Answers are grounded in this paper and cite their sources.
Reveals new inequalities in linear-ordering polytopes, suggesting structural limits for models.
For the Bradley-Terry (BT) sigmoid-tilt family on K_n, the tilted edge-marginal vector p~(u,s) = (sigma(u_i - u_j + s_ij ln 2))_ij is the natural carrier of the base-2 moment-generating functional Phi_s(u) = log2 EP*(u)[2s.X]. We give a constructive, exact-arithmetic closure of the small-n theory of this functional. The analytic engine is a Bregman identity for the independent BT exponential family, Phi_s = s.p~ - (1/ln2) KL_indBT(p~ || p*), equivalently the tilt-path-average identity Phi_s(u) = int_0^1 s.p~(u,ts) dt. Our main result is the KL-aware tied-block inequality: for every facet a.x <= b of the linear ordering polytope P^LO_n and every ordered set partition Pi of [n], T(a,Pi) - c * #inblock_supp(Pi) <= b, where T(a,Pi) is the ranking-limit value of a.p~ along Pi and c = (1/ln2) KL(2/3 || 1/2) = 5/3 - log2 3 is the single facet-independent KL penalty per in-block support edge (box-corner identity sigma(+-ln2) in {1/3, 2/3}). Via an exact ranking-limit reduction, sup_u Phi_a(u) = max_Pi [T(a,Pi) - c #inblock_supp(Pi)], so this family of inequalities is exactly the finite combinatorial content of the maximum principle sup_u Phi_a <= b on {+-1}^E. For cyclic directions the bound rests on Phi_a having no nonconstant finite local maximum: proven for acyclic a, the constant maxima handled analytically, and the nonconstant case certified numerically, exhaustively over all 396 isomorphism classes of strong tournaments at n<=7 (and sampled strong tournaments at n=8,9). Every genuine nonconstant finite critical point so tested is a strict saddle, reducing the open question to the existence of a block-constant ascent direction -- the ordered four-level quotient-saddle conjecture Lambda_4 > 0. The three-level form is refuted by a machine-certified finite counterexample on K_163, which is nonetheless a strict saddle; no spectral-average certificate can witness this saddle, and single-edge localisation (sufficient for n<=7) is itself false for n>=15. The inequality is unconditional for n<=6 (exact-arithmetic enumeration over all 910 facets of P^LO_6 and all 4683 ordered set partitions) and holds at n=7 over the 925 enumerated facets and 47,293 ordered partitions, conditional on the completeness of that classification. This yields a constructive closure of polytope membership p~(u,s) in P^LO_n at every n<=7, with Paley(7) on the boundary. The closure cannot be uniform in n: the cube [1/3,2/3]^E leaves P^LO_n at n=11 (the first exhibited failure; heavy search finds no +-1-direction witness at n=8,9,10), and the uniform-in-n form is false (the explicit large-n refutation is the companion spectral paper). Finally we prove a Theta(n^2) triangle-LP upper bound valid for all n: sup_u Phi_tau <= h_TLP(tau), with the closed form h_TLP(tau) = 1/2 sum tau + 1/6 C(n,2) for doubly regular (in particular Paley) directions; the doubly-regular hypothesis is essential (the transitive direction has h_TLP = C(n,2)).
No takes yet. Share an insight, caveat, or question.
Vladimir Riabov (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: