A First Course in Number Theory
4.5 Trace Formulae on Finite Abelian Groups
7. Letf ∈L2(G×G). Use Poisson summation to prove that
x∈G
f(x, x) = 1
|G|
χ∈G
(x,y)∈G×G
f(x, y)χ(x)χ(y).
Note that this identity is also an immediate consequence of the or- thogonality relations.
8. This is another example that shows that the lower bound in the uncer- tainty principle (Theorem 4.10) is best possible. LetH be a subgroup ofG, and define δH ∈L2(G) by
δH(x) =
1 ifx∈H 0 ifx∈H. (a) Prove that
supp(δH) =H.
(b) Prove that ifχ∈G, then δ!H(χ) =
* |H| ifχ∈GH 0 ifχ∈GH. (c) Prove that
supp(δH)supp δ!H
=|G|.
4.5 Trace Formulae on Finite Abelian Groups 145 then then×nmatrixA= (aij) = [T]Bis called the matrix of the operator T with respect to the basisB.
LetB={v1, . . . , vn} be another basis forV, and let T(vj) =
n i=1
aijvi. (4.11)
ThenA = (aij) = [T]B is the matrix ofT with respect to the basisB.
Each vectorvj ∈ B is a linear combination of the vectors in the basisB, vj=
n i=1
rijvi, (4.12)
and each vector vj ∈ B is a linear combination of the vectors in the basis B,
vj= n i=1
sijvi. (4.13)
Consider the n×n matrices R = (rij) and S = (sij). Then S = R−1 (Exercise 2). We have
T(vj) = T n
=1
rjv
= n
=1
rjT(v)
= n
=1
rj
n k=1
akvk
= n
=1
rj
n k=1
ak
n i=1
sikvi
= n i=1
n
k=1
n
=1
sikakrj
vi.
Comparing this with (4.11), we obtain aij =
n k=1
n
=1
sikakrj
for alli, j= 1, . . . , n, and so
A =SAR=R−1AR.
Identity (4.10) implies that
tr(A) =tr(R−1AR) =tr(ARR−1) = tr(A).
It follows that we can define the trace of a linear operatorT on a vector spaceV as the trace of the matrix ofT with respect to some basis forV, and that this definition does not depend on the choice of basis.
The vectorv ∈V is called aneigenvectorfor the operatorT witheigen- valueλifv= 0 andT(v) =λv. The operatorT isdiagonalizableif there exists a basis for V consisting of eigenvectors, that is, there exist nonzero vectorsv1, . . . , vn ∈V and numbersλ1, . . . , λn such thatB={v1, . . . , vn} is a basis forV andT(vi) =λivi fori= 1, . . . , n. In this case, the matrix forT with respect to the basisB is the diagonal matrix
D=
λ1 0 0 · · · 0 0 0 λ2 0 · · · 0 0 0 0 λ3 · · · 0 0
... ...
0 0 0 0 0 λn
,
and so
n i=1
aii= tr(A) = tr(D) = n i=1
λi.
We restate this important identity as a theorem.
Theorem 4.12 (Elementary trace formula) Let T be a linear opera- tor on an n-dimensional vector space V, let B be a basis for V, and let A = (aij) be the matrix of T with respect to B. If T is diagonalizable, then V has a basisB={v1, . . . , vn} of eigenvectors with T(vi) =λivi for i= 1, . . . , n, and the trace ofA is equal to the sum of the eigenvalues ofT, that is,
n i=1
aii= n i=1
λi.
We shall show that both the Fourier inversion theorem and the Poisson summation formula are consequences of this elementary trace formula.
Let G be a finite abelian group of order n, and let L2(G) be the n- dimensional vector space of complex-valued functions onG. For everya∈G there is a linear operatorTa onL2(G) defined byTa(f)(x) =f(x−a). The operatorTa is called translation bya.
Another class of operators on L2(G) are integral operators. A function K ∈ L2(G×G) induces a linear operator ΦK on the vector space L2(G) as follows: For f ∈L2(G), let
ΦK(f)(x) =
#
G
K(x, y)f(y)dy=
y∈G
K(x, y)f(y).
The map ΦK is called an integral operatoronL2(G) withkernelK(x, y).
4.5 Trace Formulae on Finite Abelian Groups 147 LetG={x1, . . . , xn}. Associated to the kernelKis a matrixA= (aij)∈ Mn(C) defined by
aij=K(xi, xj). (4.14) Conversely, to every matrixA= (aij)∈Mn(C) there is a functionK(x, y)∈ L2(G×G) defined by (4.14), and an associated integral operator ΦK. Theorem 4.13 Let G={x1, . . . , xn} be an abelian group of order n. Let K ∈L2(G×G) and let ΦK be the associated integral operator on L2(G).
The matrix ofΦK with respect to the orthonormal basis{δxi :i= 1, . . . , n} is(K(xi, xj)), and the trace of ΦK is
tr(ΦK) = n i=1
K(xi, xi). (4.15)
Proof. The matrix of the operator ΦK is (cij), where cij is defined by ΦK(δxj) =
n i=1
cijδxi.
Then
cij = ΦK(δxj)(xi) =
y∈G
K(xi, y)δxj(y) =K(xi, xj).
This completes the proof.2
Theorem 4.14 Let G be a finite abelian group. LetK∈L2(G×G)with ΦK the associated integral operator onL2(G). The operator ΦK commutes with all translationsTa, that is,
TaΦK(f) = ΦKTa(f)
for alla∈Gandf ∈L2(G), if and only if there exists a functionh∈L2(G) such thatK(x, y) =h(x−y)for allx, y∈G. In this case,ΦK is convolution by h, that is,
ΦK(f)(x) =h∗f(x) =
#
G
h(x−y)f(y)dy, and the trace ofΦK is
tr(ΦK) =nh(0).
Proof. Let f, h ∈ L2(G). We define the convolution operator Ch on L2(G) by
Ch(f)(x) =h∗f(x) =
#
G
h(x−y)f(y)dy=
y∈G
h(x−y)f(y).
(See Exercise 10 in Section 4.3.) DefineK(x, y)∈L2(G×G) byK(x, y) = h(x−y). Then
ΦK(f)(x) =
#
G
K(x, y)f(y)dy=
#
G
h(x−y)f(y)dy=Ch(f)(x), and ΦK is convolution byh. Fora, x∈G, we have
TaCh(f)(x) = Ch(f)(x−a)
=
y∈G
h(x−a−y)f(y)
=
y∈G
h(x−y)f(y−a)
=
y∈G
h(x−y)Ta(f)(y)
= ChTa(f)(x),
and so TaCh=ChTa, that is, convolution commutes with translations.
Conversely, let K(x, y)∈L2(G×G). For a, x∈ Gand f ∈L2(G), we have
TaΦK(f)(x) = ΦK(f)(x−a) =
y∈G
K(x−a, y)f(y) and
ΦKTa(f)(x) =
y∈G
K(x, y)Ta(f)(y)
=
y∈G
K(x, y)f(y−a)
=
y∈G
K(x, a+y)f(y).
If ΦK commutes with translations, thenTaΦK = ΦKTa, and
y∈G
K(x−a, y)f(y) =
y∈G
K(x, a+y)f(y).
Applying this identity to the function f(x) =δ0(x) =
1 ifx= 0 0 ifx= 0.
we obtain K(x−a,0) = K(x, a) for all a, x ∈ G. Define the function h∈L2(G) by
h(x) =K(x,0).
4.5 Trace Formulae on Finite Abelian Groups 149 Then
K(x, y) =K(x−y,0) =h(x−y)
for all x, y ∈ G, and the operator ΦK is convolution by h(x). Moreover, tr(ΦK) =nh(0) by (4.15). This completes the proof. 2
Theorem 4.15 (Trace formula) Forh∈L2(G), let Ch be the convolu- tion operator on L2(G), that is, Ch(f) = h∗f for f ∈ L2(G). The dual groupGis a basis of eigenvectors forCh. Ifχ is a character inG, thenχ has eigenvalueh(χ), that is,
Ch(χ) =h(χ)χ, and
nh(0) =
χ∈G
h(χ).
Proof. This is a straightforward calculation. Forx∈G, we have Ch(χ)(x) = h∗χ(x) =χ∗h(x)
=
y∈G
χ(x−y)h(y)
=
y∈G
h(y)χ(y)
χ(x)
= h(χ)χ(x),
and so χis an eigenvector of the convolutionCh with eigenvalueh(χ). By Theorem 4.12, since Gis a basis for L2(G), the trace ofCh is the sum of the eigenvalues, that is,
tr(Ch) =
χ∈G
h(χ).
By Theorem 4.14, we also have
tr(Ch) =nh(0).
This completes the proof.2
We can immediately deduce the Fourier inversion formula (Theorem 4.8) from Theorem 4.15. Iff ∈L2(G), then
f(0) = 1 n
χ∈G
f(χ). (4.16)
This trace formula can also be obtained by computing the Fourier series forf atx= 0. On the other hand, if we simply apply (4.16) to the function T−a(f) and use Exercise 9 in Section 4.3, then we obtain
f(a) = T−a(f)(0)
= 1
n
χ∈G
"
T−a(f)(χ)
= 1
n
χ∈G
f(χ)χ(a).
This is the Fourier inversion formula.
Next, we derive the Poisson summation formula (Theorem 4.11) from the elementary trace formula.
LetHbe a subgroup ofG, and letπ:G→G/Hbe the natural map. For x∈G, definex =π(x) =x+H ∈G/H. There is an orthonormal basis for the vector space L2(G/H) that consists of the functions δx, where
δx(y) =
1 ifx=y 0 ifx=y. Forf ∈L2(G), define the functionf∈L2(G/H) by
f(x+H) =
y∈H
f(x+y).
Let Cf be convolution by f onL2(G/H). The operatorCf has matrix f(x−y)
, with respect to the basis{δx}. By Theorem 4.14, the trace ofCf is
tr(Cf) =|G/H|f(0) = |G|
|H|
y∈H
f(y).
By Theorem 4.15, the character group G/H" is a basis of eigenvectors for the convolution operatorCy. Ifχ∈G/H" andχ=π(χ)∈GH, then
Cf(χ) =f(χ)χ, with eigenvalue
f(χ) =
x∈G/H
f(x)χ(x)
=
x∈G/H
y∈H
f(x+y)χ(x)
=
x∈G/H
y∈H
f(x+y)χ(x+y)
=
x∈G
f(x)χ(x).