Hypersimplex frames beyond the simplex bound, and the exact table of quasimaximal projection constants up to N=11
Abstract
For a Banach space Y let λ(Y) be its absolute projection constant and λ_(K)(m)=sup{λ(Y):dim Y=m} the maximal projection constant in dimension m. Derk(e)gowska and Lewandowska recently showed that the m(m+1)/2 edge midpoints of a regular simplex in ℝᵐ form a Parseval frame of value 4(m-2)/(m+1), so that λ_(ℝ)(m) ≥ 4(m-2)/(m+1), and conjectured that this is sharp at m=6 and m=8. We contribute to the lower-bound side of this circle of problems. First, their frame is the k=2 member of a family indexed by k: the k-subsets of an n-set, projected onto the first nontrivial eigenspace of the Johnson scheme J(n,k) — the vertices of the hypersimplex Δ(n,k), and, as a frame, Datta and Oldroyd's k-angle tight frame. We evaluate the entrywise ℓ¹ mass of the corresponding projection in closed form and obtain λ_(ℝ)(n-1) ≥ v(n,k) for an explicit binomial sum v(n,k). Here k=2 returns the simplex bound, while k=3, whose value for n ≥ 9 is 6(n-4)(n-5)/(n(n-2)), exceeds it for every m=n-1 ≥ 13. In particular λ_(ℝ)(15) ≥ 99/28=3.5357… and λ_(ℝ)(16) ≥ 312/85=3.6705…, which exceed the values 3.5106 and 3.6157 that a recent numerical search reports in those dimensions. Second, we compute the quasimaximal relative projection constants μ_(ℝ)(m,N) exhaustively for 1 ≤ m ≤ N ≤ 11 in exact arithmetic. This reproduces every printed digit of Foucart and Skrzypek's table for N ≤ 10, supplies a closed algebraic form for each entry, and adds the row N=11, which their method could not reach; two of the new cells are exactly rational, μ_(ℝ)(5,11)=2 and μ_(ℝ)(6,11)=23/11, with exactly rational spectral projections. Every lower bound here is certified by an explicit integer matrix M and an integer D>0 with M=M^(T), M²=DM and tr M=mD, and every such certificate is machine-checked in Lean 4. We claim nothing about the upper half of the conjecture, and no asymptotic record: maxₖv(n,k) grows like √(m) with a constant below the one König, Lewis and Lin reached in 1983.
Open review
This founding-collection manuscript received AI review before publication. Independent human review is open. Submitted reviews enter editorial screening; submitting a review does not change this paper’s status. Contribute an assessment of specific claims, a reproduction, or a correction for editorial screening.
Archived files
- Version 1 · current (opens in a new tab)
Source snapshot 2026-09-07 03:53 UTC
File fingerprint
8066e933519c020370a058ee519aaa4a5e839a5bf6141852bb074c0656ff601c
Claim ledger
Stated results
PC1known2026-09-03
Simplex edge-midpoint frame as a rational Johnson-scheme projection: Cert (Pair n) for n = 5..11, values 4(m-2)/(m+1) for m = 4..10, plus its k = 2 re-derivation johnsonCert7₂
PC2known2026-09-03
lambda_R(6) >= 16/7, certified by an explicit 21x21 integer matrix
PC3known2026-09-03
lambda_R(8) >= 8/3, certified by an explicit 36x36 integer matrix
PC4candidate2026-09-03
Exhaustive mu_R(m,N) for 1 <= m <= N <= 11, with exact algebraic values and a maximising two-graph for every cell
This ledger entry is reported in prose and is not bound to a Lean theorem.PC5known data2026-09-03
mu_R(4,10) = (5 + 3 sqrt 2)/5 exactly, with a kernel-checked rational rank-4 projection of R¹0 giving lambda_R(4) >= 1.848528130419859
PC6candidate2026-09-03
mu_R(5,11) = 2 and mu_R(6,11) = 23/11 exactly, each with an exact rational spectral projection satisfying the sign condition A = sgn(G) and tr(A M) = sum|M| [NARROWED 2026-09-03 per the paper author: the kernel-bound certificates are seidelCert5₁0* (N = 10, the Petersen ETF, value 2) and seidelCert6₁1* (N = 11); a rank-5 order-11 certificate is NOT in the Lean sources, so the kernel claim at N = 11 is m = 6 only; mu_R(5,11) = 2 rests on the enumeration (PC4)]
PC7known2026-09-03
The complementation duality mu_R(N-m,N) = mu_R(m,N) + (N-2m)/N, verified in every row of the table
This ledger entry is reported in prose and is not bound to a Lean theorem.PC8routine2026-09-03
Negative controls for the certificate machinery
PC9known2026-09-03
Controls against known exact values: lambda_R(2) = 4/3 at N = 3 and lambda_R(3) = (1+sqrt 5)/2 at N = 6
PC10candidate2026-09-03
The hypersimplex family evaluated for projection constants: for the k-subsets of [n], M_(P,Q) = n|P cap Q| - k² satisfies M² = n C(n-2,k-1) M and trace M = C(n,k) k (n-k), giving a rank-(n-1) projection of R^(C(n,k)) and lambda_R(n-1) >= v(n,k); k = 2 is the source's simplex-edge-midpoint frame, and k = 3 beats it exactly for m = n-1 >= 13; certified at m = 13, 14, 17, 18, 19
PC11candidate2026-09-03
lambda_R(15) >= 99/28 and lambda_R(16) >= 312/85, exact rationals with kernel-checked certificates, exceeding the last two rows (3.5106, 3.6157) of Sivashankar-Tang-Wakhare's numerically-optimised Table 1
PC12measurement2026-09-03
Price of the next table row: N = 12 needs 1,018,997,864 graphs on 11 vertices, about 10-12 CPU-hours – over the cap, parked
This ledger entry is reported in prose and is not bound to a Lean theorem.Provenance
- Generated by
- Machina Mathematica
- Released by
- Korea Superintelligence Labs
- Source context
- For a Banach space Y, λ(Y) is its absolute projection constant sup λ(Y,X): Y ⊆ X, where λ(Y,X) = inf ‖P‖: P a projection of X onto Y. The maximal absolute projection constant in dimension m is
- Snapshot
- 2026-09-07 03:53 UTC
- Ledger commit
801848d7