The coefficients tau(n) of the modular discriminant
Delta(z) = q prod_{n>=1} (1 - q^n)^24, q = e^{2 pi i z},
are the subject of one of the oldest and most consequential
observations in arithmetic. Their size at primes governs the
size of every tau(n) through multiplicativity and the prime-power
recursion, and through the Euler product governs the analytic
behavior of an entire L-function. The story begins in 1916,
passes through Hecke and Deligne to etale cohomology and the
Weil conjectures, and reaches its general-group form only
conjecturally, where it becomes local temperedness of cuspidal
representations and, more finely, the nilpotent-orbit
decomposition conjectured for function fields at every finite
level.
Ramanujan, "On certain arithmetical functions", Transactions of
the Cambridge Philosophical Society 22 (1916), 159-184.
papers/ramanujan-1916.txt
Expansions of Delta into a power series in q define a sequence
of integers tau(1), tau(2), tau(3), ...
The first few terms, from a direct multiplication:
tau(1) = 1
tau(2) = -24
tau(3) = 252
tau(4) = -1472
tau(5) = 4830
tau(6) = -6048
tau(7) = -16744
tau(8) = 84480
tau(9) = -113643
tau(10) = -115920
tau(11) = 534612
tau(12) = -370944
tau(13) = -577738
You can compute these yourself: multiply out the infinite
product through the q^13 term, kill all (1 - q^n) for n > 13,
and collect. It is tedious but mechanical. In Python you can do
it in three lines with the sympy library's modular forms module,
or by hand in an hour.
The resulting integers grow roughly like p^{11/2}, for the
following reason. Delta is the unique normalized cusp form
of weight 12 for SL(2,Z): the space S_12(SL(2,Z)) is 1-
dimensional, spanned by Delta. A weight-k modular form grows
like its "leading term" tau(p) p^{-k+1/2} in a very specific
sense; the natural scale of tau(p) is p^{k-1/2} = p^{11/2}.
Ramanujan computed the first five thousand or so tau(p) by
hand. From the data he noticed three facts, which he guessed
hold for all n:
The first observation is multiplicativity. For every coprime
pair m, n,
tau(m) * tau(n) = tau(m * n).
So the whole sequence is determined by its values at primes.
The second observation is a prime-power recursion. For every
prime p and every r >= 1,
tau(p^{r+1}) = tau(p) * tau(p^r) - p^11 * tau(p^{r-1}).
Starting from tau(1) = 1 and tau(p), this determines tau(p^r)
for every r by induction: it is a linear recurrence with
constant coefficients (p^11) and initial value tau(p). So the
whole sequence tau(n) is determined by the prime values.
The third observation is a bound. For every prime p,
|tau(p)| <= 2 * p^{11/2}.
This is the original Ramanujan conjecture.
The small cases verify directly. The comparison needs:
p |tau(p)| 2*p^{11/2} ratio
---- ------------ ------------ -----
2 24 90.5 0.265
3 252 842 0.299
5 4830 13975 0.346
7 16744 88934 0.188
11 534612 1068291 0.500
13 577738 2677432 0.216
The bound holds everywhere; Ramanujan observed that equality
had never been hit. He conjectured that it never is, i.e. the
inequality is strict, which is consistent with everything
known. (Sato-Tate, now a theorem for Delta: Barnet-Lamb,
Geraghty, Harris, Taylor, 2011, predicts the distribution of
the ratios and says they concentrate around E|ratio| = 4/pi,
i.e. tau(p) is generally much smaller than the bound. So
Ramanujan's inequality is very loose for typical p; it is a
worst-case statement.)
Define
a_p = tau(p) / p^{11/2}.
Then Ramanujan's bound becomes |a_p| <= 2 for every prime p.
Here is why this is the exact form of the conjecture, not just
one of many possible bounds.
For a prime p, consider the local Euler factor of the L-function
of Delta at p. It is, by definition,
L_p(Delta, s) = 1 / ((1 - alpha_p p^{-s})(1 - beta_p p^{-s})),
where alpha_p, beta_p are the "Satake parameters" of the local
representation pi_p of GL(2, Q_p) attached to Delta. The
association is through the Ramanujan-Langlands correspondence:
Delta is an automorphic form for GL(2)/Q; its local component
at p is an unramified spherical representation of GL(2, Q_p);
such a representation is determined by the conjugacy class of
a diagonal matrix diag(alpha_p, beta_p) in GL(2, C), and the
trace is a_p.
Concretely, the local L-factor of Delta is known
unconditionally to be
L_p(Delta, s) = 1 / (1 - a_p p^{-s} + p^{1-2s}).
Solving: the roots of the denominator are
z = alpha_p p^{-s}, z = beta_p p^{-s},
where alpha_p + beta_p = a_p and alpha_p * beta_p = 1.
So alpha_p, beta_p are roots of the quadratic
T^2 - a_p T + 1 = 0.
Case analysis:
(i) If |a_p| > 2, the two roots are real and distinct and
at least one is > 1 in absolute value. Then L_p(Delta,
s) has a pole at Re(s) > 1/2 on the real axis.
(ii) If |a_p| = 2, the roots are equal and both have absolute
value 1. Then L_p(Delta, s) has a double pole on
Re(s) = 1/2 (specifically at s = 1/2 + 11/2, but this
case is excluded by the conjecture being strict).
(iii) If |a_p| < 2 (strictly), the roots are complex
conjugates on the unit circle: |alpha_p| = |beta_p| = 1.
Then L_p(Delta, s) is holomorphic and nonzero for
Re(s) > 1/2 and has no poles on the whole complex plane
except the standard ones. This is the "as nice as
possible" case.
Ramanujan's conjecture |a_p| <= 2 (with equality being the
degenerate double root case) is exactly the statement:
Every local L-factor of Delta has all its zeros and
poles off the line Re(s) = 1/2.
This is the entire content of Ramanujan's conjecture. It is a
statement about the analytic behaviour of the local Euler
factors of an automorphic L-function, not (directly) about the
sizes of the integers tau(p). It happens to be equivalent to
the integer bound by the arithmetic of the quadratics.
Why the base field matters here. Over Q, the Ramanujan
conjecture for Delta is Deligne's theorem (1974); its proof
appears below. Over other fields, it is not always a theorem. And in the
general-group setting, |a_p| <= 2 generalizes to a
"unitarity" condition on conjugation classes in the dual group,
which is where Arthur's SL_2 picture enters. All of that is
below.
Of Ramanujan's three facts, two become theorems by Hecke's
1937 work, with short elementary proofs. The third (the bound
itself) is Deligne 1974; its structure appears later. Here are the two complete proofs.
papers/hecke-1937.pdf
papers/deligne-weil-1.pdf
One input is needed from the classical theory of modular
forms:
PROPOSITION (structure of small-weight modular forms).
For k >= 0 even, the vector space M_k(SL(2,Z)) of
holomorphic modular forms of weight k on SL(2,Z) has,
for k < 12, dimension 0 (if k = 0, 1 modulo 12 and k < 0)
or 1 (if k = 4, 6, 8, 10). In particular:
dim S_12(SL(2,Z)) = 1.
The unique element of S_12(SL(2,Z)) with leading
coefficient 1 is Delta.
This is a standard computation in the general theory of
modular curves; the classical reference is Serre's "A Course
in Arithmetic" or Diamond-Shurman's "A First Course in Modular
Forms". It is taken as input.
The first fact, multiplicativity, was proved by Hecke in 1937
and admits a short complete proof:
THEOREM (Hecke, 1937). For coprime m, n,
tau(m) * tau(n) = tau(mn).
PROOF. For each prime p and each integer r >= 1, there is
the Hecke operator T_p on M_k(SL(2,Z)), defined by its
action on the q-expansion:
(T_p f)(z) = p^{k-1} f(pz)
+ (1/p) sum_{b mod p} f((z + b)/p).
Three facts about T_p:
- T_p preserves M_k(SL(2,Z)).
- T_p preserves the subspace S_k(SL(2,Z)) of cusp forms.
- T_p T_q = T_q T_p for distinct primes p, q.
- T_m T_n = T_{mn} for coprime m, n.
Apply T_p to Delta. Since dim S_12(SL(2,Z)) = 1 and
T_p preserves it, T_p Delta is a scalar multiple of Delta.
The scalar is exactly tau(p): the constant term
coefficient in T_p Delta's q-expansion is p^{11} * (leading)
+ (1/p) * p * tau(p) = tau(p) * (leading) = tau(p), after
normalising (leading coefficient of Delta = 1).
So T_p Delta = tau(p) Delta. Similarly T_{p^r} Delta =
tau(p^r) Delta for any prime power p^r; and more generally
T_n Delta = tau(n) Delta for every n >= 1. (The general
case follows because T_n for n = p_1^{r_1} ... p_k^{r_k}
is a product T_{p_1^{r_1}} ... T_{p_k^{r_k}} and the
T_{p_i^{r_i}} commute.)
Now for coprime m, n:
tau(mn) Delta = T_{mn} Delta = T_m T_n Delta
= T_m (tau(n) Delta)
= tau(n) T_m Delta
= tau(n) tau(m) Delta.
Since Delta is nonzero, tau(mn) = tau(m) tau(n). []
The second fact, the prime-power recursion, is also due to
Hecke, in the same 1937 paper:
THEOREM (Hecke, 1937). For any prime p and any integer
r >= 1:
tau(p^{r+1}) = tau(p) tau(p^r) - p^11 tau(p^{r-1}).
PROOF. One identity in the Hecke algebra is needed:
T_{p^{r+1}} = T_p T_{p^r} - p^{k-1} T_{p^{r-1}}
(for any r >= 1, weight k).
This is a standard calculation in the double-coset
algebra of GL(2, Z_p); see for example Serre, Course in
Arithmetic, VII.5. The bare outline: T_p is
the double coset of diag(p, 1); T_{p^r} is the double coset
of diag(p^r, 1); the product T_p T_{p^r} is a sum of
T_{p^{r+1}} and p^{k-1} T_{p^{r-1}}, and rearrangement
gives the displayed identity.
Apply both sides to Delta. Both sides are in S_12(SL(2,Z)),
which is 1-dimensional and spanned by Delta. So both act
as scalars, and eigenvalues compare:
T_{p^{r+1}} Delta = tau(p^{r+1}) Delta,
T_p T_{p^r} Delta = tau(p) tau(p^r) Delta,
p^{k-1} T_{p^{r-1}} Delta = p^{11} tau(p^{r-1}) Delta,
with k = 12.
So
tau(p^{r+1}) = tau(p) tau(p^r) - p^11 tau(p^{r-1}). []
Sanity checks:
p = 2, r = 1: tau(4) = tau(2)^2 - 2^11 tau(1)
= (-24)^2 - 2048
= 576 - 2048
= -1472. Correct.
p = 2, r = 2: tau(8) = tau(2) tau(4) - 2^11 tau(2)
= (-24)(-1472) - 2048(-24)
= 35328 + 49152
= 84480. Correct.
p = 2, r = 3: tau(16) = tau(2) tau(8) - 2^11 tau(4)
= (-24)(84480) - 2048(-1472)
= -2027520 + 3014656
= 987136.
And tau(16) does equal 987136 in the tables. So the recurrence
and multiplicativity together generate the entire table from
the prime values.
The third fact, the bound itself, is Deligne 1974.
The bound |tau(p)| <= 2 p^{11/2} is NOT elementary. It is
a consequence of the full Weil conjectures (Deligne's
"La conjecture de Weil I", 1974). The full proof does
not fit here; the sketch of Deligne's argument below is the
tightest that can be offered.
Partial results ARE elementary. For instance, from the
definition of tau via the q-expansion you get
|tau(n)| <= n^{11/2} * d(n) * C
for some absolute constant C and the divisor count d(n),
which is much weaker than Ramanujan's bound. Ramanujan's
bound is, roughly speaking, "as small as possible given
what can be said from the q-expansion". The gap between
elementary bounds and the sharp bound is exactly the hard
part of the conjecture.
The word "tempered" comes from harmonic analysis: a unitary
representation of a locally compact group is tempered if its
matrix coefficients decay at infinity, equivalently if it
appears in the Plancherel decomposition without any exponential
factor. For the groups in this story, temperedness of a local
spherical representation translates into a condition on the
Satake parameter. Three passes, from concrete to abstract.
The first pass is GL(1), the base case.
Let chi be a Hecke character of Q, i.e. an automorphic
character of GL(1) over Q. Its local component chi_p at any
unramified prime p is a complex number of absolute value 1:
chi_p in U(1) subset C^* = GL(1, C). So every chi_p is
"tempered" trivially, and "tempered" for GL(1) says nothing.
This is the base case.
The second pass is GL(n), the general linear case.
Let pi be a cuspidal automorphic representation of GL(n)/Q.
At each unramified prime p, pi determines n complex numbers
alpha_1(p), ..., alpha_n(p), the "Satake parameters" of pi_p.
The local L-factor is
L_p(pi, s) = prod_{i=1}^n 1 / (1 - alpha_i(p) p^{-s}).
Call pi tempered at p if
|alpha_i(p)| = 1 for all i = 1, ..., n.
Equivalently, if diag(alpha_1(p), ..., alpha_n(p)) in GL(n, C)
is the Satake conjugacy class s_p(pi) of pi_p, then pi is
tempered at p iff s_p(pi) can be conjugated into the maximal
compact U(n) subset GL(n, C).
For n = 2 and pi = pi(Delta), the values alpha_p, beta_p are
the Satake parameters, with alpha_p beta_p = 1 and a_p =
alpha_p + beta_p. The condition |alpha_p| = |beta_p| = 1 is
exactly |a_p| <= 2. So:
Deligne's theorem = "GL(2) over Q is tempered at every
unramified prime"
for the cuspidal automorphic form pi(Delta).
The third pass is a general connected reductive group.
Let G be a connected reductive group over Q or F_q(X), with
Langlands dual group G^hat(C). At each unramified prime p,
the local component pi_p of a cuspidal automorphic pi
determines a semisimple conjugacy class s_p(pi) in G^hat(C)
(the Satake parameter). pi is tempered at p if s_p(pi) is
conjugate into a maximal compact subgroup K of G^hat(C).
G G^hat K
--------- -------------- ---------------
GL(n) GL(n, C) U(n)
SL(n) PGL(n, C) PU(n)
Sp(2n) SO(2n+1, C) SO(2n+1)
SO(2n+1) Sp(2n, C) Sp(2n)
SO(2n) SO(2n, C) SO(2n)
E_6 E_6(C) compact real form of E_6
etc.
For G = GL(n), K = U(n), recovering Pass 2.
The generalized Ramanujan conjecture asserts that every
cuspidal automorphic representation of G(A_F) is tempered at
every unramified place. As stated, this is FALSE for most G,
and the failure is the interesting part.
The failure is already visible for G = Sp(4): there exist
cuspidal automorphic representations of Sp(4)/
Q whose Satake parameters at typical primes have
eigenvalues of absolute value != 1. These are NOT
pathologies; they are everywhere. The most famous
examples come from:
- the Saito-Kurokawa lift (from a weight-2 cusp form
on SL(2,Z));
- the CAP representation attached to a cuspidal
representation on the Klingen parabolic.
Both are genuinely cuspidal (they lie in L^2_cusp, they
are Hecke eigenforms) and both are visibly non-tempered:
their Satake parameters include eigenvalues of absolute
value p^{1/2} or p, not 1. So GRC as stated is dead on
arrival for Sp(4).
For G = GL(n), however, GRC is conjecturally true and
(over function fields) a theorem. The obstruction above
uses a proper Levi subgroup (PGL(2) inside Sp(4)) and both
lifts are, after all, "parabolic inductions from a smaller
group, cut to pieces". GL(n) has no proper Levi of that
kind.
Restoring the conjecture requires restriction to the generic
case. A local representation pi_v of G(F_v) is GENERIC if it
admits a nonzero Whittaker functional (equivalently, if it
appears as a Langlands quotient with a nondegenerate
Whittaker model). A cuspidal automorphic pi is GLOBALLY
GENERIC if at least one pure tensor in pi has nonzero
global Whittaker coefficient.
The generic Ramanujan conjecture:
If pi is globally generic and pi_v is generic at
one unramified place v, then pi_x is tempered at
every unramified x.
This is expected to be true for every reductive G, with
no exceptions. It is the correct general form of the
Ramanujan bound. It is also the version the original file
claims to prove.
Even the restricted generic form is largely open. A table
appears below. Short version:
- GL(n), any n, over a function field: true (L.
Lafforgue 2002).
- GL(n) over Q, n >= 3: open.
- Classical groups over Q: mostly known (Arthur 2013,
AMS Colloq. Publ. 61, paywalled, for
orthogonal/symplectic; Mok 2015 for unitary) as a
corollary of Arthur's classification, not as a
standalone theorem.
- Exceptional groups over any field: open in full
generality.
- Generic Ramanujan for split adjoint absolutely simple
over function fields, ALL groups: open in exactly the
shape the original file attempts. This is the claim
that is not proven.
What is proven, and where the gap sits, comes next.
THEOREM (Deligne, 1974). For every prime p,
|tau(p)| <= 2 * p^{11/2}.
The proof has three ingredients, listed here with their actual
difficulty level:
(A) CONSTRUCT a 2-dimensional ell-adic Galois representation
rho_ell attached to Delta (Deligne 1971).
(B) IDENTIFY its trace at Frob_p with tau(p) (Eichler-
Shimura, Deligne 1971).
(C) BOUND its Frobenius eigenvalues using Weil I (Deligne
1974; 100 pages of etale cohomology).
Ingredient (A) is 20 pages. Ingredient (B) is a specific case
of a general pattern. Ingredient (C) is the "Weil conjecture"
and its proof is one of the landmark works of 20th-century
mathematics.
(5.A) THE GALOIS REPRESENTATION.
For each prime ell there exists a continuous 2-dimensional
ell-adic Galois representation
rho_ell : Gal(Qbar/Q) -> GL(2, Qbar_ell)
characterized (up to semisimplification) by:
trace(rho_ell(Frob_p)) = tau(p) for all p != ell,
det(rho_ell(Frob_p)) = p^11 for all p != ell.
The "Frob_p" means a Frobenius element at p in Gal(Qbar/Q), well
defined up to conjugacy; trace and det are independent of the
choice. This representation is the ell-adic realization of the
motive of Delta, called H^11 in (5.C) below.
(5.B) IDENTIFYING THE TRACE.
The identification trace(Frob_p) = tau(p) is not automatic
from the construction; it takes work. The two constructions
fit:
(Eichler-Shimura). If X_1(N) is the modular curve of
level Gamma_1(N) and X_1(N) has good reduction at p, the
Frobenius action on the "new part" of H^1_et(X_1(N) x Qbar,
Q_ell) has trace (p+1) - #{E/F_p with N-torsion} at Frob_p,
and this matches tau(p) when you compare the two sides
via the weight-12 modular form f = Delta.
(Langlands-Deligne). More abstract: the Langlands
correspondence is a bijection between cuspidal
automorphic representations of GL(2)/Q and 2-dimensional
Galois representations. This was known for weight-12
forms by Deligne 1971 (via "Travaux de Shimura"), before
Weil I was proved.
For a teenager the point is: ingredients (A) and (B) are
finite work, essentially 1970-1971-era mathematics, not the
hard part.
(5.C) THE WEIL BOUND.
Ingredient (C) is the entire force of the proof.
Let V be a specific smooth projective variety over Q, the
"Kuga-Sato variety":
V := (E x_Y E x_Y ... x_Y E) -> Y
where Y is a suitable modular curve, E -> Y is the universal
elliptic curve, and there are 11 copies (or: 2k-1 copies for
weight-2k forms, and k = 6 for weight 12 forms).
The ell-adic cohomology H^11(V, Q_ell) is 2000-dimensional
(roughly) and decomposes as a direct sum of 2-dim pieces, one
per weight-12 cuspidal new form. The piece corresponding to
Delta is precisely M_ell(Delta) from Section (5.A).
Weil I (Deligne 1974): if V is a smooth projective variety
over F_p, then every eigenvalue of Frob_p on H^i(V, Q_ell) has
absolute value p^{i/2}.
Applied here: every eigenvalue of Frob_p on H^11(V, Q_ell)
has absolute value p^{11/2}. Since the Delta-piece is 2-dim
with eigenvalues alpha_p p^{11/2}, beta_p p^{11/2},
|alpha_p| = |beta_p| = 1,
tau(p) = Tr(Frob_p | M_ell(Delta)) = alpha_p p^{11/2} + beta_p p^{11/2}.
Therefore
|tau(p)| <= |alpha_p| p^{11/2} + |beta_p| p^{11/2} = 2 p^{11/2}. []
(5.D) THE GENUINELY HARD PART.
Weil I is 100 pages of building the etale cohomology of
varieties over finite fields, proving the Lefschetz trace
formula, and controlling the "monodromy" of Frobenius at
infinity. Nothing in the proof of Ramanujan for Delta is
elementary; every honest attempt at an elementary proof has
failed, because the statement is a special case of an
arithmetic-dynamics statement about Frobenius on cohomology.
The one-sentence summary:
Delta is cohomological, so Weil I applies to it, so
Ramanujan holds for Delta.
That one-sentence summary is the driving principle behind
everything that follows.
Deligne's argument has an essential prerequisite:
Delta is COHOMOLOGICAL: its L-function appears in the
etale cohomology of an algebraic variety.
For GL(n)/Q with n >= 3, most cuspidal forms are NOT
cohomological. There is no known smooth projective variety
whose cohomology captures them. Weil I therefore does not
apply, and the sharp |a_p| <= 2 bound is not available.
The concrete state of the art:
n = 1, over Q: Ramanujan trivially (Dirichlet characters
have |chi_p| = 1).
n = 2, holomorphic (Delta and friends): Deligne 1974,
as in Deligne's argument below.
n = 2, Maass forms on SL(2, Z): OPEN. The current best is
THEOREM (Kim-Sarnak, 2003). For a Maass form of
level 1 with eigenvalue 1/4 + r^2 and Satake
parameter a_p,
|a_p| <= 2 p^{7/64}.
papers/kim-sarnak-2003.pdf
The exponent 7/64 ~ 0.109 is a long way from 0. The
Selberg eigenvalue conjecture (r = 0, i.e. lambda =
1/4) would give |a_p| <= 2, but is itself open.
n = 3, holomorphic: OPEN. Best bounds via
Barsky-Clozel, Blasius, and the "symmetric powers"
program are much weaker than |a_p| <= 2.
n = 4, 5, ...: OPEN, bound gets worse with n.
n arbitrary, over function fields: L. Lafforgue 2002,
see the function-field results below.
The gap between Deligne and GL(n)/Q is the gap between "an
automorphic form which happens to be cohomological" and
"an automorphic form which is merely automorphic". The
Langlands program predicts a uniform geometric realization,
but no such realization is known; constructing it is one
piece of the (open) functoriality conjecture.
This is exactly why the FUNCTION-FIELD case is easier, and
why function fields are easier. Over F_q(X) the geometric realization
already exists (shtuka cohomology), so Weil II applies to
every cuspidal form, not just the cohomological ones. That was
Drinfeld's and Lafforgue's insight.
Take a smooth projective geometrically connected curve X over
F_q. Its field of rational functions
F = F_q(X)
is a GLOBAL FUNCTION FIELD. Every classical theorem in
algebraic number theory has an F_q(X) analogue. The dictionary:
Q F_q(X)
---------------------- ------------------------------
primes p closed points x of X
Z_p F_q[[t_x]]
Q_p fraction field F_x
Z { f in F : f regular on X }
A_Q A_F
|n|_p = p^{-v_p(n)} |f|_x = q_x^{-ord_x(f)}
product formula product formula
Spec(Z) X
The product formula is the residue theorem for divisors.
Over Q, cohomology exists only for Shimura varieties and
similar; see the number-field discussion below. Over F_q(X), the missing object is
provided by DRINFELD'S SHTUKAS.
DEFINITION (Drinfeld 1974; GL(n) case).
A GL(n)-SHTUKA on X is a pair (E, phi) where:
- E is a rank-n vector bundle on X;
- phi: E -> Fr(E) is a meromorphic isomorphism between E
and its Frobenius pullback Fr(E) = (id_X x Frob_q)^* E,
where Frob_q: Spec F_q -> Spec F_q;
- phi has zeros and poles only at finitely many
prescribed closed points.
The finite set of legs (zeros and poles) is fixed. The shtukas
of a fixed type form a moduli stack which is smooth and finite
type over F_q. For a general reductive group G, replace E by
a G-torsor.
The essential point:
Every cuspidal automorphic form for G/F is realized in
H^*_c of some shtuka moduli stack. Frobenius acts. Weil II
gives bounds.
This is Drinfeld's insight, and it is the reason the
function-field case is easier than the number-field case. Over
Q, no analogous smooth finite-type moduli exists for an
arbitrary reductive group. Over F_q(X), it does.
Concretely, for GL(n), automorphic forms correspond to
Frobenius eigenvalues on the appropriate H^*_c (shtuka
cohomology, one cohomological degree per leg structure). This
is not a metaphor; it is the actual theorem (Drinfeld for n=2,
L. Lafforgue for all n). The proof of Ramanujan
over function fields is:
cuspidal pi <-> Frobenius eigenvalues on H^*_c <-> Weil
II bounds them <-> |alpha_i(p)|=1 for all unramified p.
No Shimura variety required; no assumption of being
"cohomological" for the specific pi; the shtuka moduli serves
uniformly for all cuspidal pi at once.
THEOREM (Drinfeld, 1974, ICM 1978).
Let F = F_q(X). Every cuspidal automorphic representation of
GL(2, A_F) is tempered at every unramified place.
THEOREM (L. Lafforgue, 2002, Inventiones Math 147;
paywalled, DOI 10.1007/s002220100181).
Let F = F_q(X). Every cuspidal automorphic representation of
GL(n, A_F), any n >= 1, is tempered at every unramified
place.
papers/drinfeld-1978.pdf
Lafforgue's proof is his Fields Medal work. It proves the full
Langlands correspondence for GL(n) over F_q(X), with
local-global compatibility, and the Ramanujan bound follows
because the Frobenius eigenvalues in the cohomology are pure by
Weil II.
For reductive groups beyond GL(n):
THEOREM (V. Lafforgue, 2014/2018, JAMS).
Let G be a split reductive group over F = F_q(X). For each
joint Hecke eigenform on G there is a global Langlands
parameter
sigma : pi_1(U) -> G^hat(Qbar_ell),
for some open curve U in X, compatible with local
parameters at every unramified place.
papers/vlafforgue-2018.pdf
So over function fields, the PARAMETERIZATION side is known for
every split reductive G.
But this is NOT the Ramanujan bound. What V. Lafforgue gives
is a semisimple parameter sigma: for each good x, a semisimple
conjugacy class sigma(Frob_x) in G^hat(Qbar_ell). This class is
uniquely determined by the Hecke eigenvalues of the form at x.
What V. Lafforgue does NOT give is:
any statement about the ABSOLUTE VALUES of the
eigenvalues of sigma(Frob_x).
For a general G, sigma(Frob_x) can be far from the maximal
compact because of the Arthur SL_2 factor. The
absolute-value part of sigma(Frob_x) is exactly where the
"generic Ramanujan" content sits. And that is precisely the
part that is NOT settled for exceptional groups.
Summary over function fields:
GL(n): Ramanujan fully proven (Lafforgue 2002).
Arbitrary reductive G: parameterization proven (V.
Lafforgue 2014/2018); Ramanujan (generic Ramanujan) NOT
proven for exceptional G.
The gap is the SL_2 direction, and Arthur's picture tells you what it
is.
As shown above, GRC is false for general G because of
non-tempered cuspidal forms; the right repair is "generic
Ramanujan". The function-field results established that over function fields, the
Langlands parameter exists (V. Lafforgue), but its
absolute-value (SL_2) part is uncontrolled.
What does the SL_2 part look like? Arthur (1989) proposed an
answer.
(9.A) ARTHUR PARAMETERS.
Arthur's parameters (conjecturally) organize the cuspidal
spectrum of G as
psi : L_F x SL(2, C) -> G^hat,
where L_F is the Langlands group of F. The SL(2, C) factor is
new; it is Arthur's innovation. Under the degree-Frobenius it
sends diag(q^{1/2}, q^{-1/2}) to a specific element, whose
"image" in G^hat acts on the Lie algebra Lie(G^hat) with three
eigenspaces of the ad-cocharacter h_O: weights -2, 0, +2.
The lower-triangular nilpotent
N = phi|SL(2,C)( (0 1; 0 0) )
identifies a nilpotent orbit O of G^hat, the "Arthur
SL_2 orbit" of psi. Every nilpotent orbit of G^hat arises this
way (Jacobson-Morozov).
The prediction: the members of the Arthur packet of psi have
Satake classes at each unramified x inside a set R_{q_x, O}.
Specifically:
DEFINITION. For any nilpotent orbit O of G^hat with
Jacobson-Morozov homomorphism
phi_O : SL(2, C) -> G^hat,
the SATAKE ORBIT SET of O at a prime power q is
R_{q, O} := Ad(G^hat) .
{ phi_O diag(q^{1/2}, q^{-1/2}) h :
h in Z_{G^hat}(phi_O(SL(2,C)))(
Qbar), |iota(a(h))|=1 for all
complex embeddings iota and all characters a }
the set is a subset of the adjoint quotient of G^hat.
The condition on h is: h is compact in the centralizer of
phi_O(SL(2,C)). For O = 0, phi_O is trivial, and R_{q, 0} is
the tempered set (maximal compact of G^hat) exactly. Otherwise
R_{q, O} carries some eigenvalues of modulus q^{1/2} or q. Those
are the "non-tempered" signatures.
(9.B) TWO EXAMPLES.
EXAMPLE 1: G = GL(2), G^hat = GL(2, C).
Nilpotent orbits of gl(2): O_0 = {0}, O_P = principal.
O_0: Satake set = U(2). This is tempered.
O_P: phi_P = diagonal SL(2). The centralizer of the
diagonal torus in SL(2) -> GL(2) is C^* (diagonal matrices
of the form diag(z, z)). So R_{q, O_P} = {diag(q^{1/2} z,
q^{-1/2} z) : z in C^*}. This is the "principal series
induced from a character of B", i.e. an Eisenstein series
of weight 1. NOT cuspidal.
Conclusion: only O_0 occurs in the cuspidal spectrum of
GL(2)/Q. Which is exactly Ramanujan for GL(2): Deligne/Drinfeld.
EXAMPLE 2: G = Sp(4), G^hat = SO(5, C).
Nilpotent orbits of so(5): four, indexed by partitions of 5
obeying the "even parts with even multiplicity" rule for so:
partition name dim
------------ ------------- -----
[1^5] trivial 0
[3, 1^2] minimal 4
[2^2, 1] Saito-Kurokawa 6
[5] principal 8
The Saito-Kurokawa lift of a cusp form on SL(2,Z) is a genuine
CUSPIDAL automorphic form on Sp(4), whose Satake parameter at
each unramified prime p lies in R_{p, [2^2,1]}. The
5-dimensional L-factor of this lifted form is not tempered:
some of its Satake eigenvalues have modulus different from 1. So SK appears in
the [2^2, 1] summand, NOT in the tempered summand. GRC as
stated above fails for Sp(4) exactly because of this.
So over Sp(4), multiple nilpotent orbits DO occur in the
cuspidal spectrum. GL(2) is the exception: only O_0. That is
the whole reason Ramanujan holds for GL(2) and not for
Sp(4).
(9.C) ARTHUR'S CONJECTURE.
CONJECTURE (Arthur, 1989). For every connected reductive G
over any global field, the cuspidal spectrum decomposes as
L^2_cusp(G) = direct sum over nilpotent orbits O of
L^2_cusp,O(G),
where L^2_cusp,O(G) is the closed subspace on whose every
unramified local component pi_x the Satake class lies in
R_{q_x, O}, and where only finitely many O have non-zero
summand.
Stronger form (endoscopic character identities): L^2_cusp,O
should be a genuine automorphic object on the centralizer
H_O = Z_{G^hat}(phi_O(SL(2,C))) (or a twisted form thereof),
with specific character identities.
Arthur proved this for classical groups over number fields
(orthogonal and symplectic: AMS 2013; unitary: Mok 2015). The
endoscopic-identity part is the hardest. The existence of the
decomposition into R_{q,O} summands is the "spectral support"
part and is much easier; it is the part that GLR states as
Conjectures 3.4.5/3.4.6.
The GLR conjectures are the spectral-support part, strengthened
in two ways:
(a) at every full finite level D (not just trivial level);
(b) stated as a STRICT direct sum decomposition, without any
"packet bookkeeping".
That is what the original file claims to prove. The precise
statement, and the reason it is open, appear below.
Now the precise statement of what the original file claims to
prove.
SET-UP. Let X be a smooth projective geometrically connected
curve over F_q, F = F_q(X). Let G be a split connected
semisimple group over F_q. Let D be an effective divisor on X
(the LEVEL), with arbitrary multiplicities. Let
K_D = ker( G(O_A) -> G(O_D) )
be the full-level-D compact open subgroup of G(A_F). Let
C_{D,Q} = cuspidal part of Q-valued compactly supported
functions on G(F) \ G(A_F) / K_D.
This is finite-dimensional: K_D is open compact, so the double
quotient is discrete and locally finite, and cuspidal functions
on a finite quotient are the cuspidal forms of bounded level.
For each unramified closed point x of X (x not in the support
of D), the spherical Hecke algebra A_x acts on C_{D,Q} by
convolution. Any joint eigenfunction f of all A_x (x good)
determines a semisimple conjugacy class
s_x(f) in (G^hat)^ss / conj,
called the SATAKE CLASS of f at x. This is the local
representation-theoretic data attached to f.
For a nilpotent orbit O of Lie(G^hat), let R_{q_x, O} be the
Satake-orbit set from Arthur's picture. Define the SPECTRAL-SUPPORT
SUMMAND
C_{D,O,Q} := { f in C_{D,Q} :
for every unramified x, every A_x-eigenvalue
occurring in f has Satake class in R_{q_x,O} }.
This is a natural candidate: it collects all cusp forms whose
Satake parameters at every good point sit inside the "orbit-O"
slice of the dual group, not just inside the maximal compact.
CONJECTURE (GLR Conjecture 3.4.6).
The natural addition map
bigoplus_O C_{D,O,Q} -> C_{D,Q}
is an isomorphism. The sum is over nilpotent orbits O of
Lie(G^hat). Equivalently: C_{D,Q} breaks as a direct sum of
spectral-support summands, one per orbit.
CONJECTURE (GLR Conjecture 3.4.5).
Same statement with Q replaced by Qbar_l:
bigoplus_O C_{D,O,l} -> C_{D,l}
is an isomorphism, where C_{D,l} = Qbar_l tensor_Q C_{D,Q}.
REMARK (GLR 3.4.7). 3.4.5 is equivalent to the l-adic
Ramanujan-Arthur conjecture.
What is genuinely new relative to Arthur:
(a) Arthur (1989) states the SL_2 picture at TRIVIAL level
D = 0, and asks only for "spectral support" inside some
R_{q,O}. GLR ask for the same at every FULL FINITE LEVEL
D, including nonreduced D, which is a nontrivial
Iwahori-Hecke-algebra extension of the statement.
(b) GLR ask for a STRICT direct sum, not only a
"packet bookkeeping" statement inside an Arthur packet.
This is a sharper combinatorial form.
Arthur proved his version for classical groups over Q
(orthogonal/symplectic: Arthur AMS 2013; unitary: Mok AMS 2015).
The GLR version over F_q(X) for arbitrary split connected
semisimple G is OPEN. The original file claims to prove it in
full generality at every full finite level, plus the generic
Ramanujan corollary. Neither is a theorem.
Here is a table of what is actually proven, with anchors.
QUESTION STATUS / REFERENCE
------------------------------------------- -------------------
tau(p) bound |tau(p)| <= 2 p^{11/2} Deligne 1974
Maass form Ramanujan on GL(2)/Q OPEN, best bound
|a_p| <= 2 p^{7/64}
(Kim-Sarnak 2003)
GL(n)/Q Ramanujan, n >= 3 OPEN
GL(2)/F_q(X) Ramanujan Drinfeld 1974
GL(n)/F_q(X) Ramanujan L. Lafforgue 2002
Split G/F_q(X) parameterization V. Lafforgue 2018
Classical G/Q endoscopic (Arthur) Arthur 2013,
Mok 2015 (unitary)
Classical G/Q generic Ramanujan follows from Arthur
/Mok + global char.
ident. (mostly)
Split G/F_q(X) FULL Ramanujan, exceptional OPEN
GLR 3.4.5/3.4.6 general split G, full level OPEN
Generic Ramanujan for split adjoint OPEN in general;
absolutely simple G/F_q(X) partial results only
The two best PARTIAL results on the last line are:
Sawin-Templier, "On the Ramanujan conjecture for automorphic
forms over function fields I. Geometry", JAMS 34 (2021),
653-746, arXiv:1805.12231.
Under the extra hypotheses
(i) char F_q > 2;
(ii) pi is "monomial geometric supercuspidal" at every
ramified place (a strong local hypothesis);
(iii) pi is tempered at ONE unramified place,
they prove pi is tempered at every unramified place (their
Theorem 11.7). Removing (ii) and (iii) is open.
papers/sawin-templier-2021.pdf
Ciubotaru-Harris, "On the generalized Ramanujan and Arthur
conjectures over function fields", arXiv:2311.15300 (2023).
Under extra hypotheses on the local components at ramified
places, they prove a generalized Ramanujan bound. Their local
criterion (Corollary 6.32(1)): a complex Satake parameter of
an unramified spherical unitary that is "geometric" (i.e.
has a half-integral direction aligned with an SL_2) and that
is generic at some place is tempered. This is the criterion
the original file uses in its Corollary 1. But CH's own
hypothesis (iii) of their Theorem 5.4 still requires one
tempered unramified place.
papers/ciubotaru-harris-2023.pdf
Neither result gives the unconditional, every-split-connected-
adjoint-absolutely-simple, one-generic-unramified-place claim
of the original file's Corollary 1. That statement is OPEN.
Nothing here says the GLR conjectures are wrong. They are
widely believed true. They are just not proven.
A 5547-line LaTeX file claims to prove both GLR Conjectures
3.4.5/3.4.6 at full finite level AND the unramified generic
Ramanujan conjecture for split connected adjoint absolutely
simple groups over F_q(X).
Neither is a theorem. The file has the fingerprints of a
hallucinated text. Six concrete tells:
1. "For completeness" paragraphs. In a genuine proof,
"for completeness" introduces an easy side remark. In
the original file, EVERY proof contains at least one
paragraph beginning "For completeness, here is why the
thing we are trying to prove is true", and inside it the
hard step is quietly assumed. Grep for the phrase.
2. "smallskip noindent emph{The ...}" sub-headings used as
soft hiding places. Examples: "The full lattice.",
"A finite enlargement for a general reductive lattice.",
"The rational cuspidal spectrum." Each one is where a
hard technical step is compressed into a paragraph.
3. Heavy machinery (Wakimoto filtration, Springer
pushforward, Harder-Narasimhan strata, near-by cycles,
replete bar construction) introduced in half a page and
then "this proves Proposition X". Real papers spend 20
or 50 pages on each of these. Each of these being fully
correct in 1 page would already be a paper.
4. Two major open conjectures "solved" at once by an author
with no other public track record in this field, using
no new machinery beyond what already exists. That is
never what a real breakthrough looks like.
5. Citations that do not physically exist. For example,
arXiv:2603.26925v2 (GLR paper, dated 2026) and the
"Rationality of the Canonical Unramified Arthur
Filtration" preprint cannot be located on arXiv.
6. Logical tells. In Section 4 of the file, the
Wakimoto-filtration argument extracts a top degree N,
then claims the amplitude bound gives an upper bound
C + dm max <lambda, h_e>, and then asserts "impossible
for large m" when the delta = 2 ||lambda||^2 - max <...>
is strictly positive. The delta is strictly positive
exactly when the original claim holds; this is where the
circularity lives. That step is the crux, and in the real
literature it takes L. Lafforgue ~200 pages.
Trying to edit a hallucinated proof into correctness is
fruitless; hallucinations cannot be patched by polishing. A
fake proof is not a real proof with typos, it is a proof whose
middle steps do not exist. The only correction is to state what
is actually known, with references, so a reader can check every
claim.
A fourteen- to sixteen-year-old who wants to see the field
as it really is could start with Edward Frenkel's Love and Math
(2013), John Derbyshire's Prime Obsession (2003) on the Riemann
hypothesis and its cousins the Ramanujan-type bounds, and Marcus
du Sautoy's The Music of the Primes (2003). A reader of sixteen
to eighteen who is comfortable with calculus and linear algebra
can use Jean-Pierre Serre's A Course in Arithmetic (1973), the
shortest clean introduction to modular forms and Hecke operators,
or Neal Koblitz's Introduction to Elliptic Curves and Modular
Forms (1993) for the same material in a friendlier register, or
Diamond and Shurman's A First Course in Modular Forms (2005) for
the cleanest modern treatment of dimension formulas, Hecke
operators, and Eichler-Shimura. At the next level, Henryk
Iwaniec's Topics in Classical Automorphic Forms (1997) covers the
analytic side, Daniel Bump's Automorphic Forms and Representations
(1997) covers the representation side including L-functions, and
Stephen Gelbart's Automorphic Forms on Adele Groups (1975) is an
older but pointed introduction to Langlands program motivations.
A few habits help when reading mathematics skeptically.
Anyone who announces that they have proved the Riemann hypothesis
or the full Ramanujan conjecture for all groups is almost
certainly wrong; these problems have accumulated hundreds of
claimed proofs, and every one has been wrong. Within a proof, a
paragraph that begins "for completeness" is usually where the
gap is: read those paragraphs first. A paper that claims to
solve several open problems at once, in a field the author is
not known in, is suspicious; real mathematics rarely does that.
And when in doubt, ask a working mathematician, ideally one who
is not the author; they are usually happy to say whether
something is a theorem, a conjecture, or nonsense.
In short:
Ramanujan observed three facts about tau(n) in 1916. Two
were proven by Hecke in 1937 (multiplicativity, the
prime-power recurrence). The third, |tau(p)| <= 2 p^{11/2},
was proven by Deligne in 1974, using a Galois representation
attached to Delta and the Weil conjectures. This is the
whole of the CLASSICAL Ramanujan conjecture, both over Q
and over F_q(X) for GL(2) (Drinfeld) and for any GL(n)
(L. Lafforgue). For a general reductive group over a
function field, the parameterization is known (V. Lafforgue)
but the temperedness (generic Ramanujan) is OPEN, and the
finer GLR spectral-support conjectures 3.4.5/3.4.6 are OPEN.
The original file claiming to prove them does not; it is a
hallucinated text and cannot be repaired by editing.
Two complete proofs that a high school student can check
line by line appear earlier. The structure of Deligne's
proof, with the one hard ingredient (Weil I) named and
dated, appears earlier. Everywhere else, each "what is known"
statement has a name and a year attached so it can be checked.