This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Sharp Bounds for Sets with Distinct Subset Products

Rushil Raghavan
Abstract

Let A[N]A\subseteq[N] be such that for any pair of distinct subsets B,CAB,C\subseteq A, the products bBb\prod_{b\in B}b and cCc\prod_{c\in C}c are distinct. We prove that |A|π(N)+π(N1/2)+o(π(N1/2))|A|\leq\pi(N)+\pi(N^{1/2})+o(\pi(N^{1/2})), where π\pi is the prime counting function, answering a question of Erdős.

1 Introduction

For NN\in\mathbb{N}, let f(N)f(N) denote the size of the largest subset AA of [N][N] such that for any pair of distinct subsets B,CAB,C\subseteq A, bBbcCc\prod_{b\in B}b\neq\prod_{c\in C}c. We say that a set AA satisfying these conditions has distinct subset products. Erdős [3] initiated the study of the quantity f(N)f(N), proving111See Definition 1.9 for the asymptotic notation used in this paper.

f(N)π(N)+O(π(N1/2)),f(N)\leq\pi(N)+O(\pi(N^{1/2})),

where π(x)=|{nx:n is prime}|\pi(x)=|\{n\leq x:n\text{ is prime}\}| is the prime counting function.

He also produced the following example, establishing that the above bound is sharp up to the implicit constant:

Example 1.1.

Let A={pN:p is prime}{p2:p is prime and pN1/2}.A=\{p\leq N:p\text{ is prime}\}\cup\{p^{2}:p\text{ is prime and }p\leq N^{1/2}\}. Then AA has distinct subset products, so f(N)|A|=π(N)+π(N1/2)f(N)\geq|A|=\pi(N)+\pi(N^{1/2}).

He then asked [3], [4] whether this estimate is optimal up to lower order terms. This problem is also listed at https://www.erdosproblems.com/795.

Question 1.2 (Erdős #795).

Is f(N)=π(N)+π(N1/2)+o(π(N1/2))?f(N)=\pi(N)+\pi(N^{1/2})+o(\pi(N^{1/2}))?

We answer this question affirmatively:

Theorem 1.3.

f(N)=π(N)+π(N1/2)+O(N5/12+o(1))f(N)=\pi(N)+\pi(N^{1/2})+O(N^{5/12+o(1)}).

In view of the above estimate, it is natural to ask for more precise information about the lower-order term. Erdős also considered this in [3], and speculated that a refinement of Example 1.1 may be optimal. To be more specific, for kk\in\mathbb{N}, let g(k)g(k) be the smallest element of \mathbb{N} for which there is an Ek[g(k)]E_{k}\subseteq[g(k)] of size kk such that for any distinct F,FEkF,F^{\prime}\subseteq E_{k}, nFnnFn\sum_{n\in F}n\neq\sum_{n\in F^{\prime}}n. In other words, let g(k)g(k) be the smallest possible maximal element in a set of size kk with distinct subset sums. Then, by setting

A=k=1nEk{pn:p(N1/g(k+1),N1/g(k)] is prime},A=\bigcup_{k=1}^{\infty}\bigcup_{n\in E_{k}}\{p^{n}:p\in(N^{1/g(k+1)},N^{1/g(k)}]\text{ is prime}\},

we see that the subset products of AA are distinct, and thus

f(N)k=1π(N1/g(k)).f(N)\geq\sum_{k=1}^{\infty}\pi(N^{1/g(k)}).

Since g(1)=1g(1)=1, g(2)=2g(2)=2, g(3)=4g(3)=4, and g(4)=7g(4)=7 (with E4={3,5,6,7}E_{4}=\{3,5,6,7\}), Erdős established f(N)π(N)+π(N1/2)+π(N1/4)+π(N1/7)f(N)\geq\pi(N)+\pi(N^{1/2})+\pi(N^{1/4})+\pi(N^{1/7}), and speculated that the above infinite sum may be best possible. However, we produce an example that improves upon this.

Theorem 1.4.

f(N)π(N)+π(N1/2)+13π(N1/3)O(1)f(N)\geq\pi(N)+\pi(N^{1/2})+\frac{1}{3}\pi(N^{1/3})-O(1).

It is also natural to consider the additive variant of Question 1.2, or in other words, to determine the asymptotic behavior of the function g(k)g(k). This is another problem of Erdős, which can be seen at https://www.erdosproblems.com/1. Although our methods do not address this problem, the interested reader may consult [2], [5] for the best-known lower bounds on g(k)g(k) and some history.

Since squares feature so prominently in Example 1.1, one may ask about what kinds of estimates can be obtained in the absence of squares. Our method answers this question as well, with a somewhat smaller second-order term.

Theorem 1.5.

Let h(N)h(N) be the maximal size of a subset of [N][N] consisting of squarefree integers with distinct subset products. Then h(N)π(N)+12π(N1/2)+O(N5/12+o(1))h(N)\leq\pi(N)+\frac{1}{2}\pi(N^{1/2})+O(N^{5/12+o(1)}).

This estimate is also sharp up to the error term.

Theorem 1.6.

h(N)π(N)+12π(N1/2)+o(π(N1/2))h(N)\geq\pi(N)+\frac{1}{2}\pi(N^{1/2})+o(\pi(N^{1/2})).

1.1 Notation and strategy of proof

Definition 1.7 (Subset Product Set).

Given a finite set SS\subseteq\mathbb{N}, we define the subset product set Π(S)={tTt:TS}\Pi(S)=\{\prod_{t\in T}t:T\subseteq S\}.

Throughout the proof, we will use the fact that an element of [N][N] can be divisible by at most one prime in (N1/2,N](N^{1/2},N], and at most two primes (with multiplicity) in (N1/3,N](N^{1/3},N]. We thus define

Definition 1.8 (Small, Medium, and Large Primes, Valuations).

Let 𝒫small\mathcal{P}_{\text{small}} denote the set of primes in [N1/3],[N^{1/3}], 𝒫med\mathcal{P}_{\text{med}} the set of primes in (N1/3,N1/2](N^{1/3},N^{1/2}], and 𝒫large\mathcal{P}_{\text{large}} the set of primes in (N1/2,N](N^{1/2},N]. We will also use the terms “small primes”, “medium primes”, and “large primes” to refer to elements of 𝒫small\mathcal{P}_{\text{small}}, 𝒫med\mathcal{P}_{\text{med}}, and 𝒫large\mathcal{P}_{\text{large}}, respectively.

For a fixed prime p[N]p\in[N], and nn\in\mathbb{N}, let Vp(n)V_{p}(n) denote the valuation of nn at pp, i.e., the largest nonnegative integer rr such that prp^{r} divides nn. Then define functions

𝐕small:𝒫small,𝐕med:𝒫med,𝐕large:𝒫largeby\mathbf{V}_{\text{small}}:\mathbb{N}\to\mathbb{Z}^{\mathcal{P}_{\text{small}}},\quad\mathbf{V}_{\text{med}}:\mathbb{N}\to\mathbb{Z}^{\mathcal{P}_{\text{med}}},\quad\mathbf{V}_{\text{large}}:\mathbb{N}\to\mathbb{Z}^{\mathcal{P}_{\text{large}}}\quad\text{by}
𝐕small(n)=(Vp(n))p𝒫small,𝐕med(n)=(Vp(n))p𝒫med,𝐕large(n)=(Vp(n))p𝒫large.\mathbf{V}_{\text{small}}(n)=(V_{p}(n))_{p\in\mathcal{P}_{\text{small}}},\quad\mathbf{V}_{\text{med}}(n)=(V_{p}(n))_{p\in\mathcal{P}_{\text{med}}},\quad\mathbf{V}_{\text{large}}(n)=(V_{p}(n))_{p\in\mathcal{P}_{\text{large}}}.

We will also define 𝐕large×𝐕med:𝒫large𝒫med\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}}:\mathbb{N}\to\mathbb{Z}^{\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}}} as (𝐕large×𝐕med)(n)=(Vp(n))p𝒫large𝒫med(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})(n)=(V_{p}(n))_{p\in\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}}}.

We will also use some standard asymptotic notation:

Definition 1.9 (Big-O and Little-o Notation).

Given functions FF, GG, and HH from \mathbb{N} to [0,)[0,\infty), we say

  • F(n)=O(G(n))F(n)=O(G(n)) if there is a constant C>0C>0 such that F(n)CG(n)F(n)\leq CG(n) for all nn\in\mathbb{N},

  • F(n)=o(G(n))F(n)=o(G(n)) if for all c>0c>0, there is an N0N_{0}\in\mathbb{N} such that for all nN0n\geq N_{0}, F(n)cG(N)F(n)\leq cG(N),

  • F(n)=G(n)+O(H(n))F(n)=G(n)+O(H(n)) if |F(n)G(n)|=O(H(n))|F(n)-G(n)|=O(H(n)).

To prove Theorem 1.3, we will use a graph-theoretic approach. Our graphs will be simple, i.e., containing no loops or multiple edges. We will need to pay attention to certain subgraphs in our analysis:

Definition 1.10 (Paths, Cycles, Circuits).

Let G=(V,E)G=(V,E) be a graph.

  • A path of length kk is a collection of edges of the form

    {{v1,v2},{v2,v3},,{vk,vk+1}}E\{\{v_{1},v_{2}\},\{v_{2},v_{3}\},\dots,\{v_{k},v_{k+1}\}\}\subseteq E, where v1,,vk+1v_{1},\dots,v_{k+1} are distinct vertices in VV.

  • A cycle of length kk is a collection of edges of the form

    {{v1,v2},{v2,v3},,{vk1,vk},{vk,v1}}E\{\{v_{1},v_{2}\},\{v_{2},v_{3}\},\dots,\{v_{k-1},v_{k}\},\{v_{k},v_{1}\}\}\subseteq E, where v1,,vkv_{1},\dots,v_{k} are distinct vertices in VV.

  • A circuit of length kk is a collection of edges of the form

    {{v1,v2},{v2,v3},,{vk1,vk},{vk,v1}}E\{\{v_{1},v_{2}\},\{v_{2},v_{3}\},\dots,\{v_{k-1},v_{k}\},\{v_{k},v_{1}\}\}\subseteq E, where v1,,vkv_{1},\dots,v_{k} are (not necessarily distinct) vertices in VV.

In [3], to prove f(N)π(N)+O(π(N1/2))f(N)\leq\pi(N)+O(\pi(N^{1/2})), Erdős counted the possible number of prime factorizations in elements of π(A)\pi(A), where A[N]A\subseteq[N] has distinct subset products. His proof is a counting argument based on the following three observations:

  • An element of [N][N] is divisible by at most one large prime, and at most two large or medium primes (with multiplicity).

  • The range of 𝐕small\mathbf{V}_{\text{small}} on Π([N])\Pi([N]) is small (see Proposition 2.1).

  • There are at most π(N)\pi(N) elements aa of AA such that there is a prime pp for which pp divides aa, but no other element of AA is a multiple of pp.

Optimizing his argument, one can obtain the bound f(N)π(N)+22π(N1/2)f(N)\leq\pi(N)+22\pi(N^{1/2}). We will utilize each of these observations in our approach, but we take into account some more refined information as well. For example, if p,qp,q are large primes and r,sr,s are medium primes, the argument from [3] does not use the fact that AA cannot contain the elements prpr, qrqr, psps, and qsqs (although it does show that AA cannot contain this configuration for many primes p,q,r,sp,q,r,s).

The strategy of our proof is as follows. Given a set AA with distinct subset products, we can produce a graph whose vertices correspond to large and medium primes, and whose edges correspond to (most) elements of AA. We call this graph the prime factorization graph of AA. The condition that AA has distinct subset products restricts the number of short cycles that can appear in this graph. We can exploit this fact to prove an upper bound on the number of edges in this graph, and thus an upper bound on |A||A|. In Section 2, we show how to construct this prime factorization graph. In Section 3, we use the condition that AA has distinct subset products to remove circuits and cycles from the prime factorization graph without removing too many edges. In Section 4, we estimate the number of edges in a graph without any short cycles, and finally establish Theorems 1.3 and 1.5.

In Section 5, we produce some new examples of nearly maximal sets with distinct subset products and prove Theorems 1.4 and 1.6.

Acknowledgements

The author would like to thank Terence Tao for many helpful conversations and for introducing the author to this problem.

2 Constructing the Prime Factorization Graph

Given a set AA with distinct subset products, we will construct a graph that encodes the large and medium prime factors of (most) elements of AA. For this to be effective, we first need to control how much information is lost by ignoring small primes.

Proposition 2.1.

Let Π([N])\Pi([N]) denote the subset product set of [N][N]. Then

|𝐕small(Π([N]))|exp(O(N1/3+o(1))).|\mathbf{V}_{\text{small}}(\Pi([N]))|\leq\exp(O(N^{1/3+o(1)})).
Proof.

For a fixed p𝒫smallp\in\mathcal{P}_{\text{small}} and n[N]n\in[N], Vp(n)[0,log2(N)]V_{p}(n)\in[0,\log_{2}(N)]. Since an element of Π([N])\Pi([N]) is a product of at most NN elements of [N][N], for a fixed nΠ([N])n\in\Pi([N]), Vp(n)[0,Nlog2(N)]V_{p}(n)\in[0,N\log_{2}(N)]. Thus |𝐕small(Π([N]))||\mathbf{V}_{\text{small}}(\Pi([N]))|\leq

p𝒫small(Nlog2(N)+1)=exp(|𝒫small|log(Nlog2(N)+1))=exp(O(N1/3+o(1))).\prod_{p\in\mathcal{P}_{\text{small}}}(N\log_{2}(N)+1)=\exp(|\mathcal{P}_{\text{small}}|\log(N\log_{2}(N)+1))=\exp(O(N^{1/3+o(1)})).\qed

In view of this proposition, if AA has distinct subset products, there cannot be many elements of Π(A)\Pi(A) with the same valuations at medium and large primes. In fact, the same is true for AA.

Proposition 2.2.

Given 𝐫𝒫large𝒫med\mathbf{r}\in\mathbb{Z}^{\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}}}, let A𝐫={aA:(𝐕large×𝐕med)(a)=𝐫}A_{\mathbf{r}}=\{a\in A:(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})(a)=\mathbf{r}\}. Let R={𝐫𝒫large𝒫med:|A𝐫|2}R=\{\mathbf{r}\in\mathbb{Z}^{\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}}}:|A_{\mathbf{r}}|\geq 2\}. Then

𝐫R|A𝐫|=O(N1/3+o(1)).\sum_{\mathbf{r}\in R}|A_{\mathbf{r}}|=O(N^{1/3+o(1)}).
Proof.

For each 𝐫R\mathbf{r}\in R, let B𝐫A𝐫B_{\mathbf{r}}\subseteq A_{\mathbf{r}} and C𝐫A𝐫C_{\mathbf{r}}\subseteq A_{\mathbf{r}} each have size |A𝐫|/2\lfloor|A_{\mathbf{r}}|/2\rfloor. Then

𝐫RbB𝐫band𝐫RcC𝐫c\prod_{\mathbf{r}\in R}\prod_{b\in B_{\mathbf{r}}}b\quad\text{and}\quad\prod_{\mathbf{r}\in R}\prod_{c\in C_{\mathbf{r}}}c

each have the same valuations at all large and medium primes. There are thus at most exp(O(N1/3+o(1)))\exp(O(N^{1/3+o(1)})) elements of Π(A)\Pi(A) of the form 𝐫RbB𝐫b\prod_{\mathbf{r}\in R}\prod_{b\in B_{\mathbf{r}}}b. There are at least

𝐫R(|A𝐫||A𝐫|/2)𝐫R(1.1)|A𝐫|\prod_{\mathbf{r}\in R}\binom{|A_{\mathbf{r}}|}{\lfloor|A_{\mathbf{r}}|/2\rfloor}\geq\prod_{\mathbf{r}\in R}(1.1)^{|A_{\mathbf{r}}|}

elements of this form, where we used the inequality (nn/2)(1.1)n\binom{n}{\lfloor n/2\rfloor}\geq(1.1)^{n} for all n2n\geq 2. Thus

(1.1)𝐫R|A𝐫|exp(O(N1/3+o(1))),(1.1)^{\sum_{\mathbf{r}\in R}|A_{\mathbf{r}}|}\leq\exp(O(N^{1/3+o(1)})),

and taking logarithms gives the desired inequality. ∎

Corollary 2.3.

Given A[N]A\subseteq[N] with distinct subset products, one can remove O(N1/3+o(1))O(N^{1/3+o(1)}) elements of AA to get a set AAA^{\prime}\subseteq A such that 𝐕large×𝐕med\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}} is injective when restricted to AA^{\prime}. By removing at most one more element, we may assume (𝐕large×𝐕med)(a)𝟎(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})(a)\neq\mathbf{0} for all aA.a\in A^{\prime}.

Because of this corollary, we will assume here and in the next section that 𝐕large×𝐕med\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}} is injective and nonzero on AA. Observe also that for aAa\in A, aa is divisible by at most one large prime, and at most two large or medium primes (with multiplicity). Thus the tuple (𝐕large×𝐕med)(a)(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})(a) either consists of:

  • a 1 at one index, with 0 at all other indices,

  • a 1 at each of two indices, at most one of them from 𝒫large\mathcal{P}_{\text{large}}, and a 0 at all other indices, or

  • a 2 at an index from 𝒫med\mathcal{P}_{\text{med}} and a 0 at all other indices.

Under these conditions, we may define the prime factorization graph of a set AA.

Definition 2.4 (Prime Factorization Graph).

The prime factorization graph G(A)G(A) associated to AA has vertex set 𝒫large𝒫med{1}\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}}\cup\{1\}. We connect 11 to a vertex p𝒫large𝒫medp\in\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}} by an edge if there is an aAa\in A such that Vp(a)=1V_{p}(a)=1 and Vq(a)=0V_{q}(a)=0 for all other primes q𝒫large𝒫medq\in\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}}. We connect p,q𝒫large𝒫medp,q\in\mathcal{P}_{\text{large}}\cup\mathcal{P}_{\text{med}} by an edge if there is an element of AA divisible by both pp and qq.

By the aforementioned discussion, there is a bijection between the edges of G(A)G(A) and the elements of AA which are not divisible by the square of any medium prime. We will still need to consider those elements, so we define the following.

Definition 2.5 (𝒫\mathcal{P}_{\square}, 𝒫□̸\mathcal{P}_{\not\square}).

We define 𝒫={p𝒫med:p2a for some aA}\mathcal{P}_{\square}=\{p\in\mathcal{P}_{\text{med}}:p^{2}\mid a\text{ for some }a\in A\} and 𝒫□̸=𝒫med𝒫\mathcal{P}_{\not\square}=\mathcal{P}_{\text{med}}\setminus\mathcal{P}_{\square}.

Then |A||A| is the sum of the number of edges in G(A)G(A) and |𝒫||\mathcal{P}_{\square}|.


Example 2.6.

Suppose N=121N=121, so 𝒫small={2,3}\mathcal{P}_{\text{small}}=\{2,3\}, 𝒫med={5,7,11}\mathcal{P}_{\text{med}}=\{5,7,11\}, and 𝒫large\mathcal{P}_{\text{large}} is the remaining set of primes in [N][N]. On the set A={50,105,77,55,65,26,51}={252,357,711,511,513,213,317}A=\{50,105,77,55,65,26,51\}=\{2*5^{2},3*5*7,7*11,5*11,5*13,2*13,3*17\}, 𝐕large×𝐕med\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}} is injective and nonzero. Its prime factorization graph is (omitting the isolated vertices corresponding to the primes in (17,121](17,121]):

57111311710510577775555656526265151

Each element of AA corresponds to an edge in the graph except for 5050, since 5050 is a multiple of 525^{2} and 5𝒫med5\in\mathcal{P}_{\text{med}}.

Remark. Suppose aAa\in A is divisible by p2p^{2} for some p𝒫medp\in\mathcal{P}_{\text{med}}. In our definition of G(A)G(A), we do not include an edge corresponding to aa. We could alternatively define the prime factorization graph so that aa would correspond to a loop from pp to itself. One can still make sense of the arguments in the following sections that way, but we found it clearer to express our graph-theoretic arguments using only simple graphs.

Having defined the necessary sets of primes, we will state our main estimate.

Theorem 2.7.

Let A[N]A\subseteq[N] have distinct subset products and let 𝒫={p𝒫med:p2a for some aA}\mathcal{P}_{\square}=\\ \{p\in\mathcal{P}_{\text{med}}:p^{2}\mid a\text{ for some }a\in A\}. Then

|A|π(N)+12π(N1/2)+12|𝒫|+O(N5/12+o(1)).|A|\leq\pi(N)+\frac{1}{2}\pi(N^{1/2})+\frac{1}{2}|\mathcal{P}_{\square}|+O(N^{5/12+o(1)}).

Since 𝒫=\mathcal{P}_{\square}=\emptyset if the elements of AA are all squarefree and |𝒫|π(N1/2)|\mathcal{P}_{\square}|\leq\pi(N^{1/2}) for any set AA, Theorems 1.3 and 1.5 follow immediately from Theorem 2.7.

3 Cycle Removal

Throughout this section, A[N]A\subseteq[N] will have distinct subset products and be such that 𝐕large×𝐕med\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}} is injective and nonzero on AA. Unless otherwise specified, the graph GG will denote the prime factorization graph of AA.

Lemma 3.1.

Let C1,,CnC_{1},\dots,C_{n} be the sets of edges in a maximal collection of edge-disjoint even circuits in GG, each with length at most 2N1/122N^{1/12}. Then n=O(N1/3+o(1))n=O(N^{1/3+o(1)}), so k=1n|Ck|=O(N5/12+o(1))\sum_{k=1}^{n}|C_{k}|=O(N^{5/12+o(1)}).

Proof.

For each k[n]k\in[n], let {v1k,,v|Ck|k}\{v_{1}^{k},\dots,v_{|C_{k}|}^{k}\} be the set of vertices in the cycle CkC_{k} indexed so that {{v1k,v2k},{v2k,v3k},,{v|Ck|k,v1k}}=Ck\{\{v_{1}^{k},v_{2}^{k}\},\{v_{2}^{k},v_{3}^{k}\},\dots,\{v_{|C_{k}|}^{k},v_{1}^{k}\}\}=C_{k}. Let A0kA_{0}^{k} be the subset of AA corresponding to the edges {v1k,v2k},{v3k,v4k},,{v|Ck1|k,v|Ck|k}\{v_{1}^{k},v_{2}^{k}\},\{v_{3}^{k},v_{4}^{k}\},\dots,\{v_{|C_{k-1}|}^{k},v_{|C_{k}|}^{k}\}, and let A1kA_{1}^{k} be the subset of AA corresponding to all the other edges in CkC_{k}. Then

(𝐕large×𝐕med)(aA0ka)=(𝐕large×𝐕med)(aA1ka).(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})\left(\prod_{a\in A_{0}^{k}}a\right)=(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})\left(\prod_{a\in A_{1}^{k}}a\right).

Thus, for any choice of ϵ{0,1}n\mathbf{\epsilon}\in\{0,1\}^{n},

k=1naAϵkka\prod_{k=1}^{n}\prod_{a\in A_{\epsilon_{k}}^{k}}a

has the same valuation at all large and medium primes. By Proposition 2.1, we thus have 2n=exp(O(N1/3+o(1)))2^{n}=\exp(O(N^{1/3+o(1)})), as desired. ∎

Example 3.2.

Suppose A={15,55,84,154,221,247,323,551,437,667}A=\{15,55,84,154,221,247,323,551,437,667\}, 𝒫small={2}\mathcal{P}_{\text{small}}=\{2\}, and 𝒫med𝒫large={3,5,7,11,13,17,19,23,29\mathcal{P}_{\text{med}}\cup\mathcal{P}_{\text{large}}=\{3,5,7,11,13,17,19,23,29}. Below is the prime factorization graph of AA, containing two even circuits. The product of the elements on the blue edges and the product of the elements on the red edges have the same valuations at all large and medium primes. By swapping the red-blue labeling in either circuit, we can produce four subsets whose products have the same valuations at all large and medium primes.

357111317192329151515415422122155155143743784845555247247323323667667
Corollary 3.3.

By removing O(N5/12+o(1))O(N^{5/12+o(1)}) edges from GG, we may assume that GG has no even circuits of length at most 2N1/122N^{1/12}.

Proof.

Let C1,,CnC_{1},\dots,C_{n} be the sets of edges in a maximal collection of edge-disjoint even circuits in GG, each with length at most 2N1/122N^{1/12}. We then have k=1n|Ck|=O(N5/12+o(1))\sum_{k=1}^{n}|C_{k}|=O(N^{5/12+o(1)}), so we may remove all the edges from k=1nCk\bigcup_{k=1}^{n}C_{k} from GG. After doing so, if there remains an even circuit of length at most 2N1/122N^{1/12} in GG, this would contradict the maximality of {C1,,Cn}\{C_{1},\dots,C_{n}\}. ∎

From here forward, we will assume that GG contains no even circuits of length at most 2N1/122N^{1/12}. Having removed short even circuits from GG, we now turn our focus to cycles with odd length. We first have the following proposition for a general graph GG.

Proposition 3.4.

Let GG be a graph with no even circuits of length at most 2M2M. Then any two odd cycles in GG of length at most MM are vertex-disjoint.

Proof.

We will prove the contrapositive. Suppose GG contains two odd cycles with vertex sets V1={v1,,vn}V_{1}=\{v_{1},\dots,v_{n}\} and V2={w1,,wm}V_{2}=\{w_{1},\dots,w_{m}\}, respectively, such that v1=w1v_{1}=w_{1}. Supposing also that n+m2Mn+m\leq 2M, we will show that GG has an even circuit of length at most 2M2M. Let E1={{v1,v2},,{vn,v1}}E_{1}=\{\{v_{1},v_{2}\},\dots,\{v_{n},v_{1}\}\} and E2={{w1,w2},,{wm,w1}}E_{2}=\{\{w_{1},w_{2}\},\dots,\{w_{m},w_{1}\}\} be their respective edge sets. We consider two cases.

First, assume E1E2=.E_{1}\cap E_{2}=\emptyset. Then {{v1,v2},,{vn,v1},{w1,w2},,{wm,w1}}\{\{v_{1},v_{2}\},\dots,\{v_{n},v_{1}\},\{w_{1},w_{2}\},\dots,\{w_{m},w_{1}\}\} form a circuit of length n+mn+m, which is even and at most 2M2M.

On the other hand, suppose E1E2E_{1}\cap E_{2}\neq\emptyset. Since E1E2E_{1}\neq E_{2}, we may assume without loss of generality that v1=w1v_{1}=w_{1}, {v1,vn}={w1,wm}\{v_{1},v_{n}\}=\{w_{1},w_{m}\}, but {v1,v2}{w1,w2}\{v_{1},v_{2}\}\neq\{w_{1},w_{2}\}. Let jj be the maximal index such that for all 1ij11\leq i\leq j-1, {vi,vi+1}E2\{v_{i},v_{i+1}\}\notin E_{2}. Let P1P_{1} and P2P_{2} be the edge sets from the two paths in E2E_{2} from v1v_{1} to vjv_{j}. Since |P1|+|P2||P_{1}|+|P_{2}| is odd, we can combine one of P1P_{1} or P2P_{2} with the edges from {{v1,v2},,{vj1,vj}}\{\{v_{1},v_{2}\},\dots,\{v_{j-1},v_{j}\}\} to form an even circuit of length at most n+m2Mn+m\leq 2M. ∎

Lemma 3.5.

Let C1,,CnC_{1},\dots,C_{n} be the sets of edges in a maximal collection of edge-disjoint odd circuits in GG, each with length at most N1/12N^{1/12} and each having a vertex from 𝒫\mathcal{P}_{\square}. Then n=O(N1/3+o(1))n=O(N^{1/3+o(1)}), so k=1n|Ck|=O(N5/12+o(1)).\sum_{k=1}^{n}|C_{k}|=O(N^{5/12+o(1)}).

Proof.

For each k[n]k\in[n], let {v1k,,v|Ck|k}\{v_{1}^{k},\dots,v_{|C_{k}|}^{k}\} be the set of vertices in the cycle CkC_{k} indexed so that v1k𝒫v_{1}^{k}\in\mathcal{P}_{\square} and {{v1k,v2k},{v2k,v3k},,{v|Ck|k,v1k}}=Ck\{\{v_{1}^{k},v_{2}^{k}\},\{v_{2}^{k},v_{3}^{k}\},\dots,\{v_{|C_{k}|}^{k},v_{1}^{k}\}\}=C_{k}. For each kk, let pkp_{k} be the prime corresponding to v1kv_{1}^{k}, and let aka_{k} be the element of AA such that pk2p_{k}^{2} divides aka_{k}. Note that by Proposition 3.4, the elements aka_{k} are distinct for distinct kk. Let A0kA_{0}^{k} be the set of elements of AA corresponding to the edges {v1k,v2k},{v3k,v4k},,{v|Ck|1k,v|Ck|k}\{v_{1}^{k},v_{2}^{k}\},\{v_{3}^{k},v_{4}^{k}\},\dots,\{v_{|C_{k}|-1}^{k},v_{|C_{k}|^{k}}\}, and let A1kA_{1}^{k} be the set of elements of AA corresponding to the other edges in CkC_{k}, union {ak}\{a_{k}\}.

Then

(𝐕large×𝐕med)(aA0ka)=(𝐕large×𝐕med)(aA1ka).(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})\left(\prod_{a\in A_{0}^{k}}a\right)=(\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}})\left(\prod_{a\in A_{1}^{k}}a\right).

Thus, for any choice of ϵ{0,1}n\mathbf{\epsilon}\in\{0,1\}^{n},

k=1naAϵkka\prod_{k=1}^{n}\prod_{a\in A_{\epsilon_{k}}^{k}}a

has the same valuation at all large and medium primes. By Proposition 2.1, we thus have 2n=exp(O(N1/3+o(1)))2^{n}=\exp(O(N^{1/3+o(1)})), as desired. ∎

By a similar argument to Corollary 3.3, by removing O(N5/12+o(1))O(N^{5/12+o(1)}) edges, we may assume without loss of generality that GG has no odd cycles of length at most N1/12N^{1/12} with a vertex from 𝒫\mathcal{P}_{\square}.

Example 3.6.

Suppose A={15,65,84,154,143,9}A=\{15,65,84,154,143,9\}, 𝒫small={2}\mathcal{P}_{\text{small}}=\{2\}, 𝒫med𝒫large={3,5,7,11,13}\mathcal{P}_{\text{med}}\cup\mathcal{P}_{\text{large}}=\{3,5,7,11,13\}, and 𝒫={3}\mathcal{P}_{\square}=\{3\}. Below is the prime factorization graph of AA. The product of the red elements (including 99) and the product of the blue elements have the same valuations at all large and medium primes.

11713539154154656514314384841515
Lemma 3.7.

Let EE denote the edge set of GG. There is a set of edges EEE^{\prime}\subseteq E of size at most 12(|𝒫□̸|+1)\frac{1}{2}(|\mathcal{P}_{\not\square}|+1) such that the graph with edge set EEE\setminus E^{\prime} contains no cycles of length at most N1/12N^{1/12}.

Proof.

Let C1,,CnC_{1},\dots,C_{n} be the sets of edges of all cycles of length at most N1/12N^{1/12} in GG. By Proposition 3.4, we may assume that these sets are disjoint and that they share no vertices. None of these cycles have a vertex from 𝒫\mathcal{P}_{\square}, and no vertices in 𝒫large\mathcal{P}_{\text{large}} are adjacent, so each CkC_{k} contains an edge between two vertices from 𝒫□̸{1}\mathcal{P}_{\not\square}\cup\{1\}. Since the vertex sets are disjoint, we thus have 2n|𝒫□̸{1}|2n\leq|\mathcal{P}_{\not\square}\cup\{1\}|. We may take EE^{\prime} to consist of one edge from each cycle CkC_{k}. ∎

4 Proof of Theorems 1.3 and 1.5

Having removed all cycles of length at most N1/12N^{1/12} from the prime factorization graph, we are ready to estimate the number of edges in it. We first record the following general estimate. This can be deduced from a result of Alon, Hoory, and Linial [1], but it is easier in our case, so we present a self-contained proof. The proof uses a standard breadth-first search argument; a similar argument can be found in [6], Theorem 1.6.5, for example.

Lemma 4.1.

Let GG be a graph with nn vertices and at least (1+c)n(1+c)n edges. Then GG has a cycle of length at most 2(c+1)c(log2(n)+1)\frac{2(c+1)}{c}(\log_{2}(n)+1).

Proof.

Without loss of generality, we may assume that GG is connected, since we may pass to a component satisfying |V|(1+c)|E||V|\geq(1+c)|E|. We can remove degree 11 vertices and their incident edges while preserving the inequality |V|(1+c)|E||V|\geq(1+c)|E|, so we may assume without loss of generality that each vertex has degree at least 2. If GG contains a path of length greater than c+1c\frac{c+1}{c}, then we can remove all internal edges and vertices from that path while preserving the inequality |V|(1+c)|E||V|\geq(1+c)|E|, so we may assume that GG has no such paths.

With all these assumptions set, we fix a vertex v0Vv_{0}\in V and consider a breadth-first search starting at v0v_{0}. For each kk\in\mathbb{N}, let Vk={vV: the shortest path from v to v0 has length k}V_{k}=\{v\in V:\text{ the shortest path from $v$ to $v_{0}$ has length $k$}\}. Fix kk\in\mathbb{N} with 1kc+1c(log2(n)1)1\leq k\leq\frac{c+1}{c}(\log_{2}(n)-1) and let \ell be the least integer greater than k+c+1ck+\frac{c+1}{c}.

If there is a vvv\in v_{\ell} such that there are two distinct paths of length \ell from vv to v0v_{0}, then we are done, since vv must thus contain a cycle of length at most 22\ell. Otherwise, for each vvv\in v_{\ell}, let φ(v)Vk\varphi(v_{\ell})\in V_{k} be the vertex from VkV_{k} on the shortest path from vv_{\ell} to v0v_{0}. For each vVkv\in V_{k}, we must have |φ1(v)|2|\varphi^{-1}(v)|\geq 2, since otherwise the path from vv to the only element in φ1(v)\varphi^{-1}(v) would be a path of length greater than c+1c\frac{c+1}{c} consisting only of degree two vertices. We thus have |V|2|Vk||V_{\ell}|\geq 2|V_{k}|. Letting mm be the least integer greater than c+1clog2(n)\frac{c+1}{c}\log_{2}(n), we must thus have |Vm|>2log2(n)=|V||V_{m}|>2^{\log_{2}(n)}=|V|, so there must be a vVmv\in V_{m} with at least two distinct paths from vv to v0v_{0}. There is thus a cycle of length at most 2m2m in GG, as desired. ∎

We will apply Lemma 4.1 with c=O(N1/12+o(1)).c=O(N^{-1/12+o(1)}). The following fact is required to ensure that the error term from Lemma 4.1 is multiplied by π(N1/2)\pi(N^{1/2}) instead of π(N)\pi(N) in our final estimate.

Lemma 4.2.

Let GG be the prime factorization graph of AA, containing no even circuits of length at most 2N1/122N^{1/12}. The set of vertices from 𝒫large\mathcal{P}_{\text{large}} with degree at least 22 has size at most (1+O(N1/12+o(1)))π(N1/2)(1+O(N^{-1/12+o(1)}))\pi(N^{1/2}).

Proof.

Let 𝒬𝒫large\mathcal{Q}\subseteq\mathcal{P}_{\text{large}} be the set of vertices of degree at least 22 in GG. Recall that no pair of vertices in 𝒫large\mathcal{P}_{\text{large}} is connected by an edge, so if p𝒬p\in\mathcal{Q}, then there are distinct v1,v2𝐕med{1}v_{1},v_{2}\in\mathbf{V}_{\text{med}}\cup\{1\} which are each connected to pp by an edge.

Consider the graph with vertex set {1}𝒫med\{1\}\cup\mathcal{P}_{\text{med}}, where v1v_{1} is adjacent to v2v_{2} if there is a path of length 22 from v1v_{1} to v2v_{2} in GG with the middle vertex in QQ. By construction, there is a surjection from the set of edges in this graph to 𝒬\mathcal{Q}. Moreover, this graph is simple since GG contains no 44-cycles.

Any cycle of length mm in this graph corresponds to a circuit of length 2m2m in GG. This graph thus does not contain any cycles of length at most N1/12N^{1/12}. Applying Lemma 4.1 with c=4log2(N)N1/12c=\frac{4\log_{2}(N)}{N^{1/12}}, we thus find that the number of edges is at most (1+O(N1/12+o(1)))(|𝒫med|+1)(1+O(N^{-1/12+o(1)}))(|\mathcal{P}_{\text{med}}|+1), so |𝒬|(1+O(N1/12+o(1)))π(N1/2)|\mathcal{Q}|\leq(1+O(N^{-1/12+o(1)}))\pi(N^{1/2}). ∎

We are now ready to prove Theorem 2.7, from which Theorems 1.3 and 1.5 follow immediately.

Proof of Theorem 2.7.

Let A[N]A\subseteq[N] have distinct subset products. By Corollary 2.3, we may remove O(N1/3+o(1))O(N^{1/3+o(1)}) elements from AA to ensure that 𝐕large×𝐕med\mathbf{V}_{\text{large}}\times\mathbf{V}_{\text{med}} is injective and nonzero when restricted to AA. We may now let G=(V,E)G=(V,E) be the prime factorization graph of AA.

By Corollary 3.3 and Lemma 3.5, we may remove O(N5/12+o(1))O(N^{5/12+o(1)}) elements from EE to ensure that AA has no even circuits of length at most 2N1/122N^{1/12}, and to ensure that AA has no odd cycles of length at most N1/12N^{1/12} with a vertex from 𝒫\mathcal{P}_{\square}. Then, we may apply Lemma 3.7 and remove 12|𝒫□̸|+O(1)\frac{1}{2}|\mathcal{P}_{\not\square}|+O(1) edges to remove all odd cycles of length at most N1/12N^{1/12} from GG.

Let 𝒬\mathcal{Q} be the set of all vertices in 𝒫large\mathcal{P}_{\text{large}} with degree at least two. We remove all vertices from 𝒫large𝒬\mathcal{P}_{\text{large}}\setminus\mathcal{Q} and their incident edges. Let VV^{\prime} and EE^{\prime} be the remaining sets of vertices and edges after all these removals. We have

|V|1+π(N1/2)+|𝒬||V^{\prime}|\leq 1+\pi(N^{1/2})+|\mathcal{Q}|

and

|E||E|(|𝒫large||𝒬|)12|𝒫□̸|O(N5/12+o(1)).|E^{\prime}|\geq|E|-(|\mathcal{P}_{\text{large}}|-|\mathcal{Q}|)-\frac{1}{2}|\mathcal{P}_{\not\square}|-O(N^{5/12+o(1)}).

The graph (V,E)(V^{\prime},E^{\prime}) has no cycles of length at most N1/12N^{1/12}. Applying Lemma 4.1 with c=4log2(N)N1/12c=\frac{4\log_{2}(N)}{N^{1/12}}, we find

|E|(1+O(N1/12+o(1))|V|.|E^{\prime}|\leq(1+O(N^{-1/12+o(1)})|V^{\prime}|.

This results in the inequality

|E|(|𝒫large||𝒬|)12|𝒫□̸|(1+O(N1/12+o(1)))(π(N1/2)+|𝒬|)+O(N5/12+o(1))=|E|-(|\mathcal{P}_{\text{large}}|-|\mathcal{Q}|)-\frac{1}{2}|\mathcal{P}_{\not\square}|\leq(1+O(N^{-1/12+o(1)}))(\pi(N^{1/2})+|\mathcal{Q}|)+O(N^{5/12+o(1)})=
π(N1/2)+|𝒬|+O(N5/12+o(1)),\pi(N^{1/2})+|\mathcal{Q}|+O(N^{5/12+o(1)}),

where we used Lemma 4.2 so that O(N1/12+o(1))|𝒬|=O(N5/12+o(1))O(N^{-1/12+o(1)})|\mathcal{Q}|=O(N^{5/12+o(1)}).

Finally, using the equation |A|=|E|+|𝒫|+O(N1/3+o(1))|A|=|E|+|\mathcal{P}_{\square}|+O(N^{1/3+o(1)}), we find

|A|π(N1/2)+|𝒬|+|𝒫large||𝒬|+12|𝒫□̸|+|𝒫|+O(N5/12+o(1))|A|\leq\pi(N^{1/2})+|\mathcal{Q}|+|\mathcal{P}_{\text{large}}|-|\mathcal{Q}|+\frac{1}{2}|\mathcal{P}_{\not\square}|+|\mathcal{P}_{\square}|+O(N^{5/12+o(1)})
π(N)+12π(N1/2)+12|𝒫|+O(N5/12+o(1)),\leq\pi(N)+\frac{1}{2}\pi(N^{1/2})+\frac{1}{2}|\mathcal{P}_{\square}|+O(N^{5/12+o(1)}),

since 𝒫large=π(N)π(N1/2)\mathcal{P}_{\text{large}}=\pi(N)-\pi(N^{1/2}) and |𝒫|+|𝒫□̸|π(N1/2)|\mathcal{P}_{\square}|+|\mathcal{P}_{\not\square}|\leq\pi(N^{1/2}). ∎

5 Examples

In this section, we will discuss some examples of large sets with distinct subset products.

Our graph-theoretic perspective enables us to produce some new examples of sets A[N]A\subseteq[N] with distinct subset products and |A|π(N)+π(N1/2)|A|\approx\pi(N)+\pi(N^{1/2}).

Example 5.1.

Let GG be a tree with vertex set 𝒫med𝒫small\mathcal{P}_{\text{med}}\cup\mathcal{P}_{\text{small}}. Let A=𝒫large{p2:p𝒫med𝒫small}{pq:p,q𝒫med𝒫small are connected by an edge in G}A=\mathcal{P}_{\text{large}}\cup\{p^{2}:p\in\mathcal{P}_{\text{med}}\cup\mathcal{P}_{\text{small}}\}\cup\{pq:p,q\in\mathcal{P}_{\text{med}}\cup\mathcal{P}_{\text{small}}\text{ are connected by an edge in $G$}\}. Then |A|=π(N)+π(N1/2)1|A|=\pi(N)+\pi(N^{1/2})-1 and AA has distinct subset products.

Next, we will produce an example of a set |A||A| with distinct subset products and |A|π(N)+π(N1/2)+13π(N1/3)O(1)|A|\geq\pi(N)+\pi(N^{1/2})+\frac{1}{3}\pi(N^{1/3})-O(1), establishing Theorem 1.4.

Proof of Theorem 1.4.

Construct a set AA as follows. For any prime p(N1/3,N]p\in(N^{1/3},N], let pAp\in A. For any prime p(N1/3,N1/2]p\in(N^{1/3},N^{1/2}], let p2Ap^{2}\in A. Let p,q,rp,q,r be distinct primes in [N1/3][N^{1/3}]. The set {p2q,p2r,p2,qr,p3,q3,r3}\{p^{2}q,p^{2}r,p^{2},qr,p^{3},q^{3},r^{3}\} is a subset of [N][N] with distinct subset products. We may thus divide the primes from [N1/3][N^{1/3}] into disjoint subsets of size 33 (perhaps leaving one or two primes out) and for each triple {p,q,r}\{p,q,r\}, add the 77 elements {p2q,p2r,p2,qr,p3,q3,r3}\{p^{2}q,p^{2}r,p^{2},qr,p^{3},q^{3},r^{3}\} to AA. The number of elements of AA is thus at least

π(N)π(N1/2)+2(π(N1/2)π(N1/3))+7(13π(N1/3)O(1))=\pi(N)-\pi(N^{1/2})+2(\pi(N^{1/2})-\pi(N^{1/3}))+7\left(\frac{1}{3}\pi(N^{1/3})-O(1)\right)=
π(N)+π(N1/2)+13π(N1/3)O(1).\pi(N)+\pi(N^{1/2})+\frac{1}{3}\pi(N^{1/3})-O(1).\qed

Finally, we may use our graph-theoretic perspective to produce a set of squarefree integers with distinct subset products and nearly π(N)+12π(N1/2)\pi(N)+\frac{1}{2}\pi(N^{1/2}) elements, establishing Theorem 1.6.

Proof of Theorem 1.6.

Let ϵ>0\epsilon>0, and let NN be sufficiently large (depending on ϵ\epsilon). By the prime number theorem, for every k[1,1/ϵ2]k\in[1,1/\epsilon-2], the number of primes pp satisfying the inequality

N1/2(k+1)ϵN1/2p<N1/2kϵN1/2N^{1/2}-(k+1)\epsilon N^{1/2}\leq p<N^{1/2}-k\epsilon N^{1/2}

is less than twice the number of primes pp satisfying

N1/2+(k1)ϵN1/2<pN1/2+kϵN1/2.N^{1/2}+(k-1)\epsilon N^{1/2}<p\leq N^{1/2}+k\epsilon N^{1/2}.

We may thus list all but at most 3ϵN1/23\epsilon N^{1/2} of the primes in [1,N1/2][1,N^{1/2}] as q1,r1,q2,r2,,qn,rnq_{1},r_{1},q_{2},r_{2},\dots,q_{n},r_{n} so that there are distinct primes p1,,pnp_{1},\dots,p_{n} in (N1/2,N](N^{1/2},N] with piqiNp_{i}q_{i}\leq N, piriNp_{i}r_{i}\leq N for all ii.

Let 𝒬\mathcal{Q} be the set of large primes not in {p1,,pn}\{p_{1},\dots,p_{n}\}. Let

A=𝒬i=1n{piqi,piri,qiri}i=1n{riqi+1}.A=\mathcal{Q}\cup\bigcup_{i=1}^{n}\{p_{i}q_{i},p_{i}r_{i},q_{i}r_{i}\}\cup\bigcup_{i=1}^{n}\{r_{i}q_{i+1}\}.

Then |A|π(N)+12π(N1/2)3ϵπ(N1/2)2|A|\geq\pi(N)+\frac{1}{2}\pi(N^{1/2})-3\epsilon\pi(N^{1/2})-2 and AA has distinct subset products.

The prime factorization graph of AA is as follows (omitting the vertices from 𝒬\mathcal{Q}); it is constructed so that the argument from Lemma 3.7 is sharp.

q1q_{1}r1r_{1}p1p_{1}q2q_{2}r2r_{2}p2p_{2}\dots

We have established that for all ϵ>0\epsilon>0, there is an N0N_{0}\in\mathbb{N} such that for all NN0N\geq N_{0}, h(N)π(N)+12π(N1/2)3ϵπ(N1/2)2h(N)\geq\pi(N)+\frac{1}{2}\pi(N^{1/2})-3\epsilon\pi(N^{1/2})-2. Thus, h(N)π(N)+12π(N1/2)o(π(N1/2))h(N)\geq\pi(N)+\frac{1}{2}\pi(N^{1/2})-o(\pi(N^{1/2})). ∎

References

  • [1] Noga Alon, Shlomo Hoory, and Nathan Linial. The Moore bound for irregular graphs. Graphs and Combinatorics, 18:53–57, 2002.
  • [2] Quentin Dubroff, Jacob Fox, and Max Wenqiang Xu. A note on the Erdős distinct subset sums problem. SIAM Journal on Discrete Mathematics, 35(1):322–324, 2021.
  • [3] Paul Erdős. Extremal problems in number theory II. Mat. Lapok, 17:135–155, 1966.
  • [4] Paul Erdős. Some applications of graph theory to number theory. In The Many Facets of Graph Theory: Proceedings of the Conference held at Western Michigan University, Kalamazoo/MI., October 31–November 2, 1968, pages 77–82. Springer, 1969.
  • [5] Stefan Steinerberger. Some remarks on the Erdős distinct subset sums problem. International Journal of Number Theory, 19(08):1783–1800, 2023.
  • [6] Yufei Zhao. Graph Theory and Additive Combinatorics: Exploring Structure and Randomness. Cambridge University Press, 2023.