6.10.17 a0xn + a1xn1 can be written A0xn + A1xn where a1 = A1 and a0 = A0 + A1.

What is the connection between the coecients of a0xn + a1xn1 + a2xn2 and A0xn + A1xn + A22xn? What about a0xn + a1xn1 + a2xn2 + a3xn3 and A0xn + A1xn + A22xn + A33xn? Generalize and prove that all ARMA equations can be expressed as dierence equations.
First, lets prepare a list of formulas for the nite dierences. (0x)n (1x)n (2x)n (3x)n (4x)n (k x)n = ( )n ( )n1 = (1x)n (1x)n1 = (2x)n (2x)n1 = (3x)n (3x)n1
0x 0x

= xn = xn xn1 = xn 2xn1 + xn2 = xn 3xn1 + 3xn2 xn3 = xn 4xn1 + 6xn2 4xn3 + xn4 =
k
i

= (1 z1)k xn

=0

(1) k i x
i n

and so on. Now we can try the specic cases.

A0x + A1x + A22x = A0x + A1 (x x 1) + A2 (x 2x 1 + x 2) = (A0 + A1 + A2) x (A1 + 2A2 ) x 1 + A2x 2

n n n n n n n n n n n n

from which we deduce

a0 = A0 + A1 + A2 a1 = (A1 + 2A2) a2 = A2
which can be written in matrix form. a0 1 1 1 a1 = 0 1 2 a2 0 0 1

A A A
0 1 2
n

These formulas enable us to convert every dierence equation of order 2

y = A0x + A1x + A22x

n n

into an equivalent ARMA lter

y = a0x + a1x 1 + a2x 2 by telling us how to nd a0 , a1 , a2 given A0 , A1 , A2 .

n n n n

Unfortunately, the exercise asks for us to do precisely the opposite, namely to convert ARMA lters into dierence equations, i.e. to express a0, a1 , and a2 in terms of A0, A1 and A2. To do this we have to solve the above equations (which is simple by substitution from A2 backwards to A0 ) or equivalently to invert the above matrix (which is simple due to the matrix being upper triangular).

A0 = a0 + a1 + a2 A1 = (a1 + 2a2) A2 = a2 Remarkably the equations for the A are of the same form as those for the a. A0x + A1x + A22x +A33x = A0x + A1 (x x 1) +A2 (x 2x 1 + x 2) +A3 (x 3x 1 + 3x 2 x 3) = (A0 + A1 + A2 + A3 ) x (A1 + 2A2 + 3A3) x 1 + (A2 + 3A3) x 2 A3x 3
n n n n n n n n n n n n n n n n n n

The second order case was relatively trivial, so lets nd the formulas for the third order case.

i.e.

a0 a1 a2 a3
or in matrix form.

= = = =

A0 + A1 + A2 + A3 (A1 + 2A2 + 3A3) A2 + 3A3 A3

1 1 0 0 1 2 1 0 1 3 3 1

a a a a

0 1 2 3

1 0 = 00

A A A A

0 1 2 3

Once again the matrix is triangular and the equations can be solved by substitution from the last back towards the begining. Solving we nd A0 = a0 + a1 + a2 + a3 A1 = (a1 + 2a2 + 3a3) A2 = a2 + 3a3 A3 = a3 once again of the same form as the original equations. We have proved that we can convert any ARMA lter of order 3 into dierence equation form. It would be straightforward for the reader to verify that for the fourth order case the following holds. a0 1 1 1 1 1 A0 a1 A1 0 1 2 3 4 a2 = 0 0 1 3 6 A2 a3 A3 0 0 0 1 4 a4 A4 0 0 0 0 1 Once again the matrix is upper triangular, and once again it is its own inverse. This time the matrix is large enough to start revealing its secrets. Disregarding the minus signs this matrix is simply Pascals triangle transposed, with the familiar binomial coecients running down its columns. We can jump now to the general case. Since we have Ax = A (1 z1) x

=
k

A
i

k k i

=
k

(dening the binomial coecient to be zero when i > k), we can identify the coecients. (1) a = (1) k i A
i i k

(1) k i Ax
i k n

=0

(1) k i x
i n

The conversion matrix is obviously triangular, and can be shown (after judicious use of properties of the binomial coecients) to be its own inverse. The identity to be exploited is
k

j (1) k
k

k = l

j,l

where it is sucient to take the sum over all k for which the upper index of the binomial coecients is greater than or equal to the lower index. So in the general case A = (1) k i a
i i k

and hence for any given ARMA lter with coecients a

n n n n L

y = a0x + a1x 1 + a2x 2 + . . . + a x

n

we can always nd an equivalent dierence equation

y = A0x + A1x + A22x + . . . + A 3x

n n n n L L

and vice versa, and the coecients are related by the relations found above.