Approximation and interpolation sum of exponential

functions

Jozsef Dombi
Department of Informatics
University of Szeged, Hungary

1 Learning pliant operator

Recall Dombi operator[2]

1
o(x) = o(x1 , . . . , xn ) = P   1/ (1)
(1xi )
1+ xi

xi [0, 1] If > 0 then o(x1 , . . . , xn ) is a conjunctive operator, and

if < 0 then o(x1 , . . . , xn ) is a disjunctive operator
and 6= 0
From (1)

  n  
o(x) 1 X 1 xi
= (2)
o(x) i=i
xi

Let us suppose, that is known. (In most cases is = 1)

In the pliant concept we use the Dombi operator with sigmoid function

1
i (t) = (3)
1 + ei (tdi )

so i (t) = xi
Substituting (3) into (2)
  
o(x) 1 X X
y= = ei (tdi ) = ai ei t
o(x)

where ai = ei di and i = i

2 Interpolation sum of exponential functions

Our task is now to approximate or interpolate the following function:

(t) = a1 e1 t + a2 e2 t + + an en t (4)

Let us given (t1 , y1 ), (t2 , y2 ), . . . , (t2n , y2n ) (and let us suppose that the abscissas are equidis-
tant i. e. ti+1 ti = h, i = 1, 2, . . . , 2n 1). First we deal with the interpolation case. Because
in (1) we have 2n variables (a1 , a2 , . . . , an , 1 , 2 , . . . , n ) we have to give 2n equations:

y1 = a1 e1 t1 + a2 e2 t1 + . . . + an en t1
y2 = a1 e1 t2 + a2 e2 t2 + . . . + an en t2
.. (5)
.
y2n = a1 e1 t2n + a2 e2 t2n + . . . + an en t2n .

Let us introduce new variables:

pk = ak ek tk , zk = ek h , k = 1 . . . n. (6)

Using this notation one term is in (4):

ak ek tj = ak ek (t1 +(j1)h = ak ek t1 (ek h )j1 = pk z j1 , (j = 1, . . . , 2n), (k = 1, . . . , n).

Now (5) has the following form and we get the equations:

y 1 = p1 + p2 + p3 + . . . + pn
y2 = p1 z1 + p2 z2 + p3 z3 + . . . + pn zn
y3 = p1 z12 + p2 z22 + p3 z32 + . . . + pn zn2
y4 = p1 z13 + p2 z23 + p3 z33 + . . . + pn zn3 (7)
..
.
y2n = p1 z12n1 + p2 z22n1 + p3 z32n1 + . . . + pn zn2n1 .

where p1 , p2 , p3 , . . . , pn and z1 , z2 , z3 , . . . , zn are 2n unknown quantities.

3 Ramanujan solution
Ramanujan find an interesting solution of (7)[1]. In the following we make a detailed construc-
tion of the proof. It is easy to verify that
p1
= p1 + p1 z1 + p1 z12 2 + . . .
1 z1
p2
= p2 + p2 z2 + p2 z22 2 + . . .
1 z2 (8)
..
.
pn
= pn + pn zn + pn zn2 2 + . . .
1 zn

The sum of (8) is:

X pi X X X
() = = pi + pi zi + 2 pi zi2 + . . .
1 zi

Using (8) we get:

n
X pi
() = = y1 + y2 + 2 y3 + n yn1 + + 2n1 y2n + . . . (9)
i=1
1 zi

The left hand side of (9) is a rational expression, and when it is simplified, then its numerator
is an expression of the (n 1)-th degree in , and its denominator is an expression of the n-th
degree in
Q Q Q
n (1zi ) (1zi ) (1zi )
X pi p1 1z1
+ p2 1z2
+ . . . + pn 1zn D()
() = = Q = =
1 zi (1 zi ) N()
i=1 (10)
Y1 + Y2 + Y3 2 + + Yn n1
= = y1 + y2 + y3 2 + + y2n 2n1 ,
1 + Z1 + Z2 2 + Z3 3 + + Zn n

and so

(1+Z1+Z2 2 +. . .+Zn n )(y1 +y2+y3 2 +. . .+y2n 2n1 ) = Y1 +Y2+Y3 2 +. . .+Yn n1

Equating the coefficiens of the powers of , we have

Y1 = y1
Y 2 = y 2 + y 1 Z1
Y 3 = y 3 + y 2 Z1 + y 1 Z2 (11)
..
.
Yn = yn + yn1 Z1 + yn2 Z2 + + y1 Zn1
and
0 = yn+1 + yn Z1 + + y1 Zn
0 = yn+2 + yn+1 Z1 + + y2 Zn
0 = yn+3 + yn+2 Z1 + + y3 Zn (12)
..
.
0 = y2n + y2n1 Z1 + + yn Zn .

From (12) Z1 , Z2 , . . . , Zn can easily be found, and since Y1 , Y2 , . . . , Yn in (11) depend upon
these values they can also be found.
Now, our task is splitting
n
X pi Y1 + Y2 + Y3 2 + + Yn n1
() = = =
i=1
1 zi 1 + Z1 + Z2 2 + Z3 3 + + Zn n

(where Y1 , . . . , Yn , Z1, . . . , Zn are known) into partial fractions, i.e.

q1 q2 q3 qn
() = + + ++ ,
1 r1 1 r2 1 r3 1 rn
and comparing with (1), we see that

p1 = q1 , z1 = r1 ;

p2 = q2 , z2 = r2 ;
p3 = q3 , z3 = r3 ;
..
.
pn = qn , zn = rn .
which is the desired solution.
The question, how we can get this partial fraction form.
The denominator of (10) is:
n
Y
2 n
D() = 1 + Z1 + Z2 + + Zn = (1 ri )
i=1

1
Let us substitude: =

1 1 1 1 1 1
1 + Z1 + Z2 2 + + Zn n = (1 r1 )(1 r2 ) . . . (1 rn )) =

n
1 Y
= n ( ri )
i=1
Multipling both side by n
n
Y
n n1
D () = + Z1 + + Zn = ( ri )
i=1

From D() we build D () polinom, and the roots of D are the ri .

The nominator of (10) is:

n n Qn
j=1 (1 rj )
X X
i1
N() = Yi = pi =
i=1 i=1
1 ri

p1 (1 r2 )(1 r3 ) . . . (1 rn )+
+p2 (1 r1 )(1 r3 ) . . . (1 rn )+
+p3 (1 r1 )(1 r2 ) . . . (1 rn )+
..
.
+pn (1 r2 )(1 r3 ) . . . (1 rn1 )
= Y1 + Y2 + Y3 2 + . . . + Yn n1

From this we get the following linear equation systems:

X
Y1 = pi
X X
Y2 = pi rj
i6=j,
X X
Y3 = pi rj rk (13)
i6=j,i6=k,j6=k
..
.
X Y
Yn = pi rj
i6=j

From (14) we get pi .

pi can be determined by n equation of (7)

if pi is given
pi
ai = .
e i ti

4 Ramanujans example
As an example we may solve the equations:

x + y + z + v + u = 2,

px + qy + rz + su + tv = 3,
p2 x + q 2 y + r 2 z + s2 u + t2 v = 16,
p3 x + q 3 y + r 3 z + s3 u + t3 v = 31,
p4 x + q 4 y + r 4 z + s4 u + t4 v = 103,
p6 x + q 5 y + r 5 z + s5 u + t5 v = 235,
p6 x + q 6 y + r 6 z + s6 u + t6 v = 674,
p7 x + q 7 y + r 7 z + s7 u + t7 v = 1669,
p8 x + q 8 y + r 8 z + s8 u + t8 v = 4526,
p9 x + q 9 y + r 9 z + s9 u + t9 v = 11595,
where x, y, z, u, v, p, q, r, s, t are the unknowns. Proceeding as before, we have
x y z u v
+ + + +
1 p 1 q 1 r 1 s 1 t

= 2 + 3 + 162 + 313 + 1034 + 2355 + 6746 + 16697 + 45268 + 115959 + ...

By the method of indeterminate coefficients, this can be shown to be equal to

2 + + 32 + 23 + 4
.
1 52 + 3 + 34 5

Splitting this into partial fractions, we get the values of the unknowns, as follows:
3
x = , p = 1,
5

18 + 5 3+ 5
y= ,q = ,
10 2

18 5 3 5
z= ,r = ,
10 2

8+ 5 51
u = ,s = ,
2 5 2

8 5 5+1
v = ,t = .
2 5 2
5 New solution
We start with (7):
y 1 = p1 + p2 + p3 + . . . + pn
y2 = p1 z1 + p2 z2 + p3 z3 + . . . + pn zn
y3 = p1 z12 + p2 z22 + p3 z32 + . . . + pn zn2
y4 = p1 z13 + p2 z23 + p3 z33 + . . . + pn zn3 (14)
..
.
y2n = p1 z12n1 + p2 z22n1 + p3 z32n1 + . . . + pn zn2n1 .

Let us build the polinom

z n + s1 z n1 + s2 z n2 + . . . + sn = 0 (15)

The roots of (15) are z1 , . . . , zn . The following wellknown identities are valid.
X
s1 = zi
X
s2 = zi zj
i6=j
..
.
Y
sn = (1)n zi (16)

Based on (14) we show that:

y1 sn + y2 sn1 + . . . + yn s1 + yn+1 = 0
y2 sn + y3 sn1 + . . . + yn+1 s1 + yn+2 = 0
.. (17)
.
yn sn + yn+1 sn1 + . . . + y2n1 s1 + y2n = 0

The k th row is in (17):

yk sn + yk+1 sn1 + yk+2sn2 + . . . + yn+k1s1 + yn+k = 0

Using (14):
X X X X
sn pi zik1 + sn1 pi zik + . . . + s1 pi zin+k + pi zin+k1 = 0

Let us find the coefficient of pj :

pj (sn zjk1 + sn1 zjk + . . . + zjn+k1 ) = 0

because (15) is valid.

From (17) if y1 , . . . , y2n are given we can get si . If si is given, then the roots of (15) are the
solution for zi .

lnzi
i = (18)
h

If zi are given, then from (14) we get pi .

6 Approximation sum of exponential functions

Now we have more than 2n points:

(t1 , y1), (t2 , y2), ..., (tm , ym )

and
(t) = a1 e1 t + a2 e2 t + + am em t
Suppose that the abscissas are equidistant i.e. ti+1 ti = h, i = 1, 2, . . . , 2n 1, and m > 2n.
We build the polinom (15) and using the same construction:

y1 sn + y2 sn1 + . . . + yn s1 + yn+1 = 0
y2 sn + y3 sn1 + . . . + yn+1s1 + yn+2 = 0
.. (19)
.
yn sn + yn+1 sn1 + . . . + ym1 s1 + ym = 0

we get it instead of (17).

The A matrix of (19) and the constant vector b

y1 y2 . . . yn
y2 y3 . . . yn+1
A=

.. .. .. ..
. . . .
ymn ymn+1 . . . ym1

y1
..
b=.
yn
The best approximation of (19) is based on the least square method.

(AT A)s + AT b = 0 (20)

The solution of (20) gives the s values: If s is given we can determine (15), and the roots of
(15) are the zi values, and based on (16) i can be determined.
Now we have to find the ai -values. We use also the least square method:

n
X m X
X n
2
((ti ) yi ) = ( ai ei tj yj )2 (21)
i=1 j=1 i=1

The matrix is:

e1 t1 e2 t1 . . . en t1
e1 t2 e2 t2 . . . en t2
F = ..

.. .. ..
. . . .
e1 tm e2 tm ... e n tm

And the b = (y1 , . . . , yn )

The best approximation of (21) is:

(F T F )a + F T b = 0

where a is unknown.
 








 












Figure 1: (x) = (20.09ex 13.03ex 4.481e1.5x + e1.7x + 0.0608e1.4x )103

References
[1] S. Ramanujan. Note on a set of simultaneous equations. Journal of the Indian Mathematical
Society, 94-94, 1912.

[2] J. Dombi A general class of fuzzy operators, the De Morgan class of fuzzy operators and
fuzziness measures induced by fuzzy operators. Fuzzy Sets and Systems Vol 8., 1982, pp
149-163.