A Bilinear Bogolyubov Argument in Abelian Groups
- Mathematical Institute, Serbian Academy of Sciences and Arts
- More about Luka Milićević
Editorial introduction
A bilinear Bogolyubov argument in abelian groups, Discrete Analysis 2024:20, 41 pp.
Bogolyubov’s argument states that if A is a dense subset of Zn, then 2A−2A={x+y−z−w:x,y,z,w∈A} contains a large and highly structured subset known as a Bohr set. If r1,…,rk are residue classes mod n and δ>0, then the Bohr set B(r1,…,rk;δ) is defined to be the set of all x∈Zn such that |1−exp(2πirjx/n)|≤δ for all j=1,2,…,k. (There are equivalent definitions, but this one is convenient for the discussion here.) Such sets, when k is small and δ is not too small (which we get from Bogolyubov’s argument), resemble low-codimensional subspaces of a vector space, and this additive structure has been exploited in many proofs in additive combinatorics.
Bogolyubov’s argument can be straightforwardly generalized to arbitrary finite Abelian groups. If G is a finite Abelian group, χ1,…,χk are characters on G and δ>0, then the Bohr set B(χ1,…,χk;δ) is the set of all x∈G such that |1−χj(x)|≤δ for all j=1,2,…,k. When G=Fnp, where the Bohr sets actually are low-codimensional subspaces. That is, if A⊂Fnp has density α, then 2A−2A contains a subspace of Fnp of codimension at most C(α). The question of how large C(α) needs to be is very interesting: thanks to work of Tom Sanders it is known to be polylogarithmic in 1/α, but it is conjectured to be logarithmic.
A different kind of generalization of Bogolyubov’s method was obtained by Bienvenu and Lê, and independently by Gowers and the author of this paper, again in the case G=Fnp. Speaking very loosely, Bogolyubov’s argument tells us that if we take iterated sumsets of a dense subset of Fnp, then the resulting sets will become more and more subspace-like. Suppose now that we take a dense subset A of Fnp×Fnp. Call sections of A of the form Ax∙={y∈Fnp:(x,y)∈A} the rows of A and sections of the form A∙y={x∈Fnp:(x,y)∈A} the columns of A. Define the horizontal difference set of A to be the result of replacing each row Ax∙ of A by its difference set Ax∙−Ax∙, and define the vertical difference set in a similar way for the columns.
Bogolyubov’s argument tells us that if we take the horizontal difference set twice, then all dense rows of A will contain low-codimensional subspaces, and if we take the vertical difference set twice, then all dense columns will. Thus, one might expect that if one takes a suitable composition of these two operations, one will end up with a set that is subspace-like in both the vertical and horizontal directions, though this is not trivial because one needs to show that the operations in one direction do not undo the progress made in the other direction.
What might such a set look like? An instructive example is the following. Let M be a linear map from Fnp to Fnp and let A be the set of all (x,y)∈Fnp×Fnp such that x.My=0, (where a.b stands for the mod-p inner product ∑ni=1aibi). Then for each x, the row Ax∙ is the set of all y such that x.My=0, which is a subspace of codimension 1, except when M∗x=0 when it is all of Fnp. A similar statement holds for columns, so all the rows and columns of A are subspaces.
Sets of the kind just defined are 1-dimensional bilinear varieties in Fnp×Fnp. Taking an intersection of k of them, we obtain a k-dimensional bilinear variety. The theorem of Bienvenu and Lê and of Gowers and Milićević is a bilinear version of Bogolyubov’s argument, which states that if A is a dense subset of Fnp×Fnp then there is a composition of a bounded number of operations of taking horizontal and vertical difference sets (the bound here is absolute – it does not depend on the density of A) such that the resulting set contains a bilinear variety of bounded codimension (here the bound does depend on the density of A).
Improved bounds for the bilinear Bogolyubov theorem were later obtained by Hosseini and Lovett.
This paper generalizes the bilinear Bogolyubov argument to arbitrary finite Abelian groups. For some of the proof, it follows the approach of Hosseini and Lovett, but it diverges from it at a point where their argument involves considering ranks of linear combinations of linear maps. Even before the point of divergence, the generalization is not trivial: there are many concepts that are straightforward for Fnp but that need to be generalized rather carefully in order to obtain a proof for arbitrary Abelian groups. It is likely that these definitions and associated tools will be useful for other purposes, such as proving inverse theorems for Uk norms in general Abelian groups.