Professional Documents
Culture Documents
2110 05959
2110 05959
HANKEL MATRICES
MICHAEL YIASEMIDES
arXiv:2110.05959v1 [math.NT] 12 Oct 2021
Abstract. We prove an exact formula for the variance of the divisor function over short
intervals in Fp [T ], where p is a prime integer. We use exponential sums to translate the
problem to one involving the ranks of Hankel matrices over finite fields. We prove several
results regarding the rank and kernel structure of these matrices, and thus demonstrate
their number-theoretic properties. We briefly discuss extending our method to moments
higher than the second (the variance); to the k-th divisor function; to correlations of the
divisor function with applications to moments of Dirichlet L-functions in function fields;
and to the analogous results over Fq [T ], where q is a prime power.
where the remainder satisfies ∆(x) = O(x 2 ); while, for k ≥ 2, it can be shown that
1
x=0 2
6π n=1
while Tong [18] proved that
∆3 (x)2 dx ∼ ∑ d3 (n)2 n− 3 .
X 1 ∞
X−3 ∫
5 4
x=0 2
10π n=1
We can also consider higher moments of ∆: ∫x=0 ∆(x)k dx. For this, we refer the reader to
X
the works of Heath-Brown [5] and Tsang [19]. We also mention that Heath-Brown proved
1
that x− 4 ∆(x) has distribution function. That is, there is a function f such that
X −1 meas{x ∈ [1, X] ∶ x− 4 ∆(x) ∈ I} Ð→ ∫ f (t)dt
1
t∈I
A related topic is that of the divisor function over intervals. That is, we are interested in
∑ d(n) = ∑ d(n) − ∑ d(n).
x<n≤x+H n≤x+H n≤x
Given that ∆(x) = O(x 2 ), it is clear that, for H ≤ x, the error term
1
It is the short intervals that have H ≤ x1− k that are of particular interest. We highlight some
1
of the main results in this area. Let ǫ > 0, and consider the range X ǫ < H < X 2 −ǫ . Ivić [9]
1
∑ cj log ( ) + Oǫ (X − 2 +ǫ H 2 ) + Oǫ (X ǫ H 2 ),
3
=
1 2X
j X2
∫
1 1
2
∆(x; H) dx H
X x=X j=0 H
where c0 , c1 , c2 are constants and c3 = π82 . Assuming the Riemann hypothesis, for k ≥ 3 and
the range X ǫ < H < X 1−ǫ , Milinovich and Turnage-Butterbaugh [15] obtained the upper
bound
X x=X
Asymptotic formulas can be obtained given certain restrictions on H. For k ≥ 3 (and
assuming the Lindelöf Hypothesis for k > 3) and 2 ≤ L ≤ X k(k−1) −ǫ , Lester [12] proved
1
1 1 1
∆(x; xL2 )
1
∫
− t2
β} e dt.
X 1 3
2π t=α
x 4 π82 logL L
That is, we have a Gaussian distribution function.
Let us now consider the divisor function over short intervals in the polynomial ring A ∶=
Fq [T ], where q is a prime power. Before proceeding, we define M to be the set of monic
polynomials in A; and for B = A, M we define Bn and B≤n to be the set of polynomials
in B with degree equal to n and degree ≤ n, respectively. For non-zero A ∈ A we define
∣A∣ ∶= q deg A , and we define ∣0∣ ∶= 0. It is convenient to take deg 0 = −∞, and so (unless
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 3
otherwise indicated) the range deg A ≤ n should be taken to include the zero polynomial.
The k-th divisor function is defined for N ∈ M by
dk (N) ∶= ∣{(A1 , . . . , Ak ) ∈ Mk ∶ A1 . . . Ak = N}∣.
For A ∈ Mn and 0 ≤ h ≤ n, we define the interval
(3) I(A; h) ∶= {B ∈ M ∶ deg(B − A) < h}
and define
(4) Ndk (A; h) ∶= ∑ dk (B).
B∈I(A;h)
This notation is in keeping with [11]. Although, in [11], they take ≤ h in (3), instead of < h;
and it should be noted that when we reference their results below it will be in terms of our
notation and so it will appear slightly different. We feel that our definition is more natural.
For example, it gives ∣I(A; h)∣ = q h as opposed to ∣I(A; h)∣ = q h+1 ; and, for A ∈ Mn , it gives
∣{A′ ∈ Mn ∶ I(A′ ; h) = I(A; h)}∣ = q n−h
as opposed to
∣{A′ ∈ Mn ∶ I(A′ ; h) = I(A; h)}∣ = q n−h−1 .
Continuing, it is not difficult to obtain an exact expression for the mean value of Ndk (A; h)
(see [1] for a proof):
h n+k−1
(A; = ( ).
1
∑ N
k−1
dk
h) q
q n A∈Mn
We can now define
n+k−1
∆k (A; h) ∶= Ndk (A; h) − q h ( ).
k−1
(5)
where Ik (n; n−h−3) is an integral over a group of unitary matrices, defined by (1.27) in [11].
In particular, when k = 2, and n ≥ 5 and h ≤ n2 − 1, we have
(n − 2h + 7)(n − 2h + 8)(n − 2h + 9)
∑ ∣∆2 (A; h)∣2 ∼ q h
1
(6) n
q A∈Mn 6
as q Ð→ ∞.
1.2. Statement of Results. We obtain an exact formula for the variance of the divisor
function over intervals in Fp [T ], where p is prime.
Theorem 1.2.1. Let A ∈ Fp [T ]. For n ≥ 4 we have
(p − 1)ph−1 (n−2h−1)(n−2h)(n−2h+1) for h ≤ ⌊ n2 ⌋ − 1,
∣∆(A; = {
1
∑ 2 6
for h ≥ ⌊ n2 ⌋.
(7) h)∣
pn A∈Mn 0
4 MICHAEL YIASEMIDES
We must point out that our result is not exactly consistent with (6), in that the polynomial
that appears in our result is a shift of the polynomial that appears in (6). However, we
believe our result is correct as we have tested various non-trivial cases on Mathematica.
To prove Theorem 1.2.1, we use exponential sums to express the problem in terms of Hankel
matrices over Fp . Most of this article comprises of results on Hankel Matrices over finite
fields, which can be found in Section 2. Once these results are established, the proof of
Theorem 1.2.1, in Section 3, is relatively short.
Theorem 2.3.1 gives the number of Hankel matrices of a given size and rank. Theorems 2.4.4
and 2.4.7 demonstrate the kernel structure of Hankel matrices, and we see how function
field arithmetic is incorporated in these matrices. To see this, we must view the coefficients
of a polynomial as the entries in a vector, and vice versa. For example, the polynomial
a0 + a1 T + . . . + an T n should be associated to the vector (a0 , a1 , . . . , an )T . We prove that,
generally, for a Hankel matrix H, there are polynomials A1 , A2 such that the kernel of H
consists exactly of the polynomials
B1 A1 + B2 A2
where B1 , B2 ∈ A are any polynomials satisfying a certain bound on their degrees. The
polynomials A1 , A2 are called the characteristic polynomials of H, not to be confused with
the characteristic polynomial of a square matrix. In Theorem 2.4.8 and Corollary 2.4.9, we
show that if H ′ is a top-left submatrix 1 of H, then the characteristic polynomials associated
to H ′ are the same as the polynomials that we obtain after applying a certain number of
steps of the Euclidean algorithm to A1 , A2 ; the exact number of steps is related to the size of
the submatrix H ′ . Theorem 2.4.11 demonstrates how the kernel structure of a Hankel ma-
trix changes if we extend it by a single row or column. We prove various other results as well.
It should be noted that Theorem 1.2.1 applies to A = Fp [T ] where p is a prime, while our
results on Hankel matrices apply for matrices over Fq where q is a prime power. For a brief
discussion of the generalisation of the former to prime powers, see Subsection 1.4.
1.3. Motivation. We will briefly give an explanation of how exponential sums can allow us
to express sums of the divisor function in Fp [T ] in terms of Hankel matrices. This is best
achieved by considering the sum
∑ d(A)2 .
A∈Mn
1A
top-left submatrix of an l × m matrix M is a submatrix consisting of the first l1 rows and first m1
columns of M , for some l1 ≤ l and m1 ≤ m.
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 5
We have that
2 2
∑ d(A) = ∑ ( ∑
2
∑ 1) = ∑ ( ∑ ∑ 1) .
A∈Mn A∈Mn l,m≥0 E∈Ml A∈A≤n l,m≥0 E∈Ml
l+m=n F ∈Mm l+m=n F ∈Mm
EF =A EF =A
We note that the conditions on E and F force A to be monic and of degree n, and that is
why for the last equality we were able to replace the condition A ∈ Mn with A ∈ A≤n . Now,
let us write ai , ei , fi for the i-th coefficient of A, E, F , respectively. We also write {EF }i for
the i-th coefficient of EF when we do not yet wish to express the coefficients of EF in terms
of the ei and fi . We have
2
⎜ ⎜ ⎟
i+j=k
Let us now consider only the terms involving a given ak in the sum above. Since A ∈ A≤n ,
we can see that ak ranges over Fp . We have
if αk = −βk ,
∑ exp ( − (αk + βk )ak ) = {
1 2πi 1
p ak ∈Fp p 0 if αk ≠ −βk ,
where we have used (8) with b = αk + βk .2 Essentially, by considering all k = 0, 1, . . . , n this
means α = −β, and we have effectively removed the sum over A. Due to certain symmetrical
properties, we can actually take α = β. So, we have
2
Technically, this is not quite true as the last entry of e is 1, and so it cannot take any value
in Fp . Ultimately, this actually limits the α that we can take, and so simplifies our final
calculations. However, for now, let us assume that the last entry of e can take any value in Fp .
Continuing, noting that the number of f in the kernel of Hl+1,m+1 (α) is pm+1−rank Hl+1,m+1 (α) ,
we have
2
∑ d(A) ≈p ∑ ( ∑ p
2 − rank Hl+1,m+1 (α)
).
A∈Mn α∈Fn+1
p l,m≥0
l+m=n
So, we can now see how using exponential sums allows us to express the sum ∑A∈Mn d(A)2
in terms of Hankel matrices in a concise manner, and how knowing the exact number of
Hankel matrices of a given rank and size will allow us to obtain an exact evaluation of the
original divisor sum.
Now, Theorem 1.2.1 is concerned with the variance of the divisor function. That is, it is
concerned with the sum
∣∆ (A; = ∣N (A; (n
1 2 1 h
2
∑ 2 h)∣ ∑ d2 h) − p + 1)∣
pn A∈Mn pn A∈Mn
2
= n ∑ ( − p2h (n + 1)2 ,
1
∑ d(B))
p A∈Mn B∈Mn
deg(B−A)<h
2Strictlyspeaking we should write, for example, αk ≡ −βk (mod p) instead of αk = −βk , but since αk and
βk are in Fp (and it is just in the exponentials that we are just viewing them as integers in {0, 1, . . . , p − 1})
it is simpler to write αk = −βk . We will do this later in the paper as well.
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 7
∑ ( ∑ d(B))
A∈Mn B∈Mn
deg(B−A)<h
which is similar to the sum ∑A∈Mn d(A)2 that we worked with earlier. Here, we have the
sum over A which appears outside the squared parentheses, while the sum over B appears
within. We can proceed similar to previously, and the sum over A will force αk + βk = 0 for
k = h, h+1, . . . , n. Whereas, the sum over B, will force αk = 0 and βk = 0 for k = 0, 1, . . . , h−1.
Ultimately, this means we will have to understand how many Hankel matrices there are of
a given size and rank with the first h skew-diagonals being 0.
Note that if the first h entries of α are zero, and h ≥ n2 , then the matrix H n2 +1, n2 +1 (α) is
lower skew-triangular (for simplicity we are assuming n is even and so n2 is an integer).
We can easily determine the rank of such a matrix by determining the first non-zero skew
diagonal; and, as we will see later, we can use that to easily determine the rank of all
Hl,m(α) for l + m − 2 = n. On the other hand, if h < n2 , then it is more difficult to determine
the rank of the matrix H n2 +1, n2 +1 (α), which is no longer necessarily lower skew-triangular.
This demonstrates why it is easier to work with large intervals (h ≥ n2 ) than short intervals
(h < n2 ). Note that the condition h < n2 is equivalent to ph = (pn ) 2 which is analogous to the
1
1.4. Extensions. Some of the most important kinds of sums of the divisor function are
divisor correlations, particularly those appearing in moments of Dirichlet L-functions. For
simplicity, let us assume that we are working with the fourth moment of Dirichlet L-functions
over characters of prime modulus Q = q0 +q1 T +. . .+qr T r ∈ Fp [T ]. The correlations of interest
are of the form
∑ ∑ d(KQ + N)d(N),
K N
where K ranges over the monics of a certain degree k and N ranges over the monics of a
certain degree n < deg Q. Now, if k ≥ deg Q, then an exact formula for the correlation above
is not difficult to obtain using our approach of exponential sums and Hankel matrices, and
the results we establish in Section 2. However, it is the case when k < deg Q that is of
interest for the fourth moment of the L-functions. In this case we must understand how
many α = (α0 , α1 , . . . , αk+r ) there are that satisfy te following three conditions:
● The square matrix H k+r +1, k+r +1 (α) has rank r1 ;
● The square matrix H n2 +1, n2 +1 (α′ ) has rank r2 ;
2 2
Let us now consider generalisations of Theorem 1.2.1. First, consider the analogue of The-
orem 1.2.1 for the k-th divisor function, dk . The approach is similar but we will ultimately
8 MICHAEL YIASEMIDES
be working with Hankel tensors instead of Hankel matrices. A matrix is a two dimensional
array, which appeared because we were working with the standard divisor function, d = d2 .
When working with higher divisor functions, we will work with higher dimensional arrays
(i.e. tensors). Consider the case k = 3. We will have tensors of the form
(αi+j+l−3 ) 1≤i≤i1
1≤j≤j1
1≤l≤l1
and we will need to determine how many f = (f1 , . . . , fj1 )T and g = (g1 , . . . , gl1 )T there are
such that
j1 l1
∑ ∑ αi+j+l−3 fj gl = 0
j=1 l=1
Now, in Theorem 1.2.1, we consider the variance of the divisor function, which is essentially
the second moment. We could consider higher moments, such as the third, and if we are
aiming to obtain an exact evaluation then it would be sufficient to obtain an exact formula
for
3
∑ ( ∑ d(B)) .
A∈Mn B∈Mn
deg(B−A)<h
With the variance, we needed to determine how many α, β there are such that Hl+1,m+1 (α)
and Hl+1,m+1 (β) have certain given ranks, and α + β = 0. Of course, the last condition
means the matrices are essentially identical, making the problem slightly simpler. However,
for the third moment, we will need to determine how many α, β, γ there are such that
Hl+1,m+1 (α), Hl+1,m+1 (β), and Hl+1,m+1 (γ) have certain given ranks, and α + β + γ = 0. This
last condition makes the problem more difficult. We do indicate in Remark 2.4.14 how we
can reduce this problem to special cases of α, β, γ. We are effectively taking two Hankel
matrices for which we understand their rank, and we wish to understand the rank of their
sum. Related problems have been considered for Toeplitz-plus-Hankel matrices over the
complex numbers [6, 16] (compared to Hankel-plus-Hankel matrices, which is what we are
interested in).
Now suppose we wish to extend Theorem 1.2.1 to Fq [T ], where q is a prime power. This will
require us to work with block Hankel matrices, where each block is also a Hankel matrix.
That is, if we express Fq [T ] = Fp [T ]/QFp [T ] for some irreducible Q = q0 + q1 T + . . . + qr T r ∈
Fp [T ] of degree r, then we will need to work with matrices of the form
⎛A0 A1 A2 Am ⎞
⎜A1 A2 ⎟
⎜ ⎟
⎜A2 ⎟
⎜ ⎟
⎜ ⎟,
⎜ ⎟
⎜ An−2 ⎟
⎜ ⎟
⎜ An−1 ⎟
⎜ An−2 ⎟
⎝ Al An−2 An−1 An ⎠
where each Ai is equal to a Hankel matrix Hr,r (αi ) such that Hr−1,r+1(αi ) has (q0 , q1 , . . . , qr )T
in its kernel. The reason for this is that there is polynomial multiplication taking place on
two levels. Indeed, suppose A ∈ Fq [T ]; then, d(A) counts the number of E, F ∈ Fq [T ]
satisfying EF = N. The multiplication between E and F is accounted for by the entire
matrix above. However, each coefficient of E and F can itself be viewed as a polynomial
over Fp [T ], and so when we multiply these coefficients there is polynomial multiplication
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 9
occurring there as well. This is accounted for by the individual blocks inside the matrix
above. The requirement on the kernel of these blocks ultimately stems from the fact that
the coefficients are in Fq [T ] = Fp [T ]/QFp [T ]. For block Hankel matrices over the complex
numbers we refer the reader to [7].
⎛ α0 α1 α2 αm−1 ⎞
⎜ α1 α2 ⎟
⎜ ⎟
⎜ α2 ⎟
⎜ ⎟
(αi+j−2 ) 1≤i≤l = ⎜
⎜
⎟,
⎟
⎜ αl+m−4 ⎟
(10)
⎜ ⎟
1≤j≤m
⎜ αl+m−3 ⎟
⎜ αl+m−4 ⎟
⎝αl−1 αl+m−4 αl+m−3 αl+m−2 ⎠
where α0 , . . . , αl+m−2 ∈ Fq . As we can see, all the entries on a given skew-diagonal take the
same value. We index our entries from zero, and we later see this is necessary in order to be
consistent with the indexing of coefficients in a polynomial, which also begins at zero.
Define the n × n counter-identity matrix, Jn , to be the matrix with zero entries everywhere
except for the main skew-diagonal going from bottom-left to top-right. If H is an l × m
Hankel matrix, then Jl H and HJm are l × m Toeplitz matrices. Similarly, if T is an l × m
Toeplitz matrix, then Jl T and T Jm are l × m Hankel matrices. Thus, we can see that Han-
kel matrices are inextricably linked to the more well known Toeplitz matrices. Although,
we will focus on the former. We denote the set of all l ×m Hankel matrices in Fq [T ] by Hl,m.
It is natural to consider the finite sequence (α0 , α1 , . . . , αl+m−2 ) ∈ Fql+m−1 that is associated to
the matrix (10). Generally, for α ∶= (α0 , . . . , αn ) ∈ Fn+1
q and l + m − 2 = n with l, m ≥ 1, we
define the matrices
⎛ α0 α1 α2 αm−1 ⎞
⎜ α1 α2 ⎟
⎜ ⎟
⎜ α2 ⎟
⎜ ⎟
Hl,m (α) ∶= ⎜
⎜
⎟.
⎟
⎜ αl+m−4 ⎟
(11)
⎜ ⎟
⎜ αl+m−3 ⎟
⎜ αl+m−4 ⎟
⎝αl−1 αl+m−4 αl+m−3 αl+m−2 ⎠
That is, we associate n + 1 number of matrices with α. As we will later see, there is a crucial
relationship between the kernels of these matrices. We note that we can extend the above
definition to the case where l+m−2 = n′ < n, in which case we have n′ +1 number of matrices,
10 MICHAEL YIASEMIDES
2 ⌋ and n2 ∶= ⌊ 2 ⌋.
Throughout this paper, for an integer n ≥ 0, we will always define n1 ∶= ⌊ n+2 n+3
Now consider the matrix Hn1 ,n2 (α), which is square if n is even and almost square if n is
odd. Note that it has exactly n + 1 skew-diagonals, and so each entry in α appears in the
matrix. We define
π(α) ∶= rank Hn1 ,n2 (α) − ρ(α).
We will later see that ρ(α) and π(α) will help us in understanding the rank and kernel of
not just Hn1 ,n2 (α) but all Hl,m (α) for l + m − 2 = n.
Thus, it is appropriate to extend the definitions of ρ(α) and π(α) to the matrices associated
to α: For l + m − 2 = n, we define ρ(Hl,m (α)) = ρ(α) and π(Hl,m (α)) = π(α). In particular,
these properties are dependent only on the underlying α and not on the shape of the matrix.
Note that this definition does not apply to matrices with l + m − 2 < n; in these cases, we will
need to work with an appropriate truncation α′ of α and work with ρ(α′ ) and π(α′ ) instead.
In fact, these are all attainable if n is odd, but if n is even then π(α) can only attain values in
{0, 1, . . . , max{0, n1 − ρ(α) − 1}}. That is, we cannot have π(α) = n1 − ρ(α) when ρ(α) ≠ n1 .
Indeed, for a contradiction, suppose we do have π(α) = n1 − ρ(α) and ρ(α) ≠ n1 . Then,
rank Hn1 ,n1 (α) = rank Hn1 ,n2 (α) = ρ(α) + π(α) = n1 ,
where the first equality uses the fact that n1 = n2 (since n is even). Thus, Hn1 ,n1 (α) has full
rank and so ρ(α) = n1 by definition, which obviously contradicts ρ(α) ≠ n1 . When n is odd,
we have n1 ≠ n2 and so the first relation above does not hold, meaning we do not encounter
this contradiction.
We now make several definitions for sets of finite sequences and Hankel matrices.
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 11
Remark 2.1.4. Consider Lnh (r, ρ1 , π1 ), and suppose ρ1 ≠ 0. We must have that h ≤ ρ1 − 1.
Otherwise, for any α ∈ Lnh (r, ρ1 , π1 ) we would have that Hρ1 ,ρ1 (α) is strictly lower skew-
triangular3 and thus contradicting that det Hρ1 ,ρ1 (α) ≠ 0.
Now suppose that ρ1 = 0. Then, h ≤ n + 1 − r. This can be seen from the following reasoning.
Since ρ1 = 0, we have
det H1,1 (α) = 0, ..., det Hn1 ,n1 (α) = 0,
and so, by induction (as we later demonstrate), we have that Hn1 ,n1 (α) is strictly lower-skew
triangular. Thus, the matrix Hn1 ,n2 (α) is of the form
⎛0 0 ⎞
⎜ n1 ⎟
⎛0 0 αn 1 ⎞
⎜ ⎟ ⎜0 αn1 αn1 +1 ⎟
0 0 α
⎜0 αn1 αn1 +1 ⎟ ⎜ ⎟
0
⎜ ⎟ ⎜ ⎟
0
⎜ ⎟ ⎜ ⎟
or
⎜ ⎟
⎝0 αn1 αn1 +1 αn ⎠
⎝0 αn1 αn1 +1 αn ⎠
3We say a square matrix is lower/upper skew-triangular if all entries above/below the main skew-diagonal
are zero, and it is strictly so if the entries on the main skew diagonal are also zero.
12 MICHAEL YIASEMIDES
if n is even or odd, respectively. Given that rank Hn1 ,n2 (α) = r, we can see that Hn1 ,n2 (α)
must be of the form
⎛0 0 ⎞
⎜ ⎟
⎜ ⎟
⎜0 0 ⎟
⎜ ⎟
⎜0 αn+1−r ⎟
⎜ 0 ⎟
⎜0 αn+2−r ⎟
⎜ ⎟
⎜ ⎟
0 αn+1−r
⎜ ⎟
⎜ ⎟
⎝0 0 αn+1−r αn+2−r αn ⎠
with αn+1−r ≠ 0. (If n is odd and r = n1 , then there should be no rows of zeros at the top,
and the matrix above should be interpreted as such). In particular, this forces h ≤ n + 1 − r,
as required.
Definition 2.1.5 (Quasi-regular). Suppose we have α = (α0 , α1 , . . . , αn ) ∈ Fn+1
q . We say that
α is quasi-regular if π(α) = 0. For any integers l, m ≥ 1 with l + m − 2 = n we say Hl,m (α)
is quasi-regular if α is quasi-regular. Quasi-regularity is defined in Definition 5.8 of [8].
Before proceeding, we make a couple of remarks on notation. Let M be an l × m matrix and
let l1 , l2 , m1 , m2 satisfy l1 + l2 ≤ l and m1 + m2 ≤ m. Then, we define M[l1 , −l2 ; m1 , −m2 ] to be
the submatrix of M consisting of the first l1 and last l2 rows, and the first m1 and last m2
columns. In the special cases when one or more of l1 , l2 , m1 , m2 are zero, we may not include
them. For example, M[l1 ; −m2 ] should be taken to be M[l1 , 0; 0, −m2 ].
(See Remark 2.1.2 regarding the values that π1 can take). Then, π(α) = π1 if and only if
⎧
⎪ {0} for i < (n + 1) − π1
⎪
⎪
⎪ ∗
αi ∈ ⎨ F q for i = (n + 1) − π1
⎪
⎪
⎪
⎪
⎩ Fq for i > (n + 1) − π1 .
Proof. We begin with the forward implication. Let H ∶= Hn1 ,n2 (α). Since ρ(α) = 0, we have
det H[i; i] = 0 for i = 1, . . . n1 . When i = 1 this gives
0 = det H[1; 1] = α0 .
When i = 2 it gives
⎛0 0 ⎞
⎜0 αn 1 ⎟ ⎛0 0 αn 1 ⎞
⎜ ⎟ ⎜ n1 +1 ⎟
0
⎜ n1 +1 ⎟ H =⎜ ⎟,
0 0 α α
H =⎜ ⎟ ⎜ ⎟
n1
0 0 α α
⎜ ⎟ ⎜ ⎟
n1 or
⎜ ⎟
⎝0 αn1 αn1 +1 αn ⎠
⎝0 αn1 αn1 +1 αn ⎠
respectively. Now, suppose i is the largest element in the set {1, 2, . . . , (n+1)−n1 } satisfying
α(n+1)−i ≠ 0 (such an i must exist, unless we are working with the zero matrix, in which case
we take i = 0). Then,
⎛0 0 ⎞
⎜ ⎟
⎜0 0 ⎟
⎜ ⎟
H =⎜
⎜0
⎟.
α(n+1)−i ⎟
⎜ 0 ⎟
⎜ ⎟
⎜ ⎟
⎝0 0 α(n+1)−i αn ⎠
Recalling that α(n+1)−i ≠ 0, we can clearly see that rank H = i. Since rank H = ρ(α) + π(α) =
π1 , we see that i = π1 . In particular, α(n+1)−π1 ≠ 0; while for i < (n + 1) − π1 we have αi = 0,
and for i > (n + 1) − π1 we have αi ∈ Fq . This concludes the forward implication.
The backward implication follows easily given some of the reasoning that we have established
above.
Remark 2.2.2. Let l + m − 2 = n. It is helpful to visualise what the matrix Hl,m (α) looks
like given ρ(α) = 0 and π(α) = π1 . For presentational purposes, we use 1 to denote an entry
in F∗q , we use ∗ to denote an entry in Fq , and 0 denotes 0 as usual. For l < π1 , we have
⎛0 0 1 ∗ ∗⎞
⎜0 ∗⎟
Hl,m (α) = ⎜ ⎟.
0 1 ∗
⎜ ⎟
⎜ ⎟
⎝0 0 1 ∗ ∗⎠
14 MICHAEL YIASEMIDES
This has full row rank. We can describe the kernel in some manner, although it is more
helpful when l, m ≥ π1 . In this case, we have
⎛0 0⎞
⎜ ⎟
⎜ ⎟
⎜0 0⎟
⎜ ⎟
⎜ 1⎟
Hl,m (α) = ⎜0 0 ⎟,
⎜0 ∗⎟
⎜ ⎟
⎜ ⎟
0 1
⎜ ⎟
⎜ ⎟
⎝0 0 1 ∗ ∗⎠
where there are exactly π1 number of 1s. This has rank equal to π1 , we can clearly see that
any vector in the kernel must have zeros in its last π1 positions. For m < π1 , we have
⎛0 0 0⎞
⎜ ⎟
⎜ 0⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ 1⎟
⎜ ⎟
⎜ ⎟
⎜ ∗⎟
Hl,m (α) = ⎜ 0 ⎟.
⎜ ⎟
⎜0 ⎟
⎜ 1 ⎟
⎜1 ⎟
⎜ ⎟
⎜∗ ⎟
∗
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎝∗ ∗ ∗⎠
This has full column rank and so the kernel is trivial.
Lemma 2.2.3. Suppose α = (α0 , . . . , αn ) with ρ(α) = ρ1 ∈ {1, 2, . . . , n1 − 1} and
{0, 1, . . . , n1 − ρ1 }
π(α) = π1 ∈ {
if n is odd,
{0, 1, . . . , n1 − ρ1 − 1} if n is even
(See Remark 2.1.2 regarding the values that π1 can take). Suppose l + m − 2 = n with l > ρ1 ,
and let H ∶= Hl,m (α). Define x = (x0 , . . . , xρ1 −1 )T to be the vector that satisfies
⎛ αρ1 ⎞
⎜α ⎟
H[ρ1 , ρ1 ]x = ⎜ ρ1 +1 ⎟ .
⎜ ⋮ ⎟
⎝α2ρ1 −1 ⎠
Let Ri be the i-th row of Hl,m (α). If we apply the row operations
⎛Ri−ρ1 ⎞
Ri Ð→ Ri − (x0 , . . . , xρ1 −1 ) ⎜ ⋮ ⎟ = Ri − x0 Ri−ρ1 − . . . − xρ1 −1 Ri−1
⎝ Ri−1 ⎠
for i = n1 , n1 − 1, . . . , ρ1 + 1 in that order, then we obtain a matrix
(α′ )
( ρ1 ,m ),
H
Hl−ρ1 ,m (β)
(12)
where
α′ =(α0 , . . . , αρ1 +m−2 )
β =(βρ1 , . . . , βn )
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 15
and
⎧
⎪ {0} if i < (n + 1) − π1
⎪
⎪
⎪ ∗
βi ∈ ⎨Fq if i = (n + 1) − π1
⎪
(13)
⎪
⎪
⎪
⎩ Fq if i > (n + 1) − π1 .
Furthermore, the sequence β is independent of the specific values taken by l, m (as long as
l > ρ1 ).
Remark 2.2.4. In order to keep the lemma above succinct, we avoided various explanatory
remarks. We give them here, for clarity.
The matrix H[ρ1 , ρ1 ] is the largest top-left submatrix of H that is invertible, by definition of
ρ1 . It is independent of the specific values taken by l, m. The vector (αρ1 , αρ1 +1 , . . . , α2ρ1 −1 )T
is simply the column of entries directly to the right of H[ρ1 , ρ1 ] in H, which is also equal to
the transpose of the row of entries directly below H[ρ1 , ρ1 ].
The rows operations that we apply start at the last row and end at the row just below the
submatrix H[ρ1 , ρ1 ].
The lemma states that after the row operations are applied, we are left with the matrix
⎛ α0 α1 αm−1 ⎞
⎜ α1 ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ αm+ρ1 −3 ⎟
⎜ ⎟
Hρ1 ,m (α ) ⎜α αm+ρ1 −2 ⎟
⎜ ρ −1 ⎟
( )=⎜ 1 ⎟
′ αm+ρ1 −3
Hn1 −ρ1 ,m (β) ⎜ βρ1 βρ1 +1 βm+ρ1 −1 ⎟
(14)
⎜ ⎟
⎜ βρ1 +1 ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ βn−1 ⎟
⎜ ⎟
⎝ βl−1 βn−1 βn ⎠
Of course, the lemma states that some of the βi are zero, but we do not demonstrate this above
so that we can instead see that the indexing is preserved: An entry in the i-th skew-diagonal
is always either αi or βi .
Hρ1 ,m (α′ )
H ′ ∶= ( )
Hl−ρ1 ,m (β)
looks like. Again, for presentational purposes, we use 1 to denote an entry in F∗q , we use ∗
to denote an entry in Fq , and 0 denotes 0 as usual. If l ≤ ρ1 , then the lemma above does not
apply. We simply remark that in this case H has full row rank, which follows from the fact
that H[ρ1 , ρ1 ] is invertible and thus has full rank.
16 MICHAEL YIASEMIDES
Now, we wish to consider the case ρ1 < l < ρ1 + π1 (which requires π1 ≥ 2). We can do this by
repeatedly removing a row and adding a column to (15), while maintaining that the bottom
two matrices form a Hankel matrix and that the top two matrices form a Hankel matrix.
This is permissible because, as stated in Lemma 2.2.3, the sequence β is independent of the
values of l and m. We can then see that when ρ1 < l < ρ1 + π1 we have a matrix of the form
⎛H[ρ1 , ρ1 ] H[ρ1 , −(m − ρ1 )] ⎞
⎜ ⎟
⎜ ⎟
0 0 1 ∗ ∗
H =⎜
′
⎜ 0 0 1 ∗ ∗ ⎟
⎟
⎜ ⎟
(16)
⎜ ⎟
0
⎝ 0 0 1 ∗ ∗ ⎠
Note that the 1s still appear to the right of H[ρ1 , ρ1 ]. Given this fact, and the invertibility
of H[ρ1 , ρ1 ], we see that H ′ (and H) has full row rank.
For the case m < ρ1 + π1 , we can again take (15), but this time we repeatedly remove a
column and add a row. Similar to above, we can see that we will have full column rank. In
particular, the kernel will be trivial.
Remarks 2.2.2 and 2.2.5 effectively establish the following Corollary.
Corollary 2.2.6. Let α ∈ Ln (r), and let l + m − 2 = n. Then,
if min{l, m} ≥ r,
rank Hl,m (α) = {
r
min{l, m} if min{l, m} < r.
Of course, if we are working with particular values of ρ(α) and π(α), then we can replace
r with ρ(α) + π(α).
We give a further, final remark in order to demonstrate the usefulness of the (ρ, π)-form.
Remark 2.2.7. Lemma 2.2.1 allows us to easily determine the number of α ∈ Fn+1 q with
ρ(α) = 0 and π(α) = π1 . Regarding Lemma 2.2.3, we can easily count the number of possible
values that the matrix Hl−ρ1 ,m (β) could take. It is not immediately obvious what the number
of values the matrix Hρ1 ,m (α′ ) could take, but we are working with a smaller matrix now,
and so this suggests using an inductive argument, which is what we do in Subsection 2.3.
We now proceed to prove Lemma 2.2.3.
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 17
Proof of Lemma 2.2.3. Given that row operations are only applied to rows ρ1 + 1 to n1 , it is
clear that we do indeed have Hρ1 ,m (α′ ) in the top submatrix of (12).
If a row operation is applied to an entry αi that is found on the i-th skew diagonal, then
it is mapped to αi − x0 αi−ρ1 − . . . − xρ1 −1 αi−1 . This is independent of what position on the
i-th skew diagonal the entry is found (but, of course, it must be on a row that has a row
operation applied to it). It is also independent of what the particular values of l, m are (as
long as l > ρ1 , as given in the lemma). The former demonstrates that the bottom submatrix
of (12) is indeed a Hankel matrix; while the latter demonstrates that β is independent of
the specific values taken by l, m.
Thus, all that remains to be proven is (13); and, since β is independent of the specific values
taken by l, m, it suffices to work with the case l = n1 and m = n2 . To this end, after the row
operations are applied, we have the matrix
⎛ α0 α1 ⋯ ⋯ αρ1 −1 αρ1 αρ1 +1 ⋯ ⋯ αn2 −1 ⎞
⎜ α1 ⎟
⎜ ⎟
⋮ αρ1 +1 ⋮
⎜ ⋮ ⎟
⎜ ⎟
⎜ ⋮ αn2 +ρ1 −1 ⎟
⋮ ⋮ ⋮
⎜ ⎟
⎜α ⋯ αn2 +ρ1 −1 αn2 +ρ1 −2 ⎟
α2ρ1 −3 ⋮
⎜ ⎟
H ′ = ⎜ ρ1 −1 ⎟.
⋯ ⋯ α2ρ1 −3 α2ρ1 −2 α2ρ1 −1 ⋯
⎜ βρ1 βρ1 +1 βn2 +ρ1 −1 ⎟
⎜ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯ ⎟
⎜ βρ1 +1 ⎟
⎜ ⎟
⎜ ⋮ ⎟
⋮
⎜ ⎟
⎜ ⋮ βn−1 ⎟
⋮
⎜ ⎟
⎝βn1 −1 ⋯ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯ βn−1 βn ⎠
This is the similar to (14), but here we indicate the top-left ρ1 × ρ1 submatrix, which is the
largest invertible top-left submatrix of H ′ (by definition of ρ1 ).
We recall that the row operations that we applied do not change the rank of H or any of its
top-left submatrices. We now consider the following top-left submatrices of H ′ :
H ′ [ρ1 + 1∣ρ1 + 1], H ′ [ρ1 + 2∣ρ1 + 2], . . . , H ′ [n1 ∣n1 ].
For the first, we have
⎛ α0 α1 ⋯ ⋯ αρ1 −1 αρ1 ⎞
⎜ α1 αρ1 +1 ⎟
⎜ ⋮ ⎟
⎜ ⋮ ⎟
H ′ [ρ1 + 1∣ρ1 + 1] = ⎜
⎜ ⋮
⎟
⋮ ⎟
⋮ ⋮
⎜ ⎟
⎜α α2ρ1 −1 ⎟
α2ρ1 −3
⎜ ρ1 −1 ⋯ ⋯ α2ρ1 −3 α2ρ1 −2 ⎟
⎝ βρ1 ⋯ ⋯ β2ρ1 −2 β2ρ1 −1 β2ρ1 ⎠
⎛ α0 α1 ⋯ ⋯ αρ1 −1 αρ1 ⎞
⎜ α1 αρ1 +1 ⎟
⎜ ⋮ ⎟
⎜ ⋮ ⋮ ⎟
=⎜
⎜ ⋮
⎟,
⋮ ⎟
⋮
⎜ ⎟
⎜α α2ρ1 −1 ⎟
α 2ρ1 −3
⎜ ρ1 −1 ⋯ ⋯ α2ρ1 −3 α2ρ1 −2 ⎟
⎝ 0 ⋯ ⋯ ⋯ 0 β2ρ1 ⎠
where the second equality follows from the definition of x and the row operation that we
applied to row ρ1 +1. Now, by the definition of ρ1 , we have that det H ′ [ρ1 +1∣ρ1 +1] = 0; while,
by the form of H ′ [ρ1 +1∣ρ1 +1] above, we can see that det H ′ [ρ1 +1∣ρ1 +1] = β2ρ1 ⋅det H[ρ1 ∣ρ1 ].
18 MICHAEL YIASEMIDES
Given that det H[ρ1 ∣ρ1 ] ≠ 0 (by definition of ρ1 ), we must have that β2ρ1 = 0.
0 = det H ′ [ρ1 + 2∣ρ1 + 2] = det ( ) ⋅ det H[ρ1 ∣ρ1 ] = −β2ρ1 +1 2 ⋅ det H[ρ1 ∣ρ1 ].
0 β2ρ1 +1
β2ρ1 +1 β2ρ1 +2
Given that det H[ρ1 ∣ρ1 ] ≠ 0, we must have that β2ρ1 +1 = 0.
Proceeding as above in an inductive manner, we see that βρ1 , βρ1 +1 , . . . , βn1 +ρ1 −1 = 0. That is,
H ′ [n1 , n1 ]
⎛ α0 α1 ⋯ ⋯ αρ1 −1 αρ1 αρ1 +1 ⋯ ⋯ αn1 −1 ⎞
⎜ α1 ⎟
⎜ ⎟
⋮ αρ1 +1 ⋮
⎜ ⋮ ⎟
⎜ ⎟
⎜ ⋮ αn1 +ρ1 −1 ⎟
⋮ ⋮ ⋮
⎜ ⎟
⎜α αn1 +ρ1 −1 αn1 +ρ1 −2 ⎟
α2ρ1 −3 ⋮
⎜ ⎟
= ⎜ ρ1 −1 ⎟.
(17) ⋯ ⋯ α2ρ1 −3 α2ρ1 −2 α2ρ1 −1 ⋯ ⋯
⎜ 0 0 ⎟
⎜ ⋯ ⋯ ⋯ 0 0 ⋯ ⋯ ⋯ ⎟
⎜ 0 βn1 +ρ1 ⎟
⎜ 0 0 0 ⎟
⎜ 0 βn1 +ρ1 βn1 +ρ1 +1 ⎟
⋯ ⋯ ⋯ ⋯ ⋯
⎜ ⎟
⎜ ⋮ ⎟
⋯ ⋯ ⋯ 0 0 ⋯ 0
⎜ ⋮ ⋮ ⋰ ⋰ ⋰ ⋮ ⎟
⎝ 0 ⋯ ⋯ ⋯ 0 0 βn1 +ρ1 βn1 +ρ1 +1 ⋯ β2n1 −2 ⎠
Now, if n is even, then n2 = n1 and H ′ = H ′ [n1 , n1 ], whereas if n is odd, then n2 = n1 + 1 and
we have an additional column:
(18)
H′
⎛ α0 α1 ⋯ ⋯ αρ1 −1 αρ1 αρ1 +1 ⋯ ⋯ αn1 −1 αn 1 ⎞
⎜ α1 ⎟
⎜ ⎟
⋮ αρ1 +1 ⋮
⎜ ⋮ ⎟
⎜ ⎟
⎜ ⋮ αn1 +ρ1 −2 ⎟
⋮ ⋮ ⋮
⎜ ⎟
⎜α αn1 +ρ1 −1 ⎟
α2ρ1 −3 ⋮ αn1 +ρ1 −1
⎜ ⎟
= ⎜ ρ1 −1 ⎟.
⋯ ⋯ α2ρ1 −3 α2ρ1 −2 α2ρ1 −1 ⋯ ⋯ αn1 +ρ1 −1 αn1 +ρ1 −2
⎜ 0 βn1 +ρ1 ⎟
⎜ ⋯ ⋯ ⋯ 0 0 ⋯ ⋯ ⋯ 0 ⎟
⎜ 0 βn1 +ρ1 +1 ⎟
⎜ 0 0 0 βn1 +ρ1 ⎟
⎜ 0 βn1 +ρ1 +2 ⎟
⋯ ⋯ ⋯ ⋯ ⋯
⎜ ⎟
⎜ ⋮ ⎟
⋯ ⋯ ⋯ 0 0 ⋯ 0 βn1 +ρ1 βn1 +ρ1 +1
⎜ ⋮ ⋮ ⋰ ⋰ ⋰ ⋮ ⋮ ⎟
⎝ 0 ⋯ ⋯ ⋯ 0 0 βn1 +ρ1 βn1 +ρ1 +1 ⋯ β2n1 −2 βn1 +n2 −2 ⎠
In either case, in the last n1 − ρ1 rows, all the entries are zero except for the last n2 − ρ1 − 1
skew diagonals which may or may not be zero. We now consider the first such skew-diagonal
that is non-zero: Suppose i is the largest element in the set {1, 2, . . . , (n + 1) − (n1 + ρ1 )} that
satisfies β(n+1)−i ≠ 0 (if no such i exists, then we take i = 0, and only a slight adaptation of
the following reasoning is required). Then, the bottom-right quadrant of H ′ (bounded by
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 19
the vertical and horizontal lines in (17) and (18)) is of the form
⎛0 0⎞
⎜ ⎟
⎜0 0 ⎟
⎜ ⎟
⎜ ⎟.
⎜0 0 β(n+1)−i ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎝0 0 β(n+1)−i ... βn ⎠
Given that β(n+1)−i ≠ 0, we can see that the last i columns appearing in this quadrant are
linearly independent, and that the rank of this quadrant (matrix) is i. Recall also that
H ′ [ρ1 , ρ1 ] has full column rank. So, given the form of H ′ , we can see that the first ρ1
columns and the last i columns of H ′ form a basis for its column space. In particular,
rank H ′ = ρ1 + i.
Since rank H ′ = rank H = ρ1 + π1 , we have i = π1 . This proves (13) as required.
Lemmas 2.2.1 and 2.2.3 extend upon [4, Section 5], where they prove similar lemmas but
for square Hankel matrices only. They use column operations instead of row operations. We
chose the latter in order to preserve the kernel. We will now formally give the definition of
(ρ, π)-form.
Definition 2.2.8 (The (ρ, π)-form). Let α ∈ Fn+1 and let l + m − 2 = n. Consider H ∶=
Hl,m(α). If ρ(α) = 0, then we define the (ρ, π)-form of H to be itself.
q
Claim 1: Suppose ρ1 = 0. By Remarks 2.1.2 and 2.1.4, in order for Lnh (r, ρ1 , π1 ) =
Lnh (r, 0, r) to be non-empty, we require that
0 ≤ r ≤ n1 − 1 if n is even,
0 ≤ r ≤ n1 if n is odd;
20 MICHAEL YIASEMIDES
and
r ≤ n − h + 1.
Assuming these conditions are satisfied, we have
if r = 0,
∣Lnh (r, 0, r)∣ = {
1
(q − 1)q r−1 if r > 0.
Claim 2: Suppose that ρ1 ∈ {1, 2, . . . , n1 − 1}. By Remarks 2.1.2 and 2.1.4, in order for
Lnh (r, ρ1 , π1 ) = Lnh (ρ1 + π1 , ρ1 , π1 ) to be non-empty, we require that
0 ≤ π1 ≤ n1 − ρ1 − 1 if n is even,
0 ≤ π1 ≤ n1 − ρ1 if n is odd;
and
ρ1 ≥ h + 1.
Assuming these conditions are satisfied, we have
(q − 1)q 2ρ1 −h−1 if π1 = 0,
∣Lnh (ρ1 + π1 , ρ1 , π1 )∣ = {
(q − 1)2 q 2ρ1 +π1 −h−2 if π1 > 0.
Proof. Claim 1: When r = 0, the only element in Lnh (r, 0, r) is a sequence of zeros. Thus,
∣Lnh (r, 0, r)∣ = 1. Suppose r > 0. Lemma 2.2.1 tells us α ∈ Lnh (r, 0, r) if and only if
α = (0, . . . , 0, α(n+1)−r , . . . , αn )
with α(n+1)−r ∈ F∗q and α(n+1)−r+1 , . . . , αn ∈ Fq . Thus, we have ∣Lnh (r, 0, r)∣ = (q − 1)q r−1 .
∣L2ρ1 −2 (ρ1 )∣ ⋅ q
(19)
if π1 = 0.
h
Therefore, what we must understand are the sets L2k−2 (k). We have that
j=h+1
(22)
=(q − 1)q i−1 + ∑ ∣L2k−2 (i, j, i − j)∣.
i
h
j=h+1
i=h+1 j=h+1
j=h+1 i=j
j=h+1
where the last lines applies (19). This is a recurrence relation. The initial condition is
∣L2(h+1)−2
h
(h + 1)∣ = (q − 1)q h ;
Indeed, if α ∈ L2(h+1)−2
h
(h+1) then Hh+1,h+1 (α) has all entries above the main skew-diagonal
equal to 0, and so to have rank equal to h + 1 (i.e. full rank) we must have that the entries
in the main skew-diagonal are in F∗q , while the entries in the last h skew-diagonals are free
to take any values in Fq . Now, it can easily be verified that the solution to the recurrence
relation is
∣L2k−2
h
(k)∣ = (q − 1)q 2k−h−2 .
Substituting this into (19) proves Case 3.
Claims 3 and 4: We begin with Claim 4. If r = 0, the only element in Lnh (r) is the
sequence of zeros, thus proving this case.
Suppose instead that 1 ≤ r ≤ min{h, n − h + 1}. As described in the theorem, if α ∈ Lnh (r)
then ρ(α) = 0, and so this case follows from Claim 1.
Now suppose that h+1 ≤ r ≤ n1 −1. If α ∈ Lnh (r) then ρ(α) = 0 or ρ(α) ∈ {h+1, h+2, . . . , r}.
Hence, we have
ρ1 =h+1
ρ1 =h+1
=(q − 1)q
2 2r−h−2
,
where the second equality uses Claim 2.
n1 −1
=q n−h+1 − 1 − ∑(q − 1)q r−1 − ∑ (q 2 − 1)q 2r−h−2
h
r=1 r=h+1
=q n−h+1
−q 2n1 −h−2
.
For Claim 3, if n is even, then we have Lnh (n1 , n1 , 0) = Lnh (n1 ). So, by the last case of
Claim 4, we have
∣Lnh (n1 , n1 , 0)∣ = q n−h+1 − q 2n1 −h−2 = (q − 1)q n−h .
Now suppose n is odd, and let
α =(α0 , α1 , . . . , αn ) ∈ Lnh (n1 , n1 , 0),
α′ ∶=(α0 , α1 , . . . , αn−1 ),
and H ∶= Hn1 ,n2 (α). Since ρ(α) = n1 , we have that Hn1 ,n1 (α′ ) = H[n1 ; n1 ] has full rank.
Therefore, α′ ∈ Ln−1
h
(n1 , n1 , 0) = Ln−1
h
(n1 ), of which there are (q − 1)q n−h−1 possible values
it could take (by the first case of Claim 3). Meanwhile, αn is free to take any value in Fq , of
which there are q possibilities. Thus, ∣Lnh (n1 , n1 , 0)∣ = (q − 1)q n−h , as required.
We remark that the difference between the odd case and the even case is that when n is
odd, the matrix Hn1 ,n2 (α) is not quite square; the additional column allows the matrix to
have full rank without necessarily having ρ(α) = n1 . This is why Claim 3 gives the same
result as the last case of Claim 4 only when n is even, but not when n is odd.
Claim 5: If r < min{l, m}, then Corollary 2.2.6 implies that there is a bijection between
h
Hl,m (r) and Lnh (r). The result then follows by the first three cases of Claim 4.
=q l+m−h−1
−q min{l,m}−1
,
where the third equality uses the first part of Claim 5. Now suppose that min{l, m}−1 ≥ h+1.
Note that this gives h ≤ min{l, m} − 2 ≤ n1 − 2, and so h < n − h + 1. Thus, we have
∣Hl,m
h
(r)∣ =∣Hl,m
h
( min{l, m})∣
min{l,m}−1
=∣Hl,m
h
∣− ∑ ∣Hl,m
h
(i)∣
i=0
min{l,m}−1
− 1 − (q − 1) ∑ q − (q − 1)
h
=q l+m−h−1 i−1 2
∑ q 2i−h−2
i=1 i=h+1
=q l+m−h−1
−q 2 min{l,m}−h−2
.
Again, the third equality uses the first part of Claim 5.
2.4. Kernel Structure. We will now investigate the kernel structure of Hankel matrices.
We begin with an extension to Corollary 2.2.6.
24 MICHAEL YIASEMIDES
Now, in what follows, it will be necessary to view vectors in Fk+1 (for any integer k ≥ 0)
as polynomials in A ∶= Fq [T ]. A vector (v0 , v1 , . . . , vk ) should be considered the same as
q
T
the polynomial v0 + v1 T + . . . vk T k and vice versa. Clearly, a vector has a unique polynomial
associated with it. However, a polynomial does not have a unique vector associated with
it. For example, v1 ∶= (v0 , v1 , . . . , vk )T and v2 ∶= (v0 , v1 , . . . , vk , 0)T are different vectors but
they are associated with the same polynomial. In order to avoid confusion and to ensure
everything is well defined, we will make it clear what vector space we are working with, and
its dimension will inform us of the number of zeros that should appear at the end of the
vector. It should be noted that in [8] the polynomial associated with v2 is said to have a
root at infinity (associated with the 0 in the last entry of the vector), thus distinguishing
it from the polynomial associated with v1 . However, we will not employ this. Finally, it
is helpful to keep in mind that a polynomial of degree k has k + 1 coefficients, and so any
vector associated to it must be in at least (k + 1)-dimensional space.
In this subsection we prove the following four theorems and their associated corollaries. The
proofs of the theorems are provided at the end of this subsection.
Theorem 2.4.4. Let α ∈ Ln (r, ρ1 , π1 ), where n > 0. Denote the characteristic degrees by
c1 ∶= r and c2 ∶= n − r + 2. In what follows, m is given and l should be taken such that
l + m − 2 = n.
⎪
q
⎪
⎪
⎪
⎪ {B1 A1 + B2 A2 ∶ deg B1 ≤m−c1 −1} ⊆ Fm if c2 + 1 ≤ m ≤ n + 1.
B1 ,B2 ∈A
⎪
⎪
⎩
q
deg B2 ≤m−c2 −1
of the characteristic polynomials of α, and thus it is natural and easier to have two charac-
teristic polynomials for both sequences.
Now, suppose c1 ≠ c2 . We can see that A1 is unique up to multiplication by a unit in F∗q , but
unless otherwise stated A1 should be taken to be monic. For r ≥ 2, the polynomial A2 should
be taken to be monic unless otherwise stated. However, even then it is not unique: We can
multiply it by a unit in F∗q and add B2 A2 to it, for any deg B2 ≤ c2 −c1 . Thus, if we state that
A2 is the second characteristic polynomial, it is with the understanding that it is generally
not unique. Note that all possibilities for A2 are equivalent modulo A1 ; and in particular if
ρ1 is equal to r = c1 (that is, the sequence α is quasi-regular), then we can choose A2 to be
monic and have degree less than deg A1 .
The case when c1 = c2 occurs when n is even and c1 = c2 = n1 . In this case we have
ker Hn1 ,n1 (α) ={0},
ker Hn1 −1,n1 +1 (α) ={BA + B ′ A′ ∶ B, B ′ ∈ Fq },
for some A, A′ ∈ A with at least one having degree equal to n1 . As both these polynomials first
appear in the same matrix, it is not immediately obvious how to define the first characteristic
polynomial and how to define the second. However, this can be addressed in the following
manner. We let A2 be the polynomial that is of smaller degree between A, A′ , and multiplied
by an element in F∗q so that it is monic; and we let A1 be the polynomial of higher degree,
multiplied so that it is monic. If both A, A′ have the same degree, then we take A2 to be the
smallest monic representative of A modulo A′ ; and we take A1 to be A′ , again multiplied
so that it is monic. There is more than one possibility for the specific values that A, A′
can take, with all possibilities spanning ker Hn1 −1,n1 +1 (α) as above; but regardless of which
possibility we have, the value of A2 is the same. In cases where we do not have c1 = c2 = n1 ,
this uniqueness would apply to A1 , not A2 ; however, the definition we have here is consistent
with the degree bounds on A1 , A2 given in the theorem for the other cases.
It should be noted that the characteristic degrees and characteristic polynomials of a se-
quence α completely determine the kernel structure. However, the characteristic polynomi-
als alone do not, as we will see in Theorem 2.4.11 where a sequence α = (α0 , α1 , . . . , αn ) and
a certain extension α′ = (α0 , α1 , . . . , αn , αn+1 ) can have the same characteristic polynomials
(but different characteristic degrees).
If r < m ≤ n+2−r and α is not quasi-regular (that is, π1 ≠ 0), then there are no vectors in the
kernel of H of the form (v0 , v1 , . . . , vm−1 , 1)T , for some v0 , . . . , vm−1 ∈ Fq ; that is, none of the
polynomials in the kernel are monic and of degree m. Whereas, if α is quasi-regular (that
is, π1 = 0), then exactly 1q of the vectors in the kernel of H of the form (v0 , v1 , . . . , vm−1 , 1)T ,
for some v0 , . . . , vm−1 ∈ Fq ; that is, 1q of the polynomials in the kernel are monic and of degree
m.
Theorem 2.4.7. Claim 1: Suppose we have A1 ∈ A/{0} with ρ1 ∶= deg A1 ≤ 1, and let
n ≥ ρ1 . Then, there exists a sequence α ∈ Ln (ρ1 , ρ1 , 0) with first characteristic polynomial
equal to A1 . If ρ1 = 0, then there also exists a sequence α ∈ Ln (0, 0, 1) with first character-
istic polynomial equal to A1 . The second characteristic polynomials will be as in (23).
Claim 3: Suppose we have coprime A1 , A2 ∈ A with r ∶= deg A1 ≥ 2, and let n ≥ max{r, deg A2 }
+ r − 2. Then, there exists a sequence α ∈ Ln (r, r, 0) with characteristic polynomials A1 , A2 .
Furthermore, α is unique up to multiplication by elements in F∗q .
Claim 4: Suppose we have coprime A1 , A2 ∈ A with deg A2 > deg A1 ≥ 2, and let
ρ1 ∶= deg A1
π1 ∶= deg A2 − deg A1
r ∶=ρ1 + π1
n ∶= deg A2 + r − 2.
Then, there exists a sequence α ∈ Ln (r, ρ1 , π1 ) with characteristic polynomials A1 , A2 . Fur-
thermore, α is unique up to multiplication by elements in F∗q .
This theorem demonstrates the extent to which we can take any coprime polynomials A1 , A2
and integer n such that there is a sequence α = (α0 , α1 , . . . , αn ) with characteristic polyno-
mials equal to A1 , A2 . Claims 1 and 2 address the cases where ρ1 ≤ 1. This is not difficult
and it is included for completeness. Claim 3 considers the case where α is quasi-regular, and
we can see by Theorem 2.4.4 that this allows for the possibility that deg A2 ≤ deg A2 . On the
other hand, if α is not quasi-regular, then by Theorem 2.4.4 we must have deg A2 > deg A1 ,
and this is the case that Claim 4 considers.
With regards to the definition of r in Claim 2, this follows from the fact that if A2 is to
be the second characteristic polynomial, then we need the second characteristic degree of
α, which is n + 2 − r, to be equal to deg A2 . With regards to the bounds on n, for Claim 3
we note that the characteristic degrees are r and n + 2 − r, and since the latter must be at
least as large as the former, we obtain the requirement that n ≥ r + r − 2. Furthermore, by
Theorem 2.4.4, we must have that deg A2 ≤ n + 2 − r, and thus n ≥ deg A2 + r − 2. For Claim
4, by Theorem 2.4.4 we must have that deg A2 = n + 2 − r; that is, n = deg A2 + r − 2. The
values given for r, ρ1 , π1 are also required by Theorem 2.4.4.
Theorem 2.4.8. Suppose
α = (α0 , α1 , . . . , αn ) ∈ Ln (r, r, 0)
with r ≥ 2 (note that α is quasi-regular). Let A1 , A2 be the characteristic polynomials. We
necessarily have
d1 ∶= deg A1 = r,
and we can choose A2 such that
d2 ∶= deg A2 < deg A1 .
Now, if d2 ≥ 1, then let A3 be the unique polynomial satisfying
A1 = R2 A2 + A3 and d3 ∶= deg A3 < deg A2 ,
28 MICHAEL YIASEMIDES
Case 1: If d2 ≥ 2, then
α(2) ∶= (α0 , α1 , . . . , αd1 +d2 −2 )
is in Ld1 +d2 −2 (d2 , d2 , 0) and has characteristic polynomials A2 , A3 .
Case 2: If d2 = 1, then
α(2) ∶= (α0 , α1 , . . . , αd1 )
is in Ld1 (2, 1, 1) and has characteristic polynomials A2 , A1 (note that the order is important).
Case 3: If d2 = 0, then
α(2) ∶= (α0 , α1 , . . . , αd1 )
is in Ld1 (2, 0, 2) and has characteristic polynomials A2 , A1 (note that the order is important).
The theorem above demonstrates the manifestation of the Euclidean algorithm in Hankel
matrices, and this is made clearer in the corollaries below. The final claim in the theorem is
given in order to demonstrate that even if a sequence is not quasi-regular (which is required
for the main part of the theorem) a truncation can be taken that is quasi-regular and has
the same characteristic polynomials.
Corollary 2.4.9. Let
α = (α0 , α1 , . . . , αn ) ∈ Ln (r, r, 0)
where r ≥ 2. Let the characteristic polynomials of α be A1 , A2 , where we choose A2 such
that deg A2 < deg A1 (which is possible since α is quasi-regular) and note that degree A1 = r.
Define d1 ∶= deg A1 and d2 ∶= deg A2 . Suppose the Euclidean algorithm gives
A1 =R2 A2 + A3 d3 ∶= deg A3 < deg A2 ,
A2 =R3 A3 + A4 d4 ∶= deg A4 < deg A3 ,
⋮ ⋮
At =Rt+1 At+1 + At+2 dt+2 ∶= deg At+2 < deg At+1 ,
where t is such that deg At ≥ 2 > deg At+1 ≥ 0. Finally, let α(1) ∶= α; if t ≥ 2 then let
α(2) ∶=(α0 , α1 , . . . , αd1 +d2 −2 ),
α(3) ∶=(α0 , α1 , . . . , αd2 +d3 −2 ),
⋮
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 29
We now consider the case where deg As ≥ 2, and we define As+1 = 1. We define α(s+1) as
in Corollary 2.4.9. That is, α(s+1) = (α0 , α1 , . . . , αds ) ∈ Lds−1 (2, 1, 1) and has characteristic
polynomials As+1 and As . In particular, the kernel of the matrix (α0 , α1 , . . . , αds ) contains
the polynomials
(27) As+1 , T As+1 , ... , T ds −2 As+1
and
(28) As .
Since As+1 = 1, we deduce from (27) that the first ds − 1 entries of α(s+1) are 0. We now need
only show that αds −1 is non-zero, which follows easily by a contradiction argument: If it were
zero, then (28) would imply that αds is also zero, meaning α(s+1) = 0, which contradicts that
α(s+1) ∈ Lds−1 (2, 1, 1).
30 MICHAEL YIASEMIDES
The following theorem demonstrates how the characteristic polynomials of a sequence change
if we increase the length of the sequence. The intuition behind this result is made clear in
the proof.
Theorem 2.4.11. Let
α = (α0 , α1 , . . . , αn ) ∈ Fn+1
q ,
and let c1 , c2 be the characteristic degrees and A1 , A2 be the characteristic polynomials. Now
let
α′ ∶= (α0 , α1 , . . . , αn , αn+1 ) ∈ Fn+2
q .
be an extension of α, and denote the characteristic degrees by c′1 , c′2 and the characteristic
polynomials by A′1 , A′2 . In what follows we define n′1 ∶= ⌊ (n+1)+2
2 ⌋ and n′2 ∶= ⌊ (n+1)+3
2 ⌋.
There is one value of αn+1 such that α′ ∈ Ln+1 (r, r, 0). In which case, we have c′1 = c1 = r
and c′2 = c2 + 1 = n + 3 − r, and A′1 = A1 and A′2 = A2 .
There are q − 1 values of αn+1 such that α′ ∈ Ln+1 (r + 1, r, 1). In which case, we have
c′1 = c1 + 1 = r + 1 and c′2 = c2 = n + 2 − r, and A′1 = A1 and A′2 = βA2 + T c2 −c1 A1 for some
β ∈ F∗q . There is a one-to-one correspondence between αn+1 and β.
For any value that αn+1 takes in Fq , we have α′ ∈ Ln+1 (r +1, ρ1 , π1 +1). We have c′1 = c1 +1 =
r + 1 and c′2 = c2 = n + 2 − r, and A′1 = A1 and A′2 = βT c2−c1 A1 + A2 for some β ∈ Fq . There is
a one-to-one correspondence between αn+1 and β.
Claim 3: Suppose n is even and α ∈ Ln (n1 , n1 , 0); and so c′1 , c′2 = n1 , and A1 ∈ An1 and
A2 ∈ A<n1 . For any value that αn+1 takes, we have α ∈ Ln+1 (n1 , n1 , 0). We also have
c′1 = c1 = n1 and c′2 = c2 + 1 = n + 3 − n1 = n1 + 1, and A′1 = βA2 + A1 and A′2 = A2 . There is a
one-to-one correspondence between αn+1 and β.
Suppose n is odd and α ∈ Ln (n1 , n1 , 0); and so c′1 = n1 and c′2 = n1 + 1, and A1 ∈ An1 and
A2 ∈ A<n1 .
● There is one value of αn+1 such that α′ ∈ Ln+1 (n1 , n1 , 0); in which case c′1 = c1 = n1
and c′2 = c2 + 1 = n1 + 2, and A′1 = A1 and A′2 = A2 .
● There are q − 1 values of αn+1 such that α′ ∈ Ln+1 (n1 + 1, n1 + 1, 0); in which case
c′1 = c1 + 1 = n1 + 1 and c′2 = c2 = n1 + 1, and A′1 = βA2 + T A1 and A′2 = A1 . There is a
one-to-one correspondence between αn+1 and β.
A vector
v = (v0 , . . . , vm−1 )T
is in the kernel of Hl,m (α) if and only if the vectors
(v ∣ 0) =(v0 , . . . , vm−1 , 0)T ,
(0 ∣ v) =(0, v0 , . . . , vm−1 )T
are in the kernel of Hl−1,m+1 (α). This can be extended, and expressed in terms of polynomi-
als, to give the following result:
Similarly, (0 ∣ v) is in the kernel of Hl−1,m+1 (α) if and only if v is in the kernel of Hl−1,m (α′′ ).
Again, the latter is true, because Hl−1,m (α′′ ) is the matrix we obtain by removing the first
row from Hl,m (α).
The backward implication of the first claim follows from what we have established above.
We now consider the second claim. The first claim tells us that a polynomial A ∈ A ∈ Fq [T ]
with deg A ≤ m − 1 is in the kernel of Hl,m (α) if and only if A and T A are in the kernel of
Hl−1,m+1 (α).
Using the fact that any polynomial in the kernel remains in the kernel after being multiplied
by an element of Fq , we can see that the above holds if and only if Y ⋅ A is in the kernel of
Hl−k,m+k (α) for any deg Y ≤ k.
A related lemma is the following.
Lemma 2.4.13. Let α = (α0 , α1 , . . . , αn ) ∈ Fn+1
q , and define α ∶= (α0 , α1 , . . . , αn−1 ). Also,
′
are in the kernel of Hl−1,m (α′ ) (which is just the matrix Hl,m (α) after removing the last
row).
The proof of this lemma is similar to the proof of Lemma 2.4.12.
We now give the proofs of the four theorems, beginning with Theorem 2.4.4.
Now suppose r = 1 and ρ1 = 1. Then, α0 ≠ 0, and so the matrix Hn+1,1 (α) has full column
rank and its kernel is trivial. The (ρ, π)-form of Hn,2 (α) is
⎛α 0 α1 ⎞
⎜0 0⎟
⎜ ⎟
⎜0 0 ⎟.
⎜ ⎟
⎜ ⋮ ⋮⎟
⎝0 0⎠
So, we can see that ker Hn,2 (α) = {γA1 ∶ γ ∈ Fq } for some A1 ∈ A with deg A1 = ρ1 = 1.
Lemma 2.4.12 tells us that
{B1 A1 ∶ degBB11∈A
≤m−2} ⊆ ker Hl,m (α)
If, instead, we have r = 1 and ρ1 = 0, then, α = (0, . . . , 0, αn ) with αn ≠ 0, and so the matrix
Hn+1,1 (α) has full column rank and its kernel is trivial. We have that
⎛0 0 ⎞
⎜0 0 ⎟
⎜ ⎟
Hn,2(α) = ⎜ ⋮ ⋮ ⎟ ,
⎜ ⎟
⎜0 0 ⎟
⎝0 α n ⎠
We will first address the special subcase when c1 = c2 . This occurs when n is even and
2 = n1 (which also implies that ρ1 = r). When 1 ≤ m ≤ c1 , we have ker Hl,m (α) = {0},
r = n+2
by Corollary 2.4.1. In this subcase, there are no m satisfying c1 + 1 ≤ m ≤ c2 . Suppose now
that m = c2 +1 = n1 +1 and l = c2 −1 = n1 −1. Corollary 2.4.1 tells us that dim ker Hl,m (α) = 2.
Thus, there are polynomials A1 , A2 (neither being a multiple of the other) such that
deg B2 ≤0
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 33
All that remains to be shown is that at least one of A1 , A2 have degree equal to ρ1 = r
(without loss of generality, this will be A1 ). To show this, suppose for a contradiction that
deg A1 , deg A2 < ρ1 = r. Then, the vectors associated to these polynomials are of the form
v =(v0 , v1 , . . . , vr−1 , 0),
w =(w0 , w1 , . . . , wr−1 , 0);
and so the vectors
v′ =(v0 , v1 , . . . , vr−1 ),
w′ =(w0 , w1 , . . . , wr−1 )
are in the kernel of Hl,m−1 (α′ ), where α′ ∶= (α0 , α1 , . . . , αn−1 ). In particular,
dim ker Hl,m−1 (α′ ) = 2. From this, and the fact that Hl,m−1 (α′ ) is the matrix we obtain by
removing the last row from Hl+1,m−1 (α) = Hn1 ,n1 (α), we deduce that the kernel of Hn1 ,n1 (α)
is at least one-dimensional. This contradicts that the kernel is trivial (since the matrix is
invertible).
Suppose now that c1 ≠ c2 . When 1 ≤ m ≤ c1 = r, Corollary 2.4.1 tells us that ker Hl,m (α) =
{0}.
Now suppose that m = c1 + 1 = r + 1. Corollary 2.4.1 tells us that the kernel of H ∶= Hl,m (α)
has dimension 1, and so we let v ≠ 0 be a vector that spans the kernel. The (ρ, π)-form of
H is
⎛H[ρ1 , ρ1 ] H[ρ1 , −(m − ρ1 )] ⎞
⎜ 0 ⎟
⎜ ⎟
0
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ 0 0 ⎟
⎜ 0 1 ⎟
⎜ ⎟
,
⎜ ⎟
0 0
⎜ 0 1 ∗ ⎟
⎜ ⎟
0
⎜ ⎟
⎜ ⎟
⎝ 0 1 ∗ ∗ ⎠
where 1 represents an element in F∗q and ∗ represents an element in Fq . The submatrix
H[ρ1 , ρ1 ] is invertible, and there are π1 number of 1s in the bottom-right submatrix. Of
course, if ρ1 = 0, then the top two submatrices and the bottom-left submatrix should disap-
pear. If π1 = 0, then the bottom-right submatrix should be a zero matrix. Regardless, the
(ρ, π)-form shows us that v must have zeros in its last π1 entries. That is,
⎛ v0 ⎞
⎜ ⋮ ⎟
⎜ ⎟
⎜vρ1 −1 ⎟
⎜ ⎟
v=⎜ ⎟
⎜ vρ1 ⎟ .
⎜ 0 ⎟
⎜ ⎟
⎜ ⎟
⎜ ⋮ ⎟
⎝ 0 ⎠
We must also have that vρ1 ≠ 0. When ρ1 = 0, this is clear. When ρ1 > 0, because v is in the
kernel of H, we can see that (v0 , . . . , vρ1 −1 , vρ1 )T is in the kernel of
⎛ αρ1 ⎞
⎜ ⎟
⎜H[ρ1 , ρ1 ] ⎟;
⋮
⎜ α2ρ1 −2 ⎟
⎝ α2ρ1 −1 ⎠
34 MICHAEL YIASEMIDES
and so, if vρ1 = 0, then (v0 , . . . , vρ1 −1 )T ≠ 0 is in the kernel of H[ρ1 , ρ1 ], contradicting that it
is invertible.
As previously, Lemma 2.4.12 and Corollary 2.4.1 tell us, for c1 + 1 ≤ m ≤ c2 , that
However, the left side has dimension m − c1 = n + 3 − 2r, while, by Corollary 2.4.1, the right
side has dimension 2m − n − 2 = n + 4 − 2r. Thus, there is some polynomial A2 ∈ A in the
kernel of Hl,m (α) with
Thus, we have
deg B2 ≤0
The fact that deg A2 ≤ c2 (that is, A2 has at most its first c2 + 1 coefficients being non-zero)
follows from the fact that the kernel is in (c2 + 1)-dimensional space.
We will now show that if ρ1 is not equal to r = c1 , then deg A2 is necessarily equal to c2 .
Let m = c2 + 1, and let v = (a0 , a1 , . . . , ac2 )T be the vector associated with A2 . Note that the
condition deg A2 = c2 is equivalent to ac2 ≠ 0.
Suppose for a contradiction that ρ1 ≠ r and deg A2 ≠ c2 . This means that π1 ≥ 1 and
v = (a0 , a1 , . . . , ac2 −1 , 0)T . The latter implies that v′ ∶= (a0 , a1 , . . . , ac2 −1 )T is in the ker-
nel of Hl,m−1 (α′ ), where α′ ∶= (α0 , α1 , . . . , αn−1 ). In terms of polynomials, this means
A2 ∈ Hl,m−1 (α′ ).
Note that Hl,m−1 (α′ ) is the matrix we obtain by removing the last row from Hl+1,m−1 (α).
We have already established that
ker Hl+1,m−1 (α) = ker Hc1 ,c2 (α) = {B1 A1 ∶ deg BB1 ≤c
1 ∈A
2 −c1 −1
}.
(Note that, because π1 ≥ 1 every vector in the kernel of Hl+1,m−1 (α) has a zero in its last
entry, which is a requirement for our application of Lemma 2.4.13). However, since we have
established A2 ∈ Hl,m−1 (α′ ), this implies that
A2 ∈ {B1 A1 ∶ deg B
B1 ∈A
1 ≤c2 −c1
},
Finally, it remains to consider when m > c2 + 1. By Lemma 2.4.12, we can see that
Note that (31) also proves that A1 , A2 are coprime. Indeed, if they were not then we could
write
A1 =CA′1 ,
A2 =CA′2 ,
where
deg C ≥1,
deg A′1 ≤ deg A1 − 1 = ρ1 − 1 ≤ c1 − 1 = n + 1 − c2 ,
deg A′2 ≤ deg A2 − 1 = c2 − 1 = n + 1 − c1 .
In particular, taking B1 = A′2 and B2 = −A′1 would contradict (31).
and we note that at least one of xm′ −c1 , ym′ −c2 is non-zero due to there being at least one
equality in the inequalities above. Since B1 A1 + B2 A2 = 0, we have that
(xm′ −c1 T m −c1 A1 + ym′ −c2 T m −c2 A2 )
′ ′
= − ((x0 + x1 T + . . . + xm′ −c1 −1 T m −c1 −1 )A1 + (y0 + y1 T + . . . + ym′ −c2 −1 T m −c2 −1 )A2 ).
′ ′
(Note that this is non-zero because at least one of xm′ −c1 , ym′ −c2 is non-zero, and because A2
is not a multiple of A1 ). By (30), the right side tells us that
T m −c2 Z ∈ Hl′ ,m′ (α),
′
for all 0 ≤ s ≤ m′ − c2 − 1. But then, using the second result in Lemma 2.4.12 for the first
relation below, we have that
Z ∈ Hl′ +(m′ −c2 ),m′ −(m′ −c2 ) (α) = Hc1 ,c2 (α) = {XA1 ∶ deg X≤c
X∈A
2 −c1 −1
}.
Recalling the definition of Z, we can see that this implies
A2 ∈ {XA1 ∶ deg X≤c
X∈A
2 −c1
},
which contradicts (29).
We now prove Theorem 2.4.8, which will be required for the proof of Theorem 2.4.7 after-
wards.
Proof of Theorem 2.4.8. For what follows, we define α′ ∶= (α0 , α1 , . . . , α2d1 ).
deg B2 ≤0
Recall that deg A1 = d1 and deg A2 = d2 < d1 . In particular, the vectors associated to A1 , A2
in Hd1 −1,n+3−d1 (α) have at least n + 2 − 2d1 zero entries at the end. Therefore, consider the
matrix that we obtain by removing the last n + 2 − 2d1 columns of Hd1 −1,n+3−d1 (α), which is
Hd1 −1,d1 +1 (α′ ). From the above we can see that
deg B2 ≤0
Given that Hd1 ,d1 (α′ ) = Hr,r (α′ ) is invertible, we can see that the dimension of
ker Hd1 −1,d1 +1 (α′ ) is 2 and so we must have equality above:
Now, we have deg A2 = d2 < d1 , and so the polynomial associated to A2 in ker Hd1 −1,d1 +1 (α′ )
has d1 − d2 zeros entries at the end. What we do next depends on the size of d2 .
The fact that the left side is a subset of the right side follows from (32) and successive appli-
cations of Lemma 2.4.13. Equality then follows by noting that removing a row will increase
the dimension by (at most) 1. Also, we can see that A3 is in the kernel of Hd2 −1,d1 +1 (α(2) )
and we can replace A1 with A3 :
deg B3 ≤0
We can now deduce that the characteristic polynomials of α(2) are A2 , A3 , and that α(2) ∈
Ld1 +d2 −2 (d2 , d2 , 0).
(recall that the definition of α(2) differs between the cases), and we can see by similar means
as in Case 1 that it has kernel
Note that B2 cannot have degree equal to d1 −d2 , and so A3 is not in the kernel above. Thus,
we have that the characteristic polynomials are A2 , A1 (recall that the order matters). The
fact that B2 cannot have degree equal to d1 − d2 also tells us that π(α(2) ) = 1. We deduce
that α(2) ∈ Ld1 (2, 1, 1).
We now prove the uniqueness claim that is made in the theorem. We do so for Case 1, and
the remaining cases are almost identical and only need to take into account the difference
in definition of α(2) . To this end, consider the sequences
α(2) =(α0 , α1 , . . . , αd1 +d2 −1 )
α′ =(α0 , α1 , . . . , αd1 +d2 −1 , αd1 +d2 , . . . , α2d1 ).
Note that αd1 +d2 , . . . , α2d1 form the last d1 − d2 + 1 entries in the last column of the matrix
Hd1 −1,d1 +1 (α′ ), and by assumption we must have that A1 is in the kernel of this matrix.
Thus, for d1 + d2 ≤ i ≤ 2d1 , the i-th row of Hd1 −1,d1 +1 (α′ ) is orthogonal to the vector asso-
ciated with A1 . Since this vector has non-zero final entry, we can express the last entry in
the i-th row in terms of the previous entries. An inductive argument proves that the entries
αd1 +d2 , . . . , α2d1 can be uniquely determined in terms of the entries α0 , α1 , . . . , αd1 +d2 −1 .
All that remains is to prove the final claim made in the theorem, which we do for the case
r ≥ 2. We have that
α = (α0 , α1 , . . . , αn ) ∈ Ln (r, ρ1 , π1 ).
By considering the (ρ, π)-form of the matrix Hn1 ,n2 (α) and removing last π1 columns, we
can deduce that α(1) is in Ln−π1 (ρ1 , ρ1 , 0). Given that A1 is in the kernel of Hn+1−r,r+1 (α),
and that the associated vector has π1 number of zero entries at the end, we can see that A1
is in the kernel of matrix obtained by removing the last π1 columns of Hn+1−r,r+1 (α); that
is, the matrix Hn+1−r,ρ1 +1 (α(1) ). This tells us that A1 is the first characteristic polynomial.
Similarly, since A2 is in the kernel of Hr−1,n+3−r (α), we can see that it is in the kernel
of the matrix obtained by removing the last π1 rows of Hr−1,n+3−r (α); that is, the matrix
Hρ1 −1,n+3−r (α(1) ). This tells us that A2 is the second characteristic polynomial.
We now prove Theorem 2.4.7.
38 MICHAEL YIASEMIDES
Proof of Theorem 2.4.7. The construction of the sequences α in Claims 1 and 2 are not
difficult, and so we proceed to Claims 3 and 4.
(2) The entries α1 , . . . , αdt can be expressed uniquely in terms of α0 . In fact, each such
αi can be expressed as a linear function of α0 that passes through the origin.
(3) The sequence α is uniquely determined by the sequence α(t+1) , which follows from the
uniqueness claim at the end of Corollary 2.4.9. In fact, we can show that for all i ≤ n
the entry αi can be expressed as a linear function of α0 that passes through the origin.
(4) Finally, we note from the above that if α0 = 0, then α = 0. We dismiss this case as
it does not give us α(t+1) ∈ Ldt (2, 1, 1).
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 39
Now suppose that we have dt+1 = 0 instead. We can use a similar argument as above. The
main difference will be that α0 , . . . , αdt −2 = 0, while αdt −1 will be free to take any value in F∗q ,
and for all dt ≤ i ≤ n the entry αi can be expressed as a linear function of αdt −1 that passes
through the origin.
Note that removing the last π1 rows of Hr−1,n+3−r (α) will leave us with the matrix
Hρ1 −1,n+3−r (α(1) ) from (35). Hence, by successive applications of Lemma 2.4.13, regard-
less of the way we extended α(1) , we certainly have that
The requirement that A2 is in the kernel above will actually uniquely determine the en-
tries αn−π1 +1 , . . . , αn (which form the extended part of α) in terms of α0 , . . . , αn−π1 . This
follows from the fact that αn−π1 +1 , . . . , αn appear as the final entries in the last π1 rows of
Hr−1,n+3−r (α), and the fact that A2 has degree n + 2 − r.
We will now show that α is in Ln (r, ρ1 , π1 ). This is a condition of Claim 3, but it also
allows us to determine the dimension of the left side of (36) by using Corollary 2.4.1. By
comparing this to the dimension of the right side of (36), we see that we must have equality,
and thus allowing us to deduce that A1 , A2 are the characteristic polynomials of α. All that
will be left to prove is that α is unique up to multiplication by elements in F∗q .
deg B2 ≤0
In fact, we must have equality above. This follows from the fact that the dimension of
ker H ′ is one less than that of ker H. Indeed, the number of columns remains the same, but
the rank of H ′ is one more than the rank of H, which follows by the fact that H ′ has full
row rank since ρ(α′ ) ≥ ρ(α(1) ) = ρ1 . Now, because T n+2−r−ρ1 A1 is not in the kernel of H ′ ,
we can see that π(H ′ ) = 1, and thus π(α′ ) = 1. By considering (ρ, π)-forms, we can see
that this forces π(α′ ) = π1 (indeed, extending the sequence by one entry will increase the
π-characteristic by one), and thus α ∈ Ln (r, ρ1 , π1 ).
Finally, for the uniqueness claim, this follows from the uniqueness claim in Claim 2 and the
way we formed our extension α of α(1) . Technically, we should also prove a converse result;
that any sequence α with the properties in Claim 3 has a truncation α(1) ∈ Ln−π1 (ρ1 , ρ1 , 0)
40 MICHAEL YIASEMIDES
with characteristic polynomials A1 , A2 . This is to ensure that our construction above actually
addresses all possibilities for α. This is not difficult to prove, and we have done something
similar in the proof of Theorem 2.4.8.
Finally, we prove Theorem 2.4.11.
Proof of Theorem 2.4.11. Claim 1: The cases r ≤ 1 are considerably easier than the cases
r ≥ 2; we consider only the latter.
Let us now consider the characteristic polynomials. Consider the case γ = 0 first. By defi-
nition, we have that A1 spans the kernel of Hc2 −1,c1 +1 (α). By comparing the (ρ, π)-form of
Hc2 −1,c1 +1 (α) to the (ρ, π)-form of Hc′2 −1,c′1 +1 (α′ ) = Hc2 ,c1 +1 (α′ ) (the latter having an extra
row of zeros at the bottom, compared to the former), we see that A1 spans the kernel of
Hc′2 −1,c′1 +1 (α′ ). Thus, A′1 = A1 .
Regarding A′2 , by definition it is the polynomial in the kernel of Hc′1 −1,c′2 +1 (α′ ) that is not a
multiple of A′1 . Now, let a2 be the vector in the kernel of Hc1 −1,c2 +1 (α) that is associated to
A2 . We can see that (a2 ∣ 0) is in the kernel of Hc′1 −1,c′2 +1 (α′ ) = Hc1 −1,c2 +2 (α′ ) (the latter ma-
trix has an extra column compared to the former). In terms of polynomials, this means A2 is
in the kernel of Hc′1 −1,c′2 +1 (α′ ); and since it is not a multiple of A′1 = A1 , we have that A′2 = A2 .
Now consider the case where γ ≠ 0. Let us write a1 for the vector associated to A1 in the
kernel of Hc2 −1,c1 +1 (α). Similar to above, we compare the (ρ, π)-form of Hc2 −1,c1 +1 (α) to
the (ρ, π)-form of Hc′2 −1,c′1 +1 (α′ ) = Hc2−1,c1 +2 (α′ ) (the latter having an additional column
compared to the former, with a non-zero entry in the bottom-right), and we we see that
(a1 ∣ 0) spans the kernel of Hc′2 −1,c′1 +1 (α′ ). Thus, A′1 = A1 .
For A′2 , as above, it is the polynomial in the kernel of Hc′1 −1,c′2 +1 (α′ ) = Hc1 ,c2 +1 (α′ ) that is
not a multiple of A′1 . Since A′1 = A1 is the first characteristic polynomial, we have
ker Hc′1 −1,c′2 +1 (α′ ) ⊇ {B1 A′1 ∶ deg B
B1 ∈A
′ ′ } = {B1 A1 ∶
1 ≤c2 −c1 deg B1 ≤c2 −c1 −1}.
B1 ∈A
We also have
(37) T c2 −c1 A1 ∈/ ker Hc′1 −1,c′2 +1 (α′ ).
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 41
Otherwise, by Lemma 2.4.12, we would have that A1 is in the kernel of Hc′2 −2,c′1 (α′ ), which is
a contradiction. Note also that Hc1 −1,c2 +1 (α) is the matrix we obtain after removing the last
row from Hc′1 −1,c′2 +1 (α′ ) = Hc1 ,c2 +1 (α′ ). In particular, the kernel of the latter is a subspace
of the former. That is,
ker Hc′1 −1,c′2 +1 (α′ ) ⊆ ker Hc1 −1,c2 +1 (α) = {B1 A1 + B2 A2 ∶ deg B1 ≤c2 −c1 }.
B1 ,B2 ∈A
deg B2 ≤0
{B1 A1 ∶ deg BB1 ≤c } ⊆ ker Hc′1 −1,c′2 +1 (α′ ) ⊆ {B1 A1 + B2 A2 ∶ deg B1 ≤c2 −c1 }.
B1 ,B2 ∈A
1 ∈A
2 −c1 −1
deg B2 ≤0
Corollary 2.4.1 tells us that the dimensions of adjacent vector spaces above differ by 1. Now,
A′2 is the vector in ker Hc′1 −1,c′2 +1 (α′ ) that is not a multiple of A′1 = A1 , and recall that by
Theorem 2.4.4 the degree of A′2 must be equal to c′2 . Thus, we can deduce
A′2 = βA2 + T c2 −c1 A1
for some β ∈ F∗q (β cannot be 0 by (37)). It is not difficult to see that if we change the
value of αn+1 (which appears in the last row of Hc′1 −1,c′2 +1 (α′ )), then the value of β will
have to change to ensure that the vector associated to A′2 is orthogonal to the last row of
Hc′1 −1,c′2 +1 (α′ ), and thus in the kernel of Hc′1 −1,c′2 +1 (α′ ).( Note that this makes use of the fact
that A2 is not in the kernel of Hc′1 −1,c′2 +1 (α′ ); indeed, otherwise, A2 would be orthogonal to
the last row and altering the value of β would have no effect). Hence, there is a one-to-one
correspondence between αn+1 and β.
Claim 2: Let H ∶= Hn1 ,n2 (α). The (ρ, π)-form of H is of the form
⎛H[ρ1 , ρ1 ] H[ρ1 , −(n2 − ρ1 )] ⎞
⎜ 0 ⎟
⎜ ⎟
0
⎜ ⎟
⎜ ⎟
⎜ 0 ⎟
⎜ 0 ⎟
⎜ 1 ⎟
⎜ ⎟
(38) ,
⎜ ⎟
0 0 0
⎜ ∗ ⎟
⎜ ⎟
0 0 1
⎜ ⎟
⎜ ⎟
⎝ 0 0 1 ∗ ∗ ⎠
where there are π1 number of 1s in the bottom right submatrix (recall, 1 represents an
element in F∗q , 0 represents 0 as usual, and ∗ represents an element in Fq ). Now consider
the matrix H ′ = Hn′1 ,n′2 (α′ ), and note that H ′ is the same matrix as H but it has either an
additional row or column at the end (depending on whether n is even or odd). Thus, we can
deduce that the (ρ, π)-form of H ′ is the same as (38) but with an additional row or column.
This additional row or column contributes an additional 1 in the bottom-right submatrix,
and thus we have π(α′ ) = π1 + 1. Clearly ρ(α′ ) = ρ1 , and thus α′ ∈ Ln (r + 1, ρ1 , π1 + 1), as
required. The claims on the characteristic degrees follow by definition.
Now, the proof that A′1 = A1 is similar as in the second case of Claim 1. The proof for
the second characteristic polynomial is also similar, but slightly different, and so we give an
outline of the proof.
By definition of the first characteristic polynomial, we have that A′1 = A1 spans the kernel
of Hc′2 −1,c′1 +1 (α). Thus, using Lemma 2.4.12 for the second relation below, we have that
but
(39) T c2 −c1 A1 ∈/ ker Hc′1 −1,c′2 +1 (α′ ).
We also have that
ker Hc′1 −1,c′2 +1 (α′ ) ⊆ ker Hc′1 −2,c′2 +1 (α) = ker Hc1 −1,c2 +1 (α) = {B1 A1 + B2 A2 ∶ deg B1 ≤c2 −c1 }.
B1 ,B2 ∈A
deg B2 ≤0
Thus, we have
{B1 A1 ∶ deg BB1 ≤c } ⊆ ker Hc′1 −1,c′2 +1 (α′ ) ⊆ {B1 A1 + B2 A2 ∶ deg B1 ≤c2 −c1 },
B1 ,B2 ∈A
1 ∈A
2 −c1 −1
deg B2 ≤0
and, by Corollary 2.4.1, the dimensions of adjacent vector spaces above differ by 1. Thus,
by (39), and the fact that Theorem 2.4.4 tells us that A′2 must be of degree c′2 , we have that
A′2 = βT c2−c1 A1 + A2
for some β ∈ Fq . The one-to-one correspondence between αn+1 and β follows by similar
reasoning as in Claim 2.
Claim 3: Consider the first case given in Claim 3. The fact that α ∈ Ln+1 (n1 , n1 , 0), and
the claims on the characteristic degrees, are not difficult to deduce. Thus, we restrict our
attention to the characteristic polynomials. By definition, A′1 is the polynomial that spans
the kernel of Hn1 ,n1 +1 (α′ ), and Theorem 2.4.4 tells us that deg A′1 = n1 . Now, the matrix
Hn1 ,n1 +1 (α′ ) is the same as the matrix ker Hn1 −1,n1 +1 (α) but with an additional row, and we
know that
ker Hn1 −1,n1 +1 (α) = {B1 A1 + B2 A2 ∶ B1 , B2 ∈ Fq },
and so
ker Hn1 ,n1 +1 (α′ ) ⊆ {B1 A1 + B2 A2 ∶ B1 , B2 ∈ Fq }.
Since we know deg A′1 = n1 , and that deg A1 = n1 and deg A2 < n1 , we must have that
A′1 = βA2 + A1 for some β ∈ Fq . The one-to-one correspondence between αn+1 and β follows
similarly as previously. The second characteristic polynomial is, by definition, the polyno-
mial in the kernel of Hn1 −1,n1 +2 (α′ ) that is not a multiple of A′1 . Similar to the previous
claims, we note that Hn1 −1,n1 +2 (α′ ) is the same as the matrix Hn1 −1,n1 +1 (α′ ) but with an
additional column. So, because A2 is in the kernel of the latter, we can see that it must be
in the kernel of the former. Hence A′2 = A2 .
The second case of Claim 3 is very similar to Claim 1. However, for the second bullet point,
one should note that, because α is quasi-regular (unlike in the analogous result in Claim
1), we require that deg A′1 = n1 + 1, and this is why we take A′1 = βA2 + T A1 (as opposed
to A′2 = βA2 + T A1 = βA2 + T c2 −c1 A1 which would be completely analogous to Claim 1).
However, this “swapping” of A′1 and A′2 is important and natural as it plays a role in the
manifestation of the Euclidean algorithm that we saw in Theorem 2.4.8 and its corollaries.
The third case of Claim 3 is very similar to Claim 2. Again, there is a similar “swapping”
of A′1 and A′2 .
Remark 2.4.14. Recall that in Subsection 1.4 we considered the generalisation of Theorem
1.2.1 to higher moments such as the third, and we described that we would need to determine
how many α, β, γ there are such that Hl+1,m+1 (α), Hl+1,m+1 (β), and Hl+1,m+1 (γ) have cer-
tain given ranks, and α + β + γ = 0. Now that we have established various results on Hankel
matrices, we can reduce this problem to certain special cases.
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 43
Let us write
α =(α0 , α1 , . . . , αn )
β =(β0 , β1 , . . . , βn )
γ =(γ0 , γ1 , . . . , γn )
and
α− =(α0 , α1 , . . . , αn−1 )
β − =(β0 , β1 , . . . , βn−1 )
γ − =(γ0 , γ1 , . . . , γn−1 )
Suppose we know how many α− , β − , γ − there are with certain given (ρ, π)-characteristics and
α− + β − + γ − = 0. If at least one of these sequences α− , β − , γ − has non-zero π-characteristic
(that is, not all are quasi-regular) then we can determine the (ρ, π)-characteristics of α, β, γ,
even when we impose the condition that α + β + γ = 0.
Indeed, without loss of generality, suppose that π(γ − ) ≥ 1. Theorem 2.4.11 allows us to de-
termine the (ρ, π)-characteristics of α, β based on the (ρ, π)-characteristics of α− , β − . How-
ever, if, for example, α− is quasi-regular, then the (ρ, π)-characteristic of α would depend
on αn . Theorem 2.4.11 tells us how many values αn can take so that the (ρ, π)-characteristic
of α takes a specific value. However, it does not tell us exactly what those values are, and
this is why it is important that π(γ − ) ≥ 1: Regardless of the values of αn and βn , we know
there exists γn such that αn + βn + γn = 0, and we still know what the (ρ, π)-characteristic of
γ is because it is independent of the value of γn (which uses the fact that π(γ − ) ≥ 1).
Thus we have reduce our problem to the cases where α− , β − , γ − are all quasi-regular.
We have that
2
∑ ( ∑ d(B))
A∈Mn B∈I(A;h)
2
= ∑ ( ∑ d(B))
A∈Mn B∈Mn
deg(B−A)<h
2
= ∑ ( ∑ ∑ ∑ ∑ 1EF =B )
A∈Mn B∈Mn 0≤l,m≤n E∈Ml F ∈Mm
deg(B−A)<h l+m=n
2
= ∑ ( ∑ ∑ ∑ ∑ 1EF =B ) .
A∈A≤n B∈Mn 0≤l,m≤n E∈Ml F ∈Mm
deg(B−A)<h l+m=n
44 MICHAEL YIASEMIDES
For the last line, we changed the first summation range from A ∈ Mn to all A ∈ A with
deg A ≤ n. This does not change the result because the conditions on E and F force A to
be in Mn . Continuing, we have
2
∑ ( ∑ d(B))
A∈Mn B∈I(A;h)
2
= ∑ ( ∑ ∏ 1{EF }i =bi ) .
n
∑ ∑ ∑
A∈A≤n B∈Mn 0≤l,m≤n E∈Ml F ∈Mm i=0
deg(B−A)<h l+m=n
Here, for a polynomial C, we let {C}i denote its i-ith coefficient, which is convenient
when working with products of polynomials such as EF . We also denote i-th coefficient
of A, B, E, F by ai , bi , ei , fi , respectively.
We will now express the sums over A, B, E, F by sums over their coefficients. For example
∑A∈A≤n will be expressed as ∑a0 ,...,an ∈Fp . We also note that if A′ satisfies deg(A − A′ ) < h,
then (due to the sum over B) both A and A′ give the same contribution. For a given A,
there are ph such A′ . Thus, we can multiply the whole expression by ph and consider only
the A that are of the form
Thus, using the above and applying (8) to the terms 1{EF }i =bi , we obtain
∑ ( ∑ d(B))
A∈Mn B∈I(A;h)
ph ⎛
= ∑ ∑ ∑ ∑ ∑
p2n+2 ah ,...,an ∈Fp ⎝ b0 ,...,bh−1 ∈Fp 0≤l,m≤n e0 ,...,el−1 ∈Fp f0 ,...,fm−1 ∈Fp
l+m=n el =1 fm =1
⎞
2
For any element in Fp that appears in the exponentials above, we take it as an integer in
{0, 1, . . . , p − 1}. Now, consider the sum over one of the coefficients bi , and apply (8). We
obtain,
if αi ∈ F∗p ,
∑ exp ( αi bi ) = {
1 2πi 0
p bi ∈Fp p 1 if αi = 0.
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 45
Thus, we require αi = 0 in order to have a non-zero contribution to (41). Applying this for
i = 0, . . . , h − 1, we obtain
(42)
2
∑ ( ∑ d(B))
A∈Mn B∈I(A;h)
⎛ ⎞
2
⎛ ⎞
∑ exp ( αi (ai − ∑ el1 fm1 ))
n
p3h
=
2πi
∑ ∑ ∑ ∑ ∏
ah ,...,an ∈Fp ⎝ 0≤l,m≤n e0 ,...,el−1 ∈Fp f0 ,...,fm−1 ∈Fp i=h αi ∈Fp ⎠
p 2n+2 p 0≤l1 ≤l
l+m=n el =1 fm =1 0≤m1 ≤m
l1 +m1 =i
⎛ ⎞
∏ ∑ exp ( βi (ai − gl1′ hm′1 )) .
n
2πi
∑ ∑ ∑ ∑
⎝ 0≤l′ ,m′ ≤n g0 ,...,gl′ −1 ∈Fp h0 ,...,hm′ −1 ∈Fp i=h βi ∈Fp ⎠
⋅
p 0≤l1′ ≤l′
l′ +m′ =n gl′ =1 hm′ =1 0≤m′1 ≤m′
l1′ +m′1 =i
We now consider the sum over one of the coefficients ai . Unlike the bi which appeared within
the largest parentheses, the ai appear outside. Thus, we must simultaneously consider the
terms within each pair of parentheses whose product forms the square, and that is why we
have written them explicitly in the last line above. Again, we apply (8) to obtain
if αi ≠ βi ,
∑ exp ( (αi + βi )ai ) = {
2πi 0
ai ∈Fp p p if αi = −βi .
∑ ( ∑ d(B))
A∈Mn B∈I(A;h)
⎛ ⎞
∏ exp ( − ∑ el1 fm1 )
n
=p2h−n−1
2πi
∑ ∑ ∑ ∑
αh ,αh+1 ,...,αn ∈Fp ⎝ 0≤l,m≤n e0 ,...,el−1 ∈Fp f0 ,...,fm−1 ∈Fp i=h ⎠
αi
p 0≤l1 ≤l
l+m=n el =1 fm =1 0≤m1 ≤m
l1 +m1 =i
⎛ ⎞
∏ exp ( gl1′ hm′1 )
n
2πi
∑ ∑ ∑ ∑
⎝ 0≤l′ ,m′ ≤n g0 ,...,gl′ −1 ∈Fp h0 ,...,hm′ −1 ∈Fp i=h ⎠
⋅ αi
p 0≤l1′ ≤l′
l′ +m′ =n gl′ =1 hm′ =1 0≤m′1 ≤m′
l1′ +m′1 =i
⎛ ⎞
∏ exp ( − ∑ el1 fm1 )
n
=p2h−n−1 ∑
2πi
∑ ∑ ∑
α∈Lnh ⎝ 0≤l,m≤n e0 ,...,el−1 ∈Fp f0 ,...,fm−1 ∈Fp i=h ⎠
αi
p 0≤l1 ≤l
l+m=n el =1 fm =1 0≤m1 ≤m
l1 +m1 =i
⎛ ⎞
∏ exp ( gl1′ hm′1 )
n
2πi
∑ ∑ ∑ ∑
⎝ 0≤l′ ,m′ ≤n g0 ,...,gl′ −1 ∈Fp h0 ,...,hm′ −1 ∈Fp i=h ⎠
⋅ αi
p 0≤l1′ ≤l′
l′ +m′ =n gl′ =1 hm′ =1 0≤m′1 ≤m′
l1′ +m′1 =i
⎛ ⎞
=p2h−n−1 ∑ ∑ exp ( − e Hl+1,m+1 (α)f)
2πi T
∑
α∈Lnh ⎝ 0≤l,m≤n e∈Fp l ×{1} p ⎠
l+m=n f ∈Fp m ×{1}
46 MICHAEL YIASEMIDES
⎛ ⎞
exp ( g Hl′ +1,m′ +1 (α)h)
2πi T
∑ ∑
⎝ 0≤l′ ,m′ ≤n ⎠
⋅
′ p
g∈Fp l ×{1}
l′ +m′ =n ′
h∈Fp m ×{1}
and we define
α− ∶= (α0 , α1 , . . . , αn−1 ).
The two claims below address all possible values that α could take.
Claim 1: Suppose α is such that α− ∈ Ln−1 (r, ρ1 , π1 ), where 0 ≤ r ≤ n1 − 1 and π1 ≥ 1. For
all l, m ≥ 0 with l + m = n, we have
Now suppose n is odd and α is such that α− ∈ Ln−1 (n1 , 0, 0). For all l, m ≥ 0 with l + m = n,
we have
Claim 2: Suppose α is such that α− ∈ Ln−1 (r, r, 0), where 0 ≤ r ≤ n1 − 1. We will fix α−
and consider all possible extensions α; that is, we let αn vary in Fp . We have
⎛ ⎞
∑ exp ( − e Hl+1,m+1 (α)f)
2πi T
∑ ∑
αn ∈Fp ⎝ 0≤l,m≤n e∈Fp l ×{1} p ⎠
l+m=n f ∈Fp m ×{1}
⎛ ⎞
exp ( g Hl′ +1,m′ +1 (α)h)
2πi T
∑ ∑
⎝ 0≤l′ ,m′ ≤n ⎠
⋅
′ p
g∈Fp l ×{1}
l′ +m′ =n ′
h∈Fp m ×{1}
=p2n−2r+1 (n + 1 − 2r)2 .
Proof. In what follows, ei and fi are the i-th entries of e and f, respectively.
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 47
Claim 1: Consider the first result in this claim. Since π1 ≥ 1, we can see from Theorem
2.4.4 tells us that for m ≤ n − r − 1 there is no monic polynomial of degree m in the kernel of
Hl,m+1 (α− ). That is, there is no vector f ∈ Fp m × {1} such that Hl,m+1 (α− )f = 0. Therefore,
for any f ∈ Fp m × {1} we can find a row, Ri , of Hl,m+1 (α) such that Ri f ≠ 0. Now, consider
only the sum and terms involving ei on the left side of (43); we have
∑ exp ( − ei Ri f) = 0,
2πi
ei ∈Fp p
where we have used (8) and the fact that Ri f ≠ 0. Thus, the left side of (43) is indeed 0.
Let us now consider when m ≥ n − r. Let c−1 ∶= r and c−2 ∶= n + 1 − r. Theorem 2.4.4 tells us
that
for some A1 ∈ Aρ1 and A2 ∈ Ac−2 . Note that in this case, there are monic polynomials of
degree m in the kernel of Hl,m+1 (α− ). That is, there are vectors f ∈ Fp m × {1} such that
Hl,m+1 (α− )f = 0. By similar reasoning as above, these are the only candidates for non-zero
contributions to the left side of (43).
Hence,
eT Hl+1,m+1 (α)f = γRl+1 gm ,
where Rl+1 is the last row of Hl+1,m+1 (α). Thus, we have
= exp ( − γRl+1 gm )
2πi
∑
e∈Fp l ×{1}
p
− m
f ∈(Fp ×{1})∩ker Hl+1,m+1 (α)
γ∈Fp
48 MICHAEL YIASEMIDES
for m ≤ r − 1 and m ≥ n + 1 − r.
As previously, a non-zero contribution requires that f ∈ ker Hl,m+1 (α− ). Let Rl+1 (α) be the
last row of Hl+1,m+1 (α). We can deduce the following two points from Theorem 2.4.11:
● There are p − 1 values of αn such that
Ln (n1 , n1 , 0) if n is even and r = n1 − 1,
α∈{
Ln (r + 1, r, 1) otherwise.
Theorem 2.4.4 tells us that
ker Hl,m+1 (α− ) = {B1 A1 ∶ degBB11∈A
≤m−r }
and
ker Hl+1,m+1 (α) = {B1 A1 ∶ deg BB11≤m−1−r
∈A
},
for some A1 ∈ Mr . Thus, any f ∈ (Fp m × {1}) ∩ ker Hl,m+1 (α− ) can be written as
f = f − + am for some f − ∈ ker Hl+1,m+1 (α), and am is the polynomial associated with
T m−r A1 . This gives
eT Hl+1,m+1 (α)f = Rl+1 (α)am ≠ 0.
Note that Rl+1 (α)am is independent of the values of l, m (so long as r ≤ m ≤ n − r)
This is because the non-zero entries of am occur in the last r + 1 entries, and the last
r + 1 entries Rl+1 (α) are independent of the value of l.
Similar statements hold for the sums over g, h, but we should keep in mind that
there is no negative sign in the exponential in these sums. Hence, we have
⎛ ⎞
∑ exp ( − e Hl+1,m+1 (α)f)
2πi T
∑
⎝ 0≤l,m≤n e∈Fp l ×{1} p ⎠
l+m=n f ∈Fp m ×{1}
THE VARIANCE OF THE DIVISOR FUNCTION IN Fp [T ] AND HANKEL MATRICES 49
⎛ ⎞
exp ( g Hl′ +1,m′ +1 (α)h)
2πi T
∑ ∑
⎝ 0≤l′ ,m′ ≤n ⎠
⋅
′ p
g∈Fp l ×{1}
l′ +m′ =n ′
h∈Fp m ×{1}
⎛ ⎞
exp ( − Rl+1 (α)am )
2πi
= ∑ ∑
⎝ l+m=n e∈Fp l ×{1}
p ⎠
r≤m≤n−r −
f ∈ker Hl+1,m+1 (α)
⎛ ⎞
exp ( Rl+1 (α)am )
2πi
∑ ∑
⎝ ⎠
⋅
l′ +m′ =n ′ p
g∈Fp l ×{1}
r≤m′ ≤n−r −
h ∈ker Hl′ +1,m′ +1 (α)
⎛ ⎞⎛ ⎞
= ∑ ∑ ∑ ∑
⎝ ⎠⎝ ⎠
1 1
l+m=n e∈Fp l ×{1} l′ +m′ =n ′
g∈Fp l ×{1}
r≤m≤n−r
f − ∈ker Hl+1,m+1 (α) r≤m′ ≤n−r
h− ∈ker Hl′ +1,m′ +1 (α)
=((n + 1 − 2r)pn−r ) ,
2
where we have used the fact that ∣Fp l × {1}∣ = pl and ∣ker Hl+1,m+1 (α)∣ = p(m+1)−(r+1) .
● There is one value of αn+1 such that α ∈ Lnh (r, r, 0). In which case, all f in
ker Hl,m+1 (α− ) are also in ker Hl+1,m+1 (α), thus giving Rl+1 (α)f = 0. A similar state-
ment holds for the sums over g, h. Hence, we have
⎛ ⎞
∑ exp ( − e Hl+1,m+1 (α)f)
2πi T
∑
⎝ 0≤l,m≤n e∈Fp l ×{1} p ⎠
l+m=n f ∈Fp m ×{1}
⎛ ⎞
exp ( g Hl′ +1,m′ +1 (α)h)
2πi T
∑ ∑
⎝ 0≤l′ ,m′ ≤n ⎠
⋅
′ p
g∈Fp l ×{1}
l′ +m′ =n ′
h∈Fp m ×{1}
⎛ ⎞⎛ ⎞
= ∑ ∑ ∑ ∑
⎝ ⎠⎝ ⎠
1 1
l+m=n e∈Fp l ×{1} l′ +m′ =n ′
g∈Fp l ×{1}
r≤m≤n−r −
f ∈ker Hl+1,m+1 (α) r≤m′ ≤n−r −
h ∈ker Hl′ +1,m′ +1 (α)
=((n + 1 − 2r)pn−r ) .
2
Thus, we have
⎛ ⎞
∑ exp ( − e Hl+1,m+1 (α)f)
2πi T
∑ ∑
αn ∈Fp ⎝ 0≤l,m≤n e∈Fp l ×{1} p ⎠
l+m=n f ∈Fp m ×{1}
⎛ ⎞
exp ( g Hl′ +1,m′ +1 (α)h)
2πi T
∑ ∑
⎝ 0≤l′ ,m′ ≤n ⎠
⋅
′ p
g∈Fp l ×{1}
l′ +m′ =n ′
h∈Fp m ×{1}
= ∑ ((n + 1 − 2r)pn−r )
2
αn ∈Fp
=p2n−2r+1 (n + 1 − 2r)2 .
Acknowledgements: This research was conducted during a postdoctoral fellowship funded
by the Leverhulme Trust research project grant “Moments of L-functions in Function Fields
50 MICHAEL YIASEMIDES
and Random Matrix Theory” (grant number RPG-2017-320) secured by Julio Andrade. The
author is most grateful for this support.
References
[1] J. C. Andrade, L. Bary-Soroker and Z. Rudnick. Shifted Convolution and the Titchmarsh Divisor
Problem over Fq [T ]. Phil. Trans. R. Soc. A, 373 (2015).
[2] G. Coppola and S. Salerno. On the Symmetry of the Divisor Function in Almost All Short Intervals.
Acta Arith., 113(2) (2004), 189–201.
[3] H. Cramér. Über Zwei Sätze von Herrn G. H. Hardy. Math. Z., 15 (1922), 201–210.
[4] M. Garcı́a-Armas, S. R. Ghorpade and S. Ram. Relatively Prime Polynomials and Nonsingular Hankel
Matrices over Finite Fields. J. Comb. Theory Ser. A, 118 (2011), 819–828.
[5] D. R. Heath-Brown. The Distribution and Moments of the Error Term in the Dirichlet Divisor Problem.
Acta Arith., 60(4) (1992), 389–415.
[6] G. Heinig. Kernel Structure of Toeplitz-plus-Hankel Matrices. Linear Algebra Appl., 340(1) (2002),
1–13.
[7] G. Heinig and P. Jankowski. Kernel Structure of Block Hankel and Toeplitz Matrices and Partial
Realization. Linear Algebra Appl., 175 (1992), 1–30.
[8] G. Heinig and K. Rost. Algebraic Methods for Toeplitz-like Matrices and Operators. Birkhäuser Basel,
Basel (1984).
[9] A. Ivić. On the Divisor Function and the Riemann Zeta-function in Short Intervals. Ramanujan J., 19
(2009), 207–224.
[10] M. Jutila. On the Divisor Function for Short Intervals. Studies in Honour of Arto Kustaa Salomaa on
the Occasion of his Fiftieth Birthday. Ann. Univ. Turku. Ser. A, I(186) (1984), 23–30.
[11] J. P. Keating, E. Roditty-Gershon, B. Rodgers and Z. Rudnick. Sums of Divisor Functions in Fq [t] and
Matrix Integrals. Math. Z., 288 (2018), 167–198.
[12] S. Lester. On the Variance of Sums of Divisor Functions in Short Intervals. Proc. Amer. Math. Soc.,
144(12) (2016), 5015–5027.
[13] S. Lester and N. Yesha. On the Distribution of the Divisor Function and Hecke Eigenvalues. Israel J.
Math., 212(1) (2016), 443–472.
[14] R. Meshulam. Spaces of Hankel Matrices over Finite Fields. Linear Algebra Appl., 218 (1995), 73–76.
[15] M. B. Milinovich and C. L. Turnage-Butterbaugh. Moments of Products of Automorphic L-functions.
J. Number Theory, 139 (2014), 175–204.
[16] O. Rojo. A New Algebra of Toeplitz-plus-Hankel Matrices and Applications. Comput. Math. with Appl.,
55(12) (2008), 2856–2869.
[17] E. C. Titchmarsh. The Theory of the Riemann Zeta-Function (Oxford Science Publications). Oxford
University Press, 2nd edition (1987). Revised by D. R. Heath-Brown.
[18] K.-C. Tong. On Divisor Problems III. Acta Math. Sin., 6 (1956), 515–541.
[19] K.-M. Tsang. Higher-power Moments of ∆(x), E(t) and P (x). Proc. London Math. Soc., 65(3) (1992),
65–84.