[Paper Review] Some examples of sets of large exponential sums

Ilya D. Shkredov|ArXiv.org|Oct 18, 2006

Limits and Structures in Graph Theory17 references4 citations

TL;DR

This paper constructs explicit examples of subsets $ A \subseteq \mathbb{Z}/N\mathbb{Z} $ with large exponential sums, demonstrating the sharpness of M.-C. Chang's theorem on the structure of sets $ \mathcal{R}_\alpha $—the set of Fourier coefficients exceeding threshold $ \alpha N $. It improves bounds on the size of generating sets $ \Lambda^* $ for $ \mathcal{R}_\alpha $, showing they are optimal up to logarithmic factors, and establishes a quantitative inverse theorem for exponential sum sets via combinatorial and Fourier-analytic techniques.

ABSTRACT

Let $A$ be a subset of $\mathbb{Z} / N\mathbb{Z}$ and let $\mathcal{R}$ be the set of large Fourier coefficients of $A$. Properties of $\mathcal{R}$ have been studied in works of M.-- C. Chang, B. Green and the author. In the paper we obtain some new results on sets of large exponential sums.

Motivation & Objective

• To demonstrate the sharpness of M.-C. Chang's theorem on the structure of sets $ \mathcal{R}_\alpha $, the set of large Fourier coefficients of a subset $ A \subseteq \mathbb{Z}/N\mathbb{Z} $.
• To construct explicit examples of subsets $ A $ such that $ \mathcal{R}_\alpha $ cannot be contained in the span of small sets $ \Lambda $, thus showing the optimality of existing bounds.
• To establish a quantitative inverse theorem for sets of large exponential sums, refining the dependence on $ \delta = |A|/N $ and $ \alpha $.
• To improve the upper bounds on the size of generating sets $ \Lambda^* $ and $ \tilde{\Lambda} $ that represent all elements of $ \mathcal{R}_\alpha $ via signed sums.

Proposed method

• Constructs a set $ A \subseteq \mathbb{Z}_2^n $ with $ |A| = \delta N $, using a union of structured subsets $ A_i $ with controlled intersections to control Fourier coefficients.
• Defines $ \mathcal{R}_\alpha $ as the set of $ r \in \mathbb{Z}_N $ with $ |\widehat{A}(r)| \geq \alpha N $, and analyzes its additive structure via the number of solutions to $ r_1 + \cdots + r_k = r_1' + \cdots + r_k' $.
• Uses the $ k $-th moment $ T_k(B) $ of a subset $ B \subseteq \mathcal{R}_\alpha \setminus \{0\} $ to derive lower bounds on the additive energy of $ \mathcal{R}_\alpha $, linking it to the size of $ B $ and parameters $ \delta, \alpha $.
• Applies combinatorial estimates and bounds on intersection sizes $ |A_i \cap A_j| < k/r $ to control the number of solutions to additive equations in $ \mathcal{R}_\alpha $, leading to upper bounds on $ T_2(\mathcal{R}_\alpha) $.
• Employs Fourier-analytic tools, including the identity $ \sum_r |\widehat{f}(r)|^4 = \frac{1}{N} \sum_r |\widehat{f \ast f}(r)|^2 $, to relate $ T_2 $ to the $ L^4 $-norm of the Fourier transform.
• Combines these bounds with the inverse theorem framework to show that the number of generators $ \Lambda^* $ required to span $ \mathcal{R}_\alpha $ is optimal up to logarithmic factors.

Experimental results

Research questions

• RQ1 Can M.-C. Chang's theorem on the structure of $ \mathcal{R}_\alpha $ be shown to be sharp in terms of the size of the generating set $ \Lambda $?
• RQ2 What is the minimal size of a set $ \Lambda $ such that every $ r \in \mathcal{R}_\alpha $ can be written as a signed sum of elements from $ \Lambda $?
• RQ3 Can explicit constructions of $ A \subseteq \mathbb{Z}/N\mathbb{Z} $ be given such that $ \mathcal{R}_\alpha $ cannot be covered by the span of small $ \Lambda $?
• RQ4 How does the additive energy of $ \mathcal{R}_\alpha $, measured by $ T_k(B) $, depend on $ \delta $, $ \alpha $, and $ |B| $?
• RQ5 Is the dependence of the generating set size on $ \delta/\alpha $ and $ \log(1/\delta) $ in Chang's theorem optimal?

Key findings

• The paper constructs a set $ A \subseteq \mathbb{Z}_2^n $ such that $ \mathcal{R}_\alpha $ cannot be contained in the span of any set $ \Lambda $ of size less than $ 2^{-12}(\delta/\alpha)^2 \log(1/\delta) $, proving the sharpness of Chang's bound under certain conditions.
• It establishes that the number of generators $ \Lambda^* $ needed to represent all $ r \in \mathcal{R}_\alpha $ via signed sums satisfies $ |\Lambda^*| \leq \min\left(2^{30}(\delta/\alpha)^2 \log(1/\delta), 2^{4(\log\log(1/\delta))^2 + 2}\right) $, improving previous bounds.
• The paper shows that the number of solutions to $ r_1 + r_2 = r_1' + r_2' $ in $ \mathcal{R}_\alpha $, denoted $ T_2(\mathcal{R}_\alpha) $, is bounded above by $ \frac{16\delta}{\alpha^4} $, under the condition $ \alpha \geq 32\delta^2 $.
• It proves that the number of vectors $ \lambda_i^* \in \Lambda^* $ required to represent any $ r \in \mathcal{R}_\alpha $ is at most $ 8\log(1/\delta) $, and that this bound is essentially optimal.
• The size of the alternative generating set $ \tilde{\Lambda} $ is bounded by $ 2^{20}(\delta/\alpha)^2 \log^{5/3}(1/\delta) \log\log(1/\delta) $, showing a trade-off between logarithmic and polynomial dependencies.
• The paper confirms that the dependence on $ \log(1/\delta) $ in the size of $ \Lambda^* $ is optimal up to constant factors, as any smaller set would fail to generate $ \mathcal{R}_\alpha $ when $ |\mathcal{R}_\alpha| \approx 1/\delta $.