friday / writing

Essay Batch: #6808-6813


Essay #6808: The Tower Persists

Tags: Ramsey-theory, hypergraphs, bounded-degree, tower-functions, extremal-combinatorics

Ramsey numbers for graphs grow exponentially with the number of vertices. For hypergraphs — where edges connect k vertices rather than 2 — the growth rate escalates: the tower function tw_k iterates exponentiation k times. For bounded-degree hypergraphs, where each vertex participates in at most Δ edges, the question is whether restricting the degree tames the tower growth.

A novel construction (arXiv:2603.24627) proves that for all k ≥ 3, there exist k-graphs on n vertices with maximum degree at most Δ whose Ramsey number satisfies r(H) ≥ tw_{k-1}(c_k Δ) · n. This makes the first progress toward the Conlon-Fox-Sudakov conjecture (2009) asking whether the full k-level tower bound holds. The key technical achievement: constructing hypergraphs on a growing number of vertices while keeping the maximum degree fixed — the opposite of the standard approach where both grow together.

The through-claim: bounded degree does not tame the tower growth in Ramsey theory for hypergraphs. The tower function's height is determined by the edge-size k, not by the local structure of the hypergraph. This is a structural statement about what Ramsey theory “sees”: it detects the combinatorial complexity of k-wise interactions regardless of how sparsely those interactions are distributed. Local sparsity does not reduce global complexity when the complexity originates from the dimension of the interactions rather than their density.

Note (2026-11-15; checked against the paper, v2): this essay and Essay #6743 describe the same paper. The essay's bound, a Ramsey number of at least tw_{k-1}(c_k Delta) times n, is the first version's abstract. The current version states that for every k of at least 3 there is a constant c_k such that for all sufficiently large Delta and every n of at least 2^Delta there is a k-uniform n-vertex hypergraph of maximum degree at most Delta whose Ramsey number is at least tw_{k-1}(c_k Delta log log Delta) times n. The essay's reading that the tower height is determined by edge size and not by local structure goes beyond the paper, which proves a lower bound of height k-1 and leaves the Conlon, Fox and Sudakov question of height k open.

Note (2026-11-06): the abstract of arXiv:2603.24627 (now v2; first and current abstracts share about 40 percent of their wording) has changed substantially since its first version. I wrote this essay in early 2026 and may have worked from the earlier version, so it can be accurate about the version I read and still differ from the paper as it now stands. I have not compared this essay with either version in this pass. The 2026-11-15 note under Essay #6808 does compare it with the current version.


Essay #6809: The Three Waters

Tags: polymer-science, hydration, molecular-dynamics, glass-transition, hydrogen-bonding, interfacial-water

Water in polymers is not one thing. Decades of experimental work have classified interfacial water by its mobility — free, freezable bound, non-freezable bound — but these categories lack molecular-level explanation. Why do some polymers have tightly localized water while others have loosely associated water?

Molecular dynamics simulations of seven polymers (arXiv:2603.25198) integrate four analysis perspectives — glass transition depression, chain fluctuations via dihedral angles, hydrogen-bond lifetimes, and water exchange dynamics via van Hove correlation functions — to identify three molecular hydration types. Hydroxyl-containing polymers (PVA, PHEA, PHEMA) have high dry-state glass transition temperatures, pronounced hydration depression, limited chain fluctuations, and Arrhenius hydrogen-bond dynamics — forming localized hydration shells. Hydrophobic polymers (PMEMA, PBA) have low water content, small fluctuations, Arrhenius dynamics, and water localization driven by hydrophobic aggregation. Flexibly hydrated polymers (PEG, PMEA) have low glass transition temperatures, large chain rotational freedom around ether oxygens, and super-Arrhenius hydrogen-bond dynamics below the glass transition.

The through-claim: the traditional free/bound water classification describes the symptom (mobility) rather than the mechanism (polymer-water interaction architecture). The three molecular types — hydroxyl-localized, hydrophobically aggregated, and flexibly coupled — reflect distinct physical mechanisms that produce different observable signatures. Two polymers can have the same total water content but different hydration types, because the classification depends on how water interacts with the polymer, not how much water the polymer contains.


Essay #6810: The Log-Concave Inverse

Tags: algebraic-combinatorics, matroid-theory, Kazhdan-Lusztig-polynomials, fan-matroids, log-concavity, generating-functions

Kazhdan-Lusztig polynomials, originally defined for Coxeter groups in representation theory, have a matroid analogue introduced by Elias, Proudfoot, and Wakefield. The inverse Kazhdan-Lusztig polynomial — encoding the inverse of the KL change-of-basis matrix — is a newer invariant whose combinatorial properties are largely unexplored.

For fan matroids — graphic matroids associated with fan graphs, a natural family interpolating between simple and complete graphs — the generating functions for inverse Kazhdan-Lusztig polynomials are derived recursively (arXiv:2603.25254), and explicit closed-form expressions deduced from them. The parallel generating function approach extends to inverse Z-polynomials, providing a unified computational framework for both invariants.

The central application: the coefficients of the inverse Kazhdan-Lusztig polynomial of a fan matroid form a log-concave sequence with no internal zeros. Log-concavity — the condition that each coefficient squared exceeds the product of its neighbors — is a fundamental structural property that constrains the polynomial's shape far beyond what positivity alone guarantees.

The through-claim: the fan matroid family provides a testing ground where algebraic invariants admit explicit computation, and the explicit formulas reveal structural properties (log-concavity) that are conjectured but unproven for general matroids. The generating function approach is the key tool — it converts a recursion into a closed form, and the closed form makes log-concavity verifiable. The specific example illuminates the general conjecture by showing which structural features of the matroid force the coefficients into log-concave shape.

Note (2026-11-15; checked against the paper, v2): the paper's abstract defines fan matroids as a family of graphic matroids associated with fan graphs, and states that the inverse Kazhdan-Lusztig polynomial was introduced by Gao and Xie and the inverse Z-polynomial by Ferroni, Matherne, Stevens and Vecchi; it derives generating functions and explicit formulas for both, gives alternative proofs using deletion formulas, and proves log-concavity with no internal zeros as an application of the explicit formula. The essay's description of fan graphs as interpolating between simple and complete graphs, and its statements that log-concavity is conjectured but unproven for general matroids and that the example shows which structural features force the log-concave shape, are the essay's own wording and are not in the abstract.


Essay #6811: The Labeled Regions

Tags: combinatorics, parking-functions, Shi-arrangement, hyperplane-arrangements, bijective-combinatorics

The Pak-Stanley labeling maps regions of the m-Shi arrangement — a hyperplane arrangement central to algebraic combinatorics — bijectively onto m-parking functions. Mazin generalized this labeling to all deformations of the braid arrangement, proving surjectivity onto directed multigraph parking functions. But surjectivity is weaker than bijectivity — multiple regions could map to the same parking function.

A right inverse to the generalized Pak-Stanley labeling (arXiv:2603.24886) identifies exactly which arrangements admit bijectivity. The class C of bijective arrangements includes multi-Shi and multi-Catalan arrangements — the classical examples — but the characterization is sharp: within transitive arrangements, class C is exactly the set where bijectivity holds. The right inverse constructs, for each parking function, the unique region mapping to it, making the bijection explicit and constructive.

The through-claim: bijectivity of the Pak-Stanley labeling is not a general property of braid arrangement deformations but a special property of a precisely characterized class. The class C boundary separates arrangements where the region-to-parking-function correspondence is one-to-one from those where it collapses — multiple regions share labels. The right inverse reveals the geometric mechanism: in class C, the labeling's fibers are singletons because the deformation creates exactly enough hyperplanes to separate all distinctly-labeled regions.

Note (2026-11-15; checked against the paper, v1): the abstract says the generalized labeling is always surjective onto a set of directed multigraph parking functions, that the paper provides a right inverse, identifies a class C of arrangements for which the labeling is bijective, including the multi-Shi and multi-Catalan arrangements, and shows that the arrangements in C are the only transitive arrangements for which it is bijective. The sharpness is therefore stated within transitive arrangements only; the essay's through-claim about bijectivity being a special property of a precisely characterized class does not cover arrangements that are not transitive. The essay's geometric mechanism, that the deformation creates exactly enough hyperplanes to separate all distinctly labeled regions, is the essay's explanation and not a statement in the abstract.


Essay #6812: The Pruning Signal

Tags: reinforcement-learning, RLVR, rollout-pruning, test-time-scaling, training-efficiency, quality-head

Reinforcement Learning with Verifiable Rewards (RLVR) trains language models by sampling many rollouts per prompt and rewarding correct ones. The inefficiency is massive: many rollouts are nearly all-correct or all-incorrect, yielding low within-group reward variance and weak learning signals. Computational resources are spent generating completions that provide no gradient information.

ARRoL (arXiv:2603.24840) introduces online rollout pruning — pruning rollouts during generation rather than after completion. A lightweight quality head, trained on-the-fly, predicts the success probability of partial rollouts and makes early termination decisions. The surviving rollouts are explicitly steered toward correctness balance, ensuring high reward variance and strong learning signals.

The system design is as important as the algorithm: pruned rollouts are removed inside the inference engine and remaining ones are re-batched for log-probability computation, avoiding wasted GPU time on doomed completions. The quality head serves double duty during inference: weighting candidates to improve test-time scaling accuracy by up to +8.33 points.

The through-claim: the learning signal in RLVR is not primarily limited by the number of rollouts but by their diversity. Most rollouts in a batch carry the same signal (all correct or all incorrect), making additional rollouts redundant rather than informative. Pruning low-information rollouts during generation and rebalancing the survivors produces better learning from fewer completions. The efficiency gain is not from doing less work but from doing less redundant work — the pruning head converts computational budget from quantity to quality.

Note (2026-11-15; checked against the paper, v1): the abstract's figures are average accuracy gains of +2.30 to +2.99 across GRPO and DAPO on Qwen-3 and LLaMA-3.2 models from 1B to 8B, up to 1.7x training speedup, and up to +8.33 additional average accuracy in test-time scaling; the essay gives only the +8.33 figure. The essay's closing claim that the efficiency gain is not from doing less work contradicts the paper's reported training speedup from pruning rollouts during generation, and its statement that the learning signal is limited by the diversity of rollouts and not their number is the essay's own inference.

## Essay #6813: The Anticipative Adjoint Tags: stochastic-control, Volterra-integral-equations, delay, Hida-Malliavin-calculus, maximum-principle, anticipative-equations Stochastic Volterra integral equations with delay combine two sources of analytical difficulty: the Volterra kernel creates memory (the system's present depends on its entire past through a convolution), and the delay shifts the state feedback backward in time. Together, they produce a control problem where the adjoint processes are neither standard backward SDEs nor standard Volterra equations. An approach through Hida-Malliavin calculus (arXiv:2603.25452) reveals that the adjoint processes satisfy an anticipated backward stochastic Volterra integral equation (ABSVIE) — a BSVIE that depends on future values of the solution, reflecting the anticipative nature of the delay. Both necessary and sufficient stochastic maximum principles are established, providing a complete characterization of optimal controls. The through-claim: the adjoint equation for a delayed Volterra system is anticipative not by choice of formulation but by necessity. The delay forces the adjoint to "look forward" because the state at time t affects the cost at time t + delay, and the adjoint — which propagates cost information backward — must therefore depend on its own future values. The anticipative structure is not a technical complication to be eliminated but a faithful reflection of the physical problem's information flow. The cost of delay is not just a shifted boundary condition — it is a qualitative change in the adjoint's temporal structure. ---