Additive Combinatorics and its Applications in Theoretical CS

Slides:



Advertisements
Similar presentations
Vector Spaces A set V is called a vector space over a set K denoted V(K) if is an Abelian group, is a field, and For every element vV and K there exists.
Advertisements

Shortest Vector In A Lattice is NP-Hard to approximate
On Complexity, Sampling, and -Nets and -Samples. Range Spaces A range space is a pair, where is a ground set, it’s elements called points and is a family.
How should we define corner points? Under any reasonable definition, point x should be considered a corner point x What is a corner point?
The Structure of Polyhedra Gabriel Indik March 2006 CAS 746 – Advanced Topics in Combinatorial Optimization.
Computational Geometry The art of finding algorithms for solving geometrical problems Literature: –M. De Berg et al: Computational Geometry, Springer,
Exercise 1- 1’ Prove that if a point B belongs to the affine / convex hull Aff/Conv (A 1, A 2, …, A k ) of points A 1, A 2,…, A k, then: Aff/Conv (A 1,
2.III. Basis and Dimension 1.Basis 2.Dimension 3.Vector Spaces and Linear Systems 4.Combining Subspaces.
1.4 Exercises (cont.) Definiton: A set S of points is said to be affinely (convex) independent if no point of S is an affine combination of the others.
A general approximation technique for constrained forest problems Michael X. Goemans & David P. Williamson Presented by: Yonatan Elhanani & Yuval Cohen.
Discrete Mathematics, 1st Edition Kevin Ferland
Chapter 9: Geometric Selection Theorems 11/01/2013
Induction Proof. Well-ordering A set S is well ordered if every subset has a least element. [0, 1] is not well ordered since (0,1] has no least element.
The Simplex Algorithm 虞台文 大同大學資工所 智慧型多媒體研究室. Content Basic Feasible Solutions The Geometry of Linear Programs Moving From Bfs to Bfs Organization of a.
I.4 Polyhedral Theory 1. Integer Programming  Objective of Study: want to know how to describe the convex hull of the solution set to the IP problem.
CS 103 Discrete Structures Lecture 13 Induction and Recursion (1)
SECTION 8 Groups of Permutations Definition A permutation of a set A is a function  ϕ : A  A that is both one to one and onto. If  and  are both permutations.
Summary of the Last Lecture This is our second lecture. In our first lecture, we discussed The vector spaces briefly and proved some basic inequalities.
Linear Programming Chap 2. The Geometry of LP  In the text, polyhedron is defined as P = { x  R n : Ax  b }. So some of our earlier results should.
1 Section 4.4 Inductive Proof What do we believe about nonempty subsets of N? Since  N, <  is well-founded, and in fact it is linear, it follows that.
3.3 Mathematical Induction 1 Follow me for a walk through...
1 Chapter 4 Geometry of Linear Programming  There are strong relationships between the geometrical and algebraic features of LP problems  Convenient.
Theory of Computational Complexity Probability and Computing Chapter Hikaru Inada Iwama and Ito lab M1.
7.3 Linear Systems of Equations. Gauss Elimination
Chapter 2 Sets and Functions.
Chapter 3 The Real Numbers.
Chapter 5 Limits and Continuity.
Unit-III Algebraic Structures
A sketch proof of the Gilbert-Pollak Conjecture on the Steiner Ratio
Chapter 3 The Real Numbers.
Advanced Algorithms Analysis and Design
Great Theoretical Ideas In Computer Science
Chapter 3 The Real Numbers.
Chapter 3 The Real Numbers.
Chapter 2 Sets and Functions.
Chapter 5. Optimal Matchings
Joint work with Avishay Tal (IAS) and Jiapeng Zhang (UCSD)
Vapnik–Chervonenkis Dimension
Distinct Distances in the Plane
Quantum Two.
Set Operations Section 2.2.
GROUPS & THEIR REPRESENTATIONS: a card shuffling approach
Polyhedron Here, we derive a representation of polyhedron and see the properties of the generators. We also see how to identify the generators. The results.
Sum-product theorems over finite fields
Basis and Dimension Basis Dimension Vector Spaces and Linear Systems
Depth Estimation via Sampling
Basis Hung-yi Lee.
The Curve Merger (Dvir & Widgerson, 2008)
Polyhedron Here, we derive a representation of polyhedron and see the properties of the generators. We also see how to identify the generators. The results.
Chapter 4 Sequences.
V11 Metabolic networks - Graph connectivity
Memoryless Determinacy of Parity Games
2.III. Basis and Dimension
Dilworth theorem and Duality in graph
Affine Spaces Def: Suppose
Quantum Two.
I.4 Polyhedral Theory (NW)
Quantum Foundations Lecture 3
Maths for Signals and Systems Linear Algebra in Engineering Lecture 6, Friday 21st October 2016 DR TANIA STATHAKI READER (ASSOCIATE PROFFESOR) IN SIGNAL.
Computational Geometry
Back to Cone Motivation: From the proof of Affine Minkowski, we can see that if we know generators of a polyhedral cone, they can be used to describe.
V12 Menger’s theorem Borrowing terminology from operations research
I.4 Polyhedral Theory.
Advanced Analysis of Algorithms
V11 Metabolic networks - Graph connectivity
(Convex) Cones Def: closed under nonnegative linear combinations, i.e.
V11 Metabolic networks - Graph connectivity
Vector Spaces RANK © 2012 Pearson Education, Inc..
THE DIMENSION OF A VECTOR SPACE
Locality In Distributed Graph Algorithms
Presentation transcript:

Additive Combinatorics and its Applications in Theoretical CS By Shachar Lovett Chapter 2: Set addition, Sections 2.1-2.3, pp. 3-8. Presented by Tomer Bincovich

Set addition 𝐺 an abelian group (think about 𝐺= 𝔽 𝑛 ). 𝐴⊆𝐺. The sumset of 𝐴 is 2𝐴=𝐴+𝐴= 𝑎+ 𝑎 ′ 𝑎, 𝑎 ′ ∈𝐴 Always 2𝐴 ≥ 𝐴 . When does equality hold?

When does 2𝐴 =|𝐴|? Equality holds if 𝐴 is empty, a subgroup of 𝐺 or a coset of a subgroup. For the other direction: If 𝐴≠∅, we can assume w.l.o.g. 0∈𝐴 by shifting. Then 𝐴⊆2𝐴, and since 2𝐴 = 𝐴 we have that 2𝐴=𝐴. 𝐴 is a nonempty finite subset that is closed under addition, hence is a subgroup.

Planned topics Ruzsa calculus The span of sets of small doubling The growth of sets of small doubling

Ruzsa calculus A set of basic inequalities between sizes of sets and their sumsets. For 𝐴,𝐵⊆𝐺, define: Sumset: 𝐴+𝐵= 𝑎+𝑏 𝑎∈𝐴,𝑏∈𝐵 Difference set: 𝐴−𝐵= 𝑎−𝑏 𝑎∈𝐴,𝑏∈𝐵 Claim 2.1 (Ruzsa triangle inequality): 𝐴 𝐵−𝐶 ≤ 𝐴−𝐵 𝐴−𝐶

Proof of Claim 2.1 Claim 2.1 (Ruzsa triangle inequality): 𝐴 𝐵−𝐶 ≤ 𝐴−𝐵 𝐴−𝐶 Define a map 𝑓:𝐴× 𝐵−𝐶 → 𝐴−𝐵 × 𝐴−𝐶 : for any 𝑥∈𝐵−𝐶 fix 𝑏∈𝐵,𝑐∈𝐶 s.t. 𝑥=𝑏−𝑐, and define 𝑓 𝑎,𝑥 = 𝑎−𝑏,𝑎−𝑐 . 𝑓 is injective, hence 𝐴 𝐵−𝐶 ≤ 𝐴−𝐵 𝐴−𝐶 .

The Ruzsa distance 𝑑 𝐴,𝐵 = log 𝐴−𝐵 𝐴 1 2 𝐵 1 2 Not formally a distance function. Why? Claim 2.3: The Ruzsa distance is symmetric and obeys the triangle inequality.

Proof of Claim 2.3 Symmetry: since 𝐵−𝐴 = 𝐴−𝐵 . 𝑑 𝐴,𝐵 = log 𝐴−𝐵 𝐴 1 2 𝐵 1 2 Symmetry: since 𝐵−𝐴 = 𝐴−𝐵 . Moreover, 𝑑 𝐴,𝐶 ≤𝑑 𝐴,𝐵 +𝑑 𝐵,𝐶 is equivalent to log 𝐴−𝐶 𝐴 1 2 𝐶 1 2 ≤ log 𝐴−𝐵 𝐴 1 2 𝐵 1 2 + log 𝐵−𝐶 𝐵 1 2 𝐶 1 2 𝐴−𝐶 𝐴 1 2 𝐶 1 2 ≤ 𝐴−𝐵 𝐴 1 2 𝐵 1 2 𝐵−𝐶 𝐵 1 2 𝐶 1 2 𝐵 𝐴−𝐶 ≤ 𝐵−𝐴 𝐵−𝐶 Which follows from Claim 2.1.

Corollary 2.4 If 𝐴−𝐵 ≤𝐾 𝐴 1 2 𝐵 1 2 then 𝐴−𝐴 ≤ 𝐾 2 𝐴 . For 𝐵=−𝐴 we get that if 𝐴+𝐴 ≤𝐾 𝐴 then 𝐴−𝐴 ≤ 𝐾 2 𝐴 .

Proof of Corollary 2.4 If 𝐴−𝐵 ≤𝐾 𝐴 1 2 𝐵 1 2 then 𝐴−𝐴 ≤ 𝐾 2 𝐴 . 𝐴−𝐴 𝐴 = 𝑒 𝑑 𝐴,𝐴 ≤ 𝑒 𝑑 𝐴,𝐵 +𝑑 𝐵,𝐴 = 𝐴−𝐵 𝐴 1 2 𝐵 1 2 2 ≤ 𝐾 2 𝑑 𝐴,𝐵 = log 𝐴−𝐵 𝐴 1 2 𝐵 1 2

The span of sets of small doubling

The doubling constant The doubling constant of 𝐴 is 𝐾= 𝐴+𝐴 𝐴 . We saw that 𝐾=1 corresponds to subgroups. Can we use 𝐾 to say something about the size of the smallest subgroup containing 𝐴? (or, in the case of 𝐺= 𝔽 𝑛 , the size of the span of 𝐴)

Lemma 2.5 (Laba) If 𝐴−𝐴 < 3 2 𝐴 then 𝐴−𝐴 is a subgroup. The constant 3/2 is tight: take 𝐴= 0,1 ⊆ℤ. 𝐴−𝐴= −1,0,1 is not a subgroup.

Proof of Lemma 2.5 We will show that for any 𝑥∈𝐴−𝐴, 𝐴∩ 𝐴+𝑥 > 𝐴 /2. This implies that for any 𝑥,𝑦∈𝐴−𝐴, 𝐴+𝑥 ∩ 𝐴+𝑦 ≠∅: Denoting 𝐴 𝑥 =𝐴∩ 𝐴+𝑥 , by the inclusion exclusion principle 𝐴+𝑥 ∩ 𝐴+𝑦 ≥ 𝐴 𝑥 ∩ 𝐴 𝑦 = 𝐴 𝑥 + 𝐴 𝑦 − 𝐴 𝑥 ∪ A 𝑦 > 𝐴 /2+ 𝐴 /2− 𝐴 =0 There are 𝑎 1 , 𝑎 2 ∈𝐴 s.t. 𝑎 1 +𝑥= 𝑎 2 +𝑦, thus 𝑥−𝑦= 𝑎 2 − 𝑎 1 ∈𝐴−𝐴. 𝐴−𝐴 is closed under taking difference, hence must be a subgroup.

Proof of Lemma 2.5 cont. We will show that for any 𝑥∈𝐴−𝐴, 𝐴∩ 𝐴+𝑥 > 𝐴 /2. Let 𝑥=𝑎− 𝑎 ′ ∈𝐴−𝐴. Then 𝐴∩ 𝐴+𝑥 = 𝐴−𝑎 ∩ 𝐴− 𝑎 ′ . Similarly to before: 𝐴−𝑎 ∩ 𝐴− 𝑎 ′ = 𝐴−𝑎 + 𝐴− 𝑎 ′ − 𝐴−𝑎 ∪ 𝐴− 𝑎 ′ ≥ 𝐴 + 𝐴 − 𝐴−𝐴 >|𝐴|/2 As needed.

Lemma 2.6 (Freiman) Let 𝐴⊆ ℝ 𝑛 with 𝐴+𝐴 ≤𝐾 𝐴 . Then 𝐴 lies in an affine subspace of dimension at most 2𝐾−1.

Proof of Lemma 2.6 We will prove that if 𝐴 has affine dimension 𝑑 then 𝐴+𝐴 ≥ 𝑑+1 𝐴 − 𝑑+1 2 Thus 𝑑+1 𝐴 − 𝑑+1 2 ≤ 𝐴+𝐴 ≤𝐾 𝐴 , and since 𝐴 ≥𝑑, 𝑑+1−𝐾 𝑑≤ 𝑑+1−𝐾 𝐴 ≤ 𝑑+1 2 = 𝑑+1 2 𝑑 Hence 𝐾≥ 𝑑+1 2 , or 𝑑≤2𝐾−1.

Proof of Lemma 2.6 cont. We prove by induction on 𝐴 that if 𝐴 has affine dimension 𝑑 then 𝐴+𝐴 ≥ 𝑑+1 𝐴 − 𝑑+1 2 If 𝐴 =1 then 𝐴+𝐴 =1, 𝑑=0 and the claim follows. Assume 𝐴 >1. Let 𝑀 𝐴 = 𝑎+ 𝑎 ′ 2 𝑎, 𝑎 ′ ∈𝐴 . Clearly 𝑀 𝐴 = 𝐴+𝐴 . Let 𝐶(𝐴) be the convex hull of 𝐴, which is a 𝑑-dimensional polytope by assumption. Its vertices are the points in 𝐴 that cannot be obtained as a convex combination of the other points.

Inductive proof, case (1) Fix a vertex 𝑎 ∗ ∈𝐴 of 𝐶 𝐴 , and let 𝐴 ′ =𝐴∖ 𝑎 ∗ . We have two cases: The affine dimension of 𝐴 ′ is 𝑑−1. Then the mid-points 𝑎+ 𝑎 ∗ 2 for all 𝑎∈𝐴 are outside 𝐶( 𝐴 ′ ). 𝑀 𝐴 ≥ 𝐴 + 𝑀 𝐴 ′ ≥ 𝐴 +𝑑 𝐴 ′ − 𝑑 2 = 𝑑+1 𝐴 −𝑑− 𝑑 2 = 𝑑+1 𝐴 − 𝑑+1 2 𝐴+𝐴 ≥ ? 𝑑+1 𝐴 − 𝑑+1 2

Inductive proof, case (2) Fix a vertex 𝑎 ∗ ∈𝐴 of 𝐶 𝐴 , and let 𝐴 ′ =𝐴∖ 𝑎 ∗ . We have two cases: The affine dimension of 𝐴 ′ is 𝑑. Let 𝑎 1 ,…, 𝑎 𝑡 be the neighboring vertices to 𝑎 ∗ in 𝐶 𝐴 (two vertices are neighbors if the segment connecting them is a one-dim. face of 𝐶 𝐴 ). Note that 𝑡≥𝑑 since 𝐶 𝐴 is 𝑑-dim. The points 𝑎 ∗ and 𝑎 ∗ + 𝑎 𝑖 2 for 1≤𝑖≤𝑡 are mid-points outside 𝐶 𝐴 ′ . 𝑀 𝐴 ≥ 𝑡+1 + 𝑀 𝐴 ′ ≥𝑑+1+ 𝑑+1 𝐴 ′ − 𝑑+1 2 = 𝑑+1 𝐴 − 𝑑+1 2 𝐴+𝐴 ≥ ? 𝑑+1 𝐴 − 𝑑+1 2

Lemma 2.6 – tightness Let 𝐴⊆ ℝ 𝑛 with 𝐴+𝐴 ≤𝐾 𝐴 . Then 𝐴 lies in an affine subspace of dimension at most 2𝐾−1. The dimension is tight up to lower order terms. Take 𝐴= 𝑣 1 ,…, 𝑣 2𝐾 a set of linearly independent vectors, then 𝐴+𝐴 = 2𝐾 2 +2𝐾 =𝐾(2𝐾+1) and 𝐴+𝐴 𝐴 =𝐾+ 1 2 .

Theorem 2.7 (Ruzsa) Let 𝐺 be an Abelian group of torsion 𝑟, 𝐴⊆𝐺 with 𝐴+𝐴 ≤𝐾 𝐴 . Then there exists a subgroup 𝐻<𝐺, 𝐻 ≤ 𝐾 2 𝑟 𝐾 4 𝐴 s.t. 𝐴⊆𝐻. A group has torsion 𝑟≥1 if 𝑟∙𝑔=0 for all 𝑔∈𝐺.

Example 𝐺= 𝔽 𝑝 𝑛 has torsion 𝑟=𝑝. Let 𝑈,𝑉⊆ 𝔽 𝑝 𝑛 two subspaces with 𝑈∩𝑉= 0 , and set 𝐴=𝑈+{ 𝑣 1 ,…, 𝑣 2𝐾 } where 𝑣 1 ,…, 𝑣 2𝐾 ∈𝑉 are linearly independent. 𝐴 = 𝑈 ∙2𝐾. Since 𝐴+𝐴=𝑈+ 𝑣 𝑖 + 𝑣 𝑗 1≤𝑖≤𝑗≤2𝐾 , 𝐴+𝐴 = 𝑈 ∙𝐾 2𝐾+1 and 𝐴+𝐴 𝐴 =𝐾+ 1 2 ≈𝐾. However, the size of the minimal subspace containing 𝐴 is 𝑝 2𝐾 𝑈 = 𝑝 2𝐾 2𝐾 𝐴 Thus an exponential dependency on 𝐾 is unavoidable.

Conjecture 2.8 (Ruzsa) There exists an absolute constant 𝐶≥2 s.t. for any Abelian group 𝐺 of torsion 𝑟, and any 𝐴⊆𝐺 with 𝐴+𝐴 ≤𝐾|𝐴|, the subgroup generated by 𝐴 has order ≤ 𝑟 𝐶𝐾 𝐴 . Was established with 𝐶=2 for 𝑟=2 and then extended for prime 𝑟: Theorem 2.9: Let 𝑝 be a prime, 𝐴⊆ 𝔽 𝑝 𝑛 with 𝐴+𝐴 ≤𝐾 𝐴 . Then there exists a subspace 𝐻⊆ 𝔽 𝑝 𝑛 s.t. 𝐴⊆𝐻 and 𝐻 ≤ 𝑝 2𝐾 2𝐾−1 𝐴 .

Theorem 2.10 If 2𝐴 ≤𝐾 𝐴 then 2𝐴−2𝐴 ≤ 𝐾 4 𝐴 . Used to prove Theorem 2.7. We will prove a more general result later today.

Proof of Theorem 2.7 - Sketch Let 𝐴⊆𝐺 with 2𝐴 ≤𝐾 𝐴 . We can assume 0∈𝐴 by possibly replacing 𝐴 with 𝐴−𝑎 for some 𝑎∈𝐴. Let 𝐵= 𝑏 1 ,…, 𝑏 𝑚 ⊆2𝐴−𝐴 be a maximal collection of elements s.t. 𝑏 𝑖 −𝐴 are all disjoint (“Ruzsa covering”). 𝐴 has small doubling, thus Theorem 2.10 implies that 𝐵 is bounded. We will show that ℓ𝐴−𝐴⊆ ℓ−1 𝐵+𝐴−𝐴 for all ℓ≥1. Together the theorem follows.

Proof of Theorem 2.7 cont. Since 𝑏 𝑖 −𝐴⊆2𝐴−2𝐴, 𝑚≤ 2𝐴−2𝐴 𝐴 ≤ 𝐾 4 𝑚≤ 2𝐴−2𝐴 𝐴 ≤ 𝐾 4 Therefore 𝐵 ≤ 𝐾 4 . 2𝐴 ≤𝐾 𝐴 ⇒ 2𝐴−2𝐴 ≤ 𝐾 4 𝐴

ℓ𝐴−𝐴⊆ ℓ−1 𝐵+𝐴−𝐴 ℓ=1 is trivial. For ℓ=2, we need to show 2𝐴−𝐴⊆𝐵+𝐴−𝐴. Take 𝑥∈2𝐴−𝐴, by construction there exists 𝑏∈𝐵 s.t. 𝑥−𝐴 ∩ 𝑏−𝐴 ≠∅, i.e. 𝑥−𝑎=𝑏− 𝑎 ′ for some 𝑎, 𝑎 ′ ∈𝐴. Hence 𝑥=𝑏+𝑎− 𝑎 ′ ∈𝐵+𝐴−𝐴. By induction on ℓ, ℓ𝐴−𝐴=𝐴+ ℓ−1 𝐴−𝐴 ⊆𝐴+ ℓ−2 𝐵+𝐴−𝐴 ⊆ ℓ−2 𝐵+𝐵+𝐴−𝐴= ℓ−1 𝐵+𝐴−𝐴

Concluding the proof ℓ𝐴−𝐴⊆ ℓ−1 𝐵+𝐴−𝐴 Let 𝐴 be the subgroup spanned by 𝐴, and similarly for 𝐵. 𝐴 = ℓ≥1 ℓ𝐴 ⊆ ℓ≥1 ℓ𝐴−𝐴 ⊆ ℓ≥1 ℓ−1 𝐵+𝐴−𝐴 = 𝐵 +𝐴−𝐴 Hence, using Corollary 2.4 (which implies 𝐴−𝐴 ≤ 𝐾 2 𝐴 ), 𝐴 ≤ 𝐵 ∙ 𝐴−𝐴 ≤ 𝐵 ∙ 𝐾 2 𝐴 We conclude by bounding 𝐵 . As 𝐺 has torsion 𝑟 and we saw that 𝐵 ≤ 𝐾 4 we have 𝐵 ≤ 𝑟 𝐾 4 . Finally, 𝐴 ≤ 𝐾 2 𝑟 𝐾 4 𝐴 .

The growth of sets of small doubling

Iterated sumsets ℓ𝐴= 𝑎 1 +…+ 𝑎 ℓ 𝑎 1 ,…, 𝑎 ℓ ∈𝐴 ℓ𝐴= 𝑎 1 +…+ 𝑎 ℓ 𝑎 1 ,…, 𝑎 ℓ ∈𝐴 ℓ𝐴−𝑚𝐴= 𝑎 1 +…+ 𝑎 ℓ − 𝑎 ℓ+1 −…− 𝑎 ℓ+𝑚 𝑎 1 ,…, 𝑎 ℓ+𝑚 ∈𝐴 Theorem 2.11 (Plünneke-Ruzsa): If 𝐴 = 𝐵 and 𝐴+𝐵 ≤𝐾 𝐴 then ℓ𝐴−𝑚𝐴 ≤ 𝐾 ℓ+𝑚 𝐴 . For 𝐵=𝐴 or 𝐵=−𝐴 we obtain the following: Corollary 2.12: If 𝐴+𝐴 ≤𝐾 𝐴 or 𝐴−𝐴 ≤𝐾 𝐴 then ℓ𝐴−𝑚𝐴 ≤ 𝐾 ℓ+𝑚 𝐴 .

Lemma 2.13 Let 𝐴,𝐵⊆𝐺 with 𝐴 = 𝐵 and 𝐴+𝐵 ≤𝐾 𝐴 . Let 𝐵 0 ⊆𝐵 a nonempty set minimizing the ratio 𝐾 0 = 𝐴+ 𝐵 0 𝐵 0 . Then for any 𝐶⊆𝐺, 𝐴+ 𝐵 0 +𝐶 ≤ 𝐾 0 𝐵 0 +𝐶 Notice that 𝐾 0 ≤𝐾.

Proof of Theorem 2.11 ∀𝐶 𝐴+ 𝐵 0 +𝐶 ≤ 𝐾 0 𝐵 0 +𝐶 ∀𝐶 𝐴+ 𝐵 0 +𝐶 ≤ 𝐾 0 𝐵 0 +𝐶 Let 𝐵 0 ⊆𝐵 as in Lemma 2.13. We will prove by induction on ℓ that 𝐵 0 +ℓ𝐴 ≤ 𝐾 0 ℓ 𝐵 0 ≤ 𝐾 ℓ 𝐵 0 . ℓ=0 is easy: 𝐵 0 +0𝐴 =1∙ 𝐵 0 . For ℓ>0: 𝐵 0 +ℓ𝐴 = 𝐴+ 𝐵 0 +(ℓ−1)𝐴 ≤ 𝐾 0 𝐵 0 + ℓ−1 𝐴 ≤ 𝐾 0 ∙ 𝐾 0 ℓ−1 𝐵 0 = 𝐾 0 ℓ 𝐵 0 By the Ruzsa triangle inequality (Claim 2.1), applied to − 𝐵 0 ,ℓ𝐴,𝑚𝐴: −𝐵 0 ℓ𝐴−𝑚𝐴 ≤ 𝐵 0 +ℓ𝐴 𝐵 0 +𝑚𝐴 ≤ 𝐾 ℓ+𝑚 𝐵 0 2 Hence ℓ𝐴−𝑚𝐴 ≤ 𝐾 ℓ+𝑚 𝐵 0 ≤ 𝐾 ℓ+𝑚 𝐴 𝐴 𝐵−𝐶 ≤ 𝐴−𝐵 𝐴−𝐶

Lemma 2.13 (Reminder) Let 𝐴,𝐵⊆𝐺 with 𝐴 = 𝐵 and 𝐴+𝐵 ≤𝐾 𝐴 . Let 𝐵 0 ⊆𝐵 a nonempty set minimizing the ratio 𝐾 0 = 𝐴+ 𝐵 0 𝐵 0 . Then for any 𝐶⊆𝐺, 𝐴+ 𝐵 0 +𝐶 ≤ 𝐾 0 𝐵 0 +𝐶

Proof of Lemma 2.13 Observe that by minimality of 𝐾 0 , for any 𝐵 ′ ⊆𝐵 we have that 𝐴+ 𝐵 ′ ≥ 𝐾 0 𝐵 ′ . The proof is by induction on 𝐶 . If 𝐶= 𝑥 , 𝐴+ 𝐵 0 +𝑥 = 𝐴+ 𝐵 0 = 𝐾 0 𝐵 0 = 𝐾 0 𝐵 0 +𝑥 If 𝐶 >1, write 𝐶= 𝐶 ′ ∪ 𝑥 . Define 𝐵 ′ ⊆ 𝐵 0 as 𝐵 ′ = 𝑏∈ 𝐵 0 𝐴+𝑏+𝑥⊆𝐴+ 𝐵 0 + 𝐶 ′ . We can write 𝐴+ 𝐵 0 +𝐶= 𝐴+ 𝐵 0 + 𝐶 ′ ∪ 𝐴+ 𝐵 0 +𝑥 ∖ 𝐴+ 𝐵 ′ +𝑥

Proof of Lemma 2.13 cont. 𝐴+ 𝐵 0 +𝐶= 𝐴+ 𝐵 0 + 𝐶 ′ ∪ 𝐴+ 𝐵 0 +𝑥 ∖ 𝐴+ 𝐵 ′ +𝑥 𝐴+ 𝐵 0 +𝐶 ≤ 𝐴+ 𝐵 0 + 𝐶 ′ + 𝐴+ 𝐵 0 +𝑥 ∖ 𝐴+ 𝐵 ′ +𝑥 = 𝐴+ 𝐵 0 + 𝐶 ′ + 𝐴+ 𝐵 0 +𝑥 − 𝐴+ 𝐵 ′ +𝑥 ≤ 𝐾 0 𝐵 0 + 𝐶 ′ + 𝐴+ 𝐵 0 − 𝐴+ 𝐵 ′ ≤ 𝐾 0 𝐵 0 + 𝐶 ′ + 𝐾 0 𝐵 0 − 𝐾 0 𝐵 ′ = 𝐾 0 𝐵 0 + 𝐶 ′ + 𝐵 0 − 𝐵 ′ It remains to show that: 𝐵 0 + 𝐶 ′ + 𝐵 0 − 𝐵 ′ ≤ 𝐵 0 +𝐶 ∀ 𝐵 ′ ⊆𝐵 𝐴+ 𝐵 ′ ≥ 𝐾 0 𝐵 ′

Concluding the proof It remains to show that: 𝐵 ′ = 𝑏∈ 𝐵 0 𝐴+𝑏+𝑥⊆𝐴+ 𝐵 0 + 𝐶 ′ It remains to show that: 𝐵 0 + 𝐶 ′ + 𝐵 0 − 𝐵 ′ ≤ 𝐵 0 +𝐶 Define 𝐵 ′′ = 𝑏∈ 𝐵 0 𝑏+𝑥∈ 𝐵 0 + 𝐶 ′ . Note that 𝐵 ′′ ⊆ 𝐵 ′ . We decompose 𝐵 0 +𝐶 as a disjoint union: 𝐵 0 +𝐶= 𝐵 0 + 𝐶 ′ ∪ 𝐵 0 +𝑥 ∖ 𝐵 ′′ +𝑥 Thus 𝐵 0 +𝐶 = 𝐵 0 + 𝐶 ′ + 𝐵 0 − 𝐵 ′′ ≥ 𝐵 0 + 𝐶 ′ + 𝐵 0 − 𝐵 ′ Which is what we needed to show.