Colouring lists for hereditary discrepancy
DefinitionDiscrepancy of a matrix on a set of columns#✓fully proved
For a real \(n \times m\) matrix \(A\) and a set \(S\) of columns of \(A\), the discrepancy of \(A\) on \(S\) is the minimum, over all sign vectors \(\varepsilon \in \{-1, 1\}^m\), of \(\max_i \left|\sum_{j \in S} A_{ij} \varepsilon_j\right|\), the largest absolute value of a row sum of \(A\) restricted to \(S\) and weighted by \(\varepsilon\) (taken to be \(0\) when \(A\) has no rows).
noncomputable def discrepancyOn {n m : ℕ} (A : Fin n → Fin m → ℝ)
(S : Finset (Fin m)) : ℝ :=
sInf ((fun ε : Fin m → Bool =>
⨆ i : Fin n, |∑ j ∈ S, A i j * (if ε j then (1 : ℝ) else -1)|) ''
Set.univ)
Used by 1
DefinitionHereditary discrepancy of a real matrix#✓fully proved
The hereditary discrepancy \(\operatorname{herdisc}(A)\) of a real \(n \times m\) matrix \(A\) is the maximum, over all sets \(S\) of columns of \(A\), of the discrepancy of \(A\) on \(S\).
noncomputable def hereditaryDisc {n m : ℕ} (A : Fin n → Fin m → ℝ) : ℝ :=
⨆ S : Finset (Fin m), discrepancyOn A S
noncomputable def discrepancyOn {n m : ℕ} (A : Fin n → Fin m → ℝ)
(S : Finset (Fin m)) : ℝ :=
sInf ((fun ε : Fin m → Bool =>
⨆ i : Fin n, |∑ j ∈ S, A i j * (if ε j then (1 : ℝ) else -1)|) ''
Set.univ)
WantedExt/Combinatorics/PolysizeColoringListWanted.lean:34
Uses 1 · Used by 1
DefinitionPolynomial-size colouring list conjecture#✓fully proved
The statement that there exist a real number \(C > 0\) and a natural number \(k\) such that for all positive integers \(n\) and \(m\) and every real \(n \times m\) matrix \(A\) there is a set \(L\) of at most \(m^k\) sign vectors in \(\{-1, 1\}^m\) with the following property: for every set \(S\) of columns of \(A\) some \(\varepsilon \in L\) satisfies \(\left|\sum_{j \in S} A_{ij} \varepsilon_j\right| \le C \operatorname{herdisc}(A)\) for every row \(i\), where \(\operatorname{herdisc}(A)\) is the hereditary discrepancy of \(A\).
def conjecture : Prop :=
∃ C : ℝ, 0 < C ∧ ∃ k : ℕ,
∀ (n m : ℕ), 0 < n → 0 < m →
∀ A : Fin n → Fin m → ℝ,
∃ L : Finset (Fin m → ℝ),
(L.card : ℝ) ≤ (m : ℝ) ^ k ∧
(∀ ε ∈ L, ∀ j, ε j = 1 ∨ ε j = -1) ∧
∀ S : Finset (Fin m),
∃ ε ∈ L, ∀ i, |∑ j ∈ S, A i j * ε j| ≤
C * hereditaryDisc A
noncomputable def hereditaryDisc {n m : ℕ} (A : Fin n → Fin m → ℝ) : ℝ :=
⨆ S : Finset (Fin m), discrepancyOn A S
WantedExt/Combinatorics/PolysizeColoringListWanted.lean:42
Sources
- AIM-COMBINATORICS-0047__20000922; AIM Hereditary discrepancy, item 1.85.
Uses 1 · Used by 1
TheoremThe polynomial-size colouring list conjecture is false#●ready to prove
There do not exist a real number \(C > 0\) and a natural number \(k\) such that for all positive integers \(n\) and \(m\) and every real \(n \times m\) matrix \(A\) there is a set \(L\) of at most \(m^k\) sign vectors in \(\{-1, 1\}^m\) with the following property: for every set \(S\) of columns of \(A\) some \(\varepsilon \in L\) satisfies \(\left|\sum_{j \in S} A_{ij} \varepsilon_j\right| \le C \operatorname{herdisc}(A)\) for every row \(i\), where \(\operatorname{herdisc}(A)\) is the hereditary discrepancy of \(A\).
The AIM problem list records this negative answer: for the \(0/1\) matrix with \(m\) columns whose \(\sqrt{m}\) rows are the indicators of disjoint sets of \(\sqrt{m}\) columns, the hereditary discrepancy is \(1\), yet for fixed \(C\) and lists of polynomial size some set of columns has discrepancy above \(C\) under every listed colouring.
public theorem_wanted conjecture_refuted : ¬ conjecture
def conjecture : Prop :=
∃ C : ℝ, 0 < C ∧ ∃ k : ℕ,
∀ (n m : ℕ), 0 < n → 0 < m →
∀ A : Fin n → Fin m → ℝ,
∃ L : Finset (Fin m → ℝ),
(L.card : ℝ) ≤ (m : ℝ) ^ k ∧
(∀ ε ∈ L, ∀ j, ε j = 1 ∨ ε j = -1) ∧
∀ S : Finset (Fin m),
∃ ε ∈ L, ∀ i, |∑ j ∈ S, A i j * ε j| ≤
C * hereditaryDisc A
WantedExt/Combinatorics/PolysizeColoringListWanted.lean:51
Sources
- AIM Problem Lists, Hereditary discrepancy and factorization norms, Problem 1.85 (Polysize list of colorings), https://aimpl.org/hereddiscrep/1/.
Uses 1
Loading dependency explorer…
The dependency explorer needs JavaScript and its graph data.