Skip to main content

Section 3 Linnik's theorem

We will mostly follow Ellenberg, Michel, Venkatesh. Linnik's ergodic method and the distribution of integer points on spheres.

Subsection 3.1 Linnik's theorem

Problem

Distribution of

\begin{equation*} \mathscr H_d = \{ (x,y,z) \in \ZZ^3 : x^2 + y^2 + z^2 = d\} \end{equation*}

as \(d \to \infty\text{.}\)

The \(\pm 1\pmod 5\) condition comes from a splitting condition in \(\QQ(\sqrt{d})\) that we need to find an appropriate group action.

We will rather focus on the following variant

Actually we will prove something a little strange.

Definition 3.3 Deviation

\(\infty\text{,}\) let \(\rho \gt 0\text{,}\) \(x\in S^2\) and consider the cap \(\Omega(x,\rho)\text{.}\)

\begin{equation*} \operatorname{dev}_d(\Omega) = \frac {1}{\operatorname{Vol}(\Omega)} \frac{| \operatorname{red}\inv_\infty(\Omega)|}{|\mathscr H_d|} - 1 \end{equation*}

(we want to show this is \(o(1)\)).

\(q\text{,}\) let \(\bar x \in \mathscr H_d(q)\) then

\begin{equation*} \operatorname{dev}_d(\bar x) = \frac {| \mathscr H_d(q)| | \operatorname{red}\inv_q(\bar x)|}{|\mathscr H_d|} - 1\text{.} \end{equation*}

How? Sketch of the argument:

Fix \(\delta ,\eta \gt 0\text{,}\)

\begin{equation*} B_\delta = \{ \bar x \in \mathscr H_q(d) : \operatorname{dev}_d(\bar x) \gt \delta \} \end{equation*}

and assume that

\begin{equation*} |B_\delta| \ge \eta |\mathscr H_d(q)|\text{.} \end{equation*}

Inputs

  1. There is an action of \(H(d) = \) class group of \(\QQ(\sqrt{-d})\) on \(\mathscr H_d(q)/ \specialorthogonal_3(\ZZ)\text{.}\) As \(d \equiv \pm 1 \pmod 5\) so \((5) = \ideal p \bar{\ideal p}\) in \(\QQ(\sqrt{-d})\text{:}\) This gives us a well defined dynamical system by considering the action
    \begin{equation*} \mathscr H_d(q) \ni x \xrightarrow {\ideal p} x_1 \to x_2 \to x_2 \to \cdots \end{equation*}
    and in reverse
    \begin{equation*} x_{-2} \to x_{-1} \to x \xrightarrow {\ideal p} x_1 \to x_2 \to x_2 \to \cdots \end{equation*}
    call this chain \(\gamma_X\) and for any \(l \gt 0 \in \ZZ\)
    \begin{equation*} \gamma_X^{(l)} = [x_{-l}, x_{-l + 1}, \ldots, x_{-1} , x,x_1, \ldots, x_l] \end{equation*}
  2. Consider the average on
    \begin{equation*} B_\delta = \gamma_X^{(l)} \cap B_\delta\text{.} \end{equation*}
    On one hand the expected size of \(\gamma_X^{(l)} \cap B_\delta\) is
    \begin{equation*} \frac{1}{|\mathscr H_d|} \sum_{x \in \mathscr H_d} | \gamma_X^{(l)} \cap B_\delta| \end{equation*}
    on the other hand the expectation is
    \begin{equation*} \frac{(2l+ 1)}{|\mathscr H_d|} \operatorname{red}_q\inv (B_\delta)\text{.} \end{equation*}
    i.e.
    \begin{equation*} \frac{1}{|\mathscr H_d|} \sum_{x\in \mathscr H_d} \frac{|\gamma_X^{(l)} \cap B_\delta|}{2l + 1} = \frac{| \operatorname{red}\inv_q(B_\delta)|}{|\mathscr H_d|} \end{equation*}
  3. Since we assume that the deviation set is large (\(\operatorname{dev}_d(x) \gt \delta\)) this implies on average
    \begin{equation*} \frac{| \gamma_X^{(l)} \cap B_\delta|}{2l+1} \end{equation*}
    is large, gives a lower bound for this.
  4. Then we count \(\gamma_X^{(l)}\) in another way, action of \(H(d)\) on \(\mathscr H_d\) and on \(\mathscr H_d(q)\text{.}\) action of \(\ideal p\) induces a graph structure on \(\mathscr H_d(q)\text{.}\) Then \(\gamma_X^{(l)}\) is a path on \(\mathscr H_d(q)\) which is non-backtracking. Count these: Inputs: for \(l \sim \log(d)\) the number of non-backtracking paths for which \(|\gamma_X^{(l)} \cap B_\delta|/(2l+1)\) is large is \(\gg d^{1/2 + \epsilon}\text{.}\) Count the total number of non-backtracking paths of length \(2l+1\) for which
    \begin{equation*} \left| \frac{|\gamma_X^{(l)} \cap B_\delta|}{2l+1 } - \frac{|B_\delta|}{|\mathscr H_d(q)|} \right| \gt \delta\eta/2 \end{equation*}
    There can not be too many of these. This will give a contradiction. Will take some work and depend on many things.
Class group actions
\begin{equation*} \mathscr H_d = \{ (x,y,z ) \in \ZZ^2: x^2 + y^2+ z^2 = d \} \end{equation*}

Aim:

  1. \(\widetilde {\mathscr H}_d^*\) is a principal homogeneous space for \(H(d) = \Cl(\QQ(\sqrt{-d}))\text{.}\)
  2. Describe the action explicitly.
Definition 3.5 Principal homogeneous spaces

Let \(G\) be a group, \(X \ne \emptyset\) a set. \(X\) is a principal homogeneous space for \(G\) if

\begin{equation*} G\acts X \end{equation*}

in a transitive, free manner. \(X\) is also called a \(G\)-torsor.

Remark 3.6

Of \(G\) is non-abelian we have notions of a left-torsor and right-torsor.

Example 3.7
  • \(G\) is a \(G\)-torsor
  • \(G= V\) a \(n\)-dimensional vector space then \(\aff^n\) is a \(G\)-torsor
  • \(G = \mu_n\) then \(X(2) = \{ x \in \CC: x^n = 2\}\)
Note 3.8
\begin{equation*} \specialorthogonal_3 (\ZZ)\acts \mathscr H_d \end{equation*}
\begin{equation*} v = \begin{pmatrix} x \\ y \\ z \end{pmatrix}\in \mathscr H_d,\,\gamma \in \specialorthogonal_3 (\ZZ) \leadsto \gamma v \in \mathscr H_d\text{.} \end{equation*}
  1. \begin{equation*} \# \specialorthogonal_3(\ZZ) = 24 \end{equation*}
  2. \(\specialorthogonal_3(\ZZ)\) has a subgroup \(\specialorthogonal_3^+(\ZZ)\) of index 2 acting via even permutations.
Definition 3.10
\begin{equation*} \widetilde{\mathscr H}_d = \specialorthogonal_3(\ZZ) \backslash \mathscr H_d \end{equation*}
\begin{equation*} \widetilde{\mathscr H}_d (q) = \specialorthogonal_3(\ZZ) \backslash \mathscr H_d(q) \end{equation*}
\begin{equation*} \widetilde S^2 = \specialorthogonal_3 (\ZZ) \backslash S^2 \end{equation*}
\begin{equation*} \widetilde{\mathscr H}_d^* = \begin{cases} \specialorthogonal_3^+(\ZZ) \backslash \mathscr H_d ,\amp d \equiv 1,2 \pmod 4 \\ \specialorthogonal_3(\ZZ) \backslash \mathscr H_d ,\amp d \equiv 3 \pmod 4 \end{cases} \end{equation*}

For all practical purposes we can basically assume that \(d \equiv 3 \pmod 4\text{.}\)

Digression: Hamilton Quaternions
\begin{equation*} B = \{ u + ia + jb + kc : u,a,b,c \in \RR\} \end{equation*}
\begin{equation*} \bar x = u - ia - j b - kc \end{equation*}
\begin{equation*} N(x) = x\bar x = u^2 + a^2 + b^2 + c^2 \end{equation*}
\begin{equation*} \trace (x) = x + \bar x = 2u \end{equation*}
\begin{equation*} B^{(0)} = \{ x\in B : \trace (x) = 0\} \end{equation*}

Note

\begin{equation*} N(x) =a^2 + b^2 + c^2 \text{ if } x \in B^{(0)}\text{.} \end{equation*}
\begin{equation*} B^\times = \text{units} ,\, B^1 = \{ x \in B : N(x)= 1\} \end{equation*}
\begin{equation*} PB^\times = B^\times/ Z(B^\times) \end{equation*}
Remark 3.12
\begin{equation*} B^\times \acts B^{(0)} \end{equation*}

via conjugation.

\begin{equation*} 1 \to Z(B^\times) \to B^\times \to PB^\times \simeq \specialorthogonal_3 (\QQ) \to 1 \end{equation*}
Definition 3.14 Hurwitz quaternions
\begin{equation*} B(\ZZ) = \ZZ[ i ,j, \frac{i+j+ k}{2} ] \end{equation*}

The Lipschitz quaternions are

\begin{equation*} \ZZ[ i ,j, k]\text{.} \end{equation*}
  1. \(B(\ZZ)\) is a maximal order
  2. \(B(\ZZ)\) is a Euclidean domain
  3. Lipschitz quaternions are not a Euclidean domain
  1. Show these properties.
  2. Show \(\Cl(\ZZ\lb i, j, k \rb) =2\text{.}\)

2 implies \(B(\ZZ)\) is a PID

\begin{equation*} (\QQ^3, a^2+ b^2+ c^2) \simeq (B^{(0)}, N) \end{equation*}

as metric spaces

The image of

\begin{equation*} B^\times(\ZZ) \to \specialorthogonal_3(\ZZ) \end{equation*}

is \(\specialorthogonal_3^+(\ZZ)\text{.}\)

The action of \(H(d)\)

Start with \(x = (a,b,c) \in \mathscr H_d\text{.}\) This gives rise to an embedding

\begin{equation*} \iota_x \colon \QQ(\sqrt{-d}) \hookrightarrow B \end{equation*}
\begin{equation*} \sqrt{-d} \mapsto x + ai + bj + ck \end{equation*}

note \(x^2 = -d\text{.}\)

Let \(K= \QQ(\sqrt{-d})\) \(\iota_x\) is integral ion the sense that if

\begin{equation*} \ints_x = B(\ZZ) \cap \QQ \lb x \rb \end{equation*}

then

\begin{equation*} \iota\inv (\ints_x) = \ints_K\text{.} \end{equation*}

Prove this

\(I \in H(d) \leadsto y= y(x,I)\)

Consider \(B(\ZZ) \iota_x(I) = B(\ZZ) q\) for some \(q\in B(\QQ)\text{.}\) Since \(B(\ZZ)\) is a PID.

Then \(y = q xq\inv\in \mathscr H_d\)

Note 3.17
  1. \begin{equation*} y^2 = -d \end{equation*}
  2. If we take \(\lambda I\) for \(\lambda \in K^\times\text{,}\) \(y(x, \lambda I) = y(x, I)\) so we get a well defined construction \(H(d)\text{.}\)
  3. \(q\) is defined up to conjugation by \(B(\ZZ)^\times\text{.}\)

To sum up:

For any \(x \in \mathscr H_d\) we have a well defined map

\begin{equation*} \Cl(d) \to \widetilde{\mathscr H}_d^\times \end{equation*}
\begin{equation*} I \mapsto y(x,I)\text{.} \end{equation*}

Let \(x,y \in \mathscr H_d\text{.}\)

Construct an ideal \(\Lambda_{x\mapsto y}\) s.t. \(y(x,\Lambda_{x\mapsto y}) = y\text{.}\)

Definition 3.19
\begin{equation*} \Lambda_{x\mapsto y} = \{ \lambda \in B(\ZZ) : x \lambda = \lambda y \}\text{.} \end{equation*}

\(x, y \in \mathscr H_d\) Witt's theorem implies there exists \(q \in PB^\times(\QQ) = \specialorthogonal_3(\QQ)\) s.t.

  • \begin{equation*} y = q\inv x q \end{equation*}
  • \begin{equation*} \Lambda_{x\mapsto y} = B(\ZZ) \cap \QQ [x] q \end{equation*}
Note 3.20

\(\Lambda_{x\mapsto y} \) is locally free and rank 1

\begin{equation*} \implies [\Lambda_{x \mapsto y}] \in \Pic(\ints_x) \simeq \Pic(\ints_K) \simeq \Cl(d)\text{.} \end{equation*}

Summary \(\widetilde {\mathscr H_d}^*\) is a principal homogeneous space for \(\Cl(d)\text{,}\) the class group of \(\QQ(\sqrt{-d})\text{.}\)

How? \(x\in\mathscr H_d\) defined

\begin{equation*} \QQ(\sqrt{-d}) \xhookrightarrow i B(\QQ) \end{equation*}
\begin{equation*} \sqrt{-d} \mapsto x \end{equation*}

any \(I \in \Cl(d) \leadsto B(\ZZ) i(I) = B(\ZZ) q\) for some \(q\text{.}\)

\begin{equation*} B(\ZZ) = \{ x + ai +bj + ck : a,b,c,d \in \ZZ, \text{ or all }\in \frac 12\ZZ\smallsetminus \ZZ\} \end{equation*}
\begin{equation*} y= q\inv x q \end{equation*}
\begin{equation*} I x= y \end{equation*}
\begin{equation*} [ \Lambda_{x\mapsto y}] \end{equation*}

inverse of this construction

\begin{equation*} \specialorthogonal(3, \QQ) \simeq PB^\times (\QQ) \end{equation*}

Requires checking everything works as intended, it is local.

Explicit realization
\begin{equation*} P \in \Cl(d) \end{equation*}
\begin{equation*} \leadsto M_P \in \specialorthogonal(3) \end{equation*}
\begin{equation*} x\in \mathscr H_d \implies Px = M_P x \end{equation*}
\begin{equation*} x = \begin{pmatrix} a\\b\\c \end{pmatrix} \end{equation*}

Let

\begin{equation*} P \bar P = (5) \end{equation*}

because of the assumption on \(d\text{.}\)

\(P\) action, let \(x\in\mathscr H_d\text{.}\)

\begin{equation*} N(P) = 5\leadsto q \in B(\ZZ) \end{equation*}
\begin{equation*} N(q) = 5 \implies 5 \in \{1\pm2i, 1\pm2j,1\pm2k\}B^\times(\ZZ) \end{equation*}

two of these will give you the action of \(P, \bar P\text{.}\)

Example 3.22

\(q\) will act by \(q\inv x q\) acting on \(B^{(0)}\text{,}\) enough to check the action on \(i,j,k\text{.}\) Say

\begin{equation*} q = 1+2i,\bar q = (1-2i)/5 \end{equation*}
\begin{equation*} \bar q j q = \frac{(1-2i)}{5} j (1+2i) \end{equation*}
\begin{equation*} = \frac{j- 2k}{5} 1+2i \end{equation*}
\begin{equation*} j-2k - 2k - 2j???????? \end{equation*}
Basic construction

Let \(\mathscr A_5 = \{M_1^\pm, M_2^\pm, M_3^\pm\}\) starting with a solution \(x\in \mathscr H_d\) one gets two matrices \(\omega_1, \omega_{-1} \in\mathscr A_5\)

\begin{equation*} x_{-3} \xleftarrow{\omega_{-3}} x_{-2} \xleftarrow{\omega_{-2}} x_{-1} \xleftarrow{\omega_{-1}} x \xrightarrow{\omega_{1}}x_1 \xrightarrow{\omega_{2}}x_2 \xrightarrow{\omega_{3}} x_3 \end{equation*}

where \(w_1\) comes from \(P\) and \(w_{-1}\) from \(\bar P\text{.}\)

This construction gives a well-defined (up to flip) path starting with a solution \(x \in \mathscr H_d\text{.}\)

Remark 3.23

These chains are periodic.

Back to proof

Idea: if \(x\in\mathscr H_d\) then \(x\) and these paths are well defined mod \(p\text{.}\)

We are aiming for

Next time, Linnik's lemma \(\#\{xy = d\} \le d^\epsilon\text{.}\) One more input.

In \(\QQ(\sqrt d)\) we have \((5) = \ideal p \bar{ \ideal p}\)

\begin{equation*} x \in \mathscr H_d = \{(x_1,x_2,x_3) : x_1^2 + x_2^2 + x_3^2 = d\} \end{equation*}

there are 6 matrices \(\mathcal A = \{A^\pm, B^\pm, C^\pm\}\text{,}\) \(A,B,C \in \GL_2(\ZZ)\text{,}\) reduction of these \(\pmod q \acts \mathscr H_d(q)\text{.}\)

This induces a graph structure on \(\mathscr H_d(q)\text{,}\) (multiple edges between vertices are allowed).

\begin{equation*} \mathscr H_d \to \mathscr H_d(q) \end{equation*}
\begin{equation*} [\ldots, x_{-1}, x_0, x_1, \ldots] \to \text{path in graph} \end{equation*}

for \(l \in \NN\text{,}\)

\begin{equation*} \gamma_x^{(l)} \subseteq \mathscr H_d(q)\text{.} \end{equation*}
Linnik's basic lemma

Generalisation of the following

\begin{equation*} r(d) = \sum_{a| d} 1 \le d^\epsilon\,\forall \epsilon\text{.} \end{equation*}

Short detour on representations of quadratic forms by other quadratic forms.

Definition 3.26

Let \((Q, \ZZ^m)\text{,}\) \((R, \ZZ^n)\text{,}\) \(m\ge n\) be nondegenerate quadratic forms. Then \(Q\) is said to represent \(R\) if there exists a \(\ZZ\)-linear map

\begin{equation*} \iota\colon \ZZ^n \to \ZZ^m \end{equation*}

s.t.

\begin{equation*} Q(\iota(x)) = R(x)\,\forall x\in \ZZ^n\text{.} \end{equation*}
Example 3.27
\begin{equation*} Q(x,y) = x \cdot y \end{equation*}
\begin{equation*} R(x) = d x^2 \end{equation*}

then let

\begin{equation*} \iota\colon \ZZ \to \ZZ^2 \end{equation*}
\begin{equation*} x \mapsto (ax,bx) \text{ where } ab =d \end{equation*}

then

\begin{equation*} Q(\iota(x)) = R(x)\text{.} \end{equation*}

Let \(x_1, x_2 \in \mathscr H_d\) s.t. \(x_1\cdot x_2 =e\text{.}\)

\begin{equation*} x_1=(a_1,b_1,c_1)\, x_2=(a_2,b_2,c_2) \end{equation*}

consider

\begin{equation*} \iota \colon \ZZ^2 \to \ZZ^3 \end{equation*}
\begin{equation*} (u,v) \mapsto (u a_1 + v a_2, u b_1 + u b_2, u c_1 + v c_2) = u x_1 + v x_2 \end{equation*}

let \(Q(a,b,c) = a^2 + b^2 + c^2\text{.}\)

\begin{equation*} Q(\iota(u,v)) = (u a_1 + v a_1)^2 + (u b_1 + v b_2)^2 + (u c_1 + v c_2)^2 \end{equation*}
\begin{equation*} = du^2 + 2euv + dv^2 \end{equation*}
\begin{equation*} = R_{d,e} (u,v)\text{.} \end{equation*}

So the number of representations of \(R\) by \(Q/ \specialorthogonal_3(\ZZ)\) bounds number of solutions \(x_1 x_2 = e\text{.}\)

Show this (recall \(d\) is a fundamental discriminant).

\begin{equation*} \gamma_x^{(l)} = \gamma_{x'}^{(l)} \implies x = \pm x' \pmod{q5^l} \text{(by the Shadowing lemma)} \end{equation*}
\begin{equation*} \implies (x+x') ( x+ x') \equiv 0 \pmod{q^2 5^{2l}} \end{equation*}

or

\begin{equation*} \implies (x-x') ( x- x') \equiv 0 \pmod{q^2 5^{2l}}\text{.} \end{equation*}
\begin{equation*} \implies \Sigma(d,l,q) \ll 2\#\{(x,x') : x\ne x',\, x\cdot x' = d \pmod{q^2 5^{2l}}\} \end{equation*}

recall

\begin{equation*} \sum_{|e| \lt d,\,e = d \pmod{q^2 5^{2l}} } \{(x,x') : x\cdot x' = e\} \ll_\epsilon d^\epsilon (1+ \frac{d}{q^2 5^{2l}})\text{.} \end{equation*}
Proof of the equidistribution statement

Recall:

\begin{equation*} \operatorname{dev}_d(\bar x) = \frac { | \operatorname{red}\inv_q(\bar x)|}{|\mathscr H_d|/| \mathscr H_d(q)|} - 1 \end{equation*}
\begin{equation*} \operatorname{dev}_d(\bar x) \to 0\text{ as } d \to \infty,\,d \equiv \pm 1 \pmod 5\,\, d\text{ fund. disc.}\text{.} \end{equation*}

Assume \(B_\delta = \{ \bar x : \operatorname{dev}_d)\bar x) \\gt \delta\}\text{.}\) Lower bound by counting the number of paths that lie in \(\operatorname{red}\inv_q(B_\delta)\text{.}\) Upper bound by the expansion property of \(\mathscr H_d(q)\text{.}\)

Recall 3.3.

Last time

Let

\begin{equation*} B_\delta = \{ \bar x\in \mathscr H_d (q) : \operatorname{dev}_d(\bar x) \gt \delta\} \end{equation*}

and assume that

\begin{equation*} B_\delta \ge \eta | \mathscr H_d(q)| \end{equation*}

for some \(\eta \gt 0\text{.}\)

Observe that

\begin{equation*} \frac{1}{|\mathscr H_d(q) | } \sum_{x\in \mathscr H_d } |\gamma_x^{(l)} \cap B_\delta | = (2l+ 1) \frac { | \operatorname{red}_q\inv (B_\delta)|}{|\mathscr H_d|} \end{equation*}

or

\begin{equation} \frac{1}{|\mathscr H_d(q) | } \sum_{x\in \mathscr H_d } \frac{ |\gamma_x^{(l)} \cap B_\delta | }{ (2l+ 1)} = \frac { | \operatorname{red}_q\inv (B_\delta)|}{|\mathscr H_d|}\tag{3.1} \end{equation}

Choose an \(l\) s.t.

\begin{equation*} \frac 15 |\mathscr H_d| \le q^2 5^{2l} \le 5 |\mathscr H_d| \end{equation*}

note that we can indeed choose such an \(l\text{.}\)

By definition every \(\bar x \in B_\delta\) satisfies

\begin{equation*} \operatorname{dev}_d(\bar x) \gt \delta \end{equation*}

implies

\begin{equation*} \frac{ |\operatorname{red}_q\inv(\bar x)|}{ |\mathscr H_d|/|\mathscr H_d(q)|} \gt 1 + \delta\text{.} \end{equation*}
\begin{equation*} | \operatorname{red}_q \inv(B_\delta)| \ge \frac{ |B_\delta | |\mathscr H_d| }{ |\mathscr H_d(q)|} (1+\delta) \end{equation*}

so

\begin{equation*} \frac{1}{|\mathscr H_d|} \sum_{x\in \mathscr H_d} \frac{|\gamma_x^{(l)} \cap B_\delta|}{(2l+1)} \gt \frac{|B_\delta|}{|\mathscr H_d(q)|}(1+\delta) \end{equation*}

by the assumption that

\begin{equation*} B_\delta \ge \eta | \mathscr H_d(q)| \end{equation*}

we have this is

\begin{equation*} \ge \frac{|B_\delta|}{|\mathscr H_d(q)|}+\eta\delta\text{.} \end{equation*}

Let

\begin{equation*} \Phi = \left\{ x\in \mathscr H_d : \frac{ | \gamma_x^{(l)} \cap B_\delta | }{ (2l + 1)} \gt \frac{|B_\delta|}{|\mathscr H_d(q)|} + \frac{\eta \delta }{2} \right\} \end{equation*}

Note that

\begin{equation*} \mathscr H_d = \Phi \sqcup \Phi^c \end{equation*}
\begin{equation*} \frac{1}{|\mathscr H_d|} \sum_{x\in \mathscr H_d} \frac{ | \gamma_x^{(l)} \cap B_\delta | }{ (2l + 1)} = \frac{1}{|\mathscr H_d|} \left( \sum_{x\in \Phi } \underbrace{\frac{ | \gamma_x^{(l)} \cap B_\delta | }{ (2l + 1)}}_{\le 1} + \sum_{x\in \Phi^c} \underbrace{\frac{ | \gamma_x^{(l)} \cap B_\delta | }{ (2l + 1)}}_{ \le \frac{|B_\delta|}{\mathscr H_d(q)|} + \frac{\eta \delta }{2} } \right) \end{equation*}
\begin{equation*} \le\frac{1}{|\mathscr H_d|} \left( |\Phi | + \overbrace{| \Phi^c|}^{=|\mathscr H_d| - |\Phi|} \left( \frac{|B_\delta|}{\mathscr H_d(q)|} + \frac{\eta \delta }{2} \right)\right) \end{equation*}
\begin{equation} \implies |\Phi| \ge \frac{\eta \delta }{2} | \mathscr H_d|\text{.}\label{eqn-phi-bd-linnik-pf}\tag{3.2} \end{equation}

Now we count the number of marked non-backtracking paths on \(\mathscr H_d(q)\) of length \(2l+1\) (write \(MNBP(\mathscr H_d(q))\) for this quantity), which in addition satisfy

\begin{equation*} \frac{|\gamma_x^{(l)} \cap B_\delta|}{(2l+1)} \gt \frac{|B_\delta|}{|\mathscr H_d(q) | } + \frac{\delta \eta}{2}\text{.} \end{equation*}

By (3.2) we have \(\eta \delta |\mathscr H_d|/2\) of these paths “upstairs”.

Recall Siegel implies

\begin{equation*} d^{\frac 12 - \epsilon} \lt | \mathscr H_d| \lt d^{\frac 12 + \epsilon} \end{equation*}

so upstairs we have \(\gg_\epsilon \frac{\eta \delta }{2} d^{\frac 12 - \epsilon }\) paths.

Claim 1: The previous proposition implies not mant of these paths give the same path mod \(q\)

\begin{equation*} MNBP(\mathscr H_d(q)) \gg_{\epsilon,\delta \eta} d^{\frac 12 - \epsilon}\text{.} \end{equation*}

On the other hand the total number of marked non-backtracking paths on

\begin{equation*} \mathscr H_d(q) \end{equation*}

is

\begin{equation*} | \mathscr H_d(q)| 6 \cdot 5^{2l -1} \end{equation*}

also observe that \(|\mathscr H_d(q)| \ll q^2\) so the total number \(\sim q^2 5^{2l} \ll d^{1/2 + \epsilon}\)

\begin{equation*} \frac 15 |\mathscr H_d| \lt q^2 5^{2l} \lt 5 |\mathscr H_d| \end{equation*}

Final input is a Chernoff type bound.

Claim 2: This proposition implies that the fraction

\begin{equation*} MNBP(\mathscr H_d(q)) \ll d^{\frac 12 - \tau} \end{equation*}

for some \(\tau \gt 0\text{.}\) Then

\begin{equation*} \gg_\epsilon d^{\frac 12 - \epsilon} \end{equation*}
\begin{equation*} \ll_\epsilon d^{\frac 12 - \tau} \end{equation*}

give a contradiction.

We now check some claims made above.

\begin{equation*} \#\{ (a,a') \in A^2 : a \equiv a' \pmod q\} = \sum n_{a_i} \end{equation*}

gives a lower bound on \(A \pmod q\text{.}\) To show this note that the

\begin{equation*} | A(q)| = | A | - \sum_{i=1}^{|A|} \frac{n_{a_i} - 1}{n_{a_i}} - \sum_{i=1}^{|A|} \frac 1{n_{a_i}} \end{equation*}

where \(n_{a_i} = \# \{ a \in A : a \equiv a_i \pmod q\}\text{.}\)

Subsection 3.2 The adelic picture

Notation: \(\adeles\) is the ring of adeles of \(\QQ\)

\begin{equation*} \prod_p' \QQ_p \end{equation*}
\begin{equation*} \adeles_f = \prod_{p \text{ finite}}' \QQ_p \end{equation*}

so that

\begin{equation*} \adeles = \RR \times \adeles_f\text{.} \end{equation*}
\begin{equation*} \widetilde \ZZ \colon \varprojlim_n \ZZ/n = \prod \ZZ_p\text{.} \end{equation*}
\begin{equation*} \widehat \ZZ \subseteq \adeles_f \end{equation*}

maximal compact.

The genus of a quadratic space

For us,

\begin{equation*} L\subseteq \QQ^3 \end{equation*}

a lattice, then

\begin{equation*} \GL_3(\adeles_f) \ni g = \prod g_p \end{equation*}
\begin{equation*} g\cdot L = \{ \alpha \subseteq \QQ^3 : \alpha_p = g_p L_p \,\forall p\} \end{equation*}
\begin{equation*} \operatorname{genus}_{\specialorthogonal_3} (\mathcal L_p) = \specialorthogonal_3(\adeles_f) \mathcal L_0 \end{equation*}
\begin{equation*} \mathcal L_0 = \ZZ^3,\,a^2+b^2 + c^2\text{.} \end{equation*}

Based on quaternions having class number 1.

More definitions:

\begin{equation*} P = \{ (\mathcal L, x) : \mathcal L \in \operatorname{genus}_{\specialorthogonal_3} (\mathcal L_0), x \in \mathcal L,\, x \cdot x = d\} \end{equation*}
\begin{equation*} P_{(q)} = \{ (\mathcal L, \bar x) : \mathcal L \in \operatorname{genus}_{\specialorthogonal_3} (\mathcal L_0), \bar x \in \mathcal L/q\mathcal L,\, \bar x \cdot \bar x = d \pmod q\} \end{equation*}
\begin{equation*} \widetilde{\mathscr H_d} \simeq \specialorthogonal_3 (\QQ) \backslash P \end{equation*}
\begin{equation*} K_f [q] = \ker(\specialorthogonal_3(\widehat \ZZ) \to \specialorthogonal_3(\ZZ/q\ZZ)) \end{equation*}
\begin{equation*} K'_f [q] = \{ g \in \specialorthogonal_3(\widehat \ZZ) : g \bar x_q = \bar x_q\} \end{equation*}

we fix a base point \(\bar x_q \in \mathscr H_d(q)\text{.}\) Then

\begin{equation*} \widetilde{\mathscr H_d}(q) \simeq \specialorthogonal_3 (\QQ) \backslash P_{(q)} \simeq \specialorthogonal_3 (\QQ) \backslash \specialorthogonal_3(\adeles_f) / K_f'[q] \end{equation*}

here we are heavily using the exceptional isomorphism

\begin{equation*} \specialorthogonal_3 \simeq \PGL_2\text{.} \end{equation*}
\begin{equation*} \specialorthogonal_3(\QQ) \acts P_{(q)} \text{ diagonally} \end{equation*}
\begin{equation*} \specialorthogonal_3(\adeles_f) = \specialorthogonal_3(\QQ) \specialorthogonal_3(\widehat \ZZ) \end{equation*}

every orbit has a representation \(\simeq (\mathcal L_0, \bar x),\,\bar x \in \mathscr H_d(q)\text{.}\)

\begin{equation*} (\mathcal L_0, \bar x) \simeq (\mathcal L_0, \bar x') \end{equation*}
\begin{equation*} \iff \end{equation*}
\begin{equation*} \exists \gamma \in \underbrace{\specialorthogonal_3(\widehat \ZZ) \cap \specialorthogonal_3(\QQ)}_{\specialorthogonal_3(\ZZ)} \end{equation*}

Observe

\begin{equation*} \specialorthogonal_3(\adeles_f) \acts P_{(q)} \end{equation*}

via

\begin{equation*} \mathcal L \otimes \widehat \ZZ/ q\mathcal L \otimes \widehat \ZZ \end{equation*}

This action is transitive, i.e.

\begin{equation*} P_{(q)} = \specialorthogonal_3(\adeles_f) [ \mathcal L_0, \bar x_q] \end{equation*}
Remark 3.40
\begin{equation*} \specialorthogonal_3(\ZZ) \backslash S^3 \simeq \specialorthogonal_3(\QQ)\backslash \specialorthogonal_3(\adeles)/ K_\infty \specialorthogonal_3(\widehat \ZZ) \end{equation*}

What about

\begin{equation*} \specialorthogonal_3(\ZZ) \backslash \mathscr H_d\text ? \end{equation*}

Fix another base point \(x_0\in \mathscr H_d\)

\begin{equation*} \specialorthogonal_3(\ZZ) \backslash \mathscr H_d \simeq \specialorthogonal_{x_0} (\QQ) \backslash \specialorthogonal_{x_0} (\adeles_f)/ \specialorthogonal_{x_0}(\widehat \ZZ) \end{equation*}

where \(\specialorthogonal_{x_0}\) is the stabiliser of \(x_0\text{.}\)

\begin{equation*} \xymatrix{ \mathscr H_d \acts \text{ class gp.} \ar[d] \\ \mathscr H_d(a) \acts \specialorthogonal_3(\adeles_f) } \end{equation*}

More concretely

\begin{equation*} \specialorthogonal_{x_0} : \QQ^3 \simeq B^{(0)} (\QQ) \end{equation*}
\begin{equation*} \specialorthogonal_3 \simeq PB^\times \end{equation*}
\begin{equation*} \specialorthogonal_{x_0} \simeq PB^\times_{x_0} \simeq \langle a + b x_0 \rangle / Z \end{equation*}
\begin{equation*} \simeq \Res_{K/\QQ} \mathbf G_m / \mathbf G_m \end{equation*}

where \(K = \QQ(\sqrt{-d})\text{.}\)

Recall

\begin{equation*} K^\times \backslash \adeles_{K,f}^\times / \widehat \ints_K^\times \simeq \Pic(\ints_K)\text{.} \end{equation*}
The graph structure on \(\mathscr H_d(q)\)

\(\specialorthogonal_3(\adeles_f)\) is a large group, but so is \(K_f' \lb q\rb \text{.}\) What is indeed true is

\begin{equation*} \mathscr H_d(q) \cong \Gamma_{(3,q)} \specialorthogonal_3(\QQ_5) / K_5 \end{equation*}

analogous to

\begin{equation*} \mathbf H = \specialorthogonal_2 \backslash \SL_2(\RR)/ \SL_2(\ZZ)\text{.} \end{equation*}

Where

\begin{equation*} \Gamma_{(3,q)} PB^\times(\QQ) \cap PB^\times(\QQ_5) K_f[3,q] \end{equation*}
\begin{equation*} K_{f} [3,q] = K_f[3] \cap K_f'[q] \subseteq K_f'[q] \end{equation*}

Now

\begin{equation*} \PGL_2(\QQ_5) / \PGL_2(\ZZ_5) \end{equation*}

is a Bruhat-Tits tree, a 6-regular tree.

Automorphic spectrum of \(\PGL_2\) of \(\mathbf F_q(t)\)

Analogous to \(\SL_2(\ZZ) \backslash \HH\text{.}\)

Reference I. Efrat: Automorphic spectrum on the tree of \(\PGL_2\)

Setup: \(k = \FF_q\text{,}\) \(v_\infty = - \deg\) so \(|f|_\infty = q^{\deg (f)}\text{.}\)

\begin{equation*} K = (k(t))_\infty = k((\frac 1t )) \end{equation*}
\begin{equation*} \ints_\infty = \ints = k [[ t \inv ]] \end{equation*}
\begin{equation*} \Gamma = \PGL_2(\FF_q[[t]]) \end{equation*}

Analogies:

\begin{equation*} \HH \leftrightsquigarrow \PGL_2(K) / \PGL_2(\ints) = X \end{equation*}
\begin{equation*} \SL_2(\ZZ) \leftrightsquigarrow \Gamma \end{equation*}

Problem: decompose \(L^2(\Gamma \backslash X)\text{.}\) In the case of

\begin{equation*} \SL_2(\ZZ) \backslash \HH \leadsto L^2(\SL_2(\ZZ)\backslash \HH) = \underbrace{R}_{\text{Residual}} \oplus \underbrace{C}_{\text{cusp forms}} \oplus \underbrace{E}_{\text{Eisenstein}}\text{.} \end{equation*}
Little bit on \(X\) (a \((q+1)\)-regular tree)

Ref: Serre “Trees”.

Description

  1. In terms of lattices

    \begin{equation*} L = \{ av_1 + bv_2 : a,b\in \ints\} \end{equation*}

    s.t.

    \begin{equation*} L\otimes_\ints K \simeq K^2\text{.} \end{equation*}

    Matrix

    \begin{equation*} \begin{pmatrix} v_1 \amp v_2 \end{pmatrix} \GL_2(K) \end{equation*}

    changing basis implies \(L\) a class in \(\GL_2(K) / \GL_2(\ints)\text{.}\)

    \begin{equation*} L_1 \sim L_2 \end{equation*}

    if \(\exists a \in K^\times\) s.t. \(L_1 = a L_2\text{.}\)

    \begin{equation*} X = \{ \text{equivalence classes of lattices} \} \end{equation*}

    A class \(L\) is adjacent to \(L'\) \(\iff\) \(\exists\) representations \(L_0, L_0'\) s.t. \(L_0' \subseteq L_0\) and \(L_0/L_0' \simeq \FF_q\text{.}\)

  2. Group theoretic

    \begin{equation*} B = \left\{ \begin{pmatrix} * \amp * \\ 0 \amp 1 \end{pmatrix} \right\}/K \end{equation*}

    Fact: \(\PGL_2(K) = B \cdot \PGL_2(\ints)\text{.}\) Each coset \(b \PGL_2(\ints)\) has a representative

    \begin{equation*} \begin{pmatrix} t^n \amp x \\ 0 \amp 1\end{pmatrix} \end{equation*}

    \(n\) unique and \(x \in K\text{.}\)

    Vertices of \(X\)

    \begin{equation*} \begin{pmatrix} t^n \amp x \\ 0 \amp 1\end{pmatrix} \PGL_2(\ints) \end{equation*}
    \begin{equation*} \begin{pmatrix} t^n \amp x \\ 0 \amp 1\end{pmatrix} \sim_{\text{adjacent}} \begin{cases} \begin{pmatrix} t^{n-1} \amp \xi t^n + x \\ 0 \amp 1\end{pmatrix},\, \xi \in \FF_q \begin{pmatrix} t^{n+1} \amp x \\ 0 \amp 1\end{pmatrix} \end{cases} \end{equation*}

    Fact:

    \begin{equation*} \Gamma\backslash X = \text{chain } \bullet -\bullet - \bullet - \cdots \end{equation*}

Measures: Normalise Haar on \(\PGL_2\) s.t. \(\mu(\PGL_2(\ints)) = q(q-1)\text{.}\) \(\leadsto\) On \(\Gamma\backslash X\text{,}\) \(\mu\) gives vertices weights

\begin{equation*} n=0: \frac{1}{q+1} \end{equation*}
\begin{equation*} n=1: \frac{1}{q} \end{equation*}
\begin{equation*} n=2: \frac{1}{q^2} \end{equation*}
\begin{equation*} \vdots \end{equation*}
\begin{equation*} n: \frac{1}{q^n} \end{equation*}
Remark 3.42

We have

\begin{equation*} \pair fg_\mu = \int_{\Gamma \backslash X} f \cdot \bar g \diff \mu = \frac{1}{q+1} f(0)\bar g(0) + \sum_{i=1}^\infty \frac{f(i) \bar g(i)}{q^i}\text{.} \end{equation*}
Remark 3.44
\begin{equation*} \Delta = (q+1) \id - T \end{equation*}

and \(T\) descends to \(\Gamma\backslash X\text{.}\)

Given \(f\colon \Gamma\backslash X \simeq \NN \to \CC\)

\begin{equation*} (T(f))(n) = \begin{cases} (q+1)f(1),\amp n=0 \\ qf(n-1) + f(n+1),\amp n \gt 0 \end{cases} \end{equation*}

\(T\) is self-adjoint with respect to \(\pair{}{}_\mu\text{.}\)

Eigenfunctions of \(T\)
\begin{equation*} (T f)(n) = \lambda \cdot f(n),\,\forall n \in \NN\text{.} \end{equation*}
\begin{equation*} \implies \lambda f(0) = (q+1) f(1) \end{equation*}
\begin{equation*} \lambda f(n) = q f(n-1) + f(n+1),\,\forall n \gt 0 \end{equation*}

i.e.

\begin{equation*} \begin{pmatrix} f(n+1) \\ f(n) \end{pmatrix} = \begin{pmatrix} \lambda \amp -q \\ 1 \amp 0 \end{pmatrix}^n\begin{pmatrix} f(1) \\ f(0) \end{pmatrix}\text{.} \end{equation*}

Now fix a normalization

\begin{equation*} f(0) = q+1 \end{equation*}
\begin{equation*} f(1) = \lambda\text{.} \end{equation*}

Characteristic values:

\begin{equation*} x_1, x_2 = \frac{-\lambda \pm \sqrt{\lambda^2 - 4q}}{2} \end{equation*}

note \(x_1 = x_2 \iff \lambda = \pm 2\sqrt q\text{.}\)

Eigenfunctions

\begin{equation*} f_\lambda (n) = \begin{cases} \frac{1}{\sqrt{\lambda^2 - 4q}} \left(\lambda( x_1^n - x_2^n) - q(q+1)(x_1^{n-1} - x_2^{n-1})\right),\amp n \gt0 \\ q+1,\amp n=0 \end{cases} \end{equation*}

for \(\lambda \ne \pm 2\sqrt q\) note \(f_{q+1} = q+1\text{.}\)

\begin{equation*} f_{2\sqrt q} (n) = (q+1 - (q-1) n)q^{n/2} \end{equation*}
\begin{equation*} f_{-2\sqrt q} (n) = (-1)^n f_{2\sqrt q} \end{equation*}

We now consider the space \(L^2( \Gamma\backslash X)\) with the operator \(T\text{.}\)

Question: which the \(f_\lambda\) are in \(L^2 (\Gamma\backslash X)\text{?}\)

Normalise \(f_\lambda\) by

\begin{equation*} (x_2 - x_1) f = x_1^{n-1} (\lambda x_1 - q(q+1)) - x_2^{n-1} (\lambda x_2 - q(q+1)) \end{equation*}

assume WLOG \(|x_1| \gt \sqrt q\) \(\implies\) \(|x_2| \lt \sqrt q\) (since \(|\lambda| \gt 2 \sqrt q\)) so \(x_2^n = o(q^{n/2})\text{.}\)

So \(\lambda x_1 - q(q+1) = 0\) (to be in \(L^2\)). So \(\lambda = \pm q(q+1)\text{.}\)

What about the rest?

Note: \(|\lambda| \lt 2 \sqrt q \implies x_1 = \bar x_2\text{.}\)

Renormalise: Let \(x_1 = \sqrt q e^{i\theta},x_2 = \sqrt q e^{-i\theta}\text{.}\)

\begin{equation*} \lambda = 2\sqrt q \cos \theta,\,0 \lt \theta \lt \pi \end{equation*}
\begin{equation*} \tilde f_\theta(n) = \frac{x_1 - x_2}{2\sqrt q} f_{2\sqrt q \cos \theta} (n) \end{equation*}

so

\begin{equation*} \tilde f_\theta(n) = \begin{cases} q^{n/2} i (\sin(n+1)) \theta - q\sin(n-1) \theta\\ (q+1) i \sin \theta\end{cases} \end{equation*}
Spectral decomposition of \(L^2(\Gamma \backslash X)\)

Continuous spectrum: Extend \(\tilde f_\theta\) to an odd function \(- \pi \lt \theta \lt \pi\text{.}\) Given \(\psi \in L^2(\Lambda\backslash X)\) consider

\begin{equation*} F_\psi(n) = \frac{1}{2\pi} \int_{-\pi}^\pi \psi(\theta) \tilde f_\theta (n) \diff \theta\text{.} \end{equation*}