======================================================================
THE RAMANUJAN CONJECTURE OVER FUNCTION FIELDS:
AN HONEST INTRODUCTION FOR CURIOUS TEENAGERS
======================================================================
Honest notice:
This is an EXPLANATORY ESSAY, not a proof. It exists because
ramanujan_arthur_cuspidal_original.txt (the file it is replacing)
is not a valid proof of what it claims. The original file is 5547
lines of LaTeX claiming to prove two major open conjectures:
Gaitsgory-Lafforgue-Raskin (GLR) Conjectures 3.4.5 and 3.4.6 at
every full finite level, and the unramified generic Ramanujan
conjecture for split connected adjoint absolutely simple groups
over function fields. Neither is, as far as I know, currently a
theorem. The original file has the stylistic fingerprint of an
LLM hallucination: many of its proofs contain "for completeness"
paragraphs that quietly assume the very hypothesis under attack.
A fake proof cannot be made real by editing. "More correct" is
proofreading against reality, and the reality is that these
statements are open.
What follows, then, is an honest introduction to the real math
around this problem. Every claim about "what is known" carries an
author name and year so you can check it yourself. You do not have
to trust me.
======================================================================
SECTION 1. THE CLASSICAL RAMANUJAN CONJECTURE
======================================================================
In 1916, Srinivasa Ramanujan defined a sequence of integers
tau(1)=1, tau(2)=-24, tau(3)=252, tau(4)=-1472, tau(5)=4830, ...
via the expansion of the modular discriminant
Delta(z) = q * prod_{n>=1} (1-q^n)^24 = sum_{n>=1} tau(n) q^n,
where q = e^{2 pi i z} for z in the complex upper half-plane, and
Delta is the unique weight-12 normalized cusp form on SL(2,Z).
The tau function has astonishing hidden structure. Three examples:
(a) Multiplicativity. If gcd(m,n)=1 then
tau(m) tau(n) = tau(mn).
So the whole sequence is determined by its values at primes.
(b) A recurrence. For any prime p and any r >= 1,
tau(p^{r+1}) = tau(p) tau(p^r) - p^{11} tau(p^{r-1}).
(c) A real bound. Ramanujan observed computationally that
|tau(p)| <= 2 p^{11/2}
for the primes p he could calculate, and conjectured that
this holds for every prime p.
The inequality (c) is the original Ramanujan conjecture.
Equivalent form via Hecke operators. Delta is an eigenform of
every Hecke operator T_p. If we write
a_p := tau(p) / p^{11/2},
then (c) is exactly the statement
|a_p| <= 2.
The two bounds -2 <= a_p <= 2 for all p are exactly what is needed
to conclude that numbers alpha_p, beta_p exist with
a_p = alpha_p + beta_p, alpha_p beta_p = 1, |alpha_p|=|beta_p|=1
(just solve a quadratic). Equivalently: the local L-factor
L_p(Delta, s) = 1 / (1 - a_p p^{-s} + p^{1-2s})
has no pole off the line Re(s)=1/2. So the Ramanujan conjecture is
the statement that every local L-factor of Delta is "as analytically
nice as it possibly can be".
Why this is hard. We are asking about the size (complex modulus) of
numbers that arise from infinitely many independent arithmetic
patterns. There is no a priori reason a_p should be bounded at all;
the geometry of the modular form gives no obvious constraint. The
conjecture, believed for a century, has resisted elementary attacks
and yields only to a deep cohomological machine - the Weil
conjectures - plus a heroic construction by Deligne.
How Deligne proved it (1974). Deligne observed that Delta mod p
arises as a piece of the etale cohomology of a certain modular
variety, which carries an l-adic Galois representation. The bound
|a_p| <= 2 is then equivalent to the statement that Frobenius acts
on this representation with all eigenvalues of absolute value
exactly p^{11/2} (i.e. the representation is "pure of weight 11").
That is a special case of the Weil Conjectures, proved in full
generality by Deligne:
- Pierre Deligne, "La conjecture de Weil I", Publ. Math. IHES 43
(1974), 273-307.
- Pierre Deligne, "La conjecture de Weil II", Publ. Math. IHES
52 (1980), 137-252.
So: the classical Ramanujan conjecture is a THEOREM. Everything
else in this essay is the attempt to generalize it to larger
groups, and (spoiler) that attempt fails in general, for reasons
that are themselves deep.
======================================================================
SECTION 2. WHAT DOES "TEMPERED" MEAN, AND WHY IS IT THE RIGHT
GENERALIZATION?
======================================================================
The classical bound |a_p| <= 2 does not immediately say what to
generalize. Here is the right generalization, in three passes from
concrete to abstract.
Pass 1: GL(1), i.e. characters. Let chi be a Hecke character of the
rationals (i.e. an automorphic character of GL(1) over Q - an
ideal-class character). Then chi is a continuous homomorphism from
the idele class group to C^*, and its local components at primes p
give complex numbers chi_p with |chi_p| = 1 for every unramified p.
The local L-factor is (1 - chi_p p^{-s})^{-1}, no pole off
Re(s)=1. "Tempered" in this case is trivial: every character of a
compact group has modulus 1.
Pass 2: GL(n). Let pi be a cuspidal automorphic representation of
GL(n) over Q (for example, a normalized holomorphic cusp form of
some weight, for n=2). At an unramified prime p, pi_p is
determined by n complex numbers alpha_1,...,alpha_n (the "Satake
parameters" of pi_p). The local L-factor is
L_p(pi_p, s) = prod_{i=1}^n (1 - alpha_i p^{-s})^{-1}.
pi_p is TEMPERED if
|alpha_i| = 1 for all i = 1,...,n.
Equivalently, if we put a_i = alpha_i and let A_p be the
corresponding diagonal matrix in GL_n(C), then pi_p is tempered iff
A_p lies in the maximal compact subgroup U(n) subset GL_n(C).
For GL(2), tempering at p is exactly |a_p| <= 2 with a_p = alpha+beta,
alpha beta = 1, |alpha|=|beta|=1.
Generalized Ramanujan conjecture (GRC) for GL(n): every cuspidal
automorphic representation of GL(n) is tempered at every unramified
place.
Special cases:
- n=1: trivially true (all Hecke characters have |chi_p|=1).
- n=2 over Q: Deligne 1974, as in Section 1.
- n=2 over a function field: Drinfeld 1974, see Section 3.
- n=3 over Q: not fully known unconditionally; a lot is known
conditionally.
- n>=1 over a function field: L. Lafforgue 2002, see Section 3.
- n>=3 over Q in general: still OPEN in 2025.
Pass 3: general reductive groups. Why is GL(n) the wrong level of
generality for the Ramanujan conjecture? Because number theory
naturally produces automorphic representations on many other groups:
symplectic groups Sp(2n), orthogonal groups SO(m), exceptional
groups G_2, E_6, ..., and their inner forms. Those have no
canonical embeddings into GL(N) that fix the L-functions. So we
need a statement about general reductive groups.
Setup: Let G be a connected reductive group over a number field or
function field F. Let pi = tensor'_x pi_x be a cuspidal automorphic
representation of G(A_F). For every unramified closed point x, the
spherical representation pi_x corresponds under the (classical or
geometric) Satake isomorphism to a semisimple conjugacy class
s_x(pi) in G^hat(C)
where G^hat is the complex Langlands dual group (e.g. GL(n)^hat =
GL(n,C); Sp(2n)^hat = SO(2n+1,C); SO(2n+1)^hat = Sp(2n,C)).
pi is TEMPERED at x iff the semisimple conjugacy class s_x(pi)
has a representative lying in a maximal compact subgroup K_x of
G^hat(C). For G=GL(n), G^hat=GL(n,C), and the maximal compact is
U(n); that recovers Pass 2.
Generalized Ramanujan conjecture for G: every cuspidal automorphic
representation of G(A_F) is tempered at every place. Status: open
for every nonabelian G, even over function fields, because cusp
forms on general G are not built from GL(n) forms in a way that
propagates temperedness.
Two important remarks:
(i) GRC as stated is FALSE for general G if you do not replace
"every cuspidal" with "every cuspidal that is generic". The
Arthur conjectures predict non-tempered cuspidal forms on
groups as soon as G has a proper Levi subgroup with a
cuspidal automorphic representation that can be parabolically
induced in a non-tempered way. For instance, on a symplectic
group, you can build cuspidal forms whose Satake parameters
include |alpha|^2 = p, not 1. These are predicted by Arthur's
SL_2 picture (Section 4).
(ii) The "generic Ramanujan" version - cuspidal + generic at one
place => tempered at every unramified place - is the version
that is expected to be true for every reductive group. This
is the version claimed in the original file's Corollary 1
("main:generic-ramanujan"), and it is a big open problem in
general. Over function fields it is known for GL(n) by
Lafforgue; for split adjoint absolutely simple groups it is
NOT currently known in general (partial results by Sawin-
Templier and Ciubotaru-Harris, see Section 8).
So when the original file says "we prove the unramified generic
Ramanujan conjecture for split connected adjoint absolutely simple
groups over function fields", that is a heroic claim. It cannot
honestly be asserted as a theorem without a published proof, and
the original file is not a published proof. It is a synthetic
LaTeX file.
======================================================================
SECTION 3. FUNCTION FIELDS. WHY IS F_q(X) EASIER THAN Q?
======================================================================
The original file works over a function field. That is a big hint
about what tool the paper is using, and about why function fields
are special for Ramanujan-type problems.
A function field: let X be a smooth projective geometrically
connected curve over a finite field F_q. Let F = F_q(X) be its
field of rational functions. Then F behaves like a number field:
- primes of F = closed points of X
- adeles A_F = restricted product over closed points
- a "number" in F has finitely many poles and zeros on X
and you can ask all the same automorphic questions.
Why it is easier. Over Q, the only known route to Ramanujan-type
bounds is to realize automorphic forms inside cohomology and then
apply the Weil conjectures (Deligne). But cohomology for
number fields is not fully developed: we do not have a
cohomological realization of the complete Langlands correspondence
over Q. Over function fields, however, there IS an explicit
cohomological realization: the moduli space of SHTUKAS. A shtuka
is a vector bundle E on X together with a meromorphic isomorphism
phi : E --> Fr(E)
to its Frobenius pullback (Fr is x |-> x^q induced by Frobenius
on the base F_q), allowed to have prescribed singularities at
finitely many points. Shtukas live in a moduli stack that is
locally of finite type and behaves exactly like a Shimura variety
built for function fields. Automorphic forms for G over F are
integrated against shtukas, viewed as compactly supported
cohomology classes on the shtuka stack.
The chain over function fields:
automorphic representation <--> cohomology class on a
shtuka moduli stack
<--> Frobenius eigenvalues
<--> bounds via Weil II.
Weil II gives bounds on Frobenius eigenvalues dimension-by-dimension;
the automorphic form's Satake parameters appear among those
eigenvalues. That is how the diagonal bound |alpha_i|=1 emerges
from the WEIL II purity statement, without needing any arithmetic.
Two spectacular theorems.
Theorem (Drinfeld 1974). Let F be a function field. Then every
cuspidal automorphic representation of GL(2, A_F) is tempered at
every unramified place.
Reference: V. G. Drinfeld, "Elliptic modules" (Russian), Mat. Sb.
94 (1974), 594-627; English Math. USSR Sb. 23 (1974), 561-592.
Actually the Ramanujan statement is in Drinfeld's paper "The proof
of Langlands' conjecture for GL(2) over functional fields",
Proceedings of the ICM 1978, vol. 2, 565-574.
Idiom. Drinfeld's proof was via a rather technical theory of
"elliptic modules" (now called Drinfeld modules) plus the
Langlands correspondence for GL(2) over function fields. The
Langlands correspondence is exactly
{l-adic Galois representations of pi_1(F) } <--> {cuspidal
automorphic
reps of GL(2)}
with local-global compatibility. Under this equivalence
temperedness becomes purity, which Drinfeld verified.
Theorem (L. Lafforgue 2002). The Langlands correspondence for
GL(n) over F holds for every n >= 1. Consequently GRC for GL(n)
over F holds: every cuspidal automorphic representation of GL(n,
A_F) is tempered at every unramified place.
Reference: Laurent Lafforgue, "Chtoucas de Drinfeld et
correspondance de Langlands", Inventiones Math. 147 (2002), 1-241.
This was Lafforgue's Fields Medal work. It also uses the
shtuka cohomology + Weil II machine, but now with the full
technology of perverse sheaves, the decomposition theorem, and
the L.afforgue-V. Lafforgue generalization of Drinfeld's
"excursion operators".
The local-global compatibility: for each closed point x of X, the
semisimple Satake parameter of pi_x is the semisimple part of the
Frobenius sigma_x in the Galois representation attached to pi. So
temperedness at x translates to "Frobenius has all eigenvalues of
absolute value 1 under normalized Weil-pairing", which is
automatically true if the representation is pure. And it is pure
because it is an irreducible constituent of the cohomology of a
shtuka moduli, on which Weil II applies (subject to bounds on
weights by the "support theorem" arguments of Lafforgue).
Why can't we just repeat this for general G? Three obstructions:
(i) For G != GL(n), a cuspidal automorphic representation does
NOT correspond to a single Galois representation but to a
more complicated object - a "global Langlands parameter",
which is a homomorphism pi_1(U) --> G^hat(Qbar_l) for some
open curve U subset X (V. Lafforgue, "Introduction to
chtoucas for reductive groups", arXiv:1404.6416). Proving
that this parameterization holds, and then going from the
parameter to temperedness of the corresponding automorphic
form, is harder.
(ii) Groups with interesting semantics: even when the
parameterization is settled (split reductive, V. Lafforgue
2014), the parameter can be ANY semisimple
homomorphism. It can be highly non-pure: it can have
Frobenius eigenvalues with nontrivial absolute values.
That non-purity is exactly the "non-tempered" part of the
cuspidal spectrum.
(iii) Non-tempered cuspidal forms. There exist genuine
cuspidal representations on general G (not GL(n)) that are
NOT tempered, i.e. their Satake parameters have eigenvalues
with absolute value != 1. Those are predicted by Arthur's
conjectures. The sub-question "which non-tempered forms
exist, and what do their Satake parameters look like?" is
the Arthur conjecture. It is a definite conjecture with a
definite answer, and the answer uses SL_2-subgroups of the
dual group.
So the function-field landscape is:
GL(n): GRC holds (Drinfeld, L. Lafforgue).
General G: GRC FAILS, but Arthur predicts exactly which
non-tempered forms exist.
That prediction is the Arthur SL_2 picture.
Sections 4-9 explain Arthur's picture, what the original file
claims about it, what is actually known, and why the claimed
theorem is a big open problem.
======================================================================
SECTION 4. ARTHUR'S SL_2 PICTURE AND NILPOTENT ORBITS
======================================================================
Section 3 ended with a promise: GRC fails for general G, but
Arthur predicts exactly which non-tempered forms exist. This
section says how, in enough detail that Section 5 can state the
GLR conjecture that the original file purports to prove.
The main idea. James Arthur (Asterisque 171-172, 1989) proposed
that the cuspidal spectrum of a reductive group G should be
organised by a larger parameter space than the one for GL(n).
Specifically, for any reductive G there should exist parameters of
the form
psi : L_F x SL(2, C) --> G^hat
(here L_F is the Langlands group of F; over function fields you
may think of "continuous homomorphisms from pi_1(U) to G^hat
compatible with Frobenius"; over number fields it is conjectural).
The extra SL(2,C) factor is Arthur's innovation. It is the
"non-temperedness dial": SL(2,C) carries a 2-dimensional
representation of weight (1, -1) at the diagonal torus, and
evaluating it on the element diag(q^{1/2}, q^{-1/2}) produces
the non-unit absolute values q^{1/2} that appear in non-tempered
Satake parameters.
For any such psi, let O = O(psi | SL(2,C)) be the nilpotent orbit
of the image of the SL(2,C)-part, i.e. the nilpotent element
d psi( 0 1 )
( 0 0 )
in Lie(G^hat). Every Jacobson-Morozov triple comes from such a
psi, and vice versa. So Arthur's parameters are in bijection with
pairs (a "Langlands parameter" for the reductive centralizer of
the JM homomorphism SL(2) --> G^hat, plus a nilpotent orbit).
The predicted shape of the Satake parameters. If pi is a member
of the Arthur packet attached to psi, then for every unramified
closed point x with residue cardinality q_x, the Satake class
[t_x] of pi_x lies in
R_{q_x, O} = { phi_O diag(q_x^{1/2}, q_x^{-1/2}) c :
c runs over "compact elements"
of Z_G^hat(phi_O(SL(2,C))) }
where "compact" means: after every complex embedding, conjugate
into a maximal compact subgroup of the centralizer. Equivalently,
c has all its eigenvalues of modulus 1 at every complex
embedding. So the non-tempered part of [t_x] is exactly the
q_x^{1/2} factor coming from diag(q_x^{1/2}, q_x^{-1/2}) through
phi_O.
This is exactly the definition of R_{r, O} in the original file,
Equation (intro:orbit-spectrum). The original file did not
invent it; it copied the definition from Gaitsgory-Lafforgue-
Raskin Section 3.4.
Small examples to make it concrete.
Example: G = GL(2), G^hat = GL(2,C).
Nilpotent orbits in gl(2):
- {0} (the zero orbit)
- principal (the nonzero nilpotent orbit, i.e. all
of rank 0, i.e. nonzero matrices with
square zero)
For the zero orbit: phi_0 is trivial, R_{q, 0} = {compact
elements of GL(2,C)}, i.e. |alpha_1|=|alpha_2|=1. This is
the tempered set.
For the principal orbit: phi_princ is the diagonal embedding
SL(2,C) --> GL(2,C). Then R_{q, princ} = {phi_princ
diag(q^{1/2}, q^{-1/2}) c} for c in the centralizer. But the
centralizer of a regular SL(2)-image in GL(2,C) is just
C^*, so c is a scalar. So R_{q, princ} = { z *
diag(q^{1/2}, q^{-1/2}) : z in C^* }, i.e. the point with
Satake parameters q^{1/2} z and q^{-1/2} z. That is exactly
the "1-dimensional representation parabolically induced",
i.e. a weight-1 Eisenstein series, which is NOT cuspidal
but is the model non-tempered Arthur packet component.
Over GL(2), all cuspidal representations are tempered
(Drinfeld / Deligne), so only the zero orbit occurs in the
cuspidal spectrum. That recovers GRC for GL(2) as a special
case of Arthur's picture.
Example: G = Sp(4) (split symplectic rank 2), G^hat = SO(5,C).
Nilpotent orbits in so(5): {0}, minimal, trivial-rank (i.e.
long-root type), and principal - roughly 5 orbits in the
standard labelling [A1], [A1~], [A1+A1], [A2] etc.
The zero orbit gives tempered cuspidal forms on Sp(4).
A nonzero orbit, say the one whose phi_O has image a
principal SL(2) in an so(3) = sl(2) subalgebra, gives a
cuspidal Arthur packet whose members have Satake parameters
mixing a q^{1/2} factor into one "directions" of the 5-dim
standard representation. These are the famous "CAP" or
"Saito-Kurokawa-type" or "theta-type" lifts that are
eigenforms, are cuspidal, but are visibly non-tempered.
The point of Arthur's conjecture. Arthur conjectures that the
whole cuspidal spectrum decomposes (up to the packet membership
issue) as a direct sum over nilpotent orbits,
C_cusp(G) = bigoplus_{O in Nil(G^hat)} C_O(G),
where C_O(G) is the set of Arthur packet components whose Satake
parameters lie in R_{q_x, O} at every unramified x. Moreover the
summand C_O(G) "looks like" a tempered object attached to the
centralizer of O.
Full Arthur conjecture. Arthur actually conjectures much more:
endoscopic character identities, packet structure, the fact that
the summand C_O is a genuine automorphic object on a "smaller"
group, and a precise classification of the discrete spectrum.
The version just stated is the part he calls the "SL_2 part" of
the classification, and it is what Gaitsgory-Lafforgue-Raskin
turn into a specific spectral-support conjecture.
Status of Arthur's conjecture.
- Classical groups over number fields: proven by Arthur
(orthogonal/symplectic, "The endoscopic classification of
representations: orthogonal and symplectic groups", AMS 2013)
and by Mok ("Endoscopic classification of representations of
quasi-split unitary groups", Mem. AMS 2015).
- Classical groups over function fields: via Lafforgue's
excursion operators, follow-up papers by various people
(e.g. work by Arinkin, Bezrukavnikov, Chen, Chen-Yun, D.
Gaitsgory, Q. Ho, Y. Liu, S. Raskin, ...) - this is an
active area and some cases are done.
- Exceptional groups (G_2, F_4, E_6, E_7, E_8, 2E_6, 3D_4) over
any field: OPEN in full generality, though significant work
(especially by F. Shahidi via Rankin-Selberg on "real groups
and the Arthur conjecture" for G_2 / SO(7), and by various
others) covers substantial cases.
So "split connected adjoint absolutely simple" in the original
file is a killer scope: this includes G_2, F_4, E_6, E_7, E_8,
and 2E_6, 3D_4 as inner forms. None of those has a full Arthur
classification over function fields. So a theorem claiming to
prove the GLR spectral-support conjecture for ALL such groups at
full finite level is either field-medal-if-true or (more likely)
a hallucination. Given the style, it is the latter.
======================================================================
SECTION 5. WHAT GLR CONJECTURES 3.4.5 AND 3.4.6 ACTUALLY SAY
======================================================================
The original file's main theorem is a claim to prove "GLR
Conjectures 3.4.5 and 3.4.6 at every full finite level". Here is
what those conjectures say at a conceptual level, and what the
finite-level extension in the original file is going beyond GLR.
Set-up (matching the original file). Let X be a smooth projective
geometrically connected curve over F_q, F = F_q(X), and let G be
a split connected semisimple group over F_q. Let D be an
effective divisor on X ("the level"); its multiplicities can be
arbitrary. Write K_D for the "full level D" compact open subgroup
of G(A_F): the kernel of G(O_A) --> G(O_D). Write C_{D,Q} for
the space of compactly supported Q-valued cuspidal functions on
X_D := G(F) \ G(A_F) / K_D .
This is finite-dimensional (it is the cuspidal part of the space
of functions on a finite set: K_D is open compact so the double
quotient is discrete and locally finite). For each prime l not
equal to char(F_q), let C_{D,l} = Qbar_l (x)_Q C_{D,Q}.
For every unramified closed point x of X (i.e. any point not in
the support of D), each f in C_{D,Q} has a Fourier expansion
supported on the Satake parameters appearing in the Hecke
operators at x. For each nilpotent orbit O of Lie(G^hat), let
C_{D, O, Q} := { f in C_{D,Q} : for every x not in
supp(D), every Satake
parameter appearing in
A_x f lies in R_{q_x, O} }
where A_x is the spherical Hecke algebra at x and R_{q_x, O}
is as in Section 4. So C_{D,O,Q} is the subspace of the
cuspidal space whose Satake parameters at every good point
"see" only the nilpotent orbit O.
Conjecture 3.4.6 of GLR (informally). For every nilpotent orbit
O of Lie(G^hat), C_{D,O,Q} is a well-defined direct summand of
C_{D,Q}, and the natural addition map
+_{O} : bigoplus_O C_{D, O, Q} --> C_{D, Q}
is an isomorphism. That is: the whole cuspidal space decomposes
exactly into these nilpotent-orbit-indexed pieces, and every
cusp form belongs to exactly one piece.
Conjecture 3.4.5 of GLR. The same statement but in the
l-adic setting, using Qbar_l coefficients. GLR's Remark 3.4.7
identifies this with the l-adic Ramanujan-Arthur conjecture.
The point: this is a MUCH STRONGER statement than the Arthur
SL_2 conjecture for the *discrete* spectrum. Arthur's conjecture
is about the discrete automorphic spectrum up to a packet
structure; the GLR statement is a spectral-support decomposition
of the entire cuspidal space, and it does not even need the
"packet" bookkeeping because the R_{q_x, O} are pairwise
disjoint sets. So GLR 3.4.5/3.4.6 are what you get if you:
(i) accept Arthur's SL_2 prediction for the Satake parameters,
(ii) drop the packet-bookkeeping part (packet membership,
endoscopic character identities, etc.), and
(iii) ask for a strict equality of subspaces rather than a
"modulo packets" statement.
What the original file adds beyond GLR.
(a) Full finite level D, with arbitrary multiplicities. GLR
state their conjectures "at trivial level D = 0" only. The
original file claims a "finite-level extension": instead of
only D = 0, you can take any effective D, including
non-reduced D (multiplicities > 1 at some point). That
extension is not in GLR as far as I know. It may be a
natural consequence of GLR's result (if GLR were proven),
or it may be strictly harder. Either way, it is beyond the
original published scope of the GLR conjecture.
(b) Also claims to prove the COROLLARY: split connected
adjoint absolutely simple groups over function fields,
"one unramified generic local component at v => tempered
everywhere unramified". That is the unramified generic
Ramanujan conjecture. See Section 2's Reminder (ii).
Both are massive. In particular (b) would prove a case of
Ramanujan for G_2, F_4, E_6, E_7, E_8 over function fields,
which as far as I know is not currently known at that
generality. Ciubotaru-Harris (2023) have significant partial
results under extra hypotheses; Sawin-Templier (2021) have
characteristic > 2 results under monomial supercuspidal
assumptions; but neither gives what the original file claims.
Why the original file is not this. The original file's "proof"
runs about 200 pages of LaTeX. Here is what a proof of the GLR
conjecture at full finite level would actually require, at a
minimum:
1. A derived Satake equivalence with all of the enhanced
monoidal structure, in family over the curve, with the Weil
structures completely pinned down. That is roughly what
Bézrukavnikov-Finkelberg (2008) and Ben-Zvi-Sakellaridis-
Venkatesh (2024+) give for the spherical case, but the
"derived" and "in family" parts are quite technical.
2. A Godement-Jacquet / geometric Eisenstein theory matching
the categorical central action on the automorphic category
with the action on the spectral side. That is roughly
Gaitsgory's central-sheaves construction (2001) plus the
Wakimoto filtration (Arkhipov-Bezrukavnikov 2009,
Cass-van den Hove-Scholbach 2024). But the categorical
central action on automorphic sheaves is much less understood
than the spherical/derived-Satake side.
3. Boundedness of shtuka moduli with arbitrary prescribed
local positions at finitely many points. That requires
Harder-Narasimhan theory in positive characteristic plus
careful control of the Frobenius cycle. Schieder (2015)
gives the basic HN stratification, but the shtuka-specific
arguments here are not routine.
4. "Local-to-global" induction on semisimple rank, showing
that if you know temperedness for proper Levi subgroups you
get it for the group without extra hypotheses. That requires
the categorical "spherical dg model" being able to
distinguish tempered and non-tempered spectral support
cohomologically. This is the deepest part.
5. Absolute Galois / purity arguments, delicate enough to
compare "cohomological degrees" of compactly supported
cohomology of shtukas with Satake parameters of cuspidal
automorphic forms. That is the machinery that L. Lafforgue
and V. Lafforgue developed for GL(n) and split reductive
groups respectively, but the comparison must now be uniform
in the "level", and GLR's conjectures demand a strict
spectral-support decomposition, not merely tempering at
good places.
6. Finally, an induction on semisimple rank that ties all
the above together into a size argument: "if a joint Hecke
character appears in the discrete cuspidal spectrum with an
absolute-value exponent bigger than the tempered one by a
fixed amount delta > 0, then a certain compactly supported
cohomology class would have to have degree simultaneously
>= c_1 + delta m and <= c_2 + (something smaller) * m, a
contradiction for m large".
All of this is genuinely needed for a proof. The original file
does have paragraphs corresponding to all of it, but they
collapse into 1-2 pages each in a way that papers of this
difficulty do not. In every one of the deep steps there is a
"for completeness, here is why ..." paragraph that assumes
precisely the technical thing that the deep step is supposed to
prove. This is what a plausible-looking but non-actual proof
looks like when written by a language model.
None of the above is a claim that the *conjecture* is wrong.
The GLR conjectures are well-posed, well-motivated, and widely
believed. The claim of the original file is that they have been
*proven* (and strengthened to full finite level and made to imply
unramified generic Ramanujan for all split connected adjoint
absolutely simple groups). That claim is the thing that is not
credible.
======================================================================
SECTION 6. WHAT IS ACTUALLY PROVEN (STATUS SUMMARY)
======================================================================
This is the most useful section of the essay. It is a list of
what is actually a theorem about Ramanujan-type boundedness and
Arthur-type decomposition, by group and by field, with
references. Consult this whenever somebody (me or otherwise)
says "the conjecture is proven for X".
Number fields (characteristic 0).
GL(1)/Q. GRC holds trivially (all Hecke characters have |chi_p|=1).
Reference: elementary.
GL(2)/Q. GRC holds.
Deligne, "La conjecture de Weil I", Publ. IHES 43 (1974).
GL(n)/Q, n>=3. GRC open. Symmetric-power and functoriality
results give partial bounds (e.g. Kim-Shahidi, Kim-Sarnak
-> alpha+beta <= p^{7/64} for GL(2)), but NOT the full |a_p|=1
bound. Clozel-Harris-Taylor, and more recent automorphy
lifting, only partially constrain.
Classical groups over Q (orthogonal/symplectic). Arthur's
classification (AMS 2013) proves the Arthur SL_2 decomposition
for the discrete spectrum, not a full spectral-support
decomposition a la GLR. GRC for cuspidal generic is expected
as a consequence.
Unitary groups over Q. Mok, Mem. AMS 2015.
Exceptional groups over number fields. OPEN.
Split reductive groups over number fields, general. OPEN. This
is where the "endoscopic Langlands functoriality for general
reductive groups" is itself unknown, not just the Ramanujan
bound.
Function fields (characteristic p > 0, residue field F_q).
GL(2)/F. GRC holds.
Drinfeld, ICM 1978, "The proof of Langlands' conjecture for
GL(2) over functional fields".
GL(n)/F, n>=1. GRC holds.
L. Lafforgue, "Chtoucas de Drinfeld et correspondance de
Langlands", Inventiones Math. 147 (2002), 1-241.
Split reductive / general reductive / G a Chevalley group,
parameterization side. V. Lafforgue's program:
- "Chtoucas pour les groupes réductifs et paramétrisation
de Langlands globale", arXiv:1209.5352, JAMS 2018.
- "Introduction to chtoucas for reductive groups and to the
global Langlands parameterization", arXiv:1404.6416.
This gives a global Langlands parameterization for split
reductive groups, but does NOT immediately give the Arthur
SL_2 picture, and does NOT give a full finite-level
spectral-support decomposition.
Classical groups over function fields, temperedness /
Arthur-type. Partial. Arinkin-Bezrukavnikov, Chen, Gaitsgory,
Ho, Raskin, and others have built large parts of the "geometric
Eisenstein" / "Whittaker" / "soergel bimodules" machinery.
Cases like SL(2), SL(3), split classical, are done or are
reachable. Exceptional groups are open.
Split connected adjoint absolutely simple groups over
function fields: generic Ramanujan with one generic unramified
local component -> tempered everywhere unramified. OPEN in
the generality of the original file. The best results are:
- Sawin-Templier (JAMS 2021): characteristic > 2, monomial
geometric supercuspidality hypotheses, plus a tempered-place
hypothesis. Their Theorem 11.7 has a "tempered at one
unramified place => tempered everywhere" statement, but
only for cuspidal reps on split semisimple groups.
- Ciubotaru-Harris (arXiv:2311.15300, 2023): complementary
series indexed by nilpotent orbits, generalized Ramanujan
under extra local assumptions via excursion parameters,
Frobenius weights, and the spherical unitary dual. Their
Theorem 5.4 gives generic-Ramanujan-type results but with
a tempered-place hypothesis; their Corollary 6.32 gives a
local criterion that a Satake parameter with one
half-integral geometric direction and a genericity
hypothesis is tempered.
Neither gives the unconditional, every-split-connected-adjoint-
absolutely-simple, one generic unramified component result
claimed by the original file.
GLR 3.4.5 / 3.4.6 as a *conjectured* spectral-support
decomposition at full finite level. THERE IS NO PUBLISHED
PROOF TO DATE that I am aware of.
Classification of the 5 biggest groups.
GL(n): proven Ramanujan over function fields (L. Lafforgue).
NOT proven over Q for n >= 3.
Sp(2n): Ramanujan proved for tempered-but-not-necessarily-generic
over Q via Arthur; over F_q due to Arthur + Lafforgue +
geometric Arthur in progress.
SO(2n+1), SO(2n): same status as Sp (they are dual). Arthur 2013.
Unitary U(n): Mok 2015 over Q; partial over F_q.
G_2, F_4, E_6, E_7, E_8: OPEN everywhere in full generality.
Shahidi and various have partial results.
The shape of "the truth" (conjecturally). The GLR conjectures
state this cleanly and are widely believed:
Every cuspidal automorphic form is in exactly one GLR summand
C_{D, O, Q}, indexed by a nilpotent orbit O of Lie(G^hat).
The summand C_{D, {0}, Q} is the tempered part; it contains
the cuspidal representations whose Satake parameters lie in a
maximal compact at every good point.
The shape of what is currently proven is: "for G = GL(n) with
any n >= 1 over a function field, the truth is: all cuspidal
representations are tempered, and the entire cuspidal space sits
in the {0}-summand." For any other G, only partial statements are
available, mostly structural (parameterization, temperedness at
*individual* places under genericity), not the full spectral
decomposition.
======================================================================
SECTION 7. WHY IS THIS GENUINELY HARD?
======================================================================
There are two ways to think about "why hard":
- Hard for a mathematician (what technical obstacle is in the way)
- Hard for the author of the original file (what they had to
fake)
Both matter. This section is mostly the first, with an occasional
nod to the second.
Obstacle 1. No shtuka cohomology for exceptional groups over
functions fields (that generalizes L. Lafforgue to all reductive
groups).
Over function fields, temperedness of a cuspidal automorphic
representation of GL(n) is a Frobenius-bound statement. You
realize it as an eigenvalue of a Frobenius on a cohomology
group of a specific geometric object (a Drinfeld shtuka moduli).
Weil II gives the bound.
For a general reductive group, L. Lafforgue's argument has to
be redone with the following changes:
- the cohomology groups are now on moduli of G-bundles with
shtuka structures, not GL(n)-bundles;
- the Hecke operators are not the "simple" ones (which are
actually already complicated);
- Frobenius eigenvalues no longer come out in an ordered
list of size n; they come out as semisimple conjugacy
classes in G^hat.
This is exactly what V. Lafforgue's program does, and it is
substantially more technical than the GL(n) case. For
exceptional groups, no one has run the argument.
Obstacle 2. No Arthur classification for general reductive
groups.
Arthur's classification is what gives you the SL(2) parameter
psi : L_F x SL(2,C) --> G^hat. It is proved for classical groups
over number fields (Arthur, Mok), and is in progress over
function fields. It is NOT proved for G_2, F_4, E_6, E_7,
E_8 in any setting. Without Arthur's classification, you do
not even know what the "right" assignment of cuspidal forms to
nilpotent orbits is, so the original file's statement is
in some sense assuming the answer.
Obstacle 3. Even with Arthur, GLR is finer.
Arthur's classification is about the DISCRETE part of the
spectrum (the part that contributes to L^2 in the discrete
sense) up to packet bookkeeping. GLR is about the entire
CUSPIDAL space and asks for a strict direct-sum decomposition
at every full finite level. The precise difference:
- Arthur assigns each DISCRETE automorphic form to a packet
(packet members are not distinguishable Satake-wise).
- GLR assigns each CUSPIDAL function to a summand C_{D, O}.
- Given the fact that R_{q_x, O} are open/closed conditions and
pairwise disjoint, the GLR summand is basically determined
by the Satake parameters at every good unramified point,
but this requires an argument (the "support" argument) that
the assignment is stable under all Hecke operators at
every place. That is not given by Arthur for free.
Obstacle 4. The finite-level extension.
Both Arthur and GLR are stated at TRIVIAL level (D = 0). The
original file claims a "full finite level" version. Making the
transition from level 0 to full level involves:
- Hecke operators at infinitely many places, all commuting;
- the failure of strong approximation for non-simply-connected
groups, so the "Satake parameter at x" isn't just a single
conjugacy class but a class modulo the component group of
the center;
- the level structure at x itself mixing with the local
Satake parameter at x via the Iwahori Hecke algebra.
This can be done cleanly in principle. But it involves careful
combinatorial work (the Iwahori-Hecke algebra at level x is
large) and interlocking spectral-support conditions. The
original file's proofs of this happen in several short passages
that hand-wave the "as finite level grows inside a fixed
principal congruence subgroup, the Iwahori structure
degenerates consistently" part.
Obstacle 5. Generic-implies-tempered.
The Corollary (generic Ramanujan) is a specific "one place
implies everywhere" statement. Getting it from the GLR
decomposition is NOT automatic. You have to know that:
- the specific *generic* local component at v has "geometric
Satake parameter" (which it does at many places,
essentially because unitary + a single algebraic Satake
direction at v forces polytopic structure);
- Ciubotaru-Harris Corollary 6.32(1) is a local criterion, so
you can apply it at v, get temperedness at v;
- temperedness at v (a single place) propagates to temperedness
at all unramified places. This second step is the "common
orbit" argument from GLR, and it requires the same
spectral-support decomposition.
In a real proof, the propagation step is exactly where the hard
work lives over function fields. In the original file it is
paragraph (Corollary~\ref{main:generic-ramanujan} proof), which
is short and hand-wavy.
Obstacle 6. The "why does the paper's proof not work" question.
Read the original file's Section 5, "Extraction from the discrete
spectrum". That is the section that tries to prove that a
dominant ordered Bernstein weight has compact times Jacobson-
Morozov form. Right in the middle of that proof, paragraph
after paragraph, there is one hand-waving step after another.
For example (paraphrasing):
"The Wakimoto filtration isolates the repeated highest-weight
term, and hyperspecial projections transfer its high-degree
cohomology to the spherical calculation. A separate
spherical dg model supplies an actual chain projector onto
this summand. A finite Koszul test preserves its highest
nonzero cohomological degree, which the Frobenius lower bound
and Theorem ... then bound incompatibly as m grows."
Every single sentence there is a paper's worth of work. If they
were all correct, this would be roughly equivalent to the 200
pages of L. Lafforgue's Inventiones paper, minus the Godement-
Jacquet / trace-formula machinery. If they were all correct. In
the actual file they are written in a style that quietly
collapses each into one paragraph. That is the fundamental
problem. It is not that any specific step is wrong; it is that
no specific step is given enough space to actually be right.
Obstacle 7. The style of the entire paper.
Every single proof in the original file has at least one
paragraph that starts "For completeness, here is why the thing
we are trying to prove is true." Real papers do NOT do this.
Real papers say "we prove X by constructing Y and then applying
Z", and then they either do it or the reader notices they
didn't. The other tell is that several long proofs put their
hardest step inside a "smallskip \noindent\emph{The ...}"
sub-heading with a label that could easily be turned off. These
are structural tells of LLM-generated mathematical text: hand-
waves that hide inside well-formatted sub-headings.
Conclusion. It is not that the original file is trivially wrong
at any specific step. It is that at each of the steps where a
real proof would have to spend 10-50 pages of technical
machinery, the file spends between half a page and two pages
and claims everything works. There are, in a 200-page file,
at least a dozen such choices. That is what a hallucination
looks like.
======================================================================
SECTION 8. WHAT IS THE CURRENT STATE OF THE ART FOR
GENERIC RAMANUJAN OVER FUNCTION FIELDS?
======================================================================
Section 6 listed the two main partial results touching the
central claim of the original file: Sawin-Templier (2021) and
Ciubotaru-Harris (2023). This section gives a bit more depth so
that if you (or your kids) wonder "well how close are we?", you
know the real answer.
Both papers are serious mathematics. Neither proves the full
result. Here is what they actually get.
Sawin-Templier, "On the Ramanujan conjecture for automorphic
forms over function fields I. Geometry", JAMS 34 (2021),
653-746. arXiv:1805.12231v4.
The main theorem (their Theorem 1.1) is roughly:
Let G be a split semisimple group over a function field F_q(X),
and let pi be a cuspidal automorphic representation of G(A_F).
Suppose:
(i) char F_q > 2 (their geometric machinery needs this);
(ii) pi is "monomial geometric supercuspidal" at every ramified
place (this is a strong hypothesis, essentially asking that
the local representation is induced from a character of a
finite-index subgroup);
(iii) pi is tempered at one unramified place.
Then pi is tempered at every unramified place.
Their Theorem 11.7 (which the original file cites) is the "one
tempered place implies all tempered places" part. The key point is
that (iii) is a nontrivial hypothesis: you cannot drop it and still
get the conclusion. Ciubotaru-Harris use a similar conditional
structure.
The basic technique: Sawin-Templier use excursion operators
following V. Lafforgue's framework, plus the ARTHUR SL(2) structure
to control the failure of temperedness by nilpotent orbits. Their
statement "tempered at one place implies tempered at all places"
is essentially because the "orbit assignment" from Arthur's picture
is stable under Galois and under places.
Ciubotaru-Harris, "On the generalized Ramanujan and Arthur
conjectures over function fields", arXiv:2311.15300v1 (2023).
The main theorem (their Theorem 1.11) is roughly:
Same setting. Under additional hypotheses on the local components
at ramified places (essentially requiring enough structure to
apply their local criterion), a cuspidal automorphic
representation of a split reductive group over F_q(X) is
tempered at all unramified places.
Their Theorem 5.4 is a "generic-implies-tempered" statement, but
it has a hypothesis that one place is already known to be
tempered. The proof goes through local complementary series
theory and a global propagation argument.
Their Corollary 6.32(1) gives a very useful local criterion:
if a complex Satake parameter of an unramified spherical unitary
irreducible is "geometric" (Satake parameters are in the image of
a geometric direction, i.e. half-integral in the characters),
then it is tempered. This is exactly the local criterion
applied in the original file's generic-Ramanujan corollary.
What neither paper does: eliminate the "tempered-place hypothesis"
or the "ramified-local hypothesis" or the "char > 2" assumption.
The original file's Claim 2 (generic Ramanujan for split
connected adjoint absolutely simple groups at full generality)
is strictly stronger than both.
Realistic assessment.
- For G = GL(n) over F_q(X): Ramanujan is a THEOREM
(L. Lafforgue 2002). Nothing to prove.
- For G a split classical group over F_q(X) with q > 2 (say)
and mild local conditions: partially known via Ciubotaru-
Harris, Sawin-Templier, and the geometric-Arthur-in-progress
from Arinkin-Bezrukavnikov-Gaitsgory-Ho-Raskin and others.
- For G an exceptional group (G_2, F_4, E_6, E_7, E_8, etc.) over
any field: UNKNOWN at the generality of the claim.
- For "absolutely simple adjoint" not-necessarily-classical
groups: same as exceptional.
- For "full finite level D" in the GLR sense: not even
formulated in most of the literature, though it is expected
to reduce to D = 0 by standard Hecke-algebra manipulations
(once D = 0 is understood).
That last point is important. A reasonable conejcture: if GLR
3.4.5/3.4.6 are proven for D = 0 (i.e. the trivial level), then
the full finite-level version follows by the Iwahori-Hecke-algebra
machinery of the original paper's Section 2 (which is roughly
what the paper actually does correctly, at least in shape). But
nobody has proven 3.4.5/3.4.6 at D = 0, so this reduction is
being run on top of a result that doesn't exist.
So: "how close are we?" The honest answer is:
- GL(n): done.
- Classical groups over function fields: most of the geometric
machinery is in place, but a full local-global decomposition
at unconditional level is not yet assembled.
- Exceptional groups: essentially not started for this problem.
- General split adjoint: essentially not started.
- The "one generic unramified place -> tempered everywhere"
version at full generality: requires solving the GLR
conjectures or something equivalent.
The original file claims to have done all of this. It has not.