An upper bound for weak $B_k$-sets
Extracted source text ยท arXiv 1611.06414
\documentclass[11pt,twoside]{article}
\setlength{\textwidth}{160mm} \setlength{\textheight}{210mm}
\setlength{\parindent}{8mm} \frenchspacing
\setlength{\oddsidemargin}{0pt} \setlength{\evensidemargin}{0pt}
\thispagestyle{empty}
\usepackage{mathrsfs,amsfonts,amsmath,amssymb}
\usepackage{latexsym}
\newenvironment{dedication}
{
\thispagestyle{empty}% no header and footer
%\vspace*{\stretch{1}}% some space at the top
\itshape % the text is in italics
\raggedleft % flush to the right margin
}
\pagestyle{myheadings}
%\input{polchar}
\markboth{\centerline{\sc{\small An upper bound for weak $B_k$-sets }}}
{\centerline{\sc{\small T.~Schoen, I.~D.~Shkredov }}}
\newtheorem{satz}{Theorem}
\newtheorem{proposition}[satz]{Proposition}
\newtheorem{theorem}[satz]{Theorem}
\newtheorem{lemma}[satz]{Lemma}
\newtheorem{definition}[satz]{Definition}
\newtheorem{corollary}[satz]{Corollary}
\newtheorem{remark}[satz]{Remark}
\newtheorem{example}[satz]{Example}
\newcommand{\lf}{\left\lfloor}
\newcommand{\rf}{\right\rfloor}
\newcommand{\lc}{\left\lceil}
\newcommand{\rc}{\right\rceil}
%\newcommand{\qed}{{} \hfill \mbox{$\Box$}}
%\renewcommand{\baselinestretch}{1.5}
\renewcommand{\thefootnote}{}
%\renewcommand{\labelenumi{\roman{enumi}}}
\def\proof{\par \noindent \hbox{\rm P\,r\,o\,o\,f.}~}
\def\L{\Lambda}
\def\c{\circ}
\def\no{\noindent}
\def\sbeq{\subseteq}
\def\T{\mathsf{T}}
\def\N{\mathbb {N}}
\def\Z{\mathbb {Z}}
\def\F{\mathbb {F}}
\def\E{\mathsf{E}}
\def\e{\varepsilon}
\def\l{\lambda}
\def\s{\sigma}
\def\I{{\cal I}}
\def\a{\alpha}
\def\S{\mathcal{S}}
\def\C{\mathbb{C}}
\def\P{{\cal P}}
\def\h{\widehat}
\def\p{\varphi}
\def\d{\delta}
\def\o{\omega}
\def\({\big (}
\def\){\big )}
\def\g{\gamma}
\def\G{\mathcal{G}}
\def\X{\mathcal{X}}
%\def\s{\setminus}
\def\b{\beta}
\def\ls{\leqslant}
\def\gs{\geqslant}
\def\dim{{\rm dim}}
\def\le{\leqslant}
\def\ge{\geqslant}
\def\_phi{\varphi}
\def\x{\xi}
\def\eps{\varepsilon}
\def\m{\times}
\def\Gr{{\mathbf G}}
\def\FF{\widehat}
\def\t{\times}
\def\k{\kappa}
\def\snd{{\rm {snd}}}
\def\ov{\overline}
\def\Spec{{\rm Spec\,}}
\def\R{\mathbb {R}}
\def\B{\Big}
\renewcommand{\thetheorem}{\arabic{theorem}.}
\begin{document}
\title{\bf An upper bound for weak $B_k$--sets }
%\author{ By\\ \\{\sc Tomasz Schoen\footnote{\textsc{The author is supported by NCN grant .....} }}}
\author{ By\\ \\{\sc Tomasz Schoen\footnote{The author is supported by NSC grant ...}
~ and ~ Ilya D.~Shkredov
%\footnote{This work was supported by grant RFFI NN
%06-01-00383, 11-01-00759 and grant Leading ScientificSchools N 8684.2010.1.}
} }
\date{}
\maketitle
%\begin{dedication}
%\end{dedication}
\begin{abstract} We prove that if $A\sbeq [N]$ does not contain any solution to the equation
$x_1+\dots+x_k=y_1+\dots+y_k$ with distinct $x_1,\dots,x_k,y_1,\dots,y_k\in A$, then $|A|\ll {k^{3/2}}N^{1/k}.$
\end{abstract}
\footnote{{\it 2010 Mathematics Subject Classification: } primary
11B99.}
%\centerline{\bf Introduction}
\bigskip
\section{Introduction}
A set $A\sbeq \Z$ is called a $B_k$--set, $k\ge 2,$ if
\begin{equation}\label{B_k}
x_1+\dots+x_k=y_{1}+\dots+y_{k}
\end{equation}
with $x_i,y_i\in A$ implies that $x_1,\dots,x_k$ is a permutation of $y_{1},\dots,y_{k}.$ If $A$ is a $B_k$-subsets of $[N]$, then even a simple argument argument $\binom{|A|+k-1}{k}=|kA|\le kN$ shows that
$$|A|\le (kk!)^{1/k} N^{1/k}.$$
Currently best bound
$$|A|\le (1+o(1))(1+\e_k)\frac{k}{2e}N^{1/k}$$
was proved by Green \cite{green} (see also \cite{obryant} page 14), where $\e_k\rightarrow 0$ as $k\rightarrow \infty$.
The history of the problem and bibliography can be found in \cite{obryant}.
We call a set $A\sbeq \Z$ a weak $B_k$--set ($B^*_k$--set) if
every solution to (\ref{B_k})
with $x_i\in A$ implies that among numbers $x_1,\dots,x_k,y_{1},\dots,y_{k}$ at least two are equal. In other words $A$ does not contain a solution to (\ref{B_k}) with distinct elements.
It is already non-trivial to prove that every $B^*_k$--subset of $[N]$ has size $O_k(N^{1/k}).$ It is a fact that there are equations (see Theorem 3.3 in \cite{ruzsa-I}) such that the threshold for the size of a set without solutions in distinct integers is of different order of magnitude. Ruzsa \cite{ruzsa-I} proved that for every $B^*_k$--set $A\sbeq [N]$ we have
$$|A|\le (1+o(1))k^{2-1/k}N^{1/k} \,.$$
Very recently Timmons \cite{timmons} showed that
$$|A|\le (1+o(1))(1+\e_k)\frac{k^{2}}{4}N^{1/k},$$
where $\e_k\rightarrow 0$ as $k\rightarrow \infty$. On the other hand, for $k\ge 3$ Timmons constructed a $B^*_k$--subset of $[N]$ with $(1-o(1))2^{1-1/k}N^{1/k}$ elements. Our aim is to prove the following theorem.
\begin{theorem}
Let $A\subseteq [N]$ be a $B^*_k$--set and $N\ge (2 k^{9/2})^k$.
Then
$$
|A| \le 16 k^{3/2} N^{1/k} \,.
$$
\label{t:main_introduction}
\end{theorem}
\section{The proof of the main result}
For a finite subset of integers $A$ put
$$\s_k(n) = \s_k(n;A)= |\{(x_1,\dots,x_k)\in A^k: x_1+\dots+x_ k=n\} | \,,$$
$$f(t)=f_A(t)=\sum_{a\in A}e^{-2\pi i at} \,, \quad \quad t\in [0,1]$$
further
$$M_k = M_k (A) =\sum_n \s_k(n)^2=\int_0^1|f(t)|^{2k}dt \,,$$
$$s_k(n) = s_k (n;A) = | \{(x_1,\dots,x_k)\in A^k,\, x_i\not=x_j: x_1+\dots+x_ k=n\} |,$$
$$S_k = S_k (A) =\sum_n s_k (n;A)^2 = \sum_n s_k (n)^2
=
$$
\begin{equation}\label{f:S_k_equality}
=
| \{ (x_1,\dots,x_k, y_1,\dots,y_k )\in A^{2k}~:~ x_1+\dots+x_ k = y_1 +\dots+y_k
,\, x_i\not=x_j,\, y_i \not=y_j\}| \,.
\end{equation}
Furthermore, define
\begin{eqnarray*}
S^*_k(A)= | \{ (x_1,\dots,x_k, y_1,\dots,y_k )\in A^{2k}&:&x_1+\dots+x_ k = y_{1} +\dots+y_{k}
,\\ && x_i \not=x_j,\, y_i\not=y_j,\, x_i\not=y_j,\, 1\le i,j\le k \}| \,.
\end{eqnarray*}
and for finite subsets of integers $A_1,A_2$ put
\begin{eqnarray*}
S^*_k(A_1,A_2)= | \{ (x_1,\dots,x_k, y_1,\dots,y_k )\in A_1^{k}\times A_2^k &:&x_1+\dots+x_ k = y_{1} +\dots+y_{k}
,\\ && x_i \not=x_j,\, y_i\not=y_j,\, x_i\not=y_j,\, 1\le i,j\le k \}| \,.
\end{eqnarray*}
Clearly, $A$ is a $B^*_k$--set $A$ if and only if $S^*_k (A)=0$
\medskip
Our first lemma can be extracted from Ruzsa's paper \cite{ruzsa-I}.
We will rewrite it for a convenient for us form.
\begin{lemma}\label{l_n:ruzsa}
Let $A$ be a finite set of integers with $|A| \ge 16 k^4$.
Then
\begin{equation}\label{e:ruzsa-S_k}
2^{-k} k!|A|^k \le S_k \le M_k \le 2 S_k \,.
\end{equation}
\label{l:M_k=S_k}
\end{lemma}
\begin{proof}
If $k=1$, then the result is trivial, thus suppose that $k\ge 2$.
First observed that
\begin{equation}\label{inequality-Mk}
M_k\ge S_k\ge k!|A|(|A|-1)\dots (|A|-k+1) \ge (5/4)^{-k} k! |A|^k \ge 2^{-k} k!|A|^k
\,.
\end{equation}
Ruzsa proved (see inequality (5.8) in \cite{ruzsa-I}) that
$$
\sum_n (\sigma_k (n) - s_k (n))^2 \le k^4 M_{k-1} \,,
$$
hence by the H\"{o}lder inequality
\begin{equation*}\label{tmp:03.06.2016_1}
\sum_n (\sigma_k (n) - s_k (n))^2 \le k^4 M^{\frac{k-2}{k-1}}_k |A|^{\frac{1}{k-1}} \,.
\end{equation*}
Therefore, by the triangle inequality
$$M^{1/2}_k\le S^{1/2}_k+k^2 M^{\frac{k-2}{2(k-1)}}_k |A|^{\frac{1}{2(k-1)}}\,.$$
By the middle inequality in (\ref{inequality-Mk}) and the assumption $|A| \ge 16 k^4$ we see that
$$
|M^{1/2}_k - S^{1/2}_k|^2 \le k^4 M^{\frac{k-2}{k-1}}_k |A|^{\frac{1}{k-1}}
%\ll \Delta
\le 2^{-4} M_k
%\,,
$$
so that
$$
M_k \le 2 S_k
$$ as required.
\hfill$\blacksquare$
\end{proof}
\bigskip
Observe that, actually, the inequality (\ref{e:ruzsa-S_k}) can be strengthened to $M_k = S_k (1+o(1))$ for large $|A|$.
Our next lemma provides a straightforward relation between $S_k$ and $S_l$ for $l<k$.
\begin{lemma}
Let $1\le l<k$ and let $A$ be a $B^*_k$--set with $|A| \ge 16 k^4$.
Then
$$
S_k \le (\sqrt{2} k)^{2{(k-l)}} |A|^{k-l} S_{l} \,.
$$
\label{l:S_k,S_l}
\end{lemma}
\begin{proof}
Indeed, counting the number of the solutions to (\ref{f:S_k_equality}), we see that by the assumption on $A$ we must have $x_i=y_j$ for some $i$ and $j$. There are at most $k^2$ choices for $i$ and $j$ and thus
\begin{equation}\label{ruzsa-k2_my}
S_k \le k^2 |A| S_{k-1} \le k^2 |A| M_{k-1} \,.
\end{equation}
Furthermore, by the H\"older inequality and Lemma \ref{l:M_k=S_k}
$$
M_{k-1}
=
\int_0^1 |f(t)|^{2k-2}dt
\le \B(\int_0^1 |f(t)|^{2k}dt\B)^{\frac{k-l-1}{k-l}}
\B(\int_0^1 |f(t)|^{2l}dt\B)^{\frac1{k-l}}
\le
(2S_k)^{\frac{k-l-1}{k-l}}(2S_l)^{\frac1{k-l}} \,,
$$
hence by (\ref{ruzsa-k2_my})
$$ S_k\le 2^{}k^2|A| S_k^{\frac{k-l-1}{k-l}} S_l^{\frac1{k-l}}$$
and the assertion follows.
% as required.
\hfill$\blacksquare$
\end{proof}
\bigskip
Using a simple probabilistic argument we prove that the variables from both sides of (\ref{f:S_k_equality}) can be chosen from disjoint sets.
This allows us to apply a version of a well known theorem of Erd{\H o}s--Ko--Rado, see Theorem \ref{t:EKR} below.
\begin{lemma}\label{l:S_k=S_k(A,A^c)}
Let $A$ be finite set of integers. Then there exists a set $A' \subseteq A$ such that
$$
S^*_k (A) \le
4^k S^*_k (A',A\setminus A')\,.
$$\end{lemma}
\begin{proof} Let us pick, randomly, independently, each element of $A$ with probability $1/2.$
Let $A'$ be the random set of chosen elements. Clearly,
% $$
% \mathbb{E} |A'| = |A| /2
% %\,.
% $$
% and from the large deviations inequality (see e.g. \cite{TV}, Theorem 1.8) we have $|A|/4 \le |A'| \le 3|A|/4$ with probability tending to $1$, %as $|A|\rightarrow \infty$.
% %Clearly,
% Furthermore,
\begin{equation*}\label{tmp:04.06.2016_1}
\mathbb{E} S^*_k (A',A\setminus A')
= \sum_{x_1+\dots+x_k = x'_1 + \dots +x'_k }2^{-2k} \,,
\end{equation*}
where the summation is taken over all solutions in the set $A$ with distinct integers.
Whence,
$$\mathbb{E} S^*_k (A',A\setminus A')=4^{-k}S^*_k(A)\,,$$
so that
$$\mathbb{P}\big(\,S^*_k (A',A\setminus A')\ge 4^{-k}S^*_k(A)\, \big)>0\,.$$
%and by the Markov inequality the probability that $S^*_k (A',A\setminus A')\ge 2\cdot 4^{-k}S^*_k(A)$ is less or equal $1/2.$
Therefore, there is a specific choice of $A'$ such that $S^*_k (A) \le
4^k S^*_k (A',A\setminus A').$
% Because of the sum in (\ref{tmp:04.06.2016_1}) is taken over distinct $x_1,\dots,x_k$ as well as over distinct $x'_1,\dots,x'_k$ we see that all %random variables in the formula are independent.
% Hence
%$$
% \mathbb{E} S^*_k (A',A\setminus A') = \sum^*_{x_1+\dots+x_k = x'_1 + \dots +x'_k }
% \mathbb{E} \xi_{x_1} (\omega) \dots \mathbb{E} \xi_{x_k} (\omega)
% \mathbb{E} (1-\xi_{x'_1} (\omega)) \dots \mathbb{E} (1-\xi_{x'_k} (\omega))
%$$
%$$
% =
% 2^{-2k} S^*_k (A)
% %\,.
%$
\hfill$\blacksquare$
\end{proof}
\bigskip
For a $B_k^*$--set $A$ define
$$
\mathcal{G} = \mathcal{G} (A) = \{ j \in [k] ~:~ A \text{ is a $B_j^*$--set} \}
\quad \quad \mbox{ and } \quad \quad
\mathcal{X} = \mathcal{X} (A) = [k] \setminus \mathcal{G} (A) \,.
$$
Timmons \cite{timmons} proved that a large subset of $A$ is a $B^*_{\lfloor k/2\rfloor}$ or $B^*_{\lceil k/2\rceil}$--set.
We prove that there is a large set $B\sbeq A$ such that the set $\mathcal{G}(B)$ is very large and structural. This is a crucial argument in our approach.
%
%Now we need a lemma on the structure of a set $\mathcal{G} (A_1)$ for large subset of $A$.
\bigskip
\begin{lemma}
Let $A$ be a $B_k^*$--set.
Then there is a set $B \subseteq A$,
$$
|B| \ge |A| - 2 k^3
$$
such that for $\mathcal{X} = \mathcal{X} (B)$, $\G = \G(B)$ we have for any $l$ we have $l \mathcal{X} \cap \G = \emptyset \,.$
%one has
In particular,
\begin{equation}\label{in:X}
|\X\cap [l]|\le \big\lfloor\frac{l}2\big\rfloor
\end{equation}
\label{l:structure_X}
for every $1\le l\le k.$
\end{lemma}
\begin{proof}
We prove the lemma using an iterative procedure. We start with the sets $A_0=A$, $\G_0=\emptyset,$ $\X_0=\emptyset$ and $\S_0=\emptyset.$ Suppose that we have chosen
$A_i, \G_i, \X_i$ and $\S_i.$ If $[k]=\G_i\cup \X_i$ we stop the process, otherwise put $l:=\min([k]\setminus (\G_i\cup \X_i))$. If $A_i$ contains $k$ pairwise disjoint solutions to the equation
\begin{equation}\label{eq-l}
x_1+\dots+x_{l}=y_1+\dots+y_{l}
\end{equation}
with all variables distinct, and disjoint with $\S_i,$ then we put
\begin{equation}\label{step-1}
A_{i+1}=A_i,\,~ \G_{i+1}=\G_i,\,~ \X_{i+1}=\X_i\cup \{l\},\,~\S_{i+1}=\S_{i}\cup \textsf{Sol}_i\,,
\end{equation}
where $\textsf{Sol}_i$ is the set of all numbers involved in the $k$ disjoint solutions to (\ref{eq-l}).
If this is not the case (there are less than $k$ such solutions), let $\textsf{Sol}_i$ be a maximal collection of pairwise disjoint solutions to (\ref{eq-l}) with distinct variables,
and disjoint with $\S_i$. Then
we put
\begin{equation}\label{step-2}
A_{i+1}=A_i\setminus ( \S_i\cup \textsf{Sol}_i),\,~ \G_{i+1}=\G_i\cup\{l\},\,~ \X_{i+1}=\emptyset,\,~ \S_{i+1}=\emptyset\,.
\end{equation}
Observe that then $A_{i+1}$ is a $B^*_l$--set, and $l\in \G_j,$ for all $ j\ge i+1.$ Thus,
the step (\ref{step-2}) can be applied at most $k$ times, as in each application we add one element to the set $\G_i.$ Therefore, the process must terminate after $m\le k^2$ iterations. We put $B=A_m, \G=\G_m$ and $\X=\X_m.$ Since
$|\S_i\cup \textsf{Sol}_i|\le 2k^2$
it follows that
$$|B|\ge |A|-2k^3.$$
Next, we prove that $l \mathcal{X} \cap \G = \emptyset \,.$ Suppose that it does not hold. Then there exist $n_1,\dots,n_l\in \X,\, l\le k,$ such that $n_1+\dots+n_l=n\in \G.$
By our construction there are pairwise disjoint solutions in distinct integers
\begin{eqnarray*}
x^{(1)}_1+\dots+x^{(1)}_{n_1}&=&y^{(1)}_1+\dots+y^{(1)}_{n_1}\\
&\dots&\\
x^{(l)}_1+\dots+x^{(l)}_{n_l}&=&y^{(l)}_1+\dots+y^{(l)}_{n_l}.
\end{eqnarray*}
Summing all these solutions up we obtain a solution to the equation
$$x_1+\dots+x_{n}=y_1+\dots+y_{n}$$
in distinct integers $x_i,y_i\in B$, which contradicts that $B$ is a $B^*_n$--set.
It remains to prove (\ref{in:X}). Observe that if $l\in \G$ it follows from $2\X\cap \G=\emptyset$ that
$$|\X\cap [l]|\le \big\lceil\frac{l}{2}\big\rceil-1\le \big\lfloor\frac{l-1}{2}\big\rfloor\,,$$ otherwise there are $n_1,n_2\in \X$ such that $n_1+n_2=l,$ which is impossible. Now, let $l\in \X$
and suppose that $|\X\cap [l]|> \lfloor\frac{l}{2}\rfloor$. Let $n$ be the smallest element of $\G$ bigger than $l$. Then
$$\big\lceil\frac{n}{2}\big\rceil-1\ge |\X\cap [n]|=|\X\cap [n-1]|> \big\lfloor\frac{n-1}{2}\big\rfloor\,,$$
which again is a contradiction. This completes the proof.
\hfill$\blacksquare$
\end{proof}
\bigskip
Clearly, the property $l \mathcal{X} \cap \G = \emptyset$ means that if $x_1,\dots,x_l\in \mathcal{X}$ such that $x_1+\dots+x_l \in [k]$ then automatically $x_1+\dots+x_l \in \mathcal{X}$.
Potential examples of "bad" $\mathcal{X}$ is the set of even numbers from $[k]$ or $\X=(k/2,k-1]$. However, we do not know any example of a
dense $B^*_k$--subset of $[N]$ that does not contain a large subset that is $B^*_l$--set for every $l\le k.$
In the proof of the main result we also apply a variant of a well--known theorem of Erd{\H o}s--Ko--Rado \cite{ER}, \cite{EKR}.
Let $1\le k,n$, $k< n$ be integer parameters and let $L = \{ l_1< l_2 < \dots < l_r \} \subseteq [0,k]$.
A family $\mathcal{A}$ of subsets of $[n]$ is called $(n,k,L)$--system if, each of a subset of the family has cardinality $k$ and, for all $A, B\in \mathcal{A}$ from the family we have $|A\cap B|\in L.$ The following theorem was proved in \cite{DEF}.
\begin{theorem}
Let $\mathcal{A}$ be a $(n,k,L)$--system.
Then
$$
|\mathcal{A}| \le \prod_{i=1}^{|L|} \frac{n-l_i}{k-l_i} \,.
$$
\label{t:EKR}
\end{theorem}
Now we are in position to prove Theorem \ref{t:main_introduction}.
It is a consequence of the following result.
\begin{theorem}
Let $A$ be a $B^*_k$--set, $|A| \ge 20 k^{6}$.
Then
\begin{equation}\label{f:M_k_main}
M_k (A) \le 2k 8^kk^{3k/2} |A|^k \,.
\end{equation}
\label{t:M_k_main}
\end{theorem}
To see how the theorem above implies Theorem \ref{t:main_introduction} just apply (\ref{f:M_k_main}) and note that by the Cauchy--Schwarz inequality one has
$$
|A|^{2k} = \Big( \sum_n \sigma_k (n) \Big)^2 \le |kA|\,\sum_n \sigma^2_k (n)
\le
2 k 8^k k^{3k/2} |A|^k k N\,,
$$
hence
$$
|A| \le 16 k^{3/2} N^{1/k}
$$
as required.
\bigskip
\begin{proof}
Let $B \subseteq A$ be a set given by Lemma \ref{l:structure_X}. By the Minkowski inequality we have
\begin{eqnarray*}
M^{1/2k}_k (A)&=&\big(\int_0^1|f(t)|^{2k}dt\big)^{1/2k}\le \big(\int_0^1|f_B(t)|^{2k}dt\big)^{1/2k}+\big(\int_0^1|f_{A\setminus B}(t)|^{2k}dt\big)^{1/2k}\\
&\le& M^{1/2k}_k (B) + M^{1/2k}_k (A\setminus B)\,.
\end{eqnarray*}
Since $|A\setminus B| \le 2 k^{3}, $ it follows that $M^{1/2k}_k (A\setminus B)\le 2k^3$.
Furthermore, by
%(\ref{inequality-Mk})
the first inequality in (\ref{e:ruzsa-S_k}), we get
%(more precisely see bound \ref{inequality-Mk}), we get
$$ M^{1/2k}_k (B)\ge \frac12 (20k^6-2k^3)^{1/2}\ge 2k^3 \ge M^{1/2k}_k (A\setminus B)\,,$$
%$$
% M^{1/2k}_k (B) \ge (2^{-k} k! |B|^k)^{1/2k} \ge (2^{-k-1} k! |A|^k)^{1/2k} \ge k \cdot 2k^3
% \ge k M^{1/2k}_k (A\setminus B) \,,
%$$
so that
$$
M_{k} (A) \le 2^k M_k (B) \,.
$$
%With some abuse of the notation we write $A$ for $A_1$.
Thus, by Lemma \ref{l:M_k=S_k} it is enough to estimate the quantity $S_k (B).$
%and moreover by Lemma \ref{l:S_k=S_k(A,A^c)}
%we want
% it is sufficiently
% to find an upper bound for
%the
% $S_k (B',B\setminus B')$, where $B'\subseteq B$ is a random set given by Lemma \ref{l:S_k=S_k(A,A^c)}.
Put $t=\lfloor k/2\rfloor$ and observe that (the formula below works for any $t\le k$)
\begin{equation}\label{S_k-S*_l}
S_t=S_t(B)=\sum_{l=2}^{t}\binom{t}{l}^2(t-l)!S^*_l(B)|B|^{t-l}+t!|B|^t.
\end{equation}
Indeed, notice that from every solution to the equation
\begin{equation}\label{eq-side-distinct}
x_1+\dots+x_t = y_1 + \dots+y_t
\end{equation}
in $B$ with $x_i\not=x_j$ and $y_i\not=y_j$, after cancelation of all equal variables from both sides, we obtain for some $l<t,$ a
solution to the equation
\begin{equation}\label{eq-distinct-l}
x_{i_1}+\dots+x_{i_l} = y_{i_1} + \dots+y_{i_l}
\end{equation}
with distinct integers $x_{i_j},y_{i_j}\in B$.
Conversely, each solution to the equation (\ref{eq-side-distinct}) must be constructed from a solution to (\ref{eq-distinct-l}) for some $2\le l\le t$, and by adding to both sides of the equation the same elements from $B$. This can be done on $\binom{t}{l}^2(t-l)!$ ways, as for a fixed solution $(x_{i_1},\dots,x_{i_l} , y_{i_1}, \dots,y_{i_l})$ we have to chose positions for $x_{i_j}, y_{i_j}$ on $\binom{t}{l}^2$ ways, and then to chose distinct $t-l$ elements from $B$ and order them on the right hand side of the equation on $(t-l)!$ ways. The term $t!|B|^t$ counts the number of solutions with $\{ x_1,\dots,x_t\}=\{ y_1, \dots,y_t\}.$
Next, applying Lemma \ref{l:S_k=S_k(A,A^c)} we bound $S^*_l(B)$. Let $B'$ be a set given by Lemma \ref{l:S_k=S_k(A,A^c)}. To estimate $S^*_l(B',B\setminus B')$ observe that every solution to (\ref{eq-distinct-l}), say, $(x_{1},\dots,x_{l} , y_{1}, \dots,y_{l}), x_i\in B', y_i\in B\setminus B'$
corresponds to a set $Z = \{ x_1, \dots, x_l, y_1, \dots, y_l\},$
and clearly, each such set with $\sum x_i=\sum y_i,$ gives $(l!)^2$ solutions to (\ref{eq-distinct-l}).
Thus, to bound $S^*_l(B', B\setminus B')$ it is sufficient to bound the number of such sets $Z$.
Denote by $\mathcal{A}_l$ the family of all sets $Z$ that corresponds to a solution to (\ref{eq-distinct-l}).
Put $\mathcal{X} = \mathcal{X} (B)$, $\mathcal{G} = \mathcal{G} (B)$.
% Property (\ref{f:structure_X}) gives us for any $l$
%\begin{equation}\label{f:X_property_proof}
% l \mathcal{X} \cap \mathcal{G} = \emptyset \,.
%\end{equation}
By (\ref{in:X}) we have $|\mathcal{X} \cap [l]| \le \lfloor\frac l2\rfloor$ for every $1\le l\le k.$
%assumption $k\in \mathcal{G}$ and hence taking $l=2$ in (\ref{f:X_property_proof}), we obtain
%$k\notin \mathcal{X}+\mathcal{X}$. In particular, $|\mathcal{X}| \le k/2$.
% Putting $t=[k/2]$ we see that $2t \le k$ and $|\mathcal{X} \cap [2t]| \le t$.
% Below we will consider just the intersection $\mathcal{X}$ with the segment $[2t]$ and estimate
% the quantity $S_l (B',B\setminus B')$. Theorem \ref{t:EKR}.
Suppose that we have two solutions to (\ref{eq-distinct-l})
$$
x_1+\dots+x_a + y_1+\dots+y_{l-a} = x'_1+\dots+x'_b + y'_1+\dots+y'_{l-b}
$$
$$
\tilde{x}_1+\dots+\tilde{x}_a + \tilde{y}_1+\dots+\tilde{y}_{l-a} = \tilde{x}'_1+\dots+\tilde{x}'_b + \tilde{y}'_1+\dots+\tilde{y}'_{l-b} \,,
$$
where $0\le a \le l$, $0\le b \le l$, further, $\{x_1,\dots,x_a \} = \{\tilde{x}_1,\dots,\tilde{x}_a \}$,
$\{x'_1,\dots,x'_b \} = \{\tilde{x}'_1,\dots,\tilde{x}'_b \}$
and $\{ y_1,\dots,y_{l-a}\} \cap \{ \tilde{y}_1,\dots,\tilde{y}_{l-a}\} = \emptyset$,
$\{ y'_1,\dots,y'_{l-b}\} \cap \{ \tilde{y}'_1,\dots,\tilde{y}'_{l-b}\} = \emptyset$.
The above solutions correspond to some sets $Z_1$, $Z_2 \subseteq [B]^{2l}$ and
we have $|Z_1 \cap Z_2| = a+b$.
%It gives us
Subtracting our solutions, we get
\begin{equation}\label{tmp:10.06.2016_1}
y_1+ \dots +y_{l-a} + \tilde{y}'_1+\dots+\tilde{y}'_{l-b}
=
y'_1+\dots+y'_{l-b} + \tilde{y}_1+\dots+\tilde{y}_{l-a} \,.
\end{equation}
Because of $\{ y_1,\dots,y_{l-a}\} \cap \{ \tilde{y}_1,\dots,\tilde{y}_{l-a}\} = \emptyset$,
$\{ y'_1,\dots,y'_{l-b}\} \cap \{ \tilde{y}'_1,\dots,\tilde{y}'_{l-b}\} = \emptyset$
we to see that (\ref{tmp:10.06.2016_1}) is a solution with distinct elements from $B$.
Therefore, $2l-|Z_1 \cap Z_2| \in \mathcal{X}$, so that $|Z_1 \cap Z_2| \in (2l - \mathcal{X})$.
Putting $L= 2l - \mathcal{X} = \{ l_1 < \dots< l_r\} \subseteq [0,2l]$, $r=|\mathcal{X} \cap [2l]| \le l$, we see that the family $\mathcal{A}_l$ is a $(|B|,2l,L)$--system.
Using Theorem \ref{t:EKR}, we obtain
$$
|\mathcal{A}_l| \le \prod_{i=1}^{|L|} \frac{|B|-l_i}{2l-l_i} \le \frac{|B|^r}{r!}
\le
\frac{|B|^{l}}{l!} \,.
$$
Thus,
$$
S^*_l (B',B\setminus B') \le \frac{|B|^{l}}{l!} \cdot (l!)^2 = l! |B|^l \,,
$$
so that by Lemma \ref{l:S_k=S_k(A,A^c)}
$$
S^*_l (B) \le 4^l l! |B|^l \,.
$$
Therefore by (\ref{S_k-S*_l})
$$S_t(B)\le 4^t t!|B|^t\sum_{l=0}^t\binom{t}{l} = 8^t t! |B|^t\,.$$
Applying Lemma \ref{l:S_k,S_l} with $l=t$, we get
\begin{eqnarray*}
S_k (B) &\le& (\sqrt{2} k)^{2(k-t)} |B|^{k-t}S_t(B)\\
&\le&
8^t t! (\sqrt{2} k)^{2(k-t)} |B|^{k}\\
&\le&
2^{k+2t} k^{2k-t} |B|^k\\
& \le&
k4^{k} k^{3k/2} |B|^k \,,
\end{eqnarray*}
hence by
$$M_k(A)\le 2^k M_k(B)\le 2^{k+1} S_k(B) \le 2 k 8^{k} k^{3k/2} |B|^k \le 2 k 8^{k} k^{3k/2} |A|^k\,,$$
which completes the proof.
\hfill$\blacksquare$
\end{proof}
%\bigskip
\begin{thebibliography}{99}
%\bibitem{cilleruelo}
%{\sc J.~Cilleruelo, }{\em New upper bound for finite $B_h$ sequences,} Advances in Mathematics
%159 (2001), 1--17.
\bibitem{DEF}
{\sc M. Deza, P. Erd{\H o}s, and P. Frankl, }
{\em Intersection properties of systems of finite sets, }
Proc. London Math. Soc. (3) 36 (1978), 369--384.
\bibitem{ER}
{\sc P. Erd{\H o}s, and R. Rado, }
{\em Intersection theorems for systems of sets, }
J. London Math. Soc. 35 (1960), 85--90.
\bibitem{EKR}
{\sc P. Erd{\H o}s, C. Ko, and R. Rado, }
{\em Intersection theorems for systems of finite sets, }
Quart. J. Math. Oxford (2) 12 (1961), 313--320.
\bibitem{green}
{\sc B.~Green, }{\em The number of squares and $B_h[g]$ sets,} Acta Arith. 100 (2001), 365--390.
\bibitem{obryant}
{\sc K.~O' Bryant, }{\em A complete annotated bibliography or work related to Sidon sequences,}
Electronic J. of Combinatorics DS 11, v. 575 (2004).
\bibitem{ruzsa-I}
{\sc I.~Z.~Ruzsa,}
{\em Solving a linear equation in a set of integers I,} Acta Arith. 65 (1993), 259--282.
\bibitem{timmons}
{\sc C.~M. ~Timmons, }{\em Upper and lower bounds on $B^+_k$-sets,} Integers 255 (2014), 55--70.
\bibitem{TV}
{\sc T.~Tao and V.~Vu, } {\em Additive Combinatorics, } Cambridge University Press (2006).
\end{thebibliography}
\bigskip
\noindent{T.~Schoen\\
\no{Faculty of Mathematics and Computer Science,\\ Adam Mickiewicz
University,\\ Umul\-towska 87, 61-614 Pozna\'n, Poland\\} {\tt
schoen@amu.edu.pl}
\bigskip
\noindent{I.D.~Shkredov\\
Steklov Mathematical Institute,\\
ul. Gubkina, 8, Moscow, Russia, 119991}
\\
and
\\
IITP RAS, \\
Bolshoy Karetny per. 19, Moscow, Russia, 127994
\\
and
\\
MIPT, \\
Institutskii per. 9, Dolgoprudnii, Russia, 141701\\
{\tt ilya.shkredov@gmail.com}
\end{document}
\section{Proof of Theorem 1}
\begin{lemma}\label{l:k/t-set}
Let $A$ be a $B_k$-set and let $i$ be a positive integer. Then there exists a set $A'\sbeq A$ such that $|A'|\ge |A|-k$ and $A'$ is a $B_{\lceil k/i\rceil}$-set or $B_{\lfloor k/i\rfloor}$-set.
\end{lemma}
\begin{proof} Let $k=i\lfloor k/i\rfloor+r$, hence $0\le r<i$. Then we can write
$k=(i-r)\lfloor k/i\rfloor+r\lceil k/i\rceil.$ Clearly, one cannot find $i-r$ disjoint solutions to the equation
$$x_1+\dots+x_{\lfloor k/i\rfloor}=y_1+\dots+y_{\lfloor k/i\rfloor} \,.$$
which are disjoint with
$r$ disjoint solutions to the equation
$$x_1+\dots+x_{\lceil k/i\rceil}=y_1+\dots+y_{\lceil k/i\rceil},$$
Therefore removing at most $k- \lfloor k/i \rfloor$ elements from $A$ we obtain a
$B_{\lceil k/i\rceil}$-set or $B_{\lfloor k/i\rfloor}$-set.
$\hfill\blacksquare$
\end{proof}
\bigskip
Ruzsa proved \cite{ruzsa-I}
\begin{equation}\label{ruzsa-k2}
M_k\le (1+o(1))k^2|A|M_{k-1}
\end{equation}
and
\begin{equation}\label{ruzsa-k2-final}
M_k\le (1+o(1))k^{2k-2}|A|^k
\end{equation}
\begin{lemma}\label{l:k/t-set}
Let $A$ be a $B_k$-set and let $m<k$ be a positive integer. Then we have
$$M_k\le (1+o(1))k^{2{(k-m)}}|A|^{k-m}M_{m} \,.$$
\end{lemma}
\begin{proof} If $m=k-1$ then our inequality becomes (\ref{ruzsa-k2}). Now assume that $m\le k-2.$
Then by H\"older inequality we have
$$M_{k-1}=\int_0^1 |f(t)|^{2k-2}dt\le \B(\int_0^1 |f(t)|^{2k}dt\B)^{\frac{k-m-1}{k-m}}\B(\int_0^1 |f(t)|^{2m}dt\B)^{\frac1{k-m}}=s_k^{\frac{k-m-1}{k-m}}M_m^{\frac1{k-m}},$$
hence by (\ref{ruzsa-k2})
$$M_k\le (1+o(1))k^2|A|M_k^{\frac{k-m-1}{k-m}}s_m^{\frac1{k-m}}$$
and the assertion follows. $\hfill\blacksquare$
\end{proof}
\bigskip
Put $k_0=k, k_{n+1}=0$ and for $i=2,\dots,n$ let $k_{i-1}=\lceil k/i\rceil$ if $A$ is a $B_{\lceil k/i\rceil}$-set, otherwise $A$ is a $B_{\lfloor k/i\rfloor}$-set and we set $k_{i-1}=\lfloor k/i\rfloor.$ By a simple induction one can prove that
$$M_k\le (1+o(1))\prod_{i=0}^{n-1}k_i^{2(k_i-k_{i+1})}|A|^{k-k_n}M_{k_n}\le (1+o(1))\prod_{i=0}^{n}k_i^{2(k_i-k_{i+1})} |A|^k \,.$$
Next we estimate the product on the left hand side of the inequality above
$$\prod_{i=0}^{n}k_i^{2(k_i-k_{i+1})}\le \prod_{i=0}^{n}\Big(\frac{k+1}{i}\Big) ^{2(k/i-k/({i+1})+i)}\le k^{k+n^2} (1+1/k)^{k+n^2}\exp\Big(-2k\sum_{i=1}^n \frac{\ln i}{i(i+1)}\B).
$$
Further, we have
$$\sum_{i=1}^n \frac{\ln i}{i(i+1)}>0.596498+\sum_{i=21}^n \frac{\ln i}{i(i+1)}$$
and the last sum can be bounded from above in the following way
\begin{eqnarray*}\sum_{i=21}^n \frac{\ln i}{i(i+1)}&=&
\sum_{i=21}^n \frac{\ln i}{i}-\sum_{i=21}^n \frac{\ln i}{i+1}\ge \sum_{i=21}^n \frac{\ln i}{i}-\sum_{i=21}^n \frac{\ln (i+1)}{i+1}\\
&=&\frac{\ln 21}{21}-\frac{\ln (n+1)}{n+1}>0.1449-\frac{\ln (n+1)}{n},
\end{eqnarray*}
so that
$$\exp\Big(-2k\sum_{i=1}^n \frac{\ln i}{i(i+1)}\B)\le ...$$
Thus, by Cauchy-Schwarz inequality we have
$$\frac{|A|^{2k}}{kN}\le M_k\le (1+o(1))k^{2k+n^2} (1+1/k)^{k+n^2}...,$$
hence
$$|A|\le (1+o(1))(1+1/k)^{2+n^2/k}e^{\frac{\ln(n+1)}{n}}\frac{k^{2+n^2/k}}{4.4}N^{1/k} \,.$$
Taking $n=\lfloor (\frac13k)^{1/3}\rfloor$ we obtain
$$|A|\le (1+o(1))\B(1+\frac{\ln k}{ k^{1/3}}\B)\frac{k^{2}}{4.4}N^{1/k},$$
which completes the proof. $\hfill\blacksquare$
\section{Remarks}
Let $A'$ be a subset of $A$ without disjoint solutions, $|A'|\ge |A|-k^2$.
Let $$B=\{l\le k: A' \text{ is a } B^*_l-set\} \,.$$
If $B'=[k]\setminus B$ then
clearly, $\sum(B')\cap B=\emptyset,$ in particular $|B'|\le k/2.$ Write $B=\{l_1=k,\dots,l_m\}_>$, find
$$\max_B \prod_{i=1}^{m-1}l_i^{2(l_i-l_{i+1})} \,.$$ We assume that $k$ is a fixed integer and $|A|\rightarrow \infty.$
Suppose that $A$ is $B^*_j-$set for all $j=k,k-1,\dots,l.$ Then
$$M_k\sim S_k\le (k-l)!\binom{k}{l}^2 |A|^{k-l}S_l=\frac{1}{(k-l)!}\Big(\frac{k!}{l!}\Big)^2S_l|A|^{k-l} \,.$$
If $l_1>l_2+2$ then $[l_1-1,l_2+1]\cap B=\emptyset$, so $[k-l_2-1,k-l_1+1]\sbeq B.$
Therefore,
\begin{equation}\label{e:B_k-main}
S_{k-l_2-1}\le \frac{1}{(l_1-l_2-2)!}\Big(\frac{(k-l_2-1)!}{(k-l_1+1)!}\Big)^2S_{k-l_1+1}|A|^{l_1-l_2-2}.
\end{equation}
\section{Improvement}
Let $\snd (k)$ denote the smallest non divisor of $k$.
\begin{theorem}(Lev) Let $A\sbeq [N]$ and $|A|> N/d$. Then $dA$ contains an arithmetic progression
$nq,(n+1)q,\dots,(n+l)q,$ where $q=\rm{gcd}(A-\min A)$ and $l\ge N.$
\end{theorem}
\begin{theorem} Let $A\sbeq [N]$ be a weak $B_k$-set. Then
$$|A|\le (1+o(1))(4e)^2 \frac{k^2}{{ \snd(k)}}N^{1/k} \,.$$
\end{theorem}
\begin{proof}
Let $A'(=A)$ be a set given in "Remarks" of size $|A|-k^{O(1)}$ such that
$2kB'\cap B=\emptyset.$ Put $d=\snd (k)$. We may assume that $k$ and $d$ are large enough.
Let $C=B'\cap \{1,\dots, [k/d]\}\cup \{0\}.$ We show that $|C|\le [k/d]/d+..$.
Suppose to the contradiction, then by Lev's theorem $dC$ contains an arithmetic progression
$P=\{h,2h,\dots,Lh\}$, where $h=\rm{gcd} (C)\le d$ and $L\ge k/d$. Then, clearly
$$k\in P\cup (a+P)\cup\dots \cup (ka+P)\sbeq 2kB',$$
which is a contradiction, so that $|C|\le [k/d]/d+..$.
Let $u_1>\dots>u_n$ be all positive elements of $C$
and let $I=\{i: u_i-u_{i+1}\ge 3 \}.$
Therefore by (\ref{e:B_k-main})
$$S_{u_i-1}\le \frac{1}{(u_i-u_{i+1}-2)!}\Big(\frac{(u_i-1)!}{(u_{i+1}+1)!}\Big)^2S_{u_{i+1}+1}|A|^{u_i-u_{i+1}-2}
\le \frac{k^{2(u_i-u_{i+1}-2)}}{(u_i-u_{i+1}-2)!}S_{u_{i+1}+1}|A|^{u_i-u_{i+1}-2}$$
for each $i\in I.$ We have
$$S_k\le \frac{k^{2k}}{\prod_{i\in I}(u_i-u_{i+1}-2)!}|A|^{k} \,.$$
Since $\sum_{i\in I} (u_i-u_{i+1}-2)\ge k/d-4k/d^2,$ and using the fact that if $u>v+1$ then $u!v!> (u-1)!(v+1)!$ it follows that $$\prod_{i\in I}(u_i-u_{i+1}-2)!\ge (d-3)!^{k/d-4k/d^2}\ge (d/2e)^{k-4k/d},$$ hence
$$S_k\le \frac{(2ed^{2/d}k)^{2k}}{d^{k}}|A|^k \,.$$ By (\ref{e:ruzsa-S_k}) we have
$$ \frac{|A|^{2k}}{kN}\le M_k\le \frac{(4ek)^{2k}}{d^{k}}|A|^k$$
and the assertion follows.\hfill$\blacksquare$
\end{proof}
\begin{corollary} For infinitely many $k$ we have
$$R_k(N)\ll \frac{k^2}{{\log k}}N^{{1/k}}+O(k^{O(1)}) \,.$$
\end{corollary}
%\end{proof}
%\section{New bound for the size of $B^*_k$--sets}