There is a particular thrill in mathematics that has no parallel in the physical sciences: the moment a structure you thought was rigid turns out to flex, or a chaos you thought was intractable reveals a hidden grammar. In the week of October 10–17, 2026, five papers landed in a span of seven days that together sketch a portrait of the field at full stretch — from the deepest corners of algebraic geometry to the playful arithmetic of tetration, from the combinatorics of polyhedra to the statistical machinery underpinning modern genomics. What unites them is not a single theorem or a single technique, but a shared preoccupation: the architecture of invariance, and what happens when it fails.
It is a fitting week to survey, because the results range from the abstractly beautiful to the immediately practical, and they collectively remind us that mathematics is not one discipline but a constellation of interlocking worlds, each with its own aesthetic and its own open wounds.
When the Invariant Cracks: The Haagerup Property's Quasi-Isometry Reckoning
In geometric group theory, the Haagerup property occupies a strange and important middle ground. It is a property of finitely generated groups — a kind of "nice enough to be amenable, but not quite" — that sits between the trivial amenability of abelian groups and the rigidity of property (T) groups. For years, a natural question haunted the field: is the Haagerup property preserved under quasi-isometry? Quasi-isometry is the coarse-grained notion of geometric equivalence for groups; two groups that are quasi-isometric share the same large-scale geometry, and many analytic properties (amenability, property (T), hyperbolicity) are known to be quasi-isometry invariants. The Haagerup property, however, resisted this classification.
In a paper published in the Bulletin of the London Mathematical Society, Carette, Whyte, Arnt, Pillon, and Valette deliver a definitive and, for the field, somewhat unsettling answer: neither the Haagerup property nor weak amenability is invariant under quasi-isometry of finitely generated groups [4]. The proof draws on the machinery of Cornulier–Valette and Whyte, and the authors extend the same counterexamples to show that the vanishing of equivariant compression is likewise not quasi-isometry invariant. This is not a technical footnote. It means that the large-scale geometry of a group, as captured by quasi-isometry, does not fully determine its analytic behavior in the C*-algebraic sense. The space of "reasonable" groups is more fractured than the geometric classification suggested.
The implications ripple outward. For operators on group C*-algebras, for the classification of von Neumann algebras, and for the broader program of understanding which analytic properties are geometric and which are genuinely algebraic, this result sharpens the boundary. It tells researchers that any program aiming to read analytic information off a group's coarse geometry must now account for a new layer of subtlety: groups that look the same from a distance can behave very differently up close.
Building the Modular Bridge: Tilting Sheaves and the Geometry of Duality
On the opposite end of the abstraction spectrum, a team at the intersection of algebraic geometry and representation theory has laid a critical new foundation. Achar, Makisumi, Riche, and Williamson, writing in the Memories of the American Mathematical Society, propose a construction of a monoidal category of "free-monodromic" tilting perverse sheaves on (Kac–Moody) flag varieties, living in the "mixed modular derived category" that Achar and Riche introduced [1]. The construction is designed to share most of the properties of its characteristic-zero counterpart, the category defined by Bezrukavnikov and Yun using pro-objects in triangulated categories.
To appreciate why this matters, one needs to understand the landscape. The geometric Satake equivalence and its variants have made flag varieties a central stage for the representation theory of reductive groups. In characteristic zero, the category of tilting perverse sheaves on a flag variety carries a monoidal structure that is deeply tied to the representation theory of the dual group. But in positive characteristic — the modular setting — the story was far less well understood. The Bezrukavnikov–Yun construction gave a characteristic-zero model, but translating it into the modular world required new tools.
The free-monodromic category is, in the authors' words, the main new ingredient in the construction of a "modular Koszul duality" equivalence for constructible sheaves on flag varieties, a program they began in their 2019 JAMS paper. Koszul duality is one of the great organizing principles of modern algebra: it relates a ring to a "dual" ring in a way that encodes deep homological information. Extending this duality to the modular setting, where the usual tools of characteristic zero (like the de Rham cohomology of flag varieties) break down, is a problem of first-order importance for the Langlands program in positive characteristic and for the representation theory of algebraic groups over fields of positive characteristic.
The Kac–Moody generalization is also significant. Moving from finite-dimensional reductive groups to infinite-dimensional Kac–Moody algebras opens the door to connections with vertex algebras, conformal field theory, and the representation theory of affine Lie algebras. A monoidal category of tilting sheaves in that setting would be a powerful new object, potentially linking geometric methods to problems in mathematical physics that have resisted purely algebraic attack.
Why the "free-monodromic" qualifier matters
The adjective is doing real work. In the modular derived category, the usual monodromy operators that encode Frobenius action introduce complications: the category is not semisimple, and the interaction between the monoidal structure and the mixed (Frobenius-twisted) structure is delicate. By restricting to the "free-monodromic" subcategory, the authors isolate the part of the structure that behaves well enough to support the monoidal product and the duality equivalence, while still capturing the essential representation-theoretic content. It is a surgical move — not a simplification that loses information, but a focus that makes the information accessible.
The Geometry of Frozen Digits: Taming Integer Tetration
If the previous two papers live in the world of categories and derived functors, Ripà and Di Pietro's contribution to Notes on Number Theory and Discrete Mathematics is a reminder that number theory still has room for the delightfully concrete [2]. They work with integer tetration — the iterated power tower ba = a^(a^(a^…)) of height b — and ask a question that is deceptively simple: as you increase the height b, how many of the rightmost digits of the tower freeze into place, and how fast?
The concept of congruence speed captures this: it is the rate at which the last k digits of the tower stabilize as b grows. Ripà and Di Pietro introduce a compact notation to express this speed and the cycle of rightmost non-stable digits, producing a discrete function that makes it possible to efficiently compute the exact number of frozen digits in the right tail of any nontrivial integer tetration. They also establish an improved upper bound for the minimum hyperexponent b̄(a) beyond which the congruence speed is constant — a sharpening of what was previously known.
Two results stand out for their elegance. First, they prove that for any two integers a, b > 1 whose product is not divisible by 10, the minimum of their constant congruence speeds is at most the constant congruence speed of their product. This is a submultiplicativity-type inequality for a quantity that, until now, lacked a clean algebraic description. Second, they exhibit infinitely many perfect powers whose degree matches their constant congruence speed at every height above 2, revealing a peculiar recurrence relation intrinsic to hyper-4 (tetration) that has no analogue in ordinary exponentiation.
The Appendix B generalization to all squarefree numeral systems is a quiet gem. It shows that the stabilization of congruence speed to a fixed positive value for all a > 1 not divisible by the radical of the radix is a phenomenon specific to squarefree bases, and the authors derive compact formulas for all prime-radix systems and for the composite squarefree senary (base 30) case. This is the kind of result that, while modest in scope, fills a gap in the arithmetic of iterated operations and provides a vocabulary for future work on hyper-operations in general.
A Hundred and Two Faces: The DICTO Hecatodihedron and the Art of Polyhedral Discovery
Not every mathematical discovery requires a new theorem. Sometimes, the contribution is an object: a shape that did not exist in the literature, that fits into a known family in a way no one had noticed, and that reveals a small but genuine gap in the taxonomy of convex polyhedra. The DICTO Truncated-Icosahedral Hecatodihedron is exactly that [3].
It is a convex polyhedron with 102 faces, 80 vertices, and 180 edges, arising as the seventh stage in the "Spheres of Dharma," a cycle of nested icosahedral shells. Geometrically, it is the truncated icosahedron (the familiar soccer-ball shape) with a low hexagonal pyramid — the authors call it a "Shield" — attached to each of its 20 hexagonal faces, at the unique height where the two triangles of neighboring Shields across every hexagon–hexagon edge lie in a single plane. The resulting faces are 12 regular pentagons, 30 rhombi, and 60 isosceles triangles, and the solid has only two distinct edge angles.
The combinatorial identification is the key insight: the Hecatodihedron is Conway's hexakis truncated icosahedron k6tI (which has 132 faces) with its 60 coplanar triangle pairs merged into 30 rhombi. The authors also correct an earlier claim that all 102 faces touch a single sphere: in fact, the rhombi and triangles are tangent to a sphere, but the pentagons lie slightly inside, so the solid has no insphere. This correction, while small, matters for the classification of tangential polyhedra.
Perhaps most interesting is the connection the authors draw to the pentakis dodecahedron: the 20 Shield apices are precisely the three-fold vertices of the canonical pentakis dodecahedron, meaning the Hecatodihedron can be viewed as the joined truncated icosahedron (jtI) with its 12 five-fold corners cut off at the pentagon planes. This recasts the object as a derivative of a known compound, grounding it in the classical polyhedral taxonomy while preserving its novelty as a distinct convex solid. The self-dual compound with the DICTO Ogdoecontahedron (182 faces), whose core and hull are those of the classical truncated icosahedron–pentakis dodecahedron compound, adds another layer of structural coherence.
In an era when computational geometry can enumerate millions of polyhedra, the value of a paper like this is not novelty of discovery but novelty of recognition: identifying the right height, the right merging, the right connection to known objects, and presenting the result in a form that is both precise and aesthetically satisfying.
Statistical Architecture: Lasso Meets Mixture-of-Experts in the High-Dimensional Wild
The final paper in this week's crop is the most applied, and in a sense the most urgent. Nguyen, Nguyen, Chamroukhi, and McLachlan, in the Annals of the Institute of Statistical Mathematics, tackle a problem that sits at the intersection of high-dimensional statistics, machine learning, and systems biology [5]. The setting is softmax-gated Gaussian mixture-of-experts (SGMoE) regression, where the predictor dimension P can far exceed the sample size N, and the response is itself multivariate (dimension Q), with K experts each contributing a Gaussian component.
The core result is a non-asymptotic Kullback–Leibler oracle inequality for Lasso estimation in this setting, with an explicit sufficient regularization level and explicit dependence on all the key parameters: N, P, Q, K, and the boundedness constants. This is not a large-sample limit theorem; it is a finite-sample guarantee that tells a practitioner exactly how much regularization to use and how the error scales with the dimensionality of the problem.
Two extensions are particularly valuable. First, the theory covers random covariates, moving beyond the fixed-design assumption that underlies much of the Lasso literature. Second, under model misspecification — the realistic scenario where the true data-generating process is not exactly a Gaussian SGMoE — the authors show that the Lasso estimator controls the best Gaussian SGMoE approximation in the candidate class, transferring deterministic approximation error into statistical KL-risk bounds. This is the kind of robustness guarantee that separates a usable theory from a formal one.
The applications are concrete and high-stakes: heterogeneous genotype–phenotype mapping, omics-to-trait prediction, and AI-enabled virtual-cell modelling in systems biology. In these settings, the number of candidate genetic or molecular predictors routinely dwarfs the number of samples, and the relationship between predictors and phenotypes is inherently heterogeneous — different subpopulations of cells or patients respond differently. A mixture-of-experts model captures this heterogeneity, and the Lasso provides the sparsity needed to extract signal from noise. The authors' reproducible experiments validate KL decay and density recovery, compare Lasso with SCAD in the P > N regime, and assess model-selection criteria on high-dimensional biological benchmarks.
The Bigger Picture: Structure, Invariance, and the Many Faces of Discovery
What do a counterexample in geometric group theory, a new monoidal category in algebraic geometry, a notation for tetration, a 102-faced polyhedron, and a KL oracle inequality for mixture models have in common? On the surface, almost nothing. But underneath, they share a common gesture: the identification of a structure, the testing of its limits, and the construction of a bridge to a structure one could not previously reach.
The Haagerup result [4] tells us that the bridge between coarse geometry and analytic behavior is narrower than we thought. The tilting sheaves construction [1] builds a bridge that did not exist, connecting the modular and characteristic-zero worlds through a carefully chosen subcategory. The tetration paper [2] builds a bridge between the arithmetic of iterated powers and the structure of numeral systems, revealing that the "frozen digits" phenomenon is not a curiosity but a theorem. The Hecatodihedron [3] builds a bridge between a new object and the classical polyhedral taxonomy, showing that the space of convex polyhedra is still full of unexplored corners. And the SGMoE Lasso theory [5] builds a bridge from abstract high-dimensional statistics to the concrete problem of understanding why a particular cell expresses a particular gene in a particular tissue.
Mathematics in October 2026 is not a single narrative. It is a conversation among disciplines, a negotiation between abstraction and application, a perpetual testing of whether the structures we build are as rigid as we assume or as flexible as we hope. This week's five papers, taken together, suggest that the answer is usually both: the structure holds, but only if you look at the right level, in the right category, at the right height. And when it doesn't — when the Haagerup property slips through the quasi-isometry net, when the pentagons fall just inside the sphere — the failure is not a bug. It is the next question.
References
- Pramod N. Achar, Shotaro Makisumi, Simon Riche et al. (2026). Free-Monodromic Mixed Tilting Sheaves on Flag Varieties. Memoirs of the American Mathematical Society.
- Marco Ripà, Gabriele Di Pietro (2026). A compact notation for peculiar properties characterizing integer tetration. Notes on Number Theory and Discrete Mathematics.
- DICTO (2026). The DICTO Truncated-Icosahedral Hecatodihedron. Zenodo (CERN European Organization for Nuclear Research).
- Mathieu Carette, Kevin Whyte, Sylvain Arnt et al. (2026). The Haagerup property is not invariant under quasi‐isometry. Bulletin of the London Mathematical Society.
- TrungTin Nguyen, Hien D. Nguyen, Faïcel Chamroukhi et al. (2026). Non-asymptotic oracle inequalities for the Lasso in high-dimensional mixture of experts. Annals of the Institute of Statistical Mathematics.