Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Dyadic blocks, doubling chains and a weighted pigeonhole
This proves
L213.
Write d(X) for the largest size of a dissociated subset of a finite set
X of reals.
Step 1: symmetric reduction and blocks
Let P={∣x∣:x∈A,x=0} and N=∣P∣. Changing individual signs
preserves dissociation by reversing the corresponding ternary coefficients,
and a dissociated set contains neither
0 nor two opposite elements, so d(P)=d(A); also N≥(n−1)/2≥2 for
n≥5. Let m=minP and, for integers k≥0,
Pk=P∩[2km,2k+1m).
These blocks partition P, and each satisfies maxPk<2minPk. Let
k1<⋯<kB be the indices of the nonempty blocks and ni=∣Pki∣,
so ∑ini=N.
Step 2: one block and a doubling chain
Fix i. By
L212
and d≥1, the block Pki has a dissociated subset Di of size
τi=max(1,⌊log2ni−log23⌋+1)>log2ni−log23.
Since τi≤max(1,log2N−log23+1)≤log2N, we have
⌈log2τi⌉≤λ:=⌈log2log2N⌉.
Doubling chains. If D⊆P is dissociated and
z1<⋯<zr in P∖D satisfy z1>∑D and
zj+1>2zj, then D∪{z1,…,zr} is dissociated. Indeed, by
induction ∑D+∑j<tzj<zt for every t. In a ternary relation
on the union, let t be the largest index with a nonzero coefficient on
zt; then zt is at most the sum of all the other elements, which is
impossible. So no zj occurs, and the relation lies in D, where it is
trivial.
List in increasing order the nonempty blocks with index at least
ki+1+⌈log2τi⌉, as κ1<κ2<⋯, and take
one element zj from block κ2j−1 for each j. Every element
of Di is below 2ki+1m, so
∑Di<τi2ki+1m≤2κ1m≤z1, and
zj+1≥2κ2j+1m≥2κ2j−1+2m>2zj. Among the
B−i nonempty blocks after ki, at most ⌈log2τi⌉≤λ
have an index in (ki,ki+⌈log2τi⌉], so the chain has at
least 21(B−i−λ)+ elements. Hence, for every i,
d(P)≥log2ni−log23+21(B−i−λ)+.(1)
Step 3: weighted pigeonhole
Put wi=2−(B−i−λ)+/2. At most λ+1 indices have
B−i−λ≤0, and the other indices give distinct exponents
t=B−i−λ≥1, so
W=i=1∑Bwi≤λ+1+t≥1∑2−t/2=λ+2+2.
If ni<Nwi/W held for every i, summing would give N<N. So some i
has log2ni≥log2N−log2W−21(B−i−λ)+, and (1) gives
d(A)=d(P)≥log2N−log2(λ+2+2)−log23.
Finally log2N≥log2(n−1)−1 and
λ≤⌈log2log2n⌉, which is the displayed bound of the
statement.
Step 4: the simplified form
Let n≥216 and y=log2log2n≥4. Since 2<3/2,
we have ⌈y⌉+2+2<y+9/2≤(9/4)y.
Since n≥16 also gives n−1≥15n/16, this yields
The argument uses one block's dissociated subset and single elements from
other blocks. It never combines dissociated subsets from two blocks within
a factor of order log2n of each other, and that is the whole source of
the log2log2log2n loss: a set spread with equal counts over about
log2log2n consecutive dyadic blocks is the case it handles worst.