 3, March

1977

C. F. Chen, Y. T. Tsay and T. T. Wu

-1

-1

-1

-1

-1

-1

FIG. 4. Walsh functions and Walsh matrix.

Therefore,
f +%n) df = W;~,,,G~,,,,w~,,,,r~r(,,.

I0

(13)

We define the following: W;~xrn~G~rnxrn~W~rnxrn~ p H(,,,,. Combining it with (11) gives H( mxm) =;WGW. Evaluation of the similarity transformation yields 1 1 $
. . ;,

(14)

(15)

4 1
0 1 0 m .
. 0

1 1 1
. . (j

... ... ...

... ... ;-_:

1 1 1
. . 1

3 0
. . ;,

J&mxm)

=-

(16)

Matrix H is the operational matrix as expressed in the block pulse function domain. Of course, this matrix can be directly obtained from integration of Fig. 3, by reading the magnitudes as shown in Fig. 5. It is seen that the set of the block pulse functions is more fundamental than the Walsh functions: The operational matrix for integration shown in (16) is in

272

Journal of The Frankfin Institute

Walsh Operational Matrices for Fractional Calculus

r
+
1 1 1

3
0 5. Block pulse functions and their integrations.

1 2 _

FIG.

terms

of block pulse functions and is simpler than the operational derived from Walsh. Inspecting the H matrix, we make the following decomposition: =;t;(&, H( WlXm) where
0

matrix

+Q~mx,,+Q~~!mxm,+...+Q~~~,>

(17)

1 0 0

0 1 0

0 0 1

... ... ...

..*.

0 0 0
.

0; : : .
I

0 0

1(,-l)

Qonxm)= .......
.....

0; ,__;____________

(18)

d d d d ::: i
and it is easy to see the following important
_

o/o.. .
i<m

property:
(19)

Qt mxm)

E 3
Oi: 0

0 : I ____:__rn_?for

and

Q;mxm) =qm,
Using (19) and (20), we write (17) as

for

izm.

(20)

=;

(--%,,+ (I, -Qcmx,J1>

-Qwen,)-.

=& CL +Q(ntxn,,K
Vol. 303, No. 3. March 1977

(21)
273

C. F. Chen, Y. T. Tsay and T. T. Wu We rewrite (17) by expressing Hc,,,) as a polynomial

=

of Qcmx,,,,

H(mxm)
where

hn(Qmxm> . . F-l).

(22)
(23)

h&)=~($+x+r+.

If q is an eigenvalue of Q~mxmj, it is known eigenvalue for Ht,,,, is

(6) that the corresponding

h=h,(q)=-l-%.
2m l-q

(24)

However, the eigenvalues of Qcmx,,,) in (18) are all zeros. From (24), we obtain
h=h,,,(O)=&

Therefore, we state that the eigenvalues of H(,,,,,) is 1/2m with multiplicity m. Equations (21) through (24) will be used repeatedly to derive various generalized operational matrices. IV. Operational Matrix for Differentiation Theoretically, it should not be very difficult to find the operational matrix for differentiation, when that for integration is known. However, it is much simpler to find the inversion in the block pulse domain. Let us denote the operational matrix for differentiation as Bc,xmj. We have
B (mxm)=HTAxm)=2m(L

+Q(mxm&,-Q(mxm)) ..i-lYelQK&,,,
. (26)

= 2m(I,

-2Q~,.,,+2Q&,,+. (-l)iQ&,X,,

For m = 4, we have

(27)

Similarly transformating matrix for differentiation,

back to the Walsh domain denoted by Dcmxmj:

yields the operational

274

Journal of The Franklin Institute

Walsh Operational Matrices for Fractional Calculus For m = 4, we have

(4X4) = 8

1 I
0

01 00 0

-1

-2

(29)

-12

In general, the formula is

.________:______________~_______________,__________ , : rn
;_____-mt__ ;.
.

_____:
I

-% 1 :

DC,,,)

= 2m

....I; : _______I_______.~___________7 i -2I,p, -I(?)

i

(30)

41,?:,

_____________________________:___________

From (26) the eigenvalue, eigenvalue, q, of Qcmx,,,)

b

h-,

of matrix H&,)
m-l

can be expressed

as the

=4m

++ 2 (+.ji
i-l >

(31)

or b=2m

i+q'

(32)

We could use (30) directly to evaluate the operator of differentiation or use (26) and (28) indirectly to deal with it. When the dimension number m is higher, the former is not as simple as the latter. V. Operational Matrices for Fractional Calculus Now we try to find the operational matrix for fractional differentiation. general form of (32) could be written as follows:
bl =

The

(&?)l

(33)

where 1 is an integer. Equation (33) can be expanded into polynomial of q and terminated at qm-. As the result, Eq. (33) becomes
b = (2m)p,,,(q)

(34) 275

Vol.

303, No. 3, March 1977

C. F. Chen, Y. T. Tsay and T. T, Wu where pl,,, is the polynomial of order m- 1 for l/1 differentiation. operational matrix for l/1 differentiation from (26) is given by B (mxm)= CW"h,dQ~mxm,). Thus the (35) matrix (36)

In the Walsh domain, the corresponding l/1 differentiation operational is 111 D( mxm)= (2m) W;ntxm~~,,m(Q~rnxrn))W~rnxrn). To illustrate the procedure, we try to find the operational By choosing m = 4, we have &n(q) = Then (35) becomes &x4) = &? x 4)(I (4x4r-Q~4x4)+tQf4x4)-iQ:4x4)+. However, Qi=O Therefore, we have
Bhx4j = J@)[I, 4x4) - Q(4x4)++Q:4x4) -$Qf4x4d . .I.
p2,4(q) =

matrix of Js.

( 1
=

l-q

l-q++q2-;q3+.

..

1+q

(37)

(38)

for

ir4.

In the Walsh domain, using (36) gives 3 1 3 1

(40) 3

-1

17

-5

For finding the operational matrix of fractional integration, reasoning similar to the fractional differentiation case. Generalizing (24), yields

we can use the

(41)

where n.,,, is the polynomial of order m - 1 for l/1 integration.

The operational
276

matrix for l/l-integration

in terms of the block pulse

Journal of The
Franklin Institute

Walsh Operational Matrices for Fractional Calculus

function is given by
l/I Wmxm, 1 =1/1

@ml

w,m(Q~mxm,)

(43)

and the corresponding is easily found as

l/l-integration

operational

matrix in the Walsh domain

111 GM,,,, =& 1

=m(2m)

W;~,,,y,,(Q~,xrn))W~,xrn) ,xm,7/1,m(Q(mxm,)W(,,m) WC

(44) (45)

Illustratively, i.e. Then

we try to find the operational

matrix for l/ Js, assuming m = 4,

(46)

X,,(g) = Y2.4k.I) 1+ 4 +ts +tq3. = : &,x4) = +I Js

(4x4) + Q~4x4) +fQ;4x4)

=J80011 0 From (45), we have

1

I
1 0
17 5 I -1 3

+4Q?4x4,)

1 1

1 1 0

t 1 1

G;4x4)

1 = 4.B

I WC4x4jH;4x41Wt4x4j -5 7 I 5 -1 -3 1 3 -1 -5 I 1 5 (47)

1 =1612

Of course, this result can be obtained directly from inverting the matrix of the first example. VI. Application to Distributed Systems

For many engineering systems described by partial differential equations or time-varying differential equations, when evaluated in the Laplace domain, fractional functions or transcendental functions of s result. For example, problems in thermal processes, in hole diffusion of transistors, in electromagnetic devices, in transmission lines and in percolation processes often have mathematical models involving J(s), J(s2+ l), eeJ etc. To find their inverse
Vol. 303. No. 3. March1977 277

C. F. Chen, Y. T, Tsay and T. T. Wu Laplace transform is not a trivial matter. There is a long history in attacking these types of problems. To mention a few: Pades approximation (6); Carlson and Halijaks approach (7) using regular Newtons process; Lerners work on partial analogue approximation (8); Kiloneitreva and Netushils technique (9) developing special functions; Chen and Chius use the Fast Fourier transform (10). The operational matrix method for fractional calculus developed in this paper is a general and powerful approach to the distributed problem. We now apply the method to various distributed systems. Bessel functions One of the well known functions

F(s)=

$+

1)

(48)

can be constructed as the Laplace transform of the solution of a time-varying system. We would like to find its inverse Laplace transform as the first application. Rewrite

F(s)=J(sl+l)=J(s:+l)(49)
-f&F,(s)+. Suppose we are interested in the solution for OStc8. Time scaling must be used. Note that the general rule of scaling is based on the following for where a! is the scaling factor. For this case, (Y= 8 t-+cwl

(50)

where
Fl

_s. 0 =xW8)*+
48 8

11 *

(52)

For m = 8, substituting

(32) into (Sl), yields

20 - 4)
=J(S-6q+5q2) 278 JournalThe of Franklin

(53)

Institute

Walsh Operational Matrices for Fractional Calculus

Taylor series expansion at q = 0 gives

F,(q) = 0.8944 - 0.3578q - 0.5009q20.3578q3 .. (54) -0.0429q4+0.2284q5+0.2905q6+0.1346q7+.

Replacing q by Q gives the operational matrix of this problem in the IJ domain. Note that l/s in (51) should be expressed by the JI functions, i.e. $11 Consequently, (48), becomes =z-!F, z-1J(s: 1) s 2 08
il.8944

1 1 1 1 1 1 l]&@.

(55)

-0.3578

...

0.1346

=[l

1 1 1 1 1 1 l]

-0.3578

$cg)

0.8944 = 0.8944+,+ 0.5367q1 + 0.03584& - 0.3220J/3 (56)

-0.3650~4-0.1365J/,+0.1721~6+0.3067~7.

The exact solution and (56) are compared in Fig. 6. If we use the W matrix (9a>, we can transform (56) in terms of the Walsh functions. Of course, the two approaches yield the same answer.
Generalized exponential function s-;

Consider the following function

F(s) =-

J(s)+1
F,(s) - i .

(57)

We rewrite it as
F(s) = L__.lo

1+L

JS

Assume m = 4, we are interested in the interval 0 5 t < 2.4. The scaling factor is s + s/2.4. Then we have

Vol. 303, No. 3, March

1977

279

C. F. Chen, Y. T. Tsay and T. 1. Wu

FIG. 6. Solution of Bessel function

0.25

FIG. 7. Solution of generalized exponential

function.

280

Journal of The Franklin Institute

Walsh Operational Matrices for Fractional Calculus

Using (32), gives F,(q) = 1 1+o.5477(1+q+aqz+;q3+.

0.2287q . .) 0.09602q3

= 0.6461-

- O.O3341q*-

(58)

Then replacing q by Q, gives F 1(4X4) 0.64611w, =

-0.2287Qc4,4,-0.02341Q;Lw - O.O9602Q&,,.

Finally, we obtain -0.2287 0.6461 Y-IF(s) =[l 1 1 l] 0 0 =[0.6461, -0.03341 -0.2287 0.6461 0
--d;;,, .

-0.09602 (4) (59)

0.6461

0.41074, 0.3840, 0.2880]&+. calculated by the

Figure 7 shows the exact solution and the approximation new method.
Heat conduction

For the heat conduction

problem, we easily arrive at the following function: F(s) = 5 e+.

Find the inverse Laplace transform for 0 5 t C 8. By using the scale s+s/8 and letting m =4, we have si8+iexp F(s) = (-\ii)=iexp(-dz)kiF,(q). (60)

Taylor expanding and evaluating the coefficients, give

(1+ q)-=(

1 - q)-2

10 + q)-3(l - q)-l + (-1+2q)(1+

a3F1(q) T=exp (-V/~)[(l+q)12(1-q)12+(2q-l)(l+q)-4(l-q)-2 - 3(1+ qY4(1 + (29 -

q)-yl-

q)-321

q)-l(1 q)-2(l +-

+ q)-3 - 2(1+ q)-52(1 _

q)-3/2

I)(-5/2)(1+

q)-72(1 - q)-32 + (2q -

1)(1q)-s2(3/2)(1 +

_ q)-5/2]. 281

Vol. 303, No. 3, March lY77

C. F. Chen, Y. T. Tsay and T. T. Wu Then Taylor series

I, = F,(o) 1 +$ F,(O) -q+- a2 F,(O)- q2+$ FIKO* q3 $3 2!a q
2

F,(q)= Replacing q by Q, we finally have

(l+q+$q3)e-l.

-1

e- e-l 0 0

0 e-

-c
0

-1

1
&4)

(61)

0 .O
=0.3679~o+0.7358~l+0.73584~2+0.79714~3.

e- 0

e- e- I

Figure 8 shows the exact solution and the approximation tion problem. Percolation problem It is well known that the percolation same procedure as above. Consider the function

of the heat conduc-

problem is difficult to solve. We use the

F,(s). If m = 4 and 01 t < 8, we substitute s by s/8 for scaling. Then we obtain

=e

,+~+km2+m3+ 2 2! 3!

..*

1
. .I.

=e-~[1+0.5q+0.125q2+0.0185q3+. Consequently, we have rO.6065 1 =[l 1 1 l] I 0 o 0 = 0.6065%+ 282 0.3033 0.6065 0 0

0.07582 0.6065 0.3033 0

0.011231 0.07582 I 0.3033 0.6065 (63)

Institute h4)

0.90984+, + 0.9856$2 + 0.9969*+

Journal of The Franklin

Walsh Operational Matrices for Fractional Calculus

se

I -6

FIG. 8. Solution of heat conduction

problem.

fe -zi

02.

4 1

FIG.

9. Solution of percolation

problem.

Vol 303, Yo. 3, March 1977

283

C. F. Chen,

Y. T. Tsay and T. T. Wu

The comparison Fig. 9.

between

the exact solution

and the approximation

is shown in

VIZ. Conclusion
Fractional functions in general, and the functions Js, J(s + l), e- and eeJ in particular, are inverted by the Walsh operational matrix approach. A new set of orthogonal functions &i(t), which is closely related to the Walsh set 4i(t), is derived. Various operational matrices for the typical functions are established, both in the $ domain and the + domain. The application of these operational matrices for the solution of distributed-parameter systems is demonstrated.

Acknowledgements
The first author wishes to acknowledge the support of the National Foundation grant HES75-20759. To one of the reviewers the authors express their gratitude for valuable suggestions. Science wish to

References
Solution of differential and integral equations with Walsh functions, ZEEE Trans. on Circuit Theory, Vol. CT-20, No. 5, pp. 470-476, Sept. 1973. (2) C. F. Chen and C. H. Hsiao, A state space approach to Walsh series solution of linear systems, Znt. J. Systems Sci. Vol. 6, No. 9, pp. 833-858, 1975. (3) C. F. Chen and C. H. Hsiao, Time-domain synthesis via Walsh functions, ZEE Proc. Vol. 122, pp. 565-570, May 1975. (4) C. F. Chen and C. H. Hsiao, Design of piecewise constant gains for optimal control via Walsh functions, IEEE Trans. on Automatic control, Vol. AC-20, No. 5, pp. 596-603, Oct. 1975. (5) C. F. Chen and C. H. Hsiao, Walsh series direct method for solving variational problems J. Franklin Inst., Vol. 300, No. 4, pp. 265-280, Oct. 1975. (6) 3. L. Stewart, Generalized Pade approximation, IRE Proc. pp. 2003-2008, 1960. (7) G. E. Carlson and C. A. Halijak, Approximation of fractional Capacitors (l/S) by the regular Newton Process, IEEE Trans. on Circuit Theory, Vol. m-11, pp. 210-213, 1964. (8) R. M. Lerner, The design of a constant angle or power-law magnitude impedance, ZEEE Trans. on Circuit Theory, Vol. CT-IO, pp. 98-107, 1963. (9) M. B. Kilomeitreva and A. V. Netushil, Transients in an automatic control systems with irrational transfer functions. Automation and Remote Control. Vol. 203. pp. 359-364, 1965. (10) F. Chen and R. F. Chiu, Evaluation of irrational and transcendental transfer C. functions via the fast Fourier transform, Znt. J. Electron. Vol. 35, No. 2, pp. 267-278, 1973.

(1) S. Corrington, M.

284

Journal of The Franklin

Institute