Kuznecov formulae for fractal measures
Extracted source text · arXiv 2512.18379
\documentclass[11pt]{amsart}
\usepackage{amsmath,amssymb,amsthm,mathtools}
\usepackage{mathrsfs}
\usepackage{hyperref}
\usepackage[margin=1.2in]{geometry}
\numberwithin{equation}{section}
% Theorem environments
\newtheorem{theorem}{Theorem}[section]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{remark}[theorem]{Remark}
% Shortcuts
\newcommand{\R}{\mathbb{R}}
\newcommand{\N}{\mathbb{N}}
\newcommand{\Z}{\mathbb{Z}}
\newcommand{\C}{\mathbb{C}}
\newcommand{\supp}{\operatorname{supp}}
\newcommand{\vol}{\operatorname{vol}}
\newcommand{\diam}{\operatorname{diam}}
\newcommand{\dist}{d}
\newcommand{\eps}{\varepsilon}
\newcommand{\sphere}{\mathbb{S}}
\newcommand{\ball}{\mathbb{B}}
\DeclareMathOperator{\Res}{Res}
\begin{document}
\title[Kuznecov formulae]{Kuznecov formulae for fractal measures}
\author{Yakun Xi}
\address{School of Mathematical Sciences, Zhejiang University, Hangzhou 310027, China}
\email{yakunxi@zju.edu.cn}
\begin{abstract}
Let $(M,g)$ be a compact, connected Riemannian manifold of dimension $n\ge 2$, and let
$\{e_j\}_{j=0}^\infty$ be an orthonormal basis of Laplace eigenfunctions
$-\Delta_g e_j=\lambda_j^2 e_j$. Given a finite Borel measure $\mu$ on $M$, consider the
Kuznecov sum
\[
N_\mu(\lambda):=\sum_{\lambda_j\le \lambda}\Bigl|\int_M e_j\,d\mu\Bigr|^2.
\]
Assume that $\mu$ is $s$-Ahlfors regular for some $s\in(0,n)$ and admits an averaged
$s$-density constant $A_\mu$. We prove that
\[
N_\mu(\lambda)
= (2\pi)^{-(n-s)}\,\vol(B^{\,n-s})\,A_\mu\,\lambda^{n-s}
+ o(\lambda^{n-s})
\qquad (\lambda\to\infty).
\]
The hypotheses of $s$-Ahlfors regularity and the averaged $s$-density condition are
essentially optimal for such a one-term asymptotic, and in general the remainder
$o(\lambda^{n-s})$ cannot be improved uniformly to a power-saving error term.
This extends the classical Kuznecov formulae of Zelditch \cite{Zelditch92} for smooth
submanifold measures to a broad class of singular and fractal measures.
\end{abstract}
\maketitle
\section{Introduction}
Let $(M,g)$ be a compact, connected, smooth Riemannian manifold of dimension
$n\ge 2$, and let
\[
-\Delta_g e_j = \lambda_j^2 e_j,\qquad
0 = \lambda_0 <\lambda_1 \le \lambda_2 \le \cdots,
\]
denote a complete orthonormal basis of real Laplace eigenfunctions. Given a
finite Borel measure $\mu$ on $M$, we consider the associated Kuznecov sum
\begin{equation}\label{eq:Kuz-sum}
N_\mu(\lambda) := \sum_{\lambda_j\le\lambda} \Big|\int_M e_j\,d\mu\Big|^2,
\qquad \lambda\ge 1.
\end{equation}
A basic and well-studied case is when $\mu$ is the Riemannian volume measure
on an embedded submanifold. Let $H\subset M$ be an embedded submanifold of
dimension $k$ with induced Riemannian metric and volume measure $dV_H$. In
his seminal work \cite{Zelditch92}, Zelditch proved a Kuznecov sum formula
for these coefficients, showing that
\begin{equation}\label{eq:Zelditch}
\sum_{\lambda_j\le\lambda}
\biggl|\int_H e_j\,dV_H\biggr|^2
= (2\pi)^{-(n-k)} \,\vol(B^{n-k})\,\vol(H)\,\lambda^{\,n-k}
+ O(\lambda^{n-k-1}),
\end{equation}
where $B^{n-k}\subset\R^{n-k}$ is the Euclidean unit ball and
$\vol(H)$ denotes the $k$–dimensional Riemannian volume of $H$. Thus the
leading growth of the Kuznecov sum is governed by the ambient dimension $n$,
the submanifold dimension $k$, and an explicit geometric factor.
More recently, Wyman and the author \cite{WymanXiKuz} refined Zelditch's formula by
obtaining a two-term asymptotic for \eqref{eq:Zelditch}, identifying the
next-order oscillatory term. In these works the measure is always smooth and
supported on a submanifold. On the other hand, there has been growing recent interest
in the behavior of eigenfunctions restricted to general Borel sets. For example,
Eswarathasan and Pramanik \cite{EswarathasanPramanik22} initiated a program to study
the $L^p$ norms of eigenfunctions with respect to a fractal measure $\mu$, connecting
classical results of Sogge \cite{sogge1988concerning} and Burq--G\'erard--Tzvetkov
\cite{BurqGerardTzvetkov07}. An essentially sharp version of such bounds in two
dimensions was recently obtained by Gao, Miao, and the author \cite{GaoMiaoXi24}.
The goal of this paper is to extend Zelditch's Kuznecov formulae to a
much broader class of measures, including singular and fractal measures, while
retaining the same leading term.
Throughout the paper we use the following convention for Euclidean ball
volumes. For $m>0$ (not necessarily an integer), we write $\vol(B^m)$ for
the “$m$–dimensional volume’’ of the Euclidean unit ball in formal dimension
$m$, defined by
\[
\vol(B^m)
:= \frac{\pi^{m/2}}{\Gamma\bigl(\frac{m}{2}+1\bigr)}.
\]
When $m\in\N$, this coincides with the usual Lebesgue $m$–volume of the unit
ball in $\R^m$. For an embedded $k$–dimensional submanifold $H\subset M$ we
write $\vol(H)$ for its $k$–dimensional Riemannian volume.
We now describe the small-scale geometric hypotheses on $\mu$ that will play
the role of “dimension’’ and “density’’ in our setting. Throughout,
$B(x,r)$ denotes the geodesic ball of radius $r$ centered at $x\in M$ with
respect to the Riemannian distance $d_g$.
\begin{definition}[$s$-Ahlfors regular measure]
Let $s\in(0,n)$. A finite Borel measure $\mu$ on $M$ is called
\emph{$s$-Ahlfors regular} if there exist constants $c,C>0$ and $r_0>0$
such that
\begin{equation}\label{eq:Ahlfors}
c\,r^s \le \mu(B(x,r)) \le C\,r^s,
\qquad x\in\supp\mu,\ 0<r\le r_0.
\end{equation}
\end{definition}
This condition is standard in geometric measure theory and implies that the
Hausdorff dimension of $\mu$ is $s$. See, for example, \cite{MattilaFAHD}. It
is satisfied by Riemannian volume measures on $s$-dimensional embedded
submanifolds, as well as by many self-similar and self-conformal measures on
fractals.
To obtain a one-term asymptotic for \eqref{eq:Kuz-sum}, one needs a stabilization
of the \emph{averaged} small-scale behavior of $\mu$. For $r\ge 0$ define the
distance distribution
\begin{equation}\label{eq:distance-distribution}
F(r)
:=(\mu\times\mu)\bigl(\{(x,y)\in M\times M:\ d_g(x,y)\le r\}\bigr)
=\int_M \mu\bigl(B(x,r)\bigr)\,d\mu(x).
\end{equation}
\begin{definition}[Averaged $s$-density]\label{def:DD}
Let $s\in(0,n)$. We say that $\mu$ \emph{admits an averaged $s$-density} if there
exists a constant $A_\mu\in(0,\infty)$ (called the \emph{averaged $s$-density
constant} of $\mu$) such that
\begin{equation}\label{eq:DD}
F(r)=\vol(B^s)\,A_\mu\,r^s+o(r^s)\qquad (r\to0^+).
\end{equation}
Equivalently,
\begin{equation}\label{eq:A-mu}
A_\mu
=\lim_{r\to0^+}\frac{F(r)}{\vol(B^s)\,r^s}
=\lim_{r\to0^+}\frac{1}{\vol(B^s)\,r^s}\int_M \mu(B(x,r))\,d\mu(x).
\end{equation}
\end{definition}
We can now state our main result. It shows that under the Ahlfors regularity
assumption and the averaged $s$-density condition \eqref{eq:DD}, the Kuznecov sum
\eqref{eq:Kuz-sum} has a one-term asymptotic with exponent $n-s$ and an explicit
leading constant depending only on $n$ and $s$.
\begin{theorem}\label{thm:main}
Let $(M,g)$ be a compact, connected Riemannian manifold of dimension $n\ge 2$.
Fix $s\in(0,n)$ and let $\mu$ be a finite Borel measure on $M$ which is
$s$-Ahlfors regular in the sense of \eqref{eq:Ahlfors} and admits an averaged
$s$-density constant $A_\mu$ in the sense of \eqref{eq:DD}. Then the Kuznecov sum
\eqref{eq:Kuz-sum} satisfies
\begin{equation}\label{eq:Kuz-main}
N_\mu(\lambda)
=C_{n,s}\,A_\mu\,\lambda^{\,n-s}
+ o(\lambda^{\,n-s})
\qquad (\lambda\to\infty),
\end{equation}
where
\[
C_{n,s}=(2\pi)^{-(n-s)}\,\vol(B^{\,n-s}).
\]
\end{theorem}
The proof combines a heat-kernel regularization of the Kuznecov sum with a
Tauberian argument, with the leading term determined by the small-scale
distance distribution of $\mu$.
\begin{remark}\label{rem:density-implies-DD}
The averaged $s$-density condition \eqref{eq:DD} is automatically satisfied under
the stronger pointwise density hypothesis. Namely, suppose that $\mu$ is
$s$-Ahlfors regular and admits a pointwise $s$-density $\theta$ in the sense that
\[
\theta(x)=\lim_{r\to0^+}\frac{\mu(B(x,r))}{\vol(B^s)\,r^s}
\]
exists for $\mu$-a.e.\ $x$. Then by the Ahlfors upper bound
$\mu(B(x,r))/(\vol(B^s)r^s)\lesssim 1$ and dominated convergence,
\[
\frac{F(r)}{\vol(B^s)\,r^s}
=\int_M \frac{\mu(B(x,r))}{\vol(B^s)\,r^s}\,d\mu(x)
\longrightarrow \int_M \theta(x)\,d\mu(x),
\qquad (r\to0^+).
\]
Hence \eqref{eq:DD} holds with
\[
A_\mu=\int_M \theta\,d\mu.
\]
In particular, \eqref{eq:DD} is strictly weaker than the existence of a pointwise
$s$-density.
\end{remark}
The leading exponent in \eqref{eq:Kuz-main} is $n-s$, determined by the ambient
dimension $n$ and the Hausdorff dimension $s$ of $\mu$, and the leading
coefficient factors as a universal constant times the averaged $s$-density
constant $A_\mu$. In the special case where $\mu=dV_H$ is the Riemannian volume
measure on an embedded submanifold $H$ of dimension $s$, $\mu$ admits a pointwise
$s$-density $\theta$ and hence satisfies the averaged $s$-density condition
\eqref{eq:DD} by Remark~\ref{rem:density-implies-DD}. In fact $\theta\equiv 1$
and $A_\mu=\vol(H)$, so \eqref{eq:Kuz-main} reduces to
\[
\sum_{\lambda_j\le\lambda}
\biggl|\int_H e_j\,dV_H\biggr|^2
= (2\pi)^{-(n-s)}\,\vol(B^{\,n-s})\,\vol(H)\,\lambda^{\,n-s}
+ o(\lambda^{\,n-s}),
\]
whose leading term agrees with Zelditch's Kuznecov formulae \eqref{eq:Zelditch}
with $k=s$. Thus Theorem~\ref{thm:main} extends the classical Kuznecov formulae
to a large class of singular and fractal measures. To the best of our knowledge, Theorem~\ref{thm:main} is the first Kuznecov-type asymptotic for fractal measures.
Moreover, as discussed in \cite{wyman2023can}, the left-hand side of Zelditch’s
Kuznecov formula \eqref{eq:Zelditch} can be viewed as ``audible'' spectral
information associated with the submanifold $H$, while the right-hand side
encodes geometric information about $H$, namely its dimension $k$ and its
$k$-dimensional Riemannian volume $\vol(H)$. Theorem~\ref{thm:main} extends this
audible--geometric correspondence from smooth submanifold measures to
$s$-Ahlfors regular measures admitting an averaged $s$-density constant, so that
the Hausdorff dimension $s$ and the effective ``volume'' $A_\mu$ of the measure
are also audible through the leading asymptotics of $N_\mu(\lambda)$.
\smallskip
For $0<u<n$ let
\begin{equation}\label{eq:Riesz-energy}
I_u(\mu) := \iint d_g(x,y)^{-u}\,d\mu(x)\,d\mu(y)
\end{equation}
denote the Riesz energy of $\mu$ with respect to the Riemannian distance
$d_g(x,y)$ on $M$. There is an earlier work of Hare and Roginskaya
\cite{HareRoginskaya03} which is closely related in spirit
to our results. They established energy formulae on compact Riemannian
manifolds which relate $I_u(\mu)$ to weighted sums of the projections of $\mu$
onto Laplace eigenspaces. In Section~\ref{sec:HR} we show that their
identities imply the following general upper bound without assuming the existence of an averaged $s$-density constant.
\begin{theorem}\label{thm:HR-bound}
Let $(M,g)$ be a compact Riemannian manifold of dimension $n\ge2$, and let
$\mu$ be a finite Borel measure on $M$. Fix $u\in(0,n)$ and suppose that the
Riesz energy $I_u(\mu)$ defined in \eqref{eq:Riesz-energy} is finite.
Then there exists a constant $C_1=C_1(M,g,u,\mu)$ such that
\[
\sum_{\lambda_j\le\lambda} \Bigl|\int_M e_j\,d\mu\Bigr|^2
\le C_1\,\lambda^{\,n-u}
\qquad (\lambda\ge1).
\]
In particular, if $\mu$ is $s$-Ahlfors regular for some $s\in(0,n)$, then
$I_u(\mu)<\infty$ for every $u<s$, and hence for each $\varepsilon>0$ there exists
a constant $C_2=C_2(M,g,s,\varepsilon,\mu)$ such that
\[
\sum_{\lambda_j\le\lambda} \Bigl|\int_M e_j\,d\mu\Bigr|^2
\le C_2\,\lambda^{\,n-s+\varepsilon}
\qquad (\lambda\ge1).
\]
\end{theorem}
Thus, for measures with finite Riesz energy, one also has nontrivial
polynomial upper bounds on the Kuznecov sum. However, in the formulae of
Hare--Roginskaya the exponent $u$ is fixed, and the constants appearing in
Theorem~\ref{thm:HR-bound} are not controlled uniformly as $u\to s^-$. Since
our proof of Theorem~\ref{thm:main} requires a precise description of the
singular behavior of $I_u(\mu)$ at $u=s$ in order to recover the leading
constant in \eqref{eq:Kuz-main}, one cannot simply invoke their identities
and pass to the limit $u\to s$.
\subsection*{Sharpness}
We point out that the assumptions in Theorem~\ref{thm:main} and
Theorem~\ref{thm:HR-bound} are essentially optimal. In
Section~\ref{sec:sharpness} we give examples showing, on the one hand, that one
cannot dispense with the $s$-Ahlfors regularity assumption if one wishes to
control the exponent purely in terms of the Hausdorff dimension of $\mu$, and on
the other hand, that some stabilization of the small-scale distance
distribution—such as the averaged $s$-density condition \eqref{eq:DD}—is needed
to obtain a genuine one-term asymptotic. Under $s$-Ahlfors regularity alone one
can at best expect upper and lower bounds of order $\lambda^{n-s}$, but not the
existence of a limit for $N_\mu(\lambda)/\lambda^{n-s}$.
In addition to these structural restrictions, we show in Section~\ref{sec:sharpness} that the remainder term in Theorem~\ref{thm:main} cannot, in general, be improved to a uniform power saving over the class of measures under consideration. More precisely, we construct explicit examples showing that there is no exponent $\delta>0$ such that for every $s$-Ahlfors regular measure with an averaged $s$-density constant $A_\mu$ one has
\[
N_\mu(\lambda)
= C_{n,s}A_\mu\,\lambda^{n-s} + O\bigl(\lambda^{n-s-\delta}\bigr),
\]
with an implicit constant depending only on $(M,g,n,s,\mu)$. The $o(\lambda^{n-s})$ remainder in \eqref{eq:Kuz-main} is therefore essentially sharp at this level of generality.
\subsection*{Organization}
In Section~\ref{sec:heat} we introduce a heat-kernel regularization of the Kuznecov
sum, and establish its small-time asymptotics in terms of the averaged
$s$-density constant $A_\mu$. Section~\ref{sec:Tauberian} interprets this
smoothed Kuznecov sum as a Laplace--Stieltjes transform and applies a Tauberian
theorem to prove Theorem~\ref{thm:main} and identify the constant $C_{n,s}$.
In Section~\ref{sec:HR} we revisit the energy identities of Hare and Roginskaya
to derive general polynomial upper bounds for $N_\mu(\lambda)$, proving
Theorem~\ref{thm:HR-bound}. Finally, Section~\ref{sec:sharpness} is devoted to
sharpness: we give examples showing that the Ahlfors regularity and averaged
$s$-density hypotheses in Theorem~\ref{thm:main} are essentially optimal, and
that the $o(\lambda^{n-s})$ remainder cannot, in general, be improved to a
uniform power-saving error term.
\subsection*{Acknowledgement} This project is supported by the National Key R\&D Program of China under Grant No. 2022YFA1007200, the Natural Science Foundation of China under Grant No. 12571107, and the Zhejiang Provincial Natural Science Foundation of China under Grant No. LR25A010001. The author thanks Xucheng Hu and Diankun Liu for carefully reading an earlier draft of this paper.
\subsection*{Notation}
We write $A\lesssim B$ to mean that there exists an implicit constant $C>0$ such that
\[
A\le C\,B.
\]
We write
$A\gtrsim B$ if $B\lesssim A$, and $A\approx B$ if both $A\lesssim B$ and $B\lesssim A$.
For nonnegative functions $f,g$ and $x_0\in\mathbb R\cup\{\infty\}$ we write
\[
f(x)\sim g(x)\qquad (x\to x_0)
\]
to mean
\[
\lim_{x\to x_0}\frac{f(x)}{g(x)}=1.
\]
\section{Heat kernel and a smoothed Kuznecov sum}\label{sec:heat}
In this section, we introduce a heat-kernel-regularized Kuznecov sum and
identify its small-time behavior in terms of the averaged $s$-density constant $A_\mu$.
\subsection{Heat kernel asymptotics near the diagonal}
Throughout we write $p_t(x,y)$ for the heat kernel on $(M,g)$:
\[
(\partial_t - \Delta_g)p_t(x,y) = 0,\qquad
\lim_{t\to0^+}p_t(\,\cdot\,,y) = \delta_y,
\]
so that for every $f\in C^\infty(M)$,
\[
e^{t\Delta_g}f(x)
= \int_M p_t(x,y)f(y)\,dV(y)
= \sum_{j=0}^\infty e^{-t\lambda_j^2}\langle f,e_j\rangle e_j(x).
\]
The small-time behavior of $p_t$ near the diagonal is classical. See, for
example, \cite[Ch.~2]{BGV92} or \cite[Ch.~8]{Grigoryan09}.
\begin{lemma}\label{lem:heat-local-expansion}
There exist $r_0>0$, $t_0>0$, and smooth functions $u_k(x,y)$ on
$\{(x,y)\in M\times M: d_g(x,y)< 2r_0\}$, $k=0,1,2,\dots$, with
$u_0(x,x)\equiv1$, such that for every $N\in\N$ we have
\begin{equation}\label{eq:heat-local-expansion}
p_t(x,y)
= (4\pi t)^{-n/2} e^{-d_g(x,y)^2/(4t)}
\biggl(\sum_{k=0}^{N-1} t^k u_k(x,y)\biggr)
+ R_N(t,x,y),
\end{equation}
and
\[
|R_N(t,x,y)|
\le C_N\,t^{N-n/2}
\]
for all $0<t\le t_0$ and all $x,y\in M$ with $d_g(x,y)<r_0$.
Moreover, there exist constants $c,C>0$ such that
\begin{equation}\label{eq:heat-off-diagonal}
0\le p_t(x,y) \le C\,t^{-n/2}\exp\Bigl(-\frac{d_g(x,y)^2}{ct}\Bigr),
\qquad t\in(0,1],\ x,y\in M.
\end{equation}
\end{lemma}
Since $u_0$ is smooth and $u_0(x,x)=1$, there exists $C>0$ such that
\begin{equation}\label{eq:u0-Lipschitz}
|u_0(x,y)-1|\le C\,d_g(x,y)
\end{equation}
whenever $d_g(x,y)<r_0$.
\subsection{A smoothed Kuznecov sum}
Define the heat-regularized Kuznecov sum associated with $\mu$ by
\begin{equation}\label{eq:heat-Kuz}
H_\mu(t)
:= \sum_{j=0}^\infty e^{-t\lambda_j^2}
\Bigl|\int_M e_j\,d\mu\Bigr|^2,
\qquad t>0.
\end{equation}
We first relate $H_\mu(t)$ to the heat kernel.
\begin{lemma}\label{lem:heat-spectral}
For every $t>0$,
\begin{equation}\label{eq:Hmu-heat-kernel}
H_\mu(t)
= \iint_{M\times M} p_t(x,y)\,d\mu(x)\,d\mu(y).
\end{equation}
\end{lemma}
\begin{proof}
Let
\[
\alpha_j := \int_M e_j\,d\mu,
\qquad j=0,1,2,\dots.
\]
Recall that the heat kernel satisfies
\[
p_t(x,y)
= \sum_{j=0}^\infty e^{-t\lambda_j^2}e_j(x)e_j(y).
\]
For each $J\in\mathbb N$ set
\[
p_t^{(J)}(x,y)
:= \sum_{j=0}^J e^{-t\lambda_j^2}e_j(x)e_j(y),
\qquad
S_J(t) := \sum_{j=0}^J e^{-t\lambda_j^2}|\alpha_j|^2.
\]
Since $e^{-t\lambda_j^2}\ge0$ and $|\alpha_j|^2\ge0$, the sequence $S_J(t)$ is
nondecreasing in $J$.
For each fixed $t>0$, the series converges absolutely and uniformly on $M\times M$, since $\sup_{x,y}|e_j(x)e_j(y)|\le \|e_j\|_\infty^2\lesssim \lambda_j^{n-1}$ by H\"ormander \cite{Hormander1968SpectralFunction} and $\sum_{j\ge0} e^{-t\lambda_j^2}\lambda_j^{n-1}<\infty$ by Weyl law.
In particular,
\[
\sup_{x,y\in M}\bigl|p_t(x,y)-p_t^{(J)}(x,y)\bigr|\longrightarrow 0
\qquad(J\to\infty),
\]
and $p_t$ is continuous and bounded on $M\times M$.
For each $J$ we have
\[
\begin{split}
\iint_{M\times M} p_t^{(J)}(x,y)\,d\mu(x)\,d\mu(y)
&= \iint_{M\times M}
\sum_{j=0}^J e^{-t\lambda_j^2}e_j(x)e_j(y)\,d\mu(x)\,d\mu(y) \\
&= \sum_{j=0}^J e^{-t\lambda_j^2}
\biggl(\int_M e_j(x)\,d\mu(x)\biggr)
\biggl(\int_M e_j(y)\,d\mu(y)\biggr) \\
&= S_J(t),
\end{split}
\]
where we have interchanged the finite sum and the integrals.
Since $p_t^{(J)}\to p_t$ uniformly and $\mu$ is finite, we have
\[
\lim_{J\to\infty}
\iint_{M\times M} p_t^{(J)}(x,y)\,d\mu(x)\,d\mu(y)
= \iint_{M\times M} p_t(x,y)\,d\mu(x)\,d\mu(y).
\]
Therefore $S_J(t)$ converges as $J\to\infty$, and
\[
\sum_{j=0}^\infty e^{-t\lambda_j^2}|\alpha_j|^2
= \lim_{J\to\infty} S_J(t)
= \iint_{M\times M} p_t(x,y)\,d\mu(x)\,d\mu(y),
\]
which is precisely \eqref{eq:Hmu-heat-kernel}.
\end{proof}
\subsection{A Gaussian average}
Using the distance distribution $F(r)$ introduced above, we consider the
Gaussian average
\[
G(t)
:= \iint e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y).
\]
Let $R(x,y):=d_g(x,y)$ and let $\nu:=R_*(\mu\times\mu)$ be the pushforward
of $\mu\times\mu$ under $R$. Then $\nu$ is a finite Borel measure on
$[0,\diam(M)]$ and
\[
F(r) = (\mu\times\mu)\bigl(\{(x,y): d_g(x,y)\le r\}\bigr)
= \nu([0,r]),
\]
so $F$ is the distribution function of $\nu$. In particular, for the
nonnegative Borel function $\varphi(r):=e^{-r^2/(4t)}$ we have
\[
G(t)
= \int_{M\times M} \varphi(R(x,y))\,d\mu(x)\,d\mu(y)
= \int_{[0,\diam(M)]} \varphi(r)\,d\nu(r)
= \int_0^{\diam(M)} e^{-r^2/(4t)}\,dF(r).
\]
The next lemma identifies the behavior of $G(t)$ as $t\to0^+$.
\begin{lemma}\label{lem:Gaussian-F}
Assume that $\mu$ satisfies \eqref{eq:Ahlfors} and \eqref{eq:DD}.
Then, as $t\to0^+$,
\begin{equation}\label{eq:G-asympt}
G(t)
= \gamma_s\,A_\mu\,t^{s/2} + o(t^{s/2}),
\end{equation}
where
\begin{equation}\label{eq:gamma-s}
\gamma_s := s\,2^{\,s-1}\Gamma\Bigl(\frac{s}{2}\Bigr)\,\vol(B^s).
\end{equation}
In particular, there exists $C>0$ and $t_1>0$ such that
\begin{equation}\label{eq:G-upper}
G(t)\le C\,t^{s/2}\qquad(0<t\le t_1).
\end{equation}
\end{lemma}
\begin{proof}
Write, by our assumption on $\mu$,
\[
F(r) = \vol(B^s)A_\mu r^s + \mathcal E(r),
\]
with $\mathcal E(r)=o(r^s)$ as $r\to0^+$. Fix $r_0>0$ as in
\eqref{eq:Ahlfors}. Split
\[
G(t)
= \int_0^{r_0} e^{-r^2/(4t)}\,dF(r)
+ \int_{r_0}^{\diam(M)} e^{-r^2/(4t)}\,dF(r)
=: G_0(t)+G_1(t).
\]
Since $F$ is bounded on $[r_0,\diam(M)]$, we have
\[
|G_1(t)| \le F(\diam(M))\,e^{-r_0^2/(4t)} = o(t^N)
\]
for every $N>0$ as $t\to0^+$. In particular $G_1(t)=o(t^{s/2})$.
For $G_0(t)$, decompose
\[
G_0(t)
= \vol(B^s)A_\mu\int_0^{r_0} e^{-r^2/(4t)}\,d(r^s)
+ \int_0^{r_0} e^{-r^2/(4t)}\,d\mathcal E(r)
=: J_1(t)+J_2(t).
\]
For $J_1(t)$ we compute explicitly:
\[
J_1(t)
= \vol(B^s)A_\mu s\int_0^{r_0} r^{s-1}e^{-r^2/(4t)}\,dr
= \vol(B^s)A_\mu s\int_0^\infty r^{s-1}e^{-r^2/(4t)}\,dr + O(e^{-r_0^2/(4t)}).
\]
The error term is $o(t^{s/2})$ as above. Using the standard integral
\[
\int_0^\infty r^{s-1}e^{-r^2/(4t)}\,dr
= 2^{\,s-1}\Gamma\Bigl(\frac{s}{2}\Bigr)t^{s/2},
\]
we obtain
\[
J_1(t)
= \gamma_s\,A_\mu\,t^{s/2} + o(t^{s/2})
\]
with $\gamma_s$ as in \eqref{eq:gamma-s}.
For $J_2(t)$ we integrate by parts on $[0,r_0]$:
\[
J_2(t)
= e^{-r^2/(4t)}\mathcal E(r)\Big|_{0}^{r_0}
+ \int_0^{r_0}\frac{r}{2t}e^{-r^2/(4t)}\mathcal E(r)\,dr
=: B(t)+I(t).
\]
Since $F(0)=0$ and hence $\mathcal E(0)=0$, we have
$B(t)=e^{-r_0^2/(4t)}\mathcal E(r_0)=o(t^{s/2})$.
To handle $I(t)$, fix $\eps>0$ and choose $\delta\in(0,r_0)$ so small that
$|\mathcal E(r)|\le\eps r^s$ for $0<r\le\delta$. Split
\[
I(t)
= \int_0^{\delta}\frac{r}{2t}e^{-r^2/(4t)}\mathcal E(r)\,dr
+ \int_{\delta}^{r_0}\frac{r}{2t}e^{-r^2/(4t)}\mathcal E(r)\,dr
=: I_1(t)+I_2(t).
\]
On $(0,\delta]$ we have $|\mathcal E(r)|\le\eps r^s$, so
\[
|I_1(t)|
\le \frac{\eps}{2t}\int_0^{\delta} r^{s+1}e^{-r^2/(4t)}\,dr
\le \frac{\eps}{2t}\int_0^{\infty} r^{s+1}e^{-r^2/(4t)}\,dr
= C_1\,\eps\,t^{s/2}
\]
for some $C_1$ depending only on $s$.
On $[\delta,r_0]$ we use the upper Ahlfors bound
$F(r)\le C r^s$ to deduce that $|\mathcal E(r)|\le C' r^s$ on this interval.
Thus
\[
|I_2(t)|
\le \frac{C'}{2t}\int_{\delta}^{r_0} r^{s+1}e^{-r^2/(4t)}\,dr
\le C'' t^{-1}e^{-\delta^2/(4t)},
\]
which is $o(t^{s/2})$ as $t\to0^+$. Hence $I(t)=o(t^{s/2})$, and since
$\eps>0$ was arbitrary, we also have $J_2(t)=B(t)+I(t)=o(t^{s/2})$.
Combining the estimates for $J_1(t)$ and $J_2(t)$ gives
\eqref{eq:G-asympt}. The upper bound \eqref{eq:G-upper} follows from
\eqref{eq:G-asympt} by taking $t$ sufficiently small and then enlarging the
constant $C$ to cover all $t\in(0,t_1]$.
\end{proof}
\subsection{Small-time asymptotics of the smoothed Kuznecov sum}
We can now identify the leading behavior of $H_\mu(t)$ as $t\to0^+$.
\begin{lemma}\label{lem:Hmu-asympt}
Assume that $\mu$ satisfies \eqref{eq:Ahlfors} and \eqref{eq:DD}. Then
\begin{equation}\label{eq:Hmu-asympt}
H_\mu(t)
= \gamma_{n,s}\,A_\mu\,t^{-(n-s)/2}
+ o\bigl(t^{-(n-s)/2}\bigr)
\qquad (t\to0^+),
\end{equation}
where
\begin{equation}\label{eq:gamma-ns}
\gamma_{n,s}
:= (4\pi)^{-n/2}\,\gamma_s
= (4\pi)^{-n/2}\,
s\,2^{\,s-1}\Gamma\Bigl(\frac{s}{2}\Bigr)\,\vol(B^s).
\end{equation}
\end{lemma}
\begin{proof}
By Lemma~\ref{lem:heat-spectral},
\[
H_\mu(t)
= \iint_{M\times M} p_t(x,y)\,d\mu(x)\,d\mu(y).
\]
Fix $r_0>0$ as in Lemma~\ref{lem:heat-local-expansion} and split
\[
H_\mu(t)
= \iint_{d_g(x,y)\le r_0} p_t(x,y)\,d\mu(x)\,d\mu(y)
+ \iint_{d_g(x,y)>r_0} p_t(x,y)\,d\mu(x)\,d\mu(y)
=: H_0(t)+H_1(t).
\]
By the Gaussian upper bound \eqref{eq:heat-off-diagonal},
\[
0\le H_1(t)
\le C\,t^{-n/2}\exp\Bigl(-\frac{r_0^2}{ct}\Bigr)\mu(M)^2
= o\bigl(t^{-(n-s)/2}\bigr),
\]
since $t^{(n-s)/2}e^{-r_0^2/(ct)}\to0$ as $t\to0^+$.
For $H_0(t)$ we apply \eqref{eq:heat-local-expansion} with an
integer $N\ge1$ to be chosen momentarily. For $d_g(x,y)\le r_0$ and
$0<t\le t_0$ we have
\[
p_t(x,y)
= (4\pi t)^{-n/2} e^{-d_g(x,y)^2/(4t)}
\biggl(\sum_{k=0}^{N-1} t^k u_k(x,y)\biggr)
+ R_N(t,x,y),
\]
with $|R_N(t,x,y)|\le C_N t^{N-n/2}$. Thus
\[
H_0(t)
= (4\pi t)^{-n/2}
\sum_{k=0}^{N-1} t^k
\iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}u_k(x,y)\,d\mu(x)\,d\mu(y)
+ E_N(t),
\]
where
\[
|E_N(t)|
\le C_N t^{N-n/2}\mu(M)^2.
\]
We first single out the $k=0$ term. Noting
\eqref{eq:u0-Lipschitz}, we may set $\omega:=u_0-1$, and write
\[
\begin{split}
&\iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}u_0(x,y)\,d\mu(x)\,d\mu(y) \\
&= \iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y)
+ \iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}\omega(x,y)\,d\mu(x)\,d\mu(y).
\end{split}
\]
The first integral differs from $G(t)$ by a tail integral:
\[
\begin{split}
\iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y)
&= G(t) - \iint_{d_g(x,y)>r_0} e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y) \\
&= G(t) + T(t),
\end{split}
\]
where
\[
|T(t)|
\le \mu(M)^2 e^{-r_0^2/(4t)}
= o(t^{s/2})
\]
as $t\to0^+$.
For the second integral we use \eqref{eq:u0-Lipschitz} and the Ahlfors bound
via $F(r)$:
\[
\begin{split}
\biggl|\iint_{d_g(x,y)\le r_0}
e^{-d_g(x,y)^2/(4t)}\omega(x,y)\,d\mu(x)\,d\mu(y)\biggr|
&\le C_2
\iint_{d_g(x,y)\le r_0} d_g(x,y)e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y) \\
&= C_2\int_0^{r_0} r e^{-r^2/(4t)}\,dF(r) \\
&=: C_2 J(t).
\end{split}
\]
Similar to the proof of Lemma~\ref{lem:Gaussian-F}, noting $F(0)=0$ and \eqref{eq:DD}, integration by parts gives
\[
|J(t)|
\le C r_0^{s+1}e^{-r_0^2/(4t)}
+ C\int_0^{r_0} e^{-r^2/(4t)} r^s\,dr
+ \frac{C}{t}\int_0^{r_0} e^{-r^2/(4t)} r^{s+2}\,dr.
\]
Extending the integrals to $\infty$ and changing variables $r=\sqrt{t}\,u$ gives
\[
\int_0^\infty e^{-r^2/(4t)} r^{a}\,dr
= t^{(a+1)/2}\int_0^\infty e^{-u^2/4}u^{a}\,du
\lesssim t^{(a+1)/2},
\]
so the last two terms are $O(t^{(s+1)/2})$ and $t^{-1}O(t^{(s+3)/2})=O(t^{(s+1)/2})$,
while the boundary term is $o(t^{(s+1)/2})$, and thus $J(t)\lesssim t^{(s+1)/2}$.
Hence
\[
\iint_{d_g(x,y)\le r_0}
e^{-d_g(x,y)^2/(4t)}\omega(x,y)\,d\mu(x)\,d\mu(y)
= O\bigl(t^{(s+1)/2}\bigr).
\]
Putting these estimates together, we obtain
\[
\iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}u_0(x,y)\,d\mu(x)\,d\mu(y)
= G(t) + O\bigl(t^{(s+1)/2}\bigr).
\]
Therefore
\[
(4\pi t)^{-n/2}
\iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}u_0(x,y)\,d\mu(x)\,d\mu(y)
= (4\pi t)^{-n/2}G(t) + E_0(t),
\]
where
\[
|E_0(t)|
\lesssim t^{-n/2} t^{(s+1)/2}
= t^{1/2-(n-s)/2}
= o\bigl(t^{-(n-s)/2}\bigr)
\]
as $t\to0^+$.
Next, for each $k=1,\dots,N-1$, the smoothness of $u_k$ and the upper bound
\eqref{eq:G-upper} yield
\[
\begin{split}
&\biggl|
(4\pi t)^{-n/2}t^k
\iint_{d_g(x,y)\le r_0} e^{-d_g(x,y)^2/(4t)}u_k(x,y)\,d\mu(x)\,d\mu(y)
\biggr|\\
&\qquad\le C_k t^{-n/2+k}
\iint e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y)
= C_k t^{-n/2+k}G(t)
\lesssim t^{-n/2+k} t^{s/2} \\
&\qquad= O\bigl(t^{k-(n-s)/2}\bigr)
= o\bigl(t^{-(n-s)/2}\bigr),
\end{split}
\]
since $k\ge1$.
Finally, for the remainder term we have
\[
|E_N(t)| \le C_N t^{N-n/2}\mu(M)^2.
\]
Choose an integer $N>\frac{s}{2}$. Then
\[
t^{(n-s)/2} |E_N(t)|
\lesssim t^{N-s/2}\longrightarrow 0
\qquad(t\to0^+),
\]
so $E_N(t)=o\bigl(t^{-(n-s)/2}\bigr)$.
Collecting the contributions from $k=0,\dots,N-1$, the remainder $E_N$, and
$H_1(t)$, we conclude that
\[
H_\mu(t)
= (4\pi t)^{-n/2}G(t) + o\bigl(t^{-(n-s)/2}\bigr)
\qquad (t\to0^+).
\]
Using Lemma~\ref{lem:Gaussian-F}, which gives
$G(t)=\gamma_s A_\mu t^{s/2}+o(t^{s/2})$, we obtain
\[
H_\mu(t)
= (4\pi)^{-n/2}\gamma_s A_\mu t^{-(n-s)/2}
+ o\bigl(t^{-(n-s)/2}\bigr),
\]
which is exactly \eqref{eq:Hmu-asympt} with $\gamma_{n,s}$ given by
\eqref{eq:gamma-ns}.
\end{proof}
\section{A Tauberian theorem and the proof of Theorem~\texorpdfstring{\ref{thm:main}}{main}}\label{sec:Tauberian}
We now view the smoothed Kuznecov sum $H_\mu(t)$ as a Laplace--Stieltjes
transform of a spectral counting function and apply a Tauberian theorem to
deduce the asymptotic behavior of $N_\mu(\lambda)$.
\subsection{Laplace--Stieltjes representation}
Define a nondecreasing, right-continuous function $\mathcal N:[0,\infty)\to
[0,\infty)$ by
\[
\mathcal N(S)
:= \sum_{\lambda_j^2\le S}
\Bigl|\int_M e_j\,d\mu\Bigr|^2,
\qquad S\ge0.
\]
Thus $\mathcal N$ puts mass $ |\int_M e_j\,d\mu|^2$ at the point $S=\lambda_j^2$.
By definition \eqref{eq:Kuz-sum},
\begin{equation}\label{eq:Sigma-vs-N}
\mathcal N(\lambda^2) = N_\mu(\lambda)
\qquad (\lambda\ge0).
\end{equation}
Using Lemma~\ref{lem:heat-spectral} and the spectral expansion of $p_t$, we
have
\begin{equation}\label{eq:Hmu-Laplace}
H_\mu(t)
= \sum_{j=0}^\infty e^{-t\lambda_j^2}
\Bigl|\int_M e_j\,d\mu\Bigr|^2
= \int_{[0,\infty)} e^{-tS}\,d\mathcal N(S),
\qquad t>0.
\end{equation}
In particular, $H_\mu(t)$ is the Laplace--Stieltjes transform of $\mathcal N$.
\subsection{A Tauberian theorem for Laplace--Stieltjes transforms}
We now recall a standard Tauberian theorem which connects the singular
behavior of a Laplace--Stieltjes transform at the origin to the growth of
the underlying measure. This is a special case of Karamata's theorem. See,
for example, \cite[Theorem~1.7.1]{BinghamGoldieTeugels}.
\begin{theorem}[Karamata's Tauberian Theorem]\label{thm:Tauberian-heat}
Let $\mathcal N:[0,\infty)\to[0,\infty)$ be nondecreasing and right-continuous,
and let
\[
\mathcal F(t) := \int_{[0,\infty)} e^{-tS}\,d\mathcal N(S)
\]
be finite for all $t>0$. Suppose that for some constants $C>0$ and
$\beta>0$,
\[
\mathcal F(t) \sim C\,t^{-\beta}
\qquad (t\to0^+).
\]
Then
\[
\mathcal N(S) \sim \frac{C}{\Gamma(\beta+1)}\,S^\beta
\qquad (S\to\infty).
\]
\end{theorem}
Theorem~\ref{thm:Tauberian-heat} is the Tauberian half of the usual
``Abelian $\Longleftrightarrow$ Tauberian'' correspondence for
Laplace--Stieltjes transforms with index $\beta>0$: roughly speaking,
asymptotics of $\mathcal N(S)$ as $S\to\infty$ and asymptotics of its
Laplace--Stieltjes transform $\mathcal F(t)$ as $t\to0^+$ determine each other.
A proof in this general framework can be found in
\cite[§1.7]{BinghamGoldieTeugels}.
\subsection{Proof of Theorem~\ref{thm:main}}
We now combine Lemma~\ref{lem:Hmu-asympt} with Theorem
\ref{thm:Tauberian-heat}. Recall that Lemma~\ref{lem:Hmu-asympt} shows that
\[
H_\mu(t)
= \gamma_{n,s}\,A_\mu\,t^{-(n-s)/2}
+ o\bigl(t^{-(n-s)/2}\bigr)
\qquad (t\to0^+),
\]
with $\gamma_{n,s}$ given by \eqref{eq:gamma-ns}. In the notation of
Theorem~\ref{thm:Tauberian-heat} we have
\[
\mathcal F(t) = H_\mu(t),\quad
C = \gamma_{n,s}A_\mu,\quad
\beta = \frac{n-s}{2}.
\]
Therefore
\[
\mathcal N(S)
\sim \frac{\gamma_{n,s}A_\mu}{\Gamma\bigl(\frac{n-s}{2}+1\bigr)}\,
S^{(n-s)/2}
\qquad (S\to\infty).
\]
Using \eqref{eq:Sigma-vs-N} with $S=\lambda^2$, we obtain
\[
N_\mu(\lambda)
= \mathcal N(\lambda^2)
\sim \frac{\gamma_{n,s}A_\mu}{\Gamma\bigl(\tfrac{n-s}{2}+1\bigr)}\,
\lambda^{n-s}
\qquad (\lambda\to\infty).
\]
In particular, \eqref{eq:Kuz-main} holds with
\[
C_{n,s}
= \frac{\gamma_{n,s}}{\Gamma\bigl(\tfrac{n-s}{2}+1\bigr)}.
\]
It remains to simplify this constant. Recall from \eqref{eq:gamma-ns} and
\eqref{eq:gamma-s} that
\[
\gamma_{n,s}
= (4\pi)^{-n/2}\,
s\,2^{\,s-1}\Gamma\Bigl(\frac{s}{2}\Bigr)\,\vol(B^s),
\]
and that
\[
\vol(B^m)
= \frac{\pi^{m/2}}{\Gamma\bigl(\frac{m}{2}+1\bigr)},
\qquad m>0.
\]
Since $\Gamma\bigl(\frac{s}{2}+1\bigr)=\frac{s}{2}\Gamma\bigl(\frac{s}{2}\bigr)$,
we may write
\[
\vol(B^s)
= \frac{\pi^{s/2}}{\Gamma\bigl(\frac{s}{2}+1\bigr)}
= \frac{2}{s}\,\pi^{s/2}\,\Gamma\Bigl(\frac{s}{2}\Bigr)^{-1}.
\]
Substituting this into the formula for $\gamma_{n,s}$ gives
\[
\gamma_{n,s}
= (4\pi)^{-n/2}\,s\,2^{\,s-1}\Gamma\Bigl(\frac{s}{2}\Bigr)
\cdot \frac{2}{s}\,\pi^{s/2}\,\Gamma\Bigl(\frac{s}{2}\Bigr)^{-1}
= (4\pi)^{-n/2}\,2^{\,s}\,\pi^{s/2}.
\]
Hence
\[
C_{n,s}
= \frac{(4\pi)^{-n/2}2^{\,s}\pi^{s/2}}
{\Gamma\bigl(\frac{n-s}{2}+1\bigr)}
= 2^{-(n-s)}\pi^{-(n-s)/2}
\frac{1}{\Gamma\bigl(\frac{n-s}{2}+1\bigr)}.
\]
On the other hand, taking $m=n-s$ in the volume formula,
\[
\vol(B^{n-s})
= \frac{\pi^{(n-s)/2}}{\Gamma\bigl(\frac{n-s}{2}+1\bigr)},
\]
so
\[
(2\pi)^{-(n-s)}\,\vol(B^{n-s})
= 2^{-(n-s)}\pi^{-(n-s)}\cdot
\frac{\pi^{(n-s)/2}}{\Gamma\bigl(\frac{n-s}{2}+1\bigr)}
= 2^{-(n-s)}\pi^{-(n-s)/2}
\frac{1}{\Gamma\bigl(\frac{n-s}{2}+1\bigr)}
= C_{n,s}.
\]
Thus
\[
C_{n,s}
= (2\pi)^{-(n-s)}\,\vol(B^{\,n-s}),
\]
which is exactly the constant stated in Theorem~\ref{thm:main}. This
completes the proof.
\section{formulae of Hare--Roginskaya and the proof of Theorem \ref{thm:HR-bound}}\label{sec:HR}
In this section we explain how the energy identities of Hare and Roginskaya
\cite{HareRoginskaya03} imply the polynomial upper bounds
stated in Theorem~\ref{thm:HR-bound}.
Recall that for $0<u<n$ we defined the Riesz energy of a finite Borel measure
$\mu$ on $M$ by
\[
I_u(\mu) = \iint d_g(x,y)^{-u}\,d\mu(x)\,d\mu(y),
\]
where $d_g(x,y)$ denotes the Riemannian distance on $(M,g)$, see
\eqref{eq:Riesz-energy}. Hare and Roginskaya relate $I_u(\mu)$ to a
weighted spectral sum via the kernel of a suitable elliptic operator. For
our purposes, we only need the following consequence of their results.
\begin{proposition}[\cite{HareRoginskaya03}]\label{prop:HR-energy}
Let $(M,g)$ be a compact Riemannian manifold of dimension $n\ge2$. For each
$u\in(0,n)$ there exists a constant $c_{n,u}\ge1$, depending only on $n$, $u$
and $(M,g)$, such that for every finite Borel measure $\mu$ on $M$,
\begin{equation}\label{eq:HR-energy}
c_{n,u}^{-1} I_u(\mu)
\le \sum_{j=0}^\infty (1+\lambda_j^2)^{-(n-u)/2}
\Bigl|\int_M e_j\,d\mu\Bigr|^2
\le c_{n,u}\, I_u(\mu).
\end{equation}
\end{proposition}
\begin{remark}
Estimates of the form \eqref{eq:HR-energy} are proved in
\cite{HareRoginskaya03} by expressing suitable powers of the
Laplace--Beltrami operator as pseudodifferential operators and identifying
their kernels with Riesz-type potentials. In particular, $I_u(\mu)$ is
comparable, up to multiplicative constants depending only on $(M,g,n,u)$, to
the quadratic form associated with the operator $(I-\Delta_g)^{-(n-u)/2}$
evaluated on $\mu$. See \cite{HareRoginskaya03} for details.
\end{remark}
Given Proposition~\ref{prop:HR-energy}, the proof of
Theorem~\ref{thm:HR-bound} is straightforward.
\subsection*{Proof of Theorem~\ref{thm:HR-bound}}
Let $u\in(0,n)$ and let $\mu$ be a finite Borel measure with
$I_u(\mu)<\infty$. By the upper bound in \eqref{eq:HR-energy},
\begin{equation}\label{eq:weighted-sum-bound}
\sum_{j=0}^\infty (1+\lambda_j^2)^{-(n-u)/2}
\Bigl|\int_M e_j\,d\mu\Bigr|^2
\le c_{n,u}\, I_u(\mu).
\end{equation}
Now fix $\lambda\ge1$. Since the weight
$(1+\lambda_j^2)^{-(n-u)/2}$ is decreasing in $\lambda_j$, we have
\[
(1+\lambda_j^2)^{-(n-u)/2}
\ge (1+\lambda^2)^{-(n-u)/2}
\qquad\text{whenever }\lambda_j\le\lambda,
\]
and hence
\[
\sum_{\lambda_j\le\lambda} \Bigl|\int_M e_j\,d\mu\Bigr|^2
\le (1+\lambda^2)^{\frac{n-u}{2}}
\sum_{j=0}^\infty (1+\lambda_j^2)^{-(n-u)/2}
\Bigl|\int_M e_j\,d\mu\Bigr|^2.
\]
Combining this with \eqref{eq:weighted-sum-bound} gives
\[
\sum_{\lambda_j\le\lambda} \Bigl|\int_M e_j\,d\mu\Bigr|^2
\le C_1(M,g,u,\mu)\,(1+\lambda^2)^{\frac{n-u}{2}}
\le C_1(M,g,u,\mu)\,\lambda^{\,n-u}
\]
for all $\lambda\ge1$, where we may take
\[
C_1(M,g,u,\mu)
:= 2^{\frac{n-u}{2}}c_{n,u}\,I_u(\mu).
\]
This proves the first assertion of Theorem~\ref{thm:HR-bound}.
To obtain the ``in particular'' statement, it remains to note that if $\mu$
is $s$-Ahlfors regular for some $s\in(0,n)$, then $I_u(\mu)<\infty$ for
every $u<s$. This is a standard estimate, but we include a brief argument
in our setting for completeness.
\begin{lemma}\label{lem:Ahlfors-finite-energy}
Let $\mu$ be an $s$-Ahlfors regular measure on $M$ with $s\in(0,n)$, and let
$u\in(0,s)$. Then $I_u(\mu)<\infty$.
\end{lemma}
\begin{proof}
Recall
\[
F(r)
= (\mu\times\mu)\bigl(\{(x,y): d_g(x,y)\le r\}\bigr)
= \int_M \mu(B(x,r))\,d\mu(x).
\]
Using the layer-cake representation, we have
\[
I_u(\mu)
= \int_0^{\diam(M)} u r^{-u-1} F(r)\,dr + \mu(M)^2\,\diam(M)^{-u},
\]
Split the integral at $r_0$:
\[
I_u(\mu)
= u\int_0^{r_0} r^{-u-1}F(r)\,dr
+ u\int_{r_0}^{\diam(M)} r^{-u-1}F(r)\,dr+ \mu(M)^2\,\diam(M)^{-u}.
\]
On $(0,r_0]$,
\[
u\int_0^{r_0} r^{-u-1}F(r)\,dr
\le uC\int_0^{r_0} r^{s-u-1}\,dr
= \frac{uC}{s-u}\,r_0^{\,s-u} <\infty,
\]
since $s-u>0$. On $[r_0,\diam(M)]$,
\[
u\int_{r_0}^{\diam(M)} r^{-u-1}F(r)\,dr
\le u\mu(M)^2\int_{r_0}^{\diam(M)} r^{-u-1}\,dr
<\infty.
\]
Thus $I_u(\mu)<\infty$ for every $u<s$, as claimed.
\end{proof}
Applying Lemma~\ref{lem:Ahlfors-finite-energy} with any fixed $u<s$ shows
that the first part of Theorem~\ref{thm:HR-bound} applies to $\mu$ and yields
a constant $C_2=C_2(M,g,s,u,\mu)$ such that
\[
\sum_{\lambda_j\le\lambda} \Bigl|\int_M e_j\,d\mu\Bigr|^2
\le C_2\,\lambda^{\,n-u}
\qquad (\lambda\ge1),
\]
without any density assumption. This completes the proof of
Theorem~\ref{thm:HR-bound}.
\begin{remark}
Theorem~\ref{thm:HR-bound} shows that for each fixed $u<s$ one can control the
growth of the Kuznecov sum by $\lambda^{n-u}$ under the mild hypothesis
$I_u(\mu)<\infty$, and in particular for $s$-Ahlfors regular measures. Our
main theorem, Theorem~\ref{thm:main}, goes further by identifying the sharp
exponent $n-s$ and the exact leading constant in the asymptotic
\eqref{eq:Kuz-main}, under the additional assumption of the
averaged $s$-density condition. As discussed in the introduction, the methods
of \cite{HareRoginskaya03} treat $u$ as a fixed parameter
and do not provide the uniform control in $u$ near $s$ that is needed to
pass to the endpoint $u=s$. This is why we turn to a more refined analysis
based on the heat kernel and Tauberian theory.
\end{remark}
\section{Sharpness of Theorem \ref{thm:main}}\label{sec:sharpness}
In this section, we discuss the sharpness of Theorem~\ref{thm:main}. We explain the
necessity of the two structural hypotheses, namely $s$-Ahlfors regularity and
the averaged $s$-density condition \eqref{eq:DD}, for obtaining the exponent
$n-s$ and a genuine one-term asymptotic. We also show that the remainder
$o(\lambda^{n-s})$ in \eqref{eq:Kuz-main} is essentially optimal at this level of
generality: in general it cannot be improved uniformly to a power-saving error
term over the class of measures under consideration.
\subsection{Necessity of the Ahlfors regularity assumption}
It is natural to ask whether the exponent $n-s$ in Theorem~\ref{thm:main} (or in
the upper bound of Theorem~\ref{thm:HR-bound}) could be expressed solely in
terms of the Hausdorff dimension of $\mu$, without any Ahlfors regularity
hypothesis. This is not possible in general.
Let $H_1,H_2\subset M$ be disjoint embedded submanifolds with dimensions
$s_1<s_2$, and let $\mu_i=dV_{H_i}$ and $\mu=\mu_1+\mu_2$. Then
\[
\dim_H(\mu)=\max(s_1,s_2)=s_2,
\]
but the growth of $N_\mu(\lambda)$ is governed by the \emph{smaller} dimension
$s_1$.
To see this, write
\[
a_j:=\int_{H_1} e_j\,dV_{H_1},\qquad b_j:=\int_{H_2} e_j\,dV_{H_2}.
\]
Then
\[
N_\mu(\lambda)=\sum_{\lambda_j\le\lambda}|a_j+b_j|^2
= \sum_{\lambda_j\le\lambda}|a_j|^2+\sum_{\lambda_j\le\lambda}|b_j|^2
+ 2\Re \sum_{\lambda_j\le\lambda} a_j\overline{b_j}.
\]
By Cauchy--Schwarz,
\[
\Bigl|\sum_{\lambda_j\le\lambda} a_j\overline{b_j}\Bigr|
\le \Bigl(\sum_{\lambda_j\le\lambda}|a_j|^2\Bigr)^{1/2}
\Bigl(\sum_{\lambda_j\le\lambda}|b_j|^2\Bigr)^{1/2}.
\]
Hence we have the robust comparison
\begin{equation}\label{eq:cross-term-bound}
\bigl|N_\mu(\lambda)-N_{\mu_1}(\lambda)-N_{\mu_2}(\lambda)\bigr|
\le 2\,\bigl(N_{\mu_1}(\lambda)N_{\mu_2}(\lambda)\bigr)^{1/2}.
\end{equation}
By the smooth submanifold Kuznecov formulae,
\[
N_{\mu_i}(\lambda)\sim C_{n,s_i}\,\vol(H_i)\,\lambda^{n-s_i}\qquad(i=1,2),
\]
so $N_{\mu_2}(\lambda)=o(N_{\mu_1}(\lambda))$ and the right-hand side of
\eqref{eq:cross-term-bound} is
\[
O\Bigl(\lambda^{\frac{n-s_1}{2}}\,\lambda^{\frac{n-s_2}{2}}\Bigr)
= O\bigl(\lambda^{\,n-\frac{s_1+s_2}{2}}\bigr)
= o\bigl(\lambda^{n-s_1}\bigr).
\]
Therefore $N_\mu(\lambda)=N_{\mu_1}(\lambda)+o(\lambda^{n-s_1})$, and in
particular
\[
N_\mu(\lambda)\sim C_{n,s_1}\,\vol(H_1)\,\lambda^{n-s_1}\qquad(\lambda\to\infty),
\]
even though $\dim_H(\mu)=s_2$. Thus, no exponent depending only on
$\dim_H(\mu)$ can govern $N_\mu(\lambda)$ for arbitrary finite Borel measures.
\subsection{Necessity of the averaged $s$-density condition}\label{subsec:necessity-density}
The averaged $s$-density hypothesis in Theorem~\ref{thm:main} is not merely a
technical convenience: without a stabilization of the small-scale behavior of
$\mu$, one cannot in general expect the normalized quantities
$N_\mu(\lambda)/\lambda^{n-s}$ to converge.
Recall the distance distribution
\[
F(r):=(\mu\times\mu)\bigl(\{(x,y):d_g(x,y)\le r\}\bigr),\qquad r\ge0,
\]
and the Gaussian average
\[
G(t):=\iint e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y)
=\int_0^{\diam(M)} e^{-r^2/(4t)}\,dF(r).
\]
Define the normalized coefficient
\[
a(r):=\frac{F(r)}{r^s}\qquad(r>0).
\]
If $\mu$ is $s$-Ahlfors regular then $a(r)$ is bounded above and below by
positive constants for $0<r\le r_0$.
Since $F(0)=0$, integration by parts gives, modulo a boundary term that is exponentially decaying in $1/t$, we have
\begin{equation}\label{eq:G-average-a}
G(t)
= \frac{1}{2t}\int_0^{\diam(M)} r e^{-r^2/(4t)}\,F(r)\,dr.
\end{equation}
Changing variables $r=\sqrt{t}\,u$ yields
\begin{equation}\label{eq:G-average-a2}
\frac{G(t)}{t^{s/2}}
= \int_0^{\diam(M)/\sqrt t} w_s(u)\,a(\sqrt t\,u)\,du,
\qquad
w_s(u):=\frac12\,u^{s+1}e^{-u^2/4}.
\end{equation}
The weight $w_s$ is positive and integrable on $(0,\infty)$, with
\[
\kappa_s:=\int_0^\infty w_s(u)\,du
= \frac12\int_0^\infty u^{s+1}e^{-u^2/4}\,du
= s\,2^{s-1}\Gamma\Bigl(\frac{s}{2}\Bigr).
\]
Let
\[
a_-:=\liminf_{r\to0^+} a(r),\qquad a_+:=\limsup_{r\to0^+} a(r).
\]
Using \eqref{eq:G-average-a2}, the boundedness of $a$, and $\int_0^\infty w_s<\infty$,
one obtains the general bounds
\begin{equation}\label{eq:G-liminf-limsup-general}
\kappa_s\,a_-
\le \liminf_{t\to0^+}\frac{G(t)}{t^{s/2}}
\le \limsup_{t\to0^+}\frac{G(t)}{t^{s/2}}
\le \kappa_s\,a_+.
\end{equation}
In general \eqref{eq:G-liminf-limsup-general} may be strict: the Gaussian
average is a \emph{weighted} average of the values of $a(r)$ on the scale
$r\approx\sqrt t$, so thin oscillations in $a$ can be smoothed.
However, in the model constructions relevant for sharpness one can arrange that
different limit points of $a(r)$ persist across long blocks of
scales. In that setting the endpoints in \eqref{eq:G-liminf-limsup-general} are
realized along suitable subsequences.
\begin{lemma}[Realizing $\liminf/\limsup$ from symmetric scale blocks]\label{lem:realize-liminf-limsup}
Assume $a$ is bounded on $(0,\diam (M)]$, and let
\[
a_-:=\liminf_{r\to0^+} a(r),\qquad a_+:=\limsup_{r\to0^+} a(r).
\]
\begin{enumerate}
\item If there exist $r_m\to0^+$ and $U_m\to\infty$ such that
\[
\sup_{\,r_m/U_m \le r \le U_m r_m}\,|a(r)-a_-|\longrightarrow 0,
\]
then with $t_m:=r_m^2$ one has
\[
\frac{G(t_m)}{t_m^{s/2}}\longrightarrow \kappa_s\,a_-.
\]
\item If there exist $\rho_m\to0^+$ and $V_m\to\infty$ such that
\[
\sup_{\,\rho_m/V_m \le r \le V_m \rho_m}\,|a(r)-a_+|\longrightarrow 0,
\]
then with $\tau_m:=\rho_m^2$ one has
\[
\frac{G(\tau_m)}{\tau_m^{s/2}}\longrightarrow \kappa_s\,a_+.
\]
\end{enumerate}
\end{lemma}
\begin{proof}
We prove (1); (2) is identical. Put $t_m=r_m^2$ and use \eqref{eq:G-average-a2}:
\[
\frac{G(t_m)}{t_m^{s/2}}
= \int_0^{\diam(M)/r_m} w_s(u)\,a(r_m u)\,du.
\]
Split the integral into three pieces:
\[
\int_0^{\diam(M)/r_m}
= \int_0^{1/U_m}+\int_{1/U_m}^{U_m}+\int_{U_m}^{\diam(M)/r_m}.
\]
On $u\in[1/U_m,U_m]$ we have $r_m u\in[r_m/U_m,\,U_m r_m]$, hence
$a(r_m u)\to a_-$ uniformly, so
\[
\int_{1/U_m}^{U_m} w_s(u)\,a(r_m u)\,du
= a_-\int_{1/U_m}^{U_m} w_s(u)\,du + o(1).
\]
Since $a$ is bounded, say $|a|\le K$ on $(0,r_0]$, the tails satisfy
\[
\left|\int_0^{1/U_m} w_s(u)\,a(r_m u)\,du\right|
\le K\int_0^{1/U_m} w_s(u)\,du \longrightarrow 0,
\]
because $w_s(u)\approx u^{s+1}$ near $0$, and
\[
\left|\int_{U_m}^{\diam(M)/r_m} w_s(u)\,a(r_m u)\,du\right|
\le K\int_{U_m}^{\infty} w_s(u)\,du \longrightarrow 0,
\]
because $w_s\in L^1(0,\infty)$ and $U_m\to\infty$. Finally,
\[
\int_{1/U_m}^{U_m} w_s(u)\,du \longrightarrow \int_0^\infty w_s(u)\,du=\kappa_s,
\]
so $G(t_m)/t_m^{s/2}\to \kappa_s a_-$.
\end{proof}
Lemma~\ref{lem:realize-liminf-limsup} shows that if $a(r)$ has two distinct limit
points $a_-<a_+$ and one can arrange two sequences of logarithmic blocks on which
$a(r)$ stays uniformly close to $a_-$ and $a_+$, respectively, then
$G(t)/t^{s/2}$ has two distinct subsequential limits $\kappa_s a_-$ and $\kappa_s a_+$,
and hence does not converge. By the near-diagonal reduction in
Section~\ref{sec:heat}, this forces $t^{(n-s)/2}H_\mu(t)$, and therefore
$N_\mu(\lambda)/\lambda^{n-s}$, to fail to converge as well. This explains why
some stabilization of $F(r)/r^s$ is needed for a
genuine one-term asymptotic.
\subsection{Optimality of the remainder}\label{subsec:optimal-remainder}
In the smooth submanifold case, Zelditch's Kuznecov formulae \eqref{eq:Zelditch}
come with an $O(\lambda^{n-k-1})$ remainder term. Moreover, for generic metrics
this remainder can be improved to $o(\lambda^{n-k-1})$ by a recent result of
Kaloshin, Wyman and the author \cite{kaloshin2025kuznecov}. In contrast, the asymptotic formula
in Theorem~\ref{thm:main} is stated with an $o(\lambda^{n-s})$ remainder.
It is natural to ask whether, under the assumptions of $s$-Ahlfors regularity and
the existence of an averaged $s$-density constant, one can improve this to a \emph{uniform}
power-saving error term of the form
\begin{equation}\label{eq:uniform-powersaving}
N_\mu(\lambda)=C_{n,s}A_\mu\,\lambda^{n-s}+O(\lambda^{n-s-\delta})
\qquad(\lambda\to\infty),
\end{equation}
for some $\delta>0$ depending only on $(M,g,n,s)$ and valid for all measures $\mu$
in our class. We now explain why such a uniform improvement is not available
in general.
\medskip
\noindent\textbf{Step 1: from $N_\mu$ to the Laplace--Stieltjes transform $H_\mu$.}
Write $\mathcal N(S):=N_\mu(\sqrt S)$ as in \eqref{eq:Sigma-vs-N}, so that
\[
H_\mu(t)=\int_{[0,\infty)} e^{-tS}\,d\mathcal N(S).
\]
If \eqref{eq:uniform-powersaving} holds for some $\delta>0$, then
\[
\mathcal N(S)=C_{n,s}A_\mu\,S^{\frac{n-s}{2}}+O\Bigl(S^{\frac{n-s-\delta}{2}}\Bigr),
\]
and a standard Abelian (integration-by-parts) computation yields
\begin{equation}\label{eq:Abelian-remainder}
H_\mu(t)
= C_{n,s}A_\mu\,\Gamma\Bigl(\frac{n-s}{2}+1\Bigr)\,t^{-\frac{n-s}{2}}
+ O\Bigl(t^{-\frac{n-s-\delta}{2}}\Bigr)
\qquad(t\to0^+).
\end{equation}
\medskip
\noindent\textbf{Step 2: from $H_\mu$ to the Gaussian average $G(t)$.}
Recall
\[
G(t):=\iint e^{-d_g(x,y)^2/(4t)}\,d\mu(x)\,d\mu(y).
\]
Inspecting the proof of Lemma~\ref{lem:Hmu-asympt} (using the Lipschitz bound
\eqref{eq:u0-Lipschitz}, the estimate $J(t)\lesssim t^{(s+1)/2}$, and choosing an
integer $N\ge \lceil (s+1)/2\rceil$ in the local heat-kernel parametrix) yields the
quantitative refinement
\begin{equation}\label{eq:H-vs-G-rate}
H_\mu(t)=(4\pi t)^{-n/2}G(t) + O\!\left(t^{-\frac{n-s}{2}+\frac12}\right)
\qquad(t\to0^+).
\end{equation}
Combining \eqref{eq:Abelian-remainder} and \eqref{eq:H-vs-G-rate}, and using that
$C_{n,s}=\gamma_{n,s}/\Gamma(\frac{n-s}{2}+1)$ and $\gamma_{n,s}=(4\pi)^{-n/2}\gamma_s$,
one obtains (for $\delta\in(0,1)$)
\begin{equation}\label{eq:G-power-saving}
\frac{G(t)}{t^{s/2}}
= \gamma_s A_\mu + O(t^{\delta/2})
\qquad(t\to0^+),
\end{equation}
where $\gamma_s$ is as in Lemma~\ref{lem:Gaussian-F}. Thus, a uniform power saving
in \eqref{eq:uniform-powersaving} would force a uniform power saving in $t$ for
$G(t)/t^{s/2}$.
\medskip
\noindent\textbf{Step 3: from $G(t)$ to $F(r)$.}
Recall $F(r):=(\mu\times\mu)\{(x,y):d_g(x,y)\le r\}$ and set
\[
q(r):=\frac{F(r)}{\vol(B^s)\,A_\mu\,r^s}.
\]
Then by definition, $q(r)\to1$
as $r\to0^+$. Integrating by parts in
$G(t)=\int_0^{\diam(M)} e^{-r^2/(4t)}\,dF(r)$ and changing variables $r=\sqrt t\,u$
yields
\begin{equation}\label{eq:G-average-q}
\frac{G(t)}{\vol(B^s)\,A_\mu\,t^{s/2}}
= \int_0^{\diam(M)/\sqrt t} w_s(u)\,q(\sqrt t\,u)\,du + O(e^{-c/t}),
\end{equation}
where
\[
w_s(u):=\frac12\,u^{s+1}e^{-u^2/4}
\]
is strictly positive and integrable on $(0,\infty)$.
\begin{lemma}\label{lem:block-deviation-G}
Assume $q(r)\to1$ as $r\to0^+$ and $\|q\|_{L^\infty(0,\diam(M)]}\le C_0$ for some $C_0\ge1$.
Fix $\eta\in(0,1)$ and choose $U=U(\eta)\ge2$ so that
\begin{equation}\label{eq:choose-U}
\int_U^\infty w_s(u)\,du \le \frac{\eta}{8C_0}\int_1^U w_s(u)\,du.
\end{equation}
Fix $r>0$, put $t=r^2$, and assume that $\text{for all }\rho\in[r,Ur].$
\[
q(\rho)\le 1-\eta.
\]
Then for all sufficiently small $r$ one has
\[
\left|\frac{G(t)}{t^{s/2}}-\gamma_sA_\mu\right|
\ge c_s\,A_\mu\,\eta\int_1^U w_s(u)\,du,
\]
where $c_s>0$ depends only on $s$.
\end{lemma}
\begin{proof}
Let $\kappa_s:=\int_0^\infty w_s(u)\,du$, so $\gamma_s=\vol(B^s)\kappa_s$. Using
\eqref{eq:G-average-q} with $t=r^2$ and absorbing the exponentially small term, we get
\[
\frac{G(r^2)}{\vol(B^s)\,A_\mu\,r^s}-\kappa_s
= \int_0^\infty w_s(u)\bigl(q(ru)-1\bigr)\,du + o(1)
\qquad (r\to0^+).
\]
Split the integral into $u\in(0,1)$, $u\in[1,U]$, and $u\in[U,\infty)$.
On $[1,U]$ we have $ru\in[r,Ur]$, so $q(ru)-1$ has fixed sign and $|q(ru)-1|\ge\eta$.
Hence
\[
\left|\int_1^U w_s(u)\bigl(q(ru)-1\bigr)\,du\right|
\ge \eta\int_1^U w_s(u)\,du.
\]
On $[U,\infty)$ we use $|q-1|\le C_0+1\le 2C_0$ and \eqref{eq:choose-U}:
\[
\left|\int_U^\infty w_s(u)\bigl(q(ru)-1\bigr)\,du\right|
\le 2C_0\int_U^\infty w_s(u)\,du
\le \frac{\eta}{4}\int_1^U w_s(u)\,du.
\]
Finally, since $q(\rho)\to1$ as $\rho\to0^+$ and $w_s\in L^1(0,1)$, dominated convergence gives
\[
\int_0^1 w_s(u)\bigl(q(ru)-1\bigr)\,du \longrightarrow 0
\qquad (r\to0^+).
\]
Therefore for all sufficiently small $r$,
\[
\left|\int_0^\infty w_s(u)\bigl(q(ru)-1\bigr)\,du\right|
\ge \frac{\eta}{2}\int_1^U w_s(u)\,du.
\]
Multiplying by $\vol(B^s)A_\mu$ and using $\gamma_s=\vol(B^s)\kappa_s$ yields the claim.
\end{proof}
\medskip
\noindent\textbf{Step 4: constructing arbitrarily slow convergence.}
We now construct an explicit example (for a fixed $\delta>0$) for which no
uniform remainder $O(\lambda^{n-s-\delta})$ can hold. For simplicity, we take $s$ to be an integer, and
\[
M=\mathbb T^n,\qquad H=\mathbb T^s\subset\mathbb T^n
\]
with the flat metric, so that $dV_H$ is Lebesgue measure on $\mathbb T^s$ and,
for all $\rho\in(0,1/10]$, the intrinsic ball $B_H(x,\rho)$ is a translate of
the Euclidean ball. We remark that similar constructions can be made for a general manifold--submanifold pair $(M,H)$.
Let
\[
K_\rho(y):=\mathbf 1_{B_H(0,\rho)}(y),\qquad
\mathcal A_\rho f(x):=\frac{1}{\vol(B^s)\rho^s}\int_{\mathbb T^s} f(x-y)K_\rho(y)\,dV_H(y).
\]
Then whenever $d\mu=w\,dV_H$ we have
\[
\frac{F(\rho)}{\vol(B^s)\rho^s}
=\int_H w(x)\,\mathcal A_\rho w(x)\,dV_H(x).
\]
Choose frequencies $k_m\in\mathbb Z^s\setminus\{0\}$ such that $|k_m|$ is strictly
increasing and $k_m\neq \pm k_{m'}$ for $m\neq m'$, and define
\[
\phi_m(x):=\cos(2\pi k_m\cdot x),\qquad
\|\phi_m\|_{L^\infty(H)}\le 1,\qquad
\int_H \phi_m\,dV_H=0,\qquad
\int_H \phi_m^2\,dV_H=\tfrac12\,\vol(H).
\]
For $\rho\in(0,1/10]$ the averaging convolution operator $\mathcal A_\rho$ acts diagonally on
Fourier modes: there exists a continuous function $m_s:[0,\infty)\to\mathbb R$
with
\[
m_s(0)=1,\qquad \lim_{\tau\to\infty} m_s(\tau)=0,\qquad |m_s(\tau)|\le 1,
\]
such that
\[
\mathcal A_\rho \phi_m = m_s(2\pi |k_m|\rho)\,\phi_m .
\]
Fix amplitudes $a_m>0$ such that
\begin{equation}\label{eq:am-fourier}
\sum_{m\ge1} a_m\le \tfrac14,
\qquad\text{and hence}\qquad
\sum_{m\ge1} a_m^2<\infty.
\end{equation}
Define
\[
w(x):=1+\sum_{m\ge1} a_m\,\phi_m(x),
\qquad\text{and}\qquad
d\mu:=w\,dV_H.
\]
Then $\tfrac34\le w\le \tfrac54$, so $\mu$ is $s$-Ahlfors regular. Moreover, since
$\mu$ is absolutely continuous with respect to $dV_H$ with bounded density $w$,
$\mu$ admits a pointwise $s$-density $\theta=w$. Hence, by Remark~\ref{rem:density-implies-DD},
\[
A_\mu=\int_H \theta\,d\mu=\int_H w^2\,dV_H.
\]
Using orthogonality of distinct Fourier modes,
\begin{equation}\label{eq:A-mu-fourier}
A_\mu
=\int_H w^2\,dV_H
=\vol(H)+\frac12\,\vol(H)\sum_{m\ge1} a_m^2 .
\end{equation}
Moreover, for $\rho\in(0,1/10]$,
\begin{align}\label{eq:F-fourier}
\frac{F(\rho)}{\vol(B^s)\rho^s}
&=\int_H w\,\mathcal A_\rho w\,dV_H \notag\\
&=\vol(H)
+\frac12\,\vol(H)\sum_{m\ge1} a_m^2\,m_s(2\pi |k_m|\rho).
\end{align}
Therefore
\begin{equation}\label{eq:q-fourier}
q(\rho)
=\frac{F(\rho)}{\vol(B^s)A_\mu\,\rho^s}
=\frac{1+\frac12\sum_{m\ge1} a_m^2\,m_s(2\pi |k_m|\rho)}
{1+\frac12\sum_{m\ge1} a_m^2}.
\end{equation}
Since $\sum a_m^2<\infty$ and $|m_s|\le 1$, dominated convergence and $m_s(0)=1$
give $q(\rho)\to 1$ as $\rho\to0^+$.
Choose numbers $0<\varepsilon_0<1<L_0$ such that
\begin{equation}\label{eq:m-thresholds}
|m_s(\tau)-1|\le \tfrac1{100}\quad\text{for }0\le \tau\le \varepsilon_0,
\qquad
|m_s(\tau)|\le \tfrac1{100}\quad\text{for }\tau\ge L_0.
\end{equation}
(This is possible by continuity at $0$ and $\lim_{\tau\to\infty}m_s(\tau)=0$.)
For each $m$ set
\[
\eta_m:=\frac{1}{100}\,a_m^2,
\]
and let $U_m:=U(\eta_m)\ge2$ be the scale parameter given by
Lemma~\ref{lem:block-deviation-G}. Define
\[
r_m:=\frac{L_0}{2\pi |k_m|}\,.
\]
We choose the frequencies inductively so that $|k_m|$ is strictly increasing and
\begin{equation}\label{eq:freq-sep}
2\pi |k_{m-1}|\,U_m r_m \le \varepsilon_0,
\qquad
U_m r_m \le \tfrac1{10},
\qquad
r_m^\delta \le \frac{1}{10m}\,\eta_m .
\end{equation}
This is possible because once $a_m$ (hence $\eta_m$ and $U_m$) is fixed, we are
free to take $|k_m|$ arbitrarily large, which makes $r_m$ arbitrarily small.
Fix $m$ and take any $\rho\in[r_m,U_m r_m]$. Then
\[
2\pi |k_m|\rho \ge 2\pi |k_m|r_m = L_0,
\]
so \eqref{eq:m-thresholds} gives $|m_s(2\pi |k_m|\rho)|\le 1/100$, and the same
bound holds for all $j\ge m$ since $|k_j|\ge |k_m|$. On the other hand,
\eqref{eq:freq-sep} implies, for $j<m$,
\[
2\pi |k_j|\rho \le 2\pi |k_{m-1}|\,U_m r_m \le \varepsilon_0,
\]
so \eqref{eq:m-thresholds} gives $m_s(2\pi |k_j|\rho)\ge 1-1/100$ for all $j<m$.
Let $A:=\sum_{j<m}a_j^2$ and $T:=\sum_{j\ge m}a_j^2$ (so $T\ge a_m^2$). Using the
bounds above in \eqref{eq:q-fourier} yields, for $\rho\in[r_m,U_m r_m]$,
\[
q(\rho)
\le
\frac{1+\frac12\bigl(A+\tfrac1{100}T\bigr)}
{1+\frac12(A+T)}
\le
1-\frac{1-\tfrac1{100}}{1+\frac12(A+T)}\cdot \frac{T}{2}.
\]
Since $\sum_{m\ge1}a_m^2\le \sum_{m\ge1}a_m\le \tfrac14$, we have $A+T\le \tfrac14$
and hence $1+\frac12(A+T)\le \tfrac98$. Therefore
\[
q(\rho)\le 1-\frac{4}{9}\Bigl(1-\tfrac1{100}\Bigr)T
\le 1-\eta_m,
\]
so $q(\rho)\le 1-\eta_m$ on the entire block $[r_m,U_m r_m]$.
Applying Lemma~\ref{lem:block-deviation-G} with $r=r_m$ and $\eta=\eta_m$ gives
\[
\left|\frac{G(r_m^2)}{(r_m^2)^{s/2}}-\gamma_sA_\mu\right|
=\left|\frac{G(r_m^2)}{r_m^{s}}-\gamma_sA_\mu\right|
\ge c_s\,A_\mu\,\eta_m .
\]
Finally, the last condition in \eqref{eq:freq-sep} implies
\[
\frac{\eta_m}{(r_m^2)^{\delta/2}}
=\frac{\eta_m}{r_m^\delta}\ge 10m\longrightarrow\infty,
\]
so no estimate of the form
$\bigl|G(t)/t^{s/2}-\gamma_sA_\mu\bigr|=O(t^{\delta/2})$ can hold as $t\to0^+$.
By the implication \eqref{eq:G-power-saving}, this contradicts the consequence of a
uniform remainder $O(\lambda^{n-s-\delta})$. Therefore no uniform power saving
$O(\lambda^{n-s-\delta})$ can hold over our class, and the $o(\lambda^{n-s})$
remainder in Theorem~\ref{thm:main} is essentially sharp.
\bibliographystyle{alpha}
\bibliography{kuznecov_refs}
\end{document}
\begin{thebibliography}{KWX25}
\bibitem[BGT87]{BinghamGoldieTeugels}
N.~H. Bingham, C.~M. Goldie, and J.~L. Teugels.
\newblock {\em Regular Variation}, volume~27 of {\em Encyclopedia of
Mathematics and its Applications}.
\newblock Cambridge University Press, 1987.
\bibitem[BGT07]{BurqGerardTzvetkov07}
Nicolas Burq, Patrick G{\'e}rard, and Nikolay Tzvetkov.
\newblock Restrictions of the {L}aplace--{B}eltrami eigenfunctions to
submanifolds.
\newblock {\em Duke Math. J.}, 136(1):445--486, 2007.
\bibitem[BGV92]{BGV92}
Nicole Berline, Ezra Getzler, and Mich\`ele Vergne.
\newblock {\em Heat Kernels and Dirac Operators}, volume 298 of {\em
Grundlehren der Mathematischen Wissenschaften}.
\newblock Springer-Verlag, Berlin, 1992.
\bibitem[EP22]{EswarathasanPramanik22}
Suresh Eswarathasan and Malabika Pramanik.
\newblock Restriction of {L}aplace--{B}eltrami eigenfunctions to arbitrary sets
on manifolds.
\newblock {\em Int. Math. Res. Not. IMRN}, (2):1538--1600, 2022.
\bibitem[GMX24]{GaoMiaoXi24}
Chuanwei Gao, Changxing Miao, and Yakun Xi.
\newblock Refined {$L^p$} restriction estimate for eigenfunctions on
{R}iemannian surfaces.
\newblock {\em arXiv e-prints}, 2024.
\newblock arXiv:2411.01577.
\bibitem[Gri09]{Grigoryan09}
Alexander Grigor'yan.
\newblock {\em Heat Kernel and Analysis on Manifolds}, volume~47 of {\em AMS/IP
Studies in Advanced Mathematics}.
\newblock American Mathematical Society, Providence, RI, 2009.
\bibitem[H{\"o}r68]{Hormander1968SpectralFunction}
Lars H{\"o}rmander.
\newblock The spectral function of an elliptic operator.
\newblock {\em Acta Mathematica}, 121:193--218, 1968.
\bibitem[HR03]{HareRoginskaya03}
Kathryn~E. Hare and Maria Roginskaya.
\newblock Energy of measures on compact {R}iemannian manifolds.
\newblock {\em Studia Mathematica}, 159:291--314, 2003.
\bibitem[KWX25]{kaloshin2025kuznecov}
Vadim Kaloshin, Emmett~L. Wyman, and Yakun Xi.
\newblock Kuznecov remainders and generic metrics.
\newblock {\em arXiv preprint arXiv:2507.06887}, 2025.
\bibitem[Mat15]{MattilaFAHD}
Pertti Mattila.
\newblock {\em Fourier Analysis and Hausdorff Dimension}, volume 150 of {\em
Cambridge Studies in Advanced Mathematics}.
\newblock Cambridge University Press, 2015.
\bibitem[Sog88]{sogge1988concerning}
Christopher~D Sogge.
\newblock Concerning the ${L}^p$ norm of spectral clusters for second-order
elliptic operators on compact manifolds.
\newblock {\em Journal of functional analysis}, 77(1):123--138, 1988.
\bibitem[WX23a]{wyman2023can}
Emmett~L. Wyman and Yakun Xi.
\newblock Can you hear your location on a manifold?
\newblock {\em arXiv preprint arXiv:2304.04659}, 2023.
\bibitem[WX23b]{WymanXiKuz}
Emmett~L. Wyman and Yakun Xi.
\newblock A two term {K}uznecov sum formula.
\newblock {\em Communications in Mathematical Physics}, 401:1127--1162, 2023.
\bibitem[Zel92]{Zelditch92}
Steve Zelditch.
\newblock Kuznecov sum formulae and {S}zeg{\H{o}} limit formulae on manifolds.
\newblock {\em Communications in Partial Differential Equations},
17(1--2):221--260, 1992.
\end{thebibliography}