Eight Papers, Seven Threads

Spectral gaps and expansion

The same quantity — the gap between the top of a spectrum and the rest — is the engine in five of the eight documents, under five names.

• B studies it directly: the Cheeger inequality λ/2 ≤ h ≤ √(2λ) ties spectral gap to cut structure for graphs, and the paper's whole content is that this equivalence breaks for matrix tuples — quantum expansion implies dimension expansion but not conversely (Theorem 1.8).
• E ch. 3 (nonsofic groups) runs on Kun's expander decomposition for sofic approximations of property-(T) groups; ch. 4 (Connes rigidity) needs property (T) for its ICC groups; ch. 9 (Ramsey) uses saturated random matrices. Property (T) is a uniform spectral gap over all unitary representations — B's object of study.
• D: the fractal uncertainty principle's headline application is a spectral gap for resonances on hyperbolic surfaces (Dyatlov–Jin 2017): waves cannot concentrate, hence essential spectral gap, hence decay.
• G: a sheaf's monodromy group being "big" (SL) versus "small" (²B₂) is decided by moments of Frobenius traces — M_2,2 > 2 kills SL₁₄ — i.e. by how the trace distribution spreads, a spectral criterion in disguise.
• A supplies the deep reason Ramanujan graphs exist at all: Lubotzky–Phillips–Sarnak's optimal expanders rest on Deligne's proof of the Riemann hypothesis over finite fields, the very theorem A narrates.
• H meets it as condition number (gradient descent converges at rate 1−μ/L; the gap between the smallest and largest curvature) and as Lyapunov LMIs.

Uncertainty principles: a function and its Fourier transform cannot both be small

• D is the fractal version: no function concentrates on a porous set while its transform does; Cohen's contribution is the higher-dimensional condition (line porosity) and the Beurling–Malliavin detour through complex analysis.
• E ch. 1 proves a sign uncertainty principle: for eigenfunctions of the Fourier transform, the radii A_±(d) beyond which f and f̂ can be non-negative are both (1/π + o(1))√d — and this is what pins the Cohn–Elkies sphere-packing exponent. The proof is also a strip estimate in the complex plane (Phragmén–Lindelöf on the Mellin transform), the same family of tools as Beurling–Malliavin.
• A ends with Dyson's quasicrystal speculation: a point set whose Fourier transform is also a point set. That is the exact opposite corner of the uncertainty landscape — the Poisson-summation extreme — and a quasicrystal with the zeta zeros as spectrum would be a "perfectly non-porous" object. D and A are two ends of one axis.
• H ch. 4: compressed sensing as ℓ₁ minimum-norm recovery works because of a discrete uncertainty principle (Donoho–Stark); the book uses it without naming it.

Duality as proof technology

H teaches LP duality, KKT and minimax as a toolkit for engineers. Three of the research documents are that toolkit used at the frontier.

• E ch. 1 and 2: the Cohn–Elkies bound and the Delsarte/MRRW bound are dual LP certificates — exactly the weak-duality argument of H ch. 3/6: any feasible dual point bounds the primal. The novelty in E ch. 2 is enlarging the certificate (each code point carries a harmonic subspace, via Gelfand pairs) — that is, choosing a better dual cone.
• C proves lower bounds via Yao's minimax principle, which is von Neumann's minimax, which is LP duality — the "minimax panel" of H ch. 6. C's canonical deterministic protocol is multiplicative weights, i.e. mirror descent, which H omits and which would slot naturally between its gradient and proximal chapters.
• B's Cheeger inequality is a duality gap statement: the spectral relaxation (λ) versus the combinatorial primal (h). The paper's Theorem 1.9 is the matrix-world table of such gaps.

Random matrices, universality, and equidistribution

• A poses the mystery: zeta zeros obey GUE statistics and nobody knows why. Its partial answer is Katz–Sarnak: in the function-field world, zeros are eigenvalues of Frobenius, and a family's statistics are dictated by its monodromy group — "the glue".
• G is a worked instance of that glue, in the hardest possible case: determining whether the monodromy of an explicit family is the finite Suzuki group or all of SL_D. Deligne's equidistribution theorem (Remark 9.15) is used both as evidence (all computed |Trace| = 1, impossible for SL) and as a proof device (Prop. 9.16: one trace off the unit circle forces infinite monodromy). The 7,500-CPU-hour moment computation is the same kind of object as Odlyzko's eight million zeros in A.
• B obtains dimension expanders from Haar-random unitaries (Dvir–Shpilka); E ch. 4 relies on Haar measure on a compact dual group being blind to the group law; ch. 9 uses saturated random matrices.
• C's central device is sampling — estimate an expectation with fewer samples than the naive 1/ε² by removing the first Efron–Stein term. H ch. 7 is the Bayesian/MLE view of the same estimation problem, without communication constraints.

Lattices, number fields, and the geometry of numbers

• F: the unit-distance disproof is pure geometry of numbers — a lattice in ℂ^f from a CM field with many units of absolute value 1, projected to the plane. The exponent 1 + 6·10⁻³⁸ is an honest count of lattice points in a polydisc against unit translates. The fields come from an infinite Golod–Shafarevich class-field tower, which F notes is the same construction behind good codes and sphere packings (Lenstra, Litsyn–Tsfasman).
• E contains three lattice chapters: sphere packing (ch. 1), CVP hardness (ch. 7, via Reed–Solomon interpolation lifted mod 2 to an integer lattice), and Ehrhart's conjecture (ch. 8, lattice points in a convex body via Bergman spaces). Ch. 2's codes are the discrete cousin.
• G lives in the same arithmetic: Frobenius traces in ℤ[i], Jacobi symbols, primitive prime divisors of q ± √(2q) + 1. Sawin's section of F explains why fixed-field approaches fail: zeta zeros dominate class-number effects — A's objects appearing as an obstruction in F.

Finite groups of Lie type and finite geometry

• G is about the Suzuki groups ²B₂(2²ⁿ⁺¹), which exist only in characteristic 2 and act on the Suzuki–Tits ovoid inside the symplectic generalized quadrangle W(q), q even.
• E ch. 10 disproves the Erdős–Simonovits compactness conjecture using the incidence graphs of exactly those symplectic generalized quadrangles W(q), with the q-even / q-odd dichotomy doing the work. The two documents are looking at the same finite geometry from the sheaf side and the extremal-graph side.
• E ch. 2 (Gelfand pairs, harmonic spaces) and B (Lubotzky–Zelmanov's result is stated for irreducible representations of finite groups) are representation theory of the same families.

Does solving n copies cost n times as much?

• C answers no in one direction: estimating E[f] to error ε looks like 1/ε² instances of f, but debiasing does it in Õ(R^ow(f)/ε) — a counterexample to direct-sum intuition, with a matching lower bound via a lifting theorem.
• E ch. 6 answers yes, exponentially in the other: quantum parallel repetition — the value of n copies of an entangled game decays as exp(−cε¹³n/…). Both papers are built from the same Raz/Holenstein/Yuen lineage of conditioning-and-sampling arguments; C's "cbd decomposition" and E's "postselection-stable sampleability" are cousins.
• H's ADMM consensus form and SGD mini-batching are the engineer's version: how much of a problem can be split across workers or samples before coordination cost bites.

Machine-produced mathematics — and what the folder's name is really about

Read together, these documents form a spectrum of how a computer participates in a proof, and the spectrum is the governance question.

• G — computer as witness. 7,500 CPU-hours of Magma produce one number (M_2,2 > 2) that a human argument then uses. Every step is auditable; the machine asserts nothing.
• H — computer as assistant. Brunton credits GPT-4o/5 for code and brainstorming and explicitly not for prose or equations. A declared boundary.
• D — the human baseline. Cohen's story is the one the other documents implicitly measure against: years of stuck attempts, a rediscovered 1960s theorem, "my ignorance gave me confidence." Tsimerman's line in F — AI can "play for longer in treacherous waters" — is the same virtue, industrialised.
• F — computer as author, humans as referees. Nine leading mathematicians digest, simplify (Sawin's single-split-prime version), and certify a model's proof, then disagree in print about what it means: Alon calls it an outstanding achievement; Gowers proposes "Kolmogorov complexity modulo experts" as a novelty measure; Wood warns about uncredited prior literature, unverified claims, and the need for citation norms; Wang raises arXiv's unspoken social contract with AI companies.
• E — computer as author, no referees visible. Ten results, seven of which settle named conjectures, presented in standard first-person-plural register with no methodology, no verification statement, and an "updated August 6" footnote preserving an original whose differences are not described. E is the artifact F's commentaries are worried about.