Random exponential sums and lattice points in regions
Extracted source text · arXiv 2411.06517
\documentclass[12 pt]{amsart}
\usepackage[margin=2.3cm]{geometry}
\usepackage{graphicx}
\usepackage{amssymb,latexsym}
\usepackage{amsmath,amsfonts,amstext,bbm}
\usepackage[utf8]{inputenc}
\usepackage{xcolor}
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}{Lemma}
\newtheorem{proposition}{Proposition}
\newtheorem{definition}{Definition}
\newtheorem{corollary}{Corollary}
\DeclareMathOperator{\var}{var}
\DeclareMathOperator{\Var}{Var}
\DeclareMathOperator{\E}{\mathbb{E}}
\DeclareMathOperator{\Pp}{\mathbb{P}}
\DeclareMathOperator{\R}{\mathbb{R}}
\DeclareMathOperator{\N}{\mathbb{N}}
\DeclareMathOperator{\Z}{\mathbb{Z}}
\DeclareMathOperator{\T}{\mathbb{T}}
\DeclareMathOperator{\V}{\mathbb{V}}
\DeclareMathOperator{\Oh}{\mathcal{O}}
\begin{document}
\title[RANDOM EXPONENTIAL SUMS AND LATTICE POINTS IN REGIONS]
{Random exponential sums and lattice points in regions}
\author{FARUK TEMUR, CİHAN SAHİLLİOĞULLARI}
\address{Department of Mathematics\\
İzmir Institute of Technology\\ Urla \\İzmir \\ 35430\\Turkey }
\email{faruktemur@iyte.edu.tr, cihansahilliogullari@iyte.edu.tr}
\keywords{Hardy-Littlewood majorant problem, Lattice points in regions, Mei-Chu Chang's conjecture, Random exponential sums}
\subjclass[2020]{Primary: 42A05, 42A61, 42A70, 11P05, 11P21; Secondary: 60G10,
60G50, 60G51}
\date{November 10, 2024}
\begin{abstract}
In this article we study two fundamental problems on exponential sums via randomization of frequencies with stochastic processes. These are the Hardy-Littlewood majorant problem, and $L^{2n}(\T), \ n\in \N$ norms of exponential sums, which can also be interpreted as solutions of diophantine equations or lattice points on surfaces. We establish connections to the well known problems on lattice points in regions such as the Dirichlet divisor problem.
\end{abstract}
\maketitle
\section{introduction}
In this work\footnote{The results of this article will constitute part of the second authors PhD thesis \cite{cs}. For deeper and wider background on these topics, more heuristics, and more detailed proofs we refer the reader to this thesis.} we will study random exponential sums of the form
\begin{equation}\label{int0}
\Big\|\sum_{j\in A} a_je^{2\pi i yX_j}\Big\|_{L^{p}(\T)}^{p}
\end{equation}
where $A$ is a finite subset of integers, and $X$ is a stochastic process on a probability space $(\Omega,\mathcal{F},\Pp).$ We will use these sums to investigate two esteemed problems of harmonic analysis and analytic number theory, the Hardy-Littlewood majorant problem, and $L^{2n}(\T), \ n\in \N$ norms of exponential sums. We will relate these sums for $p=2n$ and $a_j=0,1$ to problems on lattice points in regions such as the Dirichlet divisor problem. The regions that we will encounter have the form of shells around hypersurfaces. This can be seen as an analogue of the interpretation of $L^{2n}(\T), \ n\in \N$ norms of deterministic exponential sums as counting lattice points on surfaces. But in our case establishing this interpretation is rather involved.
The study of random exponential sums is a century old topic that is still immensely relevant as evidenced by the vast number of citations to \cite{k} which is a standard reference in this area. Since the pioneering works of Paley and Zygmund \cite{pz1,pz2}, exponential sums have often been randomized by multiplying the coefficients with independent identically distributed random variables $X_j$:
\begin{equation}\label{}
\Big\|\sum_{j\in A} a_jX_je^{2\pi i jy}\Big\|_{L^{p}(\T)}^{p}.
\end{equation}
And often the distribution of $X_j$ is chosen to be either Bernoulli with values $\pm 1$, or uniform on $\T$, or Gaussian. To study our problems however we must randomize the frequency set $A$, which can be done either by multiplying the coefficients with selector variables, that is Bernoulli variables that take values $0$ and $1$, or by directly randomizing the frequency as we do in this work. Selector variables were employed by Mockenhaupt and Schlag \cite{ms} to study the Hardy-Littlewood majorant problem. An advantage of our approach is that we can do randomization around a fixed set without much destroying the structure of it. For instance the sets $\{ N(j^d)\}_{j\in \N}$ for $d\geq 3$ will still look much like the set $\{j^d\}_{j\in \N}$, whereas had we randomized with selector variables they would have lost all similarity to it.
Precedents for randomization of frequency by stochastic processes exist in abundance. They emerge naturally in the study of Gaussian Fourier series, and of Brownian images. For example in Chapter 14 of \cite{k} for a process $F(t)$ obtained by Gaussian Fourier series, and $\theta$ some measure on a subset $E\subseteq\T$ the expression
\begin{equation}\label{int-2}
\E\Big\|\int_E e^{2\pi i uF(t)}d\theta(t)\Big\|_{L^{2}(\R^d)}^{2}
\end{equation}
is calculated to study the measure of $F(E).$ Also in Chapter 17 of the same work, to compute the Fourier dimension of the Brownian image $W(E)$ of a set $E$, decay estimates as $u\rightarrow \infty$ are obtained for
\begin{equation}\label{int-3}
\Big|\int_0^\infty e^{2\pi i uW(t)}d\theta(t)\Big|,
\end{equation}
where $\theta$ is a positive measure on the positive half line. Estimates similar to \eqref{int-2},\eqref{int-3} went on to be employed by researchers studying Gaussian Fourier series and Brownian images ever since, see \cite{eh,fmd,fs, gg,g,mu,x} and references therein. Note that in \eqref{int-2} and especially in \eqref{int-3} by choosing $\theta$ to be a sum of Dirac delta measures we can obtain a random exponential sum similar to those we consider in this work. Though we should also remark that owing to the purpose for which these estimates are utilized, $\theta$ in its original contexts is generally considered to be a measure supported on a larger set, so as to enable decay, whereas sums in this article are always periodic. Hence randomization of frequency by stochastic processes not only has certain advantages peculiar to itself, but is also very natural with a long and established history.
Frequency randomization by stochastic processes brings two difficulties that do not exist in coefficient randomization. The first is that some frequencies will repeat, therefore the Fourier series structure is violated and certain basic tools like the Plancherel theorem cannot be directly applied. The second is, the random variables $X_j$ obtained from a process will not be independent for many important classes of processes, the most important for us being the Poisson process. We however surmount these difficulties by concentrating on differences of the form $X_{j_2}-X_{j_1}, \ X_{k_2}-X_{k_1} $ which are independent for any independent increment process, such as Poisson and Wiener processes, whenever $(j_1,j_2)\cap (k_1,k_2)$=$\emptyset$. We delicately decompose our sums so as to obtain such independent differences whenever possible. When this is not possible, we use combinatorial and arithmetic counting arguments to count the number of intersections, and if possible reduce them to the non-intersecting case.
$L^{2n}, n\in \N$ norms of exponential sums correspond to solutions of diophantine equations: letting $A\in \Z$ be a finite set, and ${\bf j,k}\in A^n$ with ${\bf j}=(j_1,j_2,\ldots,j_n), \ {\bf k}=(k_1,k_2,\ldots,k_n)$
\begin{align}\label{int1}
\Big\|\sum_{j\in A} e^{2\pi i yj}\Big\|_{L^{2n}(\T)}^{2n} &= \big| \big\{({\bf j,k})\in A^{2n}:\ \ \sum_{i=1}^n j_i= \sum_{i=1}^nk_i \ \ \big\}\big| \\&= \sum_{m\in \Z} \big| \big\{ {\bf j}\in A^{n}:\ \sum_{i=1}^n j_i=m \big\}\big|^2 \label{int2}
\end{align}
where $|\cdot|$ denotes cardinality. Estimates on the size of the sets in \eqref{int1},\eqref{int2} in terms of $|A|$ have always been of great interest to mathematicians. We will briefly summarize the essentials we need of this broad subject, for much more detail see \cite{bhb,hb,hb2,mar,s,sw,v} and references therein. Sets in \eqref{int1} always include $\sim n! |A|^n$ trivial solutions coming from choosing ${\bf k}$ a permutation of ${\bf j}$. So the problem of estimating their size reduces to estimating the cardinality of nontrivial solutions. For certain $A,n$ these nontrivial solutions may not even exist. The sets in \eqref{int2} are representations of $m$ by elements of $A^n$, and we will denote their cardinality by $R_{A^n}(m).$ Certain of these sets may be empty, but by permutations, if a set has one element ${\bf j}$ all permutations of it are also inside.
Most often $A$ is chosen to be $d$-powers of the set $\{1,2,3\ldots,M \}$, and an estimation in terms of $M$ is demanded. In this case, the set of \eqref{int1} turns into:
\begin{equation}\label{int3}
\big| \big\{({\bf j,k}): \ 1\leq j_i,k_i\leq M, \quad j_1^d+j_2^d+\cdots +j_n^d= k_1^d+k_2^d+\cdots +k_n^d \big\}\big|,
\end{equation}
The sets in \eqref{int2}, the cardinalities of which in this particular case we will denote by $R_{n,d,M}(m)$, are representations of $m$ by $n$ $d$-powers:
\begin{equation}\label{int4}
R_{n,d,M}(m):= \big| \big\{{\bf j}: \ 1\leq j_i\leq M, \quad j_1^d+j_2^d+\cdots +j_n^d= m \big\}\big|.
\end{equation}
Plainly these problems can also be viewed as finding lattice points on a variety in affine or projective space.
The cases of small $n$, such as $n=2,3,4$ have been studied even more intensely.
To have an opinion on the size of the sets in \eqref{int3},\eqref{int4} there is the following well known heuristic argument. $A^n$ has $M^n$ elements, but any sum $ j_1^d+j_2^d+\cdots +j_n^d$ is an integer between $1$ and $nM^d.$ Therefore, on average there are $n^{-1}M^{n-d}$ representations any integer $1\leq m \leq nM^d $. So, for $d\geq n$ it is reasonable to expect that $R_{n,d,M}(m)\lesssim 1$ for any $m$, and that trivial solutions dominate in \eqref{int3}, which is referred to as paucity of nontrivial or nondiagonal solutions. This heuristics is behind the Hardy-Littlewood Hypothesis $K$, which is essentially the claim
$R_{n,n,M}(K)=\mathcal{O}(K^{\varepsilon})$. This conjecture is well known to be true for $n=2$, but was disproved by Mahler \cite{m} for $n=3.$ Yet it may still be true for various other values of $n.$ There is also the conjecture, relying on the same heuristics, that the cardinality of the set in \eqref{int3} is bounded by $C_{\varepsilon}M^{\varepsilon}\max\{M^n,M^{2n-d}\},$
see \cite{v}. For $n=2$ this is known, but even for $n=3$ it is open, and a topic of great interest. Indeed great effort \cite{bhb,hb,s} has been directed to just obtaining bounds $M^{7/2-\delta}$ for $d$ high enough. The bound $M^{3+\varepsilon}$ has been achieved for $d>25$ by Salberger in \cite{s}, and no reduction in $d$ has been possible in the intervening two decades. For higher values of $n$ even less is known, see \cite{mar,sw}. In \cite{sw} Salberger and Wooley prove the only result in the literature for generic $n$, and establish the paucity of nontrivial solutions, and hence the conjecture, for $d\geq (2n)^{4n}.$
If we take $d$-powers of an arbitrary finite set $A$, these problems become much harder, and even the $n=2$ case is wide open. In this case Chang \cite{c} conjectures that
\begin{equation}\label{chang}
\Big\|\sum_{j\in A} e^{2\pi i yj^2}\Big\|_{L^{4}(\T)}^{4}\leq C_{\varepsilon}|A|^{2+\varepsilon}.
\end{equation}
This conjecture is closely related to many other problems of harmonic analysis, additive combinatorics and number theory, see \cite{c,cg, san} for a broad view of these connections.
The bound $|A|^3$ is trivial for \eqref{chang}, and there has been no power type improvement on this, see \cite{c,san} for the improvements that have been possible. Randomized versions of this conjecture will be some of our main objects of study in this work, and we will confirm Chang's conjecture on average, and uncover similar behaviour for any $d,n.$
Another old and esteemed problem on exponential sums is the Hardy-Littlewood majorant problem which concerns precisely non-even $L^p$ norms. Let $A\subseteq \{1,2,\ldots,N\}$. Then for $|a_j|\leq 1, j\in A $ and $p=2n$ we have
\begin{equation*}\label{int5}
\Big\|\sum_{j\in A} a_je^{2\pi i yj}\Big\|_{L^{p}(\T)}\leq \Big\|\sum_{j\in A} e^{2\pi i yj}\Big\|_{L^{p}(\T)}.
\end{equation*}
Hardy and Littlewood asked what happens for arbitrary $p>2.$ To make things clearer we define the constants
\begin{equation*}
\sup_{|a_j|\leq 1 } \Big\|\sum_{j\in A} a_je^{2\pi i yj}\Big\|_{L^{p}(\T)}= B_p(A) \Big\|\sum_{j\in A} e^{2\pi i yj}\Big\|_{L^{p}(\T)},
\end{equation*}
and
\begin{equation*}
B_p(M):=\sup_{A\subseteq \{1,2,\ldots,M\}}B_p(A).
\end{equation*}
We note that $B_p(M)$ is an increasing quantity. Already Hardy and Littlewood knew that $B_3(A)>1$ for certain sets $A$. So it is not reasonable to expect in general what is true for even exponents. But it may still be expected to have $B_p(M)$ bounded by a constant independent of $M$. Bachelis \cite{bac} disproved this, and showed that for any $p$ that is not an even integer $B_p(M)\rightarrow \infty$ as $M\rightarrow \infty$. After this for a time it was hoped that $B_p(M)\leq C_{\varepsilon}M^{\varepsilon}$ may be possible. But the works of Green-Rusza \cite{gr} and Mockenhaupt-Schlag \cite{ms} disproved even this. Thus there is no hope for results that will hold for all sets, but for specific subsets of $\N$ positive results are still possible, and important. Indeed it is known by the work of Mockenhaupt \cite{moc} that for certain sets $\Gamma\subseteq \{1,2,\ldots,M\}$ obtained by discretizing and projecting the sphere $S^{n-1}$ the bound
$B_p(\Gamma)\leq C_{\varepsilon}M^{\varepsilon}$ would imply the famous restriction conjecture in harmonic analysis. For the restriction conjecture, its connections and variants see \cite{dem,ft1,ft2,ft3}. So studying the Hardy-Littlewood majorant problem for specific subsets is very important and continues to this day, see for example \cite{d,gre,kmt,kre}. This is also very meaningful, for Mockenhaupt-Schlag in an article \cite{ms} that served as the starting point of our investigations proved that generic sets satisfy the Hardy-Littlewood majorant property. In this they built upon arguments of Bourgain \cite{bou1}. More precisely let $\xi_j$ be i.i.d selector variables with distribution
$\Pp[\xi_j=1]=M^{-\delta}$ and $\Pp[\xi_j=0]=1-M^{-\delta}$ where $0<\delta<1.$ Then for any $\varepsilon>0$, as $M\rightarrow \infty$
\begin{equation}\label{intms}
\Pp\Big[ \sup_{|a_j|\leq 1}\Big\| \sum_{j=1}^{M} a_j\xi_je^{2\pi i yj} \Big\|_{L^p(\T)} \geq M^{\varepsilon} \Big\| \sum_{j=1}^{M} \xi_je^{2\pi i yj} \Big\|_{L^p(\T)} \Big]\rightarrow 0.
\end{equation}
One area of study that is very much connected with the Hardy-Littlewood majorant problem is the $\Lambda(p)$ sets. A set $A\subseteq \Z $ is a $\Lambda(p),\ p>2 $ set, if for any exponential sum on it
\begin{equation*}
\Big\|\sum_{j\in A} a_je^{2\pi i yj}\Big\|_{L^{p}(\T)} \leq C_{A,p} \Big\|\sum_{j\in A} a_je^{2\pi i yj}\Big\|_{L^{2}(\T)} = C_{A,p}\Big[ \sum_{j\in A} |a_j|^2\Big]^{1/2}.
\end{equation*}
Plainly this is true for any finite set, so this concept is nontrivial for infinite sets $A$ only. Now suppose $A$ is a $\Lambda(p)$ set. Then for coefficients $a_j$ of size at most $1$,
\begin{align*}
\Big\|\sum_{j\in A} a_je^{2\pi i yj}\Big\|_{L^{p}(\T)} \leq C_{A,p} \Big\|\sum_{j\in A} a_je^{2\pi i yj}\Big\|_{L^{2}(\T)}
\leq C_{A,p} \Big\|\sum_{j\in A} e^{2\pi i yj}\Big\|_{L^{2}(\T)}\leq C_{A,p} \Big\|\sum_{j\in A} e^{2\pi i yj}\Big\|_{L^{p}(\T)},
\end{align*}
showing that this set has the Hardy-Littlewood majorant property with $B_p(A)\leq C_{A,p}$. Bourgain \cite{bou1} proved that a generic subset of $\{1,2,\ldots,M\}$ of size $[M^{\delta}], \ 0<\delta \leq 2/p$ is a $\Lambda(p)$ set, and hence satisfies the Hardy-Littlewood majorant property. In \cite{bou2} he extended this result to generic subsets of powers and primes. Genericity meant in these works is similar to \eqref{intms}. See also the very recent work \cite{djr} extending and furthering this line of research.
Lastly, we remind the important conjecture of Rudin \cite{ru} that squares is a $\Lambda(p)$ set for $p<4$. This conjecture, if true, would imply Chang's conjecture.
Below we introduce our results, and the
following notation will help us articulate them. The notation $a\lesssim b$ means there exists a positive constant $C$ such that $a\leq Cb$, and $a\gtrsim b$ is defined analogously. We denote $a\approx b$ if both of these are true. For two functions $f,g$ the notation $f\sim g$ means $f(x)/g(x)\rightarrow 1 $ as $x\rightarrow \infty.$ We will also employ the standard asymptotic notation $ \Oh, o,\Omega $. We use $\E$ to denote the expectation and $\V$ for variance. We let $\Z_+$ denote the nonnegative integers, and $\R_+$ the nonnegative real numbers.
\begin{theorem}
Let $\{X_j\}_{j\in \mathbb{N}}$ be a stationary process taking only integer values. Let the probability mass function of the random variables in our process be denoted by $\mu:\Z\rightarrow \R.$ Let $A\subset \N$ be any finite nonempty subset. Then for any $1\leq p<\infty$ and any $\varepsilon>0$ we have
\begin{equation}\label{s1}
|A|^p \lesssim_{\mu,p} \mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{p}^p \leq |A|^p,
\end{equation}
\begin{equation}\label{s2}
\mathbb{E}\sup_{|a_j|\leq 1}\Big\| \sum_{j\in A} a_je^{2\pi i yX_j} \Big\|_{p}^p \lesssim_{\mu,p} \mathbb{E}\Big\| \sum_{j\in A} e^{2\pi i yX_j} \Big\|_{p}^p,
\end{equation}
\begin{equation}\label{s3}
\lim_{|A|\rightarrow \infty}\mathbb{P}\Big[\sup_{|a_j|\leq 1} \Big\| \sum_{j\in A} a_je^{2\pi i yX_j} \Big\|_{p} \geq |A|^{\varepsilon}\Big\| \sum_{j\in A} e^{2\pi i yX_j} \Big\|_{p} \Big]=0.
\end{equation}
Analogues of these results for $p=\infty$ also hold.
\end{theorem}
Heuristically a process being stationary means it repeats the same values with the same probabilities. Thus the exponential sum in \eqref{s1} may be expected to behave like an $|A|$-fold sum of the same exponential $e^{2\pi i yX_j}.$ Our result \eqref{s1} confirms this heuristic, and it is strong enough to yield \eqref{s2},\eqref{s3} immediately. As seen in there \eqref{s1} is about the $L^p$ norms of random exponential sums, \eqref{s2} is the Hardy-Littlewood majorant property on average, while \eqref{s3} is the Hardy-Littlewood majorant property for generic sets as in the contexts of Bourgain and Schlag-Mockenhaupt. This theorem is obtained using only basic techniques of harmonic analysis and probability, and in its proof we illustrate these techniques.
Our second result gives similar estimates for the case of the Poisson process with the same techniques. As its proof is relatively similar, it will only be summarized, with differences especially indicated. Recall that for $1<p<\infty$
\begin{equation}\label{int6}
\Big\|\sum_{j=1}^M e^{2\pi i yj}\Big\|_{p}^p\approx_p M^{p-1}.
\end{equation}
Let us see our second theorem.
\begin{theorem}
Let $\{N(t)\}_{t\geq 0}$ be a Poisson process of intensity $1$. Let $\{a_j\}_{j\in \mathbb{N}}$ be any complex sequence with $|a_j|\leq 1, \ j\in \mathbb{N} $. Then for any $2\leq p< \infty$ we have
\begin{equation}\label{p1}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yN(j)}\Big\|_{p}^p \approx_p M^{p-1}.
\end{equation}
Analogues of \eqref{s2},\eqref{s3} hold as well. Analogues of these results for $p=\infty$ also hold.
\end{theorem}
Heuristics behind this is that $N(j)\sim j$ with high probility: this follows from the well known facts $\E[N(j)]=\V[N(j)]=j$. So \eqref{p1} confirms this heuristics, and gives the same estimate as in \eqref{int6}. Again this estimate is strong enough to yield the other results of the theorem.
We can prove a similar theorem for the simple random walk, by essentially the same arguments. One difficulty is that the simple random walk is not increasing, which can be overcome by the Doob martingale inequality. One missing point of both Theorem 2 and Theorem 3 compared to Theorem 1 is that Theorem 1 is proved for an arbitrary subset $A\subseteq \{1,2,\ldots, M\}$ while the other two theorems are only valid for the whole set $\{1,2,\ldots,M\}$. It would be very nice to prove these theorems for arbitrary subsets as well, but this seems to be rather difficult. Another point is that we cede the range $1\leq p<2.$
\begin{theorem}
Let $\{R(j)\}_{j\in\Z_+}$ be a simple random walk. Then for any $2\leq p< \infty$ we have
\begin{equation}\label{r1}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yR(j)}\Big\|_{p}^p \approx_p M^{p-1/2}.
\end{equation}
Analogues of \eqref{s2},\eqref{s3} hold as well. Analogues of these results for $p=\infty$ also hold.
\end{theorem}
From the facts $\E[|R(j)|] \approx \sqrt{j}$, and $\V[R(j)]=j$ we may heuristically say $|R(j)|\approx\sqrt{j}$ with high probability, and from this guess \eqref{r1}. We can summarize \eqref{s1},\eqref{p1},\eqref{r1} heuristically as follows. If for a process $X(j)$ we have $X(j)\approx j^{1-\alpha}, \ 0\leq \alpha \leq 1$ with high probability, then we may see this as a value being repeated $M^{\alpha}$ times, and write
\begin{equation}\label{h1}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yX(j)}\Big\|_{p}^p \approx \mathbb{E}\Big\|M^{\alpha}\sum_{j=1}^{M^{1-\alpha}} e^{2\pi i yj}\Big\|_{p}^p \approx M^{p\alpha+(p-1)(1-\alpha)}= M^{p-1+\alpha}.
\end{equation}
This is a very useful heuristic to guess what happens for other processes.
We then move on to more difficult problems. In these we take the stochastic process to be Poisson, and investigate subsets of powers, subsets of what we may call equally spaced powers, and the sets introduced in Green-Ruzsa work \cite{gr} that we will call the Green-Ruzsa sets. We start with $L^4$ estimates for powers, as these illustrate our ideas in a simpler context.
\begin{theorem}
Let $\{N(t)\}_{t\geq 0}$ be a Poisson process of intensity $1$. Let $d\geq 2$ be an integer. Let $A\subset \N$ be a finite nonempty subset. Then for any $2\leq p\leq 4$ we have
\begin{equation}\label{pp1}
|A|^{p/2} \leq \mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{p}^p \lesssim_p \begin{cases}|A|^{p/2}\log^{3p/8}(1+|A|) \ &\text{for} \ d=2 \\ |A|^{p/2} \ &\text{for} \ d\geq 3. \end{cases}
\end{equation}
Analogues of \eqref{s2},\eqref{s3} hold for $2\leq p\leq 4$.
\end{theorem}
Note that the theorem is proved for an arbitrary $A\subseteq \{1,2,\ldots,M\},$ and therefore proves that Mei-Chu Chang's conjecture is correct in an average sense. The reason why this very difficult conjecture becomes reachable with randomization is that randomizing is much like averaging, that is, rather than considering just one $j^d$ we consider $N(j^d)$ which heuristically takes every integer value in $(j^d-j^{d/2},j^d+j^{d/2})$ each with $j^{-d/2}$ probability. This gives a gain of $j^{-d/2}$ in the sum we want to evaluate. Combined with a good counting argument for lattice points in shells this suffices to obtain our result.
In this theorem proving \eqref{pp1} for $p=4$ is the key step, all other claims of the theorem then follow. As carrying out this proof we will demonstrate delicate decompositons, combinatorial arguments, counting arguments and finally the connection to lattice point counting in shell type regions that we will develop further to prove subsequent theorems.
The next theorem is for $L^4$ norms of equally spaced powers, which we study mainly to compare and contrast with Theorem 4.
\begin{theorem}
Let $\{N(t)\}_{t\geq 0}$ be a Poisson process of intensity $1$. Let $r>0$ be a real number. Then
\begin{equation*}\label{lss}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yN(jM^r)}\Big\|_{2}^2 \approx M, \qquad M^{2}+M^{3-r} \lesssim \mathbb{E} \Big\|\sum_{j=1}^M e^{2\pi i yN(jM^r)}\Big\|_{4}^4 \lesssim_r M^{2}\log M+M^{3-r}.
\end{equation*}
\end{theorem}
For $r<1$, plainly $M^{3-r}$ dominates, and the logarithmic loss is unimportant. For $r>1$ it should be possible with sharper methods we used for other theorems to remove the logarithmic loss. For $r=1$ it might not be possible to remove it, for it is felt even in heuristic computations.
One issue that this theorem highlights is how randomization eliminates arithmetic progression structure so that instead of having $M^3$ for the $L^4$ estimate we have $M^{3-r}$. Also proving this theorem is somewhat easier than proving \eqref{pp1}, for the arithmetic of dealing with these equally spaced powers is easier.
Green and Ruzsa \cite{gr} introduced the following sets for which the Hardy-Littlewood majorant property fails polynomially.
For any integer $D\geq 5$ and $ k\in \N $ the Green-Ruzsa set is
\begin{equation*}
\Lambda_{D,k}:= \Big\{\sum_{j=0}^{k-1}d_j D^j \ \Big| \ d_j\in \{0,1,3\} \Big\}.
\end{equation*}
It would be interesting to see what happens if we randomize these sets via the Poisson process, and then look at the Hardy-Littlewood property. Our next theorem studies this and provides the answer that Hardy-Littlewood property holds almost surely. Indeed this follows merely from the sparsity of these sets, so our theorem is stated for a very large class of sparse sets.
\begin{theorem}
Let $\{N(t)\}_{t\geq 0}$ be a Poisson process of intensity $1$. Let $A\subset \Z_+ $ be a finite nonempty subset such that for any $M\in \N, n\in \Z$
\begin{equation*}
\big|A \cap [n-M,n+M]\big| \leq C_AM^{\alpha},
\end{equation*}
with $C_A$ a constant that depends only on $A$. If $\alpha <1/3$, then for any $2\leq p\leq 4$, and any $\varepsilon>0$ we have
\begin{equation}\label{gr1}
|A|^{p/2}\leq \mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j)}\Big\|_{p}^p \lesssim_{\alpha,C_A} |A|^{p/2},
\end{equation}
\begin{equation}\label{gr2}
\mathbb{E}\sup_{|a_j|\leq 1} \Big\| \sum_{j\in A} a_je^{2\pi i yN(j)} \Big\|_{p}^p \lesssim_{\alpha,C_A} \E \Big\| \sum_{j\in A} e^{2\pi i yN(j)} \Big\|_{p}^p,
\end{equation}
\begin{equation}\label{gr3}
\lim_{|A|\rightarrow \infty}\mathbb{P}\Big[\sup_{|a_j|\leq 1} \Big\| \sum_{j\in A} a_je^{2\pi i yN(j)} \Big\|_{p}^p \geq |A|^{\varepsilon}\Big\| \sum_{j\in A} e^{2\pi i yN(j)} \Big\|_{p}^p \Big]=0.
\end{equation}
where the limit is taken over $A$ for which $\sup_A C_A<\infty$ and $\sup_A \alpha<1/3$.
\end{theorem}
So randomization destroys the specific, fractal-like structure of the Green-Ruzsa sets that violates the Hardy-Littlewood property.
We will show in the pertinent section that the Green-Ruzsa sets satisfy the sparsity property demanded by the theorem. In the proof instead of using the arithmetic structure of the set as in the proof of Theorem 4 we use its sparsity.
We then arrive at the generic case $p=2n$. In order to prove this case we make our arguments abstract and streamlined.
\begin{theorem}
Let $\{N(t)\}_{t\geq 0}$ be a Poisson process of intensity $1$. Let $d\geq n\in \N$ be an integer. Let $A\subset \N$ be a finite nonempty subset. Then we have
\begin{equation}\label{p2n1}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{2n}^{2n} \lesssim_{d,n} \begin{cases}\max\big\{|A|^n, |A|^{2n-\frac{d}{2}-1}\log^{1+\frac{1}{d}}(1+ |A|) \big\} &\text{if} \quad d\equiv 2 \pmod 4, \\
\max\big\{|A|^n, |A|^{2n-\frac{d}{2}-1}\log^{\frac{1}{d}}(1+ |A|)\big\} &\text{else}.
\end{cases}
\end{aligned}
\end{equation}
\end{theorem}
In comparison, the deterministic results for this general case of Salberger and Wooley \cite{sw} obtain the bound $|A|^{n+\varepsilon}$ for $d\geq (2n)^{4n}$, while in our randomized setting we obtain it as soon as $d\geq 2n-2$. For the specific $n=3$ case we obtain the $|A|^{3+\varepsilon}$ bound as soon as $d>3$, while in Salberger \cite{s} this is $d>25.$ In \cite{s} improvement over even the $7/2$ exponent takes place only for $d>8.$ Also our set $A$ is arbitrary, and not just natural numbers up to $M$.
We now would like to mention the connection that is made to well-known problems on lattice points in regions. After long and arduous decompositions and dealing with auxiliary sums, the main sums of Theorems 4,7 reduce to a lattice point problem of the following form
\begin{equation}\label{ss2}
\sup_{D\leq E\leq D^2} \big|\big\{ (j,k) \in \N^2: j<k, \ |k^d-j^d-E| <D \big\}\big|
\end{equation}
for large $E,D$.
This estimate can be obtained either directly, or by viewing this set as a margin of the set
\begin{equation}\label{os}
\big\{ (j,k) \in \N^2: \ 0<k^d-j^d <D \big\}.
\end{equation}
But estimates on sets of this type do exist in number theory literature. Indeed they are some of the oldest and most famous problems in number theory. We give a brief background on these, and then state the result to be used to deduce an estimate on \eqref{ss2}.
We start with the Dirichlet divisor problem. For $x\geq 1$ the divisor summatory function $D(x)$ is defined by
\begin{equation*}
D(x):=\sum_{n\leq x}d(n)=|\{ (a,b)\in \N^2: ab\leq x \}|,
\end{equation*}
where $d(n)$ is the number of positive divisors of $n.$ Plainly this is a lattice point problem. Dirichlet \cite{dir} proved that
\begin{equation*}
D(x)=x\log x +(2\gamma-1)x+\Delta(x), \qquad \Delta(x)= \mathcal{O}(\sqrt{x}),
\end{equation*}
where $\gamma=0.577\ldots$ is the Euler-Mascheroni constant.
The Dirichlet divisor problem is then to study the error term $\Delta(x).$
Tremendous effort has been poured into this problem for two centuries. Dirichlet's result can be obtained by carefully rearranging the lattice point sum. Voronoi \cite{vor} obtained $\mathcal{O}(x^{1/3}\log x)$ in 1904. G.H. Hardy \cite{gh1,gh2} showed that $\Delta (x)=\mathcal{O}(x^\theta)$ is not possible for $\theta<1/4$. The exponent $1/3$ is important in that it can reached by relatively easy methods, and yet is very difficult to significantly improve. Indeed after a century of effort by many illustrious mathematicians, mainly using exponetial sum methods, the current best result is due to
M. N. Huxley \cite{hux3} proving $\Delta (x)=\mathcal{O}(x^{\theta}\log^{\eta} x)$
with $\theta={131}/{416},$ and $\eta=1+{18627}/{8320}.$
When $d=2$, the linear transformation $(a,b)=T(j,k)=(k-j,k+j)$ makes estimating \eqref{os} essentially equivalent to the Dirichlet divisor problem, see \cite{khl1}. For $d\geq 3$ and $x>0$ the sets
\begin{equation*}
R_d(x):=|\{ (j,k)\in \Z^2: 0<|k|^d-|j|^d\leq x \}|,
\end{equation*}
have been studied as higher power analogues of the Dirichlet divisor problem, and of course \eqref{os} can be estimated from these via symmetry.
We have by Kratzel \cite{krat2} and Nowak \cite{nowak1,nowak2}
\begin{equation}\label{R-}
R_d(x)=A_d x^{2/d}+B_d x^{1/(d-1)}+D_dF_d(x^{1/d})x^{1/d-1/d^2}+ \Delta_d(x),
\end{equation}
where the first term comes from the area, the second from the length of the boundary, and their coefficients are explicitly known constants that depend only on $d$.
The $x^{1/d-1/d^2}$ term emerges from the zero curvature points of the curve, and these do not exist when $d=2$. The coefficient $D_d$ is another explicitly known constant depending only on $d$, however $F_d$ is a bounded function of $x$
\begin{equation*}
F_d(x)=\sum_{n=1}^{\infty}n^{-1-1/d}\sin\Big(2\pi nx+\frac{\pi}{2d}\Big), \quad |F_d(x)|\leq\sum_{n=1}^{\infty}n^{-1-1/d}.
\end{equation*}
Thus the asymptotic expansion for $R_d(x)$ cannot be developed beyond this term, and this term is $\Oh(x^{1/d-1/d^2})$. The error term $\Delta_d(x)$ is analogous to the error term of the Dirichlet divisor problem, and
applying the results of Huxley \cite{hux1,hux3}, Nowak \cite{nowak1,nowak2} obtained
\begin{equation}\label{nwkr}
\Delta_d(x) =\Oh(x^{\frac{46}{73}\frac{1}{d}+\varepsilon}), \qquad
\Delta_d(x) =\Omega(x^{\frac{1}{2d}}).
\end{equation}
With the help of these results on $R_d(x)$ we can deduce the next theorem on \eqref{ss2}. But we will also show that this theorem can be proven directly without recourse to viewing the set as a margin.
\begin{theorem}\label{at1}
Let $E,D\geq 1$ be real numbers, and let $d\geq 3$ be an integer. Then
\begin{equation}\label{p6l21}
\sup_{D\leq E\leq D^2} \big|\big\{ (j,k) \in \N^2: j<k, \ |k^d-j^d-E| <D \big\}\big|\leq C_d D^{2/d}.
\end{equation}
\end{theorem}
From the literature on lattice point problems, see for example \cite{gk}, we know that bounds of order $\Oh(x^{\frac{1}{d}})$ are reachable without any recourse to exponential sums, merely by carefully decomposing the region, and judiciously choosing the direction of the lines with which we count the lattice points. We will show that the same is true for our Theorem 8. However the proof is considerably long and difficult.
Let us also look at a special case of this. In this case fixing $E=D^s, \ 1<s\leq 2$ gives some gain in the exponent.
\begin{theorem}\label{at2}
Let $D\geq 1$ be a real number, and $1<s\leq 2$. Let $d\geq 3$ be an integer. Then
\begin{equation}\label{p6sl3}
\big|\big\{ (j,k) \in \N^2: j<k, \ |k^d-j^d-D^s| <D \big\}\big|\leq \begin{cases} C_d D^{1+s(\frac{2}{d}-1)} \ \ &\text{if} \ \ 1<s\leq \frac{d^2}{d^2-d-1}, \\ C_d D^{\frac{s}{d}(1-\frac{1}{d})} \ \ &\text{if} \ \ \frac{d^2}{d^2-d-1} \leq s\leq 2. \end{cases}
\end{equation}
\end{theorem}
The arithmetic problems mentioned here concern regions bounded by hyperbolic curves, and they have elliptic analogues in the Gauss circle problem and its higher power variants. Essentially the same methods apply to both class of problems. Also analogous problems are studied in $\R^n$ for any $n\geq2$. For excellent expositions of all aspects of this topic see the books \cite{gk,hux2,krat} and the survey \cite{ikkn}.
Our investigations in this article usher new directions of research, and raise many further questions.
For example it would be interesting to prove analogues of Theorems 2,3 for arbitrary sets, or to prove Theorem 7 for processes with less variation. But the most interesting question is passing from these random sums to deterministic sums. Observe that
\begin{equation*}
\begin{aligned}
{\bf I}= \Big\|\sum_{j\in A} e^{2\pi i yj^d}\Big\|_{2n}^{2n}\lesssim \mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{2n}^{2n} + \mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yj^d}-e^{2\pi i yN(j^d)}\Big\|_{2n}^{2n}={\bf II+III}.
\end{aligned}
\end{equation*}
Since ${\bf III}\lesssim {\bf I}+{\bf II}$ is plain, our results on ${\bf II}$ imply that ${\bf III}$ is dominated by the conjectural bounds on ${\bf I}.$ But is it any easier to demonstrate
these for ${\bf III}$ than for ${\bf I}$?
The rest of the article is structured as follows. We first give a preliminaries section in which we introduce the stochastic processes we use, and prove various lemmas on their properties that will be needed later on. Then we prove Theorems 1-7 each in its own section. Our final section will prove the last two theorems on arithmetic.
%==================================================
%====================================================
%===================================================
%====================================================
%======================================================
\section{Preliminaries}
We start with briefly introducing the three types of processes that we use. As we will make frequent use of their definitions and basic properties, it will be useful to the reader to have them under hand. We assume all processes in the article to be integer valued, ensuring periodicity for the exponential sums. A process $\{X_t\}_{t\in I}$, where $I$ is an index set, is stationary if joint cumulative distribution functions remain the same under time translation. This of course means that each $X_t$ has the same distribution, and this will be the only property that we will use to prove Theorem 1.
A random variable $N$ has Poisson distribution if there exist $\mu>0$ such that for any nonnegative integer $k$
\begin{equation}\label{pd}
\Pp[N=k]=e^{-\mu}\frac{\mu^k}{k!}.
\end{equation}
Then $N$ has mean and variance $\mu$.
A process $\{N(t)\}_{t\geq 0}$ is a Poisson process with intensity $\lambda$ if three conditions hold:
{\bf 1.} $N(0)=0$ almost surely.
{\bf 2.} $N(t)$ has independent increments, that is for $0\leq t_1\leq t_2\leq \ldots\leq t_k$ the random variables $N(t_1),N(t_2)-N(t_1),\ldots, N(t_k)-N(t_{k-1})$ are independent.
{\bf 3.} For any $0 \leq s < t < \infty$ the random variable $N(t)-N(s)$ is a Poisson random variable
with mean $\lambda(t - s).$\\
In this article we will always take the intensity $\lambda$ to be 1.
Let $X_j, j\in \N$ be independent identically distributed Bernoulli variables with $\Pp[X_j=1]=\Pp[X_j=-1]=1/2$. Then simple random walk $\{R(n)\}_{n\in\Z_+}$ is the process given by
\[ R(n)=\sum_{j=1}^n X_j \]
with the convention that $R(0)=0.$
This is a mean zero process with $\V[R(n)]=n.$
For more information on these processes see any basic probability book such as \cite{ks}.
Our lemmas on the Poisson process aim to make rigorous the heuristic that for an intensity 1 Poisson process $\{N(t)\}_{t\geq 0}$, the random variable $N(j)$ only takes the integer values between $j-\sqrt{j},j+\sqrt{j}$ each with equal probability. With the lemma below, we will carry out the first part of this, that is $N(j)$ only takes the integer values between $j-\sqrt{j},j+\sqrt{j}$. Essentially this is because the Poisson process is an independent increment process, which leads to Gaussian-like concentration around the mean. As such similar results can be proved for more general classes of processes. More words about this after the proof.
\begin{lemma}\label{lem1}
Let $\{N(t)\}_{t\geq 0}$ be a Poisson process of intensity 1, and let $m\in \mathbb{N}$. Let $0<\lambda \leq \sqrt{m}.$ Then
\begin{equation*}
\mathbb{P}[|N(m)-m|>\lambda \sqrt{m}]\leq 2e^{-\lambda^2/4}.
\end{equation*}
\end{lemma}
\begin{proof}
We start with
\begin{equation*}
\begin{aligned}
\mathbb{P}[|N(m)-m|>\lambda \sqrt{m}]=\mathbb{P}[N(m)-m>\lambda \sqrt{m}]+\mathbb{P}[m-N(m)>\lambda \sqrt{m}]= {\bf I}+{\bf II}.
\end{aligned}
\end{equation*}
We first deal with {\bf I}. Let $t>0$.
\begin{equation*}
\begin{aligned}
{\bf I}=\mathbb{P}[t(N(m)-m)>t\lambda \sqrt{m}]=\mathbb{P}[e^{t(N(m)-m)}>e^{t\lambda \sqrt{m}}] &=\mathbb{P}[e^{t(N(m)-m-\lambda \sqrt{m})}>1]\\ &\leq \mathbb{E}[e^{t(N(m)-m-\lambda \sqrt{m})}] \\ &= e^{-t\lambda \sqrt{m}-tm}\mathbb{E}[e^{tN(m)}].
\end{aligned}
\end{equation*}
The last expectation is of course the moment generating function(MGF) of $N(m)$, and can easily be calculated from \eqref{pd} to be $e^{m(e^t-1)}.$ Notice then that this is just the $m$th power of the MGF of $N(1)$, and this is precisely because of the independent increment property.
Observe that for $0<t\leq 1$
\begin{equation*}
\begin{aligned}
e^t=1+t+\frac{t^2}{2}\Big[1 +\frac{t}{3}+\frac{t^2}{4\cdot 3}+\cdots \Big]\leq 1+t+\frac{t^2}{2}\Big[1 +\frac{1}{3}+\frac{1}{ 3^2}+\cdots \Big] \leq 1+t+\frac{3}{4}t^2.
\end{aligned}
\end{equation*}
Plugging in the MGF of $N(m)$ and then using this yields
\begin{equation*}
{\bf I}\leq e^{-t\lambda\sqrt{m}+m(e^t-t-1)}\leq e^{-t\lambda\sqrt{m} +\frac{3}{4}mt^2 },\qquad \qquad 0<t\leq 1.
\end{equation*}
Here choosing $t=\lambda/\sqrt{m}$ bounds ${\bf I}$ by $e^{-\lambda^2/4}.$
Exactly the same steps give
\begin{equation*}
\begin{aligned}
{\bf II}\leq e^{-t\lambda\sqrt{m}+m(e^{-t}-1+t)}.
\end{aligned}
\end{equation*}
Observe that for $0<t\leq 1$
\begin{equation*}
\begin{aligned}
e^t=1-t+\frac{t^2}{2}-\Big[\frac{t^3}{3!}-\frac{t^4}{4!}\Big]-\Big[\frac{t^5}{5!}-\frac{t^6}{6!}\Big]\cdots \leq 1-t+\frac{t^2}{2},
\end{aligned}
\end{equation*}
as each of the parantheses above is positive.
From this
\begin{equation*}
\begin{aligned}
{\bf II}\leq e^{-t\lambda\sqrt{m}+m\frac{t^2}{2}}.
\end{aligned}
\end{equation*}
Choosing $t=\lambda/\sqrt{m}$ gives the bound ${\bf II}\leq e^{-\lambda^2/2}$, finishing the proof.
\end{proof}
If a random variable $X$ for some positive constant $K_1$ satisfies
\[ \Pp[|X|\geq t] \leq 2e^{-t^2/K^2_1}\]
for all $t\geq 0$, it is called subgaussian. An equivalent definition is that $X$ is subgaussian if
\[ \E[e^{tX}] \leq e^{K^2_2t^2}\]
for all $t\in \R$ and for some constant $K_2.$ So our random variables $N(m)-m$ behave like subgaussians, but strictly speaking they are not subgaussians as is plain from their moment generating functions $e^{m(e^t-t-1)}$. But for a restricted range of $t$ we see the relevant inequalities satisfied. In \cite{cck} the authors coined the term \textquotesingle locally subgaussian\textquotesingle \ for such random variables.
The next two lemmas handle the rest of the heuristic we presented at the start of the section. They complement each other nicely.
\begin{lemma}\label{l3}
Let $a\in \N$. Then
\begin{equation*}
\sup_{t\geq 0} \Pp[N(t)=a]\leq \frac{1}{\sqrt{2\pi a}}.
\end{equation*}
\end{lemma}
\begin{proof}
We have
\begin{equation*}
\sup_{t\geq 0}\Pp[N(t)=a]= \sup_{t\geq 0} e^{-t}\frac{t^a}{a!}=\frac{1}{a!}\sup_{t\geq 0} t^ae^{-t}.
\end{equation*}
Plainly we are seeking the supremum of a smooth nonnegative function on $[0,\infty)$, and this function vanishes at the endpoints of this interval. So to find the supremum we differentiate and set it equal to zero.
\begin{equation*}
0=\frac{d}{dt} \big(t^ae^{-t}\big)=at^{a-1}e^{-t}-t^ae^{-t}=t^{a-1}e^{-t}(a-t).
\end{equation*}
Thus supremum is attained at $t=a,$ and is $a^ae^{-a}.$
A precise version of Stirling's formula due to Robbins \cite{rob} states that for any $n\in \N$
\begin{equation}\label{robs}
\frac{1}{\sqrt{2\pi n}}e^{-\frac{1}{12n}}\leq \frac{1}{n!}\cdot \frac{n^n}{e^n}\leq \frac{1}{\sqrt{2\pi n}}e^{-\frac{1}{12n+1}}\leq \frac{1}{\sqrt{2\pi n}}.
\end{equation}
This concludes the proof.
\end{proof}
\begin{lemma}\label{l4}
Let $t\geq 0$. Then
\begin{equation*}
\sup_{a\in \mathbb{Z}_+}\Pp[N(t)=a]=\Pp[N(t)=\lfloor t\rfloor ]\leq \min\big\{1,\frac{1}{\sqrt{2\pi \lfloor t\rfloor} }\big\}.
\end{equation*}
\end{lemma}
\begin{proof}
We observe that as
\[ \Pp[N(t)=a]=e^{-t}\frac{t^a}{a!} \]
the supremum over $a$ depends only on the fraction $t^a/a!$ As
\[\frac{t^{a+1}}{a+1!}=\frac{t}{a+1}\frac{t^a}{a!},\]
as long as $a+1\leq \lfloor t\rfloor $ this fraction increases, and then for $a+1>\lfloor t\rfloor $ it decreases. That is, the last multiplication must be by $t/\lfloor t\rfloor$, with which we obtain $t^{\lfloor t\rfloor }/\lfloor t\rfloor !$. On finding this we apply Lemma \ref{l3} to conclude the proof.
\end{proof}
To calculate probabilities and reduce to a lattice point problem in the general $p=2n$ case we need a number of lemmas.
We start with the following well known fact on independent random variables. For a proof see Sinai-Koralov \cite{ks}.
\begin{lemma}
Let $X_1,\ldots, X_n$ be independent random variables, $m_1+\ldots+m_k =
n$ and $f_1,\ldots, f_k$ be Borel measurable functions of $m_1, ...,m_k$ variables respectively.
Then the random variables $ f_1(X_1, ..., X_{m_1} ), f_2(X_{m_1+1}, \ldots, X_{m_1+m_2} ),\ldots,
f_k(X_{m_1+\ldots+m_{k-1}+1},\ldots, X_n)$ are independent.
\end{lemma}
Our next lemma is crucial for calculating probabilities. With this lemma we separate the random variables into two groups such that the sums over these groups are independent. After this independence is utilized next lemmas
can be brought in to calculate probabilities.
\begin{lemma}
Let $\{X_t\}_{t\geq 0}$ be an independent increment process. Let $n\geq 2$ be an integer, and let $(j_i,k_i), \ 1\leq i\leq n$ be nonempty open intervals. Let $\sigma_1\cup\sigma_2$ be a partition of indices $\{1,2,\dots,n\}$ into two nonempty subsets. Suppose no interval indexed by $\sigma_1$ intersects the intervals indexed by $\sigma_2$ and vice versa. Then the random variables
\begin{equation*}
\sum_{i\in \sigma_1} X_{k_i}-X_{j_i}, \qquad \sum_{i\in \sigma_2} X_{k_i}-X_{j_i}
\end{equation*}
are independent.
\end{lemma}
\begin{proof}
Let us list $j_i,k_i, \ i\in \sigma _1$ with respect to size, and let $t_1\leq t_2\leq\ldots\leq t_{2|\sigma_1|}$ be this list. For the intervals $(t_i,t_{i+1}), \ 1\leq i \leq 2|\sigma_1|-1,$ let $ i\in \eta_1\subseteq\{1,2,\ldots,2|\sigma_1|-1 \}$
be those indices for which $(t_i,t_{i+1})$ lies inside one of $(j_i,k_i), \ i\in \sigma_1$. Then observe that
\begin{equation*}
\sum_{i\in \sigma_1} X_{k_i}-X_{j_i}= \sum_{i\in \eta_1}d_i(X_{t_{i+1}}-X_{t_{i}})
\end{equation*}
for some natural numbers $d_i.$ We apply this same procedure to indices in $\sigma_2$ to obtain intervals $(t_i,t_{i+1}), \ i\in \eta_2$. Then we observe that the intervals indexed by $\eta_1,\eta_2$ are all disjoint. Therefore bearing in mind the independent increment property we can apply Lemma 4 to obtain our claim.
\end{proof}
The next two lemmas allow explicit computation of probabilities in Theorem 7, as we will need an analogue of Lemmas 2,3 for a linear combination of Poisson random variables. Proofs will follow from the independence of these variables and Lemma 3.
\begin{lemma}\label{l5}
Let $N_i$ be independent random variables of Poisson distribution with mean $\mu_i> 0$, and $d_i\in \N$ where $1\leq i\leq n.$ Let $\mu:=\max_{1\leq i\leq n}\mu_i$. Then for $a\in \Z_+$
\begin{equation*}
\Pp\big[\sum_{i=1}^n d_iN_i=a \big]\leq \min\big\{1, \frac{1}{\sqrt{2\pi\lfloor\mu \rfloor}} \big\}.
\end{equation*}
\end{lemma}
\begin{proof}
Without loss of generality assume $\mu_n=\mu$. We observe
\begin{equation*}
\Pp\big[\sum_{i=1}^n d_iN_i=a \big]= \sum_{\substack{a_1+a_2+\ldots+a_n=a \\ a_i\in \Z_+}}\Pp\big[ N_i=a_i/d_i, \ 1\leq i\leq n \big]=\sum_{\substack{a_1+a_2+\ldots+a_n=a \\ a_i\in \Z_+}}\prod_{i=1}^n \Pp\big[ N_i=a_i/d_i] .
\end{equation*}
Now we can use the assumption that $\mu_n=\mu$ together with Lemma 3 and reduce the number of factors in the product by one.
\begin{equation*}
\begin{aligned}
\leq\min\big\{1, \frac{1}{\sqrt{2\pi\lfloor\mu \rfloor}} \big\}\sum_{\substack{a_1+a_2+\ldots+a_n=a \\ a_i\in \Z_+}}\prod_{i=1}^{n-1} \Pp\big[ N_i=a_i/d_i].
\end{aligned}
\end{equation*}
This opens up the possibility of summing over $a_i,\ 1\leq i\leq n$.
\begin{equation*}
\begin{aligned}
\leq \min\big\{1, \frac{1}{\sqrt{2\pi\lfloor\mu \rfloor}} \big\} \sum_{\substack{a_i=0 \\ 1\leq i\leq n-1 }}^a \prod_{i=1}^{n-1} \Pp\big[ N_i=a_i/d_i] \leq
\min\big\{1, \frac{1}{\sqrt{2\pi\lfloor\mu \rfloor}} \big\}\prod_{i=1}^{n-1} \sum_{a_i=0 }^a \Pp\big[ N_i=a_i/d_i].
\end{aligned}
\end{equation*}
As each of the sums in the last term are at most 1, this concludes the proof.
\end{proof}
This more general lemma will be used in the following specific form in proving Theorem 7.
\begin{lemma}
Let $(j_i,k_i), \ 1\leq i\leq n$ be nonempty open intervals, and let $\{N(t)\}_{t\geq0}$ be a Poisson process of intensity 1. Let $m$ be an index at which $k_i-j_i$ becomes maximum. Then for any $a\in \Z_+$
\begin{equation*}
\Pp\big[ \sum_{i=1}^n N(k_i)-N(j_i)=a \big]\leq \min\big\{1,\frac{1}{\sqrt{2\pi\lfloor (k_m-j_m)/2n\rfloor}}\big\}.
\end{equation*}
\end{lemma}
\begin{proof}
We apply the procedure described in Lemma 5 to obtain the disjoint intervals $(t_i,t_{i+1}), \ i\in \eta$ for a set of indices $ \eta \subseteq\{1,2,\ldots2n-1 \}$ that enable writing
\begin{equation*}
\sum_{i=1}^{n} N(k_i)-N(j_i)= \sum_{i\in \eta}d_i[N(t_{i+1})-N(t_{i})]
\end{equation*}
for some natural numbers $d_i.$ Since the union of $(t_i,t_{i+1}), i\in \eta$ is the same as the union of $(j_i,k_i), \ 1\leq i\leq n$, at least one $(t_i,t_{i+1}), i \in \eta$ has length at least $(k_m-j_m)/2n.$
An application of Lemma 6 concludes the proof.
\end{proof}
The next two lemmas are technical and elementary in nature.
They will be needed at certain steps of the proofs of Theorems 4,7.
\begin{lemma}\label{dl}
Let $A\subseteq \N$ be a finite set, and let $a\in \R, \ b\in \N$ satisfy $0<a\leq b\leq \min A$. Let $d\geq 2$ be an integer. Let $\Phi:[a,\infty)\rightarrow (0,\infty)$ be a decreasing function. Then
\begin{equation*}
\sum_{\substack{j,k\in A \\ j<k } }\Phi(k^d-j^d) \leq \sum_{\substack{b\leq j<k\leq b+ |A|-1 } }\Phi(k^d-j^d).
\end{equation*}
\end{lemma}
\begin{proof}
We list the elements of $A$ into a strictly increasing sequence $a_1,a_2,\ldots,a_{|A|}.$
Then for any $1\leq j<k\leq |A|$ we have
$a_{j+1}-a_j\geq 1$, implying by telescoping $a_k-a_j\geq k-j.$ Choosing $j=1$, and bearing in mind that $a_1=\min A$, we obtain $a_k\geq b+k-1.$ Combining these
\begin{equation*}
\begin{aligned}
a_k^d-a_j^d=\big(a_k-a_j\big)\sum_{i=0}^{d-1}a_k^{d-1-i}a_j^{i} &\geq \big(k-j\big)\sum_{i=0}^{d-1}(b+k-1)^{d-1-i}(b+j-1)^{i}\\ &=(b+k-1)^d-(b+j-1)^d.
\end{aligned}
\end{equation*}
Since $d\geq 2$, the last term is not less than $2b$, so both the first and last term lies in the domain of $\Phi.$ As $\Phi$ is decreasing, this allows us to conclude
\begin{equation*}
\begin{aligned}
\sum_{\substack{j,k\in A \\ j<k } }\Phi(k^d-j^d)= \sum_{\substack{1\leq j,k\leq |A| \\ j<k } }\Phi(a_k^d-a_j^d)&\leq \sum_{\substack{1\leq j<k\leq |A| } }\Phi((b+k-1)^d-(b+j-1)^d) \\ &=\sum_{\substack{b\leq j<k\leq b+|A|-1 } }\Phi(k^d-j^d).
\end{aligned}
\end{equation*}
\end{proof}
\begin{lemma}\label{lem2}
Let $1\leq C\leq 10$.
{\bf a.} If $x>e^{50}$, then
\begin{equation*}
\begin{aligned}
|x-y|\leq C\sqrt{x\log x} \ \ \implies \ \ |x-y|\leq 2C\sqrt{y\log y}.
\end{aligned}
\end{equation*}
{\bf b.} If $y>e^{50},x\geq 1$ then
\begin{equation*}
\begin{aligned}
|x-y|\geq 2C\sqrt{x\log x} \ \ \implies \ \
|x-y|\geq {C}\sqrt{y\log y}.
\end{aligned}
\end{equation*}
\end{lemma}
\begin{proof}
To see {\bf a.} observe that $y\geq x-C\sqrt{x\log x}\geq x/2$, thus $x\leq 2y$.
As $\sqrt{x\log x}$ is an increasing function, we have $\sqrt{x\log x} \leq \sqrt{2y\log 2y}\leq 2\sqrt{y\log y}$, and we are done.
As for {\bf b.}, if $y\leq 2x$, owing to the same reason $\sqrt{y\log y} \leq \sqrt{2x\log 2x}\leq 2\sqrt{x\log x}$, and we are done. If on the other hand $y\geq 2x$, then $y-x\geq y/2$, and since
${C}\sqrt{y\log y}\leq y/2$, we are done.
\end{proof}
%=================================================================================================================================================================================================================================================================================
\section{Stationary processes}
This section is dedicated to proving Theorem 1. About stationary processes we will only use that the distribution remains the same in time. From harmonic analysis and probability we will use the techniques that are the mainstay of studying random and deterministic exponential sums.
\begin{proof}
Let the probability mass function of the random variables in our process be given by $\mu_{X_j}(k)=\mu(k)=\mu_k.$
We will see that \eqref{s2} and \eqref{s3} follow from \eqref{s1}, and for this we first need to investigate the cases $p=2,\infty$. We start with the latter. Clearly for every $\omega\in \Omega$ by considering a neighborhood of $y=0$
we have
\begin{equation}\label{sp1}
\Big\|\sum_{j\in A}e^{2\pi i yX_j(\omega)}\Big\|_{\infty}=|A|.
\end{equation}
From this all claims of the theorem for $p=\infty$ follow immediately. So let $1\leq p< \infty$. Also immediate from \eqref{sp1} are
\begin{equation}\label{sp-1}
\Big\|\sum_{j\in A}e^{2\pi i yX_j(\omega)}\Big\|_{p}\leq |A|, \qquad \E \Big\|\sum_{j\in A}e^{2\pi i yX_j}\Big\|_{p}^p\leq |A|^p.
\end{equation}
So we only need to show the reverse direction to obtain \eqref{s1}.
For $p=2$, to this end we write
\begin{equation}\label{sp2}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{2}^2 =\mathbb{E}\int_{\mathbb{T}}\Big|\sum_{j\in A} e^{2\pi i yX_j}\Big|^2dy&=\mathbb{E}\int_{\mathbb{T}}\sum_{j,k\in A} e^{2\pi i y(X_k-X_j)}dy\\ &= \mathbb{E}\sum_{k\in A}\sum_{j\in A}\int_{\mathbb{T}} e^{2\pi i y(X_k-X_j)}dy.
\end{aligned}
\end{equation}
For a fixed $\omega$ the integral in the last expression is $1$ if $X_j(\omega)=X_k(\omega)$ and $0$ otherwise. We define
\begin{equation*}
U(\omega,k):=|\{ j\in A: X_j(\omega)=k\}|=\sum_{j\in A} \mathbb{I}_{\{X_j=k\}}(\omega),
\end{equation*}
where $\mathbb{I}$ denotes the indicator function. For each $k$, as $U(\omega,k)$ is a finite sum of indicator functions it is measurable. Therefore we have
\begin{equation*}
\mathbb{E}U(\omega,k)=\sum_{j\in A} \mathbb{E}\ \mathbb{I}_{\{X_j=k\}}(\omega)=\sum_{j\in A} \mathbb{P}[X_j=k]=|A|\mu_k.
\end{equation*}
The Cauchy-Schwarz inequality yields $\mathbb{E}U^2(\omega,k)\geq [\mathbb{E}U(\omega,k)]^2=|A|^2 \mu_k^2$, and applying this we continue from \eqref{sp2} as follows to obtain \eqref{s1} for $p=2$
\begin{equation*}
= \mathbb{E}\sum_{j\in A} U(\omega,X_j(\omega))=\mathbb{E}\sum_{k\in \mathbb{Z}}U^2(\omega,k) =\sum_{k\in \mathbb{Z}}\mathbb{E}U^2(\omega,k)\geq |A|^2\sum_{k\in \mathbb{Z}} \mu_k^2.
\end{equation*}
For $2<p<\infty$ we utilize the $p=2,\infty$ cases:
\begin{equation*}
\mathbb{E}\Big\|\sum_{j\in A}e^{2\pi i yX_j}\Big\|_{p}^p \geq \mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{2}^p \geq \Big[ \mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{2}^2\Big]^{p/2} \geq |A|^p \Big[\sum_{k\in \mathbb{Z}} \mu_k^2\Big]^{p/2}.
\end{equation*}
For $1\leq p<2$ we again utilize the same ideas:
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{2}^2 \leq \mathbb{E}
\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{\infty}^{2-p}
\int_{\mathbb{T}}\Big|\sum_{j\in A} e^{2\pi i yX_j}\Big|^{p}dy = |A|^{2-p}\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{p}^p.
\end{aligned}
\end{equation*}
Therefore
\begin{equation*}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yX_j}\Big\|_{p}^p \geq |A|^p \sum_{k\in \mathbb{Z}} \mu_k^2 .
\end{equation*}
finishing the proof of \eqref{s1}.
To see \eqref{s2}
we just need to observe that for any $\omega$ and any sequence in $\{a_j\}_{j\in A}, \ |a_j|\leq 1$
\begin{equation*}
\Big\|\sum_{j\in A}a_j e^{2\pi i yX_j(\omega)}\Big\|_{\infty}\leq |A|, \qquad
\text{and} \qquad
\Big\|\sum_{j\in A}a_j e^{2\pi i yX_j(\omega)}\Big\|_{p}^p\leq |A|^p.
\end{equation*}
We use this to conclude that for $1\leq p<\infty$
\begin{equation*}\label{}
\mathbb{E}\sup_{ |a_j|\leq 1}\Big\|\sum_{j\in A}a_j e^{2\pi i yX_j}\Big\|_{p}^p\leq |A|^p, \ \ \ \ \ \ \mathbb{E}\sup_{ |a_j|\leq 1}\Big\|\sum_{j\in A} a_j e^{2\pi i yX_j}\Big\|_{\infty}\leq |A|.
\end{equation*}
Then using \eqref{s1} we obtain \eqref{s2}.
To handle \eqref{s3} we need to introduce another method. Let $\epsilon >0$, and let $K$ be such that
\begin{equation*}
\sum_{|k|\leq K} \mu_k >1-\epsilon.
\end{equation*}
Therefore
\begin{equation*}
\mathbb{E}\sum_{j\in A}\mathbb{I}_{\{|X_j|\leq K\}}\geq (1-\epsilon)|A| .
\end{equation*}
Then for the set
\begin{equation*}
\Omega_{\epsilon}:=\{\omega\in \Omega: \sum_{j\in A}\mathbb{I}_{\{|X_j|\leq K\}} \geq \frac{9}{10}|A| \}
\end{equation*}
we have $\mathbb{P}(\Omega_{\epsilon})\geq 1-10\epsilon$. We observe that for $\omega\in \Omega_{\epsilon}$ and $0 \leq y\leq 1/100K$ we have
\begin{equation*}
\begin{aligned}
\Big|\sum_{j\in A}e^{2\pi i yX_j(\omega)}\Big|&\geq \Big|\sum_{ |X_j(\omega)|\leq K} e^{2\pi i yX_j(\omega)}\Big|-\Big|\sum_{|X_j(\omega)|> K} e^{2\pi i yX_j(\omega)}\Big|\\ &\geq \Big|\sum_{ |X_j(\omega)|\leq K} \cos{2\pi yX_j(\omega)}\Big|-\frac{|A|}{10}\\ &\geq \frac{9|A|}{20}-\frac{|A|}{10} >\frac{|A|}{4}.
\end{aligned}
\end{equation*}
For any $\omega\in \Omega_{\epsilon}$, and any $1\leq p< \infty$ we apply this last result to obtain
\begin{equation*}\label{}
\Big\|\sum_{j\in A}e^{2\pi i yX_j(\omega)}\Big\|_{p}^p\geq \int_{0}^{{1}/{100K}}\Big|\sum_{j\in A} e^{2\pi i yX_j(\omega)}\Big|^pdy \geq \frac{|A|^p}{4^p100K}.
\end{equation*}
Thus we understand that
\begin{equation*}
\mathbb{P}\Big[ \Big\|\sum_{j\in A} e^{2\pi i yX_j(\omega)}\Big\|_{p}^p \geq \frac{|A|^p}{4^p100K} \Big]\geq 1-10\epsilon.
\end{equation*}
Therefore for $|A|$ larger than a constant depending only on $\epsilon,\varepsilon, \mu,p$
\begin{equation*}
\mathbb{P}\Big[|A|^{\varepsilon} \Big\|\sum_{j\in A} e^{2\pi i yX_j(\omega)}\Big\|_{p}^p \geq |A|^{p+\varepsilon/2} >\sup_{S_A} \Big\|\sum_{j\in A} a_j e^{2\pi i yX_j(\omega)}\Big\|_{p}^p \Big]\geq 1-10\epsilon.
\end{equation*}
Taking limits and considering that $\epsilon$ is arbitrary we obtain our result.
\end{proof}
With the extra assumption that $X_j$ are independent random variables it would be easier to compute \eqref{sp2}. In that case we would bring the expectation inside, and then compute the probabilities exactly. In proving some of our other theorems this idea will be used in a modified form.
%=============================================
%=================================================
%====================================================
%============
%===============================================
\section{Poisson processes}
In this section we prove an analogue of Theorem 1 for Poisson processes. Unfortunately in this case we have to consider the set $A$ to be of interval form, for the analogue of \eqref{s1} in this case does not hold, that is the left hand side depends not only on the cardinality but also on the structure of the set. Proving analogues of \eqref{s2},\eqref{s3} without using an estimate like \eqref{s1} is very difficult. As the proof is similar to that of Theorem 1, we present it briefly.
\begin{proof}
As before, the analogues of \eqref{s2} and \eqref{s3} will follow from \eqref{p1}, and for this we first need to investigate the cases $p=2,\infty$. We start with the latter. Clearly for every $\omega\in \Omega$ by considering a neighborhood of $y=0$
we have
\begin{equation*}\label{pp11}
\Big\|\sum_{j=1}^M e^{2\pi i yN(j)(\omega)}\Big\|_{\infty}=M.
\end{equation*}
From this all claims of the theorem for $p=\infty$ follow immediately. So we assume $2\leq p< \infty$.
For \eqref{p1} with $p=2$ we compute
\begin{equation*}\label{}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yN(j)(\omega)}\Big\|_{2}^2= \mathbb{E} \sum_{j,k=1}^M \int_{\mathbb{T}} e^{2\pi i y[N(j)(\omega)-N(k)(\omega)]}dy&= \sum_{j,k=1}^M\mathbb{P}[N(|j-k|)=0]
\\ &= \sum_{j,k=1}^Me^{-|j-k|},
\end{aligned}
\end{equation*}
and
\begin{equation*}
M\leq \sum_{j,k=1}^Me^{-|j-k|}=M+ 2\sum_{j=1}^{M-1}(M-j)e^{-j}\leq M+ 2M\sum_{j=1}^{M-1}e^{-j}\leq 3M.
\end{equation*}
This finishes \eqref{p1} for $p=2$. Then we extend this to $p>2$. For any sequence $\{a_j\}_{j=1}^M, \ |a_j|\leq 1 $
\begin{equation}\label{pp2}
\begin{aligned}
\Big\|\sum_{j=1}^M a_j e^{2\pi i yN(j)(\omega)}\Big\|_{2}^2 \leq \Big\|\sum_{j=1}^M e^{2\pi i yN(j)(\omega)}\Big\|_{2}^2,
\end{aligned}
\end{equation}
therefore
\begin{equation}\label{pp-1}
\begin{aligned}
\E\sup_{|a_j|\leq 1}\Big\|\sum_{j=1}^M a_je^{2\pi i yN(j)}\Big\|_{p}^p \leq \E
\sup_{|a_j|\leq 1}\Big\|\sum_{j=1}^M a_je^{2\pi i yN(j)}\Big\|_{\infty}^{p-2}
\int_{\mathbb{T}}\Big|\sum_{j=1}^M a_je^{2\pi i yN(j)}\Big|^{2}dy\leq 3M^{p-1}.
\end{aligned}
\end{equation}
Let $A_M:=\{N(M)>5M\}$. By Lemma 1, $\mathbb{P}[A_M] \leq 2e^{-4M}$. Therefore for $\omega\notin A_M$
\begin{equation*}
\begin{aligned}
\Big\|\sum_{j=1}^M e^{2\pi i yN(j)(\omega)}\Big\|_{p}^p \geq \int_{0}^{1/100M}\Big|\sum_{j=1}^M \cos 2\pi yN(j)(\omega)\Big|^pdy \geq
\frac{M^{p-1}}{100\cdot 2^p}.
\end{aligned}
\end{equation*}
Therefore
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yN(j)}\Big\|_{p}^p &\geq
\frac{M^{p-1}}{100\cdot 2^{p+1}}.
\end{aligned}
\end{equation*}
This concludes \eqref{p1}, and the analogue of \eqref{s2} for any $p\geq 2$.
On the other hand by \eqref{pp-1}
\begin{equation*}
\mathbb{P}\Big[\sup_{|a_j|\leq 1} \Big\|\sum_{j=1}^M a_je^{2\pi i yN(j)(\omega)}\Big\|_{p}^p \geq 3M^{p-1}\log M \Big]\leq 1/\log M.
\end{equation*}
Therefore except for a set of measure at most $1/\log M+2e^{-4M}$
\begin{equation*}
\begin{aligned}
\sup_{|a_j|\leq 1} \Big\|\sum_{j=1}^M a_je^{2\pi i yN(j)(\omega)}\Big\|_{p}^p \leq 300\cdot 2^p\log M\Big\|\sum_{j=1}^M e^{2\pi i yN(j)(\omega)}\Big\|_{p}^p,
\end{aligned}
\end{equation*}
and this proves the analogue of \eqref{s3}.
\end{proof}
%========================================0
%=========================================
%=================================================
%===============================================
%===================================================
\section{Random walks}
This section is dedicated to proving Theorem 3. The proof is broadly similar to that of Theorem 2, and the most important difference is that the simple random walk is not increasing as opposed to the Poisson process. This is relevant in obtaining lower bounds by considering a neighborhood of the origin, and will be dealt with via Doob's martingale inequality. The other differences are computational.
\begin{proof}
As before we will see the analogues of \eqref{s2} and \eqref{s3} follow from \eqref{r1}, and for this we need to investigate the cases $p=2,\infty$. We start with the latter. Clearly for every $\omega\in \Omega$ by considering a neighborhood of $y=0$
we have
\begin{equation}\label{pp12}
\Big\|\sum_{j=1}^M e^{2\pi i yR(j)(\omega)}\Big\|_{\infty}=M.
\end{equation}
From this all claims of the theorem for $p=\infty$ follow immediately. Therefore we assume $2\leq p< \infty$.
For any sequence $\{a_j\}_{j=1}^M, \ |a_j|\leq 1, $
\begin{equation}\label{rw2}
\begin{aligned}
\Big\|\sum_{j=1}^M a_j e^{2\pi i yR(j)(\omega)}\Big\|_{2}^2 \leq \Big\|\sum_{j=1}^M e^{2\pi i yR(j)(\omega)}\Big\|_{2}^2.
\end{aligned}
\end{equation}
We have
\begin{equation*}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yR(j)}\Big\|_{2}^2 =\sum_{j,k=1}^M\mathbb{P}[R(|j-k|)=0]= M+2\sum_{1\leq j<k\leq M}^M\mathbb{P}[R(k-j)=0].
\end{equation*}
Bearing in mind that the simple random walk can be zero only at even times, the last sum equals
\begin{equation*}
\sum_{j=1}^{\lfloor(M-1)/2\rfloor} (M-2j) \mathbb{P}[R(2j)=0] =
\sum_{j=1}^{\lfloor(M-1)/2\rfloor} (M-2j) \binom {2j}j 2^{-2j},
\end{equation*}
and this can be estimated via Robbins's Stirling formula \eqref{robs}
\begin{equation*}
\begin{aligned}
\leq \sum_{j=1}^{\lfloor(M-1)/2\rfloor} (M-2j) \frac{1}{\sqrt{\pi j}} \leq M+\int_{1}^{M/2} \frac{M-2x}{\sqrt{x}}dx\leq 2M^{3/2}.
\end{aligned}
\end{equation*}
Combining this estimate with \eqref{pp12},\eqref{rw2} we obtain for $2\leq p<\infty$,
\begin{equation}\label{rw-1}
\begin{aligned}
\E\sup_{ |a_j|\leq 1}\Big\|\sum_{j=1}^M a_je^{2\pi i yR(j)}\Big\|_{p}^p &\leq \E \sup_{ |a_j|\leq 1}
\Big\|\sum_{j=1}^M a_je^{2\pi i yR(j)}\Big\|_{\infty}^{p-2}\Big\|\sum_{j=1}^M a_je^{2\pi i yR(j)}\Big\|_{2}^2
\\ &\leq M^{p-2} \E
\Big\|\sum_{j=1}^M e^{2\pi i yR(j)}\Big\|_{2}^2 \leq 2M^{p-\frac{1}{2}}.
\end{aligned}
\end{equation}
We now want to obtain a lower bound. As opposed to the Poisson process, the simple random walk is not increasing, therefore it is not enough to consider $R(M)$. We turn around this difficulty using the fact that the random walk is a martingale. By Doob's martingale inequality
\begin{equation*}
\mathbb{E}\sup_{1\leq j \leq M}|R(j)|^2 \leq 4\mathbb{E}|R(M)|^2 =4M.
\end{equation*}
Therefore with probability at least 3/4 we have
$\sup_{1\leq j \leq M}|R(j)| \leq 4M^{1/2},$
and for such $\omega$
\begin{equation*}
\begin{aligned}
\Big\|\sum_{j=1}^M e^{2\pi i yR(j)(\omega)}\Big\|_{p}^p \geq \int_{0}^{1/200M^{1/2}}\Big|\sum_{j=1}^M \cos 2\pi yR(j)(\omega)\Big|^pdy \geq
\frac{M^{p-1/2}}{200\cdot 2^p}.
\end{aligned}
\end{equation*}
Therefore
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yR(j)}\Big\|_{p}^p &\geq
\frac{M^{p-1/2}}{100\cdot 2^{p+2}}.
\end{aligned}
\end{equation*}
This concludes \eqref{r1} and the analogue of \eqref{s2}. Assuming $M$ to be large, and repeating this Chebyshev type argument we see that except for a set of probability at most
$1/\log^2M$ we have
\begin{equation*}
\begin{aligned}
\Big\|\sum_{j=1}^M e^{2\pi i yR(j)(\omega)}\Big\|_{p}^p \geq
\frac{M^{p-1/2}}{200\cdot 2^p \log M}.
\end{aligned}
\end{equation*}
On the other hand from \eqref{rw-1}
\begin{equation*}
\mathbb{P}\Big[\sup_{|a_j|\leq 1}\Big\|\sum_{j=1}^M a_je^{2\pi i yR(j)(\omega)}\Big\|_{p}^p \geq 2M^{p-1/2}\log M \Big]\leq \frac{1}{\log M}.
\end{equation*}
Therefore for a set of probability at least $1-2/ \log M$
\begin{equation*}
\begin{aligned}
100\cdot 2^{p+2}\log^2 M\Big\|\sum_{j=1}^M e^{2\pi i yR(j)(\omega)}\Big\|_{p}^p \geq \sup_{|a_j|\leq 1} \Big\|\sum_{j=1}^M a_je^{2\pi i yR(j)(\omega)}\Big\|_{p}^p,
\end{aligned}
\end{equation*}
and this proves the analogue of \eqref{s3}.
\end{proof}
%================================================\\
%=========================================
%===========================================
%================================================
\section{Perturbations of powers I}
This section is dedicated to proving Theorem 4. This theorem is a specific case of Theorem 7. Our aim in proving this case separately is to illustrate our ideas in the simplest and most concrete case possible. Even in this case the proof is rather long and complicated. So briefly lying out our plan will be helpful for the reader.
The results of the theorem will follow once we prove \eqref{pp1} for $p=2,4.$ The case $p=2$ has already been done for $d=1$ in the proof of Theorem 2, and the higher $d$ cases are similar. Main issue is to prove the $p=4$ case. In this case there will be 4 variables $j_1,j_2,k_1,k_2$ instead of just 2, and this makes calculating probabilities difficult. Our strategy is to partition our sum, and first remove the pieces that in essence do not involve 4 different variables. Then we also remove the cases like $j_1<k_1,j_2<k_2$ that can only hold with small probability. Then we partition our sum into two cases in which intervals $(j_1,k_1),(k_2,j_2)$ are disjoint, or intersect. The first case by independence will enable calculating probabilities, and after removing tail estimates coming from the Poisson process, we will reduce to a lattice point counting problem. Finally there will remain the case of intersecting intervals, and we will show that this can be reduced to the case of nonintersecting intervals by removing the intersection.
\begin{proof}
We start with showing \eqref{pp1} for $p=2.$
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{2}^2 = \mathbb{E} \sum_{j,k\in A} \int_{\mathbb{T}} e^{2\pi i y[N(j^d)-N(k^d)]}dy= \sum_{j,k\in A}\mathbb{P}[N(j^d)=N(k^d)]
= \sum_{j,k\in A}e^{-|j^d-k^d|}.
\end{aligned}
\end{equation*}
We estimate this last term as
\begin{equation}\label{pp4}
\begin{aligned}
|A|\leq \sum_{j,k\in A}e^{-|j^d-k^d|}= |A|+ 2\sum_{k\in A}\sum_{\substack {j\in A\\ j<k}}e^{j^d-k^d}&\leq |A|+ 2\sum_{k\in A}\sum_{j=1}^{(k-1)^d}e^{-k^d+j} \\
& \leq |A|+ 4\sum_{j\in A}e^{-d(j-1)^{d-1}-1},
\end{aligned}
\end{equation}
and this final term is bounded by $ |A| +4.$ This concludes \eqref{pp1} for $p=2.$
We now move on to $p=4$, after which other results of the theorem will follow. The left hand side of \eqref{pp1} follows immediately from the $p=2$ case, and the Hölder inequality. So we will concentrate on the right hand side. We start with some reductions that will help us later on. If $|A|\leq 3e^{100}$, then our result is immediate. If $|A|>3e^{100}$ then we let $A_0:=A\cap[1, e^{100}]$ . We obtain
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{4}^4 &\leq 8\Big[\mathbb{E}\Big\|\sum_{j\in A_0} e^{2\pi i yN(j^d)}\Big\|_{4}^4 +\mathbb{E}\Big\|\sum_{j\in A\setminus A_0} e^{2\pi i yN(j^d)}\Big\|_{4}^4\Big].
\end{aligned}
\end{equation*}
The first term in the paranthesis can be bounded by $|A_0|^4$. If we had our result for sets that contain no element in $[1, e^{100}]$, and have more elements than $e^{100}$, then applying it would bound the second term by $C|A\setminus A_0|^2.$ Thus the whole right hand side would be bounded by $8(C+e^{200})|A|^2.$
So we may assume that $A$ contains no element in $[1, e^{100}]$ and has more elements than $e^{100}$. Here we observed this reduction for $d>2$, but the same reduction is possible for $d=2$ as well.
We start with transforming our sum
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{4}^4 &= \mathbb{E} \sum_{j_1,j_2,k_1,k_2\in A} \int_{\mathbb{T}} e^{2\pi i y[N(j_1^d)+N(j_2^d)-N(k_1^d)-N(k_2^d)]}dy\\ &= \sum_{j_1,j_2,k_1,k_2\in A}\mathbb{P}[N(j_1^d)+N(j_2^d)=N(k_1^d)+N(k_2^d)].
\end{aligned}
\end{equation*}
From this last sum we want to remove the cases when two of these variables are equal. These give rise to auxiliary sums that we deal with before going to the main sum. The set $A^4$ to which any quadruple $(j_1,j_2,k_1,k_2)$ belongs can be written as
\[A^4=A_{11}\cup A_{12}\cup A_{21}\cup A_{22}\cup A_3 . \]
Here $A_{ab}, \ a,b\in \{1,2\}$ is the set of quadruples $(j_1,j_2,k_1,k_2)$
for which $j_a=k_b.$ The set $A_3$ contains the remaining elements of $A^4$.
We have
\begin{equation*}
\begin{aligned}
\sum_{ (j_1,j_2,k_1,k_2)\in A_{11}}\mathbb{P}[N(j_1^d)+N(j_2^d)=N(k_1^d)+N(k_2^d)]&= \sum_{ j_1,j_2,k_2\in A}\mathbb{P}[N(j_2^d)=N(k_2^d)]\\ &=|A|\sum_{ j,k\in A}\mathbb{P}[N(j^d)=N(k^d)],
\end{aligned}
\end{equation*}
and this last sum has been estimated above. For $A_{12},A_{21},A_{22}$ via the same argument we get the same result. Therefore it remains to handle $A_3$. We decompose
\[A_3=A_{31}\cup A_{32}\cup A_{33} \cup A_{34}, \]
\begin{equation*}
\begin{aligned}
A_{31}&:=\{(j_1,j_2,k_1,k_2)\in A_3: j_1>k_1, j_2>k_2 \}, \quad
A_{32}:=\{(j_1,j_2,k_1,k_2)\in A_3: j_1>k_1, j_2<k_2 \},\\
A_{33}&:=\{(j_1,j_2,k_1,k_2)\in A_3: j_1<k_1, j_2>k_2 \}, \quad
A_{34}:=\{(j_1,j_2,k_1,k_2)\in A_3: j_1<k_1, j_2<k_2 \}.
\end{aligned}
\end{equation*}
Summing over $A_{31},A_{34}$ is easy, and we first handle these.
\begin{equation*}
\begin{aligned}
\sum_{ A_{31}}\mathbb{P}[N(j_1^d)+N(j_2^d)=N(k_1^d)+N(k_2^d)]&=\sum_{ A_{31}}\mathbb{P}[N(j_1^d)-N(k_1^d)=N(k_2^d)-N(j_2^d)].
\end{aligned}
\end{equation*}
As $N(t)$ is increasing we must have $N(j_1^d)-N(k_1^d)\geq 0$ and $N(k_2^d)-N(j_2^d)\leq 0$. Therefore for these to be equal they must both be zero. Hence
\begin{equation*}
\begin{aligned}
\leq \sum_{ A_{31}}\mathbb{P}[N(j_1^d)-N(k_1^d)=0]\leq |A|^2\sum_{\substack {j,k\in A\\ j<k}} \mathbb{P}[N(j^d)=N(k^d)].
\end{aligned}
\end{equation*}
We estimated this last sum above to be bounded by 2. The set $A_{34}$ is handled the same way.
We now move on to $A_{32}.$ This, together with its symmetric counterpart $A_{33}$, represent the most important, generic cases. We have
\begin{equation*}
\begin{aligned}
\sum_{ A_{32}}\mathbb{P}[N(j_1^d)+N(j_2^d)=N(k_1^d)+N(k_2^d)]&=\sum_{ A_{32}}\mathbb{P}[N(j_1^d)-N(k_1^d)=N(k_2^d)-N(j_2^d)].
\end{aligned}
\end{equation*}
We decompose $A_{32}=A_{321}\cup A_{322}$, in the first of which are contained those $(j_1,j_2,k_1,k_2)$ for which the intervals $(k_1,j_1),(j_2,k_2)$ are disjoint, and in the second those for which they intersect. We first sum over $A_{321}$ which can be written as
\begin{equation*}
\begin{aligned}
\sum_{ A_{321}} \mathbb{P}[N(j_1^d)-N(k_1^d)=N(k_2^d)-N(j_2^d)]
=&\sum_{ A_{321}}\sum_{a=0}^{\infty} \mathbb{P}[N(j_1^d)-N(k_1^d)=a=N(k_2^d)-N(j_2^d)]\\
=&\sum_{a=0}^{\infty} \sum_{ A_{321}} \mathbb{P}[N(j_1^d)-N(k_1^d)=a]\Pp[N(k_2^d)-N(j_2^d)=a].
\end{aligned}
\end{equation*}
This can be bounded by
\begin{equation*}
\begin{aligned}
\sum_{a=0}^{\infty}\sum_{ \substack{j_1,k_1 \in A \\j_1>k_1}}\sum_{\substack{j_2,k_2 \in A \\ k_2>j_2}}\mathbb{P}[N(j_1^d)-N(k_1^d)=a]\Pp[N(k_2^d)-N(j_2^d)=a]
=
\sum_{a=0}^{\infty}\Big[\sum_{ \substack{j,k\in A \\j<k}}\mathbb{P}[N(k^d-j^d)=a]\Big]^2 .
\end{aligned}
\end{equation*}
If we could show that the inner sum is bounded by a constant $C(A)$ that may depend on $A$ but is independent of $a$, then
\begin{equation*}
\begin{aligned}
\leq C(A)\sum_{a=0}^{\infty}\sum_{ \substack{j,k\in A \\j<k}}\mathbb{P}[N(k^d-j^d)=a]=C(A)\sum_{ \substack{j,k\in A \\j<k}}\sum_{a=0}^{\infty}\mathbb{P}[N(k^d-j^d)=a] \leq C(A)|A|^2.
\end{aligned}
\end{equation*}
So our aim is to obtain such an estimate.
We start with a crude estimate. For $a\in \N$ we have by Lemma \ref{l3}
\begin{equation*}
\begin{aligned}
\sum_{ \substack{j,k\in A \\j<k}} \mathbb{P}[N(k^d-j^d)=a] \leq|A|^2\sup_{t\geq 0}\Pp[N(t)=a]
\leq \frac{|A|^2}{\sqrt{2\pi a}}.
\end{aligned}
\end{equation*}
So if $a\geq |A|^4$ the inner sum is bounded by an absolute constant.
This is also true for $a=0$ as we have calculated in \eqref{pp4}. Thus we may assume $1\leq a <|A|^4.$
Let us define the function $f(x):=4 \sqrt{x\log x}$ for $x\geq 1$. We decompose the inner sum to two sums over sets $A_1,A_2$
\begin{equation*}
\begin{aligned}
A_1:=\{(j,k)\in A^2: j<k, \ |a-(k^d-j^d)|\geq 2f(a) \}, \quad A_2:=\{(j,k)\in A^2: j<k \}\setminus A_1.
\end{aligned}
\end{equation*}
We note that if $a<e^{90}$ then $A_2$ is empty, and therefore we may assume that $a$ is large when summing over that set.
We first consider the sum over $A_1,$ which is another auxiliary sum. Actually it is a tail estimate coming from the Poisson process taking every positive value with some probability, albeit small. By Lemma \ref{lem2} we have $|a-(k^d-j^d)|\geq f(k^d-j^d)$ and by Lemma \ref{lem1}
\begin{equation*}
\begin{aligned}
\mathbb{P}[N(k^d-j^d)=a]
\leq \mathbb{P}[|N(k^d-j^d)-(k^d-j^d)|\geq f(k^d-j^d)]
\leq 2(k^d-j^d)^{-4}.
\end{aligned}
\end{equation*}
Therefore
\begin{equation*}
\begin{aligned}
\sum_{A_1}\mathbb{P}[N(k^d-j^d)=a]\leq 2\sum_{\substack{j,k\in A \\ j<k}} (k^d-j^d)^{-4} \leq 2\sum_{j\in A} \sum_{k>j} k^{-4}\leq 1.
\end{aligned}
\end{equation*}
We now sum over $A_2.$ This is the main sum that will be estimated by counting lattice points.
By Lemma \ref{l3} we have
\begin{equation*}
\begin{aligned}
\sum_{ A_2}\mathbb{P}[N(j^d-k^d)=a] \leq a^{-1/2}|A_2|.
\end{aligned}
\end{equation*}
So our aim is to estimate the cardinality of $A_2$ which is a lattice point problem. We can rewrite
\begin{equation*}
\begin{aligned}
A_2= \{(j,k): j<k, \ (j^d+a-2f(a)))^{\frac{1}{d}} < k < (j^d+a+2f(a))^{\frac{1}{d}} \}.
\end{aligned}
\end{equation*}
We observe by the mean value theorem that
\begin{equation*}
\begin{aligned}
(j^d+a+2f(a))^{\frac{1}{d}}-j\leq [a+2f(a)]\frac{j^{1-d}}{d}\leq aj^{1-d}.
\end{aligned}
\end{equation*}
When $j>a^{1/(d-1)}$ this final expression is less than 1. Therefore for these $j$ we have no pair $(j,k)$ in $A_2$. Similarly
\begin{equation*}
\begin{aligned}
(j^d+a+2f(a))^{\frac{1}{d}}- (j^d+a-2f(a))^{\frac{1}{d}} \leq 4f(a) \frac{(j^d+a-2f(a))^{\frac{1}{d}-1}}{d} &\leq 4f(a)a^{\frac{1}{d}-1 }\\ &\leq 16a^{\frac{1}{d}-\frac{1}{2}}\log a.
\end{aligned}
\end{equation*}
For $d\geq 3$ this is less than one, so for each $j\leq a^{1/(d-1)}$, there can be at most 1 solution, which means $|A_2|\leq a^{1/2}.$
For $d=2$ this method of counting solutions for each fixed $j$ is too crude, and instead we will count them for each fixed value of $b=k-j$. Since when $b$ is fixed knowing $j$ immediately gives $k$, all we need to do is to count $j$.
We have $k^2-j^2=b^2+2jb$, therefore elements of $A_2$ satisfy
\begin{equation*}
\begin{aligned}
|b^2+2jb-a| < 2f(a)
\implies \frac{a}{2b}-\frac{b}{2}- \frac{f(a)}{b}&< j < \frac{a}{2b}-\frac{b}{2}+ \frac{f(a)}{b}.
\end{aligned}
\end{equation*}
So possible number of solutions $j$ for fixed $b$ is bounded by $1+2f(a)/b$,
and as $j\geq 1$, we must have
\[b^2+2jb < a+2f(a)\implies b^2<2a\implies b< \sqrt{2a}.\]
Thus summing the number of solutions over $b$
\begin{equation*}
\begin{aligned}
|A_2|\leq\sum_{b=1}^{\lfloor \sqrt{2a} \rfloor} 1+\frac{ 2f(a)}{b}\leq \sqrt{2a}+2f(a)\log a<10\sqrt{a}\log^{3/2} a.
\end{aligned}
\end{equation*}
By our assumption $a\leq |A|^4$ the sum over $A_2$ is bounded by $80\log^{3/2} |A|$. Also taking into account the contribution of $A_1$, we see that independent of $a$, the inner sum is bounded by $1+80\log^{3/2}|A|$.
This concludes the sum over $A_{321}.$
We finally consider the sum over $A_{322}$. We partition this set into four:
\begin{equation}\label{par2}
\begin{aligned}
A_{3221}&:=\{ (j_1,j_2,k_1,k_2)\in A_{322}: [k_1,j_1] \subseteq (j_2,k_2) \}\\
A_{3222}&:=\{(j_1,j_2,k_1,k_2)\in A_{322}: [j_2,k_2] \subseteq (k_1,j_1) \}\\
A_{3223}&:=\{(j_1,j_2,k_1,k_2)\in A_{322}: k_1<j_2<j_1<k_2 \}\\
A_{3224}&:=\{(j_1,j_2,k_1,k_2)\in A_{322}: j_2<k_1< k_2<j_1 \}.
\end{aligned}
\end{equation}
First two and the last two are handled in the same way so we will only consider $ A_{3221}, A_{3223}.$ On $ A_{3221}$ the condition $ [k_1,j_1] \subseteq (j_2,k_2)$ means
\begin{equation*}
\begin{aligned}
\sum_{ A_{3221}}\mathbb{P}[N(j_1^d)+N(j_2^d)=N(k_1^d)+N(k_2^d)]&=\sum_{ A_{3221}}\mathbb{P}[N(j_2^d)-N(k_1^d)=N(k_2^d)-N(j_1^d)].
\end{aligned}
\end{equation*}
Since $j_2<k_1,$ but $k_2>j_1$, and since $(j_2,k_1)\cap(j_1,k_2)=\emptyset$ the last sum is
\begin{equation*}
\begin{aligned}
=&\sum_{ A_{3221}}\mathbb{P}[N(j_2^d)-N(k_1^d)=0=N(k_2^d)-N(j_1^d)]\\ = &\sum_{ A_{3221}}\mathbb{P}[N(k_1^d)-N(j_2^d)=0]\mathbb{P}[N(k_2^d)-N(j_1^d)=0]\\
\leq &\Big[ \sum_{ k>j}\mathbb{P}[N(k^d)-N(j^d)=0] \Big]^2.
\end{aligned}
\end{equation*}
We estimated the sum inside the square by $2$. So this case is bounded by just $4$.
On $ A_{3223}$ the condition $ k_1<j_2<j_1<k_2 $ means $(k_1,j_2)\cap(j_1,k_2)=\emptyset$, and again we reduce to the case $A_{321}.$ This finishes $A_{32}$. The set $A_{33}$ is entirely symmetric to $A_{32}$, and is handled in exactly the same way. Hence we obtain \eqref{pp1} for $p=4$. By the Hölder inequality we obtain \eqref{pp1}. Then arguments similar to those in the previous proofs yield analogues of \eqref{s2},\eqref{s3}.
\end{proof}
In this proof we made every decomposition explicitly. Of course this will not be possible for the general case in Theorem 7, as it would result in too many cases to deal with. We will introduce new ideas to handle this situation. Also lattice point counting problem is harder in that case, and will be handled with Theorem 8.
%==================================================
%====================================================
%===================================================
%====================================================
%======================================================
\section{Arithmetic progressions of larger step size}
This section pertains to Theorem 5. In this theorem
our aim is to see how randomization affects an exponential sum with frequencies forming an arithmetic progression when the step size is large. Our Theorem 2 shows that it does not have much of an effect when the step size is $1$. But as the step size grows we will observe the decay induced. This decay also means the destruction of arithmetic progression structure. It is also instructive to compare this theorem to Theorem 4. Picking $r=d-1$ we obtain two sequences $\{j^d\}, \{jM^{r}\}, \ 1\leq j\leq M$ that grow roughly at the same pace, but for the first one growth starts slowly and then gains pace, while for the second it is uniform. Comparing the two theorems illuminates the effects that these differences between sequences have.
\begin{proof}
We first look at the
$p=2$ case.
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yN(jM^r)}\Big\|_{2}^2 = \sum_{j,k=1 }^Me^{-|j-k|M^r}= M+ 2\sum_{k=2}^Me^{-kM^{r}} \sum_{j< k }e^{jM^{r}}&\leq M+4(M-1)e^{-M^r} \\ &\leq M+2\min\{ M, C_r\},
\end{aligned}
\end{equation*}
where $C_r$ is a constant that depends only on $r.$
Armed with these we move on to the case $p=4$. We first show the right hand side.
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j=1}^M e^{2\pi i yN(jM^r)}\Big\|_{4}^4 = \sum_{j_1,j_2,k_1,k_2=1}^M \mathbb{P}[N(j_1M^r)+N(j_2M^r)=N(k_1M^r)+N(k_2M^r)].
\end{aligned}
\end{equation*}
Let $A:=\{1,2,3\ldots, M\}$, and we decompose the set $A^4$ to which any quadruple $(j_1,j_2,k_1,k_2)$ belongs as in the proof of Theorem 4,
\[A^4=A_{11}\cup A_{12}\cup A_{21}\cup A_{22}\cup A_3, \]
and all except $A_3$ are dealt with easily as before. For example on $A_{11}$
\begin{equation*}
\begin{aligned}
\sum_{ A_{11}}\mathbb{P}[N(j_2M^r)=N(k_2M^r)]=M\sum_{ j,k=1}^M\mathbb{P}[N(|j-k|M^r)=0] = M\sum_{j,k=1 }^Me^{-|j-k|M^r},
\end{aligned}
\end{equation*}
which as above gives a $\approx M^2$ term. Each of the $A_{ij}, \ i,j=1,2$ gives the same contribution. We thus proceed to $A_3$, and decompose
\[A_3=A_{31}\cup A_{32}\cup A_{33} \cup A_{34} \]
as in the proof of Theorem 4, and
$A_{31},A_{34}$ can be dispatched using exactly the same ideas. For example the sum on $A_{31}$ is bounded by
\begin{equation*}
\begin{aligned}
M^2 \sum_{ k>j}\mathbb{P}[N(kM^r)-N(jM^r)=0]\leq C_rM^2,
\end{aligned}
\end{equation*}
with $C_r$ being as above. The set $A_{34}$ give the same contribution.
We now move on to $A_{32}.$
We decompose $A_{32}=A_{321}\cup A_{322}$, the first of which contains the quadruples for which the intervals $(k_1,j_1),(j_2,k_2)$ are disjoint, and the second contains those for which they intersect. Mimicking the steps in section 6, the sum over $A_{321}$ is bounded by
\begin{equation*}
\begin{aligned}
\sum_{a=0}^{\infty}\Big[\sum_{ {k>j}}\mathbb{P}[N((k-j)M^r)=a]\Big]^2 = \sum_{a\leq M^{2r}}+\sum_{a> M^{2r}}\Big[\sum_{ {k>j}}\mathbb{P}[N((k-j)M^r)=a]\Big]^2={\bf I + II.}
\end{aligned}
\end{equation*}
This decomposition and the methods we employ to estimate each sum differ significantly from our previous methods.
Let $b=k-j$, and $a\in \N$. Then
\begin{equation*}
\begin{aligned}
\sum_{ {k>j}}\mathbb{P}[N((k-j)M^r)=a]=\sum_{b=1}^{M-1} (M-b)\mathbb{P}[N(bM^r)=a]&=\frac{M^{ar}}{a!}\sum_{b=1}^{M-1} (M-b)b^ae^{-bM^r}.
\end{aligned}
\end{equation*}
As $b^ae^{-bM^r}$ for $b\geq 0$ is a function that increases up to a supremum and then decreases, this can be estimated by
\begin{equation*}
\begin{aligned}
\leq \frac{M^{1+ar}}{a!}\sum_{b=1}^{M-1} b^ae^{-bM^r} \leq \frac{M^{1+ar}}{a!}2\sup_{b\geq 0} b^ae^{-bM^r}+ \frac{M^{1+ar}}{a!}\int_{0}^{\infty} b^ae^{-bM^r}db.
\end{aligned}
\end{equation*}
The supremum is attained when $b=aM^{-r}$, while the integral is clearly the Gamma function after a change of variables. Plugging these in, and using \eqref{robs} gives
\begin{equation*}
\begin{aligned}
= 2\frac{M a^a}{a!e^a} +
\frac{M^{1+ar}}{a!} \frac{\Gamma(a+1)}{M^{r(a+1)}} \leq \frac{M}{\sqrt{ a}}+M^{1-r}.
\end{aligned}
\end{equation*}
In the summation range of ${\bf I}$ this last sum is bounded by $2M/\sqrt{a}$, and thus
\begin{equation*}
\begin{aligned}
{\bf I }\leq \Big[\sum_{ {k>j}}\mathbb{P}[N((k-j)M^r)=0]\Big]^2 +\sum_{1\leq a\leq M^{2r}}\frac{4M^2}{a}\leq C_r^2+4M^2(1+2r\log M).
\end{aligned}
\end{equation*}
In the summation range of ${\bf II}$ the term $M^{1-r}$ dominates, hence
\begin{equation*}
\begin{aligned}
{\bf II }\leq 2M^{1-r} \sum_{ a> M^{2r}}\sum_{ {k>j}}\mathbb{P}[N((k-j)M^r)=a]= 2M^{1-r} \sum_{ k>j}\sum_{ a> M^{2r}}\mathbb{P}[N((k-j)M^r)=a]\leq M^{3-r}.
\end{aligned}
\end{equation*}
This finishes the estimate on $A_{321}.$
As in the previous proof, this case of nonintersection is the generic case, and intersection cases contained in $A_{322}$ reduce to the previous cases.
The set $A_{33}$ is similar to $A_{32}$. We thus obtain the upper bound $\lesssim_r M^2\log M+M^{3-r}$.
For the lower bound, we use the observation that on the torus $L^4$ norm dominates $L^2$ norm to obtain the $M^2$ term, and the argument that in a small neighborhood of the origin there is little cancellation between the terms of the exponential sum to obtain the $M^{3-r}$ term.
\end{proof}
%=============================================================
%==============================================================
%===========================================================
%===========================================================
%===========================================================
\section{Perturbations of Green-Ruzsa Sets}
In this section we will prove that if we perturb the Green-Ruzsa sets, they will satisfy the Hardy-Littlewood majorant property almost surely. For an integer $D\geq 5$ the Green-Ruzsa set $\Lambda_{D,k}$ is defined by
\begin{equation*}
\Lambda_{D,k}:= \Big\{\sum_{j=0}^{k-1}d_j D^j | \ k\in \N, \ d_j\in \{0,1,3\} \Big\}.
\end{equation*}
Thus this set consists of nonnegative integers that are when written in base $D$ have digits only $0,1,3$.
Therefore this set is sparse, and we can quantify this sparsity by the estimate below. This property is the only property of this set that will be used.
\begin{proposition}
The Green-Ruzsa set $\Lambda_{D,k}$ satisfies the sparsity condition
\begin{equation*}
\big|\Lambda_{D,k} \cap [n-M,n+M]\big|\leq 24M^{\frac{\log 3}{\log D}}
\end{equation*}
for any $M\in \N, n\in \Z$.
\end{proposition}
\begin{proof}
We may assume $n\geq 0$. If the set in question has less than 2 elements, we are done. Therefore we may assume that it has at least two elements, and we let $a$ be its least element and $b$ its largest element.
Then we can write for some $m\leq k$
\[b-a=\sum_{j=0}^{m-1}c_jD^j \quad -3\leq c_j\leq 3\]
with $c_{m-1}>0$ as $b-a>0.$ But observe that
\[ 2M\geq b-a \geq D^{m-1}- 3\sum_{j=0}^{m-2}D^j =D^{m-1}-3\frac{D^{m-1}-1}{D-1}\geq D^{m-1}-\frac{3}{4}D^{m-1}=\frac{D^{m-1}}{4}.\]
From this we deduce
\[ m\leq 1+\frac{\log 8M}{\log D}.\]
By changing the first $m$ digits of $a$ in base $D$ we might obtain elements of $\Lambda_{D,k}$ not exceeding $b$, but we cannot change the higher digits even if they exist, for numbers thus obtained will certainly lie outside $[a,b]$. Therefore there are at most $3^m$ elements of $\Lambda_{D,k}$ in $[a,b]$, which means
\[ \big|\Lambda_{D,k} \cap [n-M,n+M]\big| =\big|\Lambda_{D,k} \cap [a,b]\big| \leq 3^{1+\frac{\log 8M}{\log D}}=3(8M)^{\frac{\log 3}{\log D}}\leq 24 M^{\frac{\log 3}{\log D}}. \]
\end{proof}
We now prove Theorem 6.
\begin{proof}
We start with \eqref{gr1} for $p=2.$
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j)}\Big\|_{2}^2 = \sum_{j,k\in A}\mathbb{P}[N(|j-k|)=0]
= \sum_{j,k\in A}e^{-|j-k|}.
\end{aligned}
\end{equation*}
We estimate this last term as
\begin{equation*}
\begin{aligned}
|A|\leq \sum_{j,k\in A}e^{-|j-k|}= |A|+ 2\sum_{k\in A}\sum_{\substack {j\in A\\ j<k}}e^{j-k}&\leq |A|+ 2\sum_{k\in A}\sum_{j=0}^{k-1}e^{j-k} \leq 3|A|.
\end{aligned}
\end{equation*}
Here we note that if for example $A=\Lambda_{D,k} \cap [0,M]$, then any element $a$ of $A$ with $a\equiv 1 \pmod{D}$ there exist $b\in A$ with $b=a-1$. Also if $a\equiv 3 \pmod{D}$, there exist $b\in A$ with $b=a-2$. Therefore for large $|A|$
\begin{equation*}
\begin{aligned}
\sum_{k\in A}\sum_{\substack {j\in A\\ j<k}}e^{j-k}\geq \frac{|A|-2}{3}\big[e^{-1}+e^{-2} \big].
\end{aligned}
\end{equation*}
Thus for such sets the contribution of the nondiagonal term is comparable to that of the diagonal term, as opposed to the situation we encountered in estimates over powers. Now all claims of the theorem for $p=2$ follow.
We now move on to \eqref{gr1} for $p=4$. From this \eqref{gr1} follows for all $2\leq p\leq 4$ immediately. The left hand side of \eqref{gr1} follows immediately from the $p=2$ case, and the Hölder inequality. So we concentrate on the right hand side. We may assume that $A$ contains no element in $[0, e^{100}]$, have more elements than $e^{100}$, and whenever $j,k\in A, \ k>j$ we have $k-j\geq e^{100}$. For if we have our result under these assumptions, then for a generic set $A\subset \Z_+$ we can do the following decomposition
\begin{equation*}
A_{-1}:=A\cap [0,e^{100}], \qquad A_n=(A\setminus A_{-1})\cap(\lceil e^{100} \rceil \Z +n), \qquad 0\leq n <\lceil e^{100} \rceil.
\end{equation*}
Some $A_n, \ 0\leq n < \lceil e^{100} \rceil $ may have more elements than $e^{100}$, let $N_1$ be the set of these $n$, and let $N_2$ be the set of the other $n$, including $n=-1$. Then
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j)}\Big\|_{4}^4 &\leq e^{303}\Big[\sum_{n\in N_1} \mathbb{E}\Big\|\sum_{j\in A_n} e^{2\pi i yN(j)}\Big\|_{4}^4 + \sum_{n\in N_2} \mathbb{E}\Big\|\sum_{j\in A_n} e^{2\pi i yN(j)}\Big\|_{4}^4 \Big].
\end{aligned}
\end{equation*}
Applying our result under the assumptions above to $A_n ,\ n\in N_1$, and estimating sums over $A_n ,\ n\in N_2$ trivially
\begin{equation*}
\begin{aligned}
\leq e^{303}\Big[C\sum_{n\in N_1}|A_n|^2 + \sum_{n\in N_2} |A_n|^4 \Big]
\leq e^{303}\Big[C\sum_{n\in N_1}|A_n|^2 + e^{202}\sum_{n\in N_2} |A_n|^2 \Big] \leq e^{303}[C+e^{202}]|A|^2.
\end{aligned}
\end{equation*}
So proving our result under these assumptions suffices.
We start with
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j)}\Big\|_{4}^4 = \sum_{j_1,j_2,k_1,k_2\in A}\mathbb{P}[N(j_1)+N(j_2)=N(k_1)+N(k_2)].
\end{aligned}
\end{equation*}
As before, the set $A^4$ can be written as
$A^4=A_{11}\cup A_{12}\cup A_{21}\cup A_{22}\cup A_3, $
and for any $A_{ij},i,j=1,2$
\begin{equation*}
\begin{aligned}
\sum_{ A_{ij}}\mathbb{P}[N(j_1^d)+N(j_2^d)=N(k_1^d)+N(k_2^d)]=|A|\sum_{ j,k\in A}\mathbb{P}[N(j^d)=N(k^d)].
\end{aligned}
\end{equation*}
This last sum has been estimated above. So from these sets we have $\lesssim |A|^2$ contribution.
We now decompose
$A_3=A_{31}\cup A_{32}\cup A_{33} \cup A_{34} $ as before.
Because we lack the good bounds on the nondiagonal sum in the $p=2$ case, unlike the powers case, estimating the sets $A_{31},A_{34}$ is now more difficult. Without loss of generality we can just concentrate on $A_{31}.$
\begin{equation*}
\begin{aligned}
\sum_{ A_{31}}\mathbb{P}[N(j_1)+N(j_2)=N(k_1)+N(k_2)]&=\sum_{ A_{31}}\mathbb{P}[N(j_1)-N(k_1)=0=N(k_2)-N(j_2)].
\end{aligned}
\end{equation*}
At this stage if we just ignore one of these equations and crudely estimate by
\begin{equation*}
\begin{aligned}
\leq |A|^2\sum_{ \substack{j,k\in A \\ j<k}}\mathbb{P}[N(k)-N(j)=0],
\end{aligned}
\end{equation*}
we may get a contribution $\approx |A|^3$. Therefore we cannot follow this approach, and must employ more delicate analysis similar in vein to the one we deploy for the sets $A_{32},A_{33}.$ We partition
$ A_{31}$ into two sets $A_{311},A_{312}$ where the first contains the quadruples for which $ (k_1,j_1),(k_2,j_2)$ are disjoint, and the second those for which they intersect.
For $A_{311}$ we have
\begin{equation*}
\begin{aligned}
\sum_{ A_{311}}\mathbb{P}[N(j_1)-N(k_1)=0=N(j_2)-N(k_2)]
=& \sum_{ A_{311}}\mathbb{P}[N(j_1)-N(k_1)=0]\Pp[ N(j_2)-N(k_2)=0]
\\ \leq &\Big[\sum_{ \substack{j,k\in A \\ j<k}}\mathbb{P}[N(k)-N(j)=0] \Big]^2 .
\end{aligned}
\end{equation*}
This last sum, as we estimated above is $\lesssim |A|^2$. In $A_{312}$ we have two possibilities, either
$j_2-k_2\geq j_1-k_1$ or $j_2-k_2< j_1-k_1$. Without loss of generality we may assume the first. Then
\begin{equation*}
\begin{aligned}
\sum_{ \substack{ A_{312}\\ j_2-k_2\geq j_1-k_1 } }\mathbb{P}[N(j_1)-N(k_1)=0=N(j_2)-N(k_2)]
\leq \sum_{ \substack{ A_{312}\\ j_2-k_2\geq j_1-k_1 }}\Pp[ N(j_2)-N(k_2)=0].
\end{aligned}
\end{equation*}
The conditions that $(k_1,j_1)$ must intersect $(k_2,j_2)$, and $j_2-k_2\geq j_1-k_1$ forces $j_1,k_1$ to be within an interval of length $3(j_2-k_2)-2$. Considering the sparsity condition on $A$, there are
$ C_A [ (3(j_2-k_2)/2]^{\alpha}$
integers to choose from. Thus the last sum above is bounded by
\begin{equation*}
\begin{aligned}
\leq 2C^2_A\sum_{\substack{j,k \in A \\ k>j}} \Pp[ N(k-j)=0](k-j)^{2\alpha}\leq 2C^2_A\sum_{\substack{j,k \in A \\ k>j}} e^{j-k}(k-j)^{2\alpha} \leq 2C^2_A |A|^2\sup_{x\geq 0} xe^{-x} \leq 2C^2_A|A|^2.
\end{aligned}
\end{equation*}
Thus contribution from $A_{31},A_{34}$ is $\lesssim_{C_A} |A|^2.$
We now move on to $A_{32}.$ This, together with its symmetric counterpart $A_{33}$, which is handled in exactly the same way, represent the most important, generic cases. We decompose $A_{32}=A_{321}\cup A_{322}$ as before.
We can write the sum over $A_{321}$ as before
\begin{equation*}
\begin{aligned}
\sum_{ A_{321}} \mathbb{P}[N(j_1)+N(j_2)=N(k_1)+N(k_2)]
\leq
\sum_{a=0}^{\infty}\Big[\sum_{ \substack{j,k\in A \\j<k}}\mathbb{P}[N(k-j)=a]\Big]^2,
\end{aligned}
\end{equation*}
and use the ideas introduced above, but this is not good enough. For the extra variable $a$ introduced to relate $j_1,j_2,k_1,
k_2$ to each other leads to inefficiencies. Instead we will relate these variables to each other directly with an inequality. We further decompose $A_{321}$ into the set of quadruples $A_{3211}$ with $k_2-j_2\geq j_1-k_1$, which we consider without loss of generality, and the remaining ones comprising $A_{3212}$ . We define $f(x):=4\sqrt{x\log x}$ for $x\geq 1.$
To each quadruple $(j_1,j_2,k_1,k_2)$ we consider, we assign three events
\begin{equation*}
\begin{aligned}
& \Omega_*:=\big\{\omega\in \Omega: N(j_1)(\omega)-N(k_1)(\omega)=N(k_2)(\omega)-N(j_2)(\omega) \big\},\\
&\Omega_1:=\big\{\omega\in \Omega: |N(j_1)(\omega)-N(k_1)(\omega)-(j_1-k_1)|< f(k_2-j_2) \big\}, \\
&\Omega_2:=\big\{\omega\in \Omega: |N(k_2)(\omega)-N(j_2)(\omega)-(k_2-j_2)|< f(k_2-j_2) \big\}.
\end{aligned}
\end{equation*}
Using these sets we can write
\begin{equation*}
\begin{aligned}
\Omega_*\subseteq [\Omega_*\cap\Omega_1 \cap \Omega_2] \cup \Omega_1^c \cup \Omega_2^c,
\end{aligned}
\end{equation*}
and thus also
\begin{equation*}
\begin{aligned}
\Pp[\Omega_*]\leq \Pp[\Omega_*\cap\Omega_1 \cap \Omega_2] + \Pp[\Omega_1^c] +\Pp[\Omega_2^c].
\end{aligned}
\end{equation*}
We have by Lemma \ref{lem1}
\begin{equation*}
\begin{aligned}
\Pp[\Omega_1^c], \Pp[\Omega_2^c]\leq 2e^{-4\log (k_2-j_2)}=2(k_2-j_2)^{-4},
\end{aligned}
\end{equation*}
that is, the contribution of these terms are minor. For the main term $\Omega_*\cap\Omega_1 \cap \Omega_2$ we first observe that if this set contains even one $\omega$, we have by triangle inequality
\begin{equation*}\label{gr4}
\begin{aligned}
&|(k_2-j_2)-(j_1-k_1)|\\ =&\big|(k_2-j_2)-(j_1-k_1)+N(j_1)(\omega)-N(k_1)(\omega)-[N(k_2)(\omega)-N(j_2)(\omega)] \big| \\ \leq &|N(j_1)(\omega)-N(k_1)(\omega)-(j_1-k_1)|+ |N(k_2)(\omega)-N(j_2)(\omega)-(k_2-j_2)|\\ < &2f(k_2-j_2).
\end{aligned}
\end{equation*}
So the set $\Omega_*\cap\Omega_1 \cap \Omega_2$ is empty for quadruples violating this relation. On the other hand, the probability of $\Omega_*$ can be estimated as
\begin{equation*}
\begin{aligned}
\Pp[\Omega_*]=\sum_{a=0}^{\infty} \Pp[ N(j_1)-N(k_1)=a=N(k_2)-N(j_2) ] & =\sum_{a=0}^{\infty} \Pp[ N(j_1-k_1)=a]\Pp[N(k_2-j_2)=a ]\\ & \leq \frac{1}{\sqrt{k_2-j_2}}\sum_{a=0}^{\infty} \Pp[ N(j_1-k_1)=a]\\ & =\frac{1}{\sqrt{k_2-j_2}}.
\end{aligned}
\end{equation*}
We then can bound
\begin{equation*}
\begin{aligned}
\sum_{A_{3211}} \mathbb{P}[N(j_1)+N(j_2)=N(k_1)+N(k_2)] =
\sum_{ A_{3211}} \mathbb{P}[N(j_1)-N(k_1)=N(k_2)-N(j_2)]
\end{aligned}
\end{equation*}
by the following
\begin{equation*}
\begin{aligned}
&\sum_{ A_{3211}} \Pp[\Omega_*\cap\Omega_1 \cap \Omega_2] + \Pp[\Omega_1^c] +\Pp[\Omega_2^c]
\leq \sum_{ \substack{A_{3211}\\ |(k_2-j_2)-(j_1-k_1)|< 2f(k_2-j_2) }} \frac{1}{\sqrt{k_2-j_2}}+ \sum_{ A_{3211}} 4(k_2-j_2)^{-4}.
\end{aligned}
\end{equation*}
The second sum is easy to handle:
\begin{equation*}
\begin{aligned}
\sum_{ A_{3211}} 4(k_2-j_2)^{-4}\leq \sum_{\substack{j_2,k_2 \in A \\ j_2 <k_2 }}\sum_{\substack{j_1,k_1 \in A \\ k_1<j_1}} 4(k_2-j_2)^{-2}(j_1-k_1)^{-2} \leq 4\Big[\sum_{\substack{j,k \in A \\ j<k}} (k-j)^{-2}\Big]^2.
\end{aligned}
\end{equation*}
But we have
\begin{equation*}
\begin{aligned}
\sum_{\substack{j,k \in A \\ k>j}} (k-j)^{-2} \leq \sum _{k\in A}\sum_{1<j<k} (k-j)^{-2} < \sum _{k\in A}\frac{\pi ^2}{6}<2|A|.
\end{aligned}
\end{equation*}
So the contribution of the second sum is bounded by $16|A|^2.$
We now look at the first sum. It is bounded by
\begin{equation*}
\begin{aligned}
\sum_{\substack{j,k \in A \\ j_2<k_2}} \frac{1}{\sqrt{k_2-j_2}}|\{(j_1,k_1)\in A^2: |(k_2-j_2)-(j_1-k_1)|< 2f(k_2-j_2) \}|.
\end{aligned}
\end{equation*}
So we need to bound the cardinality of the set within the summation. The condition $|(k_2-j_2)-(j_1-k_1)|< 2f(k_2-j_2)$ implies that for a fixed $k_1 \in A$ the element $j_1$ lies in an interval of radius $2f(k_2-j_2)$. By the density condition on our set $A$, there are at most $C_A(2f(k_2-j_2))^{\alpha}$ elements $j_1$ for each $k_1 \in A$. Therefore
\begin{equation*}
\begin{aligned}
\leq 2C_A |A| \sum_{\substack{j,k \in A \\ j<k}} \sqrt{\log(k-j)} (k-j)^{\frac{\alpha-1}{2}}\leq 2C_A |A| \sum_{n=0}^{\infty} \sum_{\substack{j,k \in A \\ 2^n\leq k-j<2^{n+1}}} ((n+1)\log 2)^{\frac{1}{2}} 2^{n\frac{\alpha-1}{2}}.
\end{aligned}
\end{equation*}
To proceed we need the cardinality of $\{(j,k)\in A^2: 2^n\leq k-j<2^{n+1}\}$. For a pair $(j,k)$ to be in this set, for a fixed $j$ we must have $k$ within the interval $[j+2^{n},j+2^{n+1})$. By the density hypothesis on $A$, this means for a fixed $j$ we can have at most $C_A2^{n\alpha}$ values of $k.$
Thus
\begin{equation*}
\begin{aligned}
\leq 2C_A^2 |A|^2 \sum_{n=0}^{\infty} (n+1)^{\frac{1}{2}} 2^{n\frac{3\alpha-1}{2}}.
\end{aligned}
\end{equation*}
Plainly if $\alpha<1/3$ the sum converges to a constant $C_{\alpha}.$ This concludes the estimation of the sum $A_{321}$ with a constant that depends only on $\alpha,C_A.$
We finally consider the sum over $A_{322}$. We partition this set into $A_{322n}, \ 1\leq n\leq 4$ as in
\eqref{par2}.
The first two and the last two are handled in the same way so we will only consider $ A_{3221}, A_{3223}.$
For $ A_{3221}$ the same arguments as in Chapter 6 yield
\begin{equation*}
\begin{aligned}
\sum_{ A_{3221}}\mathbb{P}[N(j_1)+N(j_2)=N(k_1)+N(k_2)]\leq \Big[ \sum_{ k>j}\mathbb{P}[N(k-j)=0] \Big]^2.
\end{aligned}
\end{equation*}
We estimated the sum inside the square by $3|A|$. So the contribution from here is at most $9|A|^2.$
On $ A_{3223}$ the condition $ k_1<j_2<j_1<k_2 $ means $(k_1,j_2)\cap(j_1,k_2)=\emptyset$, and again we reduce to the case $A_{321}.$ Hence we obtain \eqref{gr1} for $p=4$. By the Hölder inequality we obtain \eqref{gr1}.
Plainly for any $n\in \N$
\begin{equation*}
\begin{aligned}
\Big\|\sum_{j\in A} a_je^{2\pi i yN(j)}\Big\|_{2n}^{2n} \leq \Big\|\sum_{j\in A} e^{2\pi i yN(j)}\Big\|_{2n}^{2n}.
\end{aligned}
\end{equation*}
Then \eqref{gr2} follows by using this together with \eqref{gr1}, and the Hölder inequality.
For $p\geq 2$
\begin{equation*}
\begin{aligned}
\Big\|\sum_{j\in A} e^{2\pi i yN(j)}\Big\|_{p}^{p} \geq \Big\|\sum_{j\in A} e^{2\pi i yN(j)}\Big\|_{2}^{p} \geq |A|^{p/2}.
\end{aligned}
\end{equation*}
So by the Chebyshev inequality
\begin{equation*}\label{}
\begin{aligned}
\mathbb{P}\Big[\sup_{|a_j|\leq 1} \Big\| \sum_{j\in A} a_je^{2\pi i yN(j)} \Big\|_{p}^p \geq |A|^{\varepsilon}\Big\| \sum_{j\in A} e^{2\pi i yN(j)} \Big\|_{p}^p \Big]
&\leq \mathbb{P}\Big[\sup_{|a_j|\leq 1} \Big\| \sum_{j\in A} a_je^{2\pi i yN(j)} \Big\|_{p}^p \geq |A|^{\frac{p}{2}+\varepsilon} \Big]\\
& \leq |A|^{-\frac{p}{2}-\varepsilon} \mathbb{E}\sup_{|a_j|\leq 1} \Big\| \sum_{j\in A} a_je^{2\pi i yN(j)}\Big\|_p^p .
\end{aligned}
\end{equation*}
By \eqref{gr2} this last expression is $ \lesssim_{\alpha,C_A}|A|^{-\varepsilon}.$ Therefore taking the limits completes the proof of \eqref{gr3}.
\end{proof}
%==================================================
%====================================================
%===================================================
%====================================================
%======================================================
\section{Perturbations of Powers II: $p=2n$}
In this section we prove Theorem 7. The proof handles the explicit decompositions of the proofs of Theorems 4,5,6 via abstract combinatorial arguments. Also in those proofs when two variables are equal we reduce to the $p=2$ case. In the present proof, this reduces us to $p=2n-2$, which requires an inductive argument. After finishing off the auxiliary sums, the main sum reduces to a lattice point problem, for which we appeal to Theorem 8.
\begin{proof}
With Theorem 4 we already proved the cases $p=2,4$, so we assume $p\geq 6$, and thus $d\geq n\geq 3$. As explained we will need an inductive argument, so we also assume the statement to be true for $n-1.$
We may assume that $A$ contains no element in $[1, e^{100d}]$, and have more elements than $e^{100d}$. It follows immediately that whenever $j,k\in A, \ k>j$ we have $k^d-j^d\geq e^{100d(d-1)}$.
We let ${\bf j}=(j_1,j_2,\cdots,j_n)$ and ${\bf k}=(k_1,k_2,\cdots,k_n)$ denote vectors in $A^n$. We start with transforming our sum
\begin{equation*}
\begin{aligned}
\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{2n}^{2n} = \sum_{({\bf j,k})\in A^{2n}}\mathbb{P}\Big[\sum_{i=1}^nN(j_i^d)=\sum_{i=1}^nN(k_i^d)\Big].
\end{aligned}
\end{equation*}
We decompose $A^{2n}=A^*\cup A^{**}$, where $A^{**}$ contains the vectors $({\bf j,k})$ with $j_a=k_b$ for some $1\leq a,b\leq n.$ The main contribution comes from $A^*$, and $A^{**}$ can easily be dealt with induction. We first dispatch $A^{**}$.
Let $A^{**}_{ab}$ for some fixed $1\leq a,b\leq n$ denote the subset of $A^{**}$ contsisting of vectors $({\bf j,k})$ with $j_a=k_b$. Then after relabeling
\begin{equation*}
\begin{aligned}
\sum_{ A_{ab}^{**}}\mathbb{P}\Big[\sum_{i=1}^nN(j_i^d)=\sum_{i=1}^nN(k_i^d)\Big]&=|A|\sum_{ ({\bf j',k'})\in A^{2n-2}}\mathbb{P}\Big[\sum_{i=1}^{n-1}N(j_i^d)=\sum_{i=1}^{n-1}N(k_i^d)\Big]\\ &=
|A|\mathbb{E}\Big\|\sum_{j\in A} e^{2\pi i yN(j^d)}\Big\|_{2(n-1)}^{2(n-1)}.
\end{aligned}
\end{equation*}
Now applying the inductive hypothesis gives an appropriate bound for this term.
Therefore it remains to handle $A^*$.
We consider the set
\begin{equation*}
\begin{aligned}
A_{1}^*&=:\{({\bf j,k})\in A^{2n}: j_1\leq j_2 \leq \ldots j_n, \ \ k_1\leq k_2\leq \ldots k_n \}.
\end{aligned}
\end{equation*}
Since there are $n!$ different orderings of for each of $j_i,k_i $ and since each ordering can be turned into another by a relabeling, we have
\begin{equation*}
\begin{aligned}
\sum_{ A^*}\mathbb{P}\Big[\sum_{i=1}^n(j_i^d)=\sum_{i=1}^nN(k_i^d)\Big]&\leq (n!)^2 \sum_{A_1^*}\mathbb{P}\Big[\sum_{i=1}^nN(j_i^d)=\sum_{i=1}^nN(k_i^d)\Big].
\end{aligned}
\end{equation*}
We will say that the pair $j_i,k_i$ have positive orientation if $j_i<k_i$ and negative orientation if $j_i>k_i.$ For every vector $({\bf j,k})\in A^*_1$ this divides pairs or indices $1\leq i\leq n$ of the vector into two sets $\sigma_+({\bf j,k})$ and $\sigma_-({\bf j,k}).$ We observe that open intervals $(j_a,k_a)\in \sigma_+({\bf j,k})$ and $(k_b,j_b)\in \sigma_-({\bf j,k})$ cannot intersect. For intersection implies $j_a<j_b$ and hence $a<b$, but it also implies $k_b<k_a$ and thus $b<a.$ The sum over $A_1^*$ above then becomes
\begin{equation}\label{p2k1}
\begin{aligned}
= \sum_{A_1^*}\mathbb{P}\Big[\sum_{i\in \sigma_-({\bf j,k})}N(j_i^d)-N(k_i^d)=\sum_{i\in \sigma_+({\bf j,k})}N(k_i^d)-N(j_i^d)\Big].
\end{aligned}
\end{equation}
Of course one of $\sigma_+({\bf j,k}),\sigma_-({\bf j,k})$ may be empty, in which case the sum over that set is taken to be zero. We define $f(x):=4\sqrt{nx\log x}$ for $x\geq 1.$
To each vector in $A_1^*$ we assign the events
\begin{equation*}
\begin{aligned}
& \Omega_*:=\Big\{\sum_{i\in \sigma_-({\bf j,k})}N(j_i^d)-N(k_i^d)=\sum_{i\in \sigma_+({\bf j,k})}N(k_i^d)-N(j_i^d) \Big\},\\
&\Omega_i:=\Big\{ \big|N(k_i^d)(\omega)-N(j_i^d)(\omega)-(k_i^d-j_i^d)\big|<\max_{1\leq i\leq n} f(|k_i^d-j_i^d|) \Big\}, \qquad 1\leq i \leq n.
\end{aligned}
\end{equation*}
We let $m$ be an index that maximizes $|k_i^d-j_i^d|$ and hence $f(|k_i^d-j_i^d|)$. Using these sets we define $\Omega_{**}:=\Omega_*\cap\Omega_1 \cap \Omega_2\ldots \cap \Omega_n$, and obtain
\begin{equation*}
\begin{aligned}
\Pp[\Omega_*]\leq \Pp[\Omega_{**}] + \sum_{i=1}^n\Pp[\Omega_i^c] .
\end{aligned}
\end{equation*}
By Lemma \ref{lem1},
$\Pp[\Omega_i^c] \leq 2|k_m^d-j_m^d|^{-4n},$
that is, the contribution of these terms are minor. For the main term $\Omega_{**}$ we first observe that if this set is nonempty for a vector $({\bf j,k})$ we have by the triangle inequality
\begin{equation}\label{p2k2}
\begin{aligned}
\Big| \sum_{i\in \sigma_+({\bf j,k})} (k_i^d-j_i^d) - \sum_{i\in \sigma_-({\bf j,k})} (j_i^d-k_i^d) \Big| < nf(|k_m^d-j_m^d|).
\end{aligned}
\end{equation}
So the set $\Omega_{**}$ is empty for vectors violating this relation.
Let $A_{11}^*$ be the set of vectors in $A_{1}^*$ satisfying this relation. Observe that if for a vector $({\bf j,k})$ one of the sets $\sigma_+({\bf j,k}), \sigma_-({\bf j,k})$ is empty, then that vector cannot belong to $A_{11}^*$. For vectors in $A_{11}^*$, assuming $m\in \sigma_+({\bf j,k})$, we compute the probability of the set $\Omega_{**}$ as follows. Our observation on nonintersection of pairs in sets $\sigma_-,\sigma_+$, and the independent increment property of Poisson processes allow us to apply Lemma 5 to write
\begin{equation*}
\begin{aligned}
\Pp[\Omega_{**}] &=\sum_{a\in \Z_+}\mathbb{P}\Big[\sum_{i\in \sigma_-({\bf j,k})}N(j_i^d)-N(k_i^d)=a=\sum_{i\in \sigma_+({\bf j,k})}N(k_i^d)-N(j_i^d)\Big] \\ &=\sum_{a\in \Z_+}\mathbb{P}\Big[\sum_{i\in \sigma_-({\bf j,k})}N(j_i^d)-N(k_i^d)=a\Big]\Pp\Big[ \sum_{i\in \sigma_+({\bf j,k})}N(k_i^d)-N(j_i^d)=a\Big].
\end{aligned}
\end{equation*}
We apply Lemma 7 to estimate the sum on $\sigma_+$, and sum over $a$ for the sum on $\sigma_-$
\begin{equation*}
\begin{aligned}
\leq \frac{\sqrt{n}}{\sqrt{ |k_m^d-j_m^d|}}\sum_{a\in \Z_+}\mathbb{P}\Big[\sum_{i\in \sigma_-({\bf j,k})}N(j_i^d)-N(k_i^d)=a\Big] \leq \frac{\sqrt{n}}{\sqrt{ |k_m^d-j_m^d|}} .
\end{aligned}
\end{equation*}
If $m$ belongs to $ \sigma_-({\bf j,k})$, after using independence we apply Lemma 7 to the sum on $\sigma_-$ and sum over $a$ the other sum to get the same result.
Combining all of these we continue from \eqref{p2k1}
\begin{equation*}
\begin{aligned}
\leq \sum_{ A_{1}^*}\Big[ \Pp[\Omega_{**}]+ \sum_{i=1}^n \Pp[\Omega_i^c] \Big]
& = \sum_{ A_{11}^*} \Pp[\Omega_{**}] + \sum_{ A_{1}^*} \sum_{i=1}^n \Pp[\Omega_i^c] \\ &\leq
\sum_{ A_{11}^*} \frac{\sqrt{n}}{\sqrt{ |k_m^d-j_m^d|}}+ 2n\sum_{ A_{1}^*} |k_m^d-j_m^d|^{-4n}.
\end{aligned}
\end{equation*}
The second sum is easy to handle:
\begin{equation*}
\begin{aligned}
\sum_{ A_{1}^*} |k_m^d-j_m^d|^{-4n}\leq \sum_{ A_{1}^*} \prod_{i=1}^{n}|k_i^d-j_i^d|^{-4} \leq 2^n \prod_{i=1}^{n} \sum_{\substack{j_i,k_i \in A \\ j_i<k_i}} |k_i^d-j_i^d|^{-4} \leq 2^n\Big[\sum_{k \in A } k^{-4d+5}\Big]^n\leq 1 .
\end{aligned}
\end{equation*}
We now look at the first sum. We decompose $ A_{11}^*=\bigcup_{i=1}^n A_{11i}^*$ where $m=i$ for vectors in $A_{11i}^*.$ We further decompose $ A_{11i}^*=\bigcup_{\substack{j_i,k_i\in A \\ j_i\neq k_i}} A_{11i}^{*j_ik_i}$, where $A_{11i}^{*j_ik_i}$ contains the vectors in $A_{11i}^{*}$ in which $j_i,k_i$ are fixed. Then
\begin{equation*}\label{}
\begin{aligned}
\sum_{ A_{11}^*} \frac{1}{\sqrt{|k_m^d-j_m^d|}}&\leq \sum_{i=1}^n \sum_{ A_{11i}^{*}} \frac{1}{\sqrt{ |k_i^d-j_i^d|}} \leq n\sup_{1\leq i\leq n} \sum_{ A_{11i}^{*}} \frac{1}{\sqrt{ |k_i^d-j_i^d|}}.
\end{aligned}
\end{equation*}
We will estimate the last sum for a fixed $i=r.$ As the estimation will not depend on the choice of $r$, this will suffice.
\begin{equation}\label{p2k4}
\begin{aligned}
\sum_{ A_{11r}^{*}} \frac{1}{\sqrt{ |k_r^d-j_r^d|}}= \sum_{\substack{j_r,k_r \in A \\ j_r\neq k_r}} \sum_{ A_{11r}^{*j_rk_r}} \frac{1}{\sqrt{|k_r^d-j_r^d|}}= \sum_{\substack{j_r,k_r \in A \\ j_r\neq k_r}} \frac{\big| A_{11r}^{*j_rk_r} \big|}{\sqrt{ |k_r^d-j_r^d|}}={\bf S}.
\end{aligned}
\end{equation}
We need to estimate $|A_{11r}^{*j_rk_r}|$. Let $D_i:=k_i^d-j_i^d$.
The set $A_{11r}^{*j_rk_r}$ lies within the set
\begin{equation}\label{p2k5}
\begin{aligned}
\Big\{ ({\bf j,k})\in A^{2n}\Big| \ j_r,k_r \ \text{are fixed}, \quad 0<|k_i^d-j_i^d|\leq |D_r| \ \ \ 1\leq i\leq n, \quad \eqref{p2k2} \ \text{holds} \Big\}.
\end{aligned}
\end{equation}
We estimate the size of this set by looking at the possible choices for pairs $j_i,k_i$. Number of choices for any pair $j_i,k_i$ for vectors of this set can be estimated trivially by $|A|^2$ using merely the fact that $j_i,k_i\in A$. This we call the first method. Alternatively our second method proceeds via the identity
\begin{equation*}\label{}
\begin{aligned}
\big|\big\{(j_i,k_i)\in A^2 \big| \ 0<|k_i^d-j_i^d|\leq |D_r| \ \big\} \big|=2 \big|\big\{(j_i,k_i)\in A^2 \big| \ 0<k_i^d-j_i^d\leq |D_r| \ \big\} \big|
\end{aligned}
\end{equation*}
and applying Theorem 8 with $E=D=|D_r|$, and obtains a bound $C_d|D_r|^{2/d}.$ Once $n-1$ pairs are fixed, for the remaining pair we can use the full force of our arithmetic results. This will be called the refined method.
We first crudely estimate
{\bf S}. Observe that by the second method
\begin{equation*}\label{}
{\bf S}\lesssim_{d,n} \sum_{\substack{ j_r,k_r \in A \\ j_r\neq k_r}} \frac{|D_r|^{2(n-1)/d}}{ |D_r|^{1/2}}= \sum_{\substack{ j_r,k_r \in A \\ j_r\neq k_r}} |D_r|^{\frac{2(n-1)}{d}-\frac{1}{2}}.
\end{equation*}
The exponent of $|D_r|$ in this last expression is nonpositive if $d\geq 4n-4$. In this case the contribution of this term can be controlled by $C_{d,n}|A|^2.$ So we only need to estimate ${\bf S}$ under the assumption $d<4n-4.$
To do this we will estimate the size of \eqref{p2k5} combining all our three methods. We have $n-1$ pairs to estimate in this set. One of them will be estimated by the refined method after all $n-2$ pairs are estimated and fixed by the first and the second methods, we fix an index $1\leq l\leq n$ with $l\neq r$ for this purpose. For the other $n-2$ indices we note that in general the first method gives a better bound than the second, but it is the second method that enables cancellation with the $\sqrt{|k^d-j^d|}$ term of the denominator. So we must use the second method as many times as possible to take advantage of this cancellation, and for the remaining pairs we use the first method. To this end we pick $d':= \lceil \frac{d}{4}-\frac{1}{2} \rceil-1$ indices apart from $r,l$. This $d'$ is the largest number of applications for the second method after which the power of
$|k^d-j^d|$ still remains nonpositive in the sum. Given the assumption $d<4n-4$ we also have $d'\leq n-2$. To the remaining $n-2-d'$ pairs we apply the first method. So the cardinality of \eqref{p2k5} is bounded by
\begin{equation}\label{p2k3}
\lesssim_{d,n}|D_r|^{\frac{2d'}{d}}|A|^{2(n-2-d')} \sup_{\substack{|D_i| \leq |D_r| \\ i\neq r,l}} \Big| \Big\{ ({ j_l,k_l})\in A^{2}\Big| \ j_l\neq k_l, \quad \Big| k_l^d-j_l^d +\sum_{i\neq l} D_i \Big| < nf(|D_r|) \Big\}\Big|.
\end{equation}
So only the index $l$ remains, to which we will apply the refined method. Since $\Big|\sum_{i\neq l} D_i\Big| \leq n|D_r|$ we can bound the supremum above by
\begin{equation*}
\begin{aligned}
&\leq \sup_{|D| \leq n|D_r| } \big| \big\{ ({ j_l,k_l})\in A^{2}\big| \ j_l\neq k_l, \quad | k_l^d-j_l^d +D | < nf(|D_r|) \big\}\big| \\ &\leq 2 \sup_{|D| \leq n|D_r| } \big| \big\{ ({ j,k})\in A^{2}\big| \ j< k, \quad | k^d-j^d -D | < nf(|D_r|) \big\}\big| \\ &\leq 2 \sup_{nf(|D_r|) \leq D \leq n|D_r| } \big| \big\{ ({ j,k})\in A^{2}\big| \ j< k, \quad | k^d-j^d -D | < nf(|D_r|) \big\}\big| .
\end{aligned}
\end{equation*}
To this we can now apply our Theorem 8 to continue from \eqref{p2k3}
\begin{equation*}\label{}
\begin{aligned}
\lesssim_{d,n}|A|^{2n-4-2d'}|D_r|^{\frac{2d'}{d}} f^{\frac{2}{d}}(|D_r|)\lesssim_{d,n} |A|^{2n-4-2d'}|D_r|^{\frac{2d'+1}{d}} \log^{\frac{1}{d}}|D_r| .
\end{aligned}
\end{equation*}
Plugging this in we can estimate
\begin{equation*}\label{p2k8}
\begin{aligned}
{\bf S} \lesssim_{d,n} |A|^{2n-4-2d'} \sum_{\substack{j,k\in A\\ j<k}} \frac{ \log^{\frac{1}{d}}|k^d-j^d|}{ |k^d-j^d|^{\frac{1}{2}-\frac{2d'+1}{d}}}.
\end{aligned}
\end{equation*}
In the last sum the power of $|k^d-j^d|$ is $i/2d$ with $i=\{1,2,3,4\}$ depending on $d \pmod{4}$. In any case the summand in this last sum is a decreasing function of the type described in Lemma 8 on $[e^2,\infty)$. So we can apply Lemma 8 twice to bound the last sum by
\begin{equation*}\label{}
\begin{aligned}
\leq \sum_{\substack{10\leq j<k\leq |A|+9}} \frac{ \log^{\frac{1}{d}}|k^d-j^d|}{ |k^d-j^d|^{\frac{1}{2}-\frac{2d'+1}{d}}}\lesssim_d \log^{\frac{1}{d}}|A| \sum_{\substack{1\leq j<k\leq |A|}} |k^d-j^d|^{\frac{2d'+1}{d}-\frac{1}{2}}.
\end{aligned}
\end{equation*}
To estimate the last sum we factorize, and apply a change of variables to obtain
\begin{equation*}\label{}
\begin{aligned}
\leq \sum_{1< k\leq |A| }k^{(d-1)(\frac{2d'+1}{d}-\frac{1}{2})}\sum_{\substack{1\leq j<k } }\big(k-j\big)
^{\frac{2d'+1}{d}-\frac{1}{2}}= \sum_{1< k\leq |A| }k^{(d-1)(\frac{2d'+1}{d}-\frac{1}{2})}\sum_{\substack{1\leq j<k } }j
^{\frac{2d'+1}{d}-\frac{1}{2}}.
\end{aligned}
\end{equation*}
Using integrals we bound the inner sum and obtain
\begin{equation*}\label{}
\begin{aligned}
\lesssim \sum_{1< k\leq |A| }k^{(d-1)(\frac{2d'+1}{d}-\frac{1}{2})}k
^{\frac{2d'+1}{d}+\frac{1}{2}}= \sum_{1< k\leq |A| }k^{2d'+2-\frac{d}{2}}.
\end{aligned}
\end{equation*}
The exponent in this last sum is $-1$ if $d\equiv 2 \pmod{4}$, and $> -1$ otherwise. So this sum is bounded by
\begin{equation*}\label{}
\begin{aligned}
\lesssim \begin{cases} |A|^{2d'+3-\frac{d}{2}}\log |A| \quad &\text{if} \quad d\equiv 2 \pmod{4}, \\
|A|^{2d'+3-\frac{d}{2}} \quad &\text{else}.
\end{cases}
\end{aligned}
\end{equation*}
This finishes the estimation of {\bf S}, and gives exactly the desired bounds
\begin{equation*}\label{}
\begin{aligned}
{\bf S} \lesssim_{d,n} \begin{cases} |A|^{2n-\frac{d}{2}-1}\log^{1+\frac{1}{d}} |A| \quad &\text{if} \quad d\equiv 2 \pmod{4}, \\
|A|^{2n-\frac{d}{2}-1}\log^{\frac{1}{d}} |A| \quad &\text{else}.
\end{cases}
\end{aligned}
\end{equation*}
Putting together all our estimates finishes the proof.
\end{proof}
%=====================================================================================================================================================================================================================================================================================
\section{The Arithmetic Component}
We provide two proofs of Theorem 8. The first is by viewing it as a margin of sets of type \eqref{os}. The second is directly, without using any result from the literature on lattice point problems.
\begin{proof}\label{APROOF 1}
Plainly the cardinality of our set \eqref{os} can be obtained from the equations
\begin{equation*}
\begin{aligned}
\big|\big\{ (j,k)\in \Z^2: 0<|k|^d-|j|^d\leq x \big\}\big|&= \big|\big\{ (j,k)\in \Z^2: 0<|k|^d-|j|^d< x \big \}\big| \\ &+ \big| \big\{ (j,k)\in \Z^2: |k|^d-|j|^d= x \big\}\big|
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}
\big|\big\{ (j,k)\in \Z^2: 0<|k|^d-|j|^d< x \big \}\big|&= 4\big|\big\{ (j,k)\in \N^2: 0<k^d-j^d< x \big \}\big|\\ &+ 2\big|\big\{ k\in \N: 0<k^d< x \big \}\big|.
\end{aligned}
\end{equation*}
Plainly
\begin{equation*}
\big|\big\{ k\in \N: 0<k^d< x \big\}\big|= x^{1/d}+\Oh(1).
\end{equation*}
We also have
\begin{equation*}
\big|\big\{ (j,k)\in \Z^2: |k|^d-|j|^d= x \big\}\big|=4 \big|\big\{ (j,k)\in \N^2: k^d-j^d= x \big\}\big|+\Oh(1).
\end{equation*}
For the set on the right hand side to be nonempty, $x$ must be an integer. Then $k-j$ is a divisor of $x$, so it has at most $2d(x)$ different values. For a fixed value $a=k-j$, we have $x=k^d-j^d=(j+a)^d-j^d,$ as $a\neq 0$, this means $j$ is the root of a degree $d-1$ polynomial. Once $j$ is fixed, so is $k$. Hence there are $2(d-1)d(x)$ different pairs $(j,k)$ in this set. But $2(d-1)d(x)\leq C_{d,\varepsilon}x^{\varepsilon}.$ Therefore
\begin{equation}\label{crv}
\big |\big\{ (j,k)\in \Z^2: |k|^d-|j|^d= x \big\}\big|=\Oh_{d,\varepsilon}(x^{\varepsilon}).
\end{equation}
Using these we finally obtain
\begin{equation*}
\begin{aligned}
\big |\big\{ (j,k)\in \N^2: 0<k^d-j^d< x \big \}\big| = \frac{1}{4}\big|\big\{ (j,k)\in \Z^2: 0<|k|^d-|j|^d\leq x \big \}\big|-\frac{1}{2}x^{1/d}+\Oh_{d,\varepsilon}(x^{\varepsilon}),
\end{aligned}
\end{equation*}
which by \eqref{R-}, and \eqref{nwkr}
\begin{equation}\label{eta}
\begin{aligned}
= \frac{1}{4}\Big[A_dx^{2/d}+B_d x^{1/(d-1)}+D_dF_d(x^{1/d})x^{1/d-1/d^2}\Big]-\frac{1}{2}x^{1/d}+\Oh_{d,\varepsilon}(x^{\frac{46}{73}\frac{1}{d}+\varepsilon}).
\end{aligned}
\end{equation}
Choosing $\varepsilon=1/100d$ in both \eqref{nwkr},\eqref{crv}, and considering $D_dF_d(x^{1/d})x^{1/d-1/d^2}$ as an error term we obtain the estimate
\begin{equation}\label{eta3}
= \frac{1}{4}\Big[A_dx^{2/d}+B_d x^{1/(d-1)}\Big]-\frac{1}{2}x^{1/d}+\Oh_{d}(x^{1/d-1/d^2}).
\end{equation}
From this we will deduce a theorem on the set \eqref{ss2}.
We split this estimate into two:
\begin{equation}\label{p6te2}
\begin{aligned}
\sup_{D\leq E\leq 4D} &\big|\big\{ (j,k) \in \N^2: j<k, \ |k^d-j^d-E| <D \big\}\big| \\
\sup_{4D\leq E\leq D^2} &\big|\big\{ (j,k) \in \N^2: j<k, \ |k^d-j^d-E| <D \big\}\big| .
\end{aligned}
\end{equation}
The first estimate is easier to do, and we start with this first estimate. It is bounded by
\begin{equation}\label{a5d}
\begin{aligned}
\big|\big\{ (j,k) \in \N^2: \ 0<k^d-j^d <5D \big\}\big|.
\end{aligned}
\end{equation}
We apply \eqref{eta3} to immediately bound this by $C_dD^{2/d}.$ For the second estimate in \eqref{p6te2} we fix an $E$ in the described range, and then apply \eqref{eta3} for the values $E+D$ and $E-D$, and subtract to obtain
\begin{equation*}\label{}
\big|\big\{ (j,k) \in \N^2: j<k, \ E-D\leq k^d-j^d <E+D \big\}\big|
\end{equation*}
\begin{equation}\label{fth}
\begin{aligned}
&\leq \frac{1}{4}\Big[A_d\big[(E+D)^{2/d}-(E-D)^{2/d}\big] +B_d \big[(E+D)^{1/(d-1)}-(E-D)^{1/(d-1)}\big]\Big]\\ &-\frac{1}{2}\big[ (E+D)^{1/d}-(E-D)^{1/d} \big]+\Oh_{d}(E^{\frac{1}{d}-\frac{1}{d^2}}).
\end{aligned}
\end{equation}
We have the estimates for $4D\leq E\leq D^2$ by mean value theorem
\begin{equation*}\label{}
(E+D)^{\frac{2}{d}}-(E-D)^{\frac{2}{d}}\lesssim_d DE^{\frac{2}{d}-1}\leq D^{\frac{2}{d}}, \qquad (E+D)^{\frac{1}{d-1}}-(E-D)^{\frac{1}{d-1}}\lesssim_d DE^{\frac{1}{d-1}-1}\leq D^{\frac{1}{d-1}}.
\end{equation*} Plugging these into \eqref{fth} we obtain the bound $C_dD^{2/d}$.
\end{proof}
From the literature on lattice point problems, see for example \cite{gk}, we know that for the error terms $\Delta_d$ bounds of order $\Oh(x^{\frac{1}{d}})$ are reachable without recourse to Fourier analysis, merely by carefully decomposing the region, and carefully choosing the direction of the lines with which we count the lattice points. We show now that the same is true for our Theorem 8. However the proof is considerably long and difficult, as our estimate concerns not just one fixed set but a whole class of sets. This forces a case by case analysis. Another issue is that for $d\geq 3$ the curves $y^d-x^d=m$ do not have the algebraic simplicity that they have when $d=2.$ Once $d>2$ we no longer have easy rotation and inversion properties. All this comes through in the proof, most obviously in our introduction and study of the function $g$.
\begin{proof}
We split this estimate into two as in \eqref{p6te2},
and start with the first. It is bounded by \eqref{a5d}.
We count the pairs of this set by fixing $j.$
We observe that
\begin{equation*}\label{}
\begin{aligned}
(k-j)dj^{d-1}< \big(k-j\big)\big(k^{d-1}+k^{d-2}j+\ldots +kj^{d-2}+j^{d-1}\big)=k^d-j^d <5D.
\end{aligned}
\end{equation*}
Therefore $j<5D^{\frac{1}{d-1}}$, and for each such $j$
\begin{equation*}\label{}
\begin{aligned}
1\leq k-j<5Dj^{1-d}.
\end{aligned}
\end{equation*}
Hence for a fixed $j$ there are at most $5Dj^{1-d}$ pairs $(j,k).$ We will use this estimate for $j>D^{1/d}.$ For $j\leq D^{1/d}$ we observe that
\begin{equation*}\label{}
\begin{aligned}
k^d-j^d <5D\implies k<(j^d+5D)^{1/d}\implies k<6D^{1/d}.
\end{aligned}
\end{equation*}
Thus for these $j$ there are at most $6D^{1/d}$ pairs $(j,k).$ Now summing all these we can
bound our set by
\begin{equation*}\label{}
\begin{aligned}
\sum_{1\leq j < 5D^{\frac{1}{d-1}}}\big|\big\{ k\in \N: \ 0<k^d-j^d <5D \big\}\big|
\leq \sum_{1\leq j\leq D^{1/d}} 6D^{1/d} + 5D\cdot \sum_{D^{\frac{1}{d}}<j < 5D^{\frac{1}{d-1}}} j^{1-d}.
\end{aligned}
\end{equation*}
The first sum is plainly bounded by $6D^{2/d}$. The second sum can be estimated by
\begin{equation*}\label{}
\begin{aligned}
D^{-1+\frac{1}{d}} +\int_{D^{1/d}}^{\infty} x^{1-d}dx\leq 2D^{-1+\frac{2}{d}}.
\end{aligned}
\end{equation*}
So we obtain the final bound $16D^{2/d}.$
Now we turn to the second estimate in \eqref{p6te2}.
We will count this set via differences $b=k-j$ as in the proof of Theorem 4. Since $b^d=(k-j)^d<k^d-j^d<E+D$ we must have
$1\leq b<(E+D)^{1/d}.$ Then for a fixed $E\in [4D,D^2],$ the second set in \eqref{p6te2} is bounded by
\begin{equation*}\label{}
\begin{aligned}
\leq \sum_{1\leq b<(E-2D)^{\frac{1}{d}}}+\sum_{(E-2D)^{\frac{1}{d}}\leq b<(E+D)^{\frac{1}{d}}}\big|\big\{ j \in \N: |(j+b)^d-j^d-E| <D \big\}\big|
={\bf I+II}.
\end{aligned}
\end{equation*}
To deal with these sets we introduce the function
\begin{equation*}\label{}
\begin{aligned}
g(x):=(x+b)^d-x^d-b^d=\sum_{j=1}^{d-1}{d \choose j}x^{d-j}b^j, \quad x\geq 0.
\end{aligned}
\end{equation*}
This function is strictly increasing for $x\geq 0$ with $g(0)=0$, and indeed is a bijection on $[0,\infty)$. So it has a strictly increasing inverse $g^{-1}:[0,\infty)\rightarrow [0,\infty) $. Observe that
\begin{equation}\label{comgg}
\begin{aligned}
g(x) \leq xg'(x)=\sum_{j=1}^{d-1}{d \choose j}(d-j)x^{d-j}b^j\leq dg(x), \quad x\geq 0.
\end{aligned}
\end{equation}
Also for $x\geq 0$,
\begin{equation}\label{p6l1}
\begin{aligned}
g(x)\geq dx^{d-1}b+dxb^{d-1}, \ \text{and} \ \ g'(x)\geq db^{d-1}.
\end{aligned}
\end{equation}
Furthermore from the first of these we obtain
\begin{equation}\label{p6l4}
\begin{aligned}
g(x)\geq dbx^{d-1} \implies x\geq g^{-1}(dbx^{d-1})\implies \Big(\frac{x}{db}\Big)^{\frac{1}{d-1}}\geq g^{-1}(x),\\ \
g(x)\geq dxb^{d-1} \implies x\geq g^{-1}(dxb^{d-1})\implies \frac{x}{db^{d-1}}\geq g^{-1}(x).
\end{aligned}
\end{equation}
We now apply these observations on $g$ to obtain our results. We first deal with {\bf II}. Fix $b$.
\begin{equation}\label{p6l2}
\begin{aligned}
\big|\big\{ j \in \N: |(j+b)^d-j^d-E| <D \big\} \big| = &\big|\big\{ j \in \N: E-D-b^d<g(j) <E+D-b^d \big\} \big| \\ \leq &\big|\big\{ j \in \N: 0<g(j) <E+D-b^d \big\} \big|.
\end{aligned}
\end{equation}
By \eqref{p6l1}
\begin{equation*}\label{}
\begin{aligned}
g(j) <E+D-b^d \implies j<\frac{E+D-b^d}{db^{d-1}}.
\end{aligned}
\end{equation*}
Since $j>0$, and $b^d\geq E-2D\geq 2D$ we bound the last term of \eqref{p6l2} by
\begin{equation*}\label{}
\begin{aligned}
\big|\big\{ j\in \N: 1\leq j < \frac{E+D-b^d}{db^{d-1}} \big\} \big| \leq \frac{E+D-b^d}{db^{d-1}} \leq \frac{3D}{db^{d-1}} \leq \frac{3D^{1/d}}{d}\leq D^{1/d}.
\end{aligned}
\end{equation*}
Observe that the cardinality of integers $b$ with our assumption can be bounded from
\begin{equation*}\label{}
\begin{aligned}
3D=(E+D)-(E-2D)= \big[ (E+D)^{\frac{1}{d}}-(E-2D)^{\frac{1}{d}} \big] \Big[\sum_{j=1}^{d} (E+D)^{\frac{d-j}{d}}(E-2D)^{\frac{j-1}{d}}\Big],
\end{aligned}
\end{equation*}
implying that
\begin{equation*}\label{}
\begin{aligned}
(E+D)^{\frac{1}{d}}-(E-2D)^{\frac{1}{d}} < 3D(E+D)^{\frac{1}{d}-1}<3D^{\frac{1}{d}}.
\end{aligned}
\end{equation*}
So we bound
\begin{equation*}\label{}
\begin{aligned}
{\bf II}\leq (3D^{\frac{1}{d}}+1)D^{1/d}\leq 4D^{2/d}.
\end{aligned}
\end{equation*}
We turn to ${\bf I}.$ Again we fix $b$.
\begin{equation*}\label{}
\begin{aligned}
\big|\big\{ j \in \N: |(j+b)^d-j^d-E| <D \big\} \big| &= \big|\big\{ j \in \N: E-D-b^d<g(j) <E+D-b^d \big\} \big| \\ &= \big|\big\{ j \in \N: g^{-1}(E-D-b^d) <j <g^{-1}(E+D-b^d) \big\} \big|.
\end{aligned}
\end{equation*}
We measure the cardinality of this set by the mean value theorem
\begin{equation}\label{p6l3}
\begin{aligned}
|g^{-1}(E+D-b^d)- g^{-1}(E-D-b^d)| = 2D (g^{-1})'(x) ,
\end{aligned}
\end{equation}
for some $E-D-b^d< x< E+D-b^d.$ Observe that since both $g',g^{-1}$ are strictly increasing on the positive real axis, $ (g^{-1})'$ is strictly decreasing there. So \eqref{p6l3} can be bounded by $2D (g^{-1})'(E-D-b^d).$
We have from \eqref{comgg} for $y>0$
\begin{equation*}\label{}
\begin{aligned}
(g^{-1})'(y)=\frac{1}{g'(g^{-1}(y))}\leq \frac{g^{-1}(y)}{g(g^{-1}(y))}=\frac{g^{-1}(y)}{y}.
\end{aligned}
\end{equation*}
Combining this with \eqref{p6l4}
\begin{equation}\label{p6l5}
\begin{aligned}
(g^{-1})'(E-D-b^d) \leq \frac{g^{-1}(E-D-b^d)}{E-D-b^d}\leq \min \Big\{\frac{1}{(E-D-b^d)^{\frac{d-2}{d-1}}(db)^{\frac{1}{d-1}}}, \frac{1}{db^{d-1}} \Big\} .
\end{aligned}
\end{equation}
We can bound {\bf I} as follows
\begin{equation*}\label{}
\begin{aligned}
{\bf I}\leq \sum_{1\leq b<(E-2D)^{\frac{1}{d}}}1+ {2D}(g^{-1})'(E-D-b^d)\leq (E-2D)^{\frac{1}{d}} + 2D\Big[\sum_{1\leq b<(E-2D)^{\frac{1}{d}}}\frac{g^{-1}(E-D-b^d)}{(E-D-b^d)}\Big].
\end{aligned}
\end{equation*}
As we assumed $E\leq D^2$ we can bound $(E-2D)^{1/d}\leq D^{2/d}$. We decompose the last sum in paranthesis, and bound it by \eqref{p6l5}
\begin{equation*}\label{}
\begin{aligned}
&\leq \sum_{1\leq b<D^{\frac{1}{d}}}\frac{1}{(E-D-b^d)^{\frac{d-2}{d-1}}b^{\frac{1}{d-1}}}+ \sum_{D^{\frac{1}{d}}\leq b<(E-2D)^{\frac{1}{d}}}\frac{1}{db^{d-1}}\\ &\leq \frac{2}{(E-D)^{\frac{d-2}{d-1}}}\Big[\sum_{1\leq b<D^{\frac{1}{d}}}b^{-\frac{1}{d-1}}\Big]+ \frac{1}{d}\Big[\sum_{D^{\frac{1}{d}}\leq b<(E-2D)^{\frac{1}{d}}}b^{1-d}\Big].
\end{aligned}
\end{equation*}
We can estimate these sums by
\begin{equation*}\label{}
\begin{aligned}
\sum_{1\leq b<D^{\frac{1}{d}}}b^{-\frac{1}{d-1}}\leq 1+ \int_{1}^{D^{\frac{1}{d}}} b^{-\frac{1}{d-1}}db=1+\frac{d-1}{d-2}\Big[b^{\frac{d-2}{d-1}} \Big|^{D^{\frac{1}{d}}}_1 \Big]\leq 2D^{\frac{d-2}{d(d-1)}},
\end{aligned}
\end{equation*}
and
\begin{equation*}\label{}
\begin{aligned}
\sum_{D^{\frac{1}{d}}\leq b<
(E-2D)^{\frac{1}{d}}}b^{1-d}\leq D^{\frac{1-d}{d}}+ \int_{D^{\frac{1}{d}}}^{\infty}b^{1-d}db \leq 2D^{-1+\frac{2}{d}}.
\end{aligned}
\end{equation*}
Combining all of these we can write
\begin{equation*}\label{}
\begin{aligned}
{\bf I} \leq D^{\frac{2}{d}} +2D\Big[\frac{4D^{\frac{d-2}{d(d-1)}}}{(E-D)^{\frac{d-2}{d-1}}}+ \frac{2}{d}D^{-1+\frac{2}{d}}\Big]\leq D^{\frac{2}{d}} +8D^{1+\frac{d-2}{d(d-1)}-\frac{d-2}{d-1}}+ 2D^{\frac{2}{d}}\leq 11D^{\frac{2}{d}}.
\end{aligned}
\end{equation*}
This finishes the proof.
\end{proof}
We turn to proving our last theorem. The proof follows the same lines along our first proof of Theorem 8.
\begin{proof}
Let $D^s\geq 2D.$ We apply \eqref{eta3} for the values $D^s+D$ and $D^s-D$, and subtract to bound the left hand side of \eqref{p6sl3} by
\begin{equation}\label{fths2}
\begin{aligned}
&\leq \frac{1}{4}\Big[A_d\big[(D^s+D)^{2/d}-(D^s-D)^{2/d}\big] +B_d \big[(D^s+D)^{1/(d-1)}-(D^s-D)^{1/(d-1)}\big]\Big]\\ &-\frac{1}{2}\big[ (D^s+D)^{1/d}-(D^s-D)^{1/d} \big]+\Oh_{d}(D^{\frac{s}{d}[1-\frac{1}{d}]}).
\end{aligned}
\end{equation}
We have the estimates by the mean value theorem
\begin{equation*}\label{}
(D^s+D)^{\frac{2}{d}}-(D^s-D)^{\frac{2}{d}}\leq 4 D^{1+s(\frac{2}{d}-1)}, \qquad (D^s+D)^{\frac{1}{d-1}}-(D^s-D)^{\frac{1}{d-1}}\leq 4 D^{1+s(\frac{1}{d-1}-1)}.
\end{equation*}
Of these plainly the first one dominates. Comparing it with the error term in \eqref{fths2} yields the desired result.
If $D^s< 2D$ then \eqref{eta3} can be applied with $x=3D$ to show that our set has cardinality at most $C_dD^{2/d}$. But $D^s< 2D$ leads to
\[D^{\frac{2}{d}}=DD^{\frac{2}{d}-1}\leq D(D^s/2)^{\frac{2}{d}-1}\leq 2D^{1+s(\frac{2}{d}-1)}.\]
This finishes the proof.
\end{proof}
\begin{thebibliography}{99}
\bibitem{bac}
G.F. Bachelis, \emph{On the upper and lower majorant properties in $L^p(G)$}, Quart. J. Math. Oxford Ser. (2) {\bf 24} (1973), 119-128.
\bibitem{bp}
E. Bombieri, J. Pila,\emph{The number of integral points on arcs and ovals}, Duke Math. J. {\bf 59} (1989), no. 2, 337-357.
\bibitem{bou1} J. Bourgain, \emph{Bounded orthogonal systems and the $\Lambda(p)$-set problem}, Acta Math. {\bf 162} (1989), no. 3-4, 227–245.
\bibitem{bou2} J. Bourgain, \emph{On $\Lambda (p)$-subsets of squares}, Israel J. Math. {\bf 67} (1989), no. 3, 291–311.
\bibitem{bhb}
T.D. Browning, D.R. Heath-Brown, \emph{Equal sums of three powers}, Invent. Math. {\bf 157}, (2004), 553-573.
\bibitem{c}
M-C. Chang, \emph{On problems of Erdös and Rudin}, J. Funct. Anal. {\bf 207} (2004), no. 2, 444-460.
\bibitem{cck}
P. Chareka, O. Chareka, S. Kennedy,\emph{Locally sub-gaussian random variables and the strong law of large numbers}, Atlantic Elec. J. Math. {\bf 1} (2006), no 1.
\bibitem{cg}
J. Cilleruelo, A. Granville, \emph{Lattice points on circles, squares in arithmetic progressions and sumsets of squares}, arXiv preprint math/0608109
\bibitem{d}
L. Daskalakis,\emph{Roth’s theorem and the hardy–littlewood majorant problem for
thin subsets of primes}, J. of Funct. Analysis, {\bf 286} (2024), no. 5.
\bibitem{dem}
C. Demeter,\emph{Fourier restriction, decoupling and applications}, Cambridge University Press, 2020
\bibitem{djr}
C. Demeter, H. Jung, D. Ryou, \emph{Maximal $\Lambda(p)$ subsets of manifolds}, https://arxiv.org/abs/2411.04248
\bibitem{dir}
P.G.L. Dirichlet, \emph{Über die Bestimmung der mittleren Werte der Zahlentheorie,} Abh. Konigl. Preuss. Akad.
Wiss. (1849), 69-83.
\bibitem{eh}
M. Erraoui , Y. Hakiki, \emph{ Images of fractional Brownian motion with deterministic drift: Positive Lebesgue measure and non-empty interior}, Mathematical Proceedings of the Cambridge Philosophical Society. {\bf 173} (2022), no. 3, 693-713.
\bibitem{fmd}
W. Fouch\'{a}, S. Mukeru, G. Davie, \emph{Fourier spectra of measures associated with algorithmically random Brownian motion}, Log. Methods in Comput. Sci. {\bf 10} (2014), no. 3, 1-24.
\bibitem{fs}
J. Fraser, T. Sahlsten, \emph{On the Fourier analytic structure of the Brownian graph}, Analysis and PDE, {\bf 11}
(2018) no. 1, 115-132.
\bibitem{gg}
L. Galeati, M. Gubinelli, \emph{Prevalence of $\rho$-irregularity and related properties}, arXiv preprint arXiv:2004.00872
\bibitem{gk}
S.W. Graham, G.Kolesnik, \emph{Van der Corput's method of exponential sums}, Cambridge University Press, 1991
\bibitem{gre}
P.T. Gressman, S. Guo, L.B. Pierce, J. Roos, P-L. Yung, \emph{On the Strict Majorant Property in Arbitrary Dimensions,} Quart. J. Math. {\bf 74} (2023), no. 1, 139–161.
\bibitem{gh1}
G.H. Hardy, \emph{On the expression of a number as the sum of two squares}, Quart. J. Math. {\bf 46} (1915),
263-283.
\bibitem{gh2}
G.H. Hardy, \emph{On Dirichlet’s divisor problem,} Proc. London Math. Soc. {\bf 15} (1916), no. 2, 1-25.
\bibitem{hl} G.H. Hardy and J.E. Littlewood, \emph{Notes on the theory of series (XIX): A problem concerning majorants of Fourier
series}, Quart. J. Math. (1935), no. 1, 304–315.
\bibitem{gr} B. Green and I.Z. Ruzsa, \emph{On the Hardy–Littlewood majorant problem}, Math. Proc. Cambridge Philos. Soc. {\bf 137} (2004), no. 3,
511–517.
\bibitem{hb}
D.R. Heath-Brown, \emph{The density of rational points on curves and surfaces}, Ann. Math. {\bf 155} (2002), 553-595.
\bibitem{hb2}
D.R. Heath-Brown, \emph{Counting rational points on algebraic varieties}, https://link.springer.com/content/pdf/10.1007/978-3-540-36364-4\_2.pdf
\bibitem{hux1}
M.N. Huxley, \emph{ Exponential sums and lattice points II,} Proc. London Math. Soc. {\bf 66} (1993), no. 3, 279-301.
\bibitem{hux2}
M.N. Huxley, \emph{ Area, lattice points, and exponential sums,} LMS Monographs 13, Oxford University
Press, Oxford, 1996.
\bibitem{hux3}
M.N. Huxley, \emph{ Exponential sums and lattice points III}, Proc. London Math. Soc. {\bf 87} (2003), no. 3,
591-609.
\bibitem{ikkn}
A. Ivi\'{c}, E. Kr\"{a}tzel, M. Kühleitner, W.G. Nowak, \emph{Lattice points in large regions and related arithmetic functions:
Recent developments in a very classic topic}, https://arxiv.org/pdf/math/0410522
\bibitem{k}
J-P. Kahane,\emph{Some random series of functions}, Second edition, Cambridge University Press, 1985
\bibitem{ks}
L.B. Koralov, Ya. G. Sinai,\emph{Theory of probability and random processes}, 2nd edition, Springer Universitext, 2007
\bibitem{krat}
E. Kr\"{a}tzel,\emph{ Lattice points}, Kluwer, Dordrecht-Boston-London, 1988.
\bibitem{krat2}
E. Kr\"{a}tzel, \emph{Mittlere Darstellungen natürlicher Zahlen als Differenz zweier k-ter Potenzen,} Acta Arithm. {\bf 16}
(1969), 111–121.
\bibitem{kmt}
B. Krause, Ben, M. Mirek, B. Trojan.\emph{On the Hardy-Littlewood majorant problem for arithmetic
sets.} J. Funct. Anal. {\bf271} (2016), no. 1, 164–181.
\bibitem{kre}
S. Krenedits, \emph{On Mockenhoupt’s conjecture in the Hardy-Littlewood majorant problem}, J. Contemp. Math. Anal. {\bf 48} (2013), no. 3, 91–109.
\bibitem{kuba}
G. Kuba, \emph{On sums of two k-th powers of numbers in residue classes II}, Abh. Math. Sem. Univ. Hamburg {\bf 63} (1993), 87-95.
\bibitem{khl1}
M. Kühleitner,\emph{An Omega theorem on differences of two squares,} Acta Math. Univ. Comenian. {\bf 41} (1992),
117–123.
\bibitem{g}
G. Leclerc, \emph{On oscillatory integrals with Hölder phases}, arXiv preprint arXiv:2211.08088
\bibitem{m}
K. Mahler, \emph{Note on the hypothesis K of Hardy and Littlewood}, J. London Math. Soc. {\bf 11} (1936), no. 2, 136-138.
\bibitem{mar}
O. Marmon, \emph{Sums and differences of four kth powers}, Monatsch. Math. {\bf 164} (2011), no. 1,
55-74.
\bibitem{moc}
G. Mockenhaupt, \emph{Bounds in Lebesgue spaces of oscillatory integrals}, Habilitationsschrift, Siegen, 1996
\bibitem{ms} G. Mockenhaupt and W. Schlag, \emph{On the Hardy-Littlewood majorant problem for random sets}, J. Funct. Anal. {\bf 256} (2009), no. 4, 1189-1237.
\bibitem{mu}
S. Mukeru, \emph{The Zero Set of Fractional Brownian Motion
Is a Salem Set}, J. Fourier Anal. Appl. {\bf 24} (2018), no. 4, 957–999.
\bibitem{nowak1}
W.G. Nowak, \emph{On differences of two k-th powers of integers,} Ramanujan J. {\bf 2}, (1998), no. 4, 421-440.
\bibitem{nowak2}
W.G. Nowak, \emph{On sums and differences of two relative prime cubes,} Analysis, {\bf 15} (1995), no. 4, 325-341.
\bibitem{pz1}
R. E. A. C. Paley, A. Zygmund,\emph{On some series of functions (1)}, Math. Proc. Cambridge Phil. Soc. {\bf 26} (1930), no. 3, 337-357.
\bibitem{pz2}
R. E. A. C. Paley, A. Zygmund,\emph{On some series of functions (2)}, Math. Proc. Cambridge Phil. Soc. {\bf 26} (1930), no. 4, 458-474.
\bibitem{rob}
H. Robbins, \emph{
The American Mathematical Monthly}, {\bf 62} (1955), no. 1, 26-29.
\bibitem{ru}
W. Rudin, \emph{Trigonometric series with gaps}, J. of Math. and Mech. {\bf 9} (1960), no. 2, 203-227.
\bibitem{cs}
C. Sahillioğulları, \emph{Randomization of certain operators in harmonic analysis}, İzmir Institute of Technology, PhD thesis, 2024
\bibitem{s}
P. Salberger, \emph{Counting rational points on hypersurfaces of low dimension}, Ann. Scient. Ec. Norm. Sup., $4^e$ serie, {\bf 38} (2005), 93-115.
\bibitem{sw}
P. Salberger, T.D. Wooley, \emph{Rational points on complete intersections of higher degree, and mean values of Weyl sums}, J. London Math. Soc. {\bf 82} (2010), no.2, 317-342.
\bibitem{san}
T. Sanders, \emph{On the Bogolyubov-Rusza lemma}, Analysis and PDE,Vol. {\bf 5} (2012), no. 3, 627-656.
\bibitem{ft1}
F. Temur,\emph{Linear and bilinear restriction estimates for the Fourier transform}, PhD thesis, University of Illinois at Urbana-Champaign
\bibitem{ft2}
F. Temur, \emph{A Fourier restriction estimate for surfaces of positive curvature in $\R^6$}, Rev. Mat. Iberoamericana, {\bf 30} (2014), no. 3, 1015-1036.
\bibitem{ft3}
F. Temur, \emph{An endline bilinear cone restriction estimate for mixed norms}, Math. Z. {\bf 273} (2013), no. 3, 1197-1214.
\bibitem{v}
R.C. Vaughan,\emph{The Hardy-Littlewood Method}, Second edition, Cambridge University Press, 1997
\bibitem{vor}
G. Voronoi, \emph{Sur un probleme du calcul des fonctions asymptotiques,} J. Math. {\bf 126} (1903), 241-282.
\bibitem{x}
Y. Xiao, \emph{Sample Path Properties of Anisotropic Gaussian Random Fields} In: Khoshnevisan, D., Rassoul-Agha, F. (eds) A Minicourse on Stochastic Partial Differential Equations. Lecture Notes in Mathematics, vol 1962. Springer, Berlin, Heidelberg (2009).
\end{thebibliography}
\end{document}