Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated


Use the notation and sunflower convention in the statement of L268. Distinct words give distinct transversal sets. Their sets form a sunflower exactly when every coordinate's symbols are either all equal or all distinct: if a symbol appears in two words, the corresponding point must belong to the common core and hence appear in all the words.

Linear independent representatives

Lemma. Let E1,…,ENE_1,\ldots,E_N be subspaces of a finite-dimensional vector space WW. There are linearly independent vectors vi∈Eiv_i\in E_i, one for each ii, if and only if

dim⁡(∑i∈IEi)≥∣I∣for every I⊆[N].(1)\dim\Bigl(\sum_{i\in I}E_i\Bigr)\ge |I| \quad\text{for every }I\subseteq[N]. \tag{1}

Proof. Necessity follows by counting the selected independent vectors in each sum. For sufficiency, induct on NN. The cases N=0,1N=0,1 are immediate. If a nonempty proper set II attains equality in (1), apply induction to its subspaces. Their selected vectors form a basis of WI=∑i∈IEiW_I=\sum_{i\in I}E_i. In W/WIW/W_I, any collection indexed by J⊆[N]∖IJ\subseteq[N]\setminus I spans dimension at least

dim⁡(WI+WJ)−dim⁡WI≥∣I∣+∣J∣−∣I∣=∣J∣.\dim(W_I+W_J)-\dim W_I\ge |I|+|J|-|I|=|J|.

Induction in the quotient, followed by lifting the selected vectors, completes the choice.

Otherwise every nonempty proper II has sum dimension at least ∣I∣+1|I|+1. Choose a nonzero vN∈ENv_N\in E_N, and quotient by its span. For the first N−1N-1 subspaces, each nonempty sum loses at most one dimension and still satisfies (1). Select their representatives by induction in the quotient and lift them; together with vNv_N they are independent. □\square

By repeating each EiE_i exactly ss times, the lemma also gives ss independent representatives from each EiE_i, all jointly independent, whenever every collection of the original subspaces spans dimension at least ss times its cardinality. Indeed a collection of repeated indices has at most s∣I∣s|I| entries when its distinct indices are II.

Contracting an obstructing collection of coordinates

Write V=CV=C and t=dim⁡Vt=\dim V. Let Li:V→AiL_i:V\to A_i be its coordinate maps, and let

Ei=im⁡(Li∗)≤V∗.E_i=\operatorname{im}(L_i^*)\le V^*.

Their sum is V∗V^*: a vector annihilated by every EiE_i has every coordinate zero and is therefore zero. For I⊆[n]I\subseteq[n], put

WI=∑i∈IEi,d(I)=s∣I∣−dim⁡WI.W_I=\sum_{i\in I}E_i,\qquad d(I)=s|I|-\dim W_I.

Choose JJ maximizing d(I)d(I), and among maximizers choose one with least cardinality. Since d(∅)=0d(\varnothing)=0 and d([n])=sn−t≤0d([n])=sn-t\le0, this JJ is a proper subset of [n][n]. This includes the equality case t=snt=sn: if the maximum is zero, the empty set is the chosen maximizer; if it is positive, the full set is not a maximizer.

For every K⊆[n]∖JK\subseteq[n]\setminus J, maximality gives

dim⁡(WJ+WK)−dim⁡WJ≥s∣K∣.(2)\dim(W_J+W_K)-\dim W_J\ge s|K|. \tag{2}

Apply the lemma in V∗/WJV^*/W_J to the images of the EiE_i for i∉Ji\notin J. Choose vectors ei,1,…,ei,se_{i,1},\ldots,e_{i,s} in those images that are all linearly independent. Define a linear map

ϕ:V∗/WJ⟶Fqs\phi:V^*/W_J\longrightarrow\mathbb F_q^s

by sending ei,je_{i,j} to the jjth standard basis vector, for every i∉Ji\notin J, and extend it from these independent vectors to a basis of the quotient. This assignment is consistent even though different selected vectors have the same image. Since [n]∖J[n]\setminus J is nonempty, ϕ\phi is onto.

Compose ϕ\phi with the quotient map to obtain Φ:V∗→Fqs\Phi:V^*\to\mathbb F_q^s. Its dual embeds (Fqs)∗(\mathbb F_q^s)^* as an ss-dimensional subspace UU of V∗∗=VV^{**}=V. If i∈Ji\in J then Φ(Ei)=0\Phi(E_i)=0, so Li∣U=0L_i|_U=0. If i∉Ji\notin J then Φ(Ei)=Fqs\Phi(E_i)=\mathbb F_q^s, so the dual of Li∣UL_i|_U is onto and Li∣UL_i|_U is injective.

For any distinct words in a+Ua+U, a zero coordinate map gives one common symbol and an injective coordinate map gives distinct symbols. Thus all qsq^s transversal sets from a+Ua+U form a sunflower, including the case of an empty core. This proves the subspace assertion.

The exponential bound

Let k≥3k\ge3 and suppose the affine word family a+Ca+C has no kk-sunflower. With s=⌈log⁡qk⌉s=\lceil\log_q k\rceil, a dimension at least snsn would supply a qsq^s-sunflower by the subspace assertion. Any kk of its members would be a kk-sunflower, because qs≥kq^s\ge k. Consequently

dim⁡C<sn,∣C∣≤qsn−1.(3)\dim C<sn,\qquad |C|\le q^{sn-1}. \tag{3}

If q≥kq\ge k and C≠{0}C\ne\{0\}, choose a nonzero v∈Cv\in C. The affine line a+Fqva+\mathbb F_qv is a qq-sunflower: in coordinate ii it is constant when vi=0v_i=0, and its symbols are distinct when vi≠0v_i\ne0. This is a contradiction, so C={0}C=\{0\} in this case and ∣C∣=1<k2n|C|=1<k^{2n}.

If q<kq<k, the definition of ss gives qs<qk<k2q^s<qk<k^2. Equation (3) therefore implies ∣C∣<k2n|C|<k^{2n} in this case as well. The estimates do not depend on the dimensions of the coordinate spaces. □\square