Restricted geometric Langlands, global Arthur enhancements, and generic Ramanujan

The explanation

Expand the product q(1 − q)²⁴(1 − q²)²⁴(1 − q³)²⁴ and so on forever. The coefficients that come out are Ramanujan’s τ numbers. The first few are 1, −24, 252, −1472 and 4830. At the prime 2, τ(2) = −24, and 2 × 2^5.5 is about 90.5. At the prime 3, τ(3) = 252, and 2 × 3^5.5 is about 841.8. At 5 and 7 the same limit holds. Deligne proved in 1974 that this size limit holds at every prime.

That limit is the simplest case of a Ramanujan bound. In modern terms, the τ numbers come from an automorphic representation (a highly symmetric object built from a field). The generalized Ramanujan conjecture predicts tempered local pieces for generic ones, a precise form of “not too large”Source 1. Temperedness means the local piece sits weakly inside the regular representation. These papers work over function fields instead of the rationals. A function field is built from a curve over a finite field, the way the rationals are built from whole numbers.

Over function fields, much was known. Drinfeld proved Ramanujan for GL₂, and Laurent Lafforgue for GLₙSource 2. Lomelí proved the generic case for split classical groupsSource 3. The fourth paper claims it for exceptional groups, such as G₂ and E₈, at every placeSource 4. That includes ramified places, where the representation is more complicated. There is no restriction on characteristic or ramification depthSource 4. It builds on the third paperSource 5. That paper claims temperedness at unramified places from one generic unramified place, for split adjoint simple groupsSource 6.

The geometric side asks for more than numbers. Geometric Langlands compares two categories (collections of objects with maps between them). One side holds sheaves on the space of G-bundles over a curve. The other side holds sheaves on the space of local systems, the curve’s symmetry data. In characteristic zero, a nine-author project proved the geometric Langlands theoremSource 7. In characteristic p, Gaitsgory and Raskin built an equivalence onto a union of pieces of the spectral sideSource 8. The open question was whether any piece is missingSource 8.

The fifth paper claims that no piece is missing, in two settingsSource 8. If correct, this proves Conjecture 1.3.10 of Gaitsgory and Raskin thereSource 9. One setting is a curve over the algebraic closure of a finite field. The other is any algebraically closed field, with two extra conditions on pSource 10. It must be very good for the group and must not divide the Weyl group’s order. Very good excludes small primes such as 2 and 3 for E₇Source 11.

The argument works by contradiction. A missing piece would give a nonzero sheaf in characteristic zero that vanishes on passing to characteristic pSource 12. Careful local tests with Hecke modifications show this cannot happen.

The other papers fill in pieces. One claims that a canonical Arthur filtration is defined over the rationals, as Gaitsgory, Lafforgue and Raskin conjecturedSource 13. Another builds Arthur enhancements, but only under the finite-level decomposition stated as a hypothesisSource 14. It does not claim Arthur-packet classification or multiplicity formulasSource 15. Three more construct Hecke eigensheaves with tame ramification at marked pointsSource 16.

How sure is this? No Lean formalization is listed for these papers in the openai/math release. So no computer proof checker has checked them. The papers are long, linked and rely on deep recent work. No specialist review has been published.

What it doesn't show. This does not cover number fields such as the rationals; every result here is about curves or function fields in positive characteristic.