Residue reduction of one polynomial modulo another is defined similarly to residue reduction for integers. A polynomial F(s)F(s) has a residue polynomial R(s)R(s) modulo P(sP(s) if, for a given F(s)F(s) and P(s)P(s), a Q(S)Q(S) and R(s)R(s) exist such that

with degree{R(s)}<degree{P(s)}degree{R(s)}<degree{P(s)}. The notation that will be used is

For example,

The concepts of factoring a polynomial and of primeness are an extension of these ideas for integers. For a given allowed set of coefficients (values of x(n)x(n)), any polynomial has a unique factored representation

where the Fi(s)Fi(s) are relatively prime. This is analogous to the fundamental theorem of arithmetic.

There is a very useful operation that is an extension of the integer Chinese Remainder Theorem (CRT) which says that if the modulus polynomial can be factored into relatively prime factors

then there exist two polynomials, K1(s)K1(s) and K2(s)K2(s), such that any polynomial F(s)F(s) can be recovered from its residues by

where F1F1 and F2F2 are the residues given by

and

if the order of F(s)F(s) is less than P(s)P(s). This generalizes to any number of relatively prime factors of P(s)P(s) and can be viewed as a means of representing F(s)F(s) by several lower degree polynomials, Fi(s)Fi(s).

This decomposition of F(s)F(s) into lower degree polynomials is the process used to break a DFT or convolution into several simple problems which are solved and then recombined using the CRT of Equation 9. This is another form of the “divide and conquer" or “organize and share" approach similar to the index mappings in Multidimensional Index Mapping.

One useful property of the CRT is for convolution. If cyclic convolution of x(n)x(n) and h(n)h(n) is expressed in terms of polynomials by

where P(s)=sN-1P(s)=sN-1, and if P(s)P(s) is factored into two relatively prime factors P=P1P2P=P1P2, using residue reduction of H(s)H(s) and X(s)X(s) modulo P1P1 and P2P2, the lower degree residue polynomials can be multiplied and the results recombined with the CRT. This is done by

where

and K1K1 and K2K2 are the CRT coefficient polynomials from Equation 9. This allows two shorter convolutions to replace one longer one.

Another property of residue reduction that is useful in DFT calculation is polynomial evaluation. To evaluate F(s)F(s) at s=xs=x, F(s)F(s) is reduced modulo s-xs-x.

This is easily seen from the definition in Equation 4

Evaluating s=xs=x gives R(s)=F(x)R(s)=F(x) which is a constant. For the DFT this becomes

Details of the polynomial algebra useful in digital signal processing can be found in [1], [4], [5].

The Z-transform of a number sequence x(n)x(n) is defined as

which is the same as the polynomial description in Equation 1 but with a negative exponent. For a finite length-N sequence Equation 18 becomes

This N-1N-1 order polynomial takes on the values of the DFT of x(n)x(n) when evaluated at

which gives

In terms of the positive exponent polynomial from Equation 1, the DFT is

where

is an NthNth root of unity (raising WW to the NthNth power gives one). The NN values of the DFT are found from X(s)X(s) evaluated at the N NthNth roots of unity which are equally spaced around the unit circle in the complex ss plane.

One method of evaluating X(z)X(z) is the so-called Horner's rule or nested evaluation. When expressed as a recursive calculation, Horner's rule becomes the Goertzel algorithm which has some computational advantages especially when only a few values of the DFT are needed. The details and programs can be found in [7], [2] and The DFT as Convolution or Filtering: Goertzel's Algorithm (or A Better DFT Algorithm)

Another method for evaluating X(s)X(s) is the residue reduction modulo (s-Wk)(s-Wk) as shown in Equation 17. Each evaluation requires NN multiplications and therefore, N2N2 multiplications for the NN values of C(k)C(k).

A considerable reduction in required arithmetic can be achieved if some operations can be shared between the reductions for different values of kk. This is done by carrying out the residue reduction in stages that can be shared rather than done in one step for each kk in Equation 25.

The NN values of the DFT are values of X(s)X(s) evaluated at s equal to the NN roots of the polynomial P(s)=sN-1P(s)=sN-1 which are WkWk. First, assuming NN is even, factor P(s)P(s) as

X(s)X(s) is reduced modulo these two factors to give two residue polynomials, X1(s)X1(s) and X2(s)X2(s). This process is repeated by factoring P1P1 and further reducing X1X1 then factoring P2P2 and reducing X2X2. This is continued until the factors are of first degree which gives the desired DFT values as in Equation 25. This is illustrated for a length-8 DFT. The polynomial whose roots are WkWk, factors as

where a2=ja2=j. Reducing X(s)X(s) by the first factoring gives two third degree polynomials

gives the residue polynomials

Two more levels of reduction are carried out to finally give the DFT. Close examination shows the resulting algorithm to be the decimation-in-frequency radix-2 Cooley-Tukey FFT [7], [2]. Martens [3] has used this approach to derive an efficient DFT algorithm.

Other algorithms and types of FFT can be developed using polynomial representations and some are presented in the generalization in DFT and FFT: An Algebraic View.