When almost all sets are difference dominated
Peter Hegarty, Steven J. Miller

TL;DR
This paper proves that for a wide range of random subsets of {0,...,N}, the difference set is almost always larger than the sum set, revealing a threshold phenomenon and extending results to generalized difference sets and binary linear forms.
Contribution
It confirms a conjecture that almost all such random subsets are difference dominated and characterizes the threshold behavior of sum and difference set sizes.
Findings
Almost all random subsets are difference dominated as N grows large.
Identifies a threshold at p(N) ~ N^{-1/2} where the ratio of difference to sumset size transitions.
Extends results to generalized difference sets and binary linear forms with sharp thresholds.
Abstract
We investigate the relationship between the sizes of the sum and difference sets attached to a subset of {0,1,...,N}, chosen randomly according to a binomial model with parameter p(N), with N^{-1} = o(p(N)). We show that the random subset is almost surely difference dominated, as N --> oo, for any choice of p(N) tending to zero, thus confirming a conjecture of Martin and O'Bryant. The proofs use recent strong concentration results. Furthermore, we exhibit a threshold phenomenon regarding the ratio of the size of the difference- to the sumset. If p(N) = o(N^{-1/2}) then almost all sums and differences in the random subset are almost surely distinct, and in particular the difference set is almost surely about twice as large as the sumset. If N^{-1/2} = o(p(N)) then both the sum and difference sets almost surely have size (2N+1) - O(p(N)^{-2}), and so the ratio in question is almost…
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsLimits and Structures in Graph Theory · Analytic Number Theory Research · Mathematical Approximation and Integration
When almost all sets are
difference dominated
Peter Hegarty
Mathematical Sciences, Chalmers University Of Technology and Göteborg University, Göteborg, Sweden
and
Steven J. Miller
Department of Mathematics and Statistics, Williams College, Williamstown, MA 01267
Abstract.
We investigate the relationship between the sizes of the sum and difference sets attached to a subset of , chosen randomly according to a binomial model with parameter , with . We show that the random subset is almost surely difference dominated, as , for any choice of tending to zero, thus confirming a conjecture of Martin and O’Bryant. The proofs use recent strong concentration results.
Furthermore, we exhibit a threshold phenomenon regarding the ratio of the size of the difference- to the sumset. If then almost all sums and differences in the random subset are almost surely distinct, and in particular the difference set is almost surely about twice as large as the sumset. If then both the sum and difference sets almost surely have size , and so the ratio in question is almost surely very close to one. If then as increases from zero to infinity (i.e., as the threshold is crossed), the same ratio almost surely decreases continuously from two to one according to an explicitly given function of .
We also extend our results to the comparison of the generalized difference sets attached to an arbitrary pair of binary linear forms. For certain pairs of forms and , we show that there in fact exists a sharp threshold at , for some computable constant , such that one form almost surely dominates below the threshold, and the other almost surely above it.
The heart of our approach involves using different tools to obtain strong concentration of the sizes of the sum and difference sets about their mean values, for various ranges of the parameter .
Key words and phrases:
Sum dominated sets, Binary linear forms, Strong concentration, Thresholds.
2000 Mathematics Subject Classification:
11P99 (primary), 11K99 (secondary).
1. Introduction
To know whether a random variable is strongly concentrated is an issue of fundamental importance in many areas of mathematics and statistics. In this paper we apply recent results of Kim and Vu [KiVu, Vu1, Vu2] to completely solve a combinatorial number theory question on the size of difference- and sumsets of integers. A classical strong concentration result (due to Chernoff) states that if with the i.i.d. binary random variables, then for any we have . Within number theory, this result was used by Erdős (see [AS], Chapter 8) to prove the existence of so-called thin’ bases of of order 2. The general requirement for many applications is to obtain Chernoff-like exponential deviation bounds in situations when the atom variables are not independent. For modern surveys of strong concentration inequalities see, for example, [Ta] and [Vu2]; the latter, in particular, contains a fine selection of applications in random graph theory, combinatorial number theory and finite geometry.
The specific result we shall utilise is a martingale inequality which appears as Lemma 3.1 in [Vu2]. It is an extension of the classical Azuma inequality ([AS], Chapter 7) to functions whose Lipschitz coefficients are small on average’. As remarked in [Vu2], this type of inequality is very general and robust, and is expected to be applicable in numerous situations; this is definitely true for our problem.
Let be a subset of the integers. We define the sumset and difference set by
[TABLE]
and denote the cardinality of a set by . As addition is commutative and subtraction is not, a typical pair of integers generates two differences but only one sum. It is therefore reasonable to expect a generic finite set will have a larger difference set than sumset. We say a set is sum dominated (such sets are also called more sums than differences, or MSTD, sets) if the cardinality of its sumset exceeds that of its difference set. If the two cardinalities are equal we say the set is balanced, otherwise difference dominated. Sum dominated sets exist: consider for example (see [He, MS, Na2] for additional examples). In [Na1], Nathanson wrote “Even though there exist sets that have more sums than differences, such sets should be rare, and it must be true with the right way of counting that the vast majority of sets satisfies .”
Recently Martin and O’Bryant [MO] showed there are many sum dominated sets. Specifically, let . They prove the existence of a universal constant such that, for any , at least subsets of are sum dominated (there are no sum dominated sets in ). Their proof is based on choosing a subset of by picking each independently with probability . The argument can be generalized to independently picking each with any probability , and yields the existence of a constant such that, as , a randomly chosen (with
respect to this model) subset is sum dominated with probability at least . Similarly one can prove there are positive constants and for the probability of having a difference dominated or balanced set.
While the authors remark that, perhaps contrary to intuition, sum dominated sets are ubiquitous, their result is a consequence of how they choose a probability distribution on the space of subsets of . Suppose , as in their paper. With high probability a randomly chosen subset will have elements (with errors of size ). Thus the density of a generic subset to the underlying set is quite high, typically about . Because it is so high, when we look at the sumset (resp., difference set) of a typical there are many ways of expressing elements as a sum (resp., difference) of two elements of . For example (see [MO]), if then there are roughly ways of writing as a sum of two elements in (similarly, if there are roughly ways of writing as a difference of two elements of ). This enormous redundancy means almost all numbers which can be in the sumset or difference set are. In fact, using uniform density on the subsets of (i.e., taking ), Martin and O’Bryant show that the average value of is and that of is (note each set has at most elements). In particular, it is only for near extremes that we have high probability of not having in an or an . In [MO] they prove a positive percentage of subsets of (with respect to the uniform density) are sum dominated sets by specifying the fringe elements of . Similar conclusions apply for any value of .
At the end of their paper, Martin and O’Bryant conjecture that if, on the other hand, the parameter is a function of tending to zero arbitrarily slowly, then as the probability that a randomly chosen subset of is sum dominated should also tend to zero. In this paper we will, among other things, prove this conjecture.
We shall find it convenient to adopt the following (fairly standard) shorthand notations. Let be a real-valued random variable depending on some positive integer parameter , and let be some real-valued function. We write ’ to denote the fact that, for any , there exists such that, for all ,
[TABLE]
In particular we shall use this notation when is just a function of (hence not random’). In practice will in this case be the expectation of some other random variable.
By we mean that there exist constants and such that for all , . By we mean that both and hold. Finally, if then we write .
Our main findings can be summed up in the following theorem.
Theorem 1.1**.**
Let be any function such that
[TABLE]
For each let be a random subset of chosen according to a binomial distribution with parameter . Then, as , the probability that is difference dominated tends to one.
*More precisely, let denote respectively the random variables and . Then the following three situations arise :
(i) : Then*
[TABLE]
(ii) for some : Define the function by
[TABLE]
Then
[TABLE]
(iii) : Let , . Then
[TABLE]
Remark 1.2**.**
Obviously, not all functions satisfying (1.3) conform to the requirements of (i), (ii) or (iii) above, but these are the natural functions to investigate in the current context. Similar remarks apply to Theorem 3.1 and Conjecture 4.2 below.
Theorem 1.1 proves the conjecture in [MO] and re-establishes the validity of Nathanson’s claim in a broad setting. It also identifies the function as a threshold function, in the sense of [JŁR], for the ratio of the size of the difference- to the sumset for a random set . Below the threshold, this ratio is almost surely , above it almost surely . Part (ii) tells us that the ratio decreases continuously (a.s.) as the threshold is crossed. Below the threshold, part (i) says that most sets are nearly Sidon sets’, that is, most pairs of elements generate distinct sums and differences. Above the threshold, most numbers which can be in the sumset (resp., difference set) usually are, and in fact most of these in turn have many different representations as a sum (resp., a difference). However the sumset is usually missing about twice as many elements as the difference set. Thus if we replace sums’ (resp., differences’) by missing sums’ (resp., missing differences’), then there is still a symmetry between what happens on both sides of the threshold.
We prove Theorem 1.1 in the next section. Our strategy will consist of first establishing an estimate for the expectation of the random variable or , followed by establishing sufficiently strong concentration of these variables about their mean values. For the second part of this strategy we will use different approaches for and . In the former range a fairly straightforward second moment argument works. In the latter range, however, we will employ a specialization of the Kim-Vu martingale lemma (Lemma 3.1 in [Vu2], Lemma 2.2 below).
In Section 3, we extend our result to arbitrary binary linear forms. The paper [NOORS] provides motivation for studying these objects. By a binary linear form we mean a function where , and GCD. For a set of integers we let
[TABLE]
Except in the special case we always have that whenever . Thus we refer to as a difference form and a set as a generalized difference set, whenever . Theorem 3.1 allows us to compare the sizes of and for random sets , and arbitrary difference forms and when .
Two situations arise :
(a) for some pairs of forms, the same one a.s. dominates the other for all parameters in this range. In fact every other difference form dominates , and hence also .
(b) for certain pairs and , something very nice happens. Namely, there is now a sharp threshold, in the sense of [JŁR], at , for some computable constant , depending on and . One form dominates a.s. below the threshold, and the other one a.s. above it. This fact may be considered a partial generalization of the main result of [NOORS] to random sets, partial in the sense that is only applies to certain pairs of forms. Namely, they proved that for any two forms and (including ), there exist finite sets and such that whereas .
We leave it to future work to investigate what happens as the threshold is crossed in this situation.
In Section 4 we make a brief summary of this and other remaining questions and make suggestions for other problems to study. In particular, we suggest looking at other probabilistic models for choosing random sets. This is partly motivated by the fact that our results in Section 3 only apply when . The reason is that, for faster decaying , as we shall see, the variance in the size of the random set itself swamps all other error terms, and it is meaningless to compare and ; in other words, the model itself becomes useless. This may be considered a problem when . For , the results of [GJLR] imply that all pairs in a random set a.s. generate different values , for any , so that a.s. for any and .
2. Proof of Theorem 1.1
Our strategy for proving the various assertions in Theorem 1.1 is the following. Let be one of the random variables , as appropriate. We then carry out the following two steps :
Step 1 : Prove that behaves asymptotically as asserted in the theorem.
Step 2 : Prove that is strongly concentrated about its mean.
As already mentioned, the calculations required to perform these two steps differ according as to whether or . In particular, in the former case, Step 2 is achieved by a fairly straightforward second moment argument, whereas a more sophisticated concentration inequality is used in the latter case. We thus divide the proof of the theorem into two separate cases, depending on the parameter function .
Throughout the paper we often abuse notation to save space, writing for . As we never consider the case where is constant (as this case has been analyzed in [MO]), this should not cause any confusion.
Case I : .
We first concentrate on the sumset and prove the various assertions in parts (i) and (ii) of the theorem. The proofs for the difference set will be similar. For any finite set and any integer , let
[TABLE]
In words, consists of all unordered -tuples of unordered pairs of elements of having the same sum. Let . So if is a random set, then each is a non-negative integer valued random variable. The crucial observation for our work is that, in the model we are considering, the random variables are all highly concentrated :
Lemma 2.1**.**
For we have for every that
[TABLE]
and, more significantly, whenever .
Proof.
We write for . By the central limit theorem it is clear that
[TABLE]
Each can be written as a sum of indicator variables , one for each unordered -tuple . There are two types of -tuples : those consisting of distinct elements of and those in which one element is repeated twice in one of the pairs, and the sum of each of the pairs is even. The probability of any -tuple of the former type occurring in is , whereas for -tuples of the latter type this probability is . Let there be a total of -tuples of the former type and of the latter type. Then, by linearity of expectation,
[TABLE]
We have
[TABLE]
where is the number of representations of as a sum of two distinct elements of , and hence we easily estimate
[TABLE]
[TABLE]
Thus
[TABLE]
A similar calculation shows that , hence . Since it follows that
[TABLE]
in accordance with the lemma. To complete the proof of the lemma, we need to show that, whenever , the random variable becomes highly concentrated about its mean as . We apply a standard second moment method. In the notation of [AS], Chapter 4, since we already know in this case that , it suffices to show that , where
[TABLE]
the sum being over pairs of -tuples which have at least one number in common. It is easy to see that, since , the main contribution to comes from pairs of -tuples, each of which consist of distinct elements of , and which have exactly one element in common. The number of such pairs is since there are
- •
choices for ,
- •
choices for the common element with ,
- •
choices for the sum of each pair in ,
- •
choices for the remaining elements in .
Since a total of elements of occur in total in , we have . Thus
[TABLE]
This completes the proof of Lemma 2.1. ∎
We can now prove parts (i) and (ii) of the theorem. First suppose . By (2.2) we have , whereas for and otherwise.
Since we have and thus almost surely. In other words, as , all but a vanishing proportion of pairs of element of will have distinct sums. It follows immediately that
[TABLE]
as claimed.
Now suppose for some fixed . This time we will need to consider all the together. Let be the partition on whereby and are in the same part if and only if . For each let denote the number of parts of size (as a random variable). Observe that
[TABLE]
and, for each , that
[TABLE]
(2.13) is a system of infinitely many equations in the variables , which together determine . For any , the general solution of the subsystem formed by the first equations (i.e.: ) is readily checked to be
[TABLE]
Regarding the second sum on the right of (2.14) we have that
[TABLE]
Hence it follows that, for any ,
[TABLE]
Now for , Lemma 2.1 says that
[TABLE]
and since as , another application of Lemma 2.1 implies that
[TABLE]
So to prove (1.6) it just remains to verify that
[TABLE]
which is an easy exercise.
This proves parts (i) and (ii) of Theorem 1.1 for the sumset. For the difference set one reasons in an entirely parallel manner. One now defines, for each ,
[TABLE]
In words, consists of all -tuples of ordered pairs of elements of which have the same non-zero difference. We let and in a completely analogous manner to Lemma 2.1 prove that
[TABLE]
and that whenever . We define the partition of in the obvious way and let denote the number of parts of size , for each . Since we can follow exactly the same analysis as above to deduce (1.4) and (1.6). This completes the proofs of parts (i) and (ii) of the theorem.
Case II : .
Recall . Set and ; thus (as we have ). Again we begin with the sumset. Recall the two steps to be accomplished :
Step 1 : We prove that .
Step 2 : We prove that the random variable is strongly concentrated about its mean.
We begin with the simpler Step 1. For each , let denote the event that . Thus
[TABLE]
Observe that . Since all the ways of representing any given as a sum of two elements of are independent of one another, we have, for ,
[TABLE]
Since we have , and since we have . Thus it is easy to see that
[TABLE]
as claimed. This completes Step 1.
For Step 2 we need the martingale machinery of Kim and Vu.
We use notation consistent with [Vu2]. Consider a fixed , which shall tend to infinity in our estimates. Let . Thus every subset of can be identified with an element of . We are working in the probability space where is the product measure with parameter . For each , and , define
[TABLE]
by we mean the expected value of the random variable , given that for the element is always (resp., never) in the subset if (resp., ). Let
[TABLE]
Further put
[TABLE]
and
[TABLE]
For two arbitrary positive numbers and , define the event
[TABLE]
Then the following is a specialization of a result appearing in [Vu2] :
Lemma 2.2**.**
For any positive numbers such that we have
[TABLE]
We quickly sketch how Lemma 2.2 completes the proof of assertion (iii) of Theorem 1.1. We shall take
[TABLE]
We show that for appropriately chosen we have
[TABLE]
From (2.30) and , for sufficiently small we will then be able to conclude that
[TABLE]
As (see (2.24)), assertion (iii) in Theorem 1.1 follows immediately. Thus we are reduced to proving (2.32), which we now proceed to do.
First we simplify things a little. For any and , we introduce the shorthand
[TABLE]
Let
[TABLE]
As , we see . For ,
[TABLE]
Since
[TABLE]
from (2.25) simplifies to
[TABLE]
Since for all sufficiently large , we have then
[TABLE]
Further, (2.27), (2.28) and (2.38) yield
[TABLE]
This completes our simplifications.
Recall that in order to use (2.30) from Lemma 2.2 we need to prove (2.32) (namely that ). The heart of the proof of (2.32) is to show that for an appropriate choice of , with probability all three of the following events occur:
[TABLE]
[TABLE]
[TABLE]
We claim that (2.41)-(2.43) imply (2.32). This follows immediately from applying the trivial bound and using (2.41)-(2.43) to show these two probabilities are both .
To summarize, the proof is completed by verifying (2.41)-(2.43). Notice also that (2.41) and (2.42), together with (2.40), imply (2.43) for any choice of , so it just remains to prove the former two. As in the arguments that follow there is a symmetry between and , we consider with ; the remaining follow similarly.
First, consider (2.41). Note that, depending on the parameter , this sum could be empty. This will not affect the argument to follow. The proof is by an averaging argument, i.e.: for each we first consider . This quantity has a very natural interpretation : in words, it is the expected increase in the size of the sumset brought about by the addition of the number to . For every , adding to will add to if and only if and, for any other numbers such that , either or . Let be the event that gets added to by the addition of to . Then using (2.23) we can explicitly estimate
[TABLE]
Since , the last sum is asympotically no more than
[TABLE]
By Markov’s inequality, we deduce that for any ,
[TABLE]
Then, just using a trivial union bound
[TABLE]
it follows that, with probability at least
[TABLE]
we have
[TABLE]
Then (2.41) clearly follows provided
[TABLE]
which is clearly the case for sufficiently large , since .
We now turn to (2.42). Firstly, a similar argument to the one just given shows that, even if , adding to a random set is very probably not going to add any elements at all to which are larger than . Secondly, among the numbers in , the addition to of one number cannot add to more numbers than were in already, plus maybe one more. But Chernoff’s inequality ([AS], Corollary A.14) implies that, with probability , where is some universal positive constant, . Then (2.42) follows from a simple union bound, as long as for example.
This completes the proof of the assertion of Theorem 1.1(iii) as regards the sumset.
For the difference set, we proceed in two identical steps. First consider the estimate of . Let now denote instead the event that for each . Clearly,
[TABLE]
For each we have
[TABLE]
where is the (bad) event that both and lie in and is the complementary event. These events are not independent, but the dependencies will not affect our estimates. To see this rigorously, one can for example use Janson’s inequality (see [AS], Chapter 8, though this is certainly overkill!)
[TABLE]
where all , and
[TABLE]
the sum being over dependent pairs , i.e.: pairs such that .
Note that we can take , we have and
[TABLE]
since there is a 1-1 correspondence between dependent pairs and 3-term arithmetic progressions in of common difference . It is then easy to see that this correction term can be ignored when we make the estimate
[TABLE]
as desired.
The concentration of about its mean can be established in the same way as we did with above. A little more care is required in estimating quantities analogous to , because of the dependencies between different representations of the same difference, but Janson’s inequality can again be used to see rigorously that this will not affect our estimates. We omit further details and simply note that we will again obtain the result that
[TABLE]
This completes the proof of Theorem 1.1.
3. General Binary Linear Forms
We have the following generalization of Theorem 1.1 :
Theorem 3.1**.**
*Let be a function satisfying (1.3). Let be non-zero integers with , GCD and . Put . For a positive integer , let be a random subset of obtained by choosing each independently with probability . Let denote the random variable . Then the following three situations arise :
(i) : Then*
[TABLE]
(ii) for some : Define the function by
[TABLE]
Then
[TABLE]
(iii) : Let . Then
[TABLE]
Proof.
One follows exactly the method of proof of Theorem 1.1, so we only give a sketch here.
Case I : .
We again write for . For any finite set and any integer , let
[TABLE]
Let . Then (2.21) has the following generalization :
[TABLE]
and whenever .
We shall just sketch the proof that behaves like the right-hand side of (3.6) in the case when . The proof for is similar, and the concentration of about its mean when is established by the same kind of second moment argument as in Section 2.
If then for any we have . Then
[TABLE]
where
[TABLE]
and denotes the number of solutions to the equation satisfying . For any integer , the general integer solution to is of course
[TABLE]
where . If then there are solutions in non-negative integers, and for all such solutions, . For the following three situations then arise :
(I) : then all non-negative solutions satisfy , so in this case.
(II) : we have for any such .
(III) : we have for these .
Thus it follows that
[TABLE]
[TABLE]
[TABLE]
which, together with (3.7), verifies our claim that behaves like the right-hand side of (3.6).
Once we have (3.6) then, in a similar manner to Section 2, we can prove part (i) of Theorem 3.1 by noting that almost surely when , and part (ii) by showing that
[TABLE]
when . It’s a simple exercise to check that (3.13) and (3.6) yield (3.3).
Case II : .
We give a sketch of the estimate for , the details of the concentration estimate being completely analogous to what has gone before. Let us continue to assume , the proof for being similar. As in the proof of Theorem 1.1(iii) one may check that various dependencies do not affect our estimates which, using observations (I),(II),(III) above, lead to
[TABLE]
[TABLE]
The sum is and the second term is negligible since , so , as claimed.
This completes the proof of Theorem 3.1. ∎
As mentioned earlier, the main result of [NOORS] was that, for any two binary forms and , including the case when , there exist finite sets of integers such that and . Theorem 3.1 has a number of consequences on the matter of comparing and for given and and random subsets of for large . We now reserve the notations and for two forms being compared. Unless otherwise stated, we assume neither nor is the form . A generic form will be denoted .
It is convenient to formalize a piece of terminology which we used informally in the introduction :
Definition 3.2**.**
Let be two binary linear forms as above. Let satisfy (1.3). Then we say that dominates for the parameter if, as , almost surely when is a random subset of obtained by
choosing each independently with probability .
We now consider three different regimes (depending on how rapidly decays). In the arguments below we shall write for . The most interesting behavior will be isolated afterwards as Theorem 3.3.
Regime 1 : .
Then part (iii) of Theorem 3.1 implies, in particular, that and . Hence dominates when . In particular, this is the case if . If then the theorem says that dominates if and only if , which is the case if and only if .
Regime 2 : .
Part (i) of Theorem 3.1 says that for any and . For every we have
[TABLE]
Thus
[TABLE]
The second moment method gives standard deviations
[TABLE]
In particular we have and . First of all, then, if , i.e.: if , then the uncertainty in the size of the random set itself swamps everything else, and our model is worthless.
If then, by (3.17), it is in the first instance the -term which will be decisive. By (3.6) we have
[TABLE]
Hence dominates in this range of if where
[TABLE]
Since it is easy to see that if and only if , this allows us to compare any pair of forms in the range and , except a pair . But for such a pair, our methods are entirely worthless anyway, since all the estimates in this section depend only on . Note in particular that for any so that any other form dominates .
Regime 3 : .
By part (iii) of Theorem 3.1, for a given value of the parameter , dominates if
[TABLE]
Since as , will dominate for sufficiently large values of , provided . This is as expected from Regime 1. On the other hand, the Taylor expansion of , as a function of , around , reads
[TABLE]
Thus dominates for sufficiently small values of provided . Again this is as expected, this time from Regime 2. Note that the injectivity of allows us to even compare forms with the same value of , namely : for a fixed value of , is clearly a decreasing function of . Hence if then dominates for all values of if and only if . Note that this is the same condition as in Regime 1. More generally, we have that dominates for all values of whenever and . In particular this is the case for and any other .
The most interesting phenomenon arises when we compare two forms such that
[TABLE]
Then the combined observations of Regimes 1, 2 and 3 imply that there
exists some
such that
[TABLE]
whereas
[TABLE]
This observation may be considered a partial generalization of the main result of [NOORS] to random sets, partial in the sense that it only applies to pairs of forms satisfying (3.23). Equations (3) and (3.25) say that we have a sharp threshold, below which dominates and above which dominates . We leave it to future work to determine what happens as one crosses this sharp threshold.
We close this section by summarizing the most important observations above in a theorem.
Theorem 3.3**.**
*Let and , where , GCD and . Let be the function given by (3.20). The following two situations can be distinguished :
(i) and .
Then dominates for all such that and . In particular, every other difference form dominates the form in this range.
(ii) and .
Then there exists such that (3) and (3.25) hold. Specifically, is the unique positive root of the equation*
[TABLE]
where is given by (3.2).
4. Open Problems
Here is a sample of issues which could be the subject of further investigations :
1. One unresolved matter is the comparison of arbitrary difference forms in the range where and . Here the problem is that the binomial model itself does not prove of any use. This provides, more generally, motivation for looking at other models. Obviously one could look at the so-called uniform model on subsets (see [JŁR]), but this seems a more awkward model to handle. Note that the property of one binary form dominating another is not monotone, or even convex.
2. Secondly, a very tantalizing problem is to investigate what happens while crossing a sharp threshold, whenever it arises under the conditions of Theorem 3.3(ii).
3. Thirdly, one can ask if the various concentration estimates in Theorem 1.1 can be improved. When we have only used an ordinary second moment argument, and it is possible to provide explicit estimates. Explicitly, the following follows from Chebyshev’s Theorem (see the appendix for a proof).
Theorem 4.1**.**
Let for some , . Set , and let be any function such that for all . Set and . For any subset chosen with respect to the binomial model with parameter , with probability at least the ratio of the cardinality of its difference set to the cardinality of its sumset is . Thus the probability a subset chosen with respect to the binomial model is not difference dominated is at most , which tends to zero rapidly with for .
The range seems more interesting, however. Here we proved that the random variable has expectation of order , where , and is concentrated within of its mean. Now one can ask whether the constant can be improved, or at the very least can one get rid of the logarithm?
4. Finally, it is natural to ask for extensions of our results to -linear forms in more than two variables. Let
[TABLE]
be such a form. We conjecture the following generalization of Theorem 3.1 :
Conjecture 4.2**.**
Let be a function satisfying (1.3). For a positive integer , let be a random subset of obtained by choosing each independently with probability . Let be as in (4.1) and assume that GCD. Set
[TABLE]
*Let denote the random variable . Then the following three situations arise :
(i) : Then*
[TABLE]
*(ii) for some : There is a rational function in variables, which is increasing in , and an increasing function
such that*
[TABLE]
(iii) : Let . Then
[TABLE]
Acknowledgement
We thank Devdatt Dubhashi for pointing us in the direction of Vu’s paper, the participants of CANT 2007 for interesting conversations, and the referees for comments on an earlier draft. The second named author was partly supported by NSF grant DMS0600848.
Appendix A Explicit bounds
Here we prove Theorem 4.1. The proof uses essentially only Chebyshev’s inequality. We have deliberately made this section entirely self-contained, rather than appealing to results from [AS], for the benefit of readers who may not be too familiar with discrete probability theory.
We first establish some notation, and then prove a sequence of lemmas from which Theorem 4.1 immediately follows. Our goal is to provide explicit bounds which decay like to a power.
Let denote the binary indicator variable for being in a subset (it is thus with probability and [math] otherwise), and let be the random variable denoting the cardinality of a subset (thus ). For two pairs of ordered elements and in (, ), let if , and [math] otherwise.
Lemma A.1**.**
With probability at least ,
[TABLE]
Let denotes the number of ordered pairs (with ) in a subset of chosen with respect to the binomial model. Then with probability at least we have
[TABLE]
Proof.
We have . As the are independent,
[TABLE]
Thus
[TABLE]
By Chebyshev’s inequality,
[TABLE]
For we choose so that
[TABLE]
Thus , and the probability that lies in the stated interval is at least . The second claim follows from the fact that there are ways to choose two distinct objects from objects. ∎
Remark A.2**.**
By using the Central Limit Theorem instead of Chebyshev’s inequality we may obtain a better estimate on the probability of lying in the desired interval; however, as the Central Limit Theorem is not available for some of the later arguments, there is negligible gain in using it here.
Proof of Theorem 4.1. By Lemma A.1, (A.2) holds with probability at least . The main contribution to the cardinalities of the sumset and the difference set is from ordered pairs with . With probability at least there are on the order such pairs, which is much larger than the order pairs with . The proof is completed by showing that almost all of the ordered pairs yield distinct sums (and differences). Explicitly, we shall show that for a subset chosen from with respect to the binomial model, if is the number of ordered pairs (which is of size with high probability), then with high probability the cardinality of its difference set is while the cardinality of its sumset is . This argument crucially uses (if ) then the error term is the same size as the main term, and the more delicate argument given in the main text is needed). We shall show that almost all of the ordered pairs generate distinct differences; the argument for the sums follows similarly.
Each ordered pair yields two differences ( and ). The problem is that two different ordered pairs could generate the same differences. To calculate the size of the difference set, we need to control how often two different pairs give the same differences. Consider two distinct ordered pairs and with and (as the ‘diagonal’ pairs yield the same difference, namely 0, it suffices to study the case of ordered pairs with distinct elements). Without loss of generality we may assume . If then these two pairs contribute the same differences. There are two possibilities: (1) all four indices are distinct; (2) .
We calculate the expected number of pairs of non-diagonal ordered pairs with the same difference by using our binary indicator random variables . Set
[TABLE]
If the four indices are distinct then ; if then .
The number of tuples of distinct integers satisfying our conditions is bounded by (once , and are chosen there is at most one choice for with )111Although we do not need the actual value, simple algebra yields the number of tuples is .. If instead then there are at most tuples satisfying our conditions (once and are chosen, and are uniquely determined, though they may not satisfy our conditions). Therefore
[TABLE]
as .
As is much smaller than for , most of the differences are distinct. To complete the proof, we need some control on the variance of . In Lemma A.3 we show that
[TABLE]
where
[TABLE]
While we cannot use the Central Limit Theorem (as the are not independent and also depend on ), we may use Chebyshev’s inequality to bound the probability that is close to its mean (recall the mean is at most ). We have
[TABLE]
Simple algebra shows that if we take then with probability at least we have , which is a positive power of less than . Thus an at most negligible amount of the differences are repeated.
The argument for two ordered pairs yielding the same sum proceeds similarly: if then .
For our ratio to be , two events must happen. As the probability the first does not occur is at most and the probability the second does not occur is at most , the probability that the two desired events happen is at least .
Except for the claimed estimate on , the above completes the proof of Theorem 4.1. We now prove our bound for .
Lemma A.3**.**
Let the notation be as in Theorem 4.1 and (A.10). We have
[TABLE]
Proof.
If and are two random variables, then
[TABLE]
By the Cauchy-Schwartz inequality, . Thus
[TABLE]
We may therefore write
[TABLE]
where in the -sum all four indices are distinct (with , , and ) and in the -sum all three indices are distinct (with and and ). As , we are reduced to bounding the variances of and .
We first bound . Standard algebra yields
[TABLE]
As and there are at most ordered tuples of distinct integers with , the term is bounded by .
For the covariance piece, if all eight indices () are distinct, then and are independent and thus the covariance is zero. There are four cases; in each case there are always at most choices for the tuple , but often there will be significantly fewer choices for the tuple . We only provide complete details for the first and third cases, as the other cases follow similarly.
- •
Seven distinct indices: There are at most choices for . The covariance of each such term is bounded by . To see this, note
[TABLE]
The product of the expected values is , while the expected value of the product is . Thus the covariances of these terms contribute at most .
- •
Six distinct indices: The covariances of these terms contribute at most .
- •
Five distinct indices: The covariances of these terms contribute at most (once three of the have been determined, the fourth is uniquely determined; thus there are at most choices for the first tuple and at most 1 choice for the second).
- •
Four distinct indices: The covariances of these terms contribute at most .
The -dependence from the case of seven distinct indices is greater than the -dependence of the other cases (except for the case of four distinct indices if ). We also only increase the contributions if we replace with . We therefore find
[TABLE]
Similarly we have
[TABLE]
The piece is bounded by (as there are at most tuples with ). The covariance terms vanish if the six indices are distinct. A similar argument as before yields bounds of for five distinct indices, for four distinct indices, and for three distinct indices. The largest -dependence is from the term (as ). Arguing as before and replacing with yields
[TABLE]
As , . Therefore
[TABLE]
which yields
[TABLE]
∎
Remark A.4**.**
An extreme choice of would be to choose , for some small positive constant . Since for all , we then obtain a bound of for the ratio of the cardinality of the difference set to the sumset with probability .
Remark A.5**.**
Alternatively, one can get a tighter bound on the ratio than in Theorem 4.1 at the expense of having a bound on the probability which is non-trivial only for . For example, if one instead chooses in (A.11) and , then the statement of Theorem 4.1 still holds, but is a non-trivial statement only for , i.e.: for . This is a natural choice of , given the results of [GJLR]. Specifically, if then almost surely no differences or sums are repeated, and the set is a Sidon set (and therefore we trivially have the ratio of the cardinality of the difference set to the sumset is approximately ).
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1[AS] N. Alon and J. H. Spencer, The Probabilistic Method , Wiley, 1992.
- 2[GJLR] A. P. Godbole, S. Janson, N. W. Locantore Jr. and R. Rapoport, Random Sidon sequences , J. Number Theory 75 (1999), no. 1, 7–22.
- 3[He] P. V. Hegarty, Some explicit constructions of sets with more sums than differences , Acta Arith. 130 (2007), no. 1, 61–77.
- 4[HM] P. V. Hegarty and S. J. Miller, When almost all sets are difference dominated , preprint. http://www.arxiv.org/abs/0707.3417
- 5[JŁR] S. Janson, T. Łuczak and A. Ruciński, Random Graphs , Wiley, 2000.
- 6[Ki Vu] J. H. Kim and V. H. Vu, Concentration of multivariate polynomials and its applications , Combinatorica 20 (2000), 417–434.
- 7[MO] G. Martin and K. O’Bryant, Many sets have more sums than differences , Additive combinatorics, 287–305, CRM Proc. Lecture Notes 43 , Amer. Math. Soc., Providence, RI, 2007.
- 8[MS] S. J. Miller and D. Scheinerman, Explicit constructions of infinite families of MSTD sets , preprint. http://arxiv.org/abs/0809.4621
