BULLETIN (New Series) O F THE

AMERICAN MATHEMATICAL SOCIETY

Volume 1, Number 6, November 1979

IS COMPUTING WITH THE FINITE FOURIER TRANSFORM

PURE OR APPLIED MATHEMATICS?
BY L. AUSLANDER AND R. TOLIMIERI

HISTORICAL INTRODUCTION by L. Auslander

Let me begin with my view of a bit of history.
Before the Second World War mathematics in the United States was a
servant of the needs of others and mathematicians taught service courses.
Indeed, while A. Weil was teaching at an Eastern university it would be only
a slight exaggeration to say that he was forbidden from presenting proofs in
class and was called on the carpet by a dean for breaking this structure. In
the years after the War, mathematics became a subject in its own right.
Proofs became acceptable, as the creation of the "new math" proved to the
world. Mathematicians were in demand, were men in their own right and no
one's servants.
However, this growth period had a very unfortunate side affect. While
mathematics was becoming a subject in its own right, many of its practi-
tioners wanted to rid themselves of their former servant image. They had felt
denigrated by the service role; so they denigrated service mathematics.
Unfortunately, they lumped together service mathematics and applied
mathematics. And so during this growth period of mathematics, there sprang
up a distinction between pure and applied mathematics. During these years,
the applied mathematicians felt the pure mathematicians looked down on
them, and so the communications between the pure and applied mathemati-
cians virtually dried up.
In this paper we willl show that there is really not much difference between
pure and applied mathematics. Indeed, we will cite instances of pure and
applied mathematicians doing the same or analogous mathematics, but be-
cause of the lack of communication neither knew of the others' work.
With these broad generalities stated, let me try to explain how I came to
the writing of this paper. This may perhaps serve as an example of how the
gap between pure and applied mathematicians can be bridged.
I became interested in the study of the finite Fourier transform because I
needed to know the eigenvalues of the finite Fourier transform. This arose in
the study of the multiplicity of the regular representation of a solvmanifold.
This problem was solved and the solution can be found in [8, p. 95].
Tolimieri, and Tolimieri and I, took up this problem in [18] and [3] and
related the eigenvalue problem of the finite Fourier transform to a certain
algebra of theta functions as discussed in Chapter I of this paper. I felt that

Received by the editors March 1, 1979.

AMS (MOS) subject classifications (1970). Primary 42A68; Secondary 68A20, 68A10, 10G05,
22E25.
© 1979 American Mathematical Society
0 0 0 2 - 9 9 0 4 / 7 9 / 0 0 0 0 - 0 5 0 1 / $ ! 3.75

847
848 L. AUSLANDER AND R. TOLIMIERI

the mathematicians at the IBM Watson Research Center at Yorktown

Heights, New York, might be interested in these results. They were, and they
invited me to give a talk on my work. After the talk, James Cooley was kind
enough to point out that two electrical engineers, McClellan and Parks [16],
and the applied mathematician I. J. Good [11] had written interesting papers
on this subject. I. J. Good pointed out that Gauss had studied and really
solved the problem of the eigenvalues of the finite Fourier transform. All
these ideas are presented in Chapter I.
My interest in the computational aspects of the finite Fourier transform
was aroused by the papers J. Cooley gave me. Tolimieri and I in [3] had
presented a proof of the Plancherel theorem for the reals that put the
Weil-Brezin (see [19] and [7]) mapping in a central position. I felt this would
yield a method for computing the finite Fourier transform. Indeed it did! It
yeilded the Cooley-Tukey algorithm. This inter-relation between the Cooley-
Tukey algorithm and the Weil-Brezin map is discussed in Chapter II.
All this aroused my interest in the computations of the finite Fourier
transform. I spent the Fall of 1977 at the IBM Watson Research Center
where I worked with S. Winograd. I have presented some of Winograd's ideas
in Chapter HI.

TABLE OF CONTENTS

Chapter I. The multiplicity problem

1. The Legendre symbol, quadratic reciprocity and the trace of the finite
Fourier transform
2. Equivalence of the trace and eigenvalue problems for the finite Fourier
transform
3. The algebra of the finite Fourier transform
4. Direct solutions of trace and eigenvalue problems
5. The finite Heisenberg groups and the finite Fourier transform
6. A proof of Theorem 1.3.3, nil-theta functions and theta functions
Chapter II. The Cooley-Tukey algorithm and the Weil-Brezin mapping
1. The Cooley-Tukey algorithm
2. The finite Heisenberg groups and the Cooley-Tukey algorithm
3. The Plancherel theorem for the reals and the Cooley-Tukey algorithm
Chapter III. Algebraic complexity and the finite Fourier transform
1. Basic ideas in algebraic complexity
2. Bilinear algorithms for the finite Fourier transform
3. General results on bilinear algorithms
4. Some minimal algorithms
5. Algorithms for computing the finite Fourier transform

MATHEMATICAL INTRODUCTION

In most applications the finite Fourier transform F(ri), n a positive integer,

is the n X n matrix whose entry in the a row and b column, 0 < a, b < n9 is
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 849

the number
02iriab/n

v^
where ab denotes the product of the 2 numbers a and b. Thus

V2
1 1 1
F(3) = 1 -2m/3 ,2^2/3
V3 1 e2m2/Z e2m/3

1 1
e2iri/n e2iri2/n . , , e2m{n-\)/n

92m2/n J2mA/n 92mi2{n-X)/n

F(n) =
VJt
j e2m{n-\)/n e2m2{n~\)/n . . . e2m(n- l)(/i- l)//i

The main problem involving the finite Fourier transform is the following.
Given a complex valued function /(a), 0 < a < n, we want to compute the
function Y(b), 0 < b < n, given by

no) ƒ(<>)
= F(n)
y(«-i)J [/(n-i);
We call y(6) the finite Fourier transform of the function f(a) and we will
abbreviate this by Y = F(ri)f and sometimes denote F(n)f by f .
However, on books on Harmonic Analysis the finite Fourier transform is
defined in the following, apparently, different fashion. Let Z/n denote the
group of integers mod n. Let C denote the complex numbers and Cx(ri)
denote the multiplicative group of complex numbers e?mk/n9 0 < k < n, or
what is called the group of n roots of unity. Let Z/n denote the set of
homomorphisms of Z/n into Cx(w). We make Z/n into a group by defining
(â + b)(a) = â(a) • b(a), a G Z/n, â, b GZ/n,
X
where multiplication is in C (AI). The group Z/n is called the character group
of Z/n and is well known to be isomorphic to Z/n. We introduce the
notation b(a) = (a,^è), a G Z/n, b G Z/n. Let ƒ be a complex valued func-
tion on Z/n or Z/n. We definej/) 2 = 2 ö G < 7 f(à)f(a) where G is Z/n or
Z/n. We define L2(Z/n) or L2(Z/n) as the set of complex valued functions
on Z/n or Z/n with the above norm. We define the finite Fourier transform
F(n):L\Z/n)->L\Z?n)
by
(F(n)f )(b) = 1
(1)
V7t Û£Z//I
850 L. AUSLANDER AND R. TOLIMIERI

(The multiplier l/Vn is inserted to make F(ri) a norm preserving linear

transformation or a unitary mapping or operator.)
In order to relate these two definitions of the finite Fourier transform of a
function, we have to introduce some identifications or isomorphisms. First
define r:Z/n-+ Cx(n) by
r(a) = e27Tia/n, a G Z/n.
We see that r is a homomorphism because
r(a + b) = e2™(a + b)/n = çliria/riçlmb/n^

Noting that e2>nia/n, 0 < a < n, is not equal to 1, we have that r is an

isomorphism.
Next, define s: Z/n -> Z/n as follows: For b G Z/n define s(b) G Z/n by
the formula
(a, s(b)) = e2"iab/n9 all a G Z/n.
It is straightforward to verify that s is an isomorphism. Using s to identify
Z/n and Z/n formula (1) becomes

(F(n)f )(b) = - p - S Aa)e2M,\ 0<b<n. (2)

This is the same as F(«)/.

As above, throughout this paper we have tried to begin with the computa-
tional version of a result or problem and only then to present the more
abstract or structured version of the result.
The following is a brief chapter-by-chapter survey of the contents of this
paper.
In Chapter I we study the finite Fourier transform as a linear transform,
rather than as the matrix product F(ri)f. Since F(n) is a unitary operator, and,
as we will show, F(n)4 = /, where I is the identity map, F(n) is similar to a
diagonal matrix whose eigenvalues are ± 1 and ± i. Hence as a linear
transformation, F(n) is uniquely determined by the dimension of the sub-
spaces Va, where Va consists of all vectors of functions in L2(Z/ri) such that
F(n)f=af, a = ±1, ±i.
The dimension of Va is called the multiplicity of a and the problem of finding
the dimension of Va, a = ± 1, ±i, is called the multiplicity problem of the
finite Fourier transform F(n).
In Chapter I we survey the various results, some classical and some not so
classical, that enable us to solve the multiplicity problem for F(n). One of
these methods shows that being able to find the trace of F(n) for all n is
equivalent to solving the multiplicity problem. Since the trace of F(n) is the
quadratic Gauss sum this shows one relation of the multiplicity problem to
classical mathematics. We also discuss some recent results that link the finite
Fourier transform and the theory of nil-theta functions. These results center
about an algebra structure that can be associated with the collection of ah
finite Fourier transforms F(ri), n > 0.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 851

CHAPTER I. THE MULTIPLICITY PROBLEM

1. The Legendre symbol, quadratic reciprocity, and the trace of the finite
Fourier transform. One of the simplest invariants of a linear transformation is
its trace. In this section we will define the Legendre symbol (p/q) and show
that if Tr(F(n)) denotes the trace of the finite Fourier transform then
Tr(F( M ))
Œ)-q)\p) Tr(F(p))Tr(F(q))
where p and q are odd primes. Of course the celebrated result of Gauss on
quadratic reciprocity states that
(E\(l\= f - \ ) [ ( p -0/2][(<7-l)/2]^

This shows one of the connections of the finite Fourier transform with
classical mathematics. In order to carry out this program, we will have to
introduce a representation p of the group Z/nx of units (elements with
multiplicative inverse) in the ring Z/n on L2(Z/n).
We will now start this section with a discussion of the Legendre symbol.
Let p be an odd prime. Then Z/p is a field having p elements and Z/px
consists of the nonzero elements of Z/p and is a cyclic group of order/? — 1.
Let
S={£2|£EZ//?*}.

It is easily verified that S is a subgroup of Z/px. The elements of S are called

quadratic residues mod p. Again one verifies that the order of the quotient
group Z/px/S is 2, or, for h G Z/px, h $ S, Z/px = S U hS and S n hS
is empty. Let {1, -1} be the multiplicative group of order 2 and let h be the
group homomorphism of Z/px onto {1,-1} with kernel S. For the integers
Z let p: Z-*Z/p be the homomorphism with kernel consisting of the
multiples of p. For n E Z we define
/ n\ fO if n = Omod/?,
\PJ ~\h(p{n)) if« iÉOmod/?.
We call (n/p) the Legendre symbol. Since, if n{ = n2 mod/?, (nx/p) = (n2/p)
we can consider (n/p) for n G (Z/p)x.
Our first task is to obtain an analytic formula for (n/p).
LEMMA 1.1.1. Ifn ^ 0 mod/?
2
2 e*** /p = (l) 2 e2"*2'*. (1)
0<£</> \PJo<è<p

PROOF. We will call Rp c Z a complete residue system mod/? if the

homomorphism p restricted to Rp defines a 1-1 surjection of Rp onto Z/p. If
Rp is any complete residue system mod/?, then it is easy to verify that

0<£<P teRp
852 L. AUSLANDER AND R. TOLIMIERI

Let k 2É 0 mod/? and let Rp be a complete residue system mod/? and

Then it is easily verified that kRp is a complete residue system mod/?.

Let (n/p) = 1 and n ==fc2mod/?, fc e Z//? x . Then
2
2» e2mne/p = V e ™(*02//> = 2 e27Tiri2/p = 2 e2^2^

and we have proven this case of the lemma.

Let (n/p) = - 1 or p(ri) E Z//? x , p(ri) £ S. As | runs over a complete
residue system mod/?,/?(£2) runs over £ twice and the point {0} once in Z//?.
Similarly, hi2 runs over hS twice and the point {0} once in Z//?. Hence
^ eimne/p + ^ e2™*V/> = 2 2 e2wi*/#\
|e/? p le/?,, i^Rp

The right-hand sum is well known to be zero and so

^ e2mne/p = (-1) 2 e2™*2/p = ( —) S e27"^

and we have verified our assertion.

Let ƒ be a function defined on complete residue systems mod n such that
for £ G Rn and £' E i*„' with £ = £' mod n we have ƒ(£) = ƒ(£')• We will talk of
ƒ as a function on Z/n and use the notation/©, J G Z/n.
We now begin the task of introducing the representation p of Z//? x on
2
L (Z/p) that combines with F(p) to give another formulation of Lemma
1.1.1. We begin the process of defining p by looking a little more closely at
L 2 (Z/«).
Let ¥(n) denote the complex valued functions on Z/n. Let

ƒ«(£) = (J; :_^' «,/?ez/„.

Clearly, ^(n) is an ^-dimensional complex vector space and the n functions
fa, a E Z/n, determine a basis of ^(n). We will now make ^(ri) an inner
product space as follows: For/, g E ^(w) define
<ƒ,£>= 2 ƒ(«)£(«)

where the bar denotes the complex conjugate. The resulting Hermitian inner
product space is denoted by L2(Z/n) and the n functions fa, a E Z/n, define
an orthonormal basis of L 2 (Z//î).
We will now describe the unitary representation p of Z/nx on L2(Z/ri).
Since Z/nx acting on Z/n by multiplication produces a group of automor-
phisms of the additive group Z/n, we may define for each a E Z/nx a linear
transformation p(a) of L2(Z/n) by setting
p(a)(f)s = f(as\ s E Z/n J E L 2 (Z//z).
l
Because p(ö)£ = fa~ a> where /,, a E Z/n, is the orthonormal basis defined
above, it follows that p(a) is a unitary operator on L2(Z/ri)9 a 6 Z / « X , If
U(ri) denotes the group of unitary operators on L2(Z/n) it is easily verified
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 853

that
p:Z/nx~*U(n)
is a group monomorphism.
We will now review briefly, and then extend, the material on the finite
Fourier transform that we gave in the Introduction.
Let C denote the field of complex numbers and let C,x denote the
multiplicative group of complex numbers of absolute value 1. A character on
Z/n is a group homomorphism X: Z/n^-±Cx. The set of all characters on
Z/n is denoted by Z/n. For À,, X2 e Z/n, we define
(Xj + X2)a = \x(a)\2(a), a E Z/n.
Then Z/n is a group, isomorphic to Z/n. For 0 < a < n, let Xa: Z/n -> C,x
be defined by
Aja) « e-2*ia*/*9 a e z / w .
Clearly Xa is a well defined character on Z/n and it is easily verified that Z/n
consists of the n characters Xa, 0 < a < n. Notice Xa E L 2 (Z/n) and

a,/} E Z//î- Hence the Aa, 0 < a < n, are an orthogonal basis of L2{Z/ri).
We can now define the finite Fourier transform F(n) of Z/n as the linear
mapping of L\Z/n) defined for ƒ E L\Z/n) by

(F(/z)/)(a) = 4Vn^ < ^ X « > =vn^ 2 At*)***/*

pez/n
where a E Z/n. Notice that
and
F(n)fa = - p - X ^ ^(")\*=^/a,

thus
Hn?fa=f~« and F(n) 4 = ƒ
where ƒ is the identity mapping. Also

(F(n)fa, F{n)f, > - \ <X_a, X.,) = { \ ^ " J

Hence (F(n)fa9 F(ri)fp) - </«»^3> and F(/z) is a unitary operator on L2(Z/n)
of order 4. We also have that

(i^iir I /)(«)--7=-</^..>--i 2 /(j8)e-*^/-.

Since

it follows that the matrix of F(n) with respect to the basis J ^ j S G Z / / J , is

854 L. AUSLANDER AND R. TOLIMIERI

given by
1
m(e2*ifir/n)9 0< fry <n
V7l
Although the linear transformation F(ji) is represented by different matrices
relative to different bases we will denote the above matrix also by F(ri).
Since p(a) E U(n), a E Z/nx, and F(n) E U(n) we can form
F(n)p(a)F(n)~l. Because p(a)Xa = Xaa we have immediately that
F(n)p(a)F(n)-1 = p(ayl - pia-1).
We can now relate the above results to Lemma 1.1.1. and the results of
Gauss on quadratic reciprocity and the value of quadratic Gauss sums.
If Tr( ) denotes the trace of the linear transformation in the bracket we
have immediately that
Tr(F(rt))= 2 eW/n,
a£Z/«

Tr(p(a)F(n)) = 2 ^W/*
oez/n
Now let n and m be relatively prime positive integers and let y = am + fin,
0<a<n,0</3<m.
The Chinese remainder theorem implies that the set of all such 7 is a
complete residue system mod nm. Now let fa E L 2 (Z/«), 0 < a < n, and let
ffi E L2(Z/n), 0 < p < m, be the basis of L\Z/ri) and L\Z/m) as defined
above. Let x(fa>fp) = /«m+/to> where/Y, 0 < y < nm, is a basis of L2(Z/nm)
as above. Extend x t o a bilinear mapping of L2(Z/n) X L2(Z/m)-*
L2(Z/nm). Then x induces a linear map x*- L2(Z/n) ® L2(Z/m)~*
L2(Z/nm). It is easily verified that x* is an isomorphism. A straightforward
computation then shows that the tensor product of the linear operators
p(m)F(n) and p(n)F(m) satisfies
p(m)F(n) ® p(n)F(m) = F(nm). (2)
The result of Lemma 1.1.1 can now be stated as

Tr(p(h)F(p)) = ^TT(F(P)), h 2É 0 modp, (3)

p an odd prime. Since the trace of a tensor product is the product of the
traces of the factors, we have from equations (2) and (3) that
[^){^fTx{F(p))Tx{F{q)) = Tv(F(pq))

or
TtjFjpq)) (

m= Tr(F(p))TT(F(q))
and we have verified the result we stated at the beginning of this section.
The formula
^}

primes, (5)
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 855

is Gauss' celebrated formula for quadratic reciprocity. Formula (5) has many
elementary proofs; see for instance, Hardy and Wright [12]. But Gauss also
established the following fundamental result.
THEOREM 1.1.2. Let F(ri) denote the finite Fourier transform on L2(Z/n) and
let Tr(F(n)) denote the trace of F(n). Then

i +1 ifn = 0mod4,
1 ifn = 1 mod 4, _
Tr{F{n)) =
(6)
w
0 j f / i s 2 mod 4,
i ifn = 3 mod 4.
Although formula (5) has elementary proofs Theorem 1.1.2 has, to our
knowledge no elementary proof and seems to be much deeper than quadratic
reciprocity. In [13] there is an interesting discussion of the many proofs of
quadratic reciprocity.
In the next section we will relate the problem of computing Tr(F(/i)),
n > 0, to the multiplicity problem for F(n). However, before doing this, we
pause to show what insights elementary considerations can give us about
Theorem 1.1.2.
We begin by showing that Tr(F(2r)) = 0, r odd, is easily verified. For
Tr(F(2r))= 2 e 2 "* 7 * + 2 e2^+r)2/2r.
0<£<r 0<£<r
27rir/2 27rl( +r)2/2r
Because r is odd, e = - 1 and e * = -.\e2"*2/2r and we have
established that Tr(F(2r)) = 0, r odd.
Let/? be an odd prime. We will now show that Tr(F(4/?)) = 1 + / implies

[i up = 3 mod 4.
By (2)
Tr(p(4)F(p))Tr(p(p)F{4)) = Tr(F(4/>)) - 1 + /.
Lemma 1.1.1 gives
Tr(p(4)F(p)) - (4/p)Tt(F(p)) = Tr(F(/>))
as it is easy to verify that (4/p) = 1. Thus
1+i
Tr( F(p)) =
Tr(p(p)F(4)) •
But we can easily write out the four terms of the sum Tr(p(/?)F(4)) to prove
that

Tr(P(p)F(4)) = f l + l if
^lmod4>
I 1 — * if p = 3 mod 4.
This shows that
Tr(F(^)) = ( 1 ÎJ» =
= 1J mod 4,
11 if n = 33 mod 4.
856 L. AUSLANDER AND R. TOLIMIERI

2. Equivalence of the trace and eigenvalue problems for the finite Fouriei
transform. This section relates Theorem 1.2.2 or the computation of Tr(F(n)),
n > 0, and the multiplicity problem as defined in the Introduction for F(n).
Since the trace of a linear transformation is the sum of is eigenvalues, it is
clear that a solution of the multiplicity problem for F(ri) implies Theorem
1.1.2 or Gauss' result in quadratic Gauss sums. What is very surprising is that
knowing Tr(F(n)) for all n enables us to solve the multiplicity problem for
F(n). The proof rests on a simple observation that is at the heart of Schur's
proof of the computation of Tr(F(n)) (see [6, p. 351]). Schur observed that
f1 0 • • • 0]
0 0 • • 0 1

[ 0 1 0 - • 0j
and hence the characteristic polynomial of F\n) is given by
(,-l)< n + 1 ) / 2 (f + l) ( "- , ) / 2 , «odd,
K)
( , _ i)<«+2>/2(, + i)<-2>/2, nev en.

Since the eigenvalues of F2(n) are the square of the eigenvalues of F(n),
formula (7) yields both that the possible eigenvalues of F(ri) are ± 1, ±i and
the following result on multiplicity: If F(n) has mx eigenvalues 1, m2 eigen-
values - 1 , m3 eigenvalues i and m4 eigenvalues -i' then
n+1
2 , " odd,
mx + m2 = —-z—, m3 + m4
n +2
mx + m2 = —-—, m3 + m4 ——, n even. (8)
and
Tr(F(n)) = (mx - m2) + i(m3 - m4). (9)
Now let Tr(F(n)) = a + i/3 then (8) and (9) combine to yield for odd n
mx — m2= a, m3 — m4 = fi9
n+ 1 , n- 1
mx + m2 = —-—, m3 + m4 = —-—
and a similar set of equations for even n. Clearly, we can solve for mx, w2, m3,
m4 in terms of n, a, (i and hence once we have evaluated Tr(F(n)) we have a
complete solution for the eigenvalue problem. This shows that the trace
problem for F(ri) is equivalent to solving the eigenvalue problem.
Theorem 1.1.2 can now be stated in the following equivalent form.
THEOREM 1.1.2'. Let F(n) denote the finite Fourier transform on TL/n and let
mpj — 1, 2, 3, 4, denote the multiplicities of the eigenvalues 1, - 1 , i, -i of F{n\
respectively. The value of mJyj = 1, 2, 3, 4, as a function of n is given by the
following table.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 857

n mx = 1 m2 = -1 m3 = i m4 = -i
Am m+ \ m m m— i
4m + 1 m+ 1 m m m \ '
4m + 2 m+ \ m+\ m m
4m + 3 m+ 1 m 4- 1 m+ 1 m
3. The algebra of the finite Fourier transform. Theorem 1.3.1 stated below
was first discovered and proven in [4] using nilpotent harmonic analysis and
its proof is independent of Theorem 1.1.2'. However the equivalence of
Theorems 1.1.2' and 1.3.1 is easily established and requires no nilpotent
harmonic analysis.
Let C[XX9 X2, X3] be the polynomial algebra in three inde terminants over
the complex numbers C and let C[XX, X29 X3] be the subalgebra generated by

9Lx"C[Xl9Xi9Xi]/(Xi + Xi)9
9t2 = C[Xl9 Xl Xl}/ (X% + X*X\ + Xl)9
3t3 = C | X Xl Xl}/ {Xl + XfXl),
where ( ) denotes the principal ideal in C[XV X29 X3] of the polynomial in the
bracket. Let Yx Y2b 733c, a, b, c e Z, a, b9 c > 0, denote the image in 9la,
a — 1, 2, 3, of the monomial Xx X2b X3C. Then it is easily established that
Y* ?? *33c> c = 0, 1, a9 b > 0,
is a vector space basis of the algebra 2Ia, a = 1, 2, 3. It is not difficult, using
elementary methods, to prove that
y6 i y6 y 6 i Ay 4 Ay 2 _i_ A y 6 y6 i y4 v2
^ 3 "*" A 2> A
2 "*" l 2 "^ 2> A
2 "*" A l A 2
are each irreducible in C[Xl9 X29 X\\ Hence each of the algebras 3ta, a =
1, 2, 3, has no divisors of zero.
THEOREM 1.3.1. Let 3ïa, a = 1,2, 3, be as defined above and let ^%a -* 9ta,
a = 1, 2, 3, be the linear transformation of 9ta such that
<5{Yxa Ylh Ylc) = ( - l)bieYf Ylb Ylc9 c = 0, 1; a, b > 0.
Lef F(«) 6e //*e sector subspace of 9la spanned by Yx Y2b Y\c where a + 2b +
3c = n and let ^(n) = ^ | V(ri), Then dim V(n) = n and ^(n) is equivalent to
the finite Fourier transform F{ri). Further

PROOF. Let us begin by verifying the last assertion of the theorem. To do

this we need only verify it for the basis elements. Now in 3l2 (the other cases
are handled similarly)
ya ylb __ yd yle ,__. ya + d yl(b + c)

ya y2b y3 # yd yle =L ya + d y2(6 + c) y 3

ya ylb y 3 . yd yle y 3 _ __ ya + b + 4 y2(6 + c+l) _ ya + b y2(6 + c + 3)

858 L. AUSLANDER AND R. TOLIMIERI

from which it is easy to verify that

^ ( 1 7 Ylh 733c- Ydx Yle Y?) = #(y« Y\b Ylc)%{Ydx Y\e F33/)
andsof(/g)=f(M(g).
We will now indicate the inductive proof used to prove that dim V(n) = n
and ?F(rt) and F(n) have the same eigenvalues with the same multiplicities.
This will prove that ^(n) and F(n) are equivalent. We will adopt the notation
Vx{ri), x — ± 1> ± /, for the subspace of V(n) of eigenvectors with eigenvalue
X of #(/!).
We must now prove that dim Vx(ri) satisfy the table of Theorem 1.1.2'. This
is merely a property of 2la and §(n) and does not use Theorem 1.1.2' in its
proof.
By inspection the above assertion is true for n < 4. Consider n = 4m,
m > 1. Now
^ ( y ? Y\b y33c) = 1 y? y226 y33c
if and only if c = 1, 6 is even and a + 26 = 4m — 3. Thus dim ^(4m) =
dim Vx(4m — 3) = dim Vx(4(m — 1) + 1) which by induction is equal to m.
Thus
dim l^(4m) = m.
Similarly
^ ( y ? Ylh y33c) = - 1 yf y226 y33c
if and only if c = 1, 6 is odd and a + 26 = 4m — 3. Thus
dim F_ l (4m) = dim V„x{4m - 3) = dim F ^ m - 1) + 1)
which by induction is equal to m — 1. Hence
dim F_ l (4m) = m — 1.
Next
^ ( y f y22* y 3 3c ) = - Yax Ylb Y\C
if and only if c = 0, 6 is odd and a + 26 = 4m. Let 6' = 6 — 1. Then 6' is
even and a + 26' = 4m — 2 = 4(m — 1) + 2. Thus
dim V„x(4m) = dim Vx(4(m - 1) + 2)
which by induction equals m. Thus
dim V„x(4m) = m.
Finally
^ ( y f y22* y33c) = yf y f y33c
if and only if c = 0, 6 is even and a + 26 = 4m. These constraints are
satisfied only by the following values of a and 6:
a = 4m, 6 = 0; a = 4(m — 1), 6 = 2 , . . . ; a = 0, 6 = 2m.
Hence
dim Fj(4m) = m + 1.
 859
COMPUTING WITH THE FINITE FOURIER TRANSFORM

Since
2 dim Vx(4m) = dim V(4m)

we have dim V(4m) = Am and we have proven our result for n = 4m. The
other cases are proven in a similar way and the proof is omitted.
Thus Theorem 1.1.2' and F.4.1 are equivalent.
4. Direct solutions of trace and eigenvalue problems. Gauss' original proof of
Theorem 1.1.2 can be found in H. Rademacher [17]. Gauss' proof is algebraic
and very different from the proofs of his theorem that one finds in the usual
texts on elementary number theory. We will not summarize Gauss' proof, but
we do seriously suggest that all readers examine Rademacher's account of this
remarkable achievement.
We know two different analytic proofs of Theorem 1.1.2 and these are the
proofs most often found in elementary texts. Schur's proof was the first proof
that stressed the role of the finite Fourier transform and its eigenvalues.
Accordingly, we will begin with a brief discussion of Schur's proof. We will
follow it with McClellan and Park's proof of Theorem 1.1.2'. We will then
discuss the two analytic proofs of Theorem 1.1.2.
We will present Schur's proof of Theorem 1.1.2 only when n is an odd
prime p as this makes the discussion simpler, but exhibits all the most
interesting aspects of the method of proof. For those who want the whole
story, this can be found in [6].
As in §1.2, let ml9 m2, m3, m4 denote the multiplicity for F(p) of the
eigenvalues 1, - 1 , i, — /. In §1.2 we showed that
p+ 1 p- 1
m
\ + m 2 = 2 > m3 + m4 = —y—
and
Tr(F(p)) = ml - m2 + i(m3 - m4).
Schur's method of proof is simply to obtain enough relations amongst the m's
to enable them to be computed.
We begin with a result that shows that it is the sign that is the difficult part
of Theorem 1.1.2.
LEMMA 1.4.1. Let F(p) denote the finite Fourier transform on L2(Z/p),p an
odd prime. Then
iS p mo
TK^))-f;! f ^\ ^;
[ ± i if p = 3 mod 4.
PROOF. A multiplicative character X of Z/px is a homomorphism of the
multiplicative group Z/px into Cf. We extend X to a function on Z/p by
defining A(0) = 0. Now let A be a nontrivial multiplicative character of
Z/p x ; i.e., there exists £ e Z/p x such that A(£) =£ 1. View X as an element of
L\Z/p). Since the characters
K(0 = e***'*, a E Z/p,
determine an orthogonal basis of L2(Z/p) and <Xa, Xa) = p, we can write
860 L. AUSLANDER AND R. TOLIMIERI

P lez/p
An argument similar to that given in Theorem 1.1.1 gives

Tr(F(^))=^
V~p
2iez/p\P)
(fW*.
Thus, Tr(F(p)) = V 7 # i where

and ( //?) is the Legendre symbol that was defined as a multiplicative

character.
Again let A be an arbitrary multiplicative character. For c ^ 0 mod p9
X(cx) = 2 aaK(cx)-
a6Z/p
Since A(cx) = A(c)A(x) we have

aGZ/p o6Z//)

Now Aa(c.x) = AcaO) and so if a = c""1/? we have

E tfA(c*)= S *A«(*)
aeZ/j) aeZ/>

H
/?ez//> /*ez//>

Hence ac_£ = A(c)a£ for all /?. Hence |aj| = |ac|, where | | denotes the
absolute value, c ^ 0 mod/?. Thus
(\\y=p(p-l)\al\2 = p-l
and
|«,|-l/Vjp.
Since Tr(F(p)) = V^p a j when A = (/), we have
|Tr(F(/>))|=l.
It is an elementary fact that (-l/p) = ( - l ) ^ - 0 / 2 . By Theorem 1.1.1
Tr(F(/>))-= ( - l / ^ T r ^ ) )
where the bar denotes the complex conjugate. Hence
Tt(F(p))2(-l/p) = 1
or

Tr(F(P))={ll ^l™**;
{ ±i ifp = 3 mod4.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 861

Lemma 1.4.1 implies that

mx — m2 = ±1, #*3 = m4 if p = 1 mod 4,
m = m
\ 2> m3 — m4= ±1 if/? = 3 mod 4.
The fact that det(F(/?)) = ( - \)mHm^-i)m* will enable us to obtain the last
relation we require.
Let
A = det(e 2 ^^)o<a, € <p
then
A = /?'/2det(F(/?))
and v4 is a Vandermond determinant. Hence
A = Et (e 2,rir//> - e2nis/p).
0<s,r<p
1
Let Î] = e "/*; we have

0<s,r</>

- n î?r+5 n 2/sin-^—-v.
0<s,r<p 0<s,r<p P
r = 2
Because S ^ ^ o + * P((P ~ 0/2) we have II r j r + ' = 1. Hence
2

A = ,0>~ D/>/22/>0>- D/2 JI sin (LZ-fk.

0<5<r<0 /*
Because we know the form of F(p)2 we see that
A2 = p^(-iy(p-l)/2
and
A = ±iP(p-»/2pP/2.
Since sin[(r — s)7r/p] >0for0<s<r<p — lwe must have
det ^ = jP(P-i)/2pP/2 = ^p/2 det(F(/?)).
Hence I ^ P - D / 2 = ( - l ) * ^ - ^ Thus

— - = 2m2 + m3 — m4 mod 4.
From which we have
mx — m2 = mod 4 if/? = 1 mod 4,
m3 — m4 = mod 4 if/? = 3 mod 4.
This combines with our previous results to

[i if/? = 3 mod 4.
1.4.2. MCCLELLAN AND PARK'S PROOF OF THEOREM 1.1.2'. The proof of
Theorem 1.1.2' by McClellan and Park is interesting for three reasons. First, it
862 L. AUSLANDER AND R. TOLIMIERI

is the first direct proof of Theorem I.1.2'; second, it is extremely explicity, in

that it exhibits eigenvectors; third, it rests on the use of Chebyshev sets. Using
Chebyshev sets is a very novel idea and we do not believe it would have
occurred to many "pure" mathematicians. It is also interesting to note that
McClellan and Park were ignorant of Schur's work (see [15]). We will now try
to present the flavor of their proof. For complete details see [16].
We will list some basic facts about Chebyshev sets.
DEFINITION. A set of n smooth functions on an interval is a Chebyshev set
for the interval if any nonzero element in the linear span of the n functions
has at most n — 1 distinct zeros.
THEOREM C.l.
{1, cos t,. . . , cos nt) is a Chebyshev set on [0,7r].
{sin f,..., sin nt) is a Chebyshev set on (0, TT).
THEOREM C.2. If {q>x(t\ . • . , <pn(t)} is a Chebyshev set on an interval and
*v • • • 9 *n+\ are distinct points of the interval, then the matrix
(>i(>i) ••• 4>„('i)l

|>('J • • • *„(OJ
is nonsingular. If 8, i = 1 , . . . , w + 1, are all nonzero and alternate in sign,
then
\ <J>i('i)

*#i('«+i) 8 »+ij
is a nonsingular matrix.
We will now list certain elementary facts about the finite Fourier trans-
form. Proofs can be found in [16].
DEFINITION. A function ƒ on Z / « is called even if

ƒ(<*) = ƒ ( - « ) , aGZ/ii.
A function ƒ on TL/n is called odd if
ƒ ( « ) = -ƒ(-<*), aEZ/n.
2
1. For ƒ G L (Z/n), F\n)(f)(a) = /(-a).
2. Let [x] denote the greatest integer less than x. Then L2(Z/ri) has a
v = [n/2] + 1 dimensional subspace of even functions and an n — v dimen-
sional subspace of odd functions.
3. If ƒ is an eigenvector of F(ri), then ƒ is either an even or odd function.
4. Even eigenvectors have eigenvalues ± 1. Odd eigenvectors have eigenval-
ues ±i.
5. If ƒ is an even function, then F(ri)(f) + ƒ (F(ri)(f) — ƒ) is an eigenvector
of F(n) with eigenvalue 1 (-1). If ƒ is an odd function, then iF(ri)(f) -
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 863

f(iF(n)(f) + ƒ) is an eigenvector of F{n) with eigenvalue i (— *).

6. If g is an even (odd) eigenvector of F(n)9 then there exists an even (odd)
function ƒ such that g = F(n)(f) ± ƒ (g = /F(n) T ƒ). We are now in a
position to outline the McClellan-Park proof.
Consider the cases TV = 4m, Am + 1, 4m + 2, 4m + 3 separately. Let m„
i = 1, 2, 3, 4, be as in the statement of Theorem 1.1.2'. The steps of the proof
are as follows: Exhibit fk even functions 1 < k < mx such that F(ri)fk + Z^,
k = 1, . . . , m1? are linearly independent. Since F(n)fk + /^ has eigenvalue 1,
this will prove that the multiplicity of the eigenvalue 1 is greater than or equal
to mv Exhibit/^ even functions 1 < k < m2 such that F(n)fk + fk are linearly
independent. Since F(n)fk — fk has eigenvalue - 1 , this will prove that the
multiplicity of the eigenvalue - 1 is greater than or equal to m2. Similar
statements hold for m3 and m4. Since mx + m2 + m3 + m4 = TV, this will
prove Theorem 1.1.2'.
We will indicate the method of proof by working out for the case TV = 4m
that the multiplicity of the eigenvalue 1 is greater than or equal to ml = m +
1. Let fa, a = 0, . . . , 4m — 1, be the bases of L2(Z/N) where fa is the
function that takes the value 1 at a G Z/iV and zero at all other points. Let
go=fo> 8m=flm> Zi^fi+fN-i* I = 1, . . . , W - 1.

We need to study
m
^éai(F(N)gi + gi)
i= 0
or
m

1=0
Since e~27rik/N = e2^N~k)/N the coefficients of fm, ...,f2m y i e l d m + l equa-
tions in m + 1 unknowns that can be written as

1 cos tx ••• cos(m — 1)*! (-ir

2a,

1 cos tm •• • cos(m - \)tm (-1) 2 "- 1

[l cosfm+1 • • • cos(m - l)f m+ i 2a„
(I+VN)
where tt = [(m + i - l)/m][7r/2], i = 1 , . . . , m 4- 1. We now apply Theo-
rem C.2 to conclude that the images of the vectors g0, . . . , gm are indepen-
dent and so the multiplicity of the eigenvalue 1 is greater than or equal to
m + 1 for N = 4m.
1.4.3. DIRICHLET'S PROOF OF THEOREM 1.1.2. We will now outline Dirichlet's
proof of Theorem 1.1.2. Complete details can be found, for instance, in Lang
[14].
This proof of Theorem 1.1.2 rests on the following classical result about
Fourier series.
864 L. AUSLANDER AND R. TOLIMIERI

THEOREM FA. I/O is a C' function on [0, 1] then

*(*) = 2 cme2"imx, 0 < x < 1,

awrf
c
^ z, *
where

cOT = Jr,ö(jc)e-2,r''"ut dx.

o
2 ixl/n
Now let f(x) = e " , 0 < x < 1, and let fk(x) « f(x + k), k = 0 , . . . ,
n - 1. 7%e«
.4(0) + A P ) _ e2"*2 + g2™<*+'>/«
2 2

"y A(0) + A W _ 1 + e 2 *'" e 2m/ " + e2""2'/" _

A-O 2 2 2

+
2 •
By reassociating the terms in the sum, we obtain
A(0)
"S ! / t ( 1 ) = I + " 2 e2**V" + * = VÏÏ Tr(F(n)).
2 2
A: = 0 * £= 1
Let 0 = f0 4- • • * +ƒ,_!. Then 0 is C' on [0, 1] and so, by Theorem F.l, we
have

V^Tr(F(fl))= 5 'S Cfk{x)e^imxdx.

After some elementary operations that include completing the square, we

obtain
Vn Tr(F(n)) = 2 e^inm^2 fneW*-n>»/»2/» dx.
m ~ — oo Jo
ninm2/1
If w is even e' = 1 and if m is odd e-™"2/* = r \ We split the sum
over even m and odd m. A computation that involves letting m = 2r or
m = 2r + 1 shows that the sums of the integrals over m even or m odd are
equal to

4 = C e2™^1 dy.
«' — oo

One verifies that the above improper integral converges and that In = Vn tv
From all this we otain
Trcn«))-(i + r-)/0 + r')
which is another form of Theorem LI.2.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 865

1.4.4. LANDSBERG'S PROOF OF THEOREM 1.1.2. We will now outline Lands-

berg's proof of Theorem 1.1.2 as presented in Bellman [5]. This proof is
particularly interesting in light of Theorem 1.3.3 and the fact that the algebra
c[xl9 xl xl]/ {xl + xfrl + xf)
is the algebra of theta functions of characteristic (0, 0) and period /. Lands-
berg's proof rests on the famous functional equation satisfied by the theta
constants. To be more precise, the theta constants are the first order theta
functions of period / = r + is, r > 0, evaluated at the origin. This is the
function

n— - c o
and/(0 satisfies the functional equation

*>-(ÎP£).
Clearly ƒ(/) diverges along the line Re(7) = 0. To find the relation between
Gaussian sums and the theta constant f{i), we examine/(0 in a neighborhood
of its line of divergence.
Let S(py q) = 2?^o e"***'*, (p> q) = 1. Set t = e + irip/q where e > 0.
Then

jL + ™E\ . ! + 2 f *-•**'«{ | e-^^A.

The function of e in the right-hand bracket behaves like the integral

Now as e -» 0 the above integral is asymptotic to

Hence as € ~~» 0, we have/(E 4- mp/q) is asymptotic to VTT £(ƒ>, q)/qVl.

One similarly finds that f {IT2 ft) is asymptotic to Vn S( — q,p)/qVe as
e ~» 0. Using the fundamental functional equation and the asymptotics dis-
cussed above, we find

Yq r = 0 V 7 r- 0
Choosing # odd and/? = 2 yields Theorem 1.1.2 for w odd.

5. The finite Heisenberg groups and the finite Fourier transform. In this
section we will begin to discuss the role that nilpotent harmonic analysis plays
in the theory of the finite Fourier transform. The reason for the importance of
nilpotent harmonic analysis is that certain finite nilpotent groups have the
finite Fourier transform built into their structure. The first indications of this
are given in this section and will be looked at again in Chapter II.
866 L. AUSLANDER AND R. TOLIMIERI

Let 91 be a commutative ring with identity and such that 2x = 0, x G 91,

implies x = 0. We define the 9l-Heisenberg group as the group of matrices
f 1 a c]
0 1 b\
10 0 1J
with a, b, c G 91. It is easy to see that the 9l-Heisenberg may be defined as
the set
N(9L) = 91 X 91 X 91 = {(a, b, c)\a9 b, c G 91}
with multiplication given by
(*1> b\> C\)(a2' b
2> Cl) = (al + <*2> *1 + *2> C l + C
2 + #2*2)-

The center of AT(9l), z(N(<SL)) = {(0, 0, c)\c G 91} and JV(9l)/z(JV(9l)) » 91

© 91. Clearly z(7V(9l)) is isomorphic to the 9L as an additive group.
We will now review some of the basic facts about unitary representations of
the groups N(Z/n).
Let Cm denote the m-dimensional complex vector space (c,,. . . , cm) = c.
Define the usual Hermitian structure on Cm by
m
<cd>-2c,3
1=1

where bar denotes complex conjugate. Let U(m) denote the group of linear
transformation of Cm such that for U G U(m) and c, d G Cm
<c,d> = <£/(c), t/(d)>.
U(m) is then called the group of unitary transformations of Cm. A unitary
representation p of a group G is a homomorphism of G into f/(m). A unitary
representation p is called irreducible if the only subspace of Cm invariant
under p(g), g G G, is Cm or 0. Also, p is called faithful if p is a monomor-
phism.
Let p t and p2 be unitary representations of G on Cm. We will say that px is
unitarily equivalent to p2 if there exists a unitary matrix U such that
U~lpx(g)U=p2(g), allgGG.
1/ is then called an intertwining operator for px and p2.
Let / be the identity matrix in U(m) and let x be a character on an abelian
group A. By x^ we mean the unitary representation of A defined by (xl)(à)
= x(à)I,a ^A-
Let G be a group and let z(G) denote the center of G.
THEOREM R.l. Let px and p2 be unitary representations of G and assume, in
addition, that px is irreducible. Let U be an intertwining operator for px and p2.
Then p2 is irreducible and U is unique up to multiplication by an element of C*.
THEOREM R.2. Let p be an irreducible representation of G and let z(G) denote
the center of G. Then z restricted to z(G), p/z(G) = x * I> where x *s a
character of z(G).
We will now state two results that are specific for the groups N(Z/n).
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 867

These results give a picture of the faithful irreducible unitary representations

of NÇL/ri).
THEOREM R.3. Let px and p2 be two irreducible unitary representations of
NÇL/ri) then px and p2 are unitary equivalent if and only if px\z(NÇL/ri)) =
p2|z(AT(Z/«)).
THEOREM R.4. Let A be a maximal abelian subgroup of NÇL/ri) and let x be
a character on A such that x|<z(N(Z/n)) is a faithful character. Then inducing x
from A to NÇL/ri) gives an irreducible unitary representation of NÇL/ri).
Thus, every faithful character x of z(N(Z/ri)) extends to a faithful irreduc-
ible unitary representation of NÇL/ri) and every faithful irreducible unitary
representation of NÇL/ri) restricts to a faithful character on z(N(Z/n)).
In order to introduce two irreducible unitary representations px and p2 of
NÇL/ri) that have the finite Fourier transform as intertwining operator; i.e.,
F(n)p2F(n)~~l = px
we need to define the following matrices.
J2iria-0/n n

DM) = 0 <a <n,

02m(n— \)a/n
0
0 1 0 0

• •. '. '. 0|
0 1
1 0 OJ
2vic/
Let A = (a, 0, c) and let x(0, 0, c) = e ". Inducing x from A to NÇL/ri)
gives the following irreducible unitary representation of NÇL/ri) =
{(a, b, c)\a, b,c G Z/n)
Pl(z(N(Z/n))) = x • I,
Pl(a, 0, 0) = D„(a),
p,(0, b, 0) = (Sn)b.
2mc/n
Let B = (0, b, c) and let x(0, 0, c) = e . Inducing X from B to NÇL/A)
gives the following irreducible unitary representations of NÇL/ri)
p2(z(N(Z/n))) = x • /,
p2(a, 0, 0) = (Sn)a,
p 2 (0, b, 0) = Dn(b).
It is an elementary exercise to verify that p, and p 2 are irreducible unitary
representations of NÇL/ri) and if
1
F(n) (e 2iriab/n ), 0 < a,b < n,
\Tn
868 L. AUSLANDER AND R. TOLIMIERI

that
F(n)p2F(n)~l = p,
or F(n) intertwines px and p2. By the uniqueness of intertwining operators,
F(n) is essentially determined by the representations px and p2.
At this point the appearance of F(n) as an intertwining operator for unitary
representations of N(Z/ri) is totally unexplained. In §11.3 we will see another
way of looking at p, and p2 that better explains the finite Fourier transform's
role as an intertwining operator for px and p2.
6. A proof of Theorem 133, nil-theta functions and theta functions. In this
section we will outline a proof of Theorem 1.3.3 that uses harmonic analysis
on the real Heisenberg group and is completely independent of Theorems
1.1.2 or 1.1.2'. This proof shows the deep relation between the finite Fourier
transforms, F(n), n > 0, and the algebra of theta functions with periods 1 and
V^T . A complete exposition of this material can be found in Chapter II of
[1].
Let R denote the reals and Z c R, denote the integers. Let N = JV(R) be
the R-Heisenberg group and let T = N(Z) be the Z-Heisenberg group. Then
T c N and T \ N is a compact manifold. If (x,y, z) G N, x,y, z G R, then
the 3-form dx /\dy f\dz induces a probability measure on T \ N. We form
the Hubert space L 2 (r\N) and define a unitary representation U of N on
L\T \ N) as follows: For g G N, f G L 2 G (r \ N) define
(U(g)f)(Th)=f(Thg), hGN.
It will be convenient to consider functions T \ N as functions on Af such that
f(yh) = f(h), y G T, h G N. In general, if ƒ is a function on JV and yGiVwe
set
Zr{f){h)=f{rXh)
and call £ the left action. For each m G Z, let
H{m) = {ƒ G L\T \N)\f(x,y, z + /) = e2™%x,y, z)}.
One verifies that U(g)H(m) = H(m), g G N, and that
L2(T\N)= 2 @H{m)
mez
where the sum is the orthogonal sum.
We now want to better understand the spaces H(m). To do this we
introduce the automorphism
D-.N-+N
m
given by Nm(x9y, z) = (mx, >>, mz). By letting ƒ -» ƒ ° Dm we may use Dm to
induce a linear mapping of L2(T \ N). By a slight abuse of notation we will
denote this linear mapping also by Dm. Then
Dm{H{\)) c H{m).
One verifies that
m-l
H(m) = 2 eWM(/).(fl(l))).
y-o
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 869

We will denote t{0J/m^{Dm{H{\))) by H{mJ\ 0 < j < m - 1. Then

U(g)H(m,j) = H(mJ), g G N, ally. A deeper fact is that each of the spaces
H(mJ) is irreducible under the action of U and that H(m, h) and H(n, k) are
unitary equivalent with respect to U if and only if m = n.
The notions of irreducibility and unitary equivalence are the general
Hubert space analogues of those introduced in the previous section. We will
say more about the structure of H{\) in §11, where we consider the Weil-
Brezin map.
We next observe that
J(x,y, z) = {-y, x,z- xy)
is an automorphism of N such that J(T) = T and such that J(H(m)) = H(m).
Hence J induces a unitary operator on H(m) which we call Jm. The action of
Jx on H(\) corresponds to the real Fourier transform in a sense that will be
discussed in Chapter II.
In [1] we showed that J enables us to define a first order differential
operator D(J) on T \ N such that
/ > ( / ) ƒ = />(/)(/(ƒ)), ƒ G C«>(T\N).
Let
0(m) = {ƒ G C°° n H{m)\D{J)f = 0}, m > 0.
We will now outline the main properties of the subspaces 0(m) and 0 =
2 m > 0 © 0(w). (Notice, since ƒ G 0(m) and g G ©(AI) are both C°° functions
on T \ N their product f g is a C °° function on T \ JV.) First, 0 is an algebra
and 0(m)0(rt) c 0(m + /*). Second, /(0(m)) = 0(m). Third, dim 0(m) =
m. Finally, 0 has no zero divisors.
The representation theory can be used to prove that
m~\

0(m) = 2 ®®(mJ)
where @(mj) = 0(m) n #(m,;) = t{0J/mfi){Dm{&{\))) and 0(1) has basis
<p(x,y,z) = e2™ 2 e—^ + 'te™*.
/eZ
The relationship between the algebra 0 and J and the space â = 2 „ > 0 ©
L2(Z/n) and the finite Fourier transform extend to S by F = 2 „ > 0 © ^(«)
is given by the following theorem.
THEOREM 1.6.1. There exists a unitary operator V: 0-> â satisfying
(1) F(0(«)) - L 2 ( Z / H ) ,
( 2 ) F = F/K -1 .
It follows that 2- can be given the algebra structure of 0 and since
J(fifi) = JUxVUùJvh G @> we have
HSigi) = HsùHgiX Si> & G S.
It is also proved in [4] that 0 is isomorphic to

c[xv xl xl]/ (xt + xfxi + xf)

870 L. AUSLANDER AND R. TOLIMIERI

and so Theorem 1.3.1 follows from Theorem 1.6.1.

The proof of Theorem 1.6.1 proceeds in the following way. We need to find
a basis of &(m) such that the matrix of Jm with respect to this basis is

Jm = —l— e2™^m, 0 < a,j3 < m.

Vm
We do this as follows. Let

<pm(x,y,z) = e2™ 2 e-™o> + t)2e2„uxt

/ez
Then Dm(<pm) E @(m, 0) and we can form the basis
= L
*W (0J/rn,0)(£>m(<Pm))> 0 < j < m,

of 0(m). In [1] we show that <pmJ, 0 < j < m, is the required basis.
We can also verify this result using the ideas of §1.5. First, we define a
representation Ul of the Z/m-Heisenberg group N(Z/m) on @(w) as follows.
For ƒ E 0(m) and a, b, c E Z let
Ux(m)(a9 0, 0)/ = L{a/m^oyf9
^ ( m X a ^ O ) / ^ L(0,Vm,o>/,
f / ^ m X O ^ ^ ) / - e2"i(c/m)f.

It is not hard to see that with respect to the basis <pm7, defined above, that
Ux(m) is a unitary representation of N(Z/m) on 0(m) and that the matrix
Ux(m) = pj, pj as defined in the previous section. Also one verifies that
J~xUx{m)J = p2.
Since J(m) intertwines pj and p2 it follows that /(m) = cF(m) where |c| = 1.
One then verifies that c = 1 and we have our assertion.

CHAPTER II. THE COOLEY-TUKEY ALGORITHM AND THE WEIL-BREZIN MAP

1. The Cooley-Tukey algorithm. Currently the most popular algorithm for

computing the finite Fourier transform is called the Cooley-Tukey algorithm.
The history of this algorithm has been set forth in an interesting article by
Cooley et al. [9] and the original paper is Cooley and Tukey [10]. It is our
intention in this section to analyze in some detail the basic construction upon
which the Cooley-Tukey algorithm rests. This will enable us to relate the
Cooley-Tukey algorithm to the Weil-Brezin map (see [19], [7]) and the proof
of the Plancherel theorem for the reals as given in Chapter 1 of [3].
Because it has been so important in the theory of numerical computations
and because it is so brief, we will begin by reproducing the few paragraphs in
Cooley and Tukey [10] that-aside from induction-set forth the idea of the
Cooley-Tukey algorithm.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 871

"Consider the problem of calculating the complex Fourier series

XU) = 2 A(k)WJ\ j = 0, . . ., N - 1, (1)

where the given Fourier coefficients A(k) are complex and Wis the principal
iVth root of unity
W = el7Ti/N. (2)
2
A straightforward calculation using (1) would require N operations where
'operation' means, as it will throughout this note, a complex multiplication
followed by a complex addition.
The algorithm described here iterates on the array of given complex
Fourier amplitudes and yields the result in less than 2N log2 N operations
without requiring more data storage than is required for the given array A. To
derive the algorithm, suppose N is composite, i.e., N = rxrv Then let the
indices in (1) be expressed
J = J\r\ + Jo> Jo = 0, 1, . . . , rx-l, jx = 0, 1, . . . , r 2 -l,
k = kxr2 + k0, k0 = 0, 1, . . ., r 2 ~l, kx = 0, 1, . . ., /y-1. (3)
Then, one can write

XUvJo) = 2 S A{kx, k0) W*'Wk« (4)

k0 kx

since

The inner sum, over kx, depends only ony 0 and k0 and can be defined as a
new array,

AlUo^o) = ^A(kx,k0)W^^ (6)

The result can then be written

XUvJo) » 2^iOo> kjW«>*+X>*: (7)

k0

There are N elements in the array Al9 each requiring rx operations, giving a
total of Nrx operations to obtain Av Similarly, it takes Nr2 operations to
calculate X from Av Therefore, this two-step algorithm, given by (6) and (7),
requires a total of
T=N(rl + r2)
operations."
Let us now formalize the steps of the Cooley-Tukey algorithm. We let ^ ( )
denote the complex functions whose domain is the set in the parens and
872 L. AUSLANDER AND R. TOLIMIERI

^(X9 Y) denote the mappings from X to Y. Then the Cooley-Tukey algo-

rithm rests on the commutativity of the following diagram:

^(Z/r x r 2 ) • S W i x ZA2) «^(Z/r 1 ( ^(Z/r 2 ))

1 x F(r2)

Ç&fa xZfrJ^VIr^iZIrJ)

F(rxr2) M

$(Z/r2 x Z/rx) «#(Z?r 2 ,f (Z/r,))

1 x F(rt)
D'
^ ( Z / r j r j ) *• '$(Z/r2 x Z/r,) « f (Z/rj.^Z/r!))
where the "hat" denotes the dual group or the group of characters and where
we must still define the various mappings of the above diagram.
DEFINITIONS. THE MAPPING » . Let f(x9 y) E ^(X X Y). Then for each
fixed x0 E X, f(x0,y) E ^(Y) and so f(x,y) determines an element of
^(X, ^(Y)). It is obvious that this correspondence is 1-1.
THE MAPPING C. Between sets Z/rxr2 and Z/r, X Z/r 2 define the follow-
ing homeomorphism C*. (Notice: C* is not a group homomorphism.) For
0 < k < rxr2 let k = k2 + kxr29 where 0 < k2 < r2, 0 < kx < rx. Define
C*(/c) = (kX9 /c2) E Z/rx X Z/r 2 .
For ƒ E $(Z/rxr2) define C(/) = ƒ o (C*£ E ^(Z/r^X Z/r^.
THE MAPPING Z>. Between the sets Z/*rxr2 and Z/r 2 X zj}x define the
homeomorphism Z>* as follows: For 0 < k < rxr2 let k = k2rx + kv where
0 < k2 < r2, 0 < kx < rv Define
D*(k) = (k2, kx) E Z/r2 X z7>i.
For ƒ E €(Z/rxr2) define /)(ƒ) = ƒ o (Z)*)"1 E ^(z7r 2 X Zjrx).
THE MAPPINGS 1 X F(r2) AND 1 X F(r2). An element of ^(Z/rv <$(Z/r2))
determines an rptuple of elements of ^(Z/r 2 ). The mapping 1 X F{r^)
denotes applying the Fourier transform F(r2) to each of these r r tuple of
elements of ^(Z/r2y Define 1 X F(rx) similarly.
THE MAPPING M. For 0 < a < rx and 0 < b < r2 define

M(F)(b9 a) = e2™b/r^F(a9 b)9 F E 9(Z/rx X Z/r 2 ).

For those peoplejvho believe that the Fourier transform is related to the
groups Z/n and Z/n as presented in modern texts in Harmonic Analysis, the
mappings C, D9 and M involved in the Cooley-Tukey algorithm seem at best
formal and at worst arbitrary. We will show in the next section that if we use
the representation theory of the finite Heisenberg group, then the mappings
C, D and M are natural.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 873

2. The finite Heisenberg groups and the Cooley-Tukey algorithm. Before

going into the details of this section, we will, as promised, present some of the
material in §1.5 in a slightly different language. This language has the
advantage of showing the deep inter-relation of the finite^Fourier transform
F(n), the dual pairing of Z/n and its dual group Z/n, and the finite
Heisenberg group H(n).
Let C x (n) denote the multiplicative group of complex numbers e2™k/n,
k = 0 , . . . , n - 1. Let G(n) as a set be Z/n X Z% X Cx(n) and for
(ai9 bt, Cj) G G(ri), i = 1,2, define multiplication by
(av bx, cx)(a2, b2, c2) = (ax + a2, bx + b2, cx • c2 • (ax, b2))
where < , > denotes the dual pairing of Z/n and Z/n to Cx(w). Then G(n) is
a group with this law of composition. We claim that G{n) is isomorphic to
H(n). Let
<x:Z/n-»Cx(n)
be defined by a(k) = e2™k/n9 \\ follows that a is an isomorphism between the
additive group Z/n and the multiplicative group Cx(ri).
We now define /?: Z/n -» Z/n as follows: For k EL Z/n define
</, £(&)> = e2™'*/", all / G Z/n.
Now define y: //(n) -> G(n) by
(a1? a2, a3) -» (aj, ^(âf2), a(a 3 )).
It is an elementary computation to verify that y is an isomorphism. This
shows that the finite Heisenberg group is built from the group structures on
Z/n and Z/n combined with the dual pairing of Z/n and Z/n.
In this general setting, the finite Fourier transform is the isometry F(n):
L\Z/ri) -» L\Z/ri) defined by

(F(n)(f))(â) =J L 2 f(a)<a,âX âEZ/n.

\n a(EZ/n

Define the action of Z/n on L2(Z/ri) as follows: For a,b E Z/n define
(T(a)f)(b)=f(b + a).

Similarly define T(a), for â, 6 G tfh, of ƒ G L\Z?n) by

(r(â)/)(è)=/(6+â).
Let G>r(«) be the group of linear transformations of L2(Z/ri) generated by
T(a), a G Z/n, and F'l(n)T(â)F(n)9 â G Z/n. We wish to now obtain a
matrix representation of the group GF(n) in order to understand what the
group GF(n) is really like. To do this, let fa, a = 0, . . . , n - 1, be the basis of
L2(Z/n), where fa takes the value 1 at a G Z/n and 0 at all other points.
Clearly T(l)(fa) = fa+x, and so, relative to the basis f 0 , . . . ,ƒ„_!, T(l) has the
874 L. AUSLANDER AND R. TOLIMIERI

matrix representation
[0 1 0 ••• 0]

0 1 * . 0 '
I0 1 !
[ 1 0 ••• OJ
Since (T(\)F(n)fa)(â) = <a, î><a, â}/Vn , we have
{F(n)-ln\)F(n))fa = (a,\>.
Thus, using the isomorphism y defined above, we have F(nylT(l)F(n)(fa) has
the matrix representation
f e2mO/n 0 I

I 0 e2iri(n~l)/n\

This proves that GF(ri) has the matrix representation p2 of NÇL/n). This
argument shows the deep relation between the group NÇL/n) and the finite
Fourier transform F{ri). It also proves that
F(n)p2F(nyl = px
as it is easy to see that px(N(Z/n)) is the same as the matrix group generated
by T(â), âeZ/n, and F{ri)T{a)F(riy\ a G Z/n.
We now proceed to the task of characterizing the mappings C, D and M
discussed in the previous section.
Consider the Z/n-Heisenberg group, NÇL/n), n = rxr2, rx > 1 and r2 > 1
where
N(Z/n) = {(a, b, c)\a, b, c G Z/n).
Let T(r2, r t ) and r(r„ r2) contained in NÇL/n) be defined by
r(r 2 , r t ) - {(aV2, 6'^, 0)|n' G Z / r „ V G Z/r 2 },
r ( r „ r2) = {{b'rv a'r» 0)|a' G Z / r p 6' G Z/r 2 }.
An elementary computation shows that T(r2, rx) and r(r 1 , r2) are subgroups
of NÇL/n). Let T be a subgroup of NÇL/n). Consider the homogeneous space
T \ NÇL/ri) and give this finite set the measure where each point has measure
one. Form L2(T \ NÇL/ri)). Since NÇL/n) acts on T \ NÇL/ri) by
R(g)(Tn) - Tng, n, g G N(Z/n)9
R(g), g G N(Z/ri), defines a unitary representation R of NÇL/ri) on L2(T \
N(Z/n)) by
(H(g)(F))(rn) = FÇTng), F G L2(T \ N(Z/n)).
It will often be convenient to view functions on T \ NÇL/ri) as functions F
on NÇL/n) such that
*Xw)-*•(*)> yGr, g Giv(z/ w ).
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 875

Let x be the character on the center z(N(Z/n)) given by

X(0, 0, c) = e2™/\
Then — x is the character given by
- x(0, 0, c) = e-2™'".
Let ^(x) denote the functions F(a, b, c) on N(Z/ri) such that
F(a, b,c + d) = el7rid/nF(a, b, c).
Define 9r( — x) analogously. Let
^ ( x , r(r 2 , r,)) = ff(x) n L 2 (r(r 2 , r.) \ JV(Z/n))
and
^ ( - X , r ( r „ r 2 )) = f ( - x ) n L 2 ( r ( r „ r 2 ) \ JV(Z/«)).
It is easily verified that
R(9(x, r(r 2 , r,))) = ff(x, IXf* rO),
* ( # ( - * r ( r l f r 2 ))) - * ( - * T(r l5 r 2 )).
Let i**(x) denote the restriction of R to ^(x, T(r 2 , A^)) and /t*( —x) denote
the restriction of i£ to ^ ( — x» FCv ri))-
THEOREM II.2.1. R*(x) is an irreducible unitary representation of N(Z/n)
that is unitarily equivalent to px or p 2 .
PROOF. i**(x)(0, 0, d) is e2mid by our definitions of &(x). Hence by Theo-
rems R.l and R.3 the proof of our assertion reduces to computing the
dimension of ^ ( x , T(r 2 , rx)). But the dimension of ^(x, T(r 2 , rx)) is easily seen
to be the same as the dimension of S r (Z/r 1 X Z / r 2 ) which is rxr2 or n. Hence
i?*(x) is irreducible.
(A similar argument shows that /?*(—x) is irreducible.)
Since R*(x)> is unitarily equivalent to p 2 , there exists a unitary operator
W:L\Z/n)-+9(X>T{r7,rx))
I
such that PF~ i?*(x)W^ = p2- Recall that fFis unique up to multiplication by
a complex number of absolute value 1.
We will now build W from 1 X F(r2) ° C, where C and 1 X F(r2) are as
defined earlier in this chapter. For ƒ G L2(Z/n) define
»K(/) = F(x9y, t) G ^ ( Z / w X Z / n X Z / / i )
by
jp(/) = ^ v / * 2 /0>2 + x)e2"iJr*y/n.
0<j<ri

W is the analogue of the Weil-Brezin map as defined in [19] and [7]. We will
now verify that W(f) G 3F(x, T(r 2 , rx)). It is clear that W(f) G ^(x). Hence it
remains to verify that
W{f){(ar2, brx, 0)(x,y, t)) = W(f)(x,y, t).
Now (ar2, brv 0)(x,y, i) = (ar2 + x, brx + y, t + ar2y). Thus the left side
876 L. AUSLANDER AND R. TOLIMIERI

above equals
e2m(t + ar2y)/nS£ /(j>2 + ar2 + X) e2™jr& + ^ "

= e 2 "*/"2 f((j 4- a)r2 + x)£><^ ö >^/* - ^(/X**y 9 t%

and we have shown that W{f) GÏÏ(x>T(r29 rj) or

To relate W to I X F(r2) o C, we note that W(f) restricted to the set

S = {(x,y9 0)|0 < x < r9 0 < y < rx) equals (1 X F(r2) <> C)f. Further, since
W(f) G £F(x, r(r 2 , z^)), knowing W(/) on the set S uniquely determines
W(J).
It is straightforward from the discussion above and the properties of the
mappings 1 X F(r2) and C to conclude that Wis a unitary operator.
It remains to verify that
W~lR*(X)W = p2.
But this is a formal calculation that the interested reader may easily verify.
This shows that 1 X F(r2) ° C is essentially an intertwining operator between
two irreducible unitary representations of NÇL/ri).
We come next to the mapping M of the Cooley-Tukey algorithm. To
explain M one has to introduce a bit more of the structure of N(Z/ri). To be
precise, we must introduce a particular automorphism K of the group
N(Z/n). For (x9 y9 t) G NÇL/n) let
K(x9y91) = (x9 y, -t + xy).
It is a straightforward computation to verify that K is an automorphism of
NÇL/n) and that K2 is the identity automorphism.
Consider the following general situation. Let B be a group, G a subgroup of
B and ^4 an automorphism of B with ^4(Gj) = G2. Let ƒ G £F(i?) be such that
f(gb)=f(b% bEB9gGGv
If g G Gl9 A~lg G G2, and so if f(gb) = ƒ(£), g G Gl9 we have
f{A-l(gb))-f{A-\g)A-l(b)) =f{A~\b)).
_1
Hence if f* = ƒ © ^4 we have
ƒ"( &*) = /*(*)> & e G2 and i G 5 .
Now apply this to the special case of the functions W{f) and the automor-
phism K above. Because K = K~l we have W{f) ° K is invariant under
r( r i> ^2) a n < * s o * s m ^(""X* ri» ^2)- Explicitly
W(f)(K(x9y9 t)) - W ) 0 % x, - I + v )
= 2 /(^2 + y)e27Tiar*/ne~2nit/ne2"ixy/n.
0<a<r
But the mapping W{f) ° K on (x, 7, 0) is the same as applying M to
W(f)(x9 y9 0). This supplies us with the group theoretic interpretation of M
that we sought.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 877

Now R*(~~x) a n irreducible unitary representation of NÇL/ri) on

&(-X>r\>r2)* ° n e c a n s h o w t h a t D X ° * x ^(riXx>y9 °) determines an
intertwining operator between R*(~~x) anc* a n irreducible unitary representa-
tion of N(Z/ri) on L2ÇL/rxr^. Since the discussion is similar to that given
above for 1 X F(r^) © C, we will not go any further into the specific details.
3. The Plancherai theorem for the reals and the Cooley-Tukey algorithm. We
will now show the general nature of the Cooley-Tukey algorithm by showing
how it can be used to prove the Plancherel theorem for the reals. Just as for
the Cooley-Tukey algorithm, the forthcoming proof of the Plancherel theorem
has an interpretation in terms of nilpotent harmonic analysis. We will not
present this material because a full discussion would be quite long. For the
interested reader the material in Chapter I of [3] or Chapter I, §5, of [1] can
be modified along the lines of the material in the previous section to obtain
all the group theoretic ramifications of our method of proof.
Let R denote the reals, Z the integers and T the circle group or R/Z, and
let F( ) denote the Fourier transform of the group in the parens. The diagram
of the Cooley-Tukey algorithm becomes in this setting the diagram below.

*(R)- scr, z)
1 xF(Z)

9 (J x Z)
F(R) M

ff(Z x T)

1 xF(T)
>
D~ i
ff(RV •ff(Z,T)
Recall that as groups R is isomorphic to R, Z is isomorphic to T and T is
isomorphic to Z.
Define C : R -* [0, 1) X Z, where we identify T with [0, 1), as follows: If
JC G R, x = y + n, 0 < y < 1, n G Z, let

C*{x) = (y, n).

For ƒ G #(R) define C(J) = ƒ ° C*'\ For F(y, ri) e 9ÇT, Z) define

1 X F(Z)(F(y, «)) - 2 ƒ"(/. n)*2"**, 0 < £ < 1.

M is now interchange of | and 7 and multiplication by e2™*?. This gives

((1 X FÇt)) o M o (lx F(Z)) o C)f * ff 2 ƒ•(ƒ, ^é^^e^é*** <fy.

For the time being, let us proceed formally and interchange integration and
878 L. AUSLANDER AND R. TOLIMIERI

summation to obtain

(1 X F(T)) o M o (1 x F(Z)) o c(f) = 2 CF(y> n)eM<n+*Xm+*> dy.

n = — oo •'o
Applying D ~x we obtain
F(R)(f)(2ws) = r f{x)e2"ixsds
•'-oo

where x = n + y, s = m + £.
We will now see how all this works rigorously.
Consider L2(R) with Haar measure dx and T X Z with Haar measure such
that the measure of T X 0 is one. It is easily seen that C: L2(R) -» L2(T X Z)
is a unitary operator; i.e., a norm preserving surjection. Assume that F(T):
L2(T)~>L2(Z) and F(Z): L 2 (Z)-* L2(T) are unitary operators. Since M
consists of multiplying each value of a function by a number of absolute
value 1, it is trivial to verify that M is a unitary operator. This shows that
D~l • 1 X F(T) • M- 1 X F(Z) • C: L2(R) -> L2(R)
is a unitary operator and shows that F(R) is a unitary operator.
We are left with the task of interpreting the interchange of integration and
summation used in deriving the integral formula for F(R). We will close this
chapter with a discussion of this process.
We may view
G(y, 9 - M - l X F(Z) • C(f) G L2(T X T).
As such, G(y, £) has a Fourier expansion

where convergence is in the L2 norm. Now

flG(y, i)e2™> dy = lx F(T)(G(y, ©).

But the integral on the left is easily seen to be

ƒ
We next note that if gm and g G L2(T) with lim^^^ gw = g in L2(T), then the
mth Fourier coefficient of gn converges to the mth Fourier coefficient of g.
This shows that if ƒ G L2(R), then

fairs) = jiir^ ƒ n f(x)e2™< ds

where the limit is in the L2 norm.

CHAPTER III. ALGEBRAIC COMPLEXITY AND THE FINITE FOURIER TRANSFORM

1. Basic ideas in algebraic complexity. S. Winograd's work on algebraic

complexity is very interesting at both the practical and the theoretical levels.
He has in [20], [21] produced algorithms for the finite Fourier transform that
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 879

are more efficient than Cooley-Tukey. At the theoretical level, he has also
succeeded in [22] in defining the concept of essential multiplications or
divisions, or more briefly, essential m/d. This concept is important because
experience has shown that algorithms that minimize the essential m/d, or
minimal algorithms, possess interesting algebraic structures. We begin with a
formal definition of an algorithm that will permit us to define the concept of
essential m/d.
Let G be a field, called the field of constants, and let F = G{xx, — , xn) be
the purely transcendental field extension of G obtained by adjoining the
indeterminants xx, .. . , xn to G. We use Ü = {co^i = 1, 2, 3, 4} to denote the
field operations of addition, subtraction, multiplication and division, respec-
tively. For fx, f2 e F we will use wt{fl9 f2) to denote the result of applying the
binary operation co, to fx and f2 with the convention that

Let B c F be the given objects for our algorithms. Usually, B = G \J

{xx, . . . , xn}. But often, in theoretical discussions, other given objects will
play important roles.
DEFINITION. We will define an JV-step algorithm a over (F, B) inductively.
Step 1. Choose either an element of B or choose w(l) G Ö and an ordered
pair (a(l), 6(1)) from B. Require that Oa(l) = co(l)(a(l), 6(1)) be defined and
call OJY) the output of the first step of the algorithm.
Step 2. Choose either an element of B or choose co(2) G Q and an ordered
pair (a(2), 6(2)) from B u O a (l). Require that Oa(2) = <o(2)(a(2), 6(2)) be
defined and call Oa(2) the output of Step 2 of the algorithm.
Assume the first k steps of a have been defined.
Step k + 1. Choose either an element of B or choose u(k + 1 ) G 2 and an
ordered pair (a(k + 1), b(k + 1)) from B u Oa(l) U • • • U Oa(k). Require
that
Oa(k + 1) = w(k + l)(a(k + 1), b(k + 1))
be defined, and call Oa(k + 1) the output of the k + 1 step of the algorithm.
If a has N steps, we will call it an iV-step algorithm. We call Oa(k),
1 < k < N, the output function of the algorithm a.
Two algorithms a, /? over (F, E) ^vill be said to be equivalent if
Oa(k) = Ofi(k), Kk<N.
The k step of an algorithm a is called an m/d step if co(k) is multiplication
or division; i.e., if co(k) = co3 or co4. Clearly equivalent algorithms need not
have the same number of m/d steps. (For instance, x + x = 2x and 2 • x =
2x.) This may serve to motivate the following definition.
DEFINITION. A step k for an algorithm a is called m/d essential if 0{k) is
not in the G-linear span of B u O(l) u • • • U 0{k - 1).
DEFINITION. Let f x , . . . ,fs G F. We will say that the JV-step algorithm a
over (F, B) computes fl9 . . . ,fs if for each fi9 1 < i < s, there is an integer
k(i), 1 < k{î) < N, such that Oa(k(i)) = £.
880 L. AUSLANDER AND R. TOLIMIERI

It is obvious that f l 9 . . . 9fs G F can be computed by an algorithm over

(F, B) if and only if fl9 . . ., fs are in the field generated by B.
DEFINITION. We will say that a is a minimal algorithm for computing
fl9 . . . , fk if, among all algorithms over (F, B)9 a has the minimum number of
essential m/d.
The m/d number for computing fl9. . . , fk over (F, E) is the number of
essential m/d steps in a minimal algorithm for computing ƒ,, . . . ,fk.
Let a be an TV-step algorithm over (F, E) with output function Oa(k)9
I < k < N. Let 5: 2? -> F be a mapping. Then, under certain circumstances, s
determines an TV-step algorithm s(a) over (F, s(B)). We will now describe
this.
DEFINITION. Let B c F and let [2?] be the subring of F generated by 2? and
(B) be the subfield generated by B. A mapping s: B -» F is called a basis for
substitution if the following are satisfied:
(a) There is a ring homomorphism 5*: [B] -> F such that s*\B = 5.
(Note. Since 2? generates [5] as a ring, if s* exists, it is unique.)
(b) For bl9 b2 G [B]9 if s*(62) ¥= 0, 5*(61)/^*(*2) is well defined. (Hence s*
can be extended to as much of (2?) as possible.)
Let a be an TV-step algorithm over (F, B) and 5: 5 - > F a basis of
substitution. Let the k steps of a be co(k)9 (a(k)9 b(k)). Then s(a) is defined
and has k step co(A:), 0*(a(fc)), s*(è(/c))) or s(Oa(k)) if Oa(A:) G B provided
s*(b(k)) 7*= 0 whenever o)(k) = <o4. If .s(a) is defined then s(a) is an TV-step
algorithm over (F, s(B)) and 05(a)(fc) = s*(Oa(k)).
DEFINITION. Let a be an TV-step algorithm over (F, B) and /? an M-step
algorithm over (F, 5'). Let a ° /? be the TV + M-step algorithm over (F, 2? u
5') whose k step, 1 < fc < M, is the A: step of /3 and whose k step, M + 1 < k
< M + TV, is the k — M step of a.
Let a be an TV-step algorithm over (F, 2?) that computes f l 9 . . ., fr and let
a(/c) or b(k)9 1 < A: < TV, be in the subset bl9...9bs of B. Let /? be an (F, 5')
algorithm that computes bl9 . . . , £»,. Then a ° ft is an (F, 2?') algorithm that
computesfl9 ...,ƒ,..
In particular, if 5 is a basis of substitution such that s(a) is an (F, s(2?))
algorithm and /? is an (F, 5') algorithm computing s*(aa(k)) or s*(6a(&)) for
aa(k) or 6a(/c) G B9 we have s(a) ° /? over (F, 5') computes s(a).
2. Bilinear algorithms for the finite Fourier transform. S. Winograd has
recently devised algorithms for computing the finite Fourier transform that
work much better than the Cooley-Tukey algorithm. They are also of theore-
tical interest because they are based on expressing the finite Fourier trans-
form F(p)9 p a prime, in terms of the complex group algebra of the multi-
plicative group Z/p*. We will denote this group algebra by C(Z//?*).
Let us begin by writing Winograd's algorithm for/? = 7. Let
6
Aj = S e2"Vk'\9 j = 0, . . ., 6.
*=o
We will present the algorithm as a sequence of additions, then multiplica-
tions and then additions.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 881

S , = ax + a6, ^2 = a
l *~ a6> $3 = ö
4 + a
3' ^4 = «4 " a
3'
a a
5 5 = a2 ~ a5, ^6 ^ 2 + 5> «S7 = S*! + 5 3 , Sg = 5 7 + S 6 ,
S9 = Ss + a0, *^10 =
*^1 "~ *^3' *^11 ~ *^3 "" *^5> *^12 ^ *^5 "~ S\,

^ 1 3 = = S2 "H o 4 , *$14 =
^ 1 3 "*" ^6» ^IS =
^ 2 "" ^ 4 ' ^16 =
^ 4 "" ^6>

s„- =
»$6 "~ *^2>

Let u = 2TT//7.

/ cos w + cos 2w + cos 3w A _

W = 8
' I 3 V* '
/ 2 cos w — cos 2w — cos 3w \ _
™2 = ^ 3 )SÏ0,
/ cos u — 2 cos 2w + cos 3M \ _
m3 = n
I 3 r '
/ cos u + cos 2w — 2 cos 3w \ _
m4 = ( 3 ]S 1 2 ,
./ sin u + sin 2w — sin 3w \ „
w5 = ^ 3 JSH,
A 2 sin u — sin 2M + sin 3w \ _
™6 = l[ 3 JSi5>
./ sin u — 2 sin 2w — sin 3w \ _
™7 = ' ( 3 J5.6,
./ sin w 4- sin lu + 2 sin 3w \ _
m8 = ^ JS 17 .
^18 = £9 + ml9 5 1 9 = *S18 4- m2, 5*20 == ^19 + m3, S2l — o 1 8 /W2,
5*22 = s2l - w 4 , ^23 =
^18 "~ W
3> ^24 =
^23 "*" m
4> ^25 = m
5 + m
6>

•^26 = S 25 + m7, 527 = m5 — m6, S2% = o27 — Aw8, S2$ = w 5 — w 7 ,

=: w = =
^30 029 + 8> ^31 *^20 ~*~ ^ 2 6 ' ^32 ^ ^ 2 0 ~~ *^26> ^33 *^22 "*" ^ 2 8 '
= =
•^34 " *^22 "" *^28> ^35 *^24 + ^30' ^36 ^ 2 4 ~~ ^ 3 0

^ 0 == S9, Ax = o 31 , A2 = »>33, v43 == ^36'

^ 4 == ^ 3 5 ' ^5 =
^34' ^6 =
^32-

This algorithm requires 8 multiplications instead of 8 • 3 = 24 that

Cooley-Tukey requires for ^(8).
We now will describe some of the theoretical considerations upon which
the above algorithm was built.
Let co = elmi/p,p a prime, and coJk = e2viJk/p, 0 < y, k <p, and

A = 2 e2"ijk/pak = 2 <**kak. (1)

A: = 0 A: = 0

Since c&/k = 1 fory or k = 0 we have y40 = a0 + • • • + ap_x and

Aj ~a0 = Af, Kj<p- 1, (2)
882 L. AUSLANDER AND R. TOLIMIERI

where
P-I
Af= S <**kak, Kj<p-l. (3)
k— 1
Let {mv . . . , rrip^x} be the elements of (Z/p)x ordered so that mi • ny =
m^, where k = i-j mod/?. Then

('s « s ) ( 2 ^K)"1) - 2 2 « V * (4)

because mj(mkyl = m, implies y = h k mod/?. Since S ^4*/W/ is the right side
of (4), computing the finite Fourier transform is the same as computing a
product in C(Z//?*). Because Z/px is a cyclic group of order/? — 1, we can
obtain another method for computing the terms Af, / = 1,...,/? — 1. Let
m E rL/px be such that mp~l = l G Z/px; i.e., m is a generator of the
multiplicative cyclic group Z/px. Then m = ma for some 1 < a </>-l.
Hence

=
2 «S 2 «"^ (here A: m> = ( m / )

and

where 7r(a7) • ay = 1 mod/?. Hence

2^«W = (2««W)(2>W)W').
But, multiplication on the right side is the same as multiplying the expressions
as polynomials in m and reducing modulo (mp~l-l). This directly relates the
finite Fourier transform to the group algebra C(Z//? — 1).
We may also present the above discussion in matrix language as follows:
Consider the matrix equation
a
rA\ ' \ 1
= («*) •
A* *p-l\

A* = fia.
The first row of the square matrix on the right is w 1 , . . . , cop~l. Thus there
is a permutation of columns of this square matrix so that the first row of the
resulting matrix is coa, . . ., o)aP . This permutation can be achieved by
multiplying the matrix (o)iJ) on the right by a matrix P. Noting that P~l = P',
where the superscript t denotes the transpose, we have
A* = tiPP'a.
Then the first column of fi is (co1, . . ., cop~ly. Forming
P'A* = P'tiPP'a
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 883

an elementary computation shows that

or <ov
«"-».
P'£2P = (0 (0

aP-
CO" <*T CO

Thus, if we let
Yi y\
= p<4*
P'J* = P'a = = y
1
P~\ yP-x
we have
y = P'ÏÏPy
which is called the cyclic convolution of (Yv . . . , Yp_x) and (yv ••Jt-l)-
It is now a straightforward computation to verify that
y, + y2w + + y p . 1 « /, "" 2 = (coal + (oa« + + <o«'~V>-2)
Ui + ^ - i « + ' • • +72^~2) m o d ^ - 1 - 1).

The above discussion shows that the m/d number of F(p), can be bounded
above by the m/d number of the right side of the above equation.
Let us now formalize the problem to which the above discussion has led us.
Let

&{z) = 2 xtz' and S(z) = 2 yp%

i-O i-O

be two polynomials with indeterminates as coefficients and let

T(z) = R(z)S(z).
The a + h + 1 coefficients of T are a system of bilinear forms which we will
denote by f. Let P be a polynomial over G of degree n and let
r p (z) = R(z) • S(z) mod P.
Let fp denote the n coefficients of T. Then fp is a system of bilinear forms.
Let £ = G U {*!,. ^ . , xa,yl9... ,7 6 }. Our problem becomes to compute
the m/d number of T and Tp.
We will take up this problem in the next two sections. In the final section
of this paper we will return and discuss how the above problem relates to the
m/d number of F(p), p a prime.
3. General results on bilinear algorithms. In this section, we will prove three
general theorems, that, at the present state of the art, are of fundamental
importance in the theory of bilinear algorithms.
Let G be a fixed infinite field and let xv ..., xn,yx,..., ym be indetermi-
884 L. AUSLANDER AND R. TOLIMIERI

nants. Let F = G(xv . •. 9xn,yl9... 9ym) = G(x9 y) and let

n
a
Lu = 2 ukxk> l <j <m9l <i <t9

be a set of linear forms and let

A{x)y = (Ly)

We call A(x)y a system of bilinear forms. We wish to study the m/d number
of A(x)y.
Notice that the columns (rows) of all possible matrices A(x) form a vector
space C (R) over G. We define the G-column (G-row) rank of A(x) as the
dimension of the vector subspace of C (R) spanned by the columns (rows) of
A(x).
In this section F = G(x9y) and B = G \J {xl9.,., xn9 yl9... 9ym).

THEOREM III.3.1. Let A(x)y be a system of bilinear forms. If A{x) has G-row
rank s then this system of bilinear forms has m/d number greater than or equal
to s.
This theorem is actually much weaker than what is known to be true. Since
it is not harder to prove the more general result, we will prove it.
THEOREM 111.3.1', Let fv . . , , ƒ , E G(x9y) and let a be an N-step minimal
algorithm for computing f\, . • . , ƒ / . Then the m/d number offl9 . . . , ƒ / equals
the dimension of the vector space Wa defined as follows:
Let W* c G(x9y) be the vector subspace spanned by Oa(l)9 . . . , Oa(N) and
let L be the subspace of elements of the form 2 géxê + 2 hjyj + k9g9h9k EL G.
Define Wa = W* 0 L/L.
PROOF. Assume that ft = Oa(N). (A relabelling can always achieve this.)
Then let a! be the N — 1 step algorithm consisting of the first N — 1 steps of
a. Then a' is a minimal algorithm for f l 9 . . . , f„l9 aa(N), ba(N).
We will now prove Theorem III.3.1' by induction on N. Since a 1-step
algorithm that has no essential m/d computes an element of L, we have
proven the theorem for iV = 1.
By induction the theorem is true for a' and f{9 • . . 9ft^\9 aa(N)9 ba(N). Now
if the iV step of a is not an essential m/d step, a and a' have the same
number of essential m/d steps. But, clearly, W£ = W* and the theorem is
true.
If the N step of a is an essential m/d step Wa ^ W^ or else a is not
minimal. This again proves our result.
Theorem 111.3.1' implies Theorem III.3.1 once we observe that the row
rank of A(x) is the dimension of the vector space V spanned by 2 ci^x^j in
G(x9y). But V c W* and V n L = 0. This proves dim Wa > dim V.
Before going on to Theorems III.3.2 and III.3.3 let us pause to give an
application of Theorem III.3.1 to the problem posed in §111.2. Consider the
system of bilinear forms that are the coefficients of the product of the two
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 885

polynomials

2 xizi and 2 y^3*

This may be written in the form A(x)y, where A(x) has the special form
(assume a > b)
x0 0 0
•Xi Xr\ U

A(x)~
0

0 xa
where A(x) has 6 columns and a + b + \ rows and
PV
ƒ =

yb
It is easy to compute that the row rank of our special A(x) is a + b + 1.
Thus by Theorem III.3.1 the m/d number of A(x)y is greater than or equal to
a + b + 1. Later we will see that it is actually equal to a + 6 + 1.
THEOREM III.3.2. Le* ^(*).y be a system of bilinear forms. If A(x) has G
column rank s9 then the m/d number of A(x)y is greater than or equal to s.
Again it is no harder to prove a slight generalization of this and we will do
so. This is Theorem 1 of [22].
Consider G(xv . . . , * „ ) = G(x). Let <j> be a t X m matrix over G(x). We
shall use <j>l9..., <j>m to denote the columns of $. (Note. A(x) is a t X m
matrix over F.)
THEOREM III.3.2'. Let a be an algorithm computing <py over (G(x,y),B)
where B = G(x) U {yx, . . . ,ym}- If there are s vectors in {<J>l5.. . , <j>m] such
that no nontrivial linear combination of them with coefficients in G lies in the t
dimensional vector space G\ then a has at least s m/d steps of the following
form: w3(a(k), b(k)) with a(k) and b(k) & G or co4(a(A:), b(k)) with b(k) g G.
To relate Theorem III.3.2 and Theorem III.3.2', we must first show that if
A(x)y satisfies the hypothesis of Theorem HI.3.2 then A(x)y satisfies the
hypothesis of Theorem IH.3.2'. But a linear combination over G of the
columns of A(x) is in G' if and only if it is the 0 vector. This shows that our
first requirement is satisfied.
To see that the conclusion of Theorem III.3.2' implies the conclusion of
Theorem IH.3.2 merely note that the s steps guaranteed by Theorem HI.3.2'
are essential m/d for a.
886 L. AUSLANDER AND R. TOLIMIERI

Thus it remains only to prove Theorem III.3.2'. To simplify the language of

this proof and only here, we have called those m/d not excluded by Theorem
III.3.2' essential.
We prove the theorem by induction on s. Clearly an algorithm with no
essential m/d steps can only compute elements of the form 2 gptj + ƒ,
gj G G, ƒ G G(x). But if s = 1 there exists a matrix coefficient of <j> that is not
in G and so the algorithm must compute.

2 hjyp some hj £ G.
Hence we must have at least one essential m/d in the algorithm.
Suppose the assertion holds for a = N. Assume $ is such that at least
N + 1 of the vectors {<j>v . . . , <ƒ>„} have no linear combination over G which
is in G'. Let a be a minimal algorithm computing <py and let k be the first
integer such that an essential m/d step of a occurs at step k. Then, either

<>«(*)-(2&*+/)-(2 V , + / )
or
S
O (k) = &y'+/

for gi9 ht G G, and ƒ,ƒ' G G{x). Furthermore, we may assume the labelling
has been done so that one of the ht ^ 0, for otherwise the k step of a would
not be an essential m/d step.
Clearly s(yn) = - ƒ - 2 /y>„ s the identity on yx u • • • \Jyn-x U G(x).
G(x) is a basis of substitution. However s(a) may not be an algorithm. Since
there are only a finite number of divisions in any algorithm the substitution
s(a) can fail to be an algorithm only when s* applied to some finite set
{rv . . . , rm) c G(x)[.y] is zero. Choose g G G so that
s*(rj) + g * 0, j G 1, . . . , M.

This substitution yields an algorithm that computes tfy' where <j>j = ty —

*/^/i> J = 1> 2 , . . . , n — 1 and >>' = (>>!,... ,j>„_i)'. The number of essential
m/d steps in a' is at least one less than in a. This is because the image of the
k step in a' is not an essential m/d and the algorithm /? computing
g — y _ 2 /y, has no essential m/rf. But there are at least N vectors in
{</>i, . . . , <t>n-\} s u c h th a t n o nontrivial G-linear combination is in G'. Hence
by induction a' has at least N essential m/d and so a has at least N + 1
essential m/d.
It turns out that it is much easier to study minimal algorithms that use only
essential multiplications. The following theorem indicates why this is so. This
is contained in the proof of Lemma 2 in [22].
THEOREM III.3.3. Let A{x)y be a system of bilinear forms and let a be an
algorithm that uses the minimal number s of essential multiplicatons. Then there
exist 2s linear forms Lt{xyy)y Ll(x,y% i = 1,. . . , s, such that A(x)y = Um
where U is a matrix over G and m is the column matrix
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 887

fL.L',1

We call this a presentation theorem because it shows that no matter how

complicated the original a was we can produce a minimal algorithm for
computing A(x)y of the following form: Compute the linear forms L,, L/,
i = 1, . . •, s, multiply, using essential multiplications, the linear forms to
form the quadratic forms L, • L/, i = 1 , . . . , s, form linear combinations of
the quadratic forms Lt • L/, i = 1,. . . , s, to obtain the elements A(x)y.
Notice the algorithm presented at the beginning of §11.2 for computing F(l)
was of the above form. We will call algorithms of the above form quadratic
algorithms.
PROOF. Let kv . . . , ks be the steps of a where the essential multiplications
occur.
If m is a step of the algorithm a then
Oa(m) = Lm(0) + Lm(x) + Lm(y) + Lm{x2) + Lm(y2) + Lm{xy) + ...
where Lm(0) is a constant and Lm( ) is a form in the type of term in the
bracket. We will denote OJJk^) by L,(0) + . . . . Then we can compute the
system of bilinear forms A(x)y by linear combinations over G of Oa{k^
i — 1, . . ., s. Hence there is a linear combination of the L,-(xy), i = 1 , . . . , s,
terms that equals each bilinear form in A(x)y.
It is crucial to our argument to observe that we may modify the algorithm
to obtain a new algorithm a' without introducing any new essential m so that
L;(0) is always zero. This is because L,(0) G G and subtracting by it is not an
essential m. We will henceforth assume that a = a' or that the desired
modification has been made.
Now Oa(kj) = Oa(l)Oa(m) where / and m are steps of the algorithm. Note
that we may form Ut = Lt{x) + Lt(y) and Um = Lm(x) + Lm(y) without any
essential m. Further UlUm and Oa(l)Oa(m) have the same quadratic terms.
Since forming linear combinations preserves degree, it follows that we may
replace the terms Oa(l) and Ot{m) in our algorithm by Ul and Um, respec-
tively, and still compute the bilinear forms A(x)y. This proves our theorem.
4. Some minimal algorithms. In §111.3 we established some results that
enabled us to put lower bounds on m/d numbers. In this section we will see
how to use the Chinese Remainder Theorem to produce algorithms. We will
also prove that in certain cases we can actually compute m/d numbers.
One version of the Chinese Remainder Theorem goes as follows. Consider
the polynomial ring G[z] over a field G and let Pl9..., Pk G G[z] be such
that P( and Pj are relatively prime for i ^j. Then the proof of the Chinese
Remainder Theorem assures the existence of polynomials Qi9 i = 1,. . . , k,
such that
Qt^Sy mod Pj
where 8^ = 1 if i = j and 0 otherwise. The usual statements then stress the
888 L. AUSLANDER AND R. TOLIMIERI

following: Given Bv . . . , Bk in G[z] there exists a B such that

B = Bi mod Pt.
Indeed
5 = 2 5,8,.
will do.
But it also follows that if
B = Bi mod P(
then
5 ^ 2 A Ô I m ° d P whereP = 1 1 ^ .
It is this last assertion that we will need to construct algorithms. We will
now give two applications of this idea.
Let R(z) = 2? Œ a x,z' and S(z) = *2j=0yjZj. Using the notation of §111.2 we
wish to compute T or the coefficients of the polynomial T(z) = R(z) • S(z),
We will use the Chinese Remainder Theorem to produce an algorithm that
does this in a + b + 1 essential w/rf. Combining this with the results in
§111.3 we will have proven that the m/d number of f is a + b + 1.
Choose a0, . . . , aa+b distinct elements of G and let
a+ b
Q = n (2 - «,).
!«=0

Then g is a polynomial of degree a + b + 1 and so T(z) may be identified

with T(z) mod g.
Now let Qi = (z — a,-). Then g and Qj are relatively prime for i 7*7. We
observe that if
g, = Il («/ - «,),
yV/
then
Gi = gr*]l(z-aJ)
is such that
Gi=8y mod Qj.
Hence
T(z) = T(z) mod g = ( 2 G,(* • S mod &)) mod g

1=0

The above equation shows that T(z) can be computed in the a + b + 1

essential multiplications /*(«,) • ^(a;).
A second method for computing f starts by choosing a + b elements of G,
Pu • • • 9fia+ba n d u s e s the identity
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 889

R(z) • S(z) - R(z) • 5(2) mod *II (z - ft) + xj^Ji (z - fl).

/-l 1=1
As before i?(z) * S(z) mod H?*f (2 - A) *s computed by the Chinese Re-
mainder Theorem using a + è multiplications and x a * ^ is the a + i + 1
essential multiplication.
We will say that the second method uses a0 = oo in the first method.
Now if P = zn + 2£~ô ft2'» s i n c e reducing mod P involves no essential
m/d, we can compute Tp if a > n and 6 > n in 2rc — 1 multiplications.
Let R{z) and 5(2) be of degree n~\ and let P = II*Li P, where the Pt are
pairwise relatively prime and Pg = P/ where P, is irreducible. By the Chinese
Remainder Theorem there exist Qf e G[z], i = 1 , . . . , k, such that
Qi=S(/ modP,
and
T(2) modP = ( 2 ô / ( ^ - ^ m o d P J ) ) modP
= (Sô/(^'S) modP,) mod P.
Since multiplying by Q{ and reducing mod P involve no essential multiplica-
tions, we have that all the essential multiplications occur in computing R • S
mod P,. If the degree of Pr = /*,., then 2*L 1rç?— 1 and by the above discus-
sion we can compute R • S mod Pt in 2nt — 1 essential multiplications. Then
our algorithm takes 2*» i(2w/ — 1) = 2/t — A: essential multiplications to com-
pute T(z) mod P or fp in the notation of §111.2.
We will now show that all minimal bilinear algorithms for computing T are
almost the same as the two algorithms we discussed above.
THEOREM III.4.1. Any bilinear algorithm for computing T in a •¥ b + \
essential multiplications involves computing the a + b + \ bilinear forms

where g ^ O , h ^ 0 E G, a0, . . . , aa+b E C l l o o and at =£ otj for i =£j.

{Assume a > 6.)
PROOF. Let T be the set of bilinear forms defined by

fXoO — 0]

X& Xo

Xa Xb
(5)-*
0

0 ... ox,
890 L. AUSLANDER AND R. TOLIMIERI

By our presentation theorem, we have that there exist m^ . >. 9ma+b bilinear
forms where
m( « ( 2 aixi + 2 *jty)(2 a'iXi + 2 bfo)
and an (a + b + 1) X (a + b + 1) matrix U over G such that
m0
Xy = tf I !
m. la + b
The essence of this theorem is to prove that each

m=
' {4ox>a')[hkyjaJ)
V +
(£ ' *M
or
OT
, = «*«y*- (0)
As we discussed in §111.3, all the rows of X are linearly independent over
G. Hence the set of bilinear forms Xy span an a + b + 1 dimensional space.
Hence the set of bilinear forms Um span an a + b + 1 dimensional subspace
of the space of bilinear forms. This implies that U is nonsingular and so we
may let W = U ~l and write
WXy = m. (1)
Let (H>O, . . . , wj+j,) be the ith row of fF. Then substitution in (1) yields
^o
(2)
^
This implies that the bilinear form on the right side of (2) can be computed in
1 essential multiplication. By Theorem IH.3.2 of §111.3 the column rank on
the right side must be 1, or all the forms S^.Q wj+kxp & = 0 , . . . , 6, are all
(7-multiples of one. This implies that the matrix

w:
w; M>a+\
(3)

wk wat + b
has rank 1.
We claim this can happen only under two circumstances: either Wç —
0 , . . . , wj+fc-i = 0, w^+b =£ 0; or there exists a, (which may be zero) such
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 891

that
wj = «X, j = 0,. . . , a + b where a{ = (a,)7
(when 0° = 1).
We may verify this assertion as follows: We have two cases to consider.
Case 1. M>O = 0. Then since (3) has rank 1,
WQW2 — (w[) = 0 or wj = 0 (4)
and
l
wJwj - (wrf = 0 or w 2 = 0. (5)
We may proceed by induction to verify that w^ = 0, 0 < & < a + &— 1.
Since rank of (3) is 1, wla+b =£ 0.
Case 2. H>0 T^ 0. Let w[ = Z^JWQ and w2 = ^2w^. By (4)
^2 "~ ^1
and by (5)
1 ^ 3 "~~ A' i Or K"\ ~~* /C i •

We may proceed by induction to verify that if

wj = ktwl
Then kt = k[. We let a, = kv
Since W is nonsingular at most 1 row can be of the first kind and for two
rows with WQ=£ 0 and M^^Owe must have a, ^ aJ9 i ¥=j. Thus (2) becomes

\y-0 7=0 y=0 V

n
or mi has the form (0). This proves the theorem.
0 •• • 1
Wn 0 ]| 1 ua +b
a

w=
w.a + b
1
-a
x
+b
<*a + b a+b

or
„a + b
Wn
a.
W =
W,* + *>
„a+b
1 «a + * ^a + b
REMARK. We see that the Vandermonde matrix enters into every bilinear
minimal algorithm a.
We will now prepare ourselves to prove the following result. Let P = zn +
892 L. AUSLANDER AND R. TOLIMIERI

2/.o 8iz*> where P = Pl and P is irreducible over G. If

n-\ n-\
R = 2 *,*'> 5 = 2 J^'>
i-O i-O
then the system of bilinear forms Tp has m/d number 2n 1.
Let CP be the companion matrix of P acting on the column vector space
V*.
•ft

Cp =

o 1 «.-
The minimal polynomial of CP is P itself. This means that P(CP) is the zero
matrix and any other polynomial with this property is divisible by P. Now let
v G V,v¥=0. Consider the set § of polynomial Q such that
vQ(Cp) - 0.
Clearly S is an ideal, $ D P. Since CP is nonsingular z £ 5 and 5 ^ G[z].
Let (P) be the ideal generated by P. If B and P are relatively prime then the
ideal generated by B and P is G[x]. Since 5 ^ G[z], 5 c (P). Hence if Q is
not divisible by P, i>ö(C/>) 7^ 0 any v G V* and so Ô(CP) is nonsingular.
LEMMA III.4.2. Let VP = {v G F| /Aere ex/ste a polynomial Q of degree < n
and vQ(Cp) = 0}. Then dim Fp < /z.
PROOF. Let W = {t> <E FluP'-^Cp) = 0). Claim W = F P . Clearly WKc
VP. If t>ö(Q>) = 0, then by the above discussion
Q = P'g', Ö' relatively prime to P,
where / > r > 0 and Ô'(Q>) is nonsingular. But then vPr(CP) = 0 which
implies vPl~\CP) = 0.
LEMMA III.4.3. Let CP be the companion matrix to P(z). If

t =

'n-\
then the coefficients of z S"«o 4 Z ' m ° d P(z) are
CP(t).
This is essentially the definition of CP.
LEMMA III.4.4. Let R(z) = S?!»1 xizi and S(z) = 2"~ô ^ ^ ^ fp be the
system of bilinear forms that are the coefficients of R(z) • S(z) mod P(z). Let
TP = >4(x)y as in $1113. Then
A{x) = (X,CpX9...,C^lX)
where
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 893

X=

CP is the companion matrix to P, and Cp is the a power of the matrix CP.

PROOF. We have, since R(z) = 2 xizi and S(z) = 2 yjZJ, that
S(z)-R(z)=y0^xizi+ylz'2xizi+ • • • H - ^ z ' - ' S x,z'.
By Lemma III.4.3 we have
S(z)-R(z)=yQX+yiCPX+ -• • +yn-.{C£-lX.
The coefficients of zk in S{z) • i?(z) is 2 >>*£*, where ^ is the fc entry in the
column vector CpX, k = 0, 1.
We are now in a position to prove the following theorem.
THEOREM III.4.5. Let R(z) and S(z) be polynomials of degree n — 1. Let
P = Pl where P is irreducible over G and let deg P = n. The minimum number
of multiplications needed to compute Tp is In — 1, where TP is the system of
bilinear forms that are the coefficients of R(z) • S(z) mod P(z).
PROOF. Let r be the minimum number of multiplications needed to com-
pute Tp. By Theorem III.3.3, we have
A(x)y = Um
where U is an n X r matrix over G and
L, • L\

m
L, L)
Let V be an n-dimensional G-vector space, let F* be its dual space and let
w E V*. Then wA(x) =£ 0 because its first coefficient is

Hence wUm is not zero and so wU is not zero. Since w was arbitrary, this
shows that the rank of U is n. By reordering columns, if necessary, we may
assume that we have a nonsingular n X n matrix W since
WU=(I\U')
where / is the n X n identity matrix.
Let VP be as in Lemma III.4.2. Then, because W is nonsingular, there
exists a row of W, say the first, denoted by w, which is not in VP. Then
wA(x)y is a bilinear form and
wA(x)y = (1 0 . . . 0 u\ . . . uxr_n)m.
Thus the bilinear form on the left, above, can be computed using r — n + 1
multiplications. We now claim that the n columns of wA(x) are independent.
894 L. AUSLANDER AND R. TOLIMIERI

This would imply our theorem for by Theorem III.3.2 r — n + I > n or

r > In - 1.
To show that the columns of wA{x) are independent, assume that

0 = 2 wCpX-oii = w 2 *iC*P )X.

i=0 \i=0 /
Since the elements of X are indeterminants over G

w
[%aiCH=0-
But w £ VP and the above contradicts Lemma III.4.2.
l k
COROLLARY III.4.6. Let P = P[\ . . . , P k where Pt are distinct irreducible
polynomials over G. Then TP can be computed in In — k multiplications and no
divisions.
This follows easily from the discussion at the beginning of this section.
Before coming to the final result to be proven in this paper, we would like
to remind the reader of the relation between the Chinese Remainder Theorem
and the rational form theorem for a matrix. __
Again let P = Pv . . . , Pk where Pt = P/1 and the Pt are distinct irreducible
polynomials in G[z]. Let CP be the companion matrix to P acting on V*. By
the Chinese Remainder Theorem

< ? [ * ] / * - 2 .®G(z)/P,
where equality denotes isomorphic. The rational canonical form theorem says
that there exists subspaces V* of V* such that
(a) V* = 2 0 Vf9
( b ) C P ( F ? ) - J?,
(c) Cp\Vf = CP,
where CP is the companion matrix to Pr
THEOREM III.4.7. Let P = Pv . . . , Pk where Pt = Pli and Pt are distinct
irreducible polynomials in G[z]. Further let the degree of Pé = nt and n =
2 ? . , nt. Let R(z) = 2 ^ x.z1 and S(z) = ^nr}ytz\ If T = RS then fP
cannot be computed with less than 2n — k essential multiplications and no
divisions.
We have already seen in Corollary III.4.6 that the m/d number is bounded
above by In - k. In the proof of the theorem we will need the following
purely technical lemma. We will state and prove this lemma below, but the
reader may prefer to skip directly to the proof of Theorem III.4.7 and return
to the lemma later.
LEMMA III.4.8. Let P = Pl9 . .., Pk be as in Theorem III.4.7 and let VP9
i = 1, . . . , k, be as in Lemma III.4.2. Let W be a nonsingular n X n matrix
and let Wl be the first nx columns of W, let W2 be the next n2 columns of W,
etc. Then each Wl has a row wl(j{i))J a function of i, such that w\j{î)) ^ VP.
Further, there exists a 1 X n matrix /? with nonzero entries only at the j(i),
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 895

i = 1, . . . , k9 coordinates such that

/3W=(yi,...,yk)
where yj is a I X ttj matrix j = 1,. . . , k and yy £ VP.
PROOF. Since W is nonsingular, each matrix W, j = 1, . . . , k, has rank iij.
By Lemma IH.4.2 there must be a row w\j(ï)) £ VP. Consider the set U of
1 X n vectors with nonzero entries only aty(0> i = 1, . . . , & . £ / i s a fc-dimen-
sional vector space.
In W' consider the row vectors w'(J(i))9 j(J) = 1,. . ., k, and form
S f . i Pj^wyO)). The set U(i) of p such that

2 V 0 ' ( 0 ) « VP(
i= i

is the complement of a proper linear subspace in U. Hence Pi ?= i U(i) is not

empty. Any point in H U(i) satisfies the conclusion of Lemma III.4.8.
PROOF OF THEOREM. By our discussion about the Chinese Remainder
Theorem and Rational Canonical Form Theorem, we see that we may assume
(without use of essential m/d) that fp corresponds to Ai(xi)yi and f to
A(x)y where

Ax{x') 0 y1
A{x)~ y=
0 k
M* ). y\
Let A(x)y = Urn, by Theorem III.3.3. Then U is an n X t matrix over G
where / is the number of essential multiplications. Since all rows of A(x) are
linearly independent, the rank of U is n. Therefore, there exists an n X n
nonsingular matrix W such that
WU = (ƒ | U'), where I is the n X n identity matrix.
Applying Lemma III.4.8 to W and letting /? be as in Lemma III.4.8, we
consider the bilinear form
/3WA(x)y = p(I\U')m.
We claim that at least n multiplications are needed to compute this bilinear
form. We prove this by showing that the column rank of the left side above is
n. Let y = PW and consider every nontrivial linear combination of the
column of fiWA(x) = yA(x). Substituting CJPx* for they element oiAfa*) we
obtain this linear combination as
2Y/(2^)X'.
This vanishes only if ^ 2 ctyC^ = 0 for i = 1,. . . , k. But y, £ VP and so
cty = 0 for all i andy. Thus by Theorem III.3.2 it requires at least n essential
multiplications to compute yA(x)y. But /?(/\U') has at most k + t — n
nonzero coefficients. Hence
k + t ~ n> n or t > In — k
and our theorem is proven.
896 L. AUSLANDER AND R. TOLIMIERI

5. Algorithms for computing the finite Fourier transform. In §111.2 we

presented S. Winograd's algorithm for computing F(7). We are now in a
position to outline the steps for creating that algorithm as discussed in [21].
However, before proceeding to this, let us at least mention a fundamental
result proven in [24].
THEOREM 111,5.1. The m/d number for computing the finite Fourier transform
on a prime number p is 2p — 3 — £(p — 1), where £(H) is the number of d such
that d\n.
We cannot discuss the proof of this theorem in this paper as it is too long.
It does show that the algorithm of §111.2 which uses 8 multiplications achieves
the minimum number of multiplications for computing F(7).
We will now outline the steps that are followed in creating the algorithm
for F(7).
Step 1. Consider the cyclic group (Z/7)* of order 6 and show that
e 2m3/7 = w3 i s a generator of (Z/7)* and that co1, co3, co2, co6, co4, co5, corre-
spond to 1, (co3), (co3)2, (co3)3, (co3)4, (co3)5.
Step 2. This enables us to show that the Fourier transform can be
computed from the coefficients of the polynomial
R(z) • S(z) mod z6 - 1 = (co1 + œ3z + Ù)2Z2 + co6z3 + coV + coV)
* (ax + a5z + a4z2 + a6z3 4- a2z4 + a3z5) mod z6 - 1
in the notation of §111.2.
Step 3. We verify that
x« - 1 = (x - l)(x + \){x2 + x + 1)(JC2 - x + 1)

** *V *2> * 3 > *4*

Step 4. If
Qi = 80 mod PJf ij « 1, 2, 3, 4,
as in §111.4, we compute and find

Ôi = \(x + l)(x2 + x + l)(x 2 - x + 1),

Q2 = - l(X - \)(x2 + JC + 1)(;C2 - X + 1),

Ö 4 = ~ ( f ~|)(* 2 ~l)(* 2 +*+l)-

Ste/? 5. Compute
JÇ mod Pi9

and so compute
R • S mod z6 - 1 = 2 Ôi(* mod />,.)($ mod /» ) mod z6 - 1.
 COMPUTING WITH THE FINITE FOURIER TRANSFORM 897

Step 6. Apply the Chinese Remainder Theorem twice more to compute

R3SZ mod P3 and R4S4 mod P4
where
#3 = R mod P3 and R4= R mod PA
S3 = S mod P3 and S4= S mod P4.
Step 7. Combine all the previous computations into a bilinear algorithm.
REFERENCES
1. L. Auslander, Lecture notes on nil- theta functions, CBMS Regional Conf. Ser. in Math. no.
34, Amer. Math. Soc, Providence, R. I., 1977.
2. L. Auslander and J. Brezin, Translation invariant subspaces in L2 of a compact nilmanifold. I,
Invent. Math. 20 (1973), 1-14.
3. L. Auslander and R. Tolimieri (assisted by H. E. Rauch), Abelian harmonic analysis, theta
functions and function algebras on a nilmanifold. Lecture Notes in Math., vol. 436, Springer-
Verlag, Berlin and New York, 1975.
4. , Algebraic structures for © 2 n > i L2(Z/ri) compatible with the finite Fourier
transform, Trans. Amer. Math. Soc. 244 (1978), 263-272.
5. R. Bellman, A brief introduction to theta functions, Holt, Rinehart and Winston, New York,
1961.
6. Z. I. Borevich and I. R. Shafarevich, Number theory, Academic Press, New York, 1966.
7. J. Brezin, Harmonic analysis on nilmanifolds, Trans. Amer. Math. Soc. 150 (1970), 611-618.
8. , Harmonic analysis on compact solvmanifolds, Lecture Notes in Math., vol. 602,
Springer-Verlag, Berlin and New York, 1977.
9. J. W. Cooley, P. A. W. Lewis, and P. P. Welch, Historical notes on the fast Fourier
transform, Proc. IEEE 55 (1967), 1675-1677.
10. J. W. Cooley and J. W. Tukey, An algorithm for the machine calculation of complex Fourier
series, Math. Comput. 19 (1965), 297-301.
11. I. J. Good, Analogues of Poisson9s summation formula, Amer. Math. Monthly 69 (1962),
259-266.
12. G. H. Hardy and E. M. Wright, An introduction to the theory of numbers, Clarendon Press,
Oxford, 1938.
13. K. Ireland and M. I. Rosen, Elements of number theory, Bogden and Quigley, New York,
1972.
14. S. Lang, Algebraic number theory, Addison-Wesley, Reading, Mass., 1970.
15. J. H. McClellan, Comments on "eigenvector and eigenvalue decomposition of the discrete
Fourier transform"JEEE Trans. Audio and Electroacoust. (1972), 65.
16. J. H. McClellan and T. W. Parks, Eigenvalue and eigenvector decomposition of the discrete
Fourier transform, IEEE Trans. Audio and Electroacoust. March (1972), 66-74.
17. H. Rademacher, Lectures on elementary number theory, Blaisdell, Boston, Mass., 1964.
18. R. Tolimieri, The multiplicity problem for ^-dimensional solvmanifolds, Bull. Amer. Math.
Soc. 83 (1977), 365-366.
19. A. Weil, Sur certaines groupes d'operateurs unitaires, Acta Math. I l l (1964), 143-211.
20. S. Winograd, On computing the discrete Fourier transform, Proc. Nat. Acad. Sci. U.S.A. 73
(1976), 1005-1006.
21. , On computing the discrete Fourier transform, I.B.M. Research Report, 1976.
22. , On the number of multiplications necessary to compute certain functions, Comm.
Pure Appl. Math. 23 (1970), 165-179.
23. , Some bilinear forms whose multiplicative complexity depends on the field of
constants, Math. Systems Theory 10 (1977), 169-180.
24. , On the multiplicative complexity of the discrete Fourier transform, Advances in
Math, (to appear).
DEPARTMENT OF MATHEMATICS, GRADUATE SCHOOL AND UNIVERSITY CENTER (CUNY), NEW
YORK, NEW YORK 10036

DEPARTMENT OF MATHEMATICS, UNIVERSITY OF CONNECTICUT, STORRS, CONNECTICUT 06268