AI Research Methods
New technical methods, papers, architectures, training techniques, reasoning approaches, and research findings with applied significance.
Univalent Maps Have a Uniform Reciprocal-Derivative Growth Gap Below One Quarter
An OpenAI preprint proves that the average reciprocal derivative of a normalized univalent map grows more slowly near the unit-circle boundary than Kraetzer’s proposed one-quarter exponent permits. The bound holds uniformly across all such maps, with an exponent at least some fixed positive amount below \(1/4\); the paper establishes that a gap exists but does not calculate its size.
Bounded Treewidth Guarantees Finite-Distortion L1 Embeddings
An OpenAI preprint proves that every finite connected graph with bounded treewidth embeds into finite-dimensional \(L_1\) with distortion bounded by a constant determined by the tree-decomposition bag size—not by the graph’s number of vertices or the length of its decomposition. The paper’s central challenge is ensuring that distant vertex pairs remain separated, not merely that the embedding does not stretch distances too far. It addresses this with weighted cuts, using reservations and anchor cuts to preserve separation across the graph.
Planar Schrödinger Evolution Converges at the H¹ᐟ³ Endpoint
A September 2026 OpenAI manuscript claims that planar Schrödinger evolution converges almost everywhere to its initial data for functions with exactly \(1/3\) of a Sobolev derivative, reaching a boundary where earlier results required more regularity. The authors’ proof rests on a local \(L^1\) maximal estimate that controls the evolution uniformly over time at that endpoint. They define the evolution through a Gaussian-regularized Fourier integral and show the convergence holds on one full-measure set for all times approaching zero.
Gaussian Partitions Have a Sharp First-Moment Bound in Every Dimension
An OpenAI preprint proves that, in every positive dimension, the sum of the squared first-moment vectors of any finite partition of Gaussian space is at most 9/(8π). The bound is attained by dividing a plane into three 120-degree sectors and extending them across any remaining dimensions. The paper’s main argument rules out a better partition by reducing a hypothetical improvement to geometric constraints on the moments and probabilities of its cells; it also derives a sharp bound for the expected maximum of finitely many centered Gaussian scores.
A Nonspectrahedral Hyperbolicity Cone Has an Exact 307-Variable Semidefinite Lift
An OpenAI preprint gives an exact semidefinite lift for a hyperbolicity cone that a companion paper shows cannot be represented directly by a positive semidefinite matrix in its original coordinates. The authors add 307 auxiliary variables and impose one 100-by-100 matrix inequality; projecting away those variables recovers the cone, including its singular boundary points. The construction establishes that the obstruction is to a direct representation, not to semidefinite representation by lifting.
Brennan’s Conjecture Holds for Derivative Integrals Between 4/3 and 4
An OpenAI manuscript dated September 24, 2026, argues that Brennan’s conjecture holds for conformal maps from any simply connected plane domain onto the unit disk: the integral of the map’s derivative to the power \(s\) is finite for \(4/3<s<4\), without assumptions that the domain is bounded or has a smooth boundary. Its key estimate makes one the sharp limiting radial growth exponent for the inverse-square derivative mean, with the Koebe map showing why the exponent cannot be lower.
Every Convex Body Has a Lattice Covering of Density O(n log n)
An OpenAI preprint proves that every compact convex body in dimension \(n \ge 2\) can be covered by copies arranged on a single lattice with density at most \(C n \log n\), for an absolute constant \(C\). The bound applies even to bodies that are neither symmetric nor smooth; it limits the worst case rather than describing the density needed for every shape. The proof coordinates weighted choices across many sections, then uses one global shear to keep the construction within a single lattice.
Optimal L1 Distortion for Edit Distance Grows as exp(Θ(√(log d log log d)))
An OpenAI preprint argues that edit distance on strings of length at most \(d\) requires exponentially growing distortion when represented in real \(L_1\), even if the embedding’s dimension and computation time are unrestricted. For every finite alphabet with at least two symbols, it bounds the best possible distortion above and below by \(\exp(\Theta(\sqrt{\log d\,\log\log d}))\), with the constants in the exponent unspecified; the lower bound already holds for binary strings of a common length.
A Fixed-Slope Trilinear Hilbert Transform Maps Three L³ Inputs to L¹
An OpenAI preprint proves an \(L^3\times L^3\times L^3\to L^1\) bound for the trilinear Hilbert transform at the fixed slopes 1, 2 and 3. The result is specific: it does not settle every exponent or version of the transform. The paper’s proof tackles the multiscale interactions that defeat direct estimates, using quadratic structure to obtain bounds uniform across dyadic depths.
Uniform Sub-Gaussian Projection Tails Yield Dimension-Free Log-Sobolev Bounds
An OpenAI manuscript argues that a centered, log-concave distribution has a dimension-free logarithmic Sobolev inequality when every one-dimensional projection has uniformly controlled sub-Gaussian tails at a common scale. The claimed bound applies to all smooth, compactly supported functions, with an entropy-to-gradient-energy coefficient set by that scale rather than by the ambient dimension; the authors’ proof uses a contradiction argument based on bounds on mutual information.
Some Convex Bodies Require Translative Covering Density Of Order n log n
An OpenAI manuscript claims that, in every sufficiently large dimension, some centrally symmetric convex body requires translative covering density of at least \(c n \log n\), for an absolute constant \(c>0\). The result identifies the worst-case order, matching the known \(O(n \log n)\) upper bound for all convex bodies; it does not say that every body is this difficult. The proof constructs a random body and aims to rule out every sufficiently low-density covering, including arrangements chosen after the body is known.
OpenAI Paper Claims a Fixed Zero-Free Boundary at Re s = 7/8
OpenAI’s September 2026 paper claims that the Riemann zeta function and every Dirichlet L-function have no nontrivial zeros to the right of Re s = 7/8, a fixed boundary inside the critical strip. The result is weaker than the Riemann hypothesis, which places every nontrivial zero on Re s = 1/2, but would improve on zero-free regions whose width shrinks with height. Lean has accepted a formalized version of the zeta statement; the 199-page paper itself had not received public expert review.
Crypto Faces an AI Security Risk Without a Known Attack
AI-driven advances in mathematics could threaten the cryptography that protects crypto wallets, but the risk is harder to define than the quantum threat the industry has spent years preparing for. On Diet TBPN, host John Coogan relayed warnings from Vitalik Buterin and cryptographer Matthew Green, while arguing that crypto has time to strengthen its defenses rather than rush funds into new wallets. Coogan and co-host Jordi Hays also questioned whether the open-source tools that could expose vulnerabilities should be restricted.
Mod-ℓ Milnor Products Recover Function Fields Up to Frobenius
An OpenAI preprint claims that a finitely generated field extension of an algebraically closed field, with transcendence degree at least two and characteristic different from ℓ, can be reconstructed from its nonzero elements modulo ℓ-th powers and the Milnor products among those classes. The authors argue that the product’s Steinberg relation retains enough information about addition to recover the field’s perfect closure and constant field, up to a global scalar and, in positive characteristic, powers of Frobenius. The proof uses valuations and geometry to identify one-variable subfields before recovering the field operations.
Milnor K-Theory Relations Recover Addition and the Field
An OpenAI preprint argues that a field of characteristic \(p\) can be reconstructed from its multiplicative group modulo \(p\)-th powers, provided one also retains the Steinberg relations that define Milnor \(K\)-theory modulo \(p\). Under finite-generation and transcendence-degree hypotheses, the authors show that this data recovers projective lines; preserving those lines then recovers addition and yields the field isomorphism.
A Rank-Weighted Tail Bound Establishes the Mean for Quadratic Twists
An OpenAI preprint argues that, for any fixed rational elliptic curve, the mean analytic rank of its signed squarefree quadratic twists tends to one half as the twist parameter grows. The result requires more than showing that half the twists have rank zero and half rank one: rare twists with very high rank could still raise the average. The paper’s central step is a bound on the total rank contributed by those high-rank twists, which lets the authors control the tail and establish the mean.
A Strict Bound Gap Establishes Catalan’s Constant Is Irrational
An OpenAI manuscript dated September 24, 2026, claims to prove that Catalan’s constant is irrational by constructing a sequence of determinants. The argument shows that if the constant were rational, the determinants would have to satisfy an arithmetic lower bound; integral estimates, in turn, force them below that bound. The contradiction depends on exact cancellation of another constant in the determinant entries, along with proofs that the determinants are nonzero and their denominators are controlled.
An Orbit Bound Proves Finiteness for Mixed CM Products in A₂
An OpenAI preprint proves that a fixed Hodge-generic curve in the moduli space \(A_2\) contains only finitely many rational points whose associated abelian surfaces are isogenous to a product of elliptic curves with at least one complex-multiplication factor. The authors establish this component of the Zilber–Pink conjecture by combining André–Oort in the case where both factors have CM with a Galois-orbit bound and a counting theorem in the mixed case. The result allows arbitrary isogeny degrees and imposes no additional boundary or reduction hypothesis.
Pro-Modularity at 2 Holds Without de Rham or Irreducible Residual Hypotheses
An OpenAI preprint proves pro-modularity at 2 for a broad class of two-dimensional Galois representations, without requiring them to be de Rham at 2 or to have irreducible residual representations. The authors’ claim is about occurrence in a completed Hecke algebra spanning modular forms of many weights, not realization by a single classical form. Their proof constructs one auxiliary representation with stronger local properties to establish classicality, then uses a dimension comparison to transfer Hecke support to the wider family.
Weighted Divergence Guarantees Infinitely Many Inhomogeneous Approximations
An OpenAI manuscript on the weak inhomogeneous Duffin–Schaeffer conjecture gives a sufficient condition for infinitely many shifted rational-approximation hits for almost every real number. For a fixed shift \(\gamma\) and non-negative tolerances \(\psi(q)\), the authors prove the result when \(\sum_q (\phi(q)/q)\psi(q)\) diverges, without requiring the approximating numerators to be coprime to their denominators. The proof’s main challenge is controlling the structured overlap among approximation intervals.
Largest Prime Factors of Consecutive Integers Become Independent
An OpenAI preprint on the joint Dickman law argues that the normalized sizes of the largest prime factors of consecutive integers become statistically independent when integers are counted uniformly over an expanding range. In the limit, the chance that both largest prime factors fall below specified power thresholds is the product of the two familiar Dickman probabilities. The result also implies that each neighbor has the larger largest prime factor half the time, in natural density.
Quaternionic Division Points Are Finite Across All Indefinite Algebras
An OpenAI preprint proves that a reduced, irreducible, closed curve defined over the algebraic numbers and Hodge generic in the Siegel threefold contains only finitely many points whose abelian surfaces have an indefinite quaternion division algebra as their full rational endomorphism algebra. The finiteness holds across all such algebras, rather than separately for each fixed type. The proof establishes a height bound independent of the algebra, then rules out unbounded degrees by showing they would force the curve’s period image into a locus too small to contain it.
Prime Predecessor Factors Converge to the Poisson–Dirichlet Law
An OpenAI preprint proves a conjecture about the prime factors of \(p-1\): for a uniformly sampled large prime \(p\), any fixed number of the largest prime-factor occurrences, measured by their shares of \(\log(p-1)\), converge jointly to the leading pieces of a Poisson–Dirichlet partition with parameter one. The result is a limit in distribution, not a claim that each individual prime has a typical factorization. Its proof combines estimates for factor counts in arithmetic progressions with a size-biased sampling argument that yields the stick-breaking law.
A Single Exterior Sheet Can Coexist With Disk Degrees Divisible by p
An OpenAI preprint, *Étale covers with a prescribed exterior sheet*, constructs a finite étale cover of a \(p\)-adic curve that is trivial on one sheet over the exterior of fixed disks, while every component above each disk has degree divisible by a chosen prime \(p\). The authors’ argument first builds cycles above the disks, then uses a group-theoretic quotient to enforce the degree condition on every disk component without losing the exterior sheet.