Comments on: What are dense Sidon subsets of {1,2,…,n} like?

By: Seva

Seva — Sun, 13 Jan 2013 08:38:24 +0000

(This comment is identical to that I posted yesterday, but with a proper formatting attempted.)

I wonder whether one can use Ben’s idea in conjunction with Ruzsa’s construction, as follows. Let be a prime, and consider indices (discrete logarithms) with respect to a fixed primitive root modulo . With every integer associate the residue class of modulo . The set of all resulting elements of is a Sidon set. Define by . Writing , we see that to answer Erdos’ question, it suffices to show that the full image misses some, say, consecutive elements of . A simple heuristic shows that is most certainly true (a typical element of is missing in with probability about ), but is there a hope to prove this? Notice also that we do not need to have the result for all pairs (where is a primitive root modulo ), but just for any infinite sequence of such pairs.

By: Seva

Seva — Sat, 12 Jan 2013 11:19:42 +0000

I wonder whether one can use Ben’s idea in conjunction with Ruzsa’s
construction, as follows. Let $p$ be a prime, and consider indices (discrete
logarithms) with respect to a fixed primitive root modulo $p$. With every
integer $x\in[1,p-1]$ associate the residue class of $(p-1)x+p\,{\rm
ind}\,(x)$ modulo $p(p-1)$. The set of all $p-1$ resulting elements of
${\mathbb Z}_{p(p-1)}$ is a Sidon set. Define $\varphi\colon
[1,p-1]\to{\mathbb Z}_{p-1}$ by $\varphi(x)=x+{\rm ind}\, x\pmod {p-1}$.
Writing $(p-1)x+p\,{\rm ind}\,(x)=(x+{\rm ind}\,(x))p -x$, we see that to
answer Erdos’ question, it suffices to show that the full image
$\varphi([1,p-1])$ misses some, say, $\log\log\log p$ consecutive elements of
${\mathbb Z}_{p-1}$. A simple heuristic shows that is most certainly true (a
typical element of ${\mathbb Z}_{p-1}$ is missing in $\varphi([1,p-1])$ with
probability about $1/e$), but is there a hope to \emph{prove} this? Notice
also that we do not need to have the result for all pairs $(p,g)$ (where $g$
is a primitive root modulo $p$), but just for any infinite sequence of such
pairs.

By: Another way of looking at the alephs « cartesian product

Another way of looking at the alephs « cartesian product — Mon, 17 Sep 2012 16:57:39 +0000

[...] What are dense Sidon subsets of {1,2,…,n} like? (gowers.wordpress.com) [...]

By: yan

yan — Sat, 21 Jul 2012 06:18:39 +0000

Hi Dr Gowers
my name is Yan， I read “what is implied by “implies”"you wrote
And I have a question for you. Can you just spend a little bit time and give me some help？
it is annoying question about vacuous proof.

for example， let’s prove ∅ is the subset of every set A.
This statement can be translated into another way：for every x， if x is the element of ∅，then x is the element of A.
since x is the element of ∅ is always false. So this statement is vacuous true.
My problem is why can‘t we substitue x by non-exist things，for example， what if x represents unicorn？then unicorn is an element of ∅ will be true not false？so x is the element of empty set is not always true

by this problem， can we get the conclusion that all the varibles in the statement can just be substituted by exist things? which means the scope of a variable can just be exist thing?

I appreciate your help Dr Gowers~

By: obryant

obryant — Sat, 21 Jul 2012 05:05:38 +0000

As a point of history, the Erdos-Turan proof and the Lindstrom proof both give (iirc, the “+1″ can be improved to “+1/2″), although E-T only noted .

By: obryant

obryant — Sat, 21 Jul 2012 05:03:11 +0000

There’s an idea I’ve been sitting on for years, and haven’t been able to turn it into anything. First let me give a dubious heuristic-y construction of a Sidon set, and then I’ll give an actual example to demonstrate that, at least, the idea isn’t idiotic.

Let be a dynamical system, and suppose that is ergodic, and that . Let be an arbitrary set (measurable, of course), and let be a randomly chosen point of . Now let be the set of integers where is parameter to be named later. Since is ergodic, the expected size of is ; perhaps we can take so that . I claim that is pressured to be a Sidon set, at least after removing a small number of elements.

Why? Suppose that are all in , and 0' title='a-c=d-b > 0' class='latex' />, contrary to the claimed Sidon-ness of (set ). This means that the two points , are both in , and so are “close together”. But also maps these two close points to , , which are not only close to each other but close to the original two points. My intuition is that if is sufficiently strongly mixing, then it should be very rare (i.e., very few ) that a small power of maps two points in to two points in . Caveat: I don’t know what “sufficiently strongly mixing”, “very rare”, or “small power” actually need to mean to make this work, if indeed there is any way to make it work.

So here’s the promised example, which is really a Bose-Chowla set in disguise. Let , Lebesgue measure, and define to be . This is a linear map with determinant 6, so it is ergodic. Set , and let be any of the sets with (or replace the special role played by with , that works, too). Let . (Above, I suggested taking random and fixed , but here I'm taking one and many 's. Oh well.) All 32 of the resulting sets are Sidon sets!

All of the constructions of Sidon sets that give are based on finite fields, and the error term in the maximum possible size of a Sidon set are from the gaps between prime numbers. Any construction that gives without using primes would be interesting, and possibly the first improvement since the early 1940s.

By: obryant

obryant — Sat, 21 Jul 2012 04:21:11 +0000

FYI, the constructions of Bose & Chowla (mod ), Singer (mod ), and Ruzsa (mod ) are given in my now old survey/bibliography at Electronic Journal of Combinatorics (http://www.combinatorics.org/ojs/index.php/eljc/article/view/ds11).

By: Ben Green

Ben Green — Fri, 20 Jul 2012 07:58:41 +0000

Indeed, it seems very hard to improve Lindstrom’s result in any way. Erdos conjecture is surely false, as most likely one can dilate a Sidon subset of Z/qZ of size q^{1/2} so as to have a gap of size q^{1/2}logloglog q (say), and then “unwrap” so as to get a Sidon set of size q^{1/2} in an interval of length appreciably smaller than q. However, I have no idea how to achieve this with any of the known constructions of large Sidon sets modulo q. Being able to find such large gaps is most likely a generic property (i.e. has nothing to do with being Sidon) but I have no idea how to prove that either.

By: Ben Green

Ben Green — Fri, 20 Jul 2012 07:54:48 +0000

Mark, this is an interesting idea, and I like this conjecture of Helfgott and Venkatesh very much. However I suspect that *very* dense Sidon sets, of size really close to the maximum of $\sqrt{N}$, are rather well-distributed in residue classes to small moduli.
See

http://arxiv.org/pdf/math/9808061.pdf

By: Ben Green

Ben Green — Fri, 20 Jul 2012 07:46:38 +0000

Tim, I think there are constructions of Sidon subsets of $\{1,…,N\}$ of size about $\sqrt{N}$ coming from finite fields, due to Bose and Chowla. Consider $F_p$ inside $F_{p^2}$, and let $a$ be a generator of the multiplicative group of $F_{p^2}$. Let $S \subseteq Z/(p^2 – 1)Z$ be the set of all $s$ such that $a^s – a \in F_p$. Then this is a Sidon set (in fact it is Sidon modulo $p^2 – 1$). Ruzsa has some other constructions too. I certainly believe that all large Sidon sets are “algebraic” in some way that I am unable to make precise. This example of Ruzsa shows that one cannot hope for too much in that direction. I must say that I wasn’t previously aware of it either, despite having read Ruzsa’s paper in detail some years ago.

By: David Roberts

David Roberts — Thu, 19 Jul 2012 04:53:02 +0000

Ah, I see! But the discussion at that point was about infinite Sidon sets, so I was distracted. Thanks, Benoît.

By: Benoît Régent-Kloeckner

Benoît Régent-Kloeckner — Wed, 18 Jul 2012 07:29:12 +0000

What Tim Gowers meant is that in a given interval [0,N], there are far more logs of prime than integer.

By: David Roberts

David Roberts — Tue, 17 Jul 2012 07:24:46 +0000

“the logs of the primes are not integers. Indeed, there are far more of them than there are integers.” Not quite right. There are exactly the same number of integers and logs of primes. I suspect that the ‘them’ is meant to refer to reals.

By: Neil Calkin

Neil Calkin — Mon, 16 Jul 2012 14:43:29 +0000

Erdos and Turan originally showed $n^{1/2} + O(n^{1/4})$ as an upper bound. Lindstrom gave a beautiful really short proof of an upper bound of $n^{1/2} + n^{1/4} +1$, which looks like it is really loose, and could be tightened, but it’s deceptive.
IErdos offered $500 for a proof of an upper bound of $n^{1/2} + O(1)$.

By: Thomas

Thomas — Mon, 16 Jul 2012 09:52:57 +0000

Actually, what I wrote in 2) isn’t properly correct as it stands, but I see a similar problem with the example in the post. Namely, how can one guarantee that the set actually has size at least $m/\log m$? For many different $p\leq m$ are such that the integer part of $\log p$ are the same. Surely one needs the primes to be spaced at least some multiplicative constant apart, whence one only gets a set of size around $\log m$?

By: Klas Markström

Klas Markström — Mon, 16 Jul 2012 08:39:32 +0000

Looking at the data for n=44 the size distribution for all Sidon sets is unimodal, the most common size is 7 and the maximum is 9. Here is the list of {size, number of Sidon sets} pairs for n=44.
{{0, 1}, {1, 44}, {2, 946}, {3, 13 244}, {4, 129 360}, {5, 845 408},
{6, 3 157 104}, {7, 4 748 144}, {8, 1 462 854}, {9, 27 202}}

For larger n I only had data for restricted Sidon set, the ones which are well distributed mod p, but they show a similar distribution.

Again this is for very(!) small n, but it looks like as “close” to maximum size one might need to be close by something additive rather than a factor.

By: Klas Markström

Klas Markström — Mon, 16 Jul 2012 08:17:36 +0000

One of my old programs looked for Sidon sets such that every initial segment is as balanced as possible mod p, no residue classes differing by more than 1.
There are of course fewer such sets but their sizes seem to grow at a rate comparable with that of general Sidon sets, perhaps smaller by some p-dependent factor. The same looks plausible when one uses several prime at the same time.

Here are two examples which are balanced mod 2, 3, and 5
{1, 8, 15, 34, 42, 53, 65, 96, 109, 112, 113, 114}
{1, 8, 15, 24, 47, 64, 66, 77, 85, 88, 113, 114}

By: Klas Markström

Klas Markström — Mon, 16 Jul 2012 08:10:06 +0000

I took a look at some old data on Sidon sets which I had on my laptop, I once had a long wait in an airport which led to some simple computer experiments with Sidon sets, and from those data it looks like c is 4.1… in the estimate for the number of Sidon sets.

However this is data from what my laptop could do at the time so they are only for n up to 44, for general Sidon sets.

By: Thomas

Thomas — Mon, 16 Jul 2012 06:53:40 +0000

1) I think the size of the Sidon set constructed is around in the penultimate paragraph.

Thanks — corrected now.

2) You can also construct quite nasty examples from iterating , and the floor functions. For example, you can start from the primes and instead of I think one obtains a Sidon set of roughly the same density by considering . You could also take where is taken from a Sidon set of a completely different nature. I imagine one could repeat this process. It’s difficult to guess what kind of structure is preserved by repeated logarithms and floor functions…

By: Rob

Rob — Mon, 16 Jul 2012 00:01:29 +0000

Dear Tim and Gil,

Another result in this direction was obtained recently by Kohayakawa, Lee, Rodl and Samotij (see Theorem 2.1):

https://www.dpmms.cam.ac.uk/~ws299/papers/Sidon.pdf

I think they also get some structural information about a ‘typical’ (in a weak sense) Sidon set of a given size, but it sounds like you’re hoping for something much stronger…

By: gowers

gowers — Sun, 15 Jul 2012 22:11:10 +0000

Hmm, I’ve just realized something that has a big influence on any conjecture one might wish to make, even in the case of sets of size . If you have a Sidon subset of of that size, then you can create a Sidon subset of, say, by multiplying every element of your original set by 5 and arbitrarily adding 0, 1 or -1 to it. So there’s no hope of quadratic structure: the best one can hope for is a set that can be approximated by something quadratic. (Of course, then there are other things one can do like multiplying one of these perturbed examples by some appropriate mod , so things get even worse.)

By: Mark Lewko

Mark Lewko — Sat, 14 Jul 2012 04:42:52 +0000

This is vaguely reminiscent of the some theorems and conjectures loosely attached to the name “inverse sieve.” One assertion of this form states roughly that if such that has size near and misses a constant fraction of the residue classes mod for small then is contained in the image of a quadratic polynomial. Helfgott and Venkatesh have some results in this direction. With this in mind, one might try to prove that a large Sidon set (or some appropriate subset and transformations thereof) must be ill-distributed mod for small . I have no intuition suggesting if this is plausible.

By: gowers

gowers — Fri, 13 Jul 2012 22:22:38 +0000

A recent result of David Conlon’s and mine gives at least something in this kind of direction. We show that if G is an Abelian group of order n and A is a random subset of G of size , then with high probability every Freiman homomorphism from A to another Abelian group extends to a Freiman homomorphism on G. This is completely untrue for Sidon sets, since then every map is a Freiman homomorphism. However, the property that all Freiman homomorphisms extend is much stronger than the property of not being Sidon, so the bound you can get this way on the number of Sidon sets is almost certainly far too weak. I imagine that as with other problems of this kind the number of Sidon sets is roughly comparable to the number of subsets of a single maximal Sidon set — that is, . I don’t even have a guess if you try to count Sidon sets of size for (but I haven’t thought about it, so maybe it’s easy to come up with a sensible conjecture).

By: Terence Tao

Terence Tao — Fri, 13 Jul 2012 16:17:15 +0000

A characterisation of dense Sidon sets may also lead to progress on another intractable Erdos problem, namely to determine if there is an additive basis A of the natural numbers of order 2 (i.e. A+A=N) with bounded multiplicity function r_2 (i.e. every natural number is expressible as the sum of two elements of A in a bounded number of ways).

Note though that if we relax the Sidon property to that of having multiplicity that grows at most logarithmically, then random examples start working again thanks to Chernoff. This already rules out a lot of analytical techniques (e.g. Fourier analysis), as they have a hard time distinguishing between multiplicity 1 and logarithmic multiplicity. I’m not sure what that leaves one with; algebraic methods such as the polynomial method are a remote possibility, but I haven’t seen them used for inverse sumset problems before, only for direct sumset inequalities.

By: Gil Kalai

Gil Kalai — Fri, 13 Jul 2012 16:05:50 +0000

Tim, Do you believe that every sidon set with t>0 already has structure? A second question: maybe easier than structure would be an upper bound on the number of sidon sets of the required size?