The Alpöge–Claude proof of the Riemann hypothesis
explained for the Riemann zeta function
Claude Fable 5.2, Codex Astra 6.2, and Gemini Carbon
at the behest of Michael O. Rubinstein
Working draft, 21 November 2026
Abstract. We explain the argument of Alpöge and Claude for the Riemann hypothesis in
the case of the Riemann zeta function. The analytic step shows that a one-sided bound for
a smoothed sum over prime powers, valid throughout intervals [Y, Y 3/2 ] with Y arbitrarily
large, excludes zeros to the right of the critical line. We derive this bound from one inequality
between weighted sums over primes, the main inequality (M), and two estimates for Dirichlet
characters. We prove the two estimates from theorems in the literature, and explain how the
divisibility and the size of theta lifts enter (M). The proof of (M) is taken from the Lean
development. This is a work in progress: Part III writes up parts of that proof and states
exactly the inputs it takes from the Lean development; (M) itself is not proved here.
Note. This is a working guide to Alpöge’s Lean-verified proof of the Riemann hypothesis, specialised
to the Riemann zeta function. Its goal is a human-readable exposition of the proof in the style of
classical analytic number theory. It is a work in progress, written as an aid to studying the proof.
Written up so far: the deduction of the Riemann hypothesis from one inequality, called (M)
throughout, and two estimates for Dirichlet characters (Parts I and II); proofs of those two estimates
and the construction of two auxiliary sets (§III.1–III.3); the analytic part of the bound for the size of
the theta kernels (§III.4; four passages there are conditional on inputs from the Lean development);
a local divisibility estimate for power series at inert primes, under one hypothesis from the Lean
development, proved for orbits of length two and with a further duality input for longer orbits
(§III.5); the passage from that estimate to lattices: the lemmas on lattices are proved, the prime sum
is split and several of its terms are bounded (the bounds for the remaining loss sums are quoted), and
the theorem on the glued lattice is conditional on two statements of the Lean development (§III.6);
the deduction of the size inequality from eleven named inputs, and the value of an index-corrected
mean of periods under stated hypotheses (§III.7); the deduction of the dimension bound that is
used from three statements of the Lean development (on traces, ranks and pairings) and stated
conditions on sizes and subspaces, with a finite model of the trace estimates (§III.8); and the step
from the bound for Galois cohomology to a bound for denominators, with an account of how the
Lean development proves that bound (§III.9).
Not proved here, all of it inside (M), and taken from the Lean development: (1) the integrality of the
theta kernels in the coordinates in which the divisibility is proved, and the statement on valuations
built on it; (2) the arithmetic inputs of the size bound: the inner-product formula for the theta
lifts and its local factors, the covolumes of the lattices, the local index bounds and the passage
between levels; (3) the trace-formula estimates themselves; (4) the bound for Galois cohomology
itself; (5) the construction of the maps and their weights, and the assembly of (M), which in the
Lean development is a mean over a family of maps (§III.6 treats a model case only).
The table at the start of Part III indicates which statements are proved in these pages and which
are still taken without proof from the Lean development.
Contents (entries are links)
Part I. How the proof works 4. The main inequality (M) . . . . . . . . . . 30
I.0 The strategy in one page . . . . . . . . . 3 5. From the table to a single prime sum . . . 33
I.1 From zeros to a sum over primes . . . . 4 6. Where the main inequality (M) comes from 37
I.2 Chebyshev’s method, and where it stalls 5 Sources . . . . . . . . . . . . . . . . . . . . . 53
I.3 A divisible number with smooth exponents 6
Part III. Proofs of the quoted statements
I.4 Repairing the three defects . . . . . . . 7
III.1 The auxiliary primes: proof of Estimate B 55
I.5 The inequality that comes out . . . . . 13
III.2 The large scale: proof of Estimate C . 61
I.6 From the table to G(x) . . . . . . . . . 14
III.3 The sets of columns T2 and T . . . . . 64
I.7 Where (M) comes from . . . . . . . . . 16
III.4 The size of the theta kernels (partly
I.8 Why the exponent is one half . . . . . 18 written) . . . . . . . . . . . . . . . . . 66
I.9 Dictionary with Alpöge’s announcement 19 III.5 Divisibility at inert primes . . . . . . . 80
I.10 The same proof for Dirichlet L-functions 21
III.6 Divisibility: from derivatives to lattices 91
I.11 Notes on the literature . . . . . . . . . 22
III.7 The size bound: the arithmetic inputs . 111
I.12 Summary: what is proved where . . . . 25
III.8 The images in general position: the
Part II. Details and proofs trace-formula estimates . . . . . . . . . 129
1. What a zero forces . . . . . . . . . . . . . 26 III.9 The bound for Galois cohomology . . . 144
2. How divisibility counts primes . . . . . . . 28 Computations . . . . . . . . . . . . . . . . . . 161
3. The reduction to one prime sum . . . . . . 29 References . . . . . . . . . . . . . . . . . . . . 161
What this is. In November 2026 Levent Alpöge announced a proof, found with the AI system
Claude, of the Riemann hypothesis. It exists as a formal proof in the Lean proof assistant, about 17
million lines long. This paper explains how the argument works for the Riemann zeta function, for
a reader who knows a first course in analytic number theory: the prime number theorem, Dirichlet
characters, and Mellin transforms.
How it is organised. Part I is an exposition: the strategy first, then each step with its motivation
and with examples computed on real primes. Part II gives the details and proofs to which Part
I refers. A reader can stop after Part I and know how the proof goes; Part II is there to check
each step. Part III, which is being written, is to prove the statements that Parts I and II quote.
It proves Estimates B and C and constructs the two auxiliary sets of residue classes (the “sets of
columns” of §I.5). For the main inequality (M) it proves some steps, proves others conditionally on
statements of the Lean development which it names, and takes the rest from that development; the
table at the start of Part III says which.
What is proved here and what is not. The argument rests on three statements. One of
them, the main inequality (M), is the arithmetic heart of the matter. Two further estimates (B
and C) concern auxiliary Dirichlet characters. We state B and C exactly, and (M) in the form
in which it is used, and explain why one should expect them. That the three together imply the
Riemann hypothesis is proved in Part II. Estimates B and C are proved in Part III, from theorems
on Dirichlet L-functions that are cited there. But (M) is not yet proved here: we are working
through its proof in the Lean development and adding the details to Part III section by section. Of
what it needs, Part III so far constructs the two sets of columns (§III.3) and proves: the analytic
part of the bound for the size of the theta kernels, with four conditional passages (§III.4); a local
divisibility estimate for power series at inert primes, under the integrality of the theta kernels in the
coordinates of the height-two logarithm (hypothesis (B)), for orbits of length two and with a further
2
duality input for longer orbits (§III.5); the lemmas on lattices that carry this estimate to lattices
and part of the sum over primes, the main theorem there being conditional on two statements of the
Lean development (§III.6); that the size inequality follows from eleven named inputs (§III.7); that
the dimension bound follows from three statements of the Lean development on traces, ranks and
pairings (§III.8); and the step from the bound for Galois cohomology to a bound for denominators
(§III.9). Part II proves the elementary lemmas. The rest of (M) is for now quoted: hypothesis
(B) and the statement on valuations built on it, the arithmetic inputs of the size estimate, the
trace-formula estimates themselves, the bound for Galois cohomology itself, the construction of the
maps and their weights, and the assembly of (M) from these. Statements marked [quoted] are taken
from the Lean development without proof. The full proof of (M) is where almost all of the formal
development goes, and writing it up is the work in progress. Every numerical example, except
where stated, can be reproduced with the scripts listed under Computations, after Part III.
Part I. How the proof works
I.0 The strategy in one page
Chebyshev’s idea about counting primes. Chebyshev [Che52] bounded the number of primes
using the prime factorization of factorials. In the simplest form of the argument, due to Erdős
[Erd32], one considers the middle binomial coefficient 2n . It is divisible by every prime between n
n
and 2n, and it is at most 4 . So the product of those primes is at most 4n . That is an upper bound
n
for a count of primes, obtained from two facts about one integer: what divides it, and how big it is.
One piece of notation. For a real number v we write v+ for its positive part: v+ = v if v > 0,
and v+ = 0 otherwise. For example 5+ = 5 and (−3)+ = 0. A sum of terms of the form (· · · )+ is
automatically cut off where the terms would become negative.
The plan. The proof of the Riemann hypothesis described here is Chebyshev’s argument, run
with a far better “integer”. There are five steps.
1. Zeros to primes. The Riemann hypothesis is equivalent to a bound |G(x)| ≤ Cx1/2+ε for
a certain smoothed count of primes G(x), for every ε > 0, with C = C(ε). In fact an upper
bound G(x) ≤ x1/2+ε is enough, and it is enough to have it on a sparse sequence of long
intervals. (§I.1.)
2. Why Chebyshev’s method stalls. For n! the divisibility is known exactly: the exponent of
p is h ⌊n/ph ⌋. The floor functions are the obstacle; they are why the classical method gives
P
constants but not asymptotics. (§I.2.)
3. A divisible object with no floor functions. Derivatives of functions on an elliptic
curve with complex multiplication are divisible by primes to exponents of the smooth form
a/p + a/p3 + · · · , but only at the primes that stay prime in the associated imaginary quadratic
field. Subtracting a large weight w turns each term a/ph into (a/ph − w)+ = w (x/ph − 1)+
with x = a/w: exactly the smooth kernel that G uses, magnified by w. (§I.3 and §I.4.)
4. Size, to the precision needed. “Divisibility ≤ size” now gives an upper bound for a sum
over primes. For it to be useful the size must be known not just to leading order a log a
but through the term linear in a. This forces one to replace a single number by a linear
map between lattices, built from theta series, whose size is governed by classical formulas
3
(Chowla–Selberg, the mass formula). The resulting comparison of divisibility with size is the
main inequality (M). (§I.5 and §I.7.)
5. Clean up and balance. The sum in (M) runs over primes in certain residue classes. Averaging
over the classes leaves G(x) plus three sums twisted by quadratic characters, which are small
for well-chosen auxiliary primes. Repeating the construction m times leaves two competing
√
errors, x/m and m. Taking m ≈ x gives G(x) ≤ x1/2+ε . (§I.6 and §I.8.)
Logical skeleton.
§I.6 §I.6 §I.1
(M) , B, C −−−→ Estimate A −−−→ G(x) ≤ x1/2+ε on long intervals −−−→ RH. (I.1)
|{z} | {z }
[quoted] Part III
Three statements enter: the main inequality (M), which is quoted, and two estimates for auxiliary
Dirichlet characters, one at the scale x (Estimate B) and one at a much larger scale (Estimate C),
which are proved in Part III from cited theorems. The order of choices is this. For each small ε
there are arbitrarily large Y for which a single pair of auxiliary primes q, d serves for every x in
[Y, Y 3/2 ]; the rest of the construction is then made for each x separately.
I.1 From zeros to a sum over primes
The statistic. Let Λ(n) = log p if n = ph is a prime power and 0 otherwise. Define
x
S(x) = Λ(n) −1 , G(x) = S(x) − x log x + (1 + γ)x, (I.2)
X
n≤x
n
where γ is Euler’s constant. Equivalently S(x) = n Λ(n)(x/n − 1)+ , where v+ = max(v, 0): with
P
the positive part, the condition n ≤ x under the sum can be omitted. Compared with Chebyshev’s
ψ(x) = n≤x Λ(n), the sum S weights n by the ramp (x/n − 1)+ : as a function of n it is large for
P
small n, decreases, reaches 0 at n = x, and is 0 for n > x. Also S(x) = x 1x ψ(t) dt/t2 . The two
R
subtracted terms are the main term of S, so G is a remainder.
R∞
Why this statistic. The Mellin transform of the ramp is simple: 1 (x/n − 1)+ x−s−1 dx =
n−s /(s(s − 1)) for Re s > 1. Summing over n with weights Λ(n),
ζ ′ (s) 1 1 1+γ
Z ∞
F (s) = G(x)x−s−1 dx = − · − + . (I.3)
1 ζ(s) s(s − 1) (s − 1) 2 s−1
The last two terms cancel the pole at s = 1 (both terms of its principal part). What remains are
simple poles of F at the zeros ρ of ζ, with residue −1/(ρ(ρ − 1)) at a simple zero, and a pole at
s = 0. (The trivial zeros −2, −4, . . . also give poles; they produce the term O(x−2 ) below.) Shifting
the contour in the inverse transform gives the explicit formula (compare [Dav00, Ch. 17])
xρ
G(x) = log 2π − + O(x−2 ), (I.4)
X
ρ ρ(ρ − 1)
the sum being over the nontrivial zeros. It converges absolutely, because of the factor ρ(ρ − 1);
that is the benefit of smoothing. One has ρ ρ(1−ρ)1
= 2 + γ − log 4π = 0.0461 . . . . If every zero
P
√
has real part 2 then |ρ(ρ − 1)| = ρ(1 − ρ), and therefore, for large x, |G(x) − log 2π| ≤ 0.0462 x.
1
A zero with real part β > 12 would contribute a term of size xβ .
4
A sanity check on real primes. Summing over the prime powers up to 2·106 gives G(104 ) = 1.749,
G(105 )√= 3.671, G(106 ) = 11.040, G(2 · 106 ) = −12.268, to be compared with log 2π = 1.838 and
0.0462 106 = 46.2.
One side is enough. The terms of a zero off the line and of its complex conjugate together
form a real function of x that oscillates and takes both signs. So one expects a zero with real part β
to make G(x) exceed +cxβ for some arbitrarily large x. Thus an upper bound for G should already
exclude it. That an upper bound suffices follows from a theorem of Landau on Mellin transforms of
functions of one sign [Lan05]; see [MV07, §15.1].
Lemma (one-sided criterion). Let 0 < θ < 1. If G(x) ≤ xθ for all large x, then ζ(s) ̸= 0 for
Re s > θ.
Proof. The integral H(s) = 1∞ xθ − G(x) x−s−1 dx equals 1/(s − θ) − F (s), and its integrand is
R
nonnegative for large x. Landau’s theorem says that a Mellin transform of an eventually nonnegative
function has a singularity at the real point of its abscissa of convergence. But by (I.3) the function
H is regular at every real s > θ: the only real singularities of F in s > 0 would come from real
zeros of ζ in (0, 1), and there are none. Hence the abscissa is at most θ, the integral converges for
Re s > θ, and F is holomorphic there. A zero of ζ in that half-plane would be a pole of F .
The arithmetic argument below produces the upper bound only for x in certain long intervals.
The lemma survives this weakening.
Lemma (the same, on intervals). Let 0 < θ < 1. If there are arbitrarily large Y such that G(x) ≤ xθ
for all x ∈ [Y, Y 3/2 ], then ζ(s) ̸= 0 for Re s > θ.
The proof replaces Landau’s theorem by an estimate of high derivatives of F , which see only one
interval at a time; it is one page and is given in full in Part II, §1.
Conclusion of step 1. To prove the Riemann hypothesis it suffices to show: for every ε > 0
there are arbitrarily large Y with G(x) ≤ x1/2+ε throughout [Y, Y 3/2 ]. (Zeros come in pairs ρ, 1 − ρ,
so no zeros to the right of the line means none to the left.) From now on the goal is an upper bound
for a sum over primes.
I.2 Chebyshev’s method, and where it stalls
The method. Let N = 2n n . Every prime p with n < p ≤ 2n divides (2n)! and not (n!) , so it
2
divides N . Also N ≤ (1 + 1)2n = 4n . A positive integer divisible by distinct primes is at least their
product, so
log p ≤ log N ≤ 2n log 2. (I.5)
X
n
0 is a weight attached to another place: the one reached from the given place by the
Frobenius automorphism of ph followed by complex conjugation. The display keeps only the main
term. The statement proved in the source is an inequality for sums of such exponents over the m
lifts and over the places, with several loss terms: from the digits of the integers involved, from
repeated visits of a Frobenius orbit to the same place, and from indices of lattices (Part II, §6.4)
[quoted] (see §III.5 for what is proved).
The Frobenius and the places. The field E lies inside a field generated by roots of unity.
The Frobenius of p is the automorphism of E induced by raising those roots of unity to the pth
power. It permutes the 4n embeddings of E, and it depends only on residue classes of p. The
2n places are
√ chosen embeddings, one from each pair of complex conjugates. At a prime that is
inert in Q( −d) an odd power of the Frobenius carries each chosen embedding to one that is not
chosen; composing with complex conjugation brings it back. This is the reason for the conjugation
above. Since the Frobenius depends only on a residue class of ph , residue classes of prime powers
will appear in §I.5.
Where the subtraction of w comes from. The polynomial P is a product, over the 2n
places, of minors (determinants of sub-matrices) of the matrix variable at each place. Its shape at
10
each place is described by a diagram of boxes. At the chosen place the diagram has a first row
of length about a. At another place τ ′ the polynomial contains about wτ ′ minors, where wτ ′ is
the weight parameter at τ ′ . At the scale ph the Frobenius links the chosen place to the place τh ,
and each minor there can absorb only one derivative in a linked direction: in det(X + sb, Y + tb)
there is no term in st, s2 or t2 . Each derivative absorbed at τh costs one unit of the gain at the
scale ph . There are about wh minors at τh , each absorbing at most one, and at most a/ph units
are available; so the loss at that scale is at most min(a/ph , wh ), and what survives is (a/ph − wh )+
(Part II, (6.1)). In the simplest case (a single index; Part II, (6.1a)) the loss comes from the digits
of the integers involved, by Legendre’s formula, and is of size O((log a)2 ) for each prime (Part II,
§6.4). The losses in the statement the source proves are larger and of several kinds, as said above.
Example. Take a = 81, p = 3, and wh = 2 for all h.
h 1 3 5 7
p h
3 27 243 2187
a/ph 27 3 0.33 0.04
a/ph − w 25 1 −1.67 −1.96
(a/ph − w)+ 25 1 0 0
The exponent is at least 25 + 1 = 26, up to the losses. Every term with ph > a/w = 40.5 is zero.
The choice of a and w, and what it achieves. Let x be the scale at which we want
information about primes. Put
b = exp(xε ), w ≈ b, a ≈ xb. (I.20)
Then a/w ≈ x, and the elementary identity
a x
−w =w −1 (x = a/w) (I.21)
ph + ph +
shows what has been gained. (In the example, x = 40.5 and for ph = 3 the right side is
2 · (40.5/3 − 1) = 25.)
• Only prime powers up to x are involved, at the places that carry the weight b. The right
side of (I.21) vanishes for ph ≥ x. Although we differentiate about xb times, the primes that
matter at those places are those below x. (At the places with small weights the cut-off is
much larger; see below.)
• The kernel is the ramp of §I.1. The factor (x/ph − 1)+ is exactly the weight with which prime
powers enter G(x). This is where the smoothing in the definition of G comes from.
• Rounding becomes invisible. Each prime power below x now contributes about b times its
ramp value. The bound (I.19) is used only where its main term is positive; elsewhere one
uses only that the exponent is not negative, which is the integrality of the entries. The loss
terms are of lower order, and after dividing the whole inequality by b they are absorbed in
the error of (M). The precise accounting is part of the quoted statement. This removes defect
(D1), and with it the floor-function problem of §I.2.
Why b must be exponentially large. To defeat rounding, a power of x would do. One
other step needs more. There are 2n places and each needs a weight, but only some of them can be
given the weight b (the columns that carry it form the set T2 constructed in §III.3). The others
are given very small weights w′ , so that they test primes at a scale a/w′ ≥ 13 exp(xε/3 ); only at
11
such a scale is the distribution of primes in the relevant progressions known unconditionally with a
good error term (§I.6(b)). The weights in the construction are positive integers, so w′ ≥ 1; this
forces a ≥ 13 exp(xε/3 ), and hence b exponential in a power of x. The size of b also enters the
control of denominators (§I.7). What is bounded there is a weighted sum of logarithms of orders,
ℓ log√ℓ a hℓ,a in the notation of Part II, §6.8, and b is chosen so that the available bound is at
P P
most b. After division by b this is negligible.
I.4.4 Repair of (D2): characters, and a table
Only inert primes: use a character. Let d be √ a prime with d ≡ 3 mod 4 and let χd (r) = (r/d)
be the Legendre symbol. The primes inert in Q( −d) are those with χd (p) = −1. A sum over odd
powers of inert primes is a sum over all prime powers with the coefficient 1 − χd (r):
2, p inert, h odd,
(
1 − χd (p ) =
h
(I.22)
0, all other prime powers with p ̸= d.
(With this coefficient 2 the sum matches the factor 2 in 2e log p of §I.4.2(ii).) So the divisibility
side is r Λ(r) (1 − χd (r)) (· · · ): the sum for ζ minus the sum for L(s, χd ). To isolate ζ one must
P
know that the χd -sum is small. This is arranged by choosing d well (§I.6(e)), and it is not circular.
Many places: a table. By (I.19) the weight met by a prime power r = ph depends on the
place to which the Frobenius of r leads. The field used in the proof is
−1 √
√
E = Q ζN + ζ N , −q, −d , ζN = e2πi/N , N = 4n, (I.23)
with n a large power of 2 and q ≡ 3 mod 4 a second auxiliary prime. Starting from the fixed place,
the place reached by a prime power r is determined by two things: the residue class of r modulo
−1 r + ζ −r ), and the quadratic symbol χ (r) (the
N up to sign (the Frobenius sends ζN + ζN to ζN q
√ N
Frobenius multiplies −q by χq (r); the complex conjugation, which enters for the odd powers of
inert primes that contribute here, then changes its sign, which only interchanges the names of
the two rows). For general r prime to 2q this is a labelling of residue classes; the prime d itself is
treated separately, with coefficient log d (Part II, §4.1). So the places can be arranged in a table,
with two rows (the value of χq ) and n columns (the residue class), and each prime power is assigned
to a cell of the table. The inequality that comes out of the construction, stated in §I.5, is a bound
for a sum over prime powers in which each prime power carries the weight of its cell. §I.6 then
removes the dependence on the cell by averaging.
12
I.4.5 What the reader can check, and what is quoted
Statement in this section Status
Tn , the period ϖ, the rational numbers rk , the size of Tn computed; script
e6_hecke_theta.py
∥z a ∥2 = a! and inequality (I.17) proved above
Identity (I.21) and thePexample a = 81 elementary
Leading terms agree: p inert p log p = a log a + O(a)
2a
Mertens’ theorem in progressions
A determinant absorbs one derivative per direction; the factorial proved in Part II, §6.4
count
The entries of the lift are the numbers (I.15) quoted
The exponent bound (I.19) for the theta series of the proof written up in §III.5, conditionally
on the integrality of the theta
kernels in the coordinates of the
height-two logarithm (hypothe-
sis (B)), which is taken from the
Lean development
I.5 The inequality that comes out
Here is the statement. It has the form: there exist data (weights, two sets of columns, integers m,
k) with the properties listed, such that the inequality (M) holds. The properties are given here in
outline and exactly in Part II, §4.2–4.3. The statement looks elaborate; the example after it shows
that it is elementary to state.
Data. Fix a small ε > 0 and a scale x. Let q ≡ 3 mod 8 and d ≡ 7 mod 16 be primes with
q | d + 1 (how they are chosen: §I.6). They are very large: if x lies in the interval [Y, Y 3/2 ], then
log q is about Y ε/5 and log d about Y ε/5 (log Y )2 . Let n be a power of 2 with x2+8ε ≤ n ≤ 2x2+8ε ,
and N = 4n. Let H be the group of odd residue classes modulo N with r and −r identified; it has
n elements, written [r]. Let b = exp(xε ).
The table. Picture a table with two rows v ∈ {0, 1} and n columns indexed by H. An odd
prime power r = ph , p ̸= q, is assigned to the cell
row v(r) = 12 (1 + χq (r)), column [r]. (I.24)
The 2n cells are the 2n places of §I.4.3. Each cell i carries a positive weight wi . A cell of row 1 is
distinguished when its column lies in a set of columns of density δ; its weight is then (very nearly)
b. All other weights are tiny by comparison: so small that a/wi ≥ 13 exp(xε/3 ).
The prime sum and its main term. For a derivative order a with |a/b − x| ≤ 1 (it depends
on the label t introduced below, and is written at in Part II), put
a
P = 1 − χd (ph ) log p (I.25)
X X
h
− wcell(ph ) ,
p̸=2,q h≥1
p +
L′
M = a log a − 1 − γ + (1, χd ) − ⟨log w⟩ , (I.26)
L
where ⟨log w⟩ is the average of log wi over the 2n cells. P is the divisibility side: by (I.21) it is a
sum of ramps, one for each cell, at the scale a/wi . M is its expected value if primes were perfectly
distributed among the cells (§I.6(a); proved in Part II, §5.1). So P − M measures the irregularity
of primes in the cells of the table, at the scales a/wi .
13
The primes 2 and q can be omitted without changing anything: d ≡ 7 mod 8 gives χd (2) = 1,
and q | d + 1 gives χd (q) = 1 by quadratic reciprocity, so the coefficient 1 − χd vanishes on their
powers. This is what the congruence conditions on q and d are for (Part II, §4.1). (The powers of d
do occur, with coefficient log d, but only at the large scale, since d > x.)
The inequality. The construction depends on three further choices: a translate u ∈ H (which
moves the distinguished cells to other columns), a label t in a set T ⊂ H (which shifts the column
of every prime power), and an index j = 1, . . . , m numbering the m repetitions. Write Pu,t,j , Mu,t,j
accordingly.
m
1 X X X E x xm k
≤ xCε + +m+ 2 (M)
Pu,t,j − Mu,t,j ≤ E,
n u∈H t∈T j=1 B m k n
Here B = |T | m b δ/2 is the natural normalisation (δ is the proportion of distinguished columns), k
is a further parameter of the construction (a rank), and C is an absolute constant. The inequality
is asserted for x ≥ x0 (ε) [quoted]. In words: on average over translates, the primes are not more
numerous than expected in the table, up to the four errors shown. It is an upper bound, as every
instance of (I.6) is. The set T and the set of distinguished columns (called T2 in Parts II and III)
are constructed in Part III, §III.3; how the construction uses them is part of the quoted statement.
More precisely: let Y be large, let q, d be the primes of Estimate B for Y (§I.6(e)), and let
Y ≤ x ≤ Y 3/2 . Then the construction provides m = x1/2+O(ε) , k = x9/2+O(ε) , sets T , T2 with
|T2 |/n = δ ≥ x−ε and |T |/n ≥ 21 x−ε/8 , orders at and weights as above, for which (M) holds. The
constants C and x0 (ε) do not depend on Y , q, d.
A table small enough to write out. Take q = 3, d = 23, N = 16, so n = 4 and
H = {[1], [3], [5], [7]}. The prime 7 is a square modulo 3 and 7 ≡ 7 mod 16: it goes in row 1, column
[7], with coefficient 2 log 7 because 7 is a non-square modulo 23. The primes up to 61 with non-zero
coefficient fall as follows.
v\column [1] [3] [5] [7]
0 17 − 5, 11, 53 23 (I.27)
1 − 19, 61 37, 43 7
With x = 1000, b = 10, a = 104 , weight 10 in cell (1, [1]) and 0.05 elsewhere, summing over the
prime powers gives P = 95461.616 and M = 95444.464: the excess P − M is 17.152, against a main
term of nearly 105 . Moving the distinguished cell to the other three columns gives excesses 368.21,
−129.25, −204.65. The mean of the four excesses is 12.86. (The example ignores one feature of
the construction, that the weights are integers. The identities it illustrates hold for any positive
weights.)
I.6 From the table to G(x)
This step is classical analytic number theory, and it is proved in full, from (M), B and C, in Part II,
§4–5. Here are the ideas.
(a) Each cell is an arithmetic progression. The prime powers in a given cell are those
in one pair of residue classes ±c modulo N with a√prescribed value of χq ; the coefficient 1 − χd
then selects the odd powers of primes inert in Q( −d). For fixed moduli and scale tending to
infinity, the prime number theorem in progressions gives each cell the fraction 1/(2n) of the primes,
half of them inert, and M is defined as the resulting main term, with its constants. The identity
14
P − M = i wi Ri (a/wi ), where Ri (y) is the remainder of cell i at scale y (Part II, (5.2)), is then
P
exact (Part II, (5.3)). No uniformity is claimed at this point. At the scale x there are far more
cells than primes, and a single Ri (x) need not be small; the argument uses only the average over
translates in (c). At the large scale the uniformity is Estimate C.
(b) Two scales. A cell with weight wi tests primes up to a/wi . For a distinguished cell this is
a/b ≈ x: the scale we care about. For every other cell it is at least 13 exp(xε/3 ), an enormous scale.
Why arrange this? The construction needs weights on all 2n cells, but we want information at scale
x only. At the enormous scale the distribution of primes in these progressions can be controlled
with a saving better than any power of x. This is Estimate C below. It rests on the classical
zero-free region for Dirichlet L-functions; it is available because log q and log d are much smaller
than xε/3 , and because q and d are chosen so that no exceptional real zero interferes. So those cells
contribute a negligible amount. In the example they account for 4.18 of the mean excess 12.86.
(c) Averaging over translates. As u runs over H, the distinguished cells visit every column
equally often. So the average over u no longer sees the residue class modulo N . It sees only the
two quadratic conditions: row 1 means χq (r) = 1, and the coefficient is 1 − χd (r). The average of
the distinguished part is therefore
δb X x
Λ(r) 1 + χq (r) 1 − χd (r) −1 minus its main term. (I.28)
2 r≤x r
In the example this accounts for the other 8.68 of the 12.86.
(d) Expanding the product. Since (1 + χq )(1 − χd ) = 1 + χq − χd − χq χd , the last expression
2 Rq,d (x) with
is δb
Rq,d (x) = G(x) + Gχq (x) − Gχd (x) − Gχq χd (x), (I.29)
′
where, for a primitive character χ =
̸ 1, Gχ (x) = r≤x Λ(r)χ(r)(x/r − 1) + x LL (1, χ) is the analogue
P
of G for L(s, χ). The function G appears with coefficient +1. (The construction gives this at the
scale a/b, with |a/b − x| ≤ 1; the passage to x costs O(1 + log x), see Part II, §5.4.) Dividing (M)
by B, and using Estimates B and C below to bound the error terms (Part II, §5.4), gives:
Estimate A. Let Y be large, q, d the primes of Estimate B, and Y ≤ x ≤ Y 3/2 . For the parameters
of the construction, with m ≈ x1/2 , k ≈ x9/2 , n ≈ x2+8ε ,
x xm k ′
Rq,d (x) ≤ xCε + + m + 2 ≤ x1/2+C ε (x ≥ x0 (ε)). (I.30)
m k n
(e) The three twisted sums. To pass from Rq,d to G we need:
Estimate B. Let 0 < ε < 52 1
. For all large Y there are primes q ≡ 3 mod 8, d ≡ 7 mod 16 with
q | d + 1, log q ≍ Y , log d ≍ Y ε/5 (log Y )2 , such that |Gχ (x)| ≤ x1/2+ε for χ = χq , χd , χq χd
ε/5
and all x in the interval [Y, Y 3/2 ]. (Proved in Part III as Theorem III.1.1(d), in a stronger form:
exponent 12 + 2ε , and any interval [Y, Y A ].)
Estimate C. For the same primes q, d and every x ∈ [Y, Y 3/2 ], Y large, every one of the 4n
primitive Dirichlet characters χ that occur when the cells of the table are expanded in characters
(their conductors divide 4nqd) satisfies |Gχ (y)| ≤ y exp(−κxε/8 ) for all y ≥ 13 exp(xε/3 ), with a
constant κ > 0; here Gχ = G for the trivial character. (Proved in Part III as Theorem III.2.1,
where κ is called c′ and may be any positive number once x is large. The proof uses a further
property of q and d, part (e) of Theorem III.1.1.)
15
Is this circular? It looks like assuming the Riemann hypothesis for three L-functions. It is not,
for a reason of scale. The moduli q and d are astronomically larger than x: their logarithms are
′
powers of x. The quantity Gχ (x) has two parts. One is the constant LL (1, χ), an infinite sum over
prime powers. The other is the finite sum over r ≤ x. The tail of the constant beyond Z = Y 600 is
shown to be small using a zero-density estimate for zeros close to the line Re s = 1; this discards a
small proportion of the moduli, and uses nothing about zeros near Re s = 12 . What remains is a
sum over r ≤ Z, which is still very short compared with q and d. Mean-square estimates over the
√
modulus show that for most pairs (q, d) such a sum is as small as a random sum, about x. One
only needs one good pair. For C see (b) above. Estimates B and C are proved in Part III, from
standard theorems on Dirichlet L-functions that are cited there.
′′
Conclusion. From A and B (A itself uses (M), B and C): G(x) ≤ Rq,d (x) + 3x1/2+ε ≤ x1/2+C ε
on the interval. By §I.1 this proves the Riemann hypothesis.
I.7 Where the main inequality (M) comes from
This is the part quoted from the Lean development; Part III has begun its proof: §III.4 (the
analytic part of item 2 below) and §III.5 (item 1 below, conditionally on one hypothesis from the
Lean development). Sections III.6–III.9 prove further steps, each from named statements of the
Lean development, which are quoted: the passage from derivatives to lattices, in part (§III.6); the
deduction of the size inequality from eleven inputs (§III.7); the deduction of the dimension bound
of item 3 from three statements (§III.8); and the step from the bound for Galois cohomology to the
bound for denominators (§III.9). The table at the start of Part III says what each section proves,
proves conditionally, or quotes. The construction of the maps and weights and the assembly of (M)
are not begun, and (M) is still quoted. We describe the objects and the three things that have to
be proved about them.
Theta series. A theta series counts representations by a quadratic form. For the form a2 + b2 ,
2 2
Θ(z) = sa +b = 1 + 4s + 4s2 + 4s4 + 8s5 + · · · , s = e2πiz . (I.31)
X
a,b∈Z
Theta series have two properties at once: they are sums over a lattice, so their coefficients are
integers with arithmetic meaning; and they are Gaussian in nature (the terms have absolute value
2 2
e−2πy(a +b ) , where y is the imaginary part of z), so their sizes are governed by Gaussian integrals.
These are the two properties needed in §I.4: divisibility of derivatives (§I.4.1) and Gaussian size
(the repair of (D3), §I.4.2).
A theta lift is a map defined by a theta series in two sets of variables, used as the kernel of
an integral: (T f )(h) = Θ(g, h)f (g) dg. (It is conjugate-linear in f .) In the case used here both
R
groups are definite unitary groups, for which everything is finite: a “function on the group” is a
function on a finite set of lattices L1 , . . . , Lc (the classes in a genus), the integral is a finite sum
weighted by 1/|Aut Li |, and T is a matrix. Its entries are derivatives, at the origin, of the theta
series attached to pairs of lattices. So a theta lift is a matrix of derivatives of theta functions,
between two spaces that carry natural integer lattices (functions with integer values) and natural
norms.
The role of (I.6). Let T : L → M be an injective linear map between lattices in Euclidean
spaces. If T (L) ⊂ pe M for some prime power, then, comparing volumes,
e log p ≤ log ∥T ∥ + µmax (M ) − µ(L), (I.32)
16
where ∥T ∥ is the operator norm and µ = − rank 1
log(covolume) is the slope of a lattice (µmax : the
largest slope of a sublattice). For L = M = Z this is (I.6). In the construction the lattices are
modules over the integers of a field with complex multiplication, and one works with squared norms:
(I.32) is doubled, so that 2e log p is compared with log ∥T ∥2 , as in §I.4. Three things must be
established.
1. Divisibility. The matrix of the lift is divisible by pe with e of the form h odd (a/ph − w)+
P
at inert primes. The mechanism is that of §I.3, applied to theta functions of lattices over a field
with complex multiplication: derivatives of order a gain a(1/p + 1/p3 + · · · ). The subtraction of w
has an elementary source, explained in §I.4.3: each minor can absorb only one derivative in a given
direction. Composing m lifts multiplies the matrices, so the exponents add: this is the analogue of
a function vanishing to order m in Stepanov’s proof of the Riemann hypothesis for curves over finite
fields. The divisibility at inert primes is written up in §III.5, conditionally on the integrality of the
theta kernels in the coordinates of the height-two logarithm (hypothesis (B)), which is taken from
the Lean development. The passage from this estimate to lattices is the subject of §III.6. Proved
there: the lemmas on slopes of lattices, the form of the prime sum and its splitting by classes of
primes, and bounds for several of its terms. The theorem on the glued lattice is conditional on two
statements of the Lean development. The bounds for the eleven loss sums and for their total are
quoted, and so is the statement on valuations itself.
2. Size. The squared norm of the lift has logarithm at most a(log a − 1 − log ΩE ) + · · · , with
the linear term known exactly. The leading term is the Gaussian integral ∥z a ∥2 = a!. The constants
Ω and ΩE are periods: Ω belongs to an elliptic curve, and ΩE , defined below, to the field E. A
derivative with respect to the algebraic invariant differential and one with respect to the analytic
variable differ by a constant factor, the period; an ath derivative acquires its ath power, and Ω is
the square of that factor. Its value is
2π L′
log Ω = γ + log √ − (1, χd ). (I.33)
d L
In this form √
the identity is Kronecker’s limit formula; Ω is then a geometric mean, over the ideal
classes of Q( −d), of squared periods of the corresponding elliptic curves. Expressing it through
values of the Gamma function is the formula of Lerch and of Chowla and Selberg [CS67], whose
name the source uses. For the field Q(i), with d replaced by 4 and χd by the character modulo
4, exponentiating this formula gives Ω = 2π(ϖ/π)2 = Γ( 14 )4 /(4π 2 ) exactly: that is 2π times the
square of the factor ϖ/π met in §I.4.1. The constant Ω above is the period for the elliptic curve.
The constant that enters the size bound for the field E is ΩE , with
rdE 2π L′
log ΩE = log Ω − log √ = γ + log − (1, χd ), (I.34)
d rdE L
where rdE is the root discriminant of E. Part II, §6.6 gives both normalisations; the cancellation of
the linear terms is the same. This is how γ and L′ /L(1, χd ) enter the main term M . They are the
same constants that appear in the prime sum, by §I.6(a); that the two sides produce the same linear
term is the reason the method can work at all. The remaining linear terms come from the slopes in
(I.32): the number of lattice classes counted with weights 1/|Aut| (the Siegel mass formula) and the
dimensions of the spaces (the Weyl dimension formula). After an exact cancellation of factorials,
the total is M .
3. Denominators and rank. Two things can spoil (I.32).
• Denominators. The lift need not map integer-valued functions to integer-valued functions; one
must find the right lattices. They are assembled prime by prime, using the freedom to twist
17
the lift by characters, and what cannot be removed is measured by√certain Galois cohomology
groups of the field (generalisations of class groups). A bound, by b, for the logarithm of the
product of the orders of these groups (over all primes ℓ and all twists up to x400 ) is proved
separately in the Lean development (it is quoted here); it uses the theory of automorphic
forms and Galois representations, and is a large part of the formal development. Part III,
§III.9 proves the step from this bound to the bound for denominators, and the vanishing of
the groups in some cases over Q from cited theorems; the bound itself is quoted, with an
account of its proof in the Lean development.
• Rank. A single map of small rank controls only the maximal slope of its target, which can be
far from the average and would lose the main term. One therefore uses many maps. Their
images cannot all be perpendicular to one another: the dimensions of the images add up to
more than the dimension of the target. What is shown is a bound for their total overlap (a
second moment; a trace formula gives it). It says that the images are spread evenly: the
weighted sum, over the maps, of the adjoint of the map followed by the map is close in mean
square to a multiple of the identity of the target. A lemma of linear algebra then says the
images fill almost the whole target, so that the mean slope is controlled. Part III, §III.8
proves that the dimension bound of this lemma follows from three statements of the Lean
development (on traces, ranks and pairings) and stated size conditions; the three statements,
and the theorems behind them, are quoted.
The elementary parts. Several steps in this section are elementary and are proved, with
worked examples, in Part II, §6: the persistence of integrality through a chain of maps; the minor
congruence; the factorial count (from Legendre’s formula) giving the gain a p/(p2 − 1); the Gaussian
norm; the cancellation that leaves M ; the elementary slope inequality; and the linear-algebra lemma
on images that are spread evenly in this sense. The analytic part of the size estimate is proved
in part in Part III, §III.4: the Gaussian bound for one block, for one map and for a chain, with
four passages conditional on inputs from the Lean development. Part III, §III.7 proves that the
size inequality (III.75) follows from eleven named inputs, and that an index-corrected mean of
periods equals log ΩE under hypotheses stated there. The eleven inputs are quoted, among them
the inner-product formula, the covolumes and the local index bounds. So the size bound itself (Part
II, (6.5), first line) remains quoted.
I.8 Why the exponent is one half
After division by B, the errors in (M) are, up to factors xCε ,
x xm k
m . (I.35)
m k n2
• x/m: errors made once, in the sizes of the lattices at the two ends of the chain of maps.
The gain from divisibility is collected m times, so these are divided by m. This is why one
composes many lifts.
• m: errors accumulated over the m steps. The budget of each step (ranks, weights, denomina-
tors) itself grows in proportion to m, so the total is of order m2 before normalisation; the
factor m contained in B leaves m.
• xm/k and k/n2 : costs of the auxiliary parameters k (the rank) and n (the size of the table).
√
With k ≈ x9/2 and n ≈ x2 the first is negligible (x−3 ) and the second is again of order x.
18
√ √
The terms x/m and m are balanced when m ≈ x, and both equal x. That is where the exponent
2 comes from. If the second cost were m instead of m after normalisation, one would have
1 2
x/m + m2 , balanced at m = x1/3 , and the same analytic argument would give only that ζ(s) ̸= 0
for Re s > 23 , provided the corresponding inequality could be proved with the other errors below
x2/3+ε . (Whether such an inequality can be proved without the estimate for Galois cohomology is
a separate question, which we do not address.) So the precision of the comparison of lattices is
decisive.
I.9 Dictionary with Alpöge’s announcement
The announcement1 lists twenty steps, in the order in which they were found between 6 October
and 16 November 2026. It is a record of discovery, so it contains earlier formulations and motivating
ideas as well as steps of the final argument. This section says where each step appears in this paper.
The correspondence is our reading. For steps 1, 2, 5 and 7–20, and for the Poincaré bundle in step
3, it rests on statements of the Lean development that were read in the course of this work. Steps 4
and 6, and the rest of step 3, were not found in the Lean development and are matched from the
description in the announcement.
Steps that are in this paper.
1
https://alpo.ge/remark.pdf, pages 3–4.
19
Step in the announcement In this paper
1 Stepanov determinant §I.2: a nonzero integer divisible by pep is at
Q
least that large, (I.6); the model is Stepanov’s
count of points on curves
5 Switch to n≤X Λ(n)(X/n − 1) − X log X + §I.1: this is G(X)
P
(1 + γ)X √
8 Iterated theta lifts between totally definite
√ uni- §I.7 (theta lift; composition); §I.8 (m ≈ x)
tary groups; a chain of length about X
9 Norms of the lifts; cancellation using Chowla– §I.7, “Size”; Part II, §6.6–6.7; the norms, in
Selberg, Weyl, Siegel mass part: Part III, §III.4; the size inequality from
eleven quoted inputs: §III.7
10 Lower bound for the divisibility of the m-fold §I.4.3, (I.19); §I.7, “Divisibility”; written up in
lift Part III, §III.5, conditionally on the integrality
of the theta kernels in the coordinates of the
height-two logarithm (hypothesis (B)), which is
taken from the Lean development; the passage
to lattices, in part: §III.6
12 Seesaw: the period of the iterated lift is a Bessel §I.7, “Denominators” (the freedom to twist by
period independent of the Hecke characters characters); Part II, §6.8
13 Bost slope inequality §I.7, the inequality for a map of lattices; the
elementary case is proved in Part II, §6.8; the
form with minimal slopes used in the chain is
quoted (§6.10)
14 Cost of each lift; “one would stop at 2/3”; need §I.8: x/m + m2 gives 32 , x/m + m gives 12
for “division”
15 Glue an integral structure across Hecke charac- §I.7, “Denominators”: lattices assembled prime
ters (a “gcd”) by prime; Part II, §6.8
16 Relative trace formula; approximate orthogo- §I.7, “Rank”; Part II, §6.9–6.10
nality √
17 Classes in H 1 not killed by twisting; a weaken- §I.7, “Denominators”: the bound by b for
ing of Vandiver’s conjecture Galois cohomology; Part II, §6.8
20 GRH here only ζ, the case of the trivial character
Steps that appear here in another form, or only in part.
• Step 2 (the invariant derivation D on a curve with complex multiplication satisfies Dp ≡ 0
at an inert prime; divisibility of Da Ω · θ(0)). This is the idea of §I.3 and §I.4.1: derivatives
on such a curve, the period (ϖ in §I.4.1, Ω in §I.7), and exponents involving odd powers of
p only. We could not confirm the stated exponent on the simplest example: for the theta
function of Z[i] the exponent we computed is smaller (§I.4.1). The object or its normalisation
in the announcement must differ from that model. For what is proved in this paper about
the divisibility at an inert prime, see Part III, §III.5.
• Steps 4 and 6 (positivity W (f ) ≥ 0 in Weil’s explicit formula, obtained from a determinant;
a zero-free region from such a positivity). These are an earlier form of our first step. In the
final argument the statement “an upper bound excludes zeros” is the Landau-type lemma of
§I.1.
• Step 7 (determinants containing Gan–Gross–Prasad periods; interlacing weights). In the
Lean development the period pairing at the end of the chain of lifts is identified with a Bessel
period (§I.7; Part II, §6.2 and §6.8; source file [19]), and a condition of interlacing on the
weights governs the passage between consecutive unitary groups (source file [23]). We do not
use the Gan–Gross–Prasad terminology in the text.
20
• Steps 11 and 19 (an Iwasawa tower; a construction of Behrend). The tower is the field E
with n a large power of 2, which gives the columns of the table (§I.4.4, §I.5). Behrend’s
construction is used to choose the set T2 of distinguished columns (Part II, §4.2). The set is
constructed in Part III, §III.3, following source file [27], which uses the sphere construction
of Behrend as formalised in Mathlib. (The definitions in [17] are an earlier formulation; the
final theorem does not use them.) The announcement itself expects Behrend’s construction
to be unnecessary.
• Step 18 (covolumes; the index of theta series; congruences between lifts and forms). These
are the slopes of the lattices in §I.7. The elementary inequality for slopes is proved in Part II,
§6.8; the form with minimal slopes, the covolumes and the indices are quoted.
A step largely absent from the final proof. Step 3 (an identity qN = M k + R; a quadric;
automorphy through a theta function on a sixfold cover of GL5 ; the Poincaré bundle). We found
no trace of the identity or of a sixfold cover in the Lean development (quadrics occur in some file
names, in what appear to be unrelated lemmas on orbit measures); step 16 of the announcement
refers to them as “a different half-plane argument”. A Poincaré bundle does occur in the archive,
on an auxiliary abelian scheme, in a construction of the line bundle whose section gives the theta
kernel in algebraic coordinates (§I.4.1; Part II, §6.5). The declaration we tested from that file is
not among those the final theorem depends on, so we cannot say that the final proof uses it.
Steps of the proof that the announcement does not list. The choice of the auxiliary
primes q and d, with the three character sums and the zero-density and mean-square estimates that
control them (§I.6(e)); the two scales, and the size of b (§I.4.3, §I.6(b)); and the form of Landau’s
lemma on intervals (§I.1; Part II, §1). These are in the Lean development; file names are given
in Part II. The first is proved in Part III, §III.1, the estimate at the large scale in §III.2, and the
lemma in Part II, §1.
I.10 The same proof for Dirichlet L-functions
The Lean proof treats all Dirichlet L-functions at once, and ζ is the case of the trivial character.
This section says what changes for another character. It is described from the statements of the
Lean development and is not worked out in this paper.
Let ψ be a primitive Dirichlet character of conductor M . Its powers 1, ψ, ψ 2 , . . . form a finite
cyclic group; let o be the number of distinct powers, that is, the order of ψ. For 1 ≤ j < o let ψj
be the primitive character that induces ψ j (the power ψ j itself need not be primitive), and let Gψj
be the function of §I.6 attached to it:
y L′
Gψj (y) = Λ(n)ψj (n) −1 +y (1, ψj ). (I.36)
X
n≤y
n L
The statistic G is replaced by the combination
1 o−1
Gψ (y) = G(y) + Re Gψj (y). (I.37)
X
3o + 1 j=1
For n prime to M , the term Λ(n)(y/n − 1) occurs in it with the coefficient 1 + 3o+1
1
j=1 Re ψ (n),
o−1 j
P
which equals 3o+1 if ψ(n) = 1 and 3o+1 otherwise. For n not prime to M the coefficient is
4o 3o
1 + 3o+1 j=1 Re ψj (n), which can differ from this formula, since ψj (n) may be nonzero where
1 Po−1
21
ψ j (n) = 0; it is still at least 3o+1
3o
. So every coefficient is positive. Multiplied by 3o + 1, the
combination is the statistic attached to the product of ζ(s)3o+1 and the functions L(s, ψj ), 1 ≤ j < o;
at a zero of this product the multiplicities add, so every zero of ζ or of an L(s, ψj ) gives a pole of
the transform. The lemma of §I.1 then applies with one addition. The proof for ζ uses that there is
no pole on the real segment 0 < s < 1; for L-functions a real zero β cannot be excluded in advance.
At such a zero the residue of the transform is a positive multiple of 1/(β(1 − β)). If it is the only
pole nearest to the real point σ, the derivatives at σ have size about (σ − β)−j−1 , and the fixed
sign of the residue makes them positive; this contradicts the bound of Part II, §1 when β > θ; the
Lean proof of the lemma has this case as a separate branch. The conclusion is: if for every ε > 0
the bound Gψ (y) ≤ y 1/2+ε holds on intervals [Y, Y 3/2 ] with Y arbitrarily large, then ζ and every
L(s, ψj ) have no zero with real part greater than 21 . (For o = 1 this is G and ζ.)
The rest changes as follows. The field E is enlarged so that ψ becomes one of its characters.
The columns of the table are then indexed by a larger group: it consists of o cosets of the subgroup
on which the label coming from ψ is trivial, and the common size of these cosets depends on how
much ψ shares with the residue classes modulo 4n (for the character ( 2· ) modulo 8, for instance,
the number of columns does not grow). The set T is chosen with slightly different densities on
different groups of columns; this is where the weight 3o+11
comes from. Estimate B is needed for the
j
twists ξχq χd of all characters ξ modulo M , and Estimate C for the characters of the larger field;
i
the theorems of §III.1 and §III.2 are stated in that generality. Everything else is the same: the
main inequality (M), with its theta lifts, divisibility, size, trace formula and Galois cohomology, is
the same construction for the larger field, and the passage from the table to one prime sum is the
same computation with more characters.
I.11 Notes on the literature
This paper explains an argument due to Alpöge and Claude [Alp26]. Apart from that argument it
uses, or quotes, results of many mathematicians. This section says which, and where. Two kinds of
use should be distinguished.
(a) Published results used in the proofs given in this paper.
• Chebyshev’s method (§I.0, §I.2): Chebyshev [Che52]; the form with the middle binomial
coefficient is due to Erdős [Erd32]. The exponent of a prime in n! is due to Legendre [Leg08].
Textbook accounts: [Apo76, Ch. 4], [Dav00, Ch. 7].
• Landau’s theorem on Mellin transforms (or Dirichlet series) of functions of constant sign
(§I.1): Landau [Lan05]; see [MV07, §15.1]. The version on intervals (Part II, §1) is proved in
full in this paper, following the Lean development.
• The explicit formula and the value ρ ρ(1−ρ) 1
= 2 + γ − log 4π (§I.1): von Mangoldt; see
P
[Dav00, Ch. 12 and 17]. The use of positivity in an explicit formula goes back to Weil [Wei52].
• The prime number theorem in arithmetic progressions, for a fixed modulus, and Mertens’
theorem in progressions (§I.4.2, §I.6; Part II, §5): de la Vallée Poussin; see [Dav00, Ch. 20
and 22], [MV07, Ch. 11].
• Theorems on Dirichlet L-functions used in Part III. For Estimate B (§III.1): the zero-free
region with at most one exceptional zero, Landau and Page [Pag35], [Dav00, Ch. 14]; primes
in progressions to large moduli, Thorner and Zaman [TZ21], after Linnik and Gallagher
[Gal70], [Bom87]; a zero-density estimate, Montgomery [Mon71]; Landau’s bound for L′ /L in
22
a zero-free region, [Tit86], [MV07]. For Estimate C (§III.2): the zero-free region, the count of
zeros in a window and the truncated explicit formula, [Dav00, Ch. 14, 16, 19], [MV07, Ch. 11,
12]. These are cited, not proved. The mean value over the auxiliary primes is obtained from
primes in progressions and not from a large sieve; compare [Bom65], [DH66], [HB95].
• Sets without three-term progressions (§I.9, steps 11 and 19; Part III, §III.3): Behrend [Beh46].
The construction is carried out, with proof, in §III.3.
• Stepanov’s method (§I.2, §I.7; Part II, §2), as the model for counting with multiplicity:
Stepanov [Ste69]; see Bombieri [Bom74].
• The lemniscate functions and their arithmetic (§I.3): Hurwitz [Hur99]. Formal groups with
complex multiplication, which explain the exponent a p/(p2 − 1) at inert primes: Lubin and
Tate [LT65], Honda [Hon70]. That the reduction is of this type exactly at the inert primes:
Deuring [Deu41]. Algebraicity of the lattice sums of §I.4.1: compare Damerell [Dam70].
• The Gaussian inner product on polynomials, ∥z a ∥2 = a! (§I.4.2; Part II, §6.5): the Fock, or
Bargmann, space [Bar61].
• Hadamard’s inequality, the Chinese remainder theorem and the Cauchy–Schwarz inequality
are used in Part II, §6 without further reference.
(b) Results behind the quoted statements. The statements marked [quoted] are proved in
the Lean development, which contains its own proofs of the theorems it needs. The underlying
mathematics has the following sources. We cite them for the classical statements; we have not
compared the formalised statements with the published ones in detail.
• Theta series as kernels and the representation that governs them (§I.4.1, §I.7): Weil [Wei64];
the correspondence between forms on a pair of groups: Howe [How79]. Seesaw identities:
Kudla [Kud84]. The inner product formula for theta lifts: Rallis [Ral87]. Bessel periods and
the restriction problems in which they occur: Gan, Gross and Prasad [GGP12].
• The period formula (§I.7; Part II, §6.6). In the form used, it is Kronecker’s limit formula (see
[Wei76]); the evaluation through Gamma values is due to Lerch, and to Chowla and Selberg
[CS67].
• The mass formula for definite lattices (§I.7; Part II, §6.6): Siegel [Sie35]; for Hermitian lattices,
Shimura, see Gan, Hanke and Yu [GHY01]. The dimension formula: Weyl [Wey25].
• Slopes of Euclidean lattices and the maximal slope (§I.7; Part II, §6.8, §6.10): Stuhler [Stu76],
Grayson [Gra84]. The slope inequality for maps of lattices: Bost [Bos96].
• The relative trace formula (§I.7, “Rank”): Jacquet [Jac86].
• The bound for Galois cohomology (§I.7, “Denominators”; Part II, §6.8). The announcement
describes it as a weakening of Vandiver’s conjecture. According to our reading of the Lean
development its proof uses the vanishing of the Iwasawa µ-invariant (Ferrero and Washington
[FW79]): a statement under that name is proved in the development and the final theorem
depends on it. The development also contains statements named after Kato’s Euler systems
[Kat04]; some of these are among the declarations the final theorem depends on and one
that we tested is not. The proof also uses modularity lifting theorems. We did not find the
main conjecture of Iwasawa theory (Mazur and Wiles [MW84]) stated under that name. An
upper bound for a cohomology length in terms of the valuation of a Bernoulli number with
explicit integral factors is present; it corresponds to a consequence of the main conjecture,
and we have not compared the two. A bound for the whole group H 2 at every prime is also
23
present. Finiteness of these groups is due to Soulé ([Sou79], from memory). Soulé [Sou03,
Thms. 2 and 3] gives a quantitative bound away from small primes: for the primes p ≥ 4n′ − 1,
p ≥ [L : Q] + 2, where L is the field and n′ is the twist (Soulé writes n; it is not the n of this
paper). This was compared with the article by a reviewer, not by us. We have not identified a
published estimate in the full small-prime range. Part III, §III.9 describes the statement and
the structure of its proof in the development, and says which published theorems the cases
correspond to; we have not checked how closely the formal proofs follow the published ones.
What is new. The combination described in this paper, namely iterated theta lifts as a source
of multiplicity, the comparison of divisibility with size through slopes of lattices carried to the term
linear in the order, and the resulting one-sided bound for G(x), is the contribution of [Alp26]. The
exposition, the worked examples and the computations in this paper are ours.
24
I.12 Summary: what is proved where
Statement Status
One-sided bound for G on intervals ⇒ RH (§I.1) proved (Part II, §1)
Model divisibility a p/(p2 − 1); lemniscate numbers proved for the model series (Part II, §2); the
(§I.3) numbers computed exactly
(M), B and C ⇒ Estimate A ⇒ the bound for G proved (Part II, §3–5); the identities used are
(§I.6) checked numerically on an example
Estimate B: the auxiliary primes (§I.6) proved (Part III, §III.1), from four cited theorems
Estimate C: the characters at the large scale (§I.6) proved (Part III, §III.2), from three cited theo-
rems and Theorem III.1.1(e)
The sets of columns T2 and T (§I.5) constructed (Part III, §III.3); their use in (M) is
quoted
(M), size (§I.7, item 2) the analytic estimates for one map and for a
chain are proved in an abstract setting (Part III,
§III.4); that the maps of the construction are
instances of it is quoted; four passages of §III.4
are conditional; the arithmetic inputs are quoted
(M), divisibility at inert primes (§I.7, item 1) written up in §III.5, conditionally on the inte-
grality of the theta kernels in the coordinates of
the height-two logarithm (hypothesis (B)), which
is taken from the Lean development; elementary
lemmas proved (Part II, §6)
(M), from derivatives to lattices (§I.7, item 1) proved (Part III, §III.6): lemmas on slopes of
lattices, the splitting of the prime sum, bounds
for several of its terms. Conditional on two state-
ments of the Lean development: the theorem on
the glued lattice. Quoted: the bounds for the
eleven loss sums and their total; the statement
on valuations
(M), arithmetic inputs of the size bound (§I.7, item proved (Part III, §III.7): the size inequality fol-
2) lows from eleven named inputs; a mean of peri-
ods equals log ΩE under stated hypotheses. The
eleven inputs are quoted
(M), rank (§I.7, item 3) proved (Part III, §III.8): the dimension bound
follows from three statements of the Lean develop-
ment and stated size conditions. The three state-
ments are quoted; elementary lemmas proved
(Part II, §6)
(M), denominators (§I.7, item 3) proved (Part III, §III.9): the step from the bound
for Galois cohomology to the bound for denom-
inators. The bound itself is quoted; elementary
lemmas proved (Part II, §6)
(M), the construction of the maps and weights, and quoted; not begun in Part III
the assembly
Error budget and exponent 12 (§I.8) proved from the stated form of (M)
The formal proof in Lean establishes the Riemann hypothesis for all Dirichlet L-functions at once,
and with it the statements quoted here, in its own formulation. The case of ζ described here is the
case of the trivial character; it uses the same construction.
Sources (the Lean files) are at the end of Part II; the list of scripts follows Part III.
25
Part II. Details and proofs
Part II can be read after §I.1 and §I.5: it uses G(x) from §I.1 and otherwise repeats its definitions.
It is numbered separately. Its sections are 1–6. Its displays are numbered (II.n); those referred to
often carry section tags instead: (1.1), (4.1), (5.1), . . . , (6.7), and (A), (M). The main inequality is
called (M) in all three parts, and Estimate A of Part I is formula (A) of Part II. Section 3 of Part
II states the reduction to one prime sum again, with all definitions, so that sections 4–6 can be read
on their own. References in square brackets with numbers are to the list of source files at the end of
Part II. Some letters are used with more than one meaning in Part II (for instance N , L, T , h and
a); each is defined where it is used. Formula (5.6) of Part II is Estimate C of Part I. Where sections
3–5 attribute a statement to “the source”, a proof is now in Part III for the following: the choice of
the primes q, d and the bound for the three twists (section 3; Part III, §III.1); formula (5.6) (Part
III, §III.2); the existence of the sets T , T2 with the stated densities (section 4; Part III, §III.3).
The divisibility at inert primes of sections 6.3–6.4 is written up in Part III, §III.5, conditionally
on the integrality of the theta kernels in the coordinates of the height-two logarithm (hypothesis
(B)), which is taken from the Lean development. Part III, §III.6–III.9 prove further steps towards
(M), each from named statements of the Lean development: the passage to lattices in part (§III.6),
the size inequality from eleven quoted inputs (§III.7), the dimension bound of section 6.9 from
three quoted statements (§III.8), and the step from the bound for Galois cohomology to the bound
for denominators (§III.9); the table at the start of Part III gives the status of each. The main
inequality (M) is still quoted.
1. What a zero forces
For Re s > 1, termwise integration gives
ζ ′ (s) 1 1+γ
Z ∞
F (s) = G(x)x−s−1 dx = − − + . (II.1)
1 s(s − 1)ζ(s) (s − 1)2 s−1
Indeed, the transform of (x/n − 1)+ is n−s /[s(s − 1)], where v+ = max(v, 0). The Laurent expansion
of ζ at 1 cancels both terms of the principal part there. A zero ρ of multiplicity ν in the critical
strip gives F a simple pole of residue −ν/[ρ(ρ − 1)]. The poles at s = 0 and at the trivial zeros
−2, −4, . . . lie outside the half-plane Re s > θ used below. The poles in the critical strip are nonreal,
since the alternating zeta series gives ζ(σ) < 0 for 0 < σ < 1.
A bound valid for every large x would allow the usual oscillation argument of Landau. We know
the bound only on certain intervals. To use one such interval, we differentiate the transform many
times. Its new weight is concentrated on that interval, and a nearby pole forces large derivatives.
Lemma (the one-sided criterion on intervals; §I.1). Let 0 < θ < 1. If G(x) ≤ xθ throughout
[Y, Y 3/2 ] for arbitrarily large Y , then ζ has no zero in Re s > θ.
Write u = log x, g(u) = G(eu ) and, for each integer j ≥ 0, let Aj (s) = (−1)j F (j) (s)/j! be the
signed jth Taylor coefficient. Let η = max(0.8e0.2 , 1.2e−0.2 ) = 0.9825 . . . and c = 0.008, which is
less than log((1 + η)/(2η)) = 0.00887 . . .; these are absolute constants. Fix σ > 2 so large that
(σ − θ)/(σ − 2) ≤ ec . With the constants above this holds for σ ≥ 252 and every θ in (0, 1).
Differentiating the Mellin integral gives
Z ∞
Aj (σ + it) = g(u)e−itu Kj (u) du, Kj (u) = uj e−σu /j!. (II.2)
0
26
On each interval satisfying the lemma’s bound, put U = log Y and choose j nearest to 54 (σ − 2)U .
Then Z ∞
|Aj (σ + it)| + Aj (σ) ≤ 2 max(g(u), 0)Kj (u) du ≪ (σ − θ)−j , (1.1)
0
uniformly for real t. The first inequality follows from the triangle inequality and |g|+g = 2 max(g, 0).
On [U, 3U/2] integrate eθu Kj (u).
Outside the window, use |g(u)| ≪ e2u , since partial summation of Chebyshev’s bound n≤x Λ(n) ≪
P
x gives n≤x Λ(n)/n ≪ log x and hence |G(x)| ≪ x log x for x ≥ 2. The density proportional to
P
uj e−(σ−2)u is concentrated about j/(σ − 2) ∼ 5U/4.
Claim. For U large and j nearest to 54 (σ − 2)U , the part of 0∞ uj e−(σ−2)u du outside [U, 3U/2]
R
is at most C0 e−cj times the whole, with C0 independent of U . Indeed, on putting v = (σ − 2)u/j
the density is proportional to (ve1−v )j , and the window becomes an interval in v whose ends
tend to 4/5 and 6/5 as U → ∞. The base ve1−v has its unique maximum 1 at v = 1. Off the
window it is bounded by a number η ′ < 1 that tends to η as U → ∞, so the tail integral is at
R ∞ 1−v
most η ′ j−1
0 ve dv. A small fixed interval about 1 has base at least (1 + η ′ )/2. So the fraction
outside is O (2η ′ /(1 + η ′ ))j , and this is O(e−cj ) for U large, because c < log((1 + η)/(2η)). The
tail is therefore O(e−cj (σ − 2)−j−1 ) = O((σ − θ)−j ), by the choice of σ. This proves (1.1) along a
sequence j → ∞.
Suppose now there is a pole to the right of θ. Let R be the smaller of σ − θ and the distance from
σ to the nearest pole in Re s > 0. Some pole has horizontal distance from the line Re s = σ strictly
less than R. If R = σ − θ, this follows from the supposition. Otherwise take a nearest pole to σ and
use its nonzero imaginary part. Take t0 = Im ρ for that pole ρ; then the pole nearest to σ + it0 is
at distance less than R. Only finitely many poles lie within R + 1 of it, and each pair is equidistant
from at most one point of the line Re s = σ (both lie to its left). So t near t0 can be chosen so that
z = σ + it has a unique nearest pole, at distance d < R.
Choose d < r < R so that the disk of radius r about z contains just that pole. Subtracting its
principal part and applying Cauchy’s estimate gives |Aj (z)| ≥ c1 d−j − O(r−j ) with c1 > 0 (the
principal part c/(s − ρ) contributes exactly c (z − ρ)−j−1 to Aj (z)). At the real centre σ the same
estimate gives |Aj (σ)| ≪ r−j . This contradicts (1.1) for large j, since d < r < σ − θ. This proof
uses only the poles in two bounded disks. The Lean proof [1] reaches the same statement by a
slightly different route.
27
2. How divisibility counts primes
Start with a polynomial P (X) of degree D. If P has distinct roots α1 , . . . , αs , each of multiplicity
at least m ≥ 1, then sj=1 (X − αj )m divides P . Comparing degrees gives ms ≤ D.
Q
Stepanov uses the same count for a nonzero rational function on a smooth projective curve
over a finite field. Count points over the algebraic closure of that field (the curve is assumed
geometrically irreducible), and include the points at infinity. For a nonconstant function the zeros
and poles are the inverse images of 0 and of infinity under its map to the projective line. Each
inverse image, counted with multiplicity, has the degree of that map. Thus the total multiplicity
of the zeros equals the total order D of the poles. Hence at most ⌊D/m⌋ points can be zeros of
order at least m. On the projective line the function is a quotient of polynomials. A polynomial of
positive degree D has just one pole, of order D at infinity.
Over the integers, write vp (c) for the exponent of a prime p in a nonzero rational number c
(numerator exponent minus denominator exponent), and put vp (0) = +∞. The corresponding
count is
ep log p ≤ log |∆|, ∆ ∈ Z \ {0}, 0 ≤ ep ≤ vp (∆). (II.3)
X
p
To bound the left side one needs an independent upper bound for |∆|. For a square integer matrix
of nonzero determinant ∆, for example, divisibility gives a lower bound, while the lengths of its
columns give an upper bound by Hadamard’s inequality. Section 6 explains the corresponding
comparison for lattices. (The file [2] states such an inequality for an earlier form of the argument.
That form is stated in the archive but, according to our dependency check, is not used by the final
theorem.)
Why odd powers of a prime occur. Here is the local model in rational arithmetic. For a prime
p form the series
2j
l(X) = X + (II.4)
X
X p /pj .
j≥1
2j
It has a unique inverse under substitution, U (z) = z+· · · , given recursively by U = z− j≥1 U p /pj .
P
Put L = p2 − 1. We claim that the coefficient of z a in U has valuation at least −(a − 1)/L for
each integer a ≥ 1. This holds for a = 1. In a product of r = p2j copies of U , the degrees i1 , . . . , ir
contributing to z a are positive and sum to a. Each is less than a. By induction their coefficients
have product of valuation at least −(a − r)/L. Division by pj costs j ≤ (r − 1)/L, proving the
claim.
For a series f with integer coefficients, the same product calculation gives the bound −(a − 1)/L
for the coefficient of z a in f (U (z)). Its ath derivative at zero is a! times this coefficient. Counting
multiples of p, p2 , . . . in a! gives vp (a!) = h≥1 ⌊a/ph ⌋ = a/(p − 1) + O(1 + logp a), where ⌊y⌋ is the
P
greatest integer at most y and logp a = log a/ log p. The derivative therefore has valuation at least
1 1 1 1 1
a − 2 − O(1 + logp a) = a + 3 + 5 + ··· − O(1 + logp a). (II.5)
p−1 p −1 p p p
The series models differentiation on an elliptic curve with complex multiplication (extra endomor-
phisms) at primes remaining prime in its imaginary quadratic field. The source proves a weighted
version for its arithmetic maps. At odd powers ph it yields the term (a/ph − w)+ for a positive
weight w (quoted; see section 6.4). When w = b and a = bx, this is precisely b(x/ph − 1)+ . [3]
28
3. The reduction to one prime sum
For an odd prime D let χD (r) equal 0 if D | r, 1 if r is a nonzero square modulo D, and −1
otherwise. For a nonprincipal Dirichlet character χ put
L′
cχ = (1, χ), Gχ (x) = Λ(r)χ(r)(x/r − 1) + cχ x. (II.6)
X
L r≤x
A Dirichlet character is a homomorphism from the invertible classes modulo an integer to the
nonzero complex numbers, extended by zero to other classes. It is nonprincipal if it takes a value
different from 1 on the invertible classes. The function L(s, χ) is the continuation of r≥1 χ(r)r−s
P
from Re s > 1, and the prime denotes differentiation in s.
For distinct odd primes q, d write η = χq χd for the pointwise product. Consider the single prime
sum
Sq,d (x) = Λ(r)(1 + χq (r))(1 − χd (r))(x/r − 1). (II.7)
X
r≤x
Away from q, d, the coefficient of Λ(r) before smoothing is 4 if r is a square modulo q and a
nonsquare modulo d, and zero otherwise. The exact identity obtained by expanding the two factors
is
Rq,d (x) = Sq,d (x) − x log x + (1 + γ + cχq − cχd − cη )x
(II.8)
= G(x) + Gχq (x) − Gχd (x) − Gη (x).
Thus a one-sided bound for Rq,d and absolute bounds for three quadratic character sums suffice.
The reduction. Fix 0 < ε < 1/52. By Theorems III.1.1 and III.2.1 of Part III, for every large Y
there are primes q ≡ 3 mod 8, d ≡ 7 mod 16 with q | d + 1 for which statement 2 below (Estimate
B) and formula (5.6) (Estimate C) hold for Y ≤ x ≤ Y 3/2 . Statement 1 (Estimate A) is then
deduced in sections 4 and 5 from the main inequality (M), which is quoted, together with Estimates
B and C.
1. The residue sum. There are positive integers m, k, n depending on x such that
x xm k
Rq,d (x) ≤ Cε x Cε
+ +m+ 2 . (A)
m k n
Here C is absolute and Cε depends only on ε. The orders are m = x1/2+O(ε) , k = x9/2+O(ε) ,
n = x2+O(ε) . Sections 4 and 5 derive this from (M).
2. The three twists. For χ = χq , χd , η the selection gives |Gχ (x)| ≤ x1/2+ε . To see what is
estimated, put Ψχ (t) = r≤t Λ(r)χ(r). Partial summation and r≥1 Λ(r)χ(r)/r = −cχ give
P P
Z ∞
dt
Gχ (x) = −x Ψχ (t) 2 . (II.9)
x t
The convergence here follows from the prime number theorem for a fixed nonprincipal character.
The proof bounds the proportion of bad choices by mean-square estimates, first in q, then in d for
each good q. This gives a common pair on the whole interval. It is given in Part III, §III.1, with
the exponent 12 + 2ε ; the Lean proof is in [9, 10, 29].
√
Conclusion. The terms x/m and m balance at x. With the stated orders, k/n2 = x1/2+O(ε) and
′
xm/k is smaller. By the identity above, G(x) ≤ Cε′ x1/2+C ε on these intervals, with an absolute
C ′ . Given 1/2 < θ < 1, choose ε so that 1/2 + C ′ ε < θ. On all sufficiently large selected intervals
G(x) ≤ xθ . Section 1 excludes every zero to the right of θ. Letting θ decrease to 1/2 and using the
symmetry ρ 7→ 1 − ρ from the functional equation proves RH from (M) and Estimates B and C. [1,
6, 12]
29
4. The main inequality (M)
4.1 Which prime powers are counted?
Fix an interval and primes q, d as in section 3. Fix x in the interval. For an odd prime power r = ph
with p =
̸ q, attach two labels. The first is the row v(r) = (1 + χq (r))/2, equal to 1 if r is a nonzero
square modulo q, and 0 otherwise. For the second, choose n = 2e at the scale x2+8ε ≤ n ≤ 2x2+8ε
and put N = 4n. Write [r] = {r, −r} mod N . The n pairs of odd residues form a group H under
multiplication. Its elements will label columns. Choosing a fixed t ∈ H puts r in column t[r].
The coefficient at r = ph , before inserting a weight, is
c(r) = (1 − χd (r)) log p. (II.10)
If p =
̸ d it is 2 log p for an odd power of a nonsquare modulo d and zero otherwise. At p = d it is
log d for every power. The two omitted primes 2, q have χd (2) = χd (q) = 1. Indeed d ≡ 7 mod 8
gives the first identity. For the second, q, d ≡ 3 mod 4 and d ≡ −1 mod q, so quadratic reciprocity
gives (q/d) = −(d/q) = −(−1/q) = 1. So c(r) vanishes at every power of 2 and of q; omitting these
primes changes nothing.
A table small enough to write out. Take q = 3, d = 23, N = 16 and t = [1]. The four columns
are [1], [3], [5], [7]. For example [3]2 = [9] = [7] and [3]3 = [5]. The nonzero squares modulo 23 are
1, 2, 3, 4, 6, 8, 9, 12, 13, 16, 18. In each cell below are the primes at most 61 with nonzero coefficient.
A dash denotes an empty cell in this range.
v\s [1] [3] [5] [7]
0 17 − 5, 11, 53 23 (II.11)
1 − 19, 61 37, 43 7
Among primes the density in each cell is 1/16, as shown in section 5. To place 11, observe that
11 ≡ 2 mod 3 and [11] = [5]. To place a higher power one repeats both tests on that power. The
prime 7 goes in (1, [7]) with coefficient 2 log 7. Its square disappears because χ23 (72 ) = 1. Its cube
has the same row and column as 7. At the exceptional prime 23, its first power lands in (0, [7]),
its square in (1, [1]), and each has coefficient log 23. The example only illustrates the identities of
section 5, which hold for any positive weights; its parameters do not satisfy the conditions of (M).
30
4.2 How a weight determines a cutoff
Let b = exp(xε ). The construction supplies nonempty subsets T, T2 of H, an integer at > 0 for
each t ∈ T with |at /b − x| ≤ 1, and m maps of initial rank k (the chain of theta lifts of section 6.2;
m and k are the parameters in (A)). Write |S| for the cardinality of a finite set S and δ = |T2 |/n.
For each translate u ∈ H and map 1 ≤ j ≤ m there are 2n positive weights wu,j (v, s). In the
construction they are integers. The distinguished cells are (1, s) with s ∈ uT2 . Their weights satisfy
b ≤ wu,j (1, s) ≤ b + O(b/(nx2 )). At every other cell the ratio at /wu,j (v, s) is at least exp(xε/3 )/3.
Thus the same table tests primes at two markedly different scales.
Fix u, t, j, write a = at and w = wu,j . Each power in cell i = (v, s) contributes
c(r)(a/r − wi )+ = c(r)wi a/(rwi ) − 1 + , z+ = max(z, 0). (II.12)
The contribution ends at r = a/wi . Adding all cells gives the finite sum
Pu,t,j = (1 − χd (ph )) log p at /ph − wu,j (v(ph ), t[ph ]) + . (4.1)
X X
p odd prime h≥1
p̸=q
In the running example take T = T2 = {[1]}, one table, x = 1000, b = 10, a = 10000 and u = [1].
Here δ = 1/4. The weights and cutoffs, in the same column order as above, are
wi [1] [3] [5] [7] a/wi [1] [3] [5] [7]
0 0.05 0.05 0.05 0.05 0 200000 200000 200000 200000 (II.13)
1 10 0.05 0.05 0.05 1 1000 200000 200000 200000
For example 7 contributes 2 log 7 (10000/7−0.05), whereas 232 = 529 contributes log 23 (10000/529−
10). Translating by u = [7] moves the weight 10 to column [7] in row 1. Now the subtracted weight
at 7 is 10, and at 529 it is 0.05.
For a reproducible finite calculation run through the primes p ≤ 200000 and the successive powers
p, p2 , . . . ≤ 200000. Drop 2, 3, compute the coefficient and the cell as above, and add its positive
part. This gives P[1],t,1 = 95461.6156 to four decimals. The same number will be obtained by
grouping the terms in a different order in section 5.
31
4.3 What the arithmetic construction must prove
The quantity to subtract from (4.1) is
1 X
⟨log w⟩ = log wi , Mu,t,j = at log at − 1 − γ + cχd − ⟨log w⟩ . (4.2)
2n i
Section 5 derives each term by ordinary prime summation. Section 6 explains why the arithmetic
estimate on the maps yields the same expression. The match through the term linear in at is
essential.
Every translate has seven weights equal to 0.05 and one equal to 10. Substitution in (4.2) gives
M = 10000 log 10000 − 1 − γ − 0.422120948038 . . .
(II.14)
− (log 10 + 7 log 0.05)/8 = 95444.4636 . . . .
For u = [1] the excess is 17.1520 to four decimals. The other three excesses are 368.2081, −129.2546,
and −204.6531, giving a mean of 12.8631. (M) bounds their average over u, summed over the labels
t and the maps j.
The main inequality (M). On the selected intervals the source constructs the above data
with δ ≥ x−ε , |T |/n ≥ x−ε/8 /2, and an upper error E ≥ 0. (Sets T2 , T with these densities are
constructed in Part III, §III.3. The rest of the data is quoted.) With B = |T |mbδ/2 it proves
m
1 X XX E x xm k
(Pu,t,j − Mu,t,j ) ≤ E, ≤ xCε + +m+ 2 . (M)
n u∈H t∈T j=1 B m k n
Here C is absolute, and the inequality is asserted for x ≥ x0 (ε). The first inequality compares
integral divisibility with complex size. To check the powers in the second, put h = |T |, h2 = |T2 |
and let aE be the aggregate scale in the source with 0 ≤ aE ≤ 2xb. An unnormalized error estimate
stated in [6] (a parallel formulation; see the start of Part III) is
h2 kbxε
E ≤x Cε
haE (1 + m /k) + m 2nmbx +
2 ε
(II.15)
n(nx−ε )
for an absolute C. The bound for E/B in (M) is thus taken from a statement of the archive on which
the final theorem does not depend in that form (see the start of Part III). Division by B = hmbδ/2
bounds the three bracketed contributions respectively by
4 x xm 4n n k 2ε
+ , mxε , 2 x . (II.16)
δ m k hδ h n2
Here h2 = nδ. The bounds on n/h and 1/δ give (M). This computation explains why multiplicity
matters. A cost incurred once yields x/m after division, while the lattice comparison incurs a cost
√
of order m after division. Their balance occurs at x. [3, 5, 6]
32
5. From the table to a single prime sum
5.1 Subtracting the main term cell by cell
Keep t fixed; the quantities Si , Ψi , Ci , Ri below depend on t, through the column t[r]. Let r∈i
P
denote a sum over prime powers assigned to cell i by section 4. The smoothed sum in that cell and
its unsmoothed counting function are
Si (y) = c(r)(y/r − 1)+ , Ψi (z) = (II.17)
X X
c(r).
r∈i r∈i
r≤z
For y ≥ 1, integrating each step of Ψi proves the first identity below. The second follows term by
term from (4.1). Z y
dz
Si (y)/y = Ψi (z) 2 , Pu,t,j = wi Si (a/wi ). (5.1)
X
1 z i
For fixed moduli, among primes not dividing 2qd the fraction contributing to a given cell is 1/(4n).
One half are nonsquares modulo d, one half have the specified sign modulo q, and 1/n have the
specified pair modulo N . The Chinese remainder theorem makes these conditions independent
among the invertible residue classes. The prime number theorem in progressions gives their density
among primes. The coefficient 2 log p doubles this density in Ψi . Thus Ψi (z) has main term µz,
where µ = 1/(2n).
The fixed-modulus prime number theorem gives an error OA (z/(log z)A ) as z → ∞ for each fixed
A > 0. Higher prime powers and the one exceptional prime fit within this bound. The constants
may depend on the moduli.
Set Ci = 1∞ (Ψi (z) − µz) dz/z 2 , which converges by that estimate. Subtract the integral from y to
R
infinity to obtain the exact formula
Z ∞
dz
Si (y) = µy log y + Ci y + Ri (y), Ri (y) = −y (Ψi (z) − µz) 2 . (5.2)
y z
In particular Ri (y) = o(y) for fixed moduli, meaning Ri (y)/y → 0. The uniform estimate needed
when the moduli vary with x comes later.
What is i Ci ? Write S1 (y) for the ordinary smoothed von Mangoldt sum and Sχ (y) for the same
P
sum with coefficient χ(r). Then
Si (y) = S1 (y) − Sχd (y). (II.18)
X
i
This identity includes powers of d, and both sides vanish termwise at p = 2, q. For a nonprincipal
character, partial summation and the convergent series for −L′ (1, χ)/L(1, χ) give Sχ (y)/y → −cχ .
For the ordinary sum, the constant in r≤y Λ(r)/r − log y is −γ, by the Laurent expansion of −ζ ′ /ζ
P
at 1 and partial summation. Subtracting y −1 r≤y Λ(r) → 1 gives the linear coefficient −1 − γ.
P
Comparison with (5.2) yields
Ci = −1 − γ + cχd , Pu,t,j − Mu,t,j = wi Ri (a/wi ). (5.3)
X X
i i
For the last equality expand µa i log(a/wi ) = a log a−a⟨log w⟩ and insert the sum of the constants.
P
33
5.2 Arithmetic check of the centering
For q = 3, d = 23, n = 4, t = [1], every cell has µ = 1/8. The constants and sums below allow
(5.1)–(5.3) to be checked on the individual cells. Here y0 = 1000 and y1 = 200000.
v s Ci Si (y0 ) Si (y1 ) Ri (y0 ) Ri (y1 )
0 [1] −.372219 523.602 230456.699 32.352 −251.387
0 [3] −.631093 224.352 178879.838 −8.025 −53.464
0 [5] .565967 1441.249 418604.451 11.813 259.312
0 [7] −.406444 417.613 224057.714 −39.412 194.742 (II.19)
1 [1] −.575353 291.947 190566.123 3.831 484.900
1 [3] −.202447 694.308 264017.194 33.286 −645.209
1 [5] −.272849 578.281 250761.014 −12.339 179.021
1 [7] −.104899 737.266 284065.998 −21.305 −106.090
Summing the eight constants before rounding gives −1.999336613 . . . = −1 − γ + cχ23 . For u = [1]
multiply the (1, [1]) entry at y0 by 10, and each of the other seven entries at y1 by 0.05. Doing this
in the S columns, with the entries before rounding, gives 95461.6156. Doing it in the R columns
gives 17.1520. The same operations at the four translates give
u Pu,t,1 Mu,t,1 Pu,t,1 − Mu,t,1
[1] 95461.6156 95444.4636 17.1520
[3] 95812.6718 95444.4636 368.2081 (II.20)
[5] 95315.2090 95444.4636 −129.2546
[7] 95239.8106 95444.4636 −204.6531
How to compute Ci without √ fitting data. For this table let ξ be the character of the pairs
modulo 16 with ξ([3]) = −1. For i = (v, s) put β = 2v − 1. As a function of a prime power, the
coefficient of cell i divided by Λ has the following finite character expansion:
1 X 3
1 X
β j (−1)l ξ(s)k (χj3 χl23 ξ k )prim . (II.21)
8 j,l=0 k=0
The subscript uses the character of smallest modulus inducing the product on units, so a factor
with exponent zero is the constant one even at its own prime. For units, the two sign sums are
each 2 when the signs agree and zero otherwise. The four powers of ξ sum to 4 precisely in column
s. This gives coefficient 2. At 23, terms with l = 1 vanish, giving coefficient 1 in the prescribed
cell. At 2, 3 the χ23 factor sums to zero. Substituting the linear coefficient −cχ for each nontrivial
character and −1 − γ for the trivial one gives the displayed Ci . For general t, n use the column
t−1 s, the n characters of H, and coefficient 1/(2n). [3]
34
5.3 What translation preserves
Fix t, j. First replace each distinguished weight by b. For a given column s, the equation s = uz,
z ∈ T2 , gives exactly |T2 | choices u = sz −1 . Thus for each admissible prime power r
1 X δ
1{v(r)=1, t[r]∈uT2 } = δ 1{χq (r)=1} = (1 + χq (r)). (II.22)
n u∈H 2
Here 1 equals one when its condition holds and zero otherwise. Multiplication by 1 − χd (r) gives
the right contribution at the three special primes as well, with zero understood at the two omitted
primes. Hence, with a = at ,
1X δb
bSi (a/b) = Sq,d (a/b). (5.4)
X
n u i distinguished at u 2
This is an identity for every common argument y = a/b. The map sending a sum with expansion
αy log y + βy + o(y) to its remainder after subtracting those two terms is linear. The coefficients
are unique by taking limits. Applying it to (5.4) gives
1X δb
bRi (a/b) = Rq,d (a/b). (5.5)
X
n u i distinguished at u 2
In particular, with η = χq χd , the right side involves G + Gχq − Gχd − Gη . The plus sign on G
comes from expanding (1 + χq )(1 − χd ).
In the example each of the four row-1 cells is distinguished once. From the Si (1000) column we get
4 (291.947014 + 694.307989 + 578.280814 + 737.265890) ≃ 5754.504268. (II.23)
10
Direct summation with the two quadratic signs gives the same value (10/8)S3,23 (1000). For (5.5),
the four remainders give 8.679631 . . . to six decimals, or (10/8)R3,23 (1000). There is a separate
check using the three L-values:
y G G χ3 Gχ23 Gχ3 χ23 R3,23
1000 1.531 1.040 1.332 −5.704 6.944 (II.24)
1000000 11.040 −16.153 −97.818 −657.145 749.850
To six decimals the three constants are cχ3 = 0.368282, cχ23 = −0.422121, and cη = 0.056523. They
can be computed by grouping the Dirichlet series into residue classes and differentiating its Laurent
expansion at 1. The mean excess in section 5.2 is 12.863118 to six decimals. Its distinguished part
accounts for 8.679631 and the other cells for 4.183488, rounded separately. The latter need an
estimate in the asymptotic proof. [3, 6]
35
5.4 Discarding the other scale with a uniform estimate
Divisibility uses all the cells, but only the distinguished cells should measure primes near x. Small
weights at the other cells push their cutoffs far enough out for the prime number theorem to control
them. The fixed-modulus limit Ri (y)/y → 0 gives no rate uniform in q, d, N . For the moduli of
section 3 one has, for each primitive character in the cell expansion and for y ≥ exp(xε/3 )/3,
|Gχ (y)| ≤ y ϑ, ϑ = exp(−κxε/8 ), κ > 0, (5.6)
with G for the trivial character. The constant κ does not depend on x, q or d, and the bound is
for x ≥ x0 (ε, κ). This is Estimate C. It is proved in Part III, §III.2 (Theorem III.2.1), where κ is
called c′ . At any nondistinguished cell i use (5.6) at y = a/wi . Each remainder is a combination
Ri = χ fχ Gχ of the functions Gχ for the characters of the table, with χ |fχ | = 2: the sum of
P P
the absolute values of the 4n Fourier coefficients is 4n/(2n) = 2. Therefore
1 X nx
|wi Ri (a/wi )| ≤ 2aϑ, |wi Ri (at /wi )| ≪ ϑ = o(1). (II.25)
X
Bn u,t,j i nondistinguished δ
Indeed there are at most 2n cells per table, at /b ≤ x + 1, and nx/δ grows polynomially in x.
Exponential decay in (5.6) absorbs this factor.
For a distinguished cell let wi = b + ∆i with 0 ≤ ∆i ≪ b/(nx2 ). The function w 7→ (a/r − w)+
changes by at most ∆i and only r ≤ a/b can contribute. Since 0 ≤ c(r) ≤ 2Λ(r), Chebyshev’s
estimate bounds the total change in the prime sum by O(x∆i ). In its main term aCi is unchanged,
and µa log(wi /b) ≤ µa∆i /b. Consequently
1
|wi Ri (a/wi ) − bRi (a/b)| ≪ x∆i , x∆i ≪ 1/x. (II.26)
X
δb/2 i distinguished
For the second bound, there are δn = |T2 | distinguished cells in each table. Summing over u, t, j
and dividing by Bn gives O(1/x).
Finally |at /b−x| ≤ 1. When y moves by at most 1 from x, Sq,d (y) and y log y change by O(1+log x),
since r≤x+1 Λ(r)/r ≪ log x. For χ = χq , χd , η, Theorem III.1.1(d) of Part III gives |cχ | ≤ 4 log x
P
(it also follows from Estimate B and the definition of Gχ ). The same estimate for the variation
thus holds for the centering terms. Combining (5.3), (5.5), and the two error bounds gives
1 X
(Pu,t,j − Mu,t,j ) = B Rq,d (x) + Oε B(1 + log x) . (5.7)
n u,t,j
With (M), this gives Rq,d (x) ≤ E/B + Oε (1 + log x), proving (A). Only an upper bound follows.
In the finite example the changes of distinguished weight and argument are both zero, so before
estimation (5.7) is the numerical split of the mean excess into 8.679631 and 4.183488. [6, 11]
36
6. Where the main inequality (M) comes from
Sections 6.1–6.2 construct the maps from theta series. Sections 6.3–6.4 explain why their derivatives
are divisible by powers of inert primes. Sections 6.5–6.7 bound the size of the same derivatives and
identify the main term. Sections 6.8–6.10 control denominators and show that enough independent
images survive. We prove the elementary calculations in each step and say which arithmetic
statements are quoted.
6.1 A theta lift in classical terms
A theta series records representations of integers by a quadratic form. For instance put L = Z[i],
Q(a + bi) = a2 + b2 and z in the complex upper half-plane. With s = exp(2πiz), its theta series is
ΘL (z) = sQ(λ) = 1 + 4s + 4s2 + 4s4 + 8s5 + · · · . (II.27)
X
λ∈L
The coefficient of s5 is 8 because the solutions are (±1, ±2) and (±2, ±1). Absolute convergence
follows from a,b exp(−2π ℑz (a2 + b2 )) < ∞. Fourier orthogonality recovers a coefficient by pairing:
P
for each y > 0, Z 1
e10πy ΘL (u + iy)e−2πi5u du = 8. (II.28)
0
R 1 2πi(k−5)u
Indeed 0 e du is one for k = 5 and zero for every other integer k.
To make this a lift, allow the lattice to vary within a finite collection of isometry classes Li . Attach
to each class a value f (i). Weight it by hi = 1/| Aut Li |, the reciprocal of the number of its
symmetries in the chosen group. The theta series are now a kernel depending on both i and z. The
lifted function is
Θ(f )(z) = hi f (i) ΘLi (z). (II.29)
X
i
In the rank-one Hermitian example, unitary symmetries of Z[i] are multiplication by 1, −1, i, −i. On
its single class with f = 1 this normalization gives ΘL /4, so the same Fourier pairing gives 8/4 = 2.
This example shows what integration against a theta kernel does. The high-rank construction has
different groups and test functions.
−1 √
√
For the source let E = Q(ζN + ζN , −q, −d), with ζN = exp(2πi/N ), and let F denote its
maximal real subfield. Write for complex conjugation on E. A Hermitian matrix J satisfies
t
J = J. Its unitary group consists of matrices g with g t Jg = J. It is definite if its form at every
real embedding of F is definite. At a fixed integral level and scalar weight, forms on a definite
group can be viewed as functions on finitely many lattice classes. At higher weight the values are
vectors and must transform under the finite stabilizers.
For a pair of unitary groups, the general lift has the same two operations:
Z
Θφ (g, h) = (ω(g, h)φ)(λ), Tφ f (h) = Θφ (g, h)f (g) dg. (II.30)
X
[U (J)]
λ∈V (F )
Here V is a finite-dimensional coordinate space for the theta sum; V (F ) denotes its rational points
over F . The notation [U (J)] denotes integration over the group at all places modulo its rational
points, with its invariant measure. The test function φ decays rapidly at infinity and is locally
constant of compact support at finite primes. The Weil representation ω is the action of the two
groups on these test functions. Lattice indicators at the finite primes restrict the sum to lattices.
37
In the displayed convention the lift is linear in φ and conjugate-linear in f . [13] (The linearity
statement of [13] is a statement in the archive, not used by the final theorem.)
38
6.2 Composing maps and clearing their denominators
For a finite form of the calculation let Ij be the finite set of classes at stage j, hj (i) its weights,
and Kj (i, i′ ) the matrix of the jth map. In the linear convention write
(Tj f )(i′ ) = hj (i)Kj (i, i′ )f (i). (II.31)
X
i∈Ij
Complex conjugates from a pairing are absorbed by taking the conjugate source space. Expanding
two such sums gives the composite kernel i′ Kj (i, i′ )hj+1 (i′ )Kj+1 (i′ , i′′ ). Thus successive maps
P
really do use the same intermediate index.
In the source the definite ranks are Nj = k + j(τ0 + 1) for an odd integer τ0 . The rank conditions
[26] include m even and m ≥ 2, k even and k ≥ m + 3, τ0 ≥ k + m + 4, and τ0 > 3m. For instance
m = 2, k = 6, τ0 = 13 gives ranks 6, 20, 34 and satisfies these numerical constraints. The full chain
also needs field, level, and lattice data. Since m is even and m ≥ 2, the chain has at least two steps.
At the end of the chain the complex pairing of the entire chain can be expressed as a single period
integral over a subgroup. The source proves this identity with prescribed normalization factors
and coordinate changes. Its use is that the final period can be compared at different choices of the
intermediate character twists. [7, 14]
The integrality assertion for a finite chain. Let S be a subring of a commutative ring R.
Suppose all Kj (i, i′ ) and starting values belong to S. If scalars Dj ∈ R satisfy Dj hj (i) ∈ S for
every i, then every final value multiplied by 1≤j≤m Dj belongs to S. For one map this follows by
Q
moving D1 into its finite sum. Applying the same argument to the already cleared values proves
the statement by induction. For vector values, replace S by a lattice at each stage and require each
Kj (i, i′ ) to carry the preceding lattice into the next. This is the assertion stated as ChainInt, with
its vector version, in [14] (a statement in the archive, not used by the final theorem).
For example, on one class let the successive weights be 1/2, 1/3, the kernel entries 5, 7, and the
starting value 1. The output is 35/6, and clearing the two weights gives 35. At the prime 5, two
entries divisible respectively by 52 and 53 would give a cleared output divisible by 55 . In a sum the
exponent can increase through cancellation. It is always at least the common lower bound on the
summands. This elementary induction explains how divisibility persists. It requires separate local
estimates to put significant prime powers into each kernel entry.
Where a period enters. A kernel evaluated in complex coordinates need not be integral. An
algebraic local coordinate and a complex analytic coordinate on the same elliptic curve differ at
the tangent space by a nonzero scale. An ath derivative acquires its ath power. Taking logarithms
gives a term linear in a. In the global comparison the local gains, the period, and the lattice indices
must all pertain to the same normalized maps. We next give a local example in which one can
track every derivative, and then the estimates that take its place in the chain.
39
6.3 A genuine local example and an elementary minor calculation
Let s(z) be the inverse near 0 of t 7→ 0t (1 − u4 )−1/2 du, and let Aj = s(j) (0). The lemniscate
R
curve y 2 = 1 − x4 has the symmetry (x, y) 7→ (ix, y) fixing the origin (0, 1). This yields complex
multiplication by Z[i]. Differentiating the inverse function gives
s(0) = 0, s′ (0) = 1, (s′ )2 = 1 − s4 , s′′ = −2s3 . (II.32)
The product rule gives an integer recursion, with A0 = 0, A1 = 1,
j!
Aj+2 = −2 (II.33)
X
Aa Ab Ac .
a+b+c=j
a! b! c!
For j = 3, only (1, 1, 1) contributes, giving A5 = −12. For j = 7 there are three permutations of
(5, 1, 1), giving A9 = −2 · 3 · 42 · (−12) = 3024. Uniqueness of the series solution gives s(iz) = is(z),
so only indices j ≡ 1 mod 4 can contribute. The exact computation gives
A17 = 21224560896 = 28 36 72 11 211. (II.34)
The primes 3, 7, 11 remain prime in Z[i]. Their exponents are 6, 2, 1; the model in section 2 predicts
the respective leading terms 17 · 3/8, 17 · 7/48, 17 · 11/120. At the split primes 5, 13 the exponents
here are zero. At j = 201 the computed valuations at 3, 7 are 75, 29. These valuations are computed,
not proved in general; section 2 proves the bound only for its model series.
There is a quantitative gap between this example and (4.1). At j = 101 the contribution to log |Aj |
from inert primes at most j is 167.67 out of 306.77. At j = 121 it is 203.58 out of 388.67. The
density calculation explains the scale: p≤j, p≡3 mod 4 (log p)/p ∼ 12 log j, so the model exponents
P
j/p at these primes give about 12 j log j. The main inequality (M) requires agreement even to the
term proportional to j. So the numbers Aj cannot serve in (M): they are too large by a factor 2 in
the leading term (defect (D3) of §I.3).
A local calculation from the source. Let A be a commutative ring with an element π0 satisfying
π02 = p, where p is prime. For an integer e ≥ 0 put P = pe . Let M, N ′ be square matrices with
Nbc′ ≡ M P mod p. Suppose all but the zeroth coordinates of row vectors u, u′ are divisible by π .
bc 0
Write X = uM , Y = u′ N ′ . The minor used in [15] satisfies
π0 | Yc X0P − Y0 XcP . (II.35)
Indeed, modulo √ π0 , Xc = u0 M0c and Yc′ = u01M
′
0c . Substitution makes the two products equal. For
P
p = 3, π0 = 3, take P = 3, M = N = ( 0 1 ), u = (1, π0 ), u′ = (1, 0). The minor for c = 1 is
1
1 − (1 + π0 )3 = −3π0 − 3π02 − π03 , visibly divisible by π0 . The estimate on the full lift requires many
such congruences and their behavior under differentiation.
40
6.4 How the weight cuts down the local exponent
Here is a calculation which accounts for the positive part in (4.1). A directional Hasse derivative of
a polynomial Q is the coefficient of Z β in Q(X + j bj Zj ), where bj are vectors, Zj indeterminates,
P
and β a list of nonnegative exponents. It has no factorial denominator. For a product c Qc the
Q
coefficient is the sum of products over decompositions β = c βc . If for a given set S of directions
P
the coefficient for Qc vanishes whenever j∈S (βc )j > mc , then the coefficient for the product
P
vanishes whenever j∈S βj > c mc : each term exceeds the allowance of some factor. This is the
P P
product-vanishing lemma in [21].
For a determinant the allowance has a simple origin. In det(X + sb, Y + tb), where X, Y, b are
columns of length two, the coefficient of st is det(b, b) = 0, and there is no term in s2 or t2 . Each
minor can take at most one such hit. Products of C minors can take at most C. With several
directions confined to specified columns, the number of available columns provides another upper
bound. These are constraints on nonvanishing terms before any prime estimate is made.
For an inert prime p, write γp = p/(p2 − 1) = h≥1, h odd p−h . Suppose a local expansion of
P
derivative order a supplies a valuation lower bound γp a − h ih − D0 , where D0 is a loss, and the
P
nonvanishing terms have
0 ≤ ih ≤ min(a/ph , wh ) (h odd). (II.36)
Then the gain after charging these terms is at least
X
a/ph − min(a/ph , wh ) − D0 = (a/ph − wh )+ − D0 . (6.1)
X
h odd h odd
The first bound ph ih ≤ a is a degree bound. The second is supplied by column and minor capacities
in the translated weight data. For example p = 3, a = 81, wh = 2 gives 25 at h = 1, 1 at h = 3,
and zero thereafter. Obtaining all these bounds together is the arithmetic content. In the source
the factorial valuations in the polynomial substitution give the initial lower bound, other terms
have to be paid for, and a local lattice index must also be subtracted. [19, 21]
The factorial count, with proof. The initial lower bound in the last paragraph comes from
valuations of factorials. Here is the source’s statement in full, with an elementary proof. Let p be
prime and put ρp = 1/(p2 − 1), so that γp + ρp = 1/(p − 1). For an integer n ≥ 0 let λ(n) be the
number of digits of n in base p, with λ(0) = 0. Let S be a finite set and S0 ⊂ S. Suppose given
nonnegative integers as follows:
(i) αs and ℓs for s ∈ S, with αs = 0 outside S0 ;
(ii) lists es,1 , . . . , es,r1 with sum αs ;
(iii) lists as,1 , . . . , as,r0 r1 with sum at most ℓs ;
(iv) is,h for h in a finite set of nonnegative integers, zero outside S0 , with p2h+1 is,h ≤ αs ;
(v) with A = s αs and Λ = s ℓs , one has j as,j + s,h is,h = Λ.
P P P P P
s
Let V be the sum of vp (n!) over all the es,j , as,j , is,h . Then for every integer i2 with pi2 ≤ Λ,
V ≥ γp A + ρp (A + Λ) + i2 − ∆, ∆= r1 λ(αs ) + r0 r1 λ(ℓs ) + λ(αs )2 . (6.1a)
X X
s∈S s∈S0
Proof. Legendre’s formula is vp (n!) = (n − σp (n))/(p − 1), where σp (n) is the sum of the digits of
n in base p. Each digit is at most p − 1, so vp (n!) ≥ n/(p − 1) − λ(n). The integers in the three
families add up to A + Λ, by the two hypotheses on sums. Hence V ≥ (A + Λ)/(p − 1) − λ(n).
P
41
Now count digits. Each es,j is at most αs , and there are r1 of them: at most r1 λ(αs ). Each as,j is
at most ℓs , and there are r0 r1 of them: at most r0 r1 λ(ℓs ). If is,h ≥ 1 then p2h+1 ≤ αs , so h < λ(αs )
and at most λ(αs ) values of h occur; each such is,h is at most αs . These contribute at most λ(αs )2 ,
and only for s ∈ S0 . Thus λ(n) ≤ ∆. Finally (A + Λ)/(p − 1) = γp A + γp Λ + ρp (A + Λ), and
P
γp Λ ≥ Λ/p ≥ i2 . □
In (6.1a) the term γp A is the model gain A(1/p + 1/p3 + · · · ) of section 2. The term ρp (A + Λ)
pays exactly for the denominators 1/(p2 − 1) per unit of degree met there in the inverse series, and
i2 pays for one further division. The loss ∆ is small: per index s it is at most (r1 + r0 r1 + λ)λ,
where λ is the larger of the digit counts λ(αs ), λ(ℓs ). For example, with p = 3, one index s,
r0 = r1 = 1, α = 81, ℓ = 0 and no is,h : V = v3 (81!) = 40, while γ3 A + ρ3 A = 81/2 = 40.5 and
∆ = λ(81) + λ(81)2 = 30. The inequality 40 ≥ 40.5 − 30 holds with room to spare; the main terms
agree to within one. This lemma supplies valuations of individual factorials. It does not identify
which factorials occur in the theta coefficients, which is part of the quoted local theorem. [21]
The coefficient 2 log p. For the coefficient field L0 and a prime ideal p over p, write ep , fp for the
ramification index and residue degree. The normalized local weight is λp = log Norm(p)/[L0 : Q].
The source measures a local gain y in units 2λp ep y. Since Norm(p) = pfp and p|p ep fp = [L0 : Q],
P
one unit at every place over p sums to 2 log p. For example in Q(i) above 3, e = 1, f = 2 and
λ = (log 9)/2 = log 3. The factor 2 is the squared-norm normalisation of §I.4.2(ii). At the ramified
prime d the splitting condition has half weight and gives log d. These are the coefficients c(r) of
section 4.1; section 5.2 checked that the character expansion reproduces them. [3, 22]
The concrete minor, the Hasse product rule and (6.1) are three elementary ingredients of the
local argument. Applying them to all the normalized theta kernels, choosing the usable places,
and bounding the losses after summation requires the source’s local divisibility theorem. In its
statement the left side is the sum of the odd-prime-power expressions minus the orbit and rounding
losses. The right side is the sum of actual local gains over the m maps minus the local lattice
correction. [19] The divisibility at inert primes is written up in Part III, §III.5, conditionally on
the integrality of the theta kernels in the coordinates of the height-two logarithm (hypothesis (B)),
which is taken from the Lean development; under that hypothesis §III.5 proves the lower bound for
the valuation of a product of minors applied to one such function (for Frobenius orbits of length
two; for longer orbits it also uses a duality statement cited from the Lean), with the positive parts
of (6.1), and it proves without hypothesis the estimate of section 2 for the inverse of the logarithm
in several variables; it does not prove the passage to lattices, and the congruence of section 6.3 is
not used there. The passage to lattices is treated in Part III, §III.6: the lemmas on lattices and the
splitting of the prime sum are proved there, the theorem on the glued lattice is conditional on two
statements of the source, and the bounds for the loss sums are quoted.
42
6.5 Which theta function is differentiated?
At step j let Nj = k + j(τ0 + 1) be the rank of the definite unitary group, where τ0 is the odd rank
parameter. The coordinates of the holomorphic theta function, for a pair of lattice classes L, L′ at
consecutive stages and a complex embedding of the coefficient field, are the entries of a rectangular
matrix at each weight place:
z = (zτ,α,β ), 1 ≤ α ≤ Nj , 1 ≤ β ≤ Nj+1 . (II.37)
Write its Taylor series at zero as ΘL,L′ (z) = ν cν z ν , using a multi-index ν = (νg ) with g = (τ, α, β)
P
and ν! = g νg !. For vectors in the consecutive spaces of forms there is a polynomial P (z) = ν Pν z ν
Q P
determined by their weights and pairing. In the formulation of [23] (a statement in the archive; see
the start of Part III) the local expression in a matrix entry is
P (∂)ΘL,L′ (0) = (6.2)
X
Pν ν! cν .
ν
This explains why the polynomial minors and the factorials occur together. At each place the
monomials in P have prescribed total degree. The global matrix coefficient sums the displayed
expressions over preceding classes with reciprocal stabilizer weights. The source also describes the
theta function with algebraic coordinates at finite places so that the same normalized pairing has
the local divisibility. That identification of rational structures is a theorem of the construction.
That the entries used by the final theorem are of the form (6.2) is quoted. [23]
On the complex side the polynomial has the squared Gaussian norm
∥P ∥2F = (II.38)
X
|Pν |2 ν!.
ν
2
Here is a proof of its simplest case. In the complex plane give area the density π −1 e−|z| . Integration
in angle makes different powers orthogonal, and the squared norm of z a is
Z ∞ Z ∞
2a+1 −r2
2 r e dr = ta e−t dt = a!. (II.39)
0 0
The last integral follows by a integrations by parts. For several independent variables the integrals
multiply. For a positive width Ω, the density (πΩ)−1 exp(−|z|2 /Ω) gives the squared monomial
norm a!Ωa . In the dual coefficient estimate the scale appears with the inverse power. Weighted
Cauchy–Schwarz gives the inequality stated in the archive [23]
|P (∂)Θ(0)|2 ≤ ∥P ∥2F (II.40)
X
ν!|cν |2 ,
ν∈supp P
where supp √P consists √
of its monomials with nonzero coefficients. Apply Cauchy–Schwarz to the
two lists Pν ν! and cν ν!.
The analytic estimates bound the coefficient factor using the Gaussian-weighted norm of the theta
function on a fundamental domain. The polynomial factor is bounded by a product of hook factors
of the weights times the squared norms of the two input vectors. In one row of length a the hook
factor is exactly a!. The full operator bound also contains the period, level, multiplicities and theta
mass. [23, 24] The estimates for the two factors are proved in Part III, §III.4, apart from the four
conditional passages listed at the start of that section.
Comparison with section 6.3. In the lattice comparison the archimedean map height is the
average over infinite places v of log ∥Tv ∥2op , where the operator norm uses the two chosen Hermitian
43
metrics. Divisibility by pe thus contributes 2e log p when measured at all places above p, and a degree-
a monomial contributes log(a!) = a log a − a + O(log(a + 1)) in the squared Gaussian norm. The
difference from section 6.3 is visible in Aa = a![z a ]s(z), where [z a ] means the coefficient of z a . The
nearest pole of s is at finite positive distance R; the radius formula gives lim supa |[z a ]s|1/a = 1/R.
Along a subsequence, therefore, log |Aa |2 = 2 log(a!) + O(a), while the squared Gaussian norm
contributes just one factorial. [24, 25]
44
A natural question. Why not differentiate a classical entire theta function, which already
has Gaussian size? For the lemniscatic lattice the Weierstrass function σ of y 2 = 4x3 − 4x
has integer derivatives σ (n) (0); they vanish unless n ≡ 1 mod 4, and the nonzero ones satisfy
log |σ (n) (0)| ≤ 12 n log n + O(n), by the Gaussian growth of σ and Cauchy’s formula. Exact
computation for n ≤ 241 shows that their divisibility at inert primes falls well short of the model:
the exponent of 3 is 13 at n = 101 and 28 at n = 241, against 3n/8 = 37.9 and 90.4. So Gaussian
size does not by itself bring the divisibility with it. The source’s local theorem establishes both
properties together for its theta kernels, in coordinates chosen for that purpose, and it is quoted
here. The computation is in the file e4_sigma.py listed under Computations. [21, 23]
6.6 Size of a derivative, its period, and the count of forms
2
The model for the size of a derivative is this. Let f be an entire function with |f (z)| ≤ S e|z| /2 for
√
all z, where S > 0. Cauchy’s formula on the circle |z| = a gives |f (a) (0)| ≤ S a! (e/a)a/2 for a ≥ 1.
By Stirling’s formula, if f (a) (0) ̸= 0, then log |f (a) (0)|2 ≤ a log a − a + 2 log S + O(log(a + 1)): one
factorial for the square. Rescaling the variable by the period adds −a log ΩE , with ΩE as defined
below. This is the model at the start of Part III, §III.4. It identifies the factorial and scale terms of
(6.5). The bound for the norm of an actual composed map needs additional estimates.
Here is the exact period normalization appearing in the comparison. It is the normalization for the
field E, which is the one of the source. Let rdE = |disc E|1/[E:Q] be the root discriminant of E, and
put v = rdE /2. The normalized period ΩE is the positive number with
log ΩE = γ + log π − log v − cχd = γ + log(2π/rdE ) − cχd . (6.3)
This is the Chowla–Selberg identity in the form used. The period ΩE measures squared size.
In [16] the period is not constructed from a specific curve: the statement is that a curve and a
normalisation exist for which (6.3) holds. It is important here which period normalization is used.
Rescaling the differential changes the linear term. [16]
√
Remark. For the elliptic√ curve itself, with complex multiplication by Q( −d),√the same
identity holds with v = d/2, the imaginary part of √ the quadratic lattice parameter (1+i d)/2, and
with the period Ω of the curve: log Ω = γ + log(2π/ d) − cχd = γ + log π − log v − cχd . In this√form
the identity is Kronecker’s limit formula, and Ω is a geometric mean over the ideal classes of Q( −d).
For Q(i), with d replaced by 4, exponentiating the same expression gives √ Ω = 2π(ϖ/π)2 , with ϖ the
lemniscate constant. For d = 23 the formula gives log Ω = γ√+ log(2π/ 23) + 0.422120948038 . . ..
The two normalizations differ by log Ω − log ΩE = log(rdE / d), and the sum √ of the logarithm of
the period and log v is the same in both: log ΩE + log(rdE /2) = log Ω + log( d/2). Only this sum
enters (6.6). So (6.6) is the same in either normalization, and the numerical check in section 6.7
may use either. The two lines of (6.5), taken separately, hold only with ΩE and v = rdE /2.
The other source of linear terms is a count of forms. The mass of a definite genus is the sum of
reciprocal stabilizer orders over its lattice classes, as in section 6.1. The dimension of a representation
with strictly decreasing shifted integral weights z1 , . . . , zr is
Vr (z)
Dr (z) = Qr−1 , Vr (z) = (zi − zj ). (II.41)
Y
j=0 j! 1≤i 0 collects the extra local factors:
Mr f
= (r − 1)!(v/π)r (2v)−f /2 κr . (II.42)
Mr−1
Here v = rdE /2, as in (6.3), and Mr is unrelated to the subtracted main term Mu,t,j . For two ranks
r0 < r1 (for instance two consecutive ranks Nj , Nj+1 of the chain of section 6.2), let uρ , wρ be their
decreasing weight lists at each of the f places ρ. Multiplying the mass ratios and the f dimension
ratios gives precisely
Mr1 ρ Dr1 (wρ ) Y Vr1 (wρ ) r1
Q Pr1
f r
r=r0 +1 (2v)−(r1 −r0 )f /2
= (v/π) (6.4)
Y
κr .
Mr0 ρ Dr0 (uρ ) ρ Vr0 (uρ )
Q
r=r +1
0
Indeed the extra (r − 1)! in the mass cancels the extra (r − 1)! in each dimension denominator.
Taking logarithms is the count identity on ordered weights. For a numerical algebra check take one
place, r0 = 2, r1 = 3, and u = (3, 1), w = (5, 3, 0). The two dimensions are 2 and 15. The ratio
of dimensions is 15/2. The factorial from the mass is 2, leaving 15 = V3 (w)/V2 (u) = 30/2 before
the scale and local factors. Equation (6.4) follows directly from the displayed mass recurrence.
The automorphic mass formula supplying that recurrence is quoted. The passage from the count
identity (6.4) to the second line of (6.5) below (the definition of the level T , the factors κr , the
error Ecnt ) is also quoted.
46
6.7 Why the linear term has precisely the required sign
Write T for the positive auxiliary level attached to one choice of lattice data. It is different from
the set T of derivative labels. Let ℓg = log(g0 g1 ) for the two auxiliary level primes g0 , g1 . The size
and count bounds [7, 19] have the following form at a fixed translate, with ΩE and v = rdE /2 as
in (6.3). Here hB stands for the combination µCC + h − F0 of three quantities defined after the
display, and Esize , Ecnt are the respective error allowances:
m
at (log at − 1 − log ΩE ) − at log T + Esize ,
XX X
hB ≤
t∈T j=1 t∈T
m (6.5)
log T ≥ ⟨log wu,j ⟩ + log v − log π + mℓg − Ecnt .
X
j=1
The three quantities in hB = µCC + h − F0 are these. First, h is an upper bound for the height of
the pairing map, the average over the archimedean places of the logarithm of the squared operator
norm (section 6.5). Second, µCC is the mean slope of the common lattice in which all the pairing
maps take their values. Third, F0 is the part of the local divisibility that comes from the places
other than those of the top weights.
Neither offset survives. The term F0 also appears, with the opposite sign, in the slope of the source
lattice after gluing (section 6.8), and the two cancel. The term µCC is the slope of the target in the
inequality of section 6.10, where it appears on both sides. So one may read (6.5) as a bound for the
height h alone, as in Part I. An earlier formulation [18] writes the same pair of inequalities for a
single quantity.
The local divisibility comparison bounds the prime sum by hB + mℓg t at plus losses; the second
P
term is a level correction. Insert the lower bound for log T in the first line of (6.5). Since each at is
positive, this gives an upper bound. The terms mℓg t at cancel, and the coefficient of at at step j
P
is
log at − 1 − log ΩE − ⟨log wu,j ⟩ − log v + log π
(6.6)
= log at − 1 − γ + cχd − ⟨log wu,j ⟩.
The equality is (6.3). Thus the sum of the surviving terms is exactly t,j Mu,t,j . The two errors in
P
this substitution are Esize + Ecnt t at . Other local and lattice losses must still be added to them.
P
The main term M needs all three: Stirling’s formula (log a − 1), the period (6.3) (−γ + cχd ), and
the count (−⟨log w⟩).
In the finite table one can check the cancellation even though no level or lattice has been constructed
for that table. For a = 10000 take the single-step expression on the first line of (6.6), multiply
√ by a,
and use the normalization of the elliptic curve (the remark in section 6.6): put v = 23/2. The
period combination − log Ω−log v +log π, which is the same as − log ΩE −log(rdE /2)+log π, equals
−γ − 0.422120948038 . . .. The remaining two entries are log 10000 − 1 and −(log 10 + 7 log 0.05)/8.
This is the 95444.4636 computed independently by prime summation. An uncancelled constant per
derivative in (6.6) would survive multiplication by at ≃ bx; division by B would leave an error at
least of order x.
What is quoted in this comparison. The period formula (6.3), the mass formula, the construction
and count of the auxiliary levels, and the local level correction are the arithmetic inputs; they are
quoted. The analytic norm bound is proved in part in Part III, §III.4. The factorial cancellation
(6.4) and all signs in (6.6) are algebraic consequences proved above. The next comparison explains
why the divisibility can be applied to enough independent vectors at once.
47
6.8 From one determinant to lattices
A lattice L of rank r in a Euclidean space is the integer span of r independent vectors. Its covolume,
measured in its span, is the volume of the parallelepiped on a basis. Put µ(L) = − log(covol L)/r.
Let T be an injective linear map with T (L) ⊂ M , and let ∥T ∥ be its Euclidean operator norm. (This
section uses the plain normalisation, µ and log ∥T ∥. In (6.5)–(6.6) the height h is the logarithm of
the squared norm, and µCC is normalised to match, so it is twice a slope in the present sense, up to
an additive constant depending only on the field. The two normalisations must not be mixed.) The
volume of T (L) is at most ∥T ∥r times that of L, as one sees using an orthonormal basis of the span
and Hadamard’s inequality. Taking logarithms,
µ(L) ≤ µ(T (L)) + log ∥T ∥. (II.43)
Suppose first that M has the same rank r. Then T (L) has finite index in M and covol T (L) = [M :
T (L)] covol M , so
1
µ(L) + log[M : T (L)] ≤ µ(M ) + log ∥T ∥. (II.44)
r
This is the exact form of the principle that divisibility is bounded by size: the index is on the left
and the norm on the right. For T = diag(6, 10) and L = M = Z2 the index is 60, both slopes are
zero and ∥T ∥ = 10. The inequality reads 12 log 60 ≤ log 10, that is, 60 ≤ 100. The factors 2, 3, 5 of
the index and the two column lengths refer to the same map.
If L has smaller rank than M , the image T (L) is only a sublattice of smaller rank, and the first
inequality gives
µ(L) ≤ µmax (M ) + log ∥T ∥, (II.45)
where µmax is the supremum of slopes of nonzero sublattices. A single map of small rank therefore
controls only the maximal slope of its target. Sections 6.9 and 6.10 use many maps together in
order to reach the mean slope µ(M ).
A rational matrix presents a further problem: local divisibility may apply at different primes in
different bases. The source uses equality of a period pairing under changes of intermediate twists to
place the local lattices in a common rational vector space. One can then choose a favorable lattice
at each prime and put these choices together.
Over Q the last operation has a short proof. Let Lp be a full lattice over Zp inside Qrp , equal to Zrp
outside a finite set. Here Qp is the completion of Q for the p-adic absolute value, and Zp its ring of
integers. Choose a positive integer D with DZrp ⊆ Lp ⊆ D−1 Zrp at all primes. Let L be the set of
x in D−1 Zr whose image in D−1 Zrp /DZrp lies in Lp /DZrp for every p | D. The Chinese remainder
theorem applied after multiplying by D identifies the ambient finite quotient with (Z/D2 Z)r , the
product of its prime parts. The resulting lattice has the prescribed localization at every prime. For
example allow a denominator 2 in the first coordinate only at 2, and a denominator 3 in the second
only at 3. The result is 12 Z ⊕ 31 Z. Its covolume is 1/6. This proves the gluing step over Q. The
source uses the corresponding construction over a number field, with its locally chosen character
twists.
The denominators that remain. For a prime ℓ and positive twist index a, the relevant finite
group consists of first Galois cohomology classes over E with values in Qℓ /Zℓ (−a), unramified away
from ℓ and zero locally above ℓ. The twist means that Galois acts by the inverse ath power of its
action on roots of unity of ℓ-power order. A finite group of order ℓh costs h log ℓ in a logarithmic
index. Write hℓ,a for the length of this group as a Zℓ -module. On the selected window the supplied
bound, after passage to the working scale, is hℓ,a < ∞ for 1 ≤ a < ⌊x400 ⌋ (the number of twists
48
used is at most x400 ; see Part III, §III.9) and
√
(log ℓ) (II.46)
X X
hℓ,a ≤ b
ℓ∈S 1≤a<⌊x400 ⌋
for every finite set S of primes. This is the uniform assertion needed here. (In [8] the bound is
stated for a < x401 at a slightly different scale; the form above is what is used.) Even for ζ the
bound is needed over the field E, not over Q. The statement on these extension groups, the local
estimates relating their lengths to actual lattice indices, and the independence of the normalized
global period are quoted. [8, 19] The slope inequalities proved in this section are the elementary
case; the form with minimal slopes used in the chain is quoted (section 6.10).
49
6.9 Why many independent images improve the estimate
A large norm for one map could be concentrated on one direction. A second moment distinguishes
this possibility. For a positive Hermitian matrix Q of size D, Cauchy–Schwarz on its nonzero
eigenvalues gives
rank Q ≥ (tr Q)2 / tr(Q2 ) (Q ̸= 0). (II.47)
Two maps with images the coordinate axes in C2 give Q = I when one adds their orthogonal
projections. The bound gives rank two. Two maps onto the same axis give Q = diag(2, 0), with the
same trace 2 but second moment 4, and the bound gives rank one.
Here is the version needed when some source directions have already been discarded. Let Bi :
Vi → H be maps of finite-dimensional complex inner product spaces, with D = dim H > 0,
S = i dim Vi > 0 and positive weights ci . A superscript † denotes the adjoint. Put Ki = ci Bi† Bi
P
and Q = i ci Bi Bi† . Let Ui be subspaces of Vi with i (dim Vi − dim Ui ) ≤ αS, and suppose
P P
0 ≤ e ≤ 1/4, tr(Q2 ) ≤ (1 + η)(tr Q)2 /D. (II.48)
X
∥Ki − I∥2HS ≤ eS,
i
Here ∥A∥2HS = tr(A† A). Lemma. Under these hypotheses the codimension of
P
i Bi Ui in H satisfies
X √
codim Bi Ui ≤ 2η + 4α + 4 e D. (II.49)
i
(The proof below gives the sharper bound with η in place of 2η. The form with η is the one that
is used: in section 6.10 for the chain of the source, and in Part III, §III.8, Lemma T1(d).) The
display, with 2η, is the finite-dimensional assertion FrameLA stated in [20] (a statement in the
archive, not used by the final theorem; the chain uses a variant, FrameLA1, stated for subspaces
and with η in place of 2η; see Part III, §III.8, §2).
Proof. We may take 0 ≤ α ≤ 1, since losing the entire space makes a larger bound automatic. Let
P project onto the missing target space, of dimension zD. Then P Bi vanishes on Ui . Hence tr(P Q)
is at most the sum of the traces of Ki over the omitted source directions, since ci Bi† P Bi ≤ Ki and
it vanishes on Ui . Write Ki = I + (Ki − I) there and apply Cauchy–Schwarz. Also apply it to the
total trace: √ √
tr(P Q) ≤ (α + αe)S, tr Q ≥ (1 − e)S ≥ S/2. (II.50)
Writing q̄ = tr Q/D, the second-moment hypothesis gives ∥Q − q̄I∥2HS ≤ η(tr Q)2 /D. Pairing with
P yields √
√
z ≤ tr(P Q)/ tr Q + ηz ≤ 2α + 2 e + (η + z)/2. (II.51)
√
This proves even z ≤ η + 4α + 4 e. The specialization to the actual maps requires a lower first
moment and an upper second moment of their period pairings; the latter comes from a relative
trace formula. A trace formula equates a sum over a basis of functions with a sum of geometric
integrals. Here it controls correlations between the images of different maps. These automorphic
estimates are quoted. [7, 19, 20] In the source the index i is a word: a string of k − 2 binary letters
which chooses, at every weight place, one of the weights that interlace a fixed weight of the group
in k − 1 variables; each word also has its own level prime. The space Vi is attached to the word,
and H is common to all words. The character twists of section 6.8 are a different index (one twist
for each place) and do not enter this family. There is one such family for each translate u of section
4; the weights wu,j depend on the translate and not on the word. See Part III, §III.8.
50
6.10 From the rank bound to (M)
It remains to see why a count of directions can control a mean slope. In this section H is a bound
for heights and u is a spread of slopes; they are not the target space H of section 6.9 and the
translate u of section 4. Let d1 , . . . , dD be real slopes, a a reference value, and assume dj ≥ a − H
with H ≥ 0. Put N (t) = #{j : dj < a − t}. Integrating the indicator of each deficit gives
1 X 1 X N (t)
Z H
dj ≥ a − (a − dj )+ = a − dt. (6.7)
D j D j 0 D
For example, slopes 2, 3, 5 and a = 4 give N (t) = 2 on 0 < t < 1, 1 on 1 < t < 2, and 0 thereafter.
The mean deficit is (2 + 1)/3 = 1. The lower bound 3 is below the true mean 10/3 because the
largest slope exceeded 4. This is the layer-cake formula; the form used by the final theorem is in
[19].
Here is how an integrable count is obtained in [20]. If R slopes yj have mean at least s and all are
at most s + u, u > 0, the fraction below s − t is at most u/(u + t). Indeed, with fraction z below,
their mean is at most z(s − t) + (1 − z)(s + u). Apply the rank bound of section 6.9 after discarding
those directions; so α = u/(u + t) in the lemma of section 6.9, which gives the term 4u/(u + t)
below.
(Passing from the bound on the missing directions to a count of the slopes of the target below a
given level uses the inequality for the minimal slope of the image of a map, and properties of the
filtration of a lattice by slopes. These are in [25]; they are quoted here, and are not the inequality
proved in section 6.8.)
In the notation of [20], for numbers s, h, L, u, H, η, e with u > 0 and L, H, η, e ≥ 0, the hypotheses
take the form
#{j : dj < s − h − L − t} 4u √
dj ≥ s − h − H, ≤ 2η + +4 e (0 ≤ t ≤ H). (II.52)
D u+t
√
Putting a = s − h − L in (6.7) and integrating gives an error at most (2η + 4 e)H + 4u log(1 + H/u).
In particular it proves the looser bound stated in [20] (the chain of the source uses a variant of it,
which is a hypothesis in [7] and is applied for each translate in [28]; see below)
1 X √
dj ≥ s − h − L − u 1 + 8 log(1 + H/u) − (4η + 8 e)H. (II.53)
D j
√ √
The trace estimates make e exponentially small on the chosen intervals: e = 2e0 with e0 ≤
exp(−xε/8 ), where e0 is the constant in the bound for the diagonal Gram blocks (Part III, §III.8). For
√
η the source has η = 5x−CH + 3 αram . Here αram is a bound for the proportion of source directions
√
lost at the level primes (it is not the α = u/(u+t) above), the source takes α√ ram = exp(−x
ε/8 ), and
CH is the exponent in the bound H ≤ x b. For all sufficiently large x, (3+4 2) exp(−x ) xCH ≤
CH ε/8
√ √ √
1. Since η = 5x−CH + 3 αram and e = 2e0 , this gives (η + 4 e)xCH ≤ 6, and so (η + 4 e)H ≤ 6b.
So η is bounded only by a negative power of x, chosen to match H. The power of x comes from the
number of maps, ⌈3G2 xCH ⌉, where G = exp(CG m5 f xε ) is the constant in the quoted bounds for
the Gram blocks, with a constant CG > 0 (Part III, §III.8). In the variant the chain uses, the error
√
bracket is 4u log(1 + H/u) + (η + 4 e)H, with η and not 2η. It is multiplied by 1/(1 − δp ), where
δp is the proportion of the target not spanned by its lattice (exponentially small, quoted). The
lower bound also has − log |W| as a separate term, which this factor does not multiply. Here W is
the set of words and log |W| ≤ (k − 3) log 2, which is negligible; the sizes of H and of Ecnt are part
51
of the quoted statements and are not derived here. The bound for the spread u and the arithmetic
control of L are also needed. This argument uses a mean slope of the target, where the elementary
inequality of section 6.8 by itself would only control a maximal slope.
Assembly. The local derivative bounds place the prime sum in the integral lattices. The period
identity makes it possible to compare the different local choices in one space. The height and
counting comparisons (6.5)–(6.6) give the prime sum’s main term. Finally the mean-slope bound
accounts for losses from indices and directions which must be discarded. The exact output of the
source’s comparison is this: for each translate, the required divisibility sum is at most its main term
plus an aggregated error [28]. The hypotheses of that statement are defined in [7]. The archive
also has the form averaged over the translates, in [7]; the final theorem does not use it. The size
estimates bound the error by the unnormalized expression of section 4.3. These are the inputs [5, 7,
19, 28] used at (M).
Once the quoted estimates are granted, sections 4 and 5 give (A), and section 3 concludes by
balancing x/m and m. A cost of m2 would instead give m = x1/3 and exponent 2/3, if the
corresponding inequality were available with the other errors under control.
Proved in Part II, section 6. Finite-chain integrality, the minor congruence, the Hasse product
rule, the reduction (6.1), the factorial count (6.1a), the Gaussian norm and the Cauchy–Schwarz
bound of section 6.5, the count identity (6.4) and the cancellation (6.6), the elementary slope
inequality, rational lattice gluing, the rank bound and its integration. (Sections 1–5 are proved in
full, given (M) and Estimates B, C.)
Quoted from the Lean source. The normalized theta maps and period identity, their
simultaneous local divisibilities and losses, the period and mass formulas, the level bounds, and
the norm bounds (apart from the analytic estimates of Part III, §III.4); the denominator estimates;
and the two moments for this family. Part III, §III.6–III.9 deduce parts of these from named
statements of the source, which remain quoted; see the table at the start of Part III.
52
Sources
Alpöge published the Lean source as one archive, lean.tar.gz, split into fourteen download parts at alpo.ge
(listed in footnote 1 of his announcement). Unpacked, it contains 175,187 Lean files. The SHA-256 digest
of the joined archive is 9631b4f020199e580ba86178cb56c6e9b09f690c2369a80dc4334aa93f49c7fc. The
numbered references below are files of that archive; all except [4] are cited in this paper. Each is given by its
path from the root of the archive.
[1] lean/src/Development/AutomorphicArithmetic/DedekindLFunctions/LandauTheorems/
landauWindow.lean
[2] lean/src/Formalism/ProductFormula/StatementOCCM.lean
[3] lean/src/Formalism/SizeWindowBrackets/FirstPrimeNeSecondPrime.lean
[4] lean/src/Formalism/StatementCM.lean (definitions for the earlier form of the argument; see the
remark on [2] in section 2)
[5] lean/src/Development/AutomorphicArithmetic/DedekindLFunctions/LandauTheorems/
ChTop.lean
[6] lean/src/Formalism/SizeWindowBrackets/GrowthBoundGivesRiemannHypothesis.lean
[7] lean/src/Formalism/RankGapBounds/MeanBoundSecondForm.lean
[8] lean/src/Development/AutomorphicArithmetic/CharacterBoxes/ExistsPageN.lean
[9] lean/src/Development/AutomorphicArithmetic/MeanSquareBounds/pairEx.lean
[10] lean/src/Development/AutomorphicArithmetic/CharacterBoxes/WinPairs.lean
[11] lean/src/Development/AutomorphicArithmetic/PrimitiveCosets/CoreGm.lean
[12] lean/src/Supplements/Kernels/MainTheorems.lean
[13] lean/src/Development/AutomorphicArithmetic/AdelicQuotients/ThetaLiftLinear.lean
[14] lean/src/Formalism/ThetaLiftIndependence/Statement.lean
[15] lean/src/Formalism/SharpRestriction/GrowthBounds/CoarseMultiplicity/KeyMinorDvd.lean
[16] lean/src/Development/AutomorphicSpectrum/Proofs/GeometryProofs/
ChowlaSelbergPeriodStatement.lean
[17] lean/src/Formalism/BehrendSets/StatementE.lean
[18] lean/src/Formalism/ThickAverageLoss/PartnersToFactorsAt.lean
[19] lean/src/Formalism/SharpRestriction/GlobalTheoremInputs.lean
[20] lean/src/Formalism/CorankShare/StatementFR.lean
[21] lean/src/Formalism/ThetaLogCoordinates/UniformThetaSubstitution/StatementAB.lean
[22] lean/src/Formalism/LocalRingCardinality/DataParams2 .lean (the file name contains the
subscript character U+2082)
[23] lean/src/Formalism/SharpRestriction/GlobalSchurBounds/FlipTwoThree.lean
[24] lean/src/Formalism/SharpRestriction/GlobalSchurBounds/
IntegrableOfContinuousCircleMeasure.lean
[25] lean/src/Formalism/HarderNarasimhan/StatementPl.lean
[26] lean/src/Formalism/LocalRingCardinality/Params.lean
[27] lean/src/Formalism/FrobeniusOrbits/FrobeniusOrbitsProofs/ThickExists.lean
[28] lean/src/Formalism/SizeWindowBrackets/ChainSlopeWindows/LevelOfCongruence.lean
[29] lean/src/Development/AutomorphicArithmetic/CharacterBoxes/PairGood.lean
53
Part III. Proofs of the quoted statements
Parts I and II deduce the Riemann hypothesis from statements marked [quoted]. Part III is meant
to prove those statements, in the same classical language, so that the three parts together form a
complete proof. It is being written in stages, and the table below records what has been done.
Statement of Parts I and II Status in Part III
Estimate B: the auxiliary primes q, d proved, from four cited theorems (Q1)–(Q4)
(§III.1)
Estimate C: the characters of E at the proved, from three cited theorems (Z1)–(Z3) and Theo-
large scale (§III.2) rem III.1.1(e)
The sets T2 and T of columns (§III.3) constructed, with proof (elementary); their use in (M) is not
yet written
The inequality (M), size of the theta the analytic estimates for one block, one map and a chain
kernels (§I.7, item 2; §III.4) are proved in an abstract setting; that the maps of the con-
struction are instances of it is quoted (item (vi) of the list
in §III.4); four passages are conditional (the reference mean
square, the eigenvalues of a lift, the row-by-row comparison,
the pairing at the end of the chain); the arithmetic inputs
listed at the end of the section are not proved
The inequality (M), divisibility at inert written up in §III.5, conditionally on the integrality of the
primes (§I.7, item 1; §III.5) theta kernels in the coordinates of the height-two logarithm
(hypothesis (B)), which is taken from the Lean development.
Under (B) the local estimate for power series is proved for
orbits of length two; for longer orbits it also uses a duality
statement cited from the Lean. That the theta kernels satisfy
(B), the frames and labels, and the remaining primes are not
proved
The inequality (M), from derivatives to proved: the lemmas on slopes of lattices, the form of the
lattices (§III.6) prime sum and its splitting by classes of primes, the digit
losses, the inequality at one class of primes, and bounds for
several terms of the sum over primes. Quoted: the bounds
for the eleven loss sums and for their total. Conditional on
two statements of the source (on valuations and on lattices):
the theorem on the glued lattice. A model case only: the
comparison of the two sides. Quoted: the statement on
valuations itself
The inequality (M), arithmetic inputs proved: the size inequality (III.75) follows from eleven named
of the size bound (§III.7) inputs; an index-corrected mean of periods equals log ΩE ,
under the hypotheses on the periods and the field stated
there (its identification with the periods of the input (S4)
rests on two further statements of the source and one that is
proved there). The eleven inputs are quoted, among them
the inner-product formula, the covolumes and the local index
bounds
The inequality (M), the trace-formula proved: the dimension bound of Part II, §6.9 follows from
estimates (§III.8) three statements of the source (on traces, ranks and pairings)
and the size and subspace conditions of Corollary T2; a finite
model of the trace estimates. The three statements, and the
theorems behind them, are quoted
54
Statement of Parts I and II Status in Part III
The inequality (M), the bound for Ga- proved: the step from the bound to the bound for denomina-
lois cohomology (§III.9) tors, and vanishing in some cases over Q from cited theorems.
The bound itself is quoted, with an account of its proof in
the source
The inequality (M), the rest: the con- not yet begun; quoted
struction of the maps and weights; the
identification of the abstract objects of
§III.4–III.9 with those of the construc-
tion; the assembly of (M)
So at present Parts I–III prove the Riemann hypothesis from the inequality (M) and from the cited
theorems, and (M) is quoted. Of what (M) needs, §III.4–III.9 prove the parts stated in the table,
conditionally as stated there.
The statements below are translated from the Lean development of the announcement. In Part III
“the source” always means this Lean development. For §III.1–III.3 each is a statement that the
final theorem actually depends on (for §III.4–III.9 see the remarks in those sections): we computed
the set of all declarations on which the Lean theorem riemannHypothesis depends, and checked
that these belong to it.
For §III.1–III.3 the check was also made for the lemmas of the Lean proofs that the sections
follow (for instance QH.thinH, twistK, meanSqK, bad_share, tailK, count_bad, hostThickExists,
topsChoiceH): all are among the declarations the final theorem depends on.
A correction to the source references of Part II. The same computation shows that several
statements cited in Part II are present in the archive but are not among the 1,478,198 declarations
on which riemannHypothesis depends: the Fock norm and jet definitions of [23]; the statements
ChainInt of [14], FrameLA and FrameLast of [20], and CountIdentity of [17] (a statement of the
same name elsewhere is used); the definitions concerning Behrend sets in [17]; and the inequality of
[6] in the form quoted. These are parallel or earlier formulations. The same holds for the linearity
statement of [13]. The mathematics of Part II does not depend on them, since what Part II proves
it proves directly; but for these items the phrase “the source’s statement” should be read as “a
statement in the archive”. The statements that are used include: the Landau lemma on intervals
[1], the auxiliary primes [9, 29], the weighted sums ASK and MAIN [3], the height of a map [25],
the Chowla–Selberg period [16], the minors congruence [15], the digit inequality StarArith [21],
the layer-cake formula, the Bessel identity and the local comparisons of [19], the bound for each
translate each_unit_bound_statement of [28] with the hypotheses defined in [7] (the averaged
statement MeanBoundSecondForm of [7] is not used), and the bound for Galois cohomology [8].
III.1 The auxiliary primes: proof of Estimate B
For a primitive Dirichlet character χ ̸= 1 put, as in Part I,
y L′
Gχ (y) = Λ(n)χ(n) −1 +y (1, χ). (III.1)
X
n≤y
n L
For an odd prime p write χp for the Legendre symbol modulo p.
55
Theorem III.1.1 (auxiliary primes; QH.pairEx in [9], statement in [29]). Let M ≥ 1 be an integer,
0 < ε < 52
1
and A ≥ 1. There is Y1 such that for every Y ≥ Y1 there are primes q and d with the
following properties.
(a) q ≡ 3 mod 8 and 21 Y ε/5 < log q ≤ Y ε/5 .
(b) d ≡ 7 mod 16 and 12 Y ε/5 (log Y )2 < log d ≤ Y ε/5 (log Y )2 .
(c) q divides d + 1, and so does every odd prime p ≤ Y ε/5 . Neither q nor d divides M .
(d) Let ξ be a Dirichlet character modulo M and let (i, j) ∈ {0, 1}2 , (i, j) ̸= (0, 0). Let χ be the
primitive character that induces ξχiq χjd . Then for all y with Y ≤ y ≤ Y A
L′
|Gχ (y)| ≤ y 1/2+ε/2 and (1, χ) ≤ 4 log y. (III.2)
L
(e) Let ρ be a real primitive character whose conductor divides 8M , let (i, j) ̸= (0, 0), and let χ be
the primitive character that induces ρχiq χjd . Then L(s, χ) ̸= 0 in the region
Re s > 1 − 400
1
|Im s| ≤ 8M exp Y ε/5 exp j Y ε/5 (log Y )2 . (III.3)
,
Remarks. (1) For ζ one takes M = 1 and A ≥ 32 . Then (d) gives Estimate B of Part I, with x = y,
the interval [Y, Y 3/2 ] contained in [Y, Y A ], and the exponent 12 + 2ε in place of 12 + ε.
(2) The conductors are far larger than y: log q is a power of Y . So (d) cannot be a statement
about a fixed character as y → ∞. It is a statement about well-chosen q and d.
(3) Condition (c) has a consequence that is not visible in Part I. If p is an odd prime with p ≤ Y ε/5 ,
then d ≡ −1 mod p, and since d ≡ 3 mod 4 quadratic reciprocity gives χd (p) = −1 p (−1)
(p−1)/2 = 1.
√
Also χd (2) = 1. So every prime up to Y ε/5 splits in Q( −d); the inert primes, which carry the
argument of Part I, all exceed Y ε/5 .
(4) Part (e) is used for Estimate C: it excludes exceptional real zeros for the real characters
whose conductor is divisible by q or d.
Proof of Theorem III.1.1
The proof chooses q at random from a set of primes, then d at random from a second set, and
shows that most choices work. Two things must hold for the chosen primes: a sum over n ≤ Z
must be small (this comes from a mean square over the primes), and a tail n > Z must be small
(this comes from a zero-free box, which holds for all but few primes by a zero-density theorem).
Throughout, τ = ε/2,
Z = Y 400A , z = Y ε/5 , z ′ = Y ε/5 (log Y )2 , (III.4)
and Y is large in terms of M , ε, A. Constants C, c are absolute, or depend only on A, unless
indicated.
Cited theorems.
(Q1) Zero-free region, with uniqueness (Landau, Page; [Dav00, Ch. 14], [Pag35]; in this explicit
form [TZ21, §1]). Let c1 = 131
. For every modulus m ≥ 2, the functions L(s, χ), χ modulo
m, have no zero with β ≥ 1 − c1 / log(m(|γ| + 3)), except possibly one simple real zero β1
of one real nonprincipal character χ1 . Moreover β1 ≤ 1 − c/m; this follows from the bound
β1 ≤ 1 − c′ m−1/2 recorded in [TZ21, §1]. This is the exceptional zero meant in (Q2).
56
(Q2) Primes in progressions to large moduli (Thorner and Zaman [TZ21, Corollary 1.4]; results of
this type go back to Linnik and to Gallagher [Gal70], see also [Bom87]). There is an absolute
c3 > 0 such that for coprime m ≥ 2 and b, and t ≥ m12 ,
1 tβ1
log p = Λb (t) 1 + O(κ(t)) , Λb (t) = t − χ1 (b) (III.5)
X
,
p≤t, p≡b (m)
φ(m) β1
with κ(t) = exp(−c3 log t/ log m) + exp(−c3 (log t)3/5 (log log t)−1/5 ). The term with χ1 is
present only if the exceptional zero of (Q1) exists for the modulus m.
(Q3) Zero density (Montgomery [Mon71, Theorem 12.2]). The number of primitive characters χ
of conductor at most Q for which L(s, χ) has a zero with β > 1 − 400
1
, |γ| ≤ Q, is at most
CQ .1/4
(Q4) The logarithmic derivative in a zero-free box (a standard consequence of Landau’s lemma
[Tit86, §3.9], [MV07, Lemma 6.3], and the convexity bound for L(s, χ)). Let χ ̸= 1 be
primitive of conductor at most Q, and suppose L(s, χ) ̸= 0 for Re s > 1 − 400 1
, |Im s| ≤ Q.
′
Then for each δ > 0 there is Kδ with | L (s, χ)| ≤ Kδ log Q for Re s ≥ 1− 400 +δ, |Im s| ≤ Q−1.
L 1
The form of (Q2) that is used. Let U ≥ m24 , and let π ∗ (U ; m, b) be the number of primes
p ≡ b mod m with U 1/2 < p ≤ U . Then
1
Z U
∗ dt
π (U ; m, b) − Mb ≤ k Mb , Mb = 1 − χ1 (b)tβ1 −1 , (III.6)
φ(m) U 1/2 log t
√
with k = A0 e−c4 log U/ log m + e− c4 log U . Proof. On [U 1/2 , U ] we have t ≥ m12 , and
√
(log t)3/5 (log log t)−1/5 ≥ 12 log U , so that the relative error in (Q2), with its implied constant,
is at most k/20 for suitable A0 , c4 . RWrite θ(t) for the sum in (Q2). By partial summation
π ∗ = θ(U )/ log U − θ(U 1/2 )/ log U 1/2 + UU1/2 θ(t) dt/(t log2 t). Replacing θ by Λb gives exactly Mb ,
because Λ′b (t) = (1 − χ1 (b)tβ1 −1 )/φ(m). The function Λb is positive and increasing, so the error is
RU
at most 20 k
Λb (U ) log1 U + log2 U + log2 U . Finally Λb (U ) ≤ 4 U/2 Λ′b ≤ 4 log U · Mb : if χ1 (b) = 1 the
derivative Λ′b is increasing, and otherwise it lies between 1/φ(m) and 2/φ(m).
The Lean development proves all four. For (Q2) and (Q3) it has its own proofs: (Q2) from
a zero-density estimate and the Deuring–Heilbronn phenomenon, (Q3) by a mollifier and the
Halász–Montgomery inequality. In (Q3) the published theorem is much stronger than what is used.
The sets of primes. Let m0 be a positive integer divisible by 8, a0 an integer prime to m0 ,
and U > 1. Put
P = P (m0 , a0 , U ) = {q prime : U 1/2 < q ≤ U, q ≡ a0 mod m0 }. (III.7)
Call (m0 , a0 , U ) admissible if
4 (log Y ) log m0 ≤ log U, Y ε/6 ≤ log U ≤ Y ε/4 , (III.8)
1 2
and every prime factor of m0 is either at most Y 1/4 or greater than Z 2 .
Step 1: characters over the primes of P . Let (m0 , a0 , U ) be admissible, let B > 0, and let
Y be large in terms of B as well. Then |P | ≥ U 9/10 . Moreover, among the odd squarefree s with
1 < s ≤ Z 2 and (s, m0 ) = 1, at most one fails to satisfy the following bound; if there is one, call it
n∗ : X q
≤ Y −B |P |. (III.9)
q∈P
s
57
Proof. Apply (III.6) with modulus m = m0 s. By (III.8), log m ≤ log m0 + 800A log Y and
log U/ log m ≥ c(log Y )2 (if log m0 ≤ log Y use log U ≥ Y ε/6 ). So U ≥ m24 and k ≤ Y −B−1 . Every
q ∈ P exceeds Z 2 ≥ s, since log q > 21 Y ε/6 , so q ∤ s. Split P into residue classes b modulo m0 s with
b ≡ a0 mod m0 , (b, s) = 1; the symbol ( sb ) is constant on each, and the class contains π ∗ (U ; m, b)
elements of P . So
X q X b X
= Mb + O k Mb . (III.10)
q∈P
s b
s b
The error. Since k ≤ 12 , (III.6) gives Mb ≤ 2π ∗ (U ; m, b), so b Mb ≤ 2|P |. (This is why a relative
P
error is needed: P may be small if the modulus m0 itself has an exceptional zero.)
The main terms. The part of Mb without χ1 is the same for all b, and b 7→ ( sb ) is a nontrivial
character on the classes b ≡ a0 mod m0 (s is squarefree, odd, prime to m0 and greater than 1). So
these parts cancel. The parts with χ1 cancel in the same way unless χ1 (b)( sb ) is constant on those
classes. In that case call s bad if in addition β1 > 1 − 3 log(m
c1
2 .
0Z )
At most one s is bad. If s1 , s2 are both bad, apply (Q1) to the modulus m0 s1 s2 ; since
log(3m0 s1 s2 ) ≤ 3 log(m0 Z 2 ), both exceptional zeros lie in its region, so the two exceptional
characters are induced by the same character. Hence ( su1 ) = ( su2 ) for u ≡ 1 mod m0 prime to s1 s2 ,
and s1 = s2 . So at most one s = n∗ is bad.
The other s. If s is not bad and the parts with χ1 do not cancel, then tβ1 −1 ≤ U −(1−β1 )/2 ≤
2
U 1 /(6 log(m0 Z )) ≤ Y −B−1 on the range of integration, so the part with χ1 is at most Y −B−1 times
−c
the part without it, hence at most 2Y −B−1 Mb . Altogether | q∈P ( qs )| ≤ (2k + 4Y −B−1 )|P | ≤
P
Y −B |P |.
The size of P . Apply (III.6) with m = m0 , b = a0 : |P | ≥ 12 Ma0 . By (Q1), 1 − β1 ≥ c/m0 , so
1 − tβ1 −1 ≥ c′ /m0 for t ≥ U/2, and Ma0 ≥ c′ U/(2m20 log U ). Since log m0 ≤ 4 log U/(log Y )2 , this
exceeds 2U 9/10 .
Step 2: the mean square. For y ≥ 1, an odd prime q and a function κ with |κ(n)| ≤ 1 put
y n
S(y, q, κ) = Λ(n) min 1, (III.11)
X
κ(n) .
n≤Z
n q
Let (m0 , a0 , U ) be admissible. Then for Y ≤ y ≤ Y A and every such κ,
|S(y, q, κ)|2 ≤ C y log y |P |. (III.12)
X
q∈P
Proof. Write cn = Λ(n) min(1, y/n). Expanding the square,
X nn′
(III.13)
X X
|S|2 ≤ cn cn′ .
q∈P n,n′ ≤Z q∈P
q
Every q ∈ P exceeds Z 2 ≥ nn′ , since log q > 12 Y ε/6 . Let s be the product of the primes that do
not divide m0 and occur in nn′ to an odd power. All q ∈ P lie in one residue class modulo m0 ,
′ q
and 8 | m0 ; so by quadratic reciprocity and its supplements ( nn q ) = ±( s ) with a sign that does
not depend on q. By Step 1 the inner sum is at most Y −A−1 |P | unless s = 1 or s = n∗ , when we
bound it by |P |.
The terms n, n′ are prime powers. Put σ(p) = pj ≤Z cpj . One checks σ(p) ≤ 3 log y, and
P
σ(p) ≤ 2y log p/p for p > y, whence p σ(p)2 ≤ Cy log y by Chebyshev’s bound. Call n neutral if
P
it is an even power of a prime or a power of a prime dividing m0 ; let R be the sum of cn over such
58
√
n ≤ Z. Even powers contribute at most C y. The primes dividing m0 that are at most Z are at
√
most Y 1/4 , so there are at most Y 1/4 of them, each contributing at most 3 log y. So R ≤ C y.
Pairs with s = 1. Either n, n′ are powers of the same prime, or both are neutral. Their
contribution is at most p σ(p) + R |P | ≤ Cy log y |P |.
2 2
P
Pairs with s = n∗ . Since n, n′ are prime powers, such pairs exist only if n∗ is a prime or a
product of two primes; then one of n, n′ is a power of a fixed prime while the other is a power of a
second fixed prime or is neutral. The contribution is at most (2 · 9 log2 y + 6 log y · R)|P | ≤ Cy|P |.
All other pairs. Since n≤Z cn ≤ y n≤Z Λ(n)/n ≤ Cy log Z, they contribute at most
P P
Y −A−1 (Cy · 400A log Y )2 |P | ≤ Cy|P |, because y ≤ Y A .
Step 3: all y at once. Let (m0 , a0 , U ) be admissible and κ as before. The number of q ∈ P
for which |S(y, q, κ)| > 21 y 1/2+τ for some y ∈ [Y, Y A ] is at most Cτ Y −τ |P |.
Proof. As a function of y, S is continuous and piecewise linear, with derivative D(y) =
y y. So (III.12), which is uniform in κ, gives also q |D(y)| ≤ C log y |P |/y. On
2
P
a block I = [W, min(2W, Y )] with Y ≤ W , by the Cauchy–Schwarz inequality, |S(y)|2 ≤
A
2|S(W )|2 + 2W I |D|2 , so q maxy∈I |S(y)|2 ≤ CW log W |P |. By Chebyshev’s inequality the
R P
number of q with max |S| > 12 W 1/2+τ on the block is at most CW −2τ log W |P | ≤ (C/τ )W −τ |P |.
Sum over the blocks W = 2i Y .
Step 4: the tail. For a primitive χ ̸= 1 and w ≥ 1 put
L′ X Λ(n)χ(n)
tχ (w) = − (1, χ) − , (III.14)
L n≤w
n
′
the tail of the series for − LL (1, χ). Let χ have conductor at most Q and suppose L(s, χ) ̸= 0 for
Re s > 1 − 400
1
, |Im s| ≤ Q. Then for Q ≥ 2, w ≥ 2, 0 < η ≤ 1 and 0 < δ ≤ 800 1
,
w log Q 2(wη + 1) log(2w)
|tχ (w)| ≤ Kδ′ (log Q)2 w−1/400+δ + + . (III.15)
ηQ w
Proof. Let Gχ be as above. For h = wη one has directly from the definitions
w + h
Gχ (w + h) − Gχ (w) = −h tχ (w) + Λ(n)χ(n) −1 , (III.16)
X
w 1 − 400 1
, |Im s| ≤ F U ; and |Gχ (y)| ≤ y 1/2+τ for
Y ≤ y ≤ Y A.
Proof. Since q > U 1/2 ≥ F , the prime q does not divide the conductor of ξ, so ξχq is primitive,
and q is determined by it. Hence by (Q3) the number of q ∈ P for which some primitive
character of the form ξ ′ χq , with ξ ′ of conductor at most F , has a zero in the box of height F U
is at most C(F U )1/4 ≤ CU 3/8 , which is at most 14 U 9/10 ≤ 41 |P |. By Step 3 with κ = ξ, used
for each ξ ∈ Ξ, the number of q with |S(y, q, ξ)| > 12 y 1/2+τ for some y and some ξ is at most
RCτ Y −τ |P | ≤ 14 |P |. For the remaining q, (III.18) and consequence (a) of Step 4 with Q = F U give
|Gχ (y)| ≤ 21 y 1/2+τ + Y ε ≤ y 1/2+τ .
Step 7: the choice of q and d.
The prime q. Take m0 = 8, a0 = 3, U = ez . This is admissible. Take F = 8M and for Ξ
all primitive characters of conductor dividing 8M , the trivial one included. Step 6 gives a prime
q ≡ 3 mod 8 with z2 < log q ≤ z such that every ξχq has the three properties, with box height
8M ez .
The prime d. Now take m0 = 16 q 3≤p≤z p and let a0 be the class with a0 ≡ 7 mod 16 and a0 ≡
Q
′
−1 modulo q and modulo every odd prime p ≤ z. Take U = ez . Then log m0 ≤ 4z = 4z ′ /(log Y )2 ,
and the prime factors of m0 are at most z ≤ Y 1/4 , or equal to q, which exceeds Z 2 . So (m0 , a0 , U )
is admissible. Take F = 8M ez ≤ U 1/2 and Ξ = {ξχaq } with ξ as before and a ∈ {0, 1}. Step 6 gives
′
a prime d ≡ 7 mod 16 with z2 < log d ≤ z ′ , with d ≡ −1 modulo q and modulo every odd prime up
′
to z, such that every ξχaq χd has the three properties, with box height 8M ez+z .
Verification. Parts (a)–(c) hold by construction; q, d > 8M . For (d): a character ξ modulo M is
induced by a primitive character of conductor dividing M , so the primitive character inducing ξχiq χjd
′
is one of those treated, and |Gχ (y)| ≤ y 1/2+ε/2 on [Y, Y A ]. The bound | LL (1, χ)| ≤ 4 log Y ≤ 4 log y
is consequence (b) of Step 4, with log Q ≤ log(8M ) + z + z ′ ≤ 3z ′ . Part (e) is the zero-free box.
60
Remarks. (1) There is no circularity. Nothing of the strength of the Riemann hypothesis is
assumed for the characters χ. The bound y 1/2+τ is the mean-square size of a sum of fixed length
Z = Y 400A over a family of primes q with log q ≥ 12 Y ε/6 , far beyond that length. As q varies, the
symbols ( nq ), n ≤ Z, behave like independent signs; reciprocity turns the statement into one about
primes q in progressions to moduli m0 s much smaller than q, where (Q2) applies. The zero-free
box is a weak property of the same kind, and it is proved for all but few q, by (Q3).
(2) Part (e) is proved for every primitive ρ of conductor dividing 8M , real or not.
(3) The mean value (III.12) is obtained without a large sieve inequality. A reader who knows
Heath-Brown’s large sieve for real characters [HB95] will recognise the statement; the route above
uses primes in progressions.
III.2 The large scale: proof of Estimate C
This section proves Estimate C from three standard theorems on Dirichlet L-functions, which
we cite, and from part (e) of Theorem III.1.1. No zero-density estimate and no mean value over
characters is used here. In the Lean development this is the theorem QH.thinH with the lemmas it
calls.
Notation. For a primitive character χ modulo N (the case N = 1, χ = 1 included) put
u χ=1
(
ψ0 (u, χ) = Λ(n)χ(n) − (III.20)
X
n≤u 0 χ ̸= 1.
Let Gχ be as in §III.1 for χ ̸= 1, and G1 = G. Zeros ρ = β + iγ are counted with multiplicity.
Theorem III.2.1 (Estimate C; QH.thinH). Let B ≥ 1, 0 < ε < 52 1
and c′ > 0. There is
x0 = x0 (B, ε, c ) such that the following holds for x ≥ x0 . Let q =
′ ̸ d be odd primes and Q a positive
integer prime to qd, divisible by 8, with odd part at most B, and with log(Qqd) ≤ 2xε/5 (log x)2 .
Suppose that for every real primitive character χ = ̸ 1 whose conductor divides Qqd and is divisible
by q or by d, L(β, χ) ̸= 0 for real β > 1 − 400
1
. Then for every primitive character χ whose conductor
divides Qqd,
|Gχ (y)| ≤ y exp − c′ xε/8 for all y ≥ 31 exp xε/3 . (III.21)
In the application Q is the power of 2 equal to 4n (for ζ), the characters are those of the field E,
and q, d are the primes of Theorem III.1.1 with Y ≤ x ≤ Y 3/2 ; then log(Qqd) ≤ 2xε/5 (log x)2 by
parts (a), (b) of that theorem, and the hypothesis on real zeros is its part (e) with M = 1 (see
Step 6; for a general Q one takes M divisible by the odd part of Q). In Part I and in formula (5.6)
of Part II the constant c′ is called κ.
The three theorems cited. Proofs are in Davenport [Dav00, Ch. 14, 16, 19] and Montgomery
and Vaughan [MV07]: Theorem 11.3 for (Z1), Theorem 10.17 for (Z2), and Theorems 12.5 and
12.10 with the discussion after Corollary 12.11 for (Z3), which is stated there in a slightly different
form. The Lean development proves all three.
(Z1) Zero-free region (de la Vallée Poussin, Gronwall, Landau, Titchmarsh). There is an absolute
constant c1 > 0 such that, for every N ≥ 1 and every character χ modulo N , if L(ρ, χ) = 0
and β ≥ 1 − c1 / log(N (|γ| + 2)), then χ is real and nonprincipal and ρ is real. (For N = 1
this says that ζ has no zero in the region.)
(Z2) Zeros in a window. There is an absolute C1 such that for every primitive χ modulo N and
every real t, the number of zeros with 0 < β < 1 and |γ − t| ≤ 1 is at most C1 log(N (|t| + 2)).
61
(Z3) Truncated explicit formula. There is an absolute C2 such that for every primitive χ modulo
N and 2 ≤ T ≤ u,
X uρ u log2 (N uT ) X 1
ψ0 (u, χ) + ≤ C2 + log(N u) + , (III.22)
|γ| 1, integrating the Dirichlet series term by term,
P
1 L′
Z ∞
δ
H(s) = ψ0 (u, χ) u−s−1 du = − (s, χ) − . (III.24)
1 sL s−1
By Step 4 below (which does not use the present step), |ψ0 (u, χ)|u−2 is integrable; for this one needs,
for a real χ, only some interval (1 − cχ , 1) free of zeros, which exists because L(1, χ) ̸= 0. So we
′
may let s → 1+ under the integral. This gives H(1) = − LL (1, χ) for χ = ̸ 1, and H(1) = −1 − γ for
ζ′
χ = 1, from the expansion − ζ (s) = s−1 − γ + O(s − 1). On the other hand, by partial summation,
1
Z y
du y
= Λ(n)χ(n) 1 (III.25)
X
y ψ(u, χ) − .
1 u2 n≤y
n
Hence y 1y ψ0 u−2 du equals this sum minus δy log y. Subtracting yH(1) gives −y y∞ ψ0 u−2 du =
R R
n≤y Λ(n)χ(n)(y/n − 1) − δy log y − yH(1), which is Gχ (y) in both cases. This is also where the
P
constant 1 + γ in the definition of G comes from.
Step 2: the zero-free region, as used. Let c0 > 0 and c = min(c1 , c0 ). Let χ be a
character modulo N and λ ≥ max(1, log N ). If χ is real and nonprincipal, suppose L(β, χ) ̸= 0 for
1 − c0 < β < 1. Then every zero with β < 1 satisfies
c
β ≤1− . (III.26)
λ + log(|γ| + 3)
Proof. If not, β > 1 − c1 / log(N (|γ| + 2)), so by (Z1) χ is real and nonprincipal and ρ = β is real.
But then β > 1 − c/(λ + log 3) ≥ 1 − c0 , since c ≤ c0 and λ ≥ 1; this is against the assumption.
Zeros are not close to 0. If moreover χ is primitive, a zero with 0 < β < 1 and |γ| ≤ 2 has
β ≥ c/(3λ). Indeed 1 − ρ is a zero of L(s, χ), and (III.26) applies to χ (if χ is real it equals χ); so
1 − β ≤ 1 − c/(λ + log 5), and log 5 ≤ 2λ.
Step 3: the sum of 1/|ρ|. Under the assumptions of Step 2 with χ primitive, for T ≥ 1,
1 9
≤ C3 (λ + log T )2 , C3 = 6C1 +4 . (III.27)
X
|γ| 0, depending
only on c0 , such that for u ≥ e, with v = log u,
c2 v
|ψ0 (u, χ)| ≤ C4 u (v + λ)2 exp − √ . (III.28)
λ+ v
√
Proof. Take T = exp( v) in (Z3); then 2 ≤ T ≤ u. Put c2 = min(c/3, 1) and E = exp(−c2 v/(λ +
√ √ √
v)). For |γ| < T we have log(|γ| + 3) ≤ v + 2, so by (III.26) β ≤ 1 − c/(λ + v + 2) ≤
√ √
1 − c2 /(λ + v), and |uρ | ≤ uE. By Step 3, |γ| 0 and c′ > 0. For large x, if y ≥ 13 exp(xε/3 ), 1 ≤ λ ≤
Axε/5 (log x)2 , and (III.28) holds for u ≥ y, then y y∞ |ψ0 (u, χ)|u−2 du ≤ y exp(−c′ xε/8 ).
R
√
Proof. Substitute u = ev . The integral is at most C4 V∞ (v + λ)2 exp(−c2 v/(λ + v)) dv with
R
V = log y ≥ 12 xε/3 . For v ≥ 12 xε/3 and large x we have λ ≤ v 61/100 , because v 61/100 ≥ 2−1 xε/5 xε/300
√
and 2A(log x)2 ≤ xε/300 . Also v ≤ v 61/100 and λ ≤ v. Hence the exponent is at least c22 v 39/100 ,
and (v + λ)2 ≤ 4v 2 . Now v 39/100 ≥ 12 x13ε/100 , so
exp − c22 v 39/100 ≤ exp − c82 x13ε/100 exp − c42 v 39/100 , (III.29)
and 0∞ v 2 exp(− c42 v 39/100 )dv is a finite constant. So the integral is at most a constant times
R
exp(− c82 x13ε/100 ). Since 100
13
> 81 , this is at most exp(−c′ xε/8 ) for large x.
√
Where the exponents come from. The classical saving is exp(−c log y/(log N + log y)). Here
log N is about xε/5 and log y about xε/3 , so the saving is exp(−xε/3−ε/5−o(1) ), and 13 − 51 = 15
2
> 18 .
Step 6: exceptional zeros. It remains to produce c0 > 0, depending only on B, such that
every real primitive χ ̸= 1 with conductor N | Qqd has no zero in (1 − c0 , 1).
(i) N divisible by q or d. This is the hypothesis of the theorem, with 400
1
.
(ii) N prime to qd. Then N | Q. The conductor of a real primitive character is not divisible by
16 or by the square of an odd prime. (If p2 | N with p odd, each u ≡ 1 mod N/p is a square modulo
2
N : 1 + Np t ≡ 1 + Np t p+1
2 . So χ(u) = 1 and χ is defined modulo N/p. Similarly 1 + N2 t ≡ (1 + N4 t)2
when 16 | N .) So N divides 8 times the odd part of Q, and N ≤ 8B. There are finitely many
characters of conductor at most 8B. For each, L(1, χ) ̸= 0 and L(s, χ) is continuous, so L(β, χ) ̸= 0
on some interval (1 − cχ , 1]. Let c0 be the smallest of the cχ and of 400 1
. (For ζ only the real
characters of conductor 4 and 8 occur. The constant can be made explicit in B by (Q1) of §III.1: a
real zero β of a character of conductor N ≤ 8B has 1 − β ≥ min(c1 / log(3N ), c/N ).)
How part (e) of Theorem III.1.1 gives (i). A real primitive character χ of conductor N | Qqd
divisible by q or d factors as ρ χiq χjd with (i, j) ̸= (0, 0) and ρ real primitive of conductor dividing 8
times the odd part of Q. Part (e), on the real axis, says L(β, χ) ̸= 0 for β > 1 − 400 1
.
Conclusion. Let χ be primitive of conductor N | Qqd, and λ = max(1, log N ) ≤ 3xε/5 (log x)2 .
By Step 6 the assumption of Step 2 holds with c0 . So (III.28) holds for all u ≥ e, and Steps 1 and 5
give the theorem.
Remark. The Lean proof obtains the limits in Step 1, and the convergence of the integral,
from the Siegel–Walfisz theorem. The argument above uses only (III.28) for the one character in
question.
63
III.3 The sets of columns T2 and T
The table of §I.5 has n columns, indexed by the group H of §I.5. It is cyclic of order n; fix an
isomorphism with A = Z/nZ and write it additively, so that the translates uT2 of Part II become
u0 + T2 . The construction singles out two sets of columns. Both are needed only through the
combinatorial properties proved here; how (M) uses them belongs to the part not yet written.
Throughout, x is large, 0 < ε < 52
1
, and x2+8ε ≤ n ≤ 2x2+8ε .
The set T2 : a construction of Behrend
Theorem III.3.1 (hostThickExists in [27]; hostThickOrbit, thickFrame2). Let L ≥ 2 be an
integer with L ≤ CL log x. For x large in terms of ε and CL there is a set T2 of positive integers
with the following properties.
(a) 4Ls < n for every s ∈ T2 , and |T2 | ≥ x−ε n.
(b) No relation. If s1 , s2 , s3 ∈ T2 are distinct and i, j are integers with 0 < |i|, |j| ≤ L, i =
̸ j, then
j(s2 − s1 ) ̸= i(s3 − s1 ).
(c) Arithmetic progressions meet translates of T2 rarely. Let u0 , t, h ∈ Z/nZ. If the additive order
of h exceeds 4L, then among any L consecutive terms t + ih at most two lie in u0 + T2 . If the
order of h is at most 4L, at most one element of Z/nZ on the progression lies in u0 + T2 .
For i = 1, j = 2 property (b) says that T2 contains no three-term arithmetic progression. The proof
is Behrend’s [Beh46]: points on a sphere, written in a large base.
Proof. A sphere. Let S, D ≥ 2 be integers. The DS vectors v ∈ {0, . . . , D −1}S have |v|2 < SD2 ,
so some sphere |v|2 = k contains at least DS−2 /S of them. Call this set Σ. Three points of a
sphere are never collinear. More precisely: if v1 , v2 , v3 ∈ Σ and j(v2 − v1 ) = i(v3 − v1 ) with
i, j =
̸ 0, i ≠ j, then v2 = v1 . Indeed put b = v2 − v1 , c = v3 − v1 . Since |v1 + b|2 = |v1 |2 we have
2v1 · b + |b|2 = 0, and likewise for c. Substitute b = ji c in the first relation and multiply by j 2 /i:
this gives 2j v1 · c + i|c|2 = 0. Subtract j times the second relation: (i − j)|c|2 = 0. So c = 0 and
then b = 0.
Base B = 4LD. Map v 7→ s(v) = m vm B m , and let T be the image of Σ. Suppose
P
j(s(v2 ) − s(v1 )) = i(s(v3 ) − s(v1 )) with i, j as in (b). The digits cm = j(v2,m − v1,m ) − i(v3,m − v1,m )
satisfy |cm | ≤ 2L(D − 1) < B and m cm B m = 0; hence every cm = 0 (look at the lowest nonzero
P
digit modulo B). By the statement about the sphere, v2 = v1 . This is (b) for T .
√
Parameters. Put u = log n log L, S = ⌊u/ log L⌋, a = (log n−log L−3)/S and D = ⌊ea /(4L)⌋.
For large x one has S ≥ 32, a ≥ 7 + log L, hence D ≥ ea /(8L) ≥ 2, and a ≤ 2u. Then
B S ≤ eSa < n/(4L), so 4Ls < n on T , and
DS−2 n
≥ exp − 32 log n log L . (III.30)
p
|T | ≥
S L
Since log L ≤ log log x + O(1), the right side exceeds 4x−ε n for large x. Remove 0 from T to get T2 .
This proves (a) and (b).
Property (c). Let the order of h exceed 4L, and suppose t + ah, t + bh, t + ch lie in u0 + T2
with a < b < c < a + L: they equal u0 + sa , u0 + sb , u0 + sc . The s’s are distinct because (b − a)h,
(c − a)h, (c − b)h ̸= 0. Eliminating h,
X = (c − a)(sb − sa ) − (b − a)(sc − sa ) ≡ 0 mod n. (III.31)
64
But |X| < n, because 4Ls < n for each s. So X = 0, against (b). If the order m of h is at most 4L
and two points of the progression are u0 + s1 , u0 + s2 , then n | m(s1 − s2 ) and |m(s1 − s2 )| < n, so
s1 = s2 .
The set T : a random choice
Let δ1 = x−ε/8 . The set T ⊂ A is chosen so that |T | is about δ1 n and so that, for most primes p
and most t ∈ T , the points of T on the orbit of t under the Frobenius of p are evenly spread. The
precise meaning is this.
Good orbits. Let n′ ≥ 1 and W ⊂ {0, . . . , n′ − 1}, read cyclically. Fix integers L2 , K, r ≥ 1,
and put u = ⌈1/δ1 ⌉, ℓ = (6r + 1)u, M0 = ⌈(5 + 8ε) 106r+1 log x⌉, G′ = (M0 + 5)ℓ and N1 =
⌈max(6eG′ δ1 , (5 + 8ε) log2 x)⌉ + 1. Cut {0, . . . , n′ − 1} into ⌊n′ /ℓ⌋ consecutive tiles of length ℓ (a
remainder shorter than ℓ is ignored, and tiles are numbered cyclically), and each tile into 6r + 1
cells of length u, numbered 1, . . . , 6r + 1. A tile is marked if its odd-numbered cells contain no
point of W and each even-numbered cell contains exactly one point of W , at an even position. Call
W good if
(G1) every cyclic window of length L2 contains at most K points of W ;
(G2) among any M0 + 1 consecutive tiles (cyclically) at least one is marked;
(G3) every cyclic window of length G′ contains at most N1 points of W .
Theorem III.3.2 (topsGroup, topsChoiceH). Let Lb > 0, K = ⌈8(4 + 8ε)/ε⌉ and r = K + 1.
There are Cb and x0 such that for x ≥ x0 the following holds. Let L2 ≤ Lb . Let P be a finite set of
primes, and for p ∈ P let hp ∈ A have order np ≥ xε/4 . For each p and each coset of ⟨hp ⟩ fix a
base point b on it; for t ∈ A let Wp (t, T ) = {0 ≤ i < np : b + ihp ∈ T }, b being the base point of the
coset of t. Then there is T ⊂ A with
X log p X log p
2 δ1 n ≤ |T | ≤ 2δ1 n
1
and #{t ∈ T : Wp (t, T ) is not good} ≤ 4Cb x−2 δ1 n .
p∈P
p p∈P
p
(III.32)
In the application P consists of the inert primes below exp(xε/3 ) whose Frobenius has order at least
xε/4 in A, and L2 is bounded by a constant depending only on ε (the source gives 2(620 Co /ε + 1)
for primes p > xε/(5Co ) , with a constant Co ≥ 2 that is defined in the source). That L2 is bounded
uses Remark (3) of §III.1: every inert prime exceeds Y ε/5 .
Proof. Choose T at random: each element of A belongs to T independently with probability δ1 .
(For Dirichlet characters other than the trivial one the source uses two different probabilities, δ1
and 34 δ1 , on different cosets. For ζ there is one coset.)
Size. |T | has mean δ1 n and variance at most δ1 n.
One orbit. Fix p and t, write n′ = np , and condition on t ∈ T . The other points of the orbit
remain independent. We bound the probability that W = Wp (t, T ) is not good.
(G1) fails only if some window of length L2 contains K points other than t. For one window
this has probability at most LK2 δ1K ≤ LK −Kε/8 ≤ LK x−(4+8ε) . There are n′ windows.
b x b
(G2). A given tile not containing t is marked with probability at least
3r+1 u 3r
(1 − δ1 )u 2 δ1 (1 − δ1 )
u−1
≥ 10−(6r+1) , (III.33)
65
since uδ1 is between 1 and 1.01. Among M0 + 1 consecutive tiles at most one contains t; the
other M0 are disjoint, hence independent. The probability that none is marked is at most
(1 − 10−(6r+1) )M0 ≤ x−(5+8ε) . There are at most n′ starting tiles. (Here and for (G1), (G3) one
needs L2 ≤ n′ and G′ = (M0 + 5)ℓ ≤ n′ , so that windows and runs of tiles do not wrap around;
this holds for large x because n′ ≥ xε/4 , L2 ≤ Lb and ℓ ≤ 2(6r + 1)xε/8 .)
(G3) fails only if some window of length G′ contains N1 points other than t; then it contains
′
j = N1 − 1 such points. Since j ≥ 6eG′ δ1 , the probability of the latter is at most Gj δ1j ≤
(eG′ δ1 /j)j ≤ 6−j , which is at most x−(5+8ε) . There are n′ windows.
Since n′ ≤ n ≤ 2x2+8ε , the three bounds add up to at most Cb x−2 . So the expected number of
t ∈ T with Wp (t, T ) not good is at most Cb x−2 δ1 n, for every p.
Choice. We may assume P is not empty. Let V = (|T | − δ1 n)2 /(2δ1 n) and let B be the left side
of the inequality of the theorem divided by four times its expected upper bound Cb x−2 δ1 n p logp p .
P
Then E(V ) ≤ 12 and E(B) ≤√ 14 . So some T has V + B ≤ 1. For it B ≤ 1, which is the second
assertion, and |T | − δ1 n ≤ 2δ1 n ≤ 12 δ1 n, which gives the first.
III.4 The size of the theta kernels (partly written)
This section concerns the upper bound for the size of the maps: inequality (6.5) of Part II, first
line. It is partly written. The following are proved below: the one-variable model; the coefficient
estimate in several variables and in several blocks, with the period; the bound for the polynomial
on the unit sphere; the bound for one map and for a chain of maps; the bound for a theta function
by its mean square; and the hook product. Four passages are conditional: the reference mean
square; the eigenvalues of a lift; the comparison, row by row, of the kernel degrees with the hook
product; and the norm of the pairing at the end of the chain. The last two are deductions from
inputs of the source that are named where they are used. The passage from these estimates to (6.5)
is not written. It needs the arithmetic statements listed at the end of the section, none of which
is proved in these pages. So the first line of (6.5) is still quoted. We begin
√ with the one-variable
model. It shows where the two features of the bound come from: the size n! and the sign of the
term with the period.
2
The model. Let F be an entire function with |F (z)| ≤ S e|z| /2 for all z, where S > 0. Then
for n ≥ 1
e n/2
|F (n) (0)| ≤ S n! , (III.34)
n
so that, if F (n) (0) ̸= 0,
log |F (n) (0)|2 ≤ 2 log S + n log n − n + O(log(n + 1)). (III.35)
2 √
Proof. By Cauchy’s formula on the circle |z| = r, |F (n) (0)| ≤ n! S er /2 r−n . Take r = n. The
second form follows from Stirling’s formula, log n! = n log n − n + O(log(n + 1)).
So a Gaussian bound for the function gives, for the square of its nth derivative, the size n log n − n:
this is log n!. Compare a function with a pole at finite distance, such as 1/(1 − z): its nth derivative
at 0 is n!, and the logarithm of the square is 2 log n!. That was the situation of the lemniscate sine
in §I.3. Theta functions satisfy such a Gaussian bound. This is point (i) of §I.4.2.
The period. If the coordinate is changed by a factor, z = λw with |λ|2 = 1/Ω, then the nth
derivative with respect to w is λn times that with respect to z, and the bound for the logarithm of
66
its square becomes
n log n − n − n log Ω + 2 log S + O(log(n + 1)). (III.36)
This is the shape a(log a − 1 − log Ω) of (6.5), with its sign.
√ The constant relevant there is ΩE , not
the elliptic Ω of §I.7; the two differ by the factor rdE / d.
The paragraphs that follow carry the model over to the setting of the proof: several complex
variables, a polynomial with values in linear maps in place of a single derivative, several blocks
of variables (one for each place), and sums over classes of lattices. Those up to and including the
bound for a theta function by its mean square, and the hook product, are complete as stated; the
four passages named above are conditional on inputs from the source. The first of them assumes
a Gaussian bound S for the function and a bound for the polynomial on the unit sphere; both
assumptions are then proved in an abstract setting; for the first, the mean square of one reference
theta function is bounded under the inputs (I1)–(I5) listed at the head of that paragraph. That
the actual theta kernels and polynomials of the construction are instances of this setting is taken
from the source: the hypotheses of the majorant lemma with v = rdE , the form of the constant
S with its factors β and c, and the identification of the polynomial P of the construction with a
substitution of the kind treated below.
Several variables: a coefficient estimate on a sphere. Let F (z) = α cα z α be entire
P
on Cd , d ≥ 1, with |F (z)| ≤ S exp(∥z∥2 /2), where ∥z∥2 = j |zj |2 . Let P (z̄) = |α|=n Aα z̄ α be
P P
homogeneous, with coefficients in a finite-dimensional normed space, for example a space of linear
Q α
maps. Here α is a d-tuple of nonnegative integers, |α| = j αj , z α = j zj j and α! = j αj !.
P Q
Suppose, for unit vectors u, that the following holds, factors with νi = 0 being omitted:
∥P (ū)∥ ≤ si (u)νi , si (u) ≥ 0, si (u)2 = 1, νi ≥ 0, νi = n. (III.37)
Y X X
i i i
Put h(v) = v log v − v for v > 0 and h(0) = 0. Then
!
cα α!Aα ≤ S e1/2 (2n + 2)d exp h(νi ) . (III.38)
X X
1
2
|α|=n i
Proof. Let du be probability measure on the unit sphere in Cd . Independent rotations of
the coordinates show that the mean of uβ ūα vanishes unless α = β. Its value for α = β is
2
α!(d − 1)!/(n + d − 1)!. Indeed, integrate |z α |2 against π −d e−∥z∥ first in theR separate complex
coordinates and then in polar coordinates. The two radial moments needed are 0∞ tj e−t dt = j! for
j = d − 1, n + d − 1. Thus for r > 0 the sum to be bounded equals
(n + d − 1)!
Z
F (ru)P (ū) du. (III.39)
(d − 1)! rn
Termwise integration is valid by uniform convergence on the sphere. For n > 0 the weighted
arithmetic–geometric mean inequality, with weights νi /n, gives i si (u)νi ≤ νi >0 (νi /n)νi /2 . Take
Q Q
√
r = n. After removing S exp( 12 i h(νi )), the remaining factor is
P
n+d−1
!
n!(e/n)n ≤ en(n + 1)d−1 ≤ e1/2 (2n + 2)d . (III.40)
d−1
The first inequality follows by comparing j 0. The differential is scaled so that the Weierstrass invariants
satisfy g23 − 27g32 = 1. Thus Aρ = |ω|2 y is the area of this particular lattice. The marked point and
its conjugates determine which of the three even theta functions to use.
In coordinates z = u + vτ on the torus, put
2
θ(z) = (α, β) ∈ {(0, 0), (0, 12 ), ( 12 , 0)}. (III.53)
X
eπim τ +2πim(z+β) ,
m∈Z+α
Choose the characteristic whose zero is the marked half-period in these coordinates. Write
T = |θ(0)| > 0. The normalisation of the section at the origin divides it by θ(0); this is input (I5).
In Jacobi coordinates its squared norm, averaged over the torus with total mass one, is
|θ(u + vτ )|2 −2πyv2 1
Z 1Z 1
Hρ = e du dv = √ . (III.54)
0 0 T 2 2y T 2
71
2
Indeed, integration in u kills the mixed Fourier terms. The remaining integrals of e−2πy(m+v) tile
the real line. Absolute convergence on the closed parallelogram justifies termwise integration.
The arithmetic normalisation. On the curve Y 2 = 4x3 − g2 x − g3 , let e be the x-coordinate
of the marked point, and e′ , e′′ the other two roots of the cubic. Put a = e − e′ , b = e − e′′ ,
P = ab = 3e2 − g2 /4, and X = P −3 ∈ M . All three roots are distinct. The discriminant identity
gives 16P 2 (a − b)2 = 1, while the sum of the roots gives g2 = 43 ((a − b)2 + P ). Substituting these
identities into j = 1728g23 yields
(X + 16)3 − jX = 0. (III.55)
The singular modulus j is an algebraic integer by complex multiplication. The displayed monic
equation therefore makes X a nonzero algebraic integer. This is the only integrality calculation
needed here.
For clarity, the classical Jacobi identities used to recover the absolute value of X are as follows.
The three half-period differences, up to sign, are π 2 ω −2 times the fourth powers of the three even
theta constants. The two differences through e use the other two constants. The product of all
three constants is 2η(τ )3 , where η(τ ) = eπiτ /12 n≥1 (1 − e2πinτ ). Finally, the discriminant identity
Q
1 = (2π/ω)12 η(τ )24 gives |ω| = 2π|η(τ )|2 . Hence at each embedding
|ρX| = (T /|η(τ )|)12 , Aρ = 4π 2 y|η(τ )|4 ,
log Hρ = 21 log(2π 2 ) − 12 log Aρ − 61 log |ρX|.
(III.56)
Here the same algebraic number X works at every embedding because its formula uses just the
marked coordinate e and g2 .
The conjugates of a curve with complex multiplication by OK run equally often through the
ideal classes of K (input (I2)). By the class mean √
period formula of §I.7 in the discriminant-one
scale (input (I3)), ⟨log Aρ ⟩ = log π + γ + log(2π/ d) − cd , where γ is Euler’s constant. Since
|NM/Q X| is a positive integer, averaging gives
√ log |NM/Q X| √
⟨log Hρ ⟩ = 12 (log d − γ + cd ) − ≤ 12 (log d − γ + cd ). (III.57)
6[M : Q]
′
The reference section of ranks N, N ′ has squared norm Θ0,ρ = Hρf N N , where f = 2n is the number
of weight places (n as in Part II). This last identity uses the tensor-power description of the
normalised reference section; the integral of a product on independent elliptic coordinates factors.
Thus its average logarithm is at most f N N ′ times the displayed bound. This is the conclusion
stated at the head of the paragraph.
The step that uses Lsc . This step is separate from the conclusion above. In the application
|cd | ≤ 4 log x, by Theorem III.1.1(d) with M = 1 and χ = χd . Put vhost = rdE /2, and let Lsc be
the logarithmic scale √ of the Lean development; it is defined
√ in the source. In the parameters of
the chain one has log d − γ ≤ log vhost (the inequality d ≤ vhost itself holds for N ≥ 8 but fails
for N = 4, q = 3) and log vhost + |cd | ≤ 2Lsc ; both inequalities are taken from the source. So the
contribution is at most f N N ′ Lsc .
The hook product. The hook product enters as an exact Gaussian integral. Given integers
λ1 ≥ · · · ≥ λs ≥ 0, form a diagram with λi boxes in row i, left justified. The hook length of a box
is one plus the numbers of boxes to its right in the same row and below it in the same column. Let
H(λ) be the product of all hook lengths, with empty product one. For a square complex matrix X
of size s put
s
pλ (X) = det(X1:i,1:i )λi −λi+1 , λs+1 = 0. (III.58)
Y
i=1
72
With the Gaussian probability norm used above, ∥pλ ∥2G = H(λ).
Here is a proof by removing full columns. Write X = QR as in the proof of the polynomial
estimate. Conditional on the preceding columns, Rii 2 is the squared length of a standard complex
Gaussian vector in dimension s − i + 1. Its density is ts−i e−t /(s − i)! on t > 0, by polar integration.
This conditional law is constant, so the Rii
2 are independent. Averaging first over unitary matrices
on the left as above gives
! s
Z Y (λi + s − i)!
∥pλ ∥2G = |pλ (U )|2 dU , (III.59)
U (s) i=1
(s − i)!
where dU has total mass one. Put a = λs and µi = λi − a. Then pλ = (det X)a pµ , so the unitary
integrals for λ and µ agree. The ratio of their squared Gaussian norms is consequently
s
(λi + s − i)!
(III.60)
Y
.
i=1
(λi − a + s − i)!
These are exactly the hook lengths in the first a columns, since the hook at (i, j) there has length
λi − j + 1 + s − i. Deleting those columns leaves all other hook lengths unchanged. The last row is
now empty, and pµ depends only on the upper left (s − 1) by (s − 1) matrix. Induction on s proves
the formula, including a = 0.
The eigenvalues of a lift: what the inner-product formula contributes. This paragraph
is conditional on the following hypotheses, which are taken from the source and not proved here.
The notation is explained below the list.
(E1) At each weight place the polynomial in the theta lift is a product of two polynomials of the
form pλ on disjoint sets of variables.
(E2) On each non-empty simultaneous Hecke component of the target, the Rallis inner-product
formula, evaluated with the chosen local test functions, gives T T ∗ = λI with the scalar λ of
the form displayed below; in it D0 ≥ 1, and the level index times the squared norm of the
finite test function equals 1/DS .
(E3) The parameter conditions: 1 ≤ f ≤ ϱ, 1 ≤ r ≤ ϱ/2, t ≥ r + 5, ϱ ≥ 8, L ≥ 5.5, log ϱ ≤ L, and
log(dρ + 2) ≤ 2L.
(E4) The bound for the mass M0 , the bounds for the normalising factors DS , DS c , and the
formal-degree estimate for Jρ , as displayed below.
(E5) The local parameter bound |x| ≤ (Np)(r−t−2)/2 for the parameters of the unramified Euler
factors.
(E6) The lift used at the distinguished embedding is S = P T with P an orthogonal projection on
the target, and the Hecke components span the target.
(E7) At the other embeddings the lower bound of the conclusion holds as well.
Conclusion. Under (E1)–(E7) every positive eigenvalue of SS ∗ is at least
X
exp log H(A+
ρ ) − 8f ϱ L .
2
(III.61)
ρ
The lower bound for these eigenvalues is one of the inputs listed at the end of this section. Fix
one step of the chain. At each of the f = 2n weight places ρ, the polynomial in the theta lift is a
73
product of two polynomials of the form pλ on disjoint sets of variables, by (E1). Write A+ρ and
A−ρ for their diagrams; their row lengths are the column degrees of the two products of minors,
arranged in decreasing order. Reversing rows or columns in a corner minor leaves its absolute
value unchanged after the corresponding change of variables. The Gaussian integrals over disjoint
coordinates factor, giving the squared norm ρ H(A+ ρ )H(Aρ ).
−
Q
Let T be the full theta lift between its finite-dimensional complex inner-product spaces. On
each non-empty simultaneous Hecke component of the target the Rallis inner-product formula,
evaluated with the chosen local test functions, gives T T ∗ = λI, by (E2). In the normalisation of
the source its scalar has the form
E0 Y −
λ = D0 M0 ρ )H(Aρ )Jρ .
H(A+ (III.62)
DS DS c ρ
Here D0 ≥ 1 is a product of representation dimensions, M0 is the mass (the sum of reciprocal
stabilizer orders), E0 is the absolute value of the unramified Euler product, and DS , DS c are the
positive normalising factors at the exceptional finite places and their complement. The level index
times the squared norm of the finite test function equals 1/DS . The factor Jρ is the integral of the
squared absolute value of a unit vector’s matrix coefficient on the real unitary group at ρ, for the
measure of the formula; it is defined in the source.
Here are the bounds on these factors that suffice. Here k, m and t are parameters of the
construction (in Part II, §6.2: the initial rank k, the number m of lifts, and the odd parameter
written τ0 there), and f = 2n is the number of weight places. Write r = k + m − 1, ϱ = r + t for the
rank parameter, g0 , g1 for the two level primes, b for the weight bound, and L = log(b rdE g0 g1 ϱ).
(In Part II, §6.6 the letter h denotes k + m − 1, which is r here.) By (E3) the parameter conditions
give 1 ≤ f ≤ ϱ, 1 ≤ r ≤ ϱ/2, t ≥ r + 5, ϱ ≥ 8, L ≥ 5.5 and log ϱ ≤ L. The mass bound, the finite
Euler factor estimates and the formal-degree estimate used in the source give
1 √
E0 ≥ e−2rf , DS , DS c ≤ e4rf , − log Jρ ≤ 2 + r2 log 2 2 r2 ϱ(dρ + 2) ,
M0 ≥ ,
(8f 2 )ϱ ϱ!
(III.63)
where dρ = |A+ ρ | + |A − | counts their boxes and log(d + 2) ≤ 2L. The bounds for M , D , D c and
ρ ρ 0 S S
Jρ are (E4). The bound for E0 follows from (E5): the factors of E0 are |1 − x|−1 with r parameters
at each unramified prime ideal p of E. Their absolute values satisfy |x| ≤ (Np)(r−t−2)/2 ≤ (Np)−2 ,
by the local parameter bound and t ≥ r + 5. At most [E : Q] = 2f prime ideals lie above any
rational prime p, and each has norm at least p, so p (Np)−2 ≤ 2f p p−2 ≤ 2f , the last sum being
P P
over the rational primes. Taking logarithms and using log |1 − x| ≤ |x| proves the stated bound
for E0 . The logarithmic losses are at most 4ϱL from the mass, 10rf from the finite factors, and
f (2 + 6r2 L) from the real integrals. Their sum is at most 8f ϱ2 L by the preceding inequalities.
Since every hook length is at least one, this gives
log λ ≥ log H(A+
ρ ) − 8f ϱ L. (III.64)
X
2
ρ
This deduction uses the stated arithmetic and analytic estimates for the test functions.
Finally, by (E6), the lift used at the distinguished embedding is S = P T with P an orthogonal
projection on the target, and the Hecke components span the target. Since T T ∗ is self-adjoint, the
scalar bounds on those components give T T ∗ ≥ cI, where c is the exponential of the last lower
bound. On the range of P one has ⟨SS ∗ x, x⟩ = ∥T ∗ x∥2 ≥ c∥x∥2 . This bounds every positive
eigenvalue of SS ∗ . At the other embeddings the same bound is taken from the Lean development;
this is (E7).
74
Row by row: why only the first row survives. This passage is conditional: it is a
deduction from inputs taken from the source, listed below. It compares the terms i h(νi ) of the
P
block estimate with the hook product of the preceding paragraph, one row at a time. Fix a step
1 ≤ j ≤ m of the chain (steps are numbered from 1 here, as in the source) and one weight place ρ.
Let k and t be the parameters of the preceding paragraph, with k ≥ 2. The terms used are these.
• The kernel degrees are the column degrees of the polynomial of the kernel of this step in the
block at this place. We assume, as input (o) from the source, that these are the degrees νi to
which the block estimate is applied (one block for each weight place); in the Lean development
this is the case: the degrees fed to the block estimate are identified with these numbers
(weights_rows, nu_eq_originRow). We call them also the lengths of the kernel rows. There
are at least k of them.
• A+
ρ is the positive-signature diagram of the preceding paragraph, and H(Aρ ) its hook product.
+
• The weight shift w at this step, the weight parameters u1 , . . . , uk of the domain at this place,
and the parameters v1 , . . . , vk−1 of the positive-signature rows are as in the source (Lean:
origin_rewrite, origin_interlace, with the formulas originRow, originBig). We do not
define them here.
• A weight place is ordinary or distinguished, as in the source (Lean: the two cases
RowsOriginOff and RowsOriginTop). At a distinguished place, which we denote by t, there
are two integers at ≥ 0 and bt ≥ 0; at is the number at of the inequality at the end of this
section. The index t is not the parameter t.
• ∆ is the width of the weight interval (in the source, Sw).
• B is a bound for the sum of the kernel degrees other than the first k.
• La = log(a rdE ), with a as in the inequality at the end of this section.
Put
k + (2j − 1)(t + 1)
ς= , W = w − ς, δ = (j − 1)t, (III.65)
2
k + 1 − 2i k − 2i
xi = ui − , yi = vi − . (III.66)
2 2
Assumed here, from the source:
(a) The degree formulas. W ≥ 0, the first k kernel degrees are given by the formulas for νi below,
and k − 1 of the rows of A+
ρ have the lengths Ri below:
νi = W + xi (1 ≤ i ≤ k), Ri = W + yi + δ (1 ≤ i < k). (III.67)
There may be further rows on either side. (Lean: nu_eq_originRow, aPlus_eq_originBig.)
(b) Strict interlacing. ui > vi > ui+1 for i < k.
(c) Integrality and parity. The degrees νi and Ri are nonnegative integers; k ≥ 2 is even, t ≥ 1 is
odd, and the weight w is an integer. Then ς, the xi and the yi are integers. (The argument
below uses only that xi − yi = νi − Ri + δ and yi − xi+1 = Ri − δ − νi+1 are integers, and this
follows from the integrality of νi , Ri and δ.)
(d) The endpoint bounds. At an ordinary place, x1 ≤ ∆ and xk ≥ 0. At a distinguished place,
x1 = at , xk = −bt , yk−1 ≥ 0 and y1 ≤ ∆.
75
(e) The extra degrees and the sizes. The sum of the kernel degrees other than the first k is at most
B, and B ≤ a rdE ; w + ∆ ≤ a rdE at an ordinary place, and w + at ≤ a rdE at a distinguished
place. The size hypotheses of the source permit B = m(t + 1)η, with η the bound used there
for the additional rows.
None of (a)–(e) is proved in these pages.
The comparison of rows. Strict interlacing and integrality give xi ≥ yi ≥ xi+1 , since both
differences are integers greater than −1/2. Consequently
Ri ≥ νi+1 (1 ≤ i < k). (III.68)
In particular every one of the first k kernel rows except the first has its own row of at least the
same length in A+ρ.
Recall that h(v) = v log v − v for v > 0 and h(0) = 0. For integers d ≥ 0 one has h(d) ≤ log(d!)
by integration of log x. The hooks along a row of length R are at least R, R − 1, . . . , 1. Hence
log H(A+ ρ)≥ i σj − j for i < j and the reverse inequality for i > j. The
omission excludes zero when i = j. Put Γ = λ1 − λΨ + Ψ and d1 = ψ i=1 min(λi − λi+1 , d0 ). Each
P
factor has absolute value at most Γ and there are ψd1 of them. What the source shows to be integral
is the transposed map ϑκC ∗ (Lean: BranchMap); the map ϑκC itself can have denominators. A
map and its adjoint have the same norm, so the statements about norms here are not affected. By
CC ∗ = I/κ, the squared norm of either map is ϑ2 κ.
One can also bound κ by the Gaussian calculation above. The two polynomials in its ratio,
ξ and pλ′ , have identical column degrees, and both are unchanged when a multiple of an earlier
77
column is added to a later column. So both lie in Wλ′ , the argument by X = QR applies to both,
and their radial factors cancel. Equal column degrees alone would not be enough. Every minor of a
unitary matrix has absolute value at most one. In addition, the Gaussian hook formula and the
radial moments give
!−1 Ψ Ψ
1 Y (λ′i + Ψ − i)! Y λ′i + Ψ − i
Z !
κ≤ |pλ′ (U )| dU
2
= ≤ ≤ ΓΨ(Ψ−1)/2 . (III.73)
U (Ψ) H(λ ) i=1 (Ψ − i)!
′
i=1
Ψ−i
The middle bound uses the row factorial lower bound for H(λ′ ).
The level factor is taken from the source, not derived here (Lean: brAt). At level T the map
has the additional scalar ϖ−ℓ with |ϖ|2 = T in orthonormal coordinates. Its squared norm is
consequently T −ℓ ϑ2 κ. Factors at independent places multiply. Taking logarithms, the full bound is
ℓρ log T + 2ψd1,ρ + Ψ(Ψ − 1)/2 log Γρ . (III.74)
X X
−
ρ ρ
The final functional pairs these images with the target by a weighted sum over lattice classes,
with reciprocal stabilizer orders as weights. That it has this form is an input from the source.
Cauchy–Schwarz for this finite sum gives precisely the same bound for the squared pairing on unit
vectors. Here unit vectors are tuples indexed by the classes, of norm one in the weighted class norms
p w(p)∥fp ∥ of the paragraph on sums over classes, with w(p) the reciprocal stabilizer order; and
P 2
the squared pairing is the squared absolute value of the sum over the classes p of w(p) times the
inner product of the image of fp with the pth entry of the second tuple (Lean: sum_pair_le).
What this proves, and what it assumes: it proves that CC ∗ = I/κ in the normalised norms,
so that ϑκC and its adjoint have squared norm ϑ2 κ, that |ϑ| ≤ Γψd1 and κ ≤ ΓΨ(Ψ−1)/2 , and the
displayed bound for the squared pairing; it assumes interlacing of the two weights with a common
last entry, that the last map of the construction is ϑκC up to the level factor, that this factor is
ϖ−ℓ , and that the final functional is the weighted sum over classes, and it does not substitute the
actual weights into ℓ and d1 .
Relation to the Lean proof. The estimate for one block, with the constant e1/2 (2n +
2)d exp( 12 i h(νi )) and the factor Ω−n τ , is the statement of the Lean development that the
P
τ
final theorem depends on (blockBound_of_fockJet, with gaussianCoeffIntegral); the Lean
statement is for the square of the norm, which is why the period enters as Ω−n τ
τ ; with the factor
|I| |J| from the sum over classes it becomes the bound for one lift. The source proves it through an
integral over Cd against the Gaussian measure; the proof above uses the sphere instead. For a chain
of m lifts the source bounds the norm of each map at each archimedean place in this way, multiplies
the bounds (the norm of a composition is at most the product of the norms), and averages the
logarithms over the places (chain_coupled_bound). The hook products and the lower bound for
the eigenvalues of a lift enter the assembly as a term subtracted for each map: in the source the
bound for the height has the form hA + j (hB,j − hX,j ), where hB,j comes from the block estimate
P
and hX,j from the eigenvalue bound.
What remains to be written. In the Lean development the passage from this estimate to
the first line of (6.5) is an assembly (divClausesR_of) of the following further statements, each
proved elsewhere in the development. The assembly itself is carried out in §III.7, which proves
that the first line of (6.5) follows from a list of such statements, and proves one statement on the
periods. The statements themselves are not proved in these pages. Until they are, the first line of
(6.5) remains a quoted statement.
(i) The inputs (I1)–(I5) of the reference mean square above: the integrality of the singular modulus
(classical), that the conjugates of the curve run equally often through the ideal classes, the
78
formula for the mean of log Aρ over the classes (the class mean period formula), the description
of the reference section as a tensor power, and the identification of the reference section at ρ
with θ/θ(0) and its metric. Also: that the chain of maps is defined over a field containing
M , with Hρ̄ = Hρ ; the two inequalities for vhost and Lsc used in the last step there; and the
bounds for β and c in the constant S.
(ii) The pairing at the end of the chain, and the eigenvalues of a lift. The norm of the pairing
at the end of the chain and the row-by-row comparison are proved above, conditionally on
the inputs named there. What remains: the integral realisation of the last map, with its
level factor; the substitution of the actual weights of the construction (into ℓ and d1 for the
pairing, and into the inputs (o) and (a)–(e) of the comparison); and, for the lower bound
for the eigenvalues of a lift, the hypotheses (E1)–(E7) of the conditional paragraph above:
the form of the polynomial at each weight place (E1); the Rallis inner-product formula with
the chosen test functions and the form of its scalar (E2); the parameter conditions (E3); the
bound for the mass, the bounds for the normalising factors and the formal-degree estimate
(E4); the local parameter bound (E5); the form S = P T of the lift and the spanning by Hecke
components (E6); and the bound at the other embeddings (E7).
(iii) The periods: that the average over the archimedean places of log Ωτ is the constant log ΩE
of §I.7
√ (the constant relevant here is ΩE , not the elliptic Ω; the two differ by the factor
rdE / d), and that the same constant appears in the covolume of the lattices. This is where
the Chowla–Selberg formula is used, and where the two occurrences of log Ω cancel up to a
small error.
(iv) Local index bounds at the finite places, including the term F0 of Part II, §6.7, which is supplied
by the divisibility statement (item 1 of §I.7).
(v) The passage between levels, and bounds for class numbers and for the constants of the kernels.
(vi) That the kernels and polynomials of the construction are instances of the abstract setting:
the lattice hypothesis with v = rdE , the form (III.51) of S, the identification of P with a
substitution, and the tuple representation of each map with matching norms.
The inequality that results is, in the notation of Part II, §6.7,
X
µCC +h−F0 ≤ m at log at −at −at log ΩE −at log T +C |T | a La +C m (B0 +Nc ) x2ε , (III.75)
t∈T
where a = x b (the orders at are not assumed to be at most a; Part II allows at up to (x + 1) b),
La = log(a rdE ), B0 = 2n m b xε and Nc = |T2 | k b x2ε /n2 . (In the Lean record RecNums.rhs the
last factor is xε times a constant, and the form above is weaker; the final Lean statement has the
factor xCabs ε with a free constant Cabs , and the form above is its case Cabs = 2. The value of Nc
is the one fixed in the Lean file DataArith.lean.) Here h is an upper bound for the height of
the pairing map, µCC is the mean slope of the lattice in which the pairing maps take their values
(in the source it is the dual of a fixed lattice), in the normalisation for squared norms, as in Part
II, §6.7, and F0 is as in Part II, §6.7. After division by |T | m b δ/2, where δ = |T2 |/n, the term
C |T | a La gives the error x/m of (M), the summand with B0 gives m, and the summand with Nc
gives k/n2 (each up to a factor xO(ε) ); the fourth error, xm/k, does not come from this inequality.
79
III.5 Divisibility at inert primes
This section concerns the lower bound for the exponent of an inert prime: inequality (I.19) of §I.4.3.
It proves the local estimate behind that bound, for power series, under one hypothesis of integrality,
called (B) below. In this section the prime is called ℓ, as in the Lean files; it is the inert prime p of
Part I. Several letters used here have other meanings elsewhere in the paper, among them β, λ, ε,
a, p, d, x, N , D, q, k, m, n, f , c, σ, e and ψ; in this section each has the meaning given where it is
introduced.
Proved outright. The valuation of the coefficients of the inverse of the height-two logarithm, in
several variables (Lemma 1); the integrality lemma (Lemma 2); the optimisation for the top row
(Lemma 6), which produces the sum h odd (A/ℓh − wh )+ ; weak duality for the linear program of
P
§5; and the closed form of that program for orbits of length two.
Proved conditionally on hypothesis (B). Lemmas 3 and 4: the passage from (B) to a bound for
a series β in the exponent, with the gain from 1/(ℓ2 − 1) to 1/(ℓ − 1). Proposition 5: the bound
for one Taylor coefficient of a function U · exp(±β), with weights on the variables. Theorem J, for
orbits of length two: the bound for a product of minors applied to such a function.
Conditional on (B) and on strong duality, cited from the source. Theorem J for orbits of
length four or more. The cited statement is the equality of a linear program with its dual (Lean:
strongDuality). This case was not tested end to end.
Taken from the source without proof. The transcription of the top-level inequality LinkG_at (§1
below); that the theta kernels of the construction have the form U · exp(±β) and satisfy (B); the
dictionary between sites and places; the frames and labels that Theorem J requires; the passage
from Taylor coefficients (jets) to lattices and to the inequality that the final theorem uses; and the
remaining primes. These are listed at the end.
1. The statement in the source
The top-level inequality. The final theorem depends on the Lean statement LinkG, of the
file GlobalTheoremInputs.lean. It does so through the theorems closed_linkG_of_duty and
linkG_concrete5. Its third clause is the inequality of Part II, §6.4, last paragraph. We state it.
The scale is x > 1, 0 < ε < 1/52, b = exp(xε ), a = xb, X ′ = exp(xε/3 ), b′ = a/X ′ , w′′ = b′ /(xn),
X ′′ = a/w′′ , p0 = k 2 x26 , δ1 = x−ε/8 . Here m (the number of lifts), k and t0 are integer parameters
of the construction (the fields m, k, t of D.R); n is a field of the parent structure (ord); and f = 2n
and Nm = k + m(t0 + 1) are defined quantities, not fields. T is a finite index set (D.host.TT).
The following are given data (fields of the structure SizeData2 ): integers at for t ∈ T with
a ≤ at ≤ a + a/(kx13 ); real numbers gain(j, w) and τ (w) for 1 ≤ j ≤ m and w a place of the
coefficient field; real numbers wPart(j, t, p, i) ≥ 1; subsets Tw ⊂ T ; and real numbers Sc, η, ΛG .
Put a
t
V (j, t, p) = wPart(j, (III.76)
X
− t, p, i) .
i odd
pi +
i<⌊logp at ⌋+2
This is the Lean’s vonMangoldtClassSum. For constants C and bo put
flG (p) = C f Nm
2
logp a+C f (logp a)2 , loG (p) = C (Sc+k +η) logp (2a)+pr(p)+or(p), (III.77)
80
where pr(p) and or(p) vanish unless p is of the class inertClass (defined below), and for such p,
with rG = ⌈8(4 + 8ε)/ε⌉ + 1 and ep = ⌈logp (k 2 x78 )⌉ + 1,
C rG b log x,
′ p < p0 ,
pr(p) = C b′ /x + η + Sc + (w′′ + t0 ) δ1−1 log x + bo , p0 ≤ p < X ′ ,
0, p ≥ X ′, (III.78)
Λ a
(
G a p−ep , p < p0
or(p) = C log x + (p < X ′ ), or(p) = 0 (p ≥ X ′ ).
pk a/p2 , p ≥ p0
In the definition of pr(p) the three cases are taken in this order, and the first that applies is used,
as in the Lean (an if ... else if ... else). This matters only when X ′ ≤ p0 .
The third clause of LinkG_at says: for every place w whose residue characteristic p satisfies
p ≤ X ′′ and is of class signal or inertClass, one has 0 ≤ gain(j, w) ≤ 4 |T | a/p for 1 ≤ j ≤ m,
and m m
X X
V (j, t, p) − loG (p) − m flG (p) ≤ gain(j, w) − τ (w). (III.79)
X
j=1 t∈Tw j=1
The statement LinkG asserts this, with four further clauses, for all large x, with a constant C and
with bo = bδ(x) √
for some function δ(x) → 0. (The classes: a prime not dividing the conductor, inert
in the field Q( −d) of Part I (this d is the prime of Part I; everywhere else in this section d is
the length of an orbit), is of class signal or inertClass according to the value +1 or −1 of one
further sign; DataHost.lean.)
What this statement is. The inequality (III.79) is an inequality between the numbers gain, τ ,
wPart. In the Lean it is not a statement about valuations. A separate field of the family (fills, of
type ParametersFill) identifies these numbers with explicit formulas (gainP, tauP), and the proof
of (III.79) (linkG_at_core) rewrites by ParametersFill and concludes by linear arithmetic from
a hypothesis AuxBlock8.Conclusion1, which already is the inequality for the formulas, together
with Conclusion0, lo_le_loG and fl_le_flG. The calculation with the formulas is in stmtF. The
statements which identify gain with a valuation of theta kernels are other hypotheses and other
theorems of the development. We do not claim more than this about (III.79). The Lean theorem in
which a valuation of Taylor coefficients is bounded below by a sum of the shape (III.76) is ThmAOp
(StatementI.lean, proved in ThmAOp.lean); the final theorem depends on its proof thmAOp: this
is confirmed by the dependency computation (see §6 (iii)). The rest of this section is about that
theorem.
One function, several places. Let K be a field of characteristic zero with a non-archimedean
valuation v, normalised by v(ℓ) = 1. Let d ≥ 2 be even. The elements of Z/d are called sites; they
stand for the places in one orbit of the Frobenius of ℓ, the site s + 1 being reached from s by the
Frobenius followed by complex conjugation (this identification is made in the source and is assumed
here). Let N0 , N1 ≥ 1. The variables are Zs,q,c and Xs,i,c , for s a site, q, i < N0 , c < N1 . A frame is
a family ξ(s, i, q) ∈ K with v ≥ 0; it gives the substitution Zs,q,c = i ξ(s, i, q)Xs,i,c . The function
P
is X
Θ = U g(Z) · exp ± β (Zs,q,c )s , (III.80)
q,c
where U is a power series in the variables Z with coefficients of valuation ≥ 0, where g is the inverse
of the logarithm (III.84) below, taken for the following data (Lean: entLog): the index set Y of
(III.84) is the set of the variables Zs,q,c , and π(s, q, c) = (s − 2, q, c) (for d = 2 this π is the identity,
and the logarithm is the one-variable logarithm in each variable); and where β is a power series
81
in d variables, one for each site, subject to the hypothesis (B) of §3. A shape is, for each site s,
a list of column lengths len(s, 1), . . . , len(s, ν1 (s)) in [1, N0 ]. Put ν2 (s) = #{c : len(s, c) ≥ 2} and
size(s) = c len(s, c). Let TT be a set of sites (the top sites; the Lean uses the letters TT both for
P
this set and for the index set T above). The polynomial is P = s c det Xs,i,Is,c (k) i,kdtop (s)
where ā ≥ 1 is an integer with size(s) ≤ ā for all s, and F (D) ≥ 0 is the value of a linear program
defined in §5. ThmAOp asserts v P (∂)Θ(0) ≥ B under hypotheses on the frame stated in Theorem J;
here P (∂)Θ(0) = κ Pκ κ! Θκ as in (6.2) of Part II, with Θ expanded in the variables X.
P
One top site. If TT = {s0 } then dtop (s0 ) = ∞, because d is even. With A = ν1 (s0 ) and
wh = ν1 (s0 + h) the bound reads
X A
v P (∂)Θ(0) − γW (s0 ) + F (D) − dN0 N1 (logℓ ā + 1). (III.83)
≥ h
− wh
h odd
ℓ +
This is (I.19): A is the length of the first row at the chosen place, and wh is the number of minors
at the place reached by the hth power of the Frobenius followed by conjugation.
2. Weights, and the inverse of the logarithm
Weights. Let f = n fn X n be a power series in finitely many variables Xi , with coefficients in K.
P
Let r = (ri ) be real numbers ≥ 0 and w a real number. Write n · r = i ni ri and |n| = i ni . We
P P
say that f has weight w at radii r, and write f ∈ W(r, w), if v(fn ) ≥ w − n · r for all n. This says
that the Gauss norm of f on the polydisc v(Xi ) ≥ ri is at most ℓ−w . (Lean: WtGe.) The following
rules are immediate from v(x + y) ≥ min(v(x), v(y)).
(W1) If f ∈ W(r, w) and h ∈ W(r, w′ ) then f h ∈ W(r, w + w′ ). If r ≤ r′ then W(r, w) ⊂ W(r′ , w).
(W2) If f ∈ W(r, w) and the series ai have no constant term and ai ∈ W(r′ , ri ), then f (a) ∈
W(r′ , w). (A monomial fn an lies in W(r′ , w − n · r + n · r).)
(W3) If Y has no constant term and Y ∈ W(r, ℓ−1
1
), then exp(Y ) − 1 and log(1 + Y ) lie in
W(r, ℓ−1 ).
1
(W4) The coefficient of X κ in f h, in exp(Y ) and in log(1 + Y ) depends only on the coefficients of
X n with n ≤ κ; so (W1) and (W3) hold with “for all n” replaced by “for all n ≤ κ”.
82
For (W3): Y k ∈ W(r, ℓ−1 k
), and v(k) ≤ v(k!) = (k − σℓ (k))/(ℓ − 1) ≤ (k − 1)/(ℓ − 1), where
σℓ (k) ≥ 1 is the sum of the digits of k in base ℓ. So Y k /k! and Y k /k lie in W(r, ℓ−1
1
). (Lean:
MulWtLe, RadMono, SubstWt, LogStep. The Lean statement ExpWtLe is not (W3): it gives weight
0 for the exponential of a series, truncated at κ, which is what Proposition 5 uses.)
Lemma 1 (inverse of the logarithm). Let Y be a finite set and π : Y → Y a map. For
y ∈ Y put
2e
Ly (X) = ℓ−e Xπℓ e (y) . (III.84)
X
e≥0
There is a unique family g = (gy ) of power series without constant term with Ly (g) = Xy for all y.
It satisfies gy ∈ W(ρ, ρ), where the radius is ρ for every variable:
v coefficient of X n in gy ≥ − |n| − 1 ρ. (III.85)
2e
Proof. The equation is gy = Xy − e≥1 ℓ−e gπℓe (y) . It determines the coefficients by induction on |n|,
P
since the terms with e ≥ 1 involve only coefficients of lower degree. The bound holds for |n| = 1.
Let |n| ≥ 2 and r = ℓ2e , e ≥ 1. The coefficient of X n in an rth power is a sum of products of r
coefficients, of degrees |n1 |, . . . , |nr | ≥ 1 with sum |n|. By induction such a product has valuation
at least −(|n| − r)ρ. Division by ℓe costs e, and e ≤ 1 + ℓ2 + · · · + ℓ2(e−1) = (r − 1)ρ.
This is Part II, §2, in several variables. (Lean: GenLogWeight, generalized_log_weight; the
series (III.84) is genLog. Also g(L(X)) = X: GenLogInverse.) In one variable (π the identity) a
computation confirms (III.85) for ℓ = 3, 5, 7 and all degrees up to 801, 1201, 1201: the minimum of
v(gn ) + (n − 1)ρ is 0, attained exactly at n = ℓ2e . The bound cannot be improved.
3. From integrality to a bound for β
In (III.80) the factor U (g(Z)) lies in W(ρ, 0), by Lemma 1 and (W2). After multiplication by κ!
this gives the exponent |κ| ( ℓ−1
1
− ρ) = |κ|γ of §I.3. For the exponential factor to do as well, by
(W3) the series β should have weight ℓ−1 1
at radii ρ. Integrality gives directly only the weight ρ.
The gain from ρ to ℓ−1 is the content of this paragraph. It uses that the degrees of the monomials
1
are divisible by ℓ + 1. The proof given below also uses that the coefficients lie in a field with value
group Z (an unramified field); the source carries this hypothesis (Unram, in BHyp), and it is kept as
(B1). The statements do not need it: see the Remarks after Theorem J.
Hypothesis (B). Let K0 ⊂ K be a subfield with v(K0× ) = Z. Let ζ ∈ K be a primitive root
of unity of order ℓd − 1. For a site s, represented by 0 ≤ s < d, put lw(s) = (−1)s ℓs (the letters
are those of the Lean). The series β, in variables Xs (s ∈ Z/d), satisfies: (B1) its coefficients lie in
K0 ; (B2) β (ζ lw(s) Xs )s = β(X); (B3) every monomial of β has degree ≥ 1 in the variables at even
sites, degree ≥ 1 in the variables at odd sites, and total degree ≥ 3; (B4) for the two series
∆1 β = β(xev + yev , xod ) − β(xev , xod ) − β(yev , xod ),
(III.86)
∆2 β = β(xev , xod + yod ) − β(xev , xod ) − β(xev , yod ),
in which the variables at the even (resp. odd) sites are doubled, the series exp(∆i β) ◦ L has all
coefficients of valuation ≥ 0. Here L is (III.84) for the doubled set of variables, with π the shift
s 7→ s − 2 of the site. (Lean: BHyp; ∆1 , ∆2 are de1, de2; L is letterLog.) Hypothesis (B4) says
that exp(∆i β) becomes integral when written in the coordinates T = g(X) of the formal group. It
is assumed here, from the source.
Lemma 2 (quantization). Let c ∈ K0 and let N be an integer divisible by ℓ + 1. If
1
v(c) ≥ ρ − N ρ, then v(c) ≥ ℓ−1 − N ρ.
83
Proof. Let c ̸= 0, k = v(c) ∈ Z and N = (ℓ + 1)m. Multiply k ≥ ρ − (ℓ + 1)mρ by ℓ2 − 1:
(ℓ + 1) k(ℓ − 1) + m ≥ 1. (III.87)
The number z = k(ℓ − 1) + m is an integer and (ℓ + 1)z ≥ 1; so z ≥ 1. Divide z ≥ 1 by ℓ − 1:
1
k ≥ ℓ−1 m
− ℓ−1 = ℓ−1
1
− N ρ.
Here k is the valuation of the coefficient and m is its degree divided by ℓ+1. The hypothesis loses the
factor ℓ+1 because N ρ = m/(ℓ−1) has denominator ℓ−1 only. (Lean: Quant, weightQuantization;
the step (ℓ + 1)z ≥ 1 ⇒ z ≥ 1 is at lines 83–91 of OneLetterRadiusIdentity.lean. The Lean
statement Quant has as hypothesis the congruence of the torus, that ℓd − 1 divides x nx lw(sx )
P
in the notation of the proof of Lemma 3; it does not have the hypothesis that ℓ + 1 divides N .)
Without the divisibility the conclusion fails: for N = 1 the value k = 0 is allowed and ℓ−11
− ρ > 0.
Lemma 3. Assume (B). Then ∆1 β and ∆2 β lie in W(ρ, ℓ−1
1
).
Proof. Let ∆ be one of them and E = exp(∆) ◦ L. By (B3) ∆ has no constant term, so E
has constant term 1. Since L ◦ g is the identity, exp(∆) − 1 = (E − 1) ◦ g. By (B4), Lemma 1
and (W2), a monomial of degree j ≥ 1 of E − 1 contributes a series in W(ρ, jρ) ⊂ W(ρ, ρ). So
exp(∆) − 1 ∈ W(ρ, ρ). Its coefficients lie in K0 , by (B1). By (B2), ∆ and exp(∆) are unchanged
when each variable at the site s is multiplied by ζ lw(s) . So a monomial X n with nonzero coefficient
has nx lw(sx ) ≡ 0 modulo ℓd − 1, where sx is the site of the variable x. Now ℓ + 1 divides ℓd − 1,
P
because d is even, and lw(s) ≡ (−1)s (−1)s = 1 modulo ℓ + 1. Hence ℓ + 1 divides |n|. Lemma 2
gives exp(∆) − 1 ∈ W(ρ, ℓ−1
1
), and (W3) gives the same for ∆ = log(1 + (exp(∆) − 1)).
(Lean: asmS_diff in ConstantCoeffEvenDerivativeEqZero.lean, with ExpDeInv for the divisi-
bility.) A computation tests the chain of Lemmas 1–3 in one variable: for Y = φ(g(X)) with φ
an integral series in T ℓ+1 , and ∆ = log(1 + Y ), the minimum over n ≤ 240 of v(∆n ) − ( ℓ−1
1
− nρ)
is 0 for ℓ = 3, 5, 7, attained at n = ℓ + 1. For φ(T ) = T the bound fails (the minimum is −11/8,
−29/24, −55/48).
Lemma 4 (from the differences to β). Assume (B). Call a monomial exceptional if it is
α α′
Xsℓ Xsℓ′ with s even, s′ odd, (α, α′ ) ̸= (0, 0) and s + α = s′ + α′ in Z/d. Then every coefficient cn
of β satisfies
1
v(cn ) ≥ − |n|ρ − ex(n), ex(n) = 1 if X n is exceptional, ex(n) = 0 otherwise. (III.88)
ℓ−1
′
Proof. Write X n = Xev u X u with multi-indices u, u′ on the even and the odd sites. By (III.86),
od ′
the coefficient of xev yev xod
j u−j u in ∆ β is c Q us , for every j with 0 ≤ j ≤ u, j = ̸ 0, j ≠ u. By
1 n s js
Lemma 3, v(cn ) ≥ ℓ−1 − |n|ρ − µ(u), where µ(u) is the least valuation of such a product of binomial
1
coefficients; this holds if |u| ≥ 2. If u involves two sites, take j equal to u at one site and 0 elsewhere:
µ(u) = 0. Let u be supported at one site, with us not a power of ℓ. If us has at least two nonzero
digits in base ℓ, let j be its leading digit times the corresponding power of ℓ; if us is ℓk times one
digit between 2 and ℓ − 1, let j = ℓ . By Lucas’s congruence j is prime to ℓ, so µ(u) = 0. If
k us
us = ℓα with α ≥ 1 then µ(u) = 1, attained at j = ℓα−1 . The same holds for u′ and ∆2 β. By (B3),
|u| ≥ 2 or |u′ | ≥ 2. So ex(n) = 0 unless u and u′ are each a power of ℓ at one site, not both equal
α α′ ′ ′
to 1. In that case X n = Xsℓ Xsℓ′ , and if cn ̸= 0 then (B2) gives ℓs+α ≡ ℓs +α modulo ℓd − 1, that
is s + α = s′ + α′ in Z/d.
(Lean: the statement is LemmaS, proved as lemmaS; the step of Lemma 4 is BinomExtract. The
proof of Lemma 4 above is ours; the Lean proof of BinomExtract was not read. The facts on
binomial coefficients were checked for us ≤ 1500.) So β has the weight ℓ−1
1
that is needed, except
84
at the exceptional monomials, where one unit is missing. The exceptional monomials pair a site s
with the site s′ = s + (α − α′ ), at an odd distance. They are the reason for the weights of the next
paragraph.
4. One Taylor coefficient, with weights on the variables
Proposition 5. Assume (B). Let Xi (i ∈ ι) be finitely many variables. Let Zy = i My,i Xi be
P
linear forms with v(My,i ) ≥ 0, one for each variable Zy of (III.80), and let Θ be (III.80) written in
the variables X. Let B = q,c β (Zs,q,c )s = n Bn X n . Let κ be a multi-index and let ri ≥ 0 be
P P
real numbers such that
′ α α′ ′
ℓα ri +ℓα rj ≥ 1 whenever Xiℓ Xjℓ divides X κ and v Bℓα ei +ℓα′ ej < ℓ−1
1
−(ℓα +ℓα )ρ. (III.89)
α α′
The case i = j is included; the monomial is then Xiℓ +ℓ . Then, with λ(k) = ⌊logℓ k⌋ + 1 for k ≥ 1
and λ(0) = 0,
v κ! Θκ ≥ κi (γ − ri ) − λ(κi ). (III.90)
X X
i i
′
Proof. First, Bn has valuation ≥ ℓ−1 1
− |n|ρ unless n = ℓα ei + ℓα ej for some i, j, α, α′ (i = j
allowed); and ≥ ℓ−1 − 1 − |n|ρ always. Indeed, by (III.88) and (W2) only the exceptional monomials
1
α α′ ℓα = P M ℓα X ℓα + ℓR, with
c Xsℓ Xsℓ′ of β matter. For a linear form Z = i Mi Xi one has Z
P
i i i
α α′
R a polynomial with coefficients of valuation ≥ 0. So all monomials of Z ℓ Z ′ℓ other than the
α α′
Xiℓ Xjℓ carry a factor ℓ. Now put ri′ = ρ + ri . For n ≤ κ we get v(Bn ) ≥ ℓ−1 1
− n · r′ : this is clear
′ ′
if v(Bn ) ≥ ℓ−1
1
− |n|ρ, and otherwise n = ℓα ei + ℓα ej and n · r′ = |n|ρ + ℓα ri + ℓα rj ≥ |n|ρ + 1 by
(III.89). By (W3) and (W4) the coefficients of X n in exp(±B), n ≤ κ, have valuation ≥ −n · r′ . The
factor U (g(Z)) lies in W(ρ, 0) ⊂ W(r′ , 0) by Lemma 1 and (W2), (W1). By (W1), v(Θκ ) ≥ −κ · r′ .
Finally v(k!) = (k − σℓ (k))/(ℓ − 1) ≥ ℓ−1
k
− λ(k), and ℓ−1
1
− ρ = γ.
(Lean: LemmaR, with PureCross for the first step and FactVal for the last; the Lean loss is
|ι| (logℓ |κ| + 1). LemmaR has one more hypothesis than Proposition 5: each variable Xi belongs
to one site. Proposition 5 does not need it. The statement LemmaR is in the archive, proved as
lemmaR_of, but the dependency computation shows that LemmaR is not among the declarations the
final theorem depends on. Proposition 5 is proved here independently. By occurrences of names
the chain uses LemmaRA.) Each variable pays its radius ri out of the gain γ. The radii are needed
only where an exceptional monomial is present with a coefficient that is not divisible by one more
power of ℓ.
One variable. Take one variable X and Θ = U (g(X)) exp(B(X)), where v(Bn ) ≥ ℓ−1 1
− nρ
′
for n ∈/ S and v(Bn ) ≥ ℓ−1 − 1 − nρ for n ∈ S = {ℓ + ℓ : α + α odd}. The second half of the
1 α α ′
proof of Proposition 5 applies to such a series: what is needed is nr ≥ 1 for n ∈ S, n ≤ κ. The
least element of S is ℓ + 1, so r = ℓ+1
1
will do, and γ − ℓ+1
1
= ρ. Thus
v κ! Θκ ≥ κγ − λ(κ) if B has no terms in S, v κ! Θκ ≥ κρ − λ(κ) in general. (III.91)
In the first case the factorial contributes κ/(ℓ − 1) = κ(γ + ρ) and the function costs κρ. In the
second the exceptional terms can cost a further κ/(ℓ + 1). (Lean, for the second: OneLetterRadius.
The Lean statement OneLetterR is for several variables through an integral matrix, with exponent
(sgn/2) β(Z, ep · Z).) A computation tests (III.91) with exact rational arithmetic for ℓ = 3, 5, 7
P
and all κ ≤ 240, with U a random series with integer coefficients that are units at ℓ, and Bn of the
least valuation allowed. The minimum over κ of the left side minus the right side is 14 , 16 , 18 in the
85
first case and 34 , 65 , 87 in the second. In the second case the first bound fails: at κ = 240 the left side
is below κγ − λ(κ) by 57, 37, 28. So the cost κ/(ℓ + 1) is real in this model (240/(ℓ + 1) = 60, 40, 30).
5. The choice of the weights
In Proposition 5 the bound is i κi (γ − ri ), and the radii are free subject to (III.89). The exponent
P
(III.82) comes from a choice of radii. Two elementary optimisations are involved.
Lemma 6 (the top radius). Let A ≥ 0, let cj ≥ 0 for odd j ≥ 1, and let J ≥ 1 be odd. There
is ω ≥ ℓ−J with
A A
Aω + cj (1 − ℓj ω)+ ≤ min cj , j + J . (III.92)
X X
j odd j odd
ℓ ℓ
jJ
Proof. Let j0 be the least odd j < J with cj ≥ A/ℓj , and j0 = J if there is none. Take ω = ℓ−j0 .
On the left of (III.92) the terms with j ≥ j0 vanish, and each term with j < j0 is at most
cj = min(cj , A/ℓj ). The term Aω = A/ℓj0 equals min(cj0 , A/ℓj0 ) if j0 < J, and is the last term on
the right if j0 = J. For (III.93) use γA = j odd A/ℓj and x − min(c, x) = (x − c)+ .
P
(The archive contains this statement as TopRadius, proved as top_radius_holds with the same
choice of ω. The dependency computation shows that neither is among the declarations the final
theorem depends on. Lemma 6 is proved here independently. The declaration the final theorem
uses for the top-row term has been identified by reading the proof terms: it is asmA_top, which
the proof of thmAOp reaches through thmAJetA and rowRadiusAssembly; the bound for one Taylor
coefficient on that route is lemma_R_log_loss. The lemma was checked on 20000 random cases.)
Here ω is the radius of the variables of a top row of length A, and cj bounds the degree in those
variables at the site s + j which are linked to the top row by an exceptional monomial. Such a
variable needs the radius (1 − ℓj ω)+ , by (III.89). The right side of (III.93), with both of its sums,
is top(s) of (III.82) for A = ν1 (s), cj = cap(s + j) and J = dtop (s): the second sum of (III.93),
over odd j > J, is the second sum in the definition of top(s). The lemma asserts that a suitable ω
exists, not that (III.93) holds for every ω ≥ ℓ−J . If dtop (s) = ∞ and only the j with ℓj ≤ A occur
on the left, the same choice (with ω = 0 if there is no j0 ) gives j odd (A/ℓj − cj )+ on the right.
P
The linear program. Let D : Z/d → R≥0 . A flow is a family ψ(s, i) ≥ 0, for sites s and odd
i < d, with
ℓi ψ(s, i) + for all s. (III.94)
X X
ψ(s − i, i) ≤ D(s)
i i
Let maxFlow(D) be the supremum of s,i ψ(s, i) over flows, and F (D) = γ s D(s) − maxFlow(D).
P P
A dual is a family 0 ≤ σ(s) ≤ 1 with ℓi σ(s) + σ(s + i) ≥ 1 for all s and all odd i < d. (Lean:
IsFlow, Fnum, IsDual; and lpF_eq_Fnum for the form used elsewhere.)
Weak duality, proved here. For a flow ψ and a dual σ,
X
ψ(s, i) ℓi σ(s)+σ(s+i) =
X X X X X
ℓi ψ(s, i)+
ψ(s, i) ≤ σ(s) ψ(s−i, i) ≤ σ(s)D(s).
s,i s,i s i i s
(III.95)
So s D(s)(γ − σ(s)) ≤ F (D) for every dual σ.
P
86
Strong duality, cited. For ℓ ≥ 3 and d even there is a dual σ with s D(s)(γ − σ(s)) = F (D),
P
and with ℓi σ(s) + σ(s + i) ≥ 1 for all odd i ≥ 1 (Lean: strongDuality, of type LemmaFa). For a
dual with the last property, if s + α = s′ + α′ and α + α′ is odd, then
′
ℓα σ(s) + ℓα σ(s′ ) ≥ 1 : (III.96)
′ ′
if α > α′ this is ℓα times ℓα−α σ(s) + σ(s + (α − α′ )) ≥ 1; if α′ > α exchange the roles of (s, α)
and (s′ , α′ ) (and α ≠ α′ , since α + α′ is odd). So such a σ, used as radius for all variables at the
site s, satisfies (III.89).
Orbits of length two, proved here. Let d = 2. Then
ρ (D0 + D1 ), D1 /ℓ ≤ D0 ≤ ℓD1 ,
D0 + D 1
maxFlow(D) = min D0 , D1 , , F (D) = γD0 − (1 − γ)D1 , ℓD1 ≤ D0 ,
ℓ+1
γD − (1 − γ)D , ℓD ≤ D .
1 0 0 1
(III.97)
−D1 ℓD1 −D0
Indeed, in the first case ψ = ℓDℓ20 −1 , ℓ2 −1 is a flow and σ = ( ℓ+1 1 1
, ℓ+1 ) a dual, both of value
ℓ+1 ; in the second ψ = (D1 , 0) and σ = (0, 1), of value D1 ; the third is symmetric. By (III.95)
D0 +D1
these are optimal. These duals satisfy the inequalities for all odd i. (Checked exactly on 20000
cases, and strong duality numerically for d = 2, 4, 6.) Note that F (D) is not h odd (D0 /ℓh − D1 )+ :
P
for ℓ = 5 and D = (1000, 10) the two are 200.42 and 190; for D = (1000, 250) they are 52.08 and
0. One has F (D) ≥ h odd (D0 /ℓh − D1 )+ , with equality only for D1 = 0. In the source the sum
P
of the shape (III.76) is the term top(s). Here it comes from Lemma 6; on the route of the final
theorem it is the statement asmA_top. The linear program governs the rows other than the top
rows.
Theorem J. Assume (B), with ℓ an odd prime and d ≥ 2 even. Let a shape, a set TT , the
polynomial P and the integer ā be as in §1. Call a row (s, i) ordinary if it is not a top row. Call a
top site s0 , an odd integer j and an ordinary row (s, i) linked if s = s0 + j, and ℓj ≤ ν1 (s0 ) if j > 0,
and ℓ−j ≤ size(s) if j < 0. Let spec be an injective map from a set of ordinary rows to pairs (s0 , j),
such that s0 , j, (s, i) are linked whenever spec(s, i) = (s0 , j); the pair is the label of the row. Here
j is an integer, not a class modulo d: (s0 , j) and (s0 , j + d) are different labels for the same site.
Suppose that whenever s0 , j, (s, i) are linked and (s0 , j) is not the label of (s, i), the frame satisfies
−j
X j
X
v ξ(s0 , 0, q)ℓ ξ(s, i, q) ≥ 1 (j > 0), v ξ(s0 , 0, q) ξ(s, i, q)ℓ ≥1 (j < 0). (III.98)
q q
Then v P (∂)Θ(0) ≥ B ′ , where B ′ = s∈TT top(s) − γW (s) + F (D) − maxκ y λ(κy ), the
P P
maximum being over the monomials X κ of P and the sum over the variables y. This is the loss
that the proof gives. Since y λ(κy ) ≤ dN0 N1 (logℓ ā + 1), one has B ′ ≥ B with B as in (III.82),
P
which is the exponent of the Lean statement. For d = 2 this is proved here; for d ≥ 4 the proof uses
the cited strong duality.
Proof. Since P has integer coefficients, it suffices to prove v(κ!Θκ ) ≥ B ′ for each monomial X κ
of P . A determinant is linear in each row. So the degree of X κ in the variables of the row (s, i)
is the number of c with len(s, c) > i. It is ν1 (s) for i = 0, and at most ν2 (s) for i ≥ 1. Hence the
degree is ν1 (s) in a top row, at most cap(s) in an ordinary row, and D(s) in all ordinary rows at s
together; and each κy ≤ ā.
Radii. Let σ be an optimal dual for D, as above. For s ∈ TT let ω(s) be given by Lemma 6 with
A = ν1 (s), cj = cap(s + j) and J = dtop (s) (and by the remark after it if dtop (s) = ∞). Give the
variables of a top row the radius ω(s). Give the variables of an ordinary row (s, i) the radius σ(s)
87
if the row has no label; max(σ(s), 1 − ℓj ω(s0 )) if its label is (s0 , j) with j > 0; and max(σ(s), ℓ−|j| )
if j < 0. (Lean: rowRad.)
′
The condition (III.89). With Zs,q,c = ℓα ℓα
i ξ(s, i, q)Xs,i,c , a monomial Xy Xy ′ , y = (s, i, c),
P
α α′
y ′ = (s′ , i′ , c′ ), has in B the coefficient cn ξ(s, i, q)ℓ ξ(s′ , i′ , q)ℓ if c = c′ , and 0 otherwise, where
P
q
α α′
cn is the coefficient of Xsℓ Xsℓ′ in β (it is 0 by (B3) if s and s′ have the same parity). By (III.88)
the condition concerns only the case s + α = s′ + α′ , α + α′ odd, and only if the sum over q has
valuation < 1. If both rows are ordinary, their radii are at least σ(s), σ(s′ ), and (III.96) applies.
′
If both are top rows, say with α > α′ , then dtop (s) ≤ α − α′ , so ℓα ω(s) ≥ ℓα ℓ−(α−α ) ≥ 1. There
remains the case of a top row and an ordinary row; by symmetry let (s, 0) be the top row and
(s′ , i′ ) the ordinary row, and j = α − α′ , so s′ = s + j. Let j > 0. Since the monomial divides
′ ′
X κ , ℓj ≤ ℓα ≤ ν1 (s). If the label of (s′ , i′ ) is (s, j), then ℓα ω(s) + ℓα (1 − ℓj ω(s)) = ℓα ≥ 1. If
j ℓα′ j ℓ α′
not, (III.98) holds, and q ξ(s, 0, q)ℓ ξ(s′ , i′ , q) ≡ q ξ(s, 0, q) ξ(s , i , q)
ℓ ′ ′ modulo ℓ has
P P
′
valuation ≥ 1: the condition is void. Let j < 0. Then ℓ ≤ ℓ ≤ size(s ). If the label is (s, j) then
|j| α ′
′
ℓα ℓ−|j| = ℓα ≥ 1. If not, (III.98) applies in the same way.
The sum. By Proposition 5 it remains to show y κy (γ − ry ) ≥ s∈TT (top(s) − γW (s)) + F (D).
P P
Use max(σ, u) ≤ σ + u+ . The terms with σ give s D(s)(γ − σ(s)) = F (D). The labels (s0 , j) with
P
j > 0 cost at most cap(s0 + j)(1 − ℓj ω(s0 ))+ each, because spec is injective and one row has degree
at most cap; with the top row this gives top(s0 ) by (III.93). The labels (s0 , j) with j < 0 cost at
most cap(s0 − |j|) ℓ−|j| ≤ W (s0 ) ℓ−|j| , and j odd ℓ−j = γ. The loss in (III.90) is y λ(κy ), as in
P P
B ′ . It is at most dN0 N1 (logℓ ā + 1), because there are dN0 N1 variables and each κy ≤ ā.
(Lean: ThmAJetA for one κ, ThmAAssembly for the sum, ThmARadii for the condition, ThmAOp for
the result. The Lean statement ThmAOp is more general: it allows an integral linear operator on the
ordinary rows before the minors are formed. That generality is not proved here. The inequality for
the sum was tested on 3000 random shapes.)
Remarks. The following are observations of the test of this section. They are recorded without
proof, and nothing above depends on them.
(1) The field K0 . The field K0 , hypothesis (B1) and Lemma 2 are not used in Lemmas 3–4,
Proposition 5 and Theorem J, except in the proof of Lemma 3 as written. The reason is this. In
that proof the series E = exp(∆) ◦ L is invariant under the torus, because L commutes with it.
Hence every monomial of E − 1 has degree j divisible by ℓ + 1, so j ≥ ℓ + 1, and it contributes a
series in W(ρ, jρ) ⊂ W(ρ, (ℓ + 1)ρ) = W(ρ, ℓ−11
). This is the conclusion of Lemma 3. The Lean
statement BinomExtract has no K0 either.
(2) Strong duality. The strong duality cited from the source is the duality theorem of linear
programming together with two elementary reductions. First, replacing σ by min(σ, 1) keeps the
constraints and does not increase s σ(s)D(s). Second, for odd i ≥ d write i = i′ + kd with i′ odd
P
′
and 1 ≤ i′ < d; then ℓi σ(s) + σ(s + i) ≥ ℓi σ(s) + σ(s + i′ ) ≥ 1, so every dual satisfies the inequalities
for all odd i ≥ 1. The equality was confirmed in every instance computed (d = 2, . . . , 10).
(3) The prime 2. That ℓ is odd is not used in the proofs of this section, and Lemma 1 holds
for ℓ = 2. No series with (B) was constructed for ℓ = 2, so Lemmas 3–4 and Theorem J were not
tested for ℓ = 2.
An example. The test of this section constructed an explicit series β for orbits of length two
2
(d = 2). Let ℓ ≡ 3 mod 4. Let Ec be the curve Y 2 = X 3 − cX, with c in Z[i] and c(ℓ −1)/4 ≡ −1
modulo ℓ (c = −i for ℓ = 3 and 11, c = −1 − i for ℓ = 7). It is a twist of the lemniscate curve,
the curve of the lemniscate sine of §I.3. Let σc be the Weierstrass σ-function of Ec (with g2 = 4c,
88
g3 = 0), let t = −X/Y , and put m(z) = log σc (z)/t(z) and β̃(z, w) = m(z + w) − m(z) − m(w).
The series β(X0 , X1 ) is the part of β̃ that is invariant under (X0 , X1 ) 7→ (ζX0 , ζ −ℓ X1 ); the field
is K0 = Qℓ (i). Hypothesis (B4) was checked in exact arithmetic: the series exp(∆1 β) ◦ L has
coefficients integral at ℓ up to total degree 84 for ℓ = 3, 72 for ℓ = 7 and 144 for ℓ = 11. That (B)
holds in all degrees was not proved; the test has a sketch of an argument, which was not written
out as a proof. On this β, Lemmas 3 and 4 were checked up to degree 800, and equality holds in
(III.88) at exceptional monomials: there the valuation of the coefficient is exactly ℓ−1 1
− |n|ρ − 1.
So the tested coefficients attain the loss of one unit. This shows that the term ex(n) is sharp for
the finite part of the series that was verified; sharpness under (B) in all degrees would follow if (B4)
were proved for this example. Theorem J was checked on this β in exact arithmetic, for minors
of size one, with ℓ = 3, 7, 11, 19: the bound holds, both with B and with B ′ , in each of 1 041 934
comparisons, of which 147 296 have B > 0 and 266 970 have B ′ > 0. It also holds in every case
computed with minors of size two at one site. These checks use β in degrees beyond those to which
(B4) was checked.
Comparison with §I.4.3. The description there is (III.83). The weight wh is cap(s0 + h):
the number of minors at the place reached by the hth power of the Frobenius composed with
conjugation. It enters in two steps. First, an exceptional monomial of β links the top row at s0
to a row at s0 + h, by (III.88); only such a link forces a radius, by (III.89). Second, a monomial
of a product of minors has degree at most cap in one row; this is the statement of §I.4.3 that a
minor absorbs one derivative in a linked direction. Lemma 6 then balances the radius of the top
row against the radii of the linked rows, and the result is (A/ℓh − wh )+ at each odd h. The labels
with (III.98) express that at each h only one row at the place s0 + h is linked. The congruence of
Part II, §6.3 is a statement of the same kind as (III.98). The source does not use it by name at
this point: see §6 (ii).
6. Assumed here, from the source
(i) The form (III.80) of the theta kernels, and hypothesis (B). In the Lean this is the statement
Loc.AdmB, of the file ProductZerothFirstParts.lean. It says that there are a unit e and a
series β such that the kernel, in the coordinates e · L, equals U · exp(± β) ◦ L with U integral
P
(clause B0; also OrbitThetaExpansion), with coefficients in the fraction field of the Witt ring
(B1), torus invariance (B2), the condition on degrees (B3), and the integrality (B4) (B4). The
hypothesis hServ.pdv (SelectionServed) of the final theorem does not contain Loc.AdmB
literally: it gives BeOK (B1–B4 for each orbit representative) and ThetaExpansionAllPlaces.
How these give Loc.AdmB was not read. Nothing of it is proved here. The passage from
Loc.AdmB to BHyp (tieB) was not read.
(ii) The frames. Theorem J needs a labelling and (III.98). In the Lean the theorem for all
frames at once is ThmAGauss: every coefficient of the polynomial frameJet in the frame
variables has valuation ≥ B, if 2 logℓ ā + 2 ≤ N0 − 1. Its proof (thmAGauss_of) names
residuallyGenericExists: there are frames whose entries are residually independent units
(IsGeneric). Three further inputs are supplied to it as hypotheses, in thmAGauss: a trans-
fer of the hypotheses between fields (admissibleTransfer, of type TransferB), the eval-
uation of the frame polynomial (FrameJetEval, GaussPow), and the existence of labels
(labelsExist_statement, of type LabelsExist). The statement TransferB contains no
frame: it transports Unram, BHyp, the inverse g of the logarithm and the integrality of U along
a norm-preserving homomorphism of fields. The proofs of these inputs were not read. The
congruence key_minor_dvd of Part II, §6.3 is not used by name in this step: no file under
89
Formalism/GaussJetBounds mentions it. Its only use by name is in its own file; that file
is imported by Loc4ResOfLine and LiftSubstStart, in the directory CoarseMultiplicity,
which were not read.
(iii) From Taylor coefficients to the statement used. The comparison of F (D) and top(s) with the
weights of the construction (normCor7; split_bound_eleven_one, conditional on Loc.AdmB);
the factorial count StarArith and the inequality BCore, which is of the form (6.1) of
Part II; the identification of sites with places and of ν1 with the weights. Then the loss
terms loG , flG of (III.77); the lattice clauses of LinkG_at; stLoss7; placesF_min’. For
stLoss7 and placesF_min’ only the statements were read. On the route from LinkG to
the final theorem that was read, the clauses about valuations of kernels are proved for
the zero chain (linkG_without_kernels); how ThmAOp enters the final theorem was not
read. The dependency computation shows that thmAOp is among the declarations the fi-
nal theorem depends on. By occurrences of names it reaches the final theorem through
thmAGauss, thmA_of, split_bound_eleven_one and sum_over_primes_service, that is,
through SelectionServed.
(iv) The other primes. The primes of class inertClass are at least 13 for large x (thirteen_le,
which has a hypothesis that x is large and uses the assumed splitting of the small primes,
split_small). The primes of the other classes, and the primes above X ′′ , are treated by
other clauses, not discussed here.
What this section proves, and what it assumes. Proved outright: Lemma 1 (the inverse
of the logarithm), Lemma 2 (integrality), Lemma 6 (the optimisation for the top row), weak duality
(III.95), and the closed form (III.97) for orbits of length two. Proved conditionally on hypothesis
(B): Lemmas 3 and 4, Proposition 5, and Theorem J for d = 2. Conditional on (B) and on strong
duality, which is cited from the source (strongDuality): Theorem J for d ≥ 4; this case was not
tested end to end. Taken from the source without proof: the transcription of the top-level inequality
LinkG_at in §1; that the theta kernels satisfy (B) (item (i)); the dictionary between sites and
places; the frames and labels (item (ii)); the passage from jets to lattices (item (iii)); the remaining
primes (item (iv)). Theorem J is proved without the operator on the ordinary rows that the Lean
statement ThmAOp allows.
Relation to the declarations the final theorem depends on. By the dependency computa-
tion, the following declarations named or used in this section are among them: strongDuality,
weightQuantization, lemmaS_of, thmAOp, binomial_coefficient_extraction, lpF_eq_Fnum,
residuallyGenericExists, theorem_A_generic and gauss_sum_powers. The four statements
TopRadius, top_radius_holds, LemmaR and ThmABound are in the archive but are not used by the
final theorem. So Lemma 6 and the weighted bound for one Taylor coefficient (Proposition 5) are
proved here independently of the route of the final theorem, and the declarations the final theorem
uses instead are asmA_top and lemma_R_log_loss. Their statements are consistent with Lemma 6
and Proposition 5: for a finite cut-off, and for the set of all odd j < J, asmA_top is (III.93) with
top(s) moved to the other side, and its proof takes the same radius (for an infinite cut-off it is the
remark after Lemma 6; see §III.6, §4), and Proposition 5 is the stronger of the two bounds for a
coefficient.
What remains to be written. The passage from the estimate of this section to lattices is
taken up in §III.6; some of the items below are treated there, as that section states.
(i) A proof of strong duality for d ≥ 4 (in the Lean: compactness and a separation theorem); see
Remark (2) after Theorem J.
90
(ii) Theorem J with the operator on the ordinary rows, and the deduction of the statement for all
frames (ThmAGauss); the role of the congruence of Part II, §6.3.
(iii) Hypothesis (B) for the theta kernels of the construction: the height-two formal group at an
inert prime, the series β, and the integrality (B4). This is the arithmetic input of the whole
section.
(iv) The passage from Theorem J to (III.79): items (iii) and (iv) above, with the loss terms.
(v) A proof that the series β of the example after Theorem J satisfies (B) in all degrees, and an
example for d ≥ 4 with an exceptional monomial; none is known to us. For d ≥ 4 Lemmas
3–4 and Theorem J were tested only on a series without exceptional monomials and on the
combinatorial steps.
(vi) To follow, in the source, how the statement for one Taylor coefficient on the route of the final
theorem, lemma_R_log_loss, is used in place of Proposition 5. (The statements TopRadius,
top_radius_holds, LemmaR and ThmABound of the archive are not used by it.)
III.6 Divisibility: from derivatives to lattices
This section concerns the passage from the local estimate of §III.5 to the prime sum Pu,t,j of (4.1)
of Part II, the left side of the main inequality (M). In §III.5 one function at one prime is studied.
What (M) needs is a statement about lattices over a number field, at all primes at once, with
the losses added up. The section states that passage, proves its elementary parts, and names the
statements of the source that carry the arithmetic. For the prime sum it does this much: it splits
the sum and bounds several of its terms. The estimates for the eleven loss sums and their total
bound (III.138) are quoted. Theorem III.3.2 alone does not give that bound. The section does not
prove the divisibility side of (M). The divisibility of the period pairing is assumed, as statement
(Val) of §2. Also assumed are a statement on lattices, the inequality (III.79) at the primes of one
of the two classes, the estimates for the eleven loss sums with their total bound, and the slope
inequality for many maps.
Notation. The notation of §III.5, §1 is kept: the scale x, ε, b, a, b′ , X ′ , X ′′ , p0 , δ1 ; the parameters
m, k, t0 , n, f, Nm ; the set T ; the numbers at , gain, τ , wPart, V (j, t, p), loG , flG , Sc, η, and the
constant C of (III.77). So a is the number xb, as in §III.4 and §III.5, with a ≤ at . And τ (w) is the
number in (III.79); it is not the modulus τ of §III.4. The letter u is the translate of Part II, §4.2,
as in Pu,t,j . The letters introduced here are listed at the end of §1.
The letters are local. The letters of this section have the meanings given here in this section
only. Other sections of Part III use some of them differently, and some objects have another name
there. The main cases are these. τ is a weight place in §III.7. T is the level in §III.4 and §III.7,
and T (w) in §III.8 is the level prime of a word, which is Lvα here. The members α are the words
w of §III.8, where α is a constant. U is a power series in §III.5 and a space on the target side in
§III.7. γ is ℓ/(ℓ2 − 1) in §III.5 and in the paragraph on asmA_top of §4, and Euler’s constant in §6.
kaj is the number ς of §III.4, and cχd is the cd of §III.4 and §III.7. λw is written wt(y) in §III.7,
and the λw of §III.8 is another number. µ0,α is not the µ0 of §III.7 or of §III.9. Ξ is an operator in
§III.7 and §III.8. L0 is the field called K in §III.9, where L0 is another field.
Proved outright. Three lemmas on lattices over the ring of integers of a number field: the
slope of a sublattice (Lemma L1), the slope of a lattice multiplied by a fractional ideal, with
the cost of rounding (Lemma L2), and the integrality of a pairing after such a multiplication
(Lemma L3). The bound for the slope when the rank drops (Lemma L5). The identity between the
91
prime sum (4.1) and sums over single primes (Lemma L6). Three elementary comparisons of losses
(Lemmas L7 to L9). An inequality of the shape of (III.79) at the primes of class signal, for the
explicit formulas of the source and with explicit loss terms (Proposition L10); it gives (III.79) itself
when the constants satisfy the conditions (III.130). The splitting of the prime sum by classes of
primes (Proposition L11) and three of the estimates that bound its terms (Lemmas L12 to L14;
the last uses Theorem III.3.2 and one cited estimate).
Proved from two statements of the source. Theorem L4. Given the valuation statement (Val) of
§2, the glued lattice pairs integrally with the target. Given the lattice statement (Lat) of §2, its
slope exceeds that of the source lattice by the total gain, minus explicit losses.
A model, not a statement about the source. Proposition L15 shows how the terms cancel when
there is one injective map between spaces of equal dimension. That hypothesis does not hold in the
source.
Taken from the source without proof. The statement (Val), that is, the divisibility of the period
pairing of the chain of maps. In the Lean it rests on three assumptions (adm2, the hypotheses Tied
and CutUp of val2AllAtRF, and the field l1b; see §2). It is where Theorem J of §III.5 is used,
and it is conditional on hypothesis (B) there. The statement (Lat). The inequality (III.79) at the
primes of class inertClass. The estimates for the eleven sums over primes of the losses, and their
total bound (III.138). The slope inequality for many maps (Part II, §6.9–6.10). These are listed in
§7.
1. Lattices over a number field: three lemmas
Places and weights. Let L0 be a number field of degree n0 , totally complex, with ring of integers
o. (In the source L0 is the coefficient field, LL.) A place w is a nonzero prime ideal pw of o; pw is its
residue characteristic, ew its ramification index over pw , vw the valuation with vw (pw ) = 1, and ow
the localisation of o at pw , a discrete valuation ring. For an ow -submodule Λw of an L0 -vector space
and an integer z (which may be negative) we write pzw Λw for (pw ow )z Λw . As in Part II, §6.4, put
log Norm(pw )
λw = λw ew = log p, szw (σ) = 2λw ew σ (σ ∈ R). (III.99)
X
,
n0 w|p
The middle identity holds because the sum over w | p of ew times the residue degree of w is
n0 . So a gain of σ at every place above p, measured in units szw , has total 2σ log p. (Lean: lw,
SizeData2 .sz, FactEtaMean. The Lean states FactEtaMean for a field that is Galois over Q. The
identity holds for every number field, by the argument just given.)
Lattices and slopes. Let U be an L0 -vector space of finite dimension dim U ≥ 1, and fix a
Haar measure on the real vector space U ⊗Q R. A lattice in U is a finitely generated o-submodule Λ
that spans U over L0 . Its rank is rk Λ = dim U ; it is a discrete subgroup of U ⊗ R of rank n0 dim U ,
with a covolume. The slope divides by the rank. So wherever a slope is taken, in Lemmas L1, L2
and L5 and in Theorem L4, the space has positive dimension and the lattice has rank ≥ 1. Put
2 log |disc L0 |
µ(Λ) = − log covol(Λ) + cstL0 , cstL0 = − log 2 + . (III.100)
n0 rk Λ n0
This is the slope in the normalisation for squared norms: twice the slope of Part II, §6.8, up to
the constant. (Lean: muDef, with the measure given by Hermitian forms at the complex places.
In Lemmas L1–L5 only differences µ(Λ) − µ(Λ′ ) for lattices in the same space occur; they do
not depend on the measure or on cstL0 .) For a finite o-module write lgw of it for the length of
92
its localisation at w: the number of factors o/pw in a composition series. For a lattice Λ write
Λw = ow Λ ⊂ U . Here Λ always denotes a lattice; the number ΛG of (III.78) is not used in this
section.
Lemma L1 (a sublattice). Let Λ′ ⊂ Λ be lattices in U , with dim U ≥ 1. Then
′ 2 X
[Λ : Λ′ ] = Norm(pw )lgw (Λ/Λ ) , µ(Λ′ ) = µ(Λ) − λw lgw (Λ/Λ′ ). (III.101)
Y
w rk Λ w
Proof. The quotient Λ/Λ′ is a finite o-module. It is the direct sum of its localisations, and each
composition factor o/pw has Norm(pw ) elements. This gives the index. Then covol(Λ′ ) = [Λ :
Λ′ ] covol(Λ) and log Norm(pw ) = n0 λw .
(Lean: muDef_index, mu_index_sum.)
Lemma L2 (multiplication by an ideal). Let Λ be a lattice in U , with dim U ≥ 1, let zw be
integers, zero for almost all w, and put I = w p−z
Q
w , a fractional ideal.
w
(i) IΛ is a lattice with (IΛ)w = p−z
w Λw for every w, and it is the only lattice with these
w
localisations.
(ii) µ(IΛ) = µ(Λ) + 2 w λ w zw .
P
(iii) Let σw be real numbers, zero for almost all w, let zw = ⌊σw ⌋, and let Rd be a finite set of
primes containing pw for every w with σw ∈/ Z. Then
2 λw σw − 2 log p ≤ µ(IΛ) − µ(Λ) ≤ 2 (III.102)
X X X
λ w σw .
w p∈Rd w
Proof. (i) Localisation commutes with products, and Iow = p−z w ow . A lattice is determined by its
w
localisations: if y ∈ U lies in Λw for all w, the ideal {β ∈ o : βy ∈ Λ} is contained in no pw , so it is o
and y ∈ Λ. (ii) Suppose first that all zw ≤ 0, so that I ⊂ o and IΛ ⊂ Λ. The module Λw is finitely
generated and torsion-free over the discrete valuation ring ow , hence free of rank rk Λ. So the
localisation of Λ/IΛ at w is a sum of rk Λ copies of ow /p−z w ow , of length −zw rk Λ, and Lemma L1
w
−1
gives (ii). In general write I = I1 I2 with ideals I1 , I2 ⊂ o. Then I2 · (IΛ) = I1 Λ. The case just
proved, applied to (I2 , IΛ) and to (I1 , Λ), gives µ(IΛ) − 2 w λw vw (I2 ) = µ(Λ) − 2 w λw vw (I1 ),
P P
and vw (I1 ) − vw (I2 ) = −zw . (iii) By (ii) the middle term is 2 w λw zw . Now 0 ≤ σw − zw < 1, and
P
σw − zw = 0 unless pw ∈ Rd. Finally w|p λw ≤ w|p λw ew = log p.
P P
(Lean: ShiftEx, ShiftMu with its proof shift_slope_mu, RoundingCost; the ideal is shiftIdeal,
the lattice IΛ is shiftLat. The factor 2 in (ii) is the squared normalisation.) The lemma and its
proof hold for any number field. For example over Q, where λp = log p: for Λ = Z2 and I = 12 Z the
covolume of IΛ is 14 , and the slope (III.100) rises by 2 log 2.
Lemma L3 (integrality after the multiplication). Let U, U ′ be L0 -vector spaces and
⟨ , ⟩ : U × U ′ → L0 a bilinear map. Let Λ ⊂ U and R ⊂ U ′ be lattices. For each place w let
Genw ⊂ U ′ be a set that generates Rw over ow , and let zw be integers, zero for almost all w, such
that
⟨y, y ′ ⟩ = 0 or vw ⟨y, y ′ ⟩ ≥ zw for all y ∈ Λ, y ′ ∈ Genw . (III.103)
Let I = w p−z
w . Then ⟨ỹ, ỹ ⟩ ∈ o for all ỹ ∈ IΛ and ỹ ∈ R. If moreover vw0 (⟨y, y ⟩) = zw0 for
′ ′ ′
Q w
some place w0 , some y ∈ Λ and some y ∈ Genw0 , then pw0 IΛ does not have this property.
′ −1
Proof. An element of L0 lies in o if its valuation is ≥ 0 at every place. Fix w. The element ỹ
is a finite sum of products βy with β ∈ I and y ∈ Λ. The element ỹ ′ is a finite sum of products
β ′ y ′ with β ′ ∈ ow and y ′ ∈ Genw . So ⟨ỹ, ỹ ′ ⟩ is a sum of terms ββ ′ ⟨y, y ′ ⟩, and each term is zero
93
or has valuation at least −zw + 0 + zw = 0 at w. For the second assertion take β ∈ p−1 w0 I with
vw0 (β) = −zw0 − 1, and β ′ ∈ o ∖ pw0 with β ′ y ′ ∈ R. Then vw0 (⟨βy, β ′ y ′ ⟩) = −1.
No slope occurs in Lemma L3, and it holds also for spaces of dimension 0. (Lean: ShiftPairing,
DivIntoDualShift. There the hypothesis is stated with real numbers and integers at most their
ceilings, which is the same since valuations are integers; and the local forms differ from ⟨ , ⟩ by
units at w, which changes nothing.)
Remarks. (1) Places are independent. By Lemma L2 (i) the localisation of IΛ at w depends
only on zw , and by (ii) the gains add over the places. This holds also for two places above the
same prime. It is the form over a number field of the gluing over Q of Part II, §6.8, where the
Chinese remainder theorem was used. (2) Maps compose. Suppose that there are m maps and m + 1
lattices, and that for 1 ≤ j ≤ m the jth map sends the localisation at w of the lattice before it into
z
pww,j times the localisation of the lattice after it. Then the composite sends the first localisation
z +···+zw,m
into pww,1 times the last. So the exponents of the m maps of the chain add. This is the
integrality assertion of Part II, §6.2. (3) A sublattice costs. If Λ′w ⊂ pzw Λw with an integer z ≥ 0,
then lgw (Λ/Λ′ ) ≥ z rk Λ, and by Lemma L1 the slope drops by at least 2λw z at w.
√
Checks. The three lemmas were tested in exact arithmetic over Q( −5), whose class number is
2, with the ramified prime above 2, the two primes above 3 and above 7, and the inert prime 11;
lattices of rank 1 to 3, not all free. In 60 cases the two sides of Lemma L2 (ii) agree to 2 · 10−15 ;
with λw zw in place of 2 λw zw they do not. In 25 cases the index formula of Lemma L1 holds
P P
exactly, the lengths being computed place by place. In 12 cases the pairing of Lemma L3 is integral
on IΛ and is not on p−1 w0 IΛ.
Letters introduced in this section. The field L0 , o, n0 ; a place w with pw , pw , ew , vw , ow ; λw , szw ;
spaces U , U ′ ; a lattice Λ, rk, µ, cstL0 , lgw ; integers zw , real numbers σw , an ideal I, a set of primes
Rd; elements y, ỹ of U and y ′ , ỹ ′ of U ′ , and scalars β, β ′ in L0 ; a member α, with Uα , Λsrc α , Λα ,
div
Λα , Iα , Lvα , rkα , Rdα ; the lattice R, the sets Genw ; sgα (w), ϕ(j, w), err(α, j, w), io(w); Gtop , F0 ,
gl ∗
Eerr (α), Lo(α); the group H, hp , np , an exponent i of a prime, cell, wtj , V∞ , an orbit Q, T (Q),
n(Q), κ(Q), Srv(p), kaj , gainplain , gainsrv , Pr, Ex, dl, C1 , C2 , digp , the constants Cpl , Clv , Cden of
§2 and Cr of §5, oddp , real numbers A, c, ci , ψi , counts nvar , nst , ntop ; sets of primes Ξ, Ξs , Ξ1 , Ξc ;
def(p), Tp , Sig(p), bw, θ, a′ , Y ; and in §6 Perα , htα and the error terms Esize , Elev , Ecnt , Esrc . The
number of digits of an integer in base p, written λ(k) in §III.5, is written digp (k) here, because λw
is the weight of a place. The level of a member, written T in Part II, §6.7, is written Lvα here,
because T is the set of labels, as in §III.5.
2. The glued lattice: statement and proof
The data. Fix x and one choice of the data of the construction (a “record” of the source). The
record contains one translate u ∈ H (Lean: the field Host.u0). It is the translate u of Part II, §4.2,
and the weights of the record are the weights for this u (§4 below). The maps of the record are
indexed by a finite set. Its elements α are the members (Lean: Word). A member is not a translate.
All members of a record have the same translate, and the prime sum Pu,t,j does not depend on the
member (the Lean statement Loss_at contains no member). A member enters through its lattices,
its level and its error terms. For each member α the source gives:
• an L0 -vector space Uα of dimension ≥ 1 with a lattice Λsrc α , the source lattice (Lean: MramL;
it is the preimage of the first lattice of the chain of maps; the positive dimension is the field
a_pos of the inputs of divClausesR_of, §3);
• a lattice Λdiv
α ⊂ Λα (Lean: NuLat);
src
94
• a space U ′ and a lattice R ⊂ U ′ , the same for all α (Lean: RR), and a bilinear pairing
⟨ , ⟩α : Uα × U ′ → L0 , the period pairing of the whole chain (Lean: perCol; the bilinear form
is col.B);
• a prime Lvα , the level of the member (the level T of Part II, §6.7), and an integer rk∗α ≥ rk Λsrc
α
(Lean: duS).
Independently of α except where shown, there are real numbers, zero outside a finite set of places:
gain(j, w) ≥ 0, ϕ(j, w), err(α, j, w), τ (w) ≥ 0, io(w) ≥ 0, for 1 ≤ j ≤ m (Lean: the fields gain, F,
E, tau, io of SizeData2 ). Here gain and τ are the numbers of (III.79). Put
m
sgα (w) = gain(j, w) + ϕ(j, w) − err(α, j, w) , (III.104)
X
j=1
X
Gtop = szw gain(j, w) − τ (w) , F0 = szw ϕ(j, w) , Eerr (α) = szw err(α, j, w) .
X X X
w j j,w j,w
(III.105)
(Lean: sg, Ga, Fo, Ee. In the Lean the gain is first set to zero at the places above a set of “dropped”
primes, see §5; we assume this done.) So
szw sgα (w) − τ (w) = Gtop + F0 − Eerr (α). (III.106)
X
w
The number F0 is the F0 of Part II, §6.7. The glued lattice is
Λgl
α = Iα Λα , Iα = p−⌊e (III.107)
Y
div w sgα (w)⌋
w .
w
(Lean: Nu, Iu, and the statement NprimeDef, which is (III.107).) The factor ew converts an
exponent of pw into an exponent of pw .
The two statements taken from the source.
(Val) Valuation. For every member α and every place w there is a set Genw ⊂ U ′ generating Rw
over ow with
⟨y, y ′ ⟩α = 0 or vw ⟨y, y ′ ⟩α ew sgα (w) (y ∈ Λdiv ′
α , y ∈ Genw ). (III.108)
≥
(Lat) Lattices. For every member α and every place w with pw ̸= Lvα : τ (w) is an integer,
ew τ (w)
(Λdiv
α )w ⊂ pw (Λsrc
α )w , and
lgw Λsrc div
≤ rk∗α ew τ (w) + io(w) . (III.109)
α /Λα
Note what (Lat) does not give. It has rk∗α , not rk Λsrc
α , on the right of (III.109). And it says nothing
at the places above Lvα .
The statement (Val) in the Lean. It is not one Lean statement. It is assumed here; in the Lean
it is obtained from three assumptions.
(1) LocalQuadraticParts.adm2. Let µw be the twist chosen at w (muAt). The letters of this
item are those of the Lean: µw , with a place as index, is a twist and not a slope; the element
F below has nothing to do with the number F0 ; and M f, V0 have nothing to do with Mu,t,j
and V . The statement is: for every member α, every place w, every y ∈ Λdiv α , and every F
that lies both in the lattice M
f(µw ) (Mtilde) and in the subspace V0 (µw ) (V0) of the ambient
95
space of the twist, there is an element of L0 whose image under the fixed embedding of L0
in C (embLL) is the period of y against F ; and this element is zero or has valuation at least
⌊ew sgα (w)⌋ at w. Both restrictions on F are part of the statement. It says nothing for an
element of the lattice outside V0 (µw ), and nothing for an element of V0 (µw ) outside the lattice.
That the period is the embedded value of an element of L0 is also part of the statement; (Val)
presupposes it.
(2) The hypotheses of the theorem val2AllAtRF. This theorem turns (1) into the same bound
for the matrix form of the pairing, on a lattice Nu given in coordinates. Besides (1) it has
three hypotheses. Two are carried by structures. ColBrF.Tied says that the matrix form is
tied to the period: the twist at w is the one of the matrix form, each element of the lattice
of the matrix form maps into V0 and is an integral combination of the generators of M f,
and the matrix pairing has the period as its embedded value. CutUp, for every twist that
is used, says that an element of V0 that is an integral combination of the generators of M f
is an integral combination of elements of M ∩ V0 . This is what the two restrictions in (1)
f
require. The third hypothesis is the compatibility with the source of (1): every element of Nu
is mapped by the coordinate map eNu into NuLat, that is, into Λdiv α . It is not an assumption.
Where val2AllAtRF is applied (in the files RecBridge.lean and RecBridgeTied.lean) it is
supplied by the theorem nu_into, from the definition (III.107) (NprimeDef) and two identities
of lattices: shift, a field of the package, and lat, a field of the structure HLatTie, which the
package carries for each member as its field hNu.
(3) The structure field l1b (of type PairingFactorsThroughPeriods). It says that the local
form at w equals a unit at w times ⟨y, y ′ ⟩α , where y ′ is the image of the element of the local
lattice under a fixed linear map.
That the sets Genw generate Rw is not an assumption. The set Genw is the image of the local
lattice of (1). The lattice R is defined as the lattice glued from these images (RR, by glueLat),
and the theorem latticeGlue proves that this is a lattice with these localisations (Lemma L2 (i)
is the case of an ideal). The passage to one bilinear pairing is faithful: the Lean theorem intoDual
applies DivIntoDualShift (Lemma L3) with one L0 -bilinear form. The quantifiers of (Val) are
those of the Lean: all members, all places, all y ∈ Λdiv
α . The final theorem depends on adm2 (the
dependency computation confirms this). It also depends on val2AllAtRF and on latticeGlue: in
the seventh run of the dependency computation both are marked as used. Whether the hypotheses
Tied and CutUp of (2) and the field of (3) are discharged for the construction on the route of the
final theorem was not determined. The dependency computation does not say how hypotheses are
discharged, and the proof terms were not followed that far.
The statement (Lat) in the Lean. It is the part of the first clause of LinkG_at that is used
here, read with NuDef. That clause is for the places w with pw ̸= Lvα and has three parts. The
first is the inclusion, with the exponent ⌊ew τ (w)⌋. The second is lenMN(α, w) ≤ rk∗α (τ (w) + io(w)).
The third says: if the set VV of the Lean at the prime pw is empty, then τ (w) = 0, io(w) = 0 and
lenMN(α, w) = 0. The third part is not used in Theorem L4. The number lenMN is not the local
length. The last clause of NuDef says that lgw (Λsrc
α /Λα ) = lenMN(α, w) · ew . With the second part
div
this gives (III.109), and it is where the factor ew there comes from. For the integrality of τ (w) two
statements are needed. The theorem exists_tau_eq_natCast says that the number given by the
formula tauP is a natural number. It does not mention the τ of the record. That τ (w) equals tauP
at pw is part of the hypothesis fills of §3.
Theorem L4 (the glued lattice). Let α be a member. Write rk = rk Λsrc
α and rk = rkα ; so
∗ ∗
rk ≥ rk ≥ 1.
∗
96
(i) Assume (Val). Then ⟨y, y ′ ⟩α ∈ o for all y ∈ Λgl
α and y ∈ R. That is, the pairing maps Λα
′ gl
into the dual lattice R = {ξ ∈ HomL0 (U , L0 ) : ξ(R) ⊂ o}.
∨ ′
(ii) Without any assumption,
2 X
α ) = µ(Λα ) + 2λw ew sgα (w) − λw lgw Λsrc (III.110)
X
µ(Λgl src div
α /Λα .
w rk w
(iii) Assume (Lat). Let Rdα be the set of primes pw with ew sgα (w) ∈
/ Z. Then
α ) ≥ µ(Λα ) + Gtop + F0 − Eerr (α) − Lo(α),
µ(Λgl (III.111)
src
where
rk∗ io(w) + (rk∗ − rk) τ (w)
Lo(α) = szw
X
w: pw ̸=Lvα
rk
2 (III.112)
+ λw lgw Λsrc − szw τ (w) +2 log p.
X X
div
α /Λα
w: pw =Lvα
rk p∈Rdα
The three terms of Lo(α) are: the loss of the lattice Λdiv α beyond τ , with the cost of the rank
rk∗ ≥ rk; the places above the level, where (Lat) gives no bound and the length is kept as it is; and
the rounding. If rk∗ = rk the first term is the sum of szw (io(w)) over the places w with pw ̸= Lvα .
Proof. (i) This is Lemma L3 with Λ = Λdiv
α and zw = ⌊ew sgα (w)⌋; the hypothesis (III.103) is
(III.108).
(ii) Lemma L2 (ii) for the pair (Iα , Λdiv
α ) gives α ) = µ(Λα )+
µ(Λgl div
w 2λw ⌊ew sgα (w)⌋. Lemma L1
P
for Λα ⊂ Λα gives µ(Λα ) = µ(Λα ) − rk w λw lgw (Λα /Λdiv
div src div src 2 P src
α ).
(iii) We bound the two sums in (III.110). The first sum. Apply Lemma L2 (iii) with σw =
ew sgα (w) and Rd = Rdα . These numbers vanish outside a finite set of places, and 2λw σw =
szw (sgα (w)). So
2λw ew sgα (w) ≥ szw sgα (w) − 2 log p. (III.113)
X X X
w w p∈Rdα
The second sum, at a place w with pw ̸= Lvα . By (III.109) and Lemma L5 below (whose proof is
one line and uses nothing of this theorem),
2 rk∗ io(w) + (rk∗ − rk) τ (w)
λw lgw Λsrc div
sz (w) + sz (III.114)
/Λ ≤ w τ w .
rk α α
rk
The second sum, at a place w with pw = Lvα . Here nothing is assumed. We write the term
as szw (τ (w)) plus the difference rk
2
λw lgw (Λsrc
α /Λα ) − szw (τ (w)). Conclusion. Insert (III.113),
div
(III.114) and the last identity into (III.110). All sums are finite. The result is
α ) ≥ µ(Λα ) + szw sgα (w) − τ (w) − Lo(α), (III.115)
X
µ(Λgl src
w
with Lo(α) as in (III.112). By (III.106) this is (III.111).
e τ (w)
Of (Lat) the proof of (iii) uses only (III.109). The inclusion (Λdiv
α )w ⊂ pw
w
(Λsrc
α )w explains the
signs. The gain gain(j, w) enters Λα with a plus sign and τ (w) with a minus sign for the following
gl
reason. The divisibility (III.108) holds on Λdiv α , not on Λα . By the inclusion and Remark (3),
src
Λα is smaller than Λα by at least τ (w) at w, in the units szw . The net exponent at w is
div src
j gain(j, w) − τ (w): the right side of (III.79).
P
97
Lemma L5 (when the rank drops). Let Λ′ ⊂ Λ be lattices, w a place, rk = rk Λ ≥ 1, and
let rk∗ ≥ rk, τ (w) ≥ 0 and io(w) ≥ 0 be real numbers with lgw (Λ/Λ′ ) ≤ rk∗ ew (τ (w) + io(w)). Then
2 rk∗ io(w) + (rk∗ − rk) τ (w)
λw lgw (Λ/Λ′ ) − szw τ (w) ≤ szw (III.116)
.
rk rk
∗ ∗
Proof. The left side is at most 2λw ew rk
rk (τ (w) + io(w)) − τ (w) , and rk (τ (w) + io(w)) − τ (w) =
rk
∗ ∗
rk io(w)+(rk −rk)τ (w)
rk .
(Lean: ClaimLow. There the rank deficit is a number de with rk = (1 − de) rk∗ , and the right side
of (III.116) reads szw (io(w) + de τ (w))/(1 − de) , which is the same number.)
√
Checks. Theorem L4 was tested in exact arithmetic over Q( −5) on 200 lattices of rank 2,
with rk∗ from 2 to 5, with the level prime 2, 3, 7 or 11, with a sublattice that is smaller at a place
above the level, and with every hypothesis of (Lat) verified on the data. The identity (III.110)
holds exactly in all cases, and (III.111) holds in all cases. In 60 cases the data were chosen with
equality in (III.109) and with ew sgα (w) integral; then (III.111) is an equality, exactly, in all 60. On
these 60 cases three weaker forms of Lo(α) were tried. With w szw (io(w)) in place of the first two
P
terms (the loss one would get from (III.109) with rk in place of rk∗ at all places) the inequality
fails in 48. Without the term at the level it fails in 15, which are all the cases where that term is
positive. Without the factor for the rank it fails in 46, which are all the cases where rk∗ > rk and
τ + io is not zero. On the other 140 cases, without the rounding term it fails in 15. Lemma L5 was
also tested on 20000 random data in the form of the Lean statement ClaimLow; without the factor
for the rank the inequality fails in a third of the cases.
The Lean statement, and what it compares with. The Lean statement of the slope
bound (SlopesN_at, proved from Places_at by lower_bound_simplified) is not (III.111) word
for word. It reads
a
α ) ≥ µ0,α +Gtop +F0 −Eerr (α)− Cpl λw ew io(w)+Cpl |T | m +b′ log Lvα +Cpl X ′′ +1 ,
X
µ(Λgl
w k
(III.117)
where Cpl is a constant of the source (it is not the constant C of (III.77)), and µ0,α is the slope of
the first lattice of the chain, not of Λsrc
α , which is its preimage. It comes with an upper bound of the
same shape for the largest slope of a sublattice of Λgl α . The passage from (III.111) to (III.117) uses
four statements of the source, which are assumed here (Lean: adm26, HypLow2, places_at_of).
(a) The rank. rk = (1 − 1/(Lvα − 1))f rk∗ (adm26). Hence (rk∗ − rk)/rk∗ ≤ 2f /Lvα , which the
source bounds by 12 . So rk∗ /rk ≤ 2, and the part of the first term of (III.112) with io is at
most 4 w λw ew io(w). This factor is absorbed in the constant Cpl of (III.117); it is not part
P
of the “+1”.
(b) The places above the level. There τ (w) = 0; this is the hypothesis htauT of places_at_of,
not a clause of adm26. The second term of (III.112) is then at most Cpl |T |m(a/k + b′ ) log Lvα
(the last clause of adm26). This is the second term of (III.117).
(c) The rounding. 2 p∈Rdα log p ≤ Cpl X ′′ . The theorem places_at_of does not prove this: it
P
has the bound as a hypothesis (NprimeMuR, the statement that µ(Λgl α ) − µ(Λα ) differs from
div
w szw (sgα (w)) by at most Cpl X ). The source proves that hypothesis elsewhere (round_le,
P ′′
nprimeMu_at, nprimeMuR_at, linkMu). There the rounding is counted at every place with
sgα (w) ̸= 0; the primes below these places include Rdα . The set Rdα is not contained in
[1, X ′′ ]. By two statements of the source (ZeroAbove for gain and err, the field F_above for
ϕ; together sg_eq_zero), sgα (w) = 0 if pw > X ′′ , unless pw is Lvα , g0 , g1 , d or a prime factor
98
of the denominator Dα of §3. So
log p ≤ X ′′ log 4 + log Lvα + log g0 + log g1 + log d + log Dα , (III.118)
X
p∈Rdα
by Chebyshev’s bound for the primes up to X ′′ (sum_log_primes_le), and because the sum
of log p over the prime factors of Dα is at most log Dα (sum_log_primeFactors_le). This
display is the statement nprimeMu_at. The source then uses that each of log g0 , log g1 , log d
is at most X ′′ (from the size conditions SizeG, SizeQD), and two further hypotheses on the
family: log Lvα ≤ Clv X ′′ and log Dα ≤ Cden X ′′ (MuTu, MuDGam). This gives the constant
2(log 4 + 3 + Clv + Cden ) in NprimeMuR (nprimeMuR_at, linkMu). It also assumes that the
glued lattices are finitely generated and have positive rank (MuFG, MuRank). These hypotheses
are assumed here. The names of this item were not put to the dependency computation.
(d) The first lattice of the chain, and the “+1”. The source has µ(Λsrc α ) = µ0,α − f log(Lvα −
1)/(Lvα − 2) exactly (adm26), and w szw (τ (w)) ≤ b2 (a hypothesis of places_at_of; the
P
source proves it as ta_le). With (a), the part of the first term of (III.112) with τ is at most
4f b2 /Lvα . The source requires 4f b2 /Lvα + f log(Lvα − 1)/(Lvα − 2) ≤ 1 (HypLow2). So the
cost of the rank on τ and the change from µ(Λsrc α ) to µ0,α are together at most 1. This is the
single “+1” of (III.117). There is one such term, not two.
3. What LinkG is, and where the valuation statement lives
The statement LinkG of the source has five clauses (LinkG_at). They are: (1) the lattice clause, of
which (Lat) is the part used in §2; (2) a bound for the auxiliary numbers Sv; (3) the inequality
(III.79) of §III.5, with 0 ≤ gain(j, w) ≤ 4|T |a/pw , at the places of class signal or inertClass
with pw ≤ X ′′ ; (4) a clause that holds by definition; (5) a bound for the mean of certain numbers
κ(Q), used in §5. None of them mentions a theta function, a Taylor coefficient, or a pairing. As
§III.5 already said, (III.79) is an inequality between the numbers gain, τ, wPart. A hypothesis on
the family (the field fills, of type ParametersFill) says that these numbers are given by explicit
formulas (gainP, tauP, fill, partner, TTP); they are recalled in §4. With that hypothesis, clause
(3) is arithmetic about the formulas. Proposition L10 below carries this out at the primes of class
signal, with its own loss terms.
The zero chain. The proof of LinkG in the source (linkG_without_kernels) goes through an
intermediate statement admG1, one part of which does speak of the valuation of a chain of maps on
a local lattice (vKge). On the route to the final theorem that part is applied to the chain in which
every map is zero (zeroCh). The valuation of 0 is infinite, so the part holds for a trivial reason
(vKge_zero). So LinkG is proved without any information on the theta kernels. This is consistent:
LinkG is the bookkeeping, and the lattice clause (1).
Where the divisibility is asserted. It is the statement (Val). The final theorem depends on
LocalQuadraticParts.adm2. The route, read downward from the place where it is consumed, is
this.
(a) divClausesR_of takes a structure of inputs, DivInputsR, and produces four clauses. One
field of DivInputsR is val2, that is (III.108). The other fields are independent of it.
From val2 comes the first clause only: the integrality of Theorem L4 (i), by intoDual
and divIntoDualShift (Lemma L3), which use the field l1b. The second clause bounds
the squares of the operator norms of the pairing at the complex places by numbers whose
mean logarithm is the height of the record minus a constant. It comes from the height inputs
(the fields ch, hn, nP, nKcb, eig6). The third clause is the inequality of sizes, first line of
99
(6.5) of Part II, with its term −F0 . It comes from the size inputs (among them L3main, Qb,
deQ, errB, and the inequality constant_ge_twenty_eight). The fourth clause is the slope
identity of Lemma L2 (ii), by shiftMu; it uses that Nu is a lattice of positive rank (Nu_lat,
a_pos). So the valuation statement does not imply the height bound or the inequality of sizes.
The inequality of sizes is the subject of §III.4.
(b) val2AllAtRF deduces val2 from adm2, by linearity, under the hypotheses described in §2. In
the seventh run of the dependency computation val2AllAtRF is marked as used by the final
theorem. The archive has a second name, DivisionPackage.val2. The query for it returned
an empty list. This is inconclusive: it does not show which fully qualified name was intended,
and it does not show that such a declaration is unused.
(c) admissible_level_two proves adm2, place by place. At the places above Lvα and above
the two auxiliary level primes g0 , g1 of Part II, §6.7 it uses separate inputs. At all other
places it uses the statement Conclusion1 of the file WeilJacquetShalikaLaw.lean: for
every y in the local lattice at w that defines Λdiv α , the image of y under the composite of
the m maps of the chain is divisible, in the last lattice of the chain, by pw to the power
(gain(j, + err(α, + (D ) Here Dα is an integer denominator
− ·
P
ew j w) ϕ(j, w) j, w)) sp v pw α .
of the member, vpw is the pw -adic valuation, and sp is a real parameter of the Lean statement.
On the route sp is 1: the statement Serv contains Conclusion1 with this value. The term
with Dα does not reach (III.108). The theorem period_geTU, which passes from the chain to
the period, lowers the exponent by exactly vpw (Dα ) (vDG), and the two cancel. This is the
statement about kernels.
(d) Conclusion1 is part of ServF, which sum_over_primes_service proves from the structure
SumOverPrimesHypotheses, which has the fields twU and pdv, a bound for the denominators
Dα (DGamSize), and the theorem split_bound_eleven_one. The latter is the local estimate:
it rests on thmAOp, the Lean form of Theorem J of §III.5. The hypothesis pdv contains
hypothesis (B) of §III.5 for the kernels of the construction.
So there are two branches. One (LinkG) says what the numbers gain and τ are and that they
dominate the sums V ; it needs no kernel. The other (ServF, then (Val)) says that the kernels are
divisible by these numbers; it needs Theorem J and hypothesis (B), and it is assumed here. They
meet in Theorem L4 and in §5.
4. From Theorem J to a gain at a place
Cells, and sums over one prime. Let H be the group of Part II, §4.1, of order n. It is the
group of §III.3, and it is written additively here, as there and as in the Lean; T ⊂ H. A cell is a
pair (0, t) or (1, t) with t ∈ H; the first entry is the row. There are f = 2n cells. For an odd prime
p let hp ∈ H be its class, of order np . The classes of primes are those of the Lean (primeClass): a
prime p ∤ 2qd is splitK if χd (p) = 1; if χd (p) = −1 it is signal when χq (p) = 1 and inertClass
when χq (p) = −1. For an odd prime p ̸= q, t ∈ H and an integer i ≥ 1 put
1 + χ (p)i
q
cell(p, t, i) = , t + i hp . (III.119)
2
This is the cell of the prime power pi in the table of Part II, §4.1, for the column t. For odd i the
row is 1 if p is signal and 0 if p is inertClass. (Lean, for odd i: partner.) For each j there are
weights wtj > 0 on the cells (Lean: fill); these are the weights of Part II, §4.2, for the translate u
100
of the record and the map j. The hypothesis fills says wPart(j, t, p, i) = wtj (cell(p, t, i)). Put
X at
V∞ (j, t, p) = i
− wtj cell(p, t, i) , (III.120)
i odd
p +
the sum being over all odd i ≥ 1. The sum V (j, t, p) of (III.76) has the same terms, but only those
with pi−1 ≤ at . If all the weights are ≥ 1, then V∞ = V : the terms with pi > at vanish, and (III.76)
contains every i with pi ≤ at . In the source the weights that occur are ≥ 1 (§III.5, §1). Without
this hypothesis the two sums can differ; see the example below.
Lemma L6 (the prime sum). Let the weights be any positive numbers. The sum (4.1) of
Part II, for these weights, is
X at
Pu,t,j = 2 log p V∞ (j, t, p) + log d − wtj cell(d, t, i) (III.121)
X
.
p signal i≥1
di +
or inertClass
Proof. In (4.1) the coefficient of pi is (1 − χd (pi )) log p. For p =
̸ d it is 2 log p if χd (p) = −1 and i is
odd, and 0 otherwise. The row and the column of p are those of (III.119). The prime d gives the
i
last term, with coefficient log d for every i. The primes 2 and q are absent, and split primes have
coefficient 0.
In the running example of Part II, §4.2 (q = 3, d = 23, n = 4, at = 10000) the weights are 0.05
and 10. Computing the right side of (III.121) gives 95461.6156 for the translate [1], the value of
Part II, and the four values for the four translates agree with (4.1) to 10−10 . The example uses
V∞ , not V : its weights 0.05 are below 1, and with V in place of V∞ the right side of (III.121) is
smaller by 0.309064 for the translate [1] (the powers 57 and 115 are lost). With the weights 10 and
1 the difference is 0. The primes of class inertClass contribute 51234.4133 for every translate:
their odd powers lie in row 0, which contains no distinguished cell. The primes of class signal
contribute 42833.7633 for the translate [1]. The distinguished weights carry the information on the
zeros. Among the unramified primes, only signal primes sample a distinguished weight; this is the
reason for the name. The powers of d contribute the separate last term of Lemma L6, and that
term does depend on the translate. Here 232 ≡ 1 modulo 3 and modulo 16, so the cell of 232 for
the column t is (1, t), in row 1, and 10000/232 > 10. The last term is 1393.4390 for the translate
[1] and 1424.6372 for the translates [3], [5], [7]. The difference is (10 − 0.05) log 23.
The formulas for the gain. From here on the weights are ≥ 1, as in the source, so that
V∞ = V . Let p be signal or inertClass. The group generated by hp (by 2hp if p is signal) acts
on H by translation. For an orbit Q write T (Q) = T ∩ Q. (The Lean fixes a base point on each
orbit, reps.) Put kaj = 12 k + (2j − 1)(t0 + 1) , with the integer parameter t0 of §III.5. The source
has 0 ≤ kaj ≤ Sc for 1 ≤ j ≤ m (kaJ_nonneg for the lower bound, kaJ_le for the upper bound).
For an orbit Q put
a − 8 Sc
t
gainplain (j, Q) = − wtj cell(p, t, i) + kaj − 4 Sc − dl(Q), (III.122)
X X
t∈T (Q) i odd
pi +
i∈Ex(p,t)
/
2
dl(Q) = C1 n(Q) Nm
2
logp (2a) + C2 |T (Q)| logp (2a) . (III.123)
Here n(Q) is the number of sites of the orbit. For an inertClass prime the sites are the points
of the orbit, and n(Q) = np . For a signal prime a site is a pair consisting of the parity of an
exponent i ≥ 0 and the column t + i hp , with t ∈ Q; there are two parities and, for each, as many
columns as Q has points, so n(Q) is twice the order of 2hp . (Lean: the sites are the values of
101
partner, and n(Q) is nP. The sites are not the cells (III.119) of the powers of p: the two agree
for odd i only.) The numbers C1 , C2 ≥ 0 are constants, and Ex(p, t) is the set of odd i for which
cell(p, t, i) = (0, t′ ) with t′ ∈ T , t′ ̸= t, and i is the least positive exponent with this property for t′ .
(Lean: g0, dl, Excl, al, capU, kaJ; the number 4 Sc is cG.) The gain is
gain(j, p) = max gainsrv (j, Q) − Pr(j, Q), gainplain (j, Q), 0
X
Q∈Srv(p)
(III.124)
+ max gainplain (j, Q), 0
X
Q∈Srv(p)
/
for p ≤ X ′′ , and gain(j, p) = 0 for p > X ′′ and for inertClass primes with hp = 0. Here Srv(p)
is the set of served orbits, defined in §5; it is empty unless p is inertClass and p < X ′ . The
numbers gainsrv and Pr are further explicit expressions (Lean: gU, Pr): gainsrv is (III.122) without
the exclusion of Ex and with the full (at − 8 Sc) p/(p2 − 1) for the points t at even positions on the
orbit, by the identity (6.1) of Part II. By fills, gain(j, w) = gain(j, pw ) and τ (w) depends only
on pw ; we write τ (p) for it.
Comparison with Theorem J. Theorem J of §III.5 is applied to one orbit: the sites are the
sites of the orbit just defined, the top sites are the sites whose column lies in T (Q) (for a signal
prime: those of even parity among them), the number at −8 Sc plays the role of the length of the first
row at the top site t, and wtj −kaj +4 Sc that of the capacity at a site. The term top of (III.82) then
has the shape of the double sum in (III.122), and the last term of (III.82) has the shape of the first
term of (III.123) (Lemma L8). This dictionary, the treatment of the excluded exponents, the other
terms of (III.82), and the passage from one function to the lattices of the chain are in the source
(normCor7, split_bound_eleven_one, orbit_entry_bounds_hold, lift_bounds_of_essential,
then WeilJacquetShalikaLaw.Proof.stmt_proved); they are assumed here.
The statement asmA_top. The final theorem uses this statement for the top row (it is marked as
used in the seventh run of the dependency computation). It was read and compared with Lemma 6
of §III.5. In this paragraph s is a site, ℓ is the prime, and ν1 , cap, top, dtop are as in §III.5, §1; i is
an odd integer. The Lean statement is: for every site s and every finite set of odd integers i ≥ 1
there is a real number ω ≥ 0, with ω ≥ ℓ−dtop (s) if dtop (s) is finite, such that
ℓ
top(s) + ν1 (s) ω + cap(s + i) 1 − ℓi ω + ≤ ν1 (s) 2 (III.125)
X
.
i in the set
ℓ −1
It relates to Lemma 6 of §III.5 as follows.
(a) If dtop (s) is finite, take the set of all odd i < dtop (s). Then (III.125) is (III.93) of that lemma,
with A = ν1 (s), ci = cap(s + i) and with dtop (s) as the odd bound J of the lemma: the
same statement. Here top(s) is the whole of the second line of (III.82): the sum over odd
h < dtop (s) and the sum of ν1 (s)/ℓh over odd h > dtop (s) (the Lean’s topTerm has both).
It is the right side of (III.93), and ν1 (s) ℓ/(ℓ2 − 1) = γA, with γ = ℓ/(ℓ2 − 1) as in §III.5.
So (III.125) is (III.93) with top(s) moved to the left. Both assert only that some ω exists;
neither holds for every ω ≥ ℓ−dtop (s) . For example, for ν1 (s) = 9, ℓ = 3, dtop (s) = 1 one has
top(s) = 38 and γA = 27 8 ; both hold with equality at ω = 3 and both fail at ω = 8 . A larger
1 3
set gives nothing more, because the terms with i ≥ dtop (s) vanish when ω ≥ ℓ top (s) . The
−d
Lean proof takes the same radius as the proof of the lemma: ω = ℓ−i0 , where i0 is the least
odd i < dtop (s) with cap(s + i) ≥ ν1 (s)/ℓi , and i0 = dtop (s) if there is none.
(b) If dtop (s) = ∞, the Lean statement is not Lemma 6, which has a finite bound. It is the
remark after Lemma 6. The Lean allows any finite set of odd i; the remark has the restriction
ℓi ≤ ν1 (s), and the Lean does not.
102
(c) The Lean has ℓ any integer ≥ 2, and ν1 , cap natural numbers.
(Tested: 25600 exact cases for (a), 30000 for the radius; and 40000 further exact cases in which the
two inequalities were evaluated from the definitions at the radius of the lemma and at an arbitrary
ω, and agreed.)
The statement lemma_R_log_loss. The final theorem also uses this statement (seventh run).
It is the bound for one Taylor coefficient on the route of the final theorem. Its comparison with
Proposition 5 is in §III.5, in the paragraph on the declarations the final theorem depends on:
Proposition 5 is the stronger of the two bounds for a coefficient. The formal statement (LemmaRA,
in the file GaussJetBounds/StatementG.lean) is for a multi-index whose entries are all at most
an integer a ≥ 1, and it assumes a test on the pure monomials below the multi-index. Its exponent
has the explicit loss |ι| (logℓ a + 1), where ι is the finite index set of the variables and ℓ is the prime.
(In this sentence a is the bound of the Lean statement.)
We now prove what is elementary: three comparisons of losses, and then an inequality of the
shape of (III.79) at the signal primes.
Lemma L7 (shifted positive parts). Let p ≥ 2, A ≥ 0, c ≥ 0, and for each odd i ≥ 1 let
ψi ≥ 1 and ci ≥ 0 be real numbers. Let oddp (A) be the number of odd i ≥ 1 with pi ≤ A. Then
X A − 2c X A 2
− ψi + ci − c ≥ − ψi −c 1+ oddp (A). (III.126)
i odd
pi +
i odd
p i + p
The left side may be +∞. If A ≥ 1, then oddp (A) ≤ 21 (logp A + 1).
Proof. For each i, A−2c
pi
− ψi + ci − c ≥ pAi − ψi − c(1 + p2i ). The positive part is increasing,
and subtracting a number ≥ 0 lowers a positive part by at most that number. If pi > A then
A/pi − ψi < 0, so the term on the right is 0 and nothing is lost. The odd i with pi ≤ A are the odd
i with 1 ≤ i ≤ logp A; for A ≥ 1 their number is at most 12 (logp A + 1).
The bound for oddp (A) needs A ≥ 1: for 0 < A < 1/p the count is 0 and 12 (logp A + 1) is negative.
In the use below A = at ≥ 1. The left side of (III.126) is finite if ci ≤ c + ψi for all large i; this
holds in the use below.
Lemma L8 (three digit losses, one bound). Let p be a prime and ā ≥ p an integer. Then
digp (ā) ≤ logp ā + 1 ≤ 2 logp ā, and:
(i) (the loss of §III.5) the sum of digp over any nvar integers in [0, ā] is at most nvar digp (ā); in
the last term of (III.82) the number nvar of variables is the number nst of sites times the two
numbers N0 , N1 of §III.5;
(ii) (the loss ∆ of (6.1a) of Part II) if the index set there has nst elements, its subset has ntop
elements, the multiplicities there are r0 , r1 , and all the integers whose digits are counted are
≤ ā, then ∆ ≤ nst (r1 + r0 r1 ) digp (ā) + ntop digp (ā)2 ;
(iii) (the loss of the Lean) if N0 , N1 , r0 , r1 ≤ Nm , both bounds are at most 2 nst Nm
2 dig (ā) +
p
ntop digp (ā) . With nst = n(Q), ntop = |T (Q)| and ā ≤ 2a this is at most (III.123) with
2
C1 = 4, C2 = 4.
Proof. digp is increasing and digp (ā) = ⌊logp ā⌋ + 1. (i) and (ii) are term-by-term bounds; for (ii)
recall that ∆ is a sum over the index set of r1 digit counts and r0 r1 digit counts, plus a sum over the
subset of squares of digit counts. (iii) N0 N1 ≤ Nm 2 , r + r r ≤ 2N 2 and dig (ā) ≤ 2 log (2a).
1 0 1 m p p
The paper has three digit losses: one digit count per variable (§III.5); one per factorial, with a
square for the top sites (Part II, (6.1a); Lean: StarArith, and the loss in gainA); and the closed
103
form (III.123) per orbit. They are three different quantities. The lemma shows only that (III.123)
is an upper bound for the first two. For example, with p = 3, ā = 80, two sites, N0 = N1 = 2 and
r0 = r1 = 1, the largest value of the first loss is 32 and that of the second, when both sites are in
the subset, is 48.
Lemma L9 (the sum over the orbits). For every prime p of class signal or inertClass,
Q n(Q) ≤ f and Q |T (Q)| = |T |, the sums being over the orbits. Hence
P P
2
dl(Q) ≤ C1 f Nm logp (2a) + C2 |T | logp (2a) . (III.127)
X
2
Q
Proof. The orbits partition H. For inertClass the orbit has np points and n(Q) = np , so the sum
is n = f /2. For signal the orbits are the cosets of the group generated by 2hp , and n(Q) is twice
their length, so the sum is 2n = f .
The right side of (III.127) is the term flG (p) of (III.77), up to the constant and to logp (2a) ≤ 2 logp a
(which holds when a ≥ 2), since |T | ≤ n = f /2. This is why flG is subtracted once for each of the
m maps in (III.79), and not once for each t.
Proposition L10 (the shape of (III.79) at a signal prime). Let p ≤ X ′′ be of class
signal, let the weights be ≥ 1, and let gain(j, p) be given by (III.124) with 0 ≤ kaj ≤ Sc. Then
Srv(p) = ∅, and for 1 ≤ j ≤ m
X 2 2
gain(j, p) ≥ V (j, t, p) − 4 Sc 1 +oddp (at ) − C1 f Nm
2
logp (2a) − C2 |T | logp (2a) ,
t∈T
p
(III.128)
and 0 ≤ gain(j, p) ≤ 43 t∈T at /p ≤ 4|T |a/p. Summing (III.128) over j gives, at the places w above
P
p,
m X m
V (j, t, p) − lo′ (p) − m fl′ (p) ≤ gain(j, w) − τ (w), (III.129)
X X
j=1 t∈T j=1
with τ (w) = 0, lo′ (p) = 2 Sc (1 + p2 )(logp (2a) + 1) and fl′ (p) = C1 f Nm
2 log (2a) + C |T |(log (2a))2 .
p 2 p
This is (III.79) with Tw = T and with lo , fl in place of loG , flG . If moreover
′ ′
C ≥ 20
3 , C ≥ 2C1 , C ≥ 2C2 , a ≥ 2, η ≥ 0, p ≤ 2a, (III.130)
then lo′ (p) ≤ loG (p) and fl′ (p) ≤ flG (p), and (III.79) itself holds at the places above p.
Proof. Served orbits exist only for inertClass primes, so Srv(p) = ∅ and gain(j, p) =
Q max(gain
P plain (j, Q), 0) ≥ P gainplain (j, Q). For a signal prime and odd i, cell(p, t, i) lies in
Q
row 1, so Ex(p, t) is empty. Apply Lemma L7 with A = at , c = 4 Sc, ψi = wtj (cell(p, t, i)) and
ci = kaj , sum over t ∈ T (Q) and over the orbits Q, and use Lemma L9. Since the weights are
≥ 1, the sum on the right of (III.126) is V (j, t, p). This gives (III.128). For the upper bound,
wtj − kaj + 4 Sc ≥ 0, so each inner sum in (III.122) is at most i odd at /pi = at p/(p2 − 1) ≤ 34 at /p,
P
and at ≤ 2a, 83 ≤ 4. For (III.129): τ (w) = 0 when Srv(p) = ∅ (the definition of tauP), and
oddp (at ) ≤ 12 (logp (2a) + 1) because 1 ≤ at ≤ 2a. For the last assertion, a signal prime is
odd, so 1 + p2 ≤ 35 ; and p ≤ 2a, so logp (2a) + 1 ≤ 2 logp (2a). Hence lo′ (p) ≤ 20 3 Sc logp (2a) ≤
C(Sc + k + η) logp (2a) = loG (p), since Sc, k, η ≥ 0 and the terms pr, or of (III.78) vanish at
a signal prime. For a ≥ 2, logp (2a) ≤ 2 logp a. So C1 f Nm 2 log (2a) ≤ 2C f N 2 log a, and
p 1 m p
C2 |T |(logp (2a))2 ≤ 4C2 |T |(logp a)2 ≤ 2C2 f (logp a)2 because |T | ≤ n = f /2.
The proposition does not show that the constants of the source satisfy (III.130). In the Lean the
comparison of the losses is lo_le_loG and fl_le_flG. They say: for every real number C∗ , the
104
Lean’s own losses lo and fl with the constant C∗ are at most loG and flG with the constant
4|C∗ | + 1. This is the source’s version of the conditions (III.130). (The sums in (III.122) are finite
because the capacities are ≥ 0. Proposition L10 was tested on 3000 random toy data, with H
cyclic of order 8, 16, 32, and again on 30170 exact cases: no violation, and the lower bound is
attained; without the loss terms the inequality fails in nearly all cases.) For the primes of class
inertClass the same computation applies to the orbits that are not served, but there Ex(p, t) is
not empty and the labels on these orbits do not belong to Tw . On the served orbits the source
uses gainsrv − Pr and pays τ (w); these orbits satisfy the conditions (G1)–(G3) of §III.3. That case
is assumed (AuxBlock8.Pf.stmtF, linkG_at_core). The terms pr(p) and or(p) of (III.78) occur
only there.
5. The sum over the primes
In this subsection the weights are ≥ 1, the record and its translate u are fixed, and gain(j, w) and
τ (w) depend only on pw . The set Tw of (III.79) also depends only on pw . From here on it is written
Tp , with p = pw .
Proposition L11 (splitting). Let Ξ be a finite set of primes of class signal or inertClass.
Suppose that V (j, t, p) = 0 for all j, t and all primes p ∈
/ Ξ of these two classes, and that
gain(j, w) = 0 (1 ≤ j ≤ m) and τ (w) = 0 at every place w with pw ∈
/ Ξ. (III.131)
Put def(p) = t∈T V (j, t, p) − j gain(j, p) − τ (p) . Then
P P P
j
m m X at
Pu,t,j − Gtop = 2 log p def(p) + log d − wtj cell(d, t, i) (III.132)
XX X XX
.
t∈T j=1 p∈Ξ t∈T j=1 i≥1
di +
Let Ξs ⊂ Ξ be a set of primes at which (III.79) holds with a set Tp ⊂ T , and suppose gain(j, p) =
τ (p) = 0 for p ∈ Ξ ∖ Ξs . Then
def(p) ≤ V (j, t, p) + m |Tp | loG (p) + m flG (p) (p ∈ Ξs ),
X X
j t∈T ∖Tp
(III.133)
def(p) = V (j, t, p) (p ∈ Ξ ∖ Ξs ).
XX
j t∈T
Proof. By (III.131) only the places above the primes of Ξ contribute to Gtop , and by (III.99),
Gtop = p∈Ξ 2 log p ( j gain(j, p) − τ (p)). With Lemma L6 and V∞ = V this is (III.132). For
P P
p ∈ Ξs split t∈T into Tp and its complement and insert (III.79).
P
The hypothesis (III.131) is needed: Gtop is a sum over all places, and Ξ contains only primes of
the two classes. In the source it holds at the places above the split primes, above 2, q, d, g0 , g1
and above the levels. The Lean field not_inert gives the vanishing at every place whose prime is
of neither of the two classes. That the level primes, among them g0 and g1 , are of class splitK
is another field (level_split). At g0 , g1 the gain vanishes also by the field gain_g. Without it
(III.132) fails: a gain of 1 at one place above a split prime changes the left side and not the right.
(Lean: SplitEq, SplitClasses. Tested on 15000 random data with (III.131): no failure. Without
it (III.132) fails in all 15000. If the labels outside Tp are forgotten, (III.133) fails.) So the prime
sum exceeds Gtop by four kinds of terms: the losses loG , flG at the primes where (III.79) is used;
the labels not served at those primes; the whole of V at the other primes; and the prime d.
Lemma L12 (one prime). For p ≥ 2, V (j, t, p) ≤ at p/(p2 − 1) ≤ 34 at /p. Proof. −i =
P
i odd p
p/(p2 − 1), and p2 /(p2 − 1) ≤ 43 .
105
The classes of primes in the sum. The source sorts the primes of class signal or inertClass
as follows (clAt, NoGain). Say that a prime p is in a band if p ≥ p0 and some power pi lies within
a factor 1 ± bw of X ′ , for a small number bw > 0 of the source. Below, Cr is a further constant of
the source (Lean: Cr; it is not the constant C of (III.77)). A candidate is an inertClass prime
p < X ′ whose class hp has order np ≥ xε/4 and which is not in a band. For a candidate put
Sig(p) = m κ(Q), the sum over the good orbits Q, those on which T satisfies (G1)–(G3) of §III.3
P
with the parameters there; κ(Q) ≥ 0 is a number attached to the orbit (kapv). The set of served
orbits Srv(p) of a candidate is the set of good orbits if Sig(p) ≤ x−ε , and is empty otherwise; for all
other primes Srv(p) = ∅. The set of served labels is
Tp = T if p is signal or p ≥ X ′ , Tp = {t ∈ T : the orbit of t is in Srv(p)} otherwise.
(III.134)
There are four classes. A prime belongs to the first class in the following list whose condition it
satisfies; read in this order the list is a partition. (As conditions, “high” and “first” overlap.)
1. high: p > X ′′ . The gain is 0. All weights are at least a/X ′′ (Lean: the field off_thin of
the structure Inst, with a/X ′′ = w′′ ; the other cells have larger weights, by off_top and
off_thick). Since moreover at ≤ a(1 + 1/(kx13 )), V vanishes unless p ≤ X ′′ (1 + 1/(kx13 )): a
short interval.
2. second: candidates with Sig(p) > x−ε . The gain is set to 0.
3. first: inertClass, p < X ′ (1 + Cr bw), with np < xε/4 or p in a band; and inertClass with
np odd and np > 1. The gain is set to 0.
4. served: all others. There (III.79) holds with Tp as in (III.134).
The class “served” contains the inertClass primes with hp = 0 and p ≥ X ′ (1 + Cr bw). Their gain
is 0 by (III.124), and V vanishes there, because at /p is below the weights of the cells (0, t), t ∈ T .
These weights exceed b′ = a/X ′ (Lean: the field bT_ge of SizeData2 for the label, and off_top of
Inst for the rest of the weight, which is at least b′ /x). The primes of the classes high, second and
first are the “dropped” primes of §2. The other primes need no gain: split primes have coefficient 0
in (4.1), and so do 2 and q. Every odd prime up to xε/(5Co ) splits, where Co is the constant of the
source that appears after Theorem III.3.2 (Lean: split_small, an assumed property of d, with
three_le_split; the Lean writes the exponent as ez/5, and ez is ε divided by a constant J of the
scale, so Co is J). The paragraph after Theorem III.3.2 writes Co ≥ 2. The Lean structure of the
scale has only 1 ≤ J (one_le_J). Whether J ≥ 2 on the route was not determined; the statement
that these primes split does not depend on it. So the sum starts above that bound. At the split
primes, at g0 , g1 and at the levels the numbers gain, τ , io are 0, and only ϕ and err occur. The
cut-offs have this meaning: X ′ = a/b′ is where a/p falls below the weights of the cells (0, t), t ∈ T ;
X ′′ is where it falls below the smallest weights; below p0 the source uses a different local argument.
Lemma L13 (a label that is not served lies on a bad orbit). Let p be a served prime
and t ∈ T ∖ Tp . Then p is of class inertClass, p < X ′ , np ≥ xε/4 , and the orbit of t under hp is
not good.
Proof. By (III.134), p is inertClass and p < X ′ . Since p is neither first nor second, np ≥ xε/4 ,
p is not in a band, and Sig(p) ≤ x−ε . So p is a candidate and Srv(p) is the set of good orbits.
(Lean: unserved_bad, which the final theorem uses by the seventh run of the dependency compu-
tation.) This is where the conditions (G1)–(G3) are used in the sum over primes. Since served
orbits are good by definition, the conditions are also available in the proof of (III.79) on the served
orbits (§4); that proof was not read.
106
Lemma L14 (the three small sums). Let Ξ1 be the set of inertClass primes p < X ′ with
np ≥ xε/4 .
(i) Let Lb , Cb , x0 and L2 be as in Theorem III.3.2, with L2 ≤ Lb , and let x ≥ x0 . Let T be the
set T of that theorem, for the set of primes Ξ1 . If x2ε/3 ≥ 16Cb and p 0 and Sig(p) ≥ 0 on a finite set Ξc of primes. Then
log p X log p 2 X log p
2 2 · [Sig(p) > θ] + 4 Sig(p) ≤ +4 Sig(p).
X
≤
p∈Ξ , Sig(p)>θ
p p∈Ξ
p θ p∈Ξ
p
c c c
(III.136)
Here [ · ] is 1 if the condition holds and 0 otherwise. Hence, if the middle sum is at most x−9 ,
the primes of the second class cost at most 83 m a′ |T | x−9 .
Proof. (i) Theorem III.3.2 bounds the left side by 4Cb x−2 δ1 n · 2xε/3 , and δ1 n ≤ 2|T |. So it is at
most 16Cb x−2+ε/3 |T | ≤ x−2+ε |T |. (ii) By Lemma L12 each V is at most 43 a′ /p; by Lemma L13 the
labels counted are those of (III.135). (iii) The first inequality is clear. For the second, [Sig(p) >
θ] ≤ Sig(p)/θ. The cost of the second class is, by Lemma L12, at most 38 ma′ |T | second logp p .
P
(Lean: ChBad_at is the inequality of Theorem III.3.2; ChBadp_at is (III.135); OutD is (ii); lhsW,
ChW_at, term_le, SecondD are (iii). The bound p 1 and ε > 0 are real numbers; Lsc = xε , b = eLsc , a = x b, La = log(a rdE ),
B0 = f m b xε , and Nc ≥ 0. (In the application f = 2n and Nc = |T2 | k b x2ε /n2 ; the proofs below
use only Nc ≥ 0.) As in §III.4, h(u) = u log u − u for u > 0 and h(0) = 0. So at log at − at means 0
when at = 0 (the convention 0 · log 0 = 0).
Index sets. J = {1, . . . , m} indexes the lifts. P is the set of weight places; it has at most f
elements. T ⊆ P is the set of distinguished places, with a real number at ≥ 0 for each t ∈ T . When
an element of T is called τ this number is written aτ . For a statement S the symbol [S] is 1 if S
holds and 0 otherwise.
The field and its places. K is the number field of the first remark above. In the source it is
totally complex. v runs over its archimedean places, s is their number, and ⟨u⟩ = s−1 v u(v) for a
P
real function u of v.
The chain and its end. A Hermitian norm at v on a K-vector space W means a positive definite
Hermitian norm on the complex space W ⊗K,v C. V0 , . . . , Vm are finite-dimensional K-vector
spaces with a Hermitian norm at each v, and Kj : Vj−1 → Vj (j ∈ J ) are K-linear maps. U is a
finite-dimensional K-vector space with a positive definite Hermitian norm at each v, and U0 ⊆ U
is a K-subspace. Ū = U/U0 , and ū ∈ Ū is the class of u ∈ U . Ū carries at each v the quotient
norm, ∥x∥ = min{∥u∥ : ū = x}, where x is in Ū ⊗K,v C and u runs over U ⊗K,v C (the minimum
is attained there). Y is a K-vector space with a Hermitian norm at each v, and ι : Ū → Y is a
K-linear bijection. Π : Vm → U ∗ is a K-linear map into the dual space (the pairing at the end
113
of the chain). Dual spaces carry the dual norms. The map of the inequality is the K-linear map
Φ : V0 → Y ∗ determined by
Φ(φ) ι ū = Π Km · · · K1 φ (u) (φ ∈ V0 , u ∈ U ). (III.141)
It exists as soon as the right side vanishes for u ∈ U0 ; this is part of input (S3). If Φ ̸= 0, its height
is ht(Φ) = ⟨log ∥Φ∥2v ⟩; §III.4 writes h(P ) for the height of a map P . In the source the chain is a
chain of matrix steps, and one step may contain several lifts (Lean: a map lf from the lifts to the
steps); the bound (S1) is then a bound for the matrix of a step. We state everything for one lift in
each step. The argument is the same.
Lattices and slopes. A lattice L is a finitely generated OK -module that spans a K-vector space
with norms at every v. µ(L) denotes the normalised slope of the source (muDef): twice the slope of
L for these norms, plus a constant depending only on K. It is the normalisation for squared norms.
L∨ is the dual lattice in the dual space, with the dual norms, and µmax (L) is the largest normalised
slope of a sublattice. R ⊂ Y is the lattice whose dual receives the values of Φ, and µCC = µ(R∨ ).
There is a second lattice R0 , in a space Y0 of the same kind, built from the same data at a lower
level; µ0 = µ(R∨0 ).
The record. The following real numbers are given: log T ≥ 0; log ΩE ; for j ∈ J and τ ∈ P
nonnegative kernel degrees νjτ i (1 ≤ i ≤ rν ) with sum njτ = i νjτ i , a number log Hjτ (in the
P
application the logarithm of the hook product H(A+ τ ) of step j), a number wjτ (the weight shift),
and two error terms ejτ , djτ ; for j ∈ J a number pj ; and parameters k, Nm , Sc , t0 , η, w′′ , b′ , λ0 , ϱ, L
and constants C0 , ce , Ch . (In the source the kernel degrees are padded by zeros to rν = Nm .) Put
hA = − at log T + Ch |T | b′ Nm + f (mη + Sc )Nm + f Nm (III.142)
X
2
La ,
t∈T
X X
hB,j = h(νjτ i ) − njτ log ΩE + Ch f Nm
2
La + m3 Lsc , (III.143)
τ ∈P i
hX,j = log Hjτ − Ch f ϱ2 L, (III.144)
X
τ ∈P
H = hA + (III.145)
X
hB,j − hX,j .
j∈J
These are the quantities hA , hB,j , hX,j of the paragraph “Relation to the Lean proof” of §III.4
(Lean: RecNums.hA, hB, hX, hu).
2. The inputs
(S1) The chain bound. For j ∈ J and every v, log ∥Kj ∥2v ≤ Σj (v), where
X X
Σj (v) = h(νjτ i ) − njτ log Ωτ (v) + |P| gj log Θ1 (v) (III.146)
τ ∈P i
+ |P| gj log 8 + 32 |P| gj vh + 2 log βj∗ (v) (III.147)
+ 1 + 2gj log(2njτ + 2) + log clj + log cl′j . (III.148)
X
τ ∈P
Here Ωτ (v) > 0 (the local squared period), Θ1 (v) > 0 (the mean square of the reference theta
function, per variable), βj∗ (v) > 0 (the constant of the kernel), clj , cl′j > 0 (numbers of classes),
114
gj and vh are real numbers, vh ≥ 0, with, for constants C2 , Cth , C4 , c5 , c6 , CT , C8 , Ce ≥ 0 and
a real number Lg (in the application log(g0 g1 )):
log βj∗ (v) ≤ C2 f m2 Nm
2
(Lg + log T ) + Ce f Nm
2
Lg , ⟨log Θ1 ⟩ ≤ Cth Lsc , (III.149)
log clj , log cl′j ≤ C4 f Nm
2
(Lsc + log T + Lg ), 0 ≤ gj ≤ Nm
2
, (III.150)
log 8 + 32|P|gj vh ≤ c5 La , log(2njτ + 2) ≤ c6 La , (III.151)
Lg ≤ C8 Lsc , log T ≤ CT mLsc , (III.152)
and Cth + c5 + 2c6 + 1 + 2C4 + 2Ce C8 ≤ Ch , 2(C2 + C4 )(CT + C8 ) ≤ Ch . Since gj ≥ 0 and
vh ≥ 0, the number 8 + 32|P|gj vh is at least 8; so its logarithm is defined. The condition
vh ≥ 0 is assumed here for this reason only. The record of the source does not contain it; in
the source’s construction vh = rdE /2 > 0 (see subsection 5).
(S2) The pairing at the end. There is a real number hΠ with ∥Π∥2v ≤ ehΠ for every v, and there
are real numbers ℓτ , λ1,τ , d0,τ , d1,τ , bτ , mτ , m′τ (τ ∈ P) and eΠ , CΠ ≥ 0 with
ℓτ log T + 2 mτ + m′τ + eΠ , ℓτ = λ1,τ − d0,τ , (III.153)
X X
hΠ ≤ −
τ ∈P τ ∈P
λ1,τ = aτ − 2 (τ ∈ T ),
Nm −k
λ1,τ ≥ Sc − Nm2−1 (τ ∈
/ T ), (III.154)
′
d0,τ ≤ Sc + mη + [τ ∈ T ] bτ , bτ ≤ 2b (τ ∈ T ), d1,τ ≤ 3d0,τ + Sc + η, (III.155)
mτ ≤ (Nm − 1) d1,τ log(2a), m′τ ≤ Nm (N2m −1) log(2a), 2
eΠ ≤ CΠ f Nm La , (III.156)
and 2(16 + CT + CΠ ) ≤ Ch . The number eΠ is one real number; it does not depend on
v. We read the first line as the bound of the paragraph “The pairing at the end of the
chain” of §III.4 at the place τ , with ψ ≤ Nm − 1, Ψ ≤ Nm and Γ ≤ 2a, and with mτ , m′τ
in the place of log |ϑ|, log κ there. This reading is ours: in the source mτ and m′τ are by
definition the two majorants (Nm − 1) d1,τ log(2a) and Nm (N2m −1) log(2a) (Lean: logc, logxi
in lastSide_of_logT_le).
(S3) The eigenvalues of a lift. The right side of (III.141) vanishes for u ∈ U0 . There are a finite
nonempty index set I, K-subspaces Ūp ⊆ Ū (p ∈ I) with p Ūp = Ū , and positive numbers
P
κv and λj,p (v) such that at every v
⟨ιx, ιx′ ⟩Y,v = κv λj,p (v) ⟨x, x′ ⟩Ū ,v (x ∈ Ūp , x′ ∈ Ū ). (III.157)
Y
j∈J
There are real numbers l6,j with l6,j ≤ log λj,p (v) for all j, p, v, and j hX,j ≤ j l6,j .
P P
(S4) The periods. For every τ ∈ P, ⟨log Ωτ ⟩ = log ΩE .
(S5) The slope at the lower level. There are real numbers l3,j (j ∈ J ), F and Eloc with
l3,j + 2F + Eloc . (III.158)
X
µ0 − ⟨log κ⟩ ≤
j∈J
Here l3,j is an upper bound for the normalised logarithmic covolume, divided by dim Y , of
the lattice attached to the jth lift; F is the sum of the local index terms; Eloc is the sum of
the local errors. This is the form of the source: its theorem has the bound by l3,j among its
hypotheses and the display as its conclusion (subsection 5).
(S6) The covolumes. There is a real number E3 with
pj wjτ log ΩE + E3 . (III.159)
X X X
l3,j ≤
j∈J j∈J τ ∈P
115
(S7) The passage between levels. There are real numbers de ≥ 0, la, Hmax , Q0 ≥ 1 with
µmax (R0 ) ≤ Hmax − ⟨log κ⟩, and such that, if de ≤ 13 and H∗ is a real number with
µmax (R0 ) ≤ H∗ , then
µCC ≤ µ0 + 3 de µ0 + H∗ + 28 la + 28 la. (III.160)
Moreover 3 de Q0 ≤ 1, j l3,j + 2F + Eloc + Hmax + 28 la ≤ Q0 , la ≤ λ0 and C0 ≥ 28. Here
P
⟨log κ⟩ is the mean of the numbers log κv of (S3), as in (S5). (The source uses the same
scalars κv for R0 and for R; this explains why one number occurs at both levels. It is a
comment, not a hypothesis.)
(S8) The size of the local errors. E3 + Eloc ≤ ce m xε (B0 + Nc ).
(W1) Rows against hooks. For all j, τ: i h(νjτ i ) − log Hjτ ≤ [τ ∈ T ] h(aτ ) + ejτ , with ejτ ≤
P
C0 wjτ + Sc + Nm
2 + m(t + 1)η L .
0 a
(W2) The total degree. For all j, τ : njτ = [τ ∈ T ] aτ + k wjτ + djτ , with |djτ | ≤ C0 wjτ + k(Nm +
Sc ) + m(t0 + 1)η + [τ ∈ T ] k b′ .
(G) Sizes of the parameters. rdE ≥ 1, log rdE ≤ Lsc , log x ≤ Lsc ; m, f, k ≥ 1, k + m ≤ Nm ;
Sc , t0 , η, b′ , λ0 , ϱ, L ≥ 0, ϱ ≤ Nm ; 0 ≤ wjτ ≤ 2b; |pj − k| ≤ m; | log ΩE | ≤ La ; and
mλ0 ≤ b, 2
Nm ≤ b, m3 Nm
2
Lsc ≤ b, m η Nm Lsc ≤ b, kNm ≤ b, (III.161)
′′ ′′ ′′ ′ ′
kSc ≤ w , m(t0 + 1)η ≤ w , Nm w ≤ b, b Nm ≤ a, m k b ≤ a, (III.162)
1 ≤ C0 ≤ ce , 0 ≤ Ch ≤ C0 , Ch L ≤ C0 La . (III.163)
Remarks on the list. (1) (W1) and (W2) follow from “Row by row”. In the source ejτ and djτ
are defined by equality in the first halves of (W1) and (W2) (Lean: eEN, dN). Fix j and τ and write
w = wjτ . From its inputs (o), (a)–(e) the paragraph “Row by row” of §III.4 derives
k + (2j − 1)(t0 + 1)
ejτ ≤ (w + ∆ + B) La , |djτ | ≤ k(ς + ∆) + w + B, ς= , (III.164)
2
where ∆ ≥ 0 is the width of the weight interval and B bounds the extra degrees. (At a distinguished
place it gives (w + B)La for the first.) Three facts are taken from the source (Lean: the fields
constant_weight_sum_bound, constant_succ_le, blk of NumFacts, and Nm_cast): 2∆ ≤ C1 Sc
for a constant C1 with C1 + 1 ≤ C0 ; B = m(t0 + 1)η; and Nm = k + m(t0 + 1). So ∆ ≤ C0 Sc ,
because Sc ≥ 0, and ς ≤ Nm , because 2j − 1 ≤ 2m. With w ≥ 0, B ≥ 0, k ≥ 1 and C0 ≥ 1 from
(G),
w + ∆ + B ≤ w + C0 Sc + B ≤ C0 w + Sc + Nm
2
+ m(t0 + 1)η , (III.165)
k(ς + ∆) + w + B ≤ kNm + C0 kSc + w + B ≤ C0 w + k(Nm + Sc ) + m(t0 + 1)η . (III.166)
Multiply the first line by La ≥ 0. These are the bounds of (W1) and (W2); the term [τ ∈ T ] k b′ of
(W2) is not needed. So (W1) and (W2) hold under the inputs (o), (a)–(e) of “Row by row” and the
three facts. In the source this is stepE_num, used in RecNums.ofNum. The theorem below assumes
(W1) and (W2) as stated.
(2) In the source pj = k + (2j − 1 − m) (Lean: ClosedScale3.l3m), so that |pj − k| ≤ m − 1.
(3) The condition C0 ≥ 28 stands in (S7) because the source has it there (Lean: the field
constant_ge_twenty_eight). The proof uses it once, in Step 3, for 28 la ≤ C0 m λ0 . It is not an
extra condition: 2(16 + CT + CΠ ) ≤ Ch of (S2) and Ch ≤ C0 of (G) give C0 ≥ 32. Two other
conditions are also used once each: k + m ≤ Nm of (G) only through m ≤ Nm in Step 2, and
C0 ≤ ce only for the last display of the theorem.
116
3. The assembly
Theorem III.7.1. Assume (S1)–(S8), (W1), (W2) and (G). Then:
(a) if Φ ̸= 0, then ht(Φ) ≤ H − ⟨log κ⟩;
(b) with U1 = |T | a La and U2 = m xε (B0 + Nc ),
X
µCC − ⟨log κ⟩ + H − 2F ≤ m h(at ) − at log ΩE − at log T
t∈T (III.167)
+ 2C0 U1 + 47C0 + 7 + ce U2 .
In particular, with h = H − ⟨log κ⟩, F0 = 2F and C = max(50C0 + 10, 60ce ),
X
µCC + h − F0 ≤ m at log at − at − at log ΩE − at log T
t∈T (III.168)
+ C |T | a La + C m B0 x 2ε
+ C m Nc x . 2ε
The last display is the inequality (III.75) at the end of §III.4. In it at log at means 0 when at = 0,
since h(0) = 0; part (b) is written with h(at ). The constant C is at least 1610, because C0 ≥ 32.
So the inequality has content only when m is large compared with C: otherwise the error term
C |T | a La can exceed the main term. The source states (b) with the error (50C0 + 10) U1 + 60ce U2
(Lean: RecNums.rhs); this is weaker than (b), since 47C0 + 7 + ce ≤ 55ce . The terms h and F0 of
the final inequality are thus H − ⟨log κ⟩ and 2F. The term F0 enters in two places. In (S5) it is
added in the bound for the slope, and it is subtracted on the left. It also occurs in the condition of
(S7) that bounds Q0 from below. The proof takes the rest of this subsection. It has four steps: a
bound for ∥Φ∥v at one place; the average over the places; the slope; the cancellation.
Step 1: a map that factors through a quotient.
Lemma (descent). Fix a place v and extend all spaces to C by the embedding at v. Assume
the first two sentences of (S3). Then
∥Π∥2v j∈J ∥Kj ∥2v
Q
∥Φ∥2v ≤ . (III.169)
κv minp∈I j∈J λj,p (v)
Q
Proof. Write sp = κv j λj,p (v) and smin = minp sp . Let Ξ be the operator on Ū ⊗ C with
Q
⟨ιx, ιx′ ⟩Y = ⟨Ξx, x′ ⟩Ū for all x, x′ . It is Hermitian. The hypothesis says ⟨Ξx, x′ ⟩ = sp ⟨x, x′ ⟩ for
x ∈ Ūp and all x′ , that is, Ξx = sp x on Ūp . The spaces Ūp ⊗ C span Ū ⊗ C, so Ξ is diagonalisable
with all eigenvalues among the sp . Eigenspaces of a Hermitian operator for different eigenvalues are
orthogonal. Hence ∥ιx∥2Y = ⟨Ξx, x⟩ ≥ smin ∥x∥2Ū for every x. Now let y ∈ Y ⊗ C and put x = ι−1 y.
Choose u with ū = x and ∥u∥U = ∥x∥Ū ; this is possible by the definition of the quotient norm. By
(III.141),
∥y∥Y
|Φ(φ)(y)| = Π(Km · · · K1 φ)(u) ≤ ∥Π∥ . (III.170)
Y Y
∥Kj ∥ ∥φ∥ ∥u∥U ≤ ∥Π∥ ∥Kj ∥ ∥φ∥ √
j j
smin
The norm of Φ(φ) in the dual of Y is the supremum of |Φ(φ)(y)|/∥y∥Y .
(Lean: HeightBoundConstant, proved as height_constant_bound and used in height_chain_atR;
there for a chain of matrix steps, as said in subsection 1.)
117
Assume now (S1), (S2), (S3) and Φ ̸= 0. Then ∥Φ∥v > 0 at every v, because a nonzero entry of
the matrix of Φ stays nonzero under an embedding of K. Taking logarithms in the lemma and
using l6,j ≤ log λj,p (v),
log ∥Φ∥2v ≤ hΠ + Σj (v) − log κv − for every v. (III.171)
X X
l6,j
j∈J j∈J
Step 2: the average over the places.
Lemma (the height record). Assume the numerical parts of (S1) and (S2), the identity
(S4), and (G). Then
hΠ ≤ hA and ⟨Σj ⟩ ≤ hB,j (j ∈ J ). (III.172)
Proof. From (G): Lsc ≥ 1 since x > 1; La = log x + Lsc + log rdE , so 1 ≤ Lsc ≤ La ≤ 3Lsc ;
0 ≤ log(2a) ≤ 2La ; and m ≤ Nm .
The first inequality. For τ ∈ T put D̄τ = Sc + mη + 2b′ , and for τ ∈
/ T put D̄τ = Sc + mη; then
d0,τ ≤ D̄τ and d1,τ ≤ 3D̄τ + Sc + η ≤ 4Sc + 4mη + [τ ∈ T ] 6b . Since (Nm − 1) log(2a) ≥ 0 and the
′
last bound is nonnegative,
2 mτ ≤ 16Nm (Sc + mη) + [τ ∈ T ] 24 b′ Nm La , m′τ ≤ Nm
2
(III.173)
La .
Next, −ℓτ log T = (d0,τ − λ1,τ ) log T and 0 ≤ log T ≤ CT mLa . For τ ∈ T the coefficient is
−aτ + Nm2−k +d0,τ ≤ −aτ + N2m +Sc +mη+2b′ ; for τ ∈
/ T it is at most Nm2−1 −Sc +Sc +mη ≤ N2m +mη.
With m ≤ Nm this gives
2
Nm ′
−ℓτ log T ≤ −[τ ∈ T ] aτ log T + CT 2 + Nm (Sc + mη) + [τ ∈ T ] 2b Nm La . (III.174)
Sum over the at most f places τ , of which |T | are in T , and add eΠ ≤ CΠ f Nm
2L :
a
at log T + (24 + 2CT )|T | b′ Nm + (16 + CT )f (mη + Sc )Nm + 1 + C2T + CΠ f Nm
X
2
hΠ ≤ − La .
t∈T
(III.175)
Each of the three coefficients is at most 2(16 + CT + CΠ ) ≤ Ch . This is hΠ ≤ hA .
The second inequality. Average Σj (v) over v. By (S4) the first sum becomes τ ( i h(νjτ i ) −
P P
njτ log ΩE ), exactly. For the other terms, using |P| ≤ f , 0 ≤ gj ≤ Nm 2, 1 ≤ L
sc ≤ La and
Lg + log T ≤ (C8 + CT )mLsc :
|P| gj ⟨log Θ1 ⟩ ≤ Cth f Nm
2
La , (III.176)
|P| gj log(8 + 32|P|gj vh ) ≤ c5 f Nm
2
La , (III.177)
1 + 2gj log(2njτ + 2) ≤ (1 + 2c6 )f Nm
2
(III.178)
P
τ La ,
2⟨log βj∗ ⟩ ≤ 2C2 (CT + C8 )f m3 Nm
2
Lsc + 2Ce C8 f Nm
2
La , (III.179)
log clj + log cl′j ≤ 2C4 f Nm
2
La + 2C4 (CT + C8 )f m3 Nm2
Lsc . (III.180)
The sum of the coefficients of f Nm2 L is C + c + 2c + 1 + 2C + 2C C ≤ C , and that of
a th 5 6 4 e 8 h
f m Nm Lsc is 2(C2 + C4 )(CT + C8 ) ≤ Ch .
3 2
(Lean: height_bound_at_set, applied through HtNums.le_hu. A parallel statement for the first
inequality, stepDiiiRecord, is in the archive but is not used by the final theorem.)
Proof of Theorem III.7.1(a). Average (III.171) over v and apply the lemma and the last
inequality of (S3):
ht(Φ) ≤ hΠ + l6,j −⟨log κ⟩ ≤ hA + hX,j −⟨log κ⟩ = H −⟨log κ⟩. (III.181)
X X X X
⟨Σj ⟩− hB,j −
j j j j
118
This is the first place where log ΩE enters: through (S4), each hB,j contains − τ njτ log ΩE . The
P
hook products enter through (S3), as the subtracted terms hX,j .
Step 3: the slope. Put µ′0 = µ0 − ⟨log κ⟩ and µC = µCC − ⟨log κ⟩. By (S5),
µ′0 ≤ l3,j + 2F + Eloc , (III.182)
X
j
and then by (S6) and (S8),
µ′0 ≤ pj wjτ log ΩE + 2F + Eb , Eb = E3 + Eloc ≤ ce m xε (B0 + Nc ). (III.183)
X X
j∈J τ ∈P
This is the second place where log ΩE enters. For the passage, 3 de Q0 ≤ 1 and Q0 ≥ 1 give de ≤ 13 .
Apply (S7) with H∗ = Hmax − ⟨log κ⟩ and subtract the number ⟨log κ⟩ from both sides:
µC ≤ µ′0 + 3 de µ′0 + Hmax + 28 la + 28 la. (III.184)
By (III.182) and (S7), µ′0 + Hmax + 28 la ≤ Q0 ; as de ≥ 0, the middle term is at most 3 de Q0 ≤ 1.
Also 28 la ≤ 28λ0 ≤ C0 mλ0 . So
µC ≤ µ′0 + Lpass , Lpass ≤ 1 + C0 m λ0 . (III.185)
Step 4: the cancellation.
Lemma (cancellation). Assume (W1), (W2), (G), and let µC , µ′0 , F, Eb , Lpass be real numbers
satisfying (III.183) and (III.185). Then
X
µC + H − 2F ≤ m h(at ) − at log ΩE − at log T + 2C0 U1 + (47C0 + 7 + ce ) U2 . (III.186)
t∈T
Proof. Add (III.185), (III.183) and the definition of H:
µC + H − 2F ≤ Lpass + Eb + hA + EK + EX (III.187)
X
+ pj wjτ log ΩE + h(νjτ i ) − njτ log ΩE − log Hjτ , (III.188)
X
j,τ i
where EK = m Ch f Nm 2 (L + m3 L ) and E = m C f ϱ2 L. By (W1) and (W2) the summand is at
a sc X h
most
[τ ∈ T ] h(aτ ) − aτ log ΩE + ejτ + (pj − k) wjτ log ΩE − djτ log ΩE . (III.189)
Here the two occurrences of the period cancel: +pj wjτ log ΩE from the slope against −k wjτ log ΩE
from the height. What is left of them is (pj − k)wjτ log ΩE . The sum over j ∈ J and τ ∈ P of the
first bracket is m t∈T (h(at ) − at log ΩE ), and hA contributes − t at log T . So the left side is at
P P
most the main term plus
E1 + · · · + E8 , (III.190)
where
E1 = Lpass , E2 = Eb , E3 = EK , E4 = EX , (III.191)
′
E5 = Ch |T |b Nm + f (mη + Sc )Nm + f Nm
2
E6 = (III.192)
P
La , j,τ ejτ ,
E7 = j,τ |pj − k| wjτ | log ΩE |, E8 = j,τ |djτ | | log ΩE |. (III.193)
P P
119
Put V = f m b Lsc . From (G): Lsc ≥ 1, b ≥ 1, La ≤ 3Lsc ; Sc ≤ kSc ≤ w′′ ≤ Nm w′′ ≤ b and
Sc Nm ≤ b; m(t0 + 1)η ≤ w′′ ≤ b; b ≤ V; and
U2 ≥ m xε B0 = m Lsc V ≥ mV ≥ V. (III.194)
There are m values of j and at most f values of τ . Then:
E1 ≤ 1 + C0 b ≤ (1 + C0 )V; (III.195)
E3 ≤ C0 mf (b · 3Lsc + b) ≤ 4C0 V; (III.196)
2
E4 ≤ C0 mf Nm La ≤ 3C0 V (by Ch L ≤ C0 La , ϱ ≤ Nm );
(III.197)
′
E5 ≤ C0 |T |aLa + C0 f 3b + 3bLsc + 3bLsc ≤ C0 U1 + 9C0 V (by b Nm ≤ a, mηNm Lsc ≤ b);
(III.198)
E6 ≤ mf C0 (2b + b + b + b) 3Lsc = 15C0 V; (III.199)
E7 ≤ mf · m · 2b · 3Lsc = 6mV; (III.200)
′
E8 ≤ 15C0 V + m|T | C0 k b La ≤ 15C0 V + C0 U1 (by kNm ≤ b, kSc ≤ b, mkb′ ≤ a).
(III.201)
Finally E2 ≤ ce U2 . The sum is at most 2C0 U1 + (47C0 + 1)V + 6mV + ce U2 ≤ 2C0 U1 + (47C0 +
7 + ce ) U2 .
(Lean: divisor_inequality_explicit, used through divIneqRecord and RecNums.divIneq; it
is divisor_inequality_assembly, the exact rearrangement, followed by error_absorption, the
estimates of E1 , . . . , E8 . The source’s constants are 50C0 + 10 and 60ce .)
Proof of Theorem III.7.1(b). Steps 3 and 4. For the last display of the theorem: h(at ) =
at log at − at ; 2C0 ≤ C; 47C0 + 7 + ce ≤ 55ce ≤ C because 1 ≤ C0 ≤ ce ; and U2 = m xε (B0 + Nc ) ≤
m x2ε (B0 + Nc ) because x > 1.
Where each term comes from.
Term of the inequality Source in the proof
m h(at ) (W1): at a distinguished place the hook product cancels all rows of the
kernel but the first, once for each of the m lifts
−m at log ΩE (S4) and (W2): the part [τ ∈ T ] aτ of the degree njτ , times the mean
period, for each lift
−at log T (S2): the level factor of the last map, through ℓτ = λ1,τ − d0,τ and
λ1,τ = aτ − Nm2−k
cancelled: k wjτ log ΩE −k wjτ log ΩE from (S4), (W2) in the height; +pj wjτ log ΩE from (S6)
in the slope; remainder E7
C |T | a La E5 (the term |T | b′ Nm of the end pairing) and E8 (the term k b′ of the
degree error at distinguished places, times | log ΩE |)
C m B0 x2ε E1 (level passage), E3 (constants of the kernels, class numbers, reference
mean square), E4 (loss in the eigenvalue bound), the rest of E5 , E6 (rows
against hooks), E7 (the periods cancel only up to |pj − k| ≤ m), the rest
of E8 ; and the part of E2 with B0
C m Nc x2ε E2 only: the local errors E3 + Eloc , by (S8)
Among the terms of the second kind, E7 ≤ 6mV is the only one whose bound carries the factor
m; so it is the only one that needs the full factor m in U2 ≥ mLsc V. It is the cost of pj ̸= k. Its
bound is not the largest in general: the bounds for E6 and E8 are 15C0 V, and 6m > 15C0 only
when m > 2.5 C0 .
120
4. The mean of the periods of the integral lines
This subsection proves a formula for a mean of periods (the Proposition below). It does not prove
(S4); what is missing is listed at the end. In this subsection ρ denotes an embedding of a field M
into C, as in the paragraph “The reference mean square” of §III.4, and τ , τ ′ denote weight places.
√
Let kd = Q( −d), embedded in C. Its discriminant is −d, because d is squarefree and
d ≡ 3 mod 4 (the d of the paper is a prime with d ≡ 7 mod 8); Lemma A uses this. Let E ⊃ kd
be the field of the construction, of degree [E : kd ] = f . The weight places are the f embeddings
τ : E → C that extend the given embedding of kd . So here P has exactly f elements; subsection 3
used only |P| ≤ f . Let D be the different of E over kd , and let the bar be complex conjugation on
E.
A fact about the field E. There is√
an integral ideal b of E with D = b b̄. This holds for the field
−1 √
of the paper, E = Q(ζN + ζN , −q, −d), where N = 4n, n is a power of 2, and q is a prime with
q ≡ 3 mod 4 that divides d + 1. The proof has two steps.
First step: every prime of E that divides D lies over 2 or over q. (This is (P2) below.) Put
−1 √ √
F = Q(ζN + ζN , −q). Then E = kd F . Since q ≡ 3 mod 4, the number −q lies in Q(ζq ) (the
quadratic Gauss sum). So F ⊆ Q(ζN q ), and F is unramified over Q at every prime number that
does not divide 2q. Let Q be a prime of E that does not lie over 2 or q. An extension that is
unramified at a prime stays unramified when the ground field is enlarged; so E = kd F is unramified
over kd at Q. A prime that is unramified does not divide the different. (For n = 1 the field F is
√
Q( −q), and only primes over q can divide D.) The Lean has this step as diffSupport.
Second step. 2 and q split in kd , because d ≡ 7 mod 8 and −d ≡ 1 mod q. Complex conjugation
exchanges the two primes of kd over 2, and the two over q. So a prime Q of E that divides D and
its conjugate Q̄ lie over different primes of kd , and Q̄ =
̸ Q. Conjugation maps E to E and kd to kd ;
so D̄ = D, and Q, Q̄ occur in D with the same exponent. Take for b the product of one prime
power from each pair. (Lean: principal in KcDiff.lean, from diffSupport and kc_below_ne;
it is used through Arith.bb_ex to fill the field b_spec of the record Host.)
This b is coprime to b̄, as is the ideal of the source. There are other integral choices when an
exponent e is larger than 1: one may take Qi Q̄e−i with 0 ≤ i ≤ e from a pair. The Proposition
below holds for every b with D = b b̄; its proof uses nothing else. The coprimality is not used in
this section.
The splitting of 2 and q in kd is sufficient for the existence of b. It is not necessary. For a
field E with complex multiplication over an imaginary quadratic field, an integral b with D = b b̄
exists if and only if every prime of D that is equal to its conjugate has an even exponent. (If
Q̄ = Q, the exponent of Q in b b̄ is twice its exponent in b. Conversely, halve the exponents of
these primes√ and√treat the other
√ primes in pairs as above.) So the statement fails for some fields:
for E = Q( −7, −3) over Q( −7) the different is the prime over 3, with exponent 1, and it is
its own conjugate.
Fix such a b and put a0 = b−1 . Let M ⊂ C be a number field that contains τ (E) for every τ
and a field of definition of the elliptic curve of §III.4, with lattice Λ normalised by g23 − 27g32 = 1.
For an embedding ρ of M , Aρ is the area of the lattice of the conjugate curve, as in §III.4.
Let M ⊂ M f be the OM -module spanned by the vectors (τ (α))τ , α ∈ a0 , and let
lτ = {z ∈ M : z eτ ∈ M}, (III.202)
where eτ is the τ th coordinate vector. This is the integral line at the place τ (Lean: ellT). For a
nonzero z ∈ lτ the local squared period at (τ, ρ) is Aρ /(π|ρ(z)|2 ), and the mean of its logarithm is
121
defined, with a correction for the index, by
1 X Aρ
log Ωτ = log + 2 log : (III.203)
lτ OM z .
[M : Q] ρ π |ρ(z)|2
(Lean: meanLogOm, logMet.) If lτ = OM zτ is principal, this is the plain mean over ρ of
log Aρ /(π|ρ(zτ )|2 ) .
Proposition (the mean period). The number (III.203) does not depend on z, on τ , or on
M , and
1 Aρ rdE
log Ωτ = log − log √ . (III.204)
X
[M : Q] ρ π d
Under the inputs (I2), (I3) of §III.4, it equals γ + log(2π/rdE ) − cd = log ΩE .
The proof uses four statements about ideals. N is the absolute norm of ideals of E.
Lemma A (the tower formula). |disc E| = d f N(D). Consequently [E:Q]
1
log N(D) =
√
log rdE − log d.
Proof. Cited: the different is multiplicative in towers, DE/Q = D · Dkd /Q OE , and the absolute
norm of the different of a number field is the absolute value of its discriminant (J. Neukirch, Algebraic
Number Theory, 1999, Ch. III, (2.2)(i) and (2.9)). The norm of Dkd /Q OE is N(Dkd /Q )[E:kd ] = df ,
since the discriminant of kd is −d. For the second statement divide the logarithm by [E : Q] =
2f .
Lemma B (the integral line). For a nonzero fractional ideal c of E let Mc be the OM -module
spanned by the vectors (τ ′ (α))τ ′ , α ∈ c. Then the line {z ∈ M : z eτ ∈ Mc } is the OM -module
generated by τ (c D). In particular lτ = τ (a0 D)OM = τ (b̄)OM .
Proof. Let c∗ = {α∗ ∈ E : TrE/kd (α∗ c) ⊆ Okd } = c−1 D−1 (the definition of the different),
and let M∗ be the OM -module spanned by the vectors (τ ′ (α∗ ))τ ′ , α∗ ∈ c∗ . Use the bilinear
form ⟨m, m′ ⟩ = τ ′ mτ ′ m′τ ′ on M f . For α ∈ c and α∗ ∈ c∗ the pairing of their vectors is
P
τ ′ τ (αα ) = TrE/kd (αα ) ∈ Okd ⊆ OM . (Every τ restricts to the given embedding of kd ; so the
P ′ ∗ ∗ ′
map α 7→ (τ ′ (α))τ ′ is kd -linear, and the sum over τ ′ is the trace to kd .) We claim that Mc is exactly
the set of m with ⟨m, m′ ⟩ ∈ OM for all m′ ∈ M∗ . It is enough to check this after localising at each
prime of OM . Above a prime p of kd the module cp is free over the local ring of kd , with a basis
α1 , . . . , αf , and c∗p has the dual basis α1∗ , . . . , αf∗ for the trace form. The corresponding vectors ai ,
ai∗ of M f satisfy ⟨ai , ai∗′ ⟩ = δii′ , so they are two bases of M f , dual to each other, and they generate
the localisations of Mc and M∗ . This proves the claim. Now z eτ pairs with the vector of α∗ to
z τ (α∗ ). So z eτ ∈ Mc if and only if z τ (α∗ ) ∈ OM for all α∗ ∈ c∗ , that is, if and only if z lies in
the inverse of the ideal τ (c∗ )OM , which is τ (c D)OM .
Lemma C (the index). For nonzero z ∈ lτ , [lτ : OM z] = |NM/Q (z)| · N(b̄)−[M :Q]/[E:Q] .
Proof. For a nonzero ideal A of OM and z ∈ A one has [A : zOM ] = |NM/Q z|/NM (A), by
multiplicativity of the norm. By Lemma B, A = lτ is the extension to M of the ideal τ (b̄) of τ (E),
and the norm of an extended ideal is the norm raised to the power [M : τ (E)].
ρ log |ρ(z)|
2 = 2 log |N
Lemma D (the product formula). For nonzero z ∈ M , M/Q (z)|.
P
Proof. NM/Q (z) = ρ ρ(z), the product over all embeddings.
Q
Proof of the Proposition. By Lemmas C and D the terms with z cancel:
2[M : Q]
2 log[lτ : OM z] − log |ρ(z)|2 = − log N(b̄). (III.205)
X
ρ [E : Q]
122
In particular a unit, or any other change of z, does not change (III.203). Since N(b) = N(b̄) and
D = bb̄, 2 log N(b̄) = log N(D). By Lemma A,
1 Aρ 1 1 Aρ rdE
log Ωτ = log log N(D) = log − log √ . (III.206)
X X
−
[M : Q] ρ π [E : Q] [M : Q] ρ π d
The first term does not depend on M . Indeed, let M0 ⊆ M be a field that contains the coefficients
g2 , g3 of the curve. The number Aρ depends only on the restriction of ρ to M0 , and every embedding
of M0 extends to exactly [M : M0 ] embeddings of M . So the mean over the embeddings of M is the
mean over the embeddings of M0 , and two fields M are compared through their compositum. √ This
proves the first assertion. By (I2) and (I3) of §III.4 the mean of log Aρ is log π + γ + log(2π/ d) − cd .
Subtracting, log Ωτ = γ + log(2π/rdE ) − cd , which is log ΩE by its definition in §I.7.
√
So the constant ΩE differs from the elliptic Ω by the factor N(D)1/[E:Q] = rdE / d. Two things
give this factor. The first is Lemma B: the integral line at a weight place is generated by τ (a0 D),
not by τ (a0 ). The second is the choice a0 = b−1 , for which N(a0 D)2 = N(D). For an arbitrary
nonzero fractional ideal a0 the same proof gives
1 Aρ 2
log Ωτ = log log N(a0 D); (III.207)
X
−
[M : Q] ρ π [E : Q]
√
with a0 = OE the factor would be (rdE / d)2 . (Lean: meanPeriodAt, from mean_log_period,
br_meanLogOm_eq and two_log_brN; Lemma B is lineR_eq, Lemma C is index_lineR, Lemma
D is br_sum_log_norm_sq, Lemma A is natAbs_discr_eq. The source proves Lemma A from the
same two facts.)
What the Proposition proves, and what it does not. It proves that the number (III.203), which
is defined by the ideal a0 and the areas Aρ , equals log ΩE , under (I2), (I3) and the fact about
E. Input (S4) is a different statement. It concerns the numbers Ωτ (v) of (S1), averaged over the
archimedean places v of K. To deduce (S4) from the Proposition one needs three further statements.
(P1) and (P3) are assumed, from the source, and not proved here. (P2) was proved above:
(P1) the period that occurs in the block estimate of §III.4 for the kernel conjugated by ρ, at the
weight place τ , is Aρ /(π|ρ(zτ )|2 ) for a generator zτ of lτ (Lean: logOmAt, ExtraTie.om_eq).
Generators need not exist over a given M . An example: M = E, q = 3, d = 23, N = 4, 8 or
16 (the class number of E is 3, 6, 60). There the ideal τ (b̄)OE of Lemma B is not principal
for any of the choices of b coprime to b̄. This was computed with Sage 10.9, with class
groups computed in its mode proof=False; so the example is correct under GRH and was
not checked without it. Nothing in this section depends on the example. It depends on the
choice of b: for N = 8 two of the eight integral b, which are not coprime to their conjugates,
give a principal line. The existence of generators over a suitable larger field is the hypothesis
IsDgen of mean_logMet_eq;
(P2) every prime of E that divides D lies over 2 or over q (Lean: diffSupport). This is the first
step of the fact about E above; it is proved there for the field E of the paper. With it the
ideal b exists;
(P3) the mean over the embeddings ρ of M can be replaced by the mean over the places v
of K: the source takes M ⊇ K and for log Ωτ (v) an average, with weights of sum 1, of
the numbers log Aρ /(π|ρ(zτ )|2 ) , the weights having the same total for every ρ (Lean:
kerSide_of_meansS, exists_weights).
123
5. The remaining inputs: statements and status
Each statement below is assumed in this section. We give its form in the source and say what a
proof needs. The Lean names are collected in subsection 6.
(S1), the chain bound. The analytic content is proved in §III.4: the bound for one block
with the period (“Several blocks, and the period”), the bound for one map (“Sums over classes”),
and the bound for the Gaussian constant by a mean square. Compare with the formula for log bj (v)
and with the constant S of (III.51) there. |P| gj is the number g = 2nN N ′ of variables, with N ,
′
N ′ the ranks of the pair as in §III.4. βj∗ (v) is the prefactor β c2[E:Q]N N , and clj , cl′j count classes.
Θ1 (v) is a mean square per variable. In §III.4 the reference mean square at an embedding ρ is
′
Θ0,ρ = Hρf N N , with Hρ the mean square of one elliptic theta function. It enters S 2 once, with
⟨log Θ0,ρ ⟩ ≤ f N N ′ Lsc . In (S1) the same term is written |P| gj log Θ1 (v), with ⟨log Θ1 ⟩ ≤ Cth Lsc .
So Θ1 corresponds to Hρ , not to Θ0,ρ . What is assumed: that the maps of the construction are
instances (item (vi) of the list in §III.4), and the eight numerical bounds listed under (S1). Of
these, the bound for ⟨log Θ1 ⟩ is the conclusion of the paragraph “The reference mean square” of
§III.4 (conditional on (I1)–(I5)), divided by the number of variables; the bounds for βj∗ and for the
numbers of classes are statements about the finite data of the construction (coefficients, modulus,
class numbers of Hermitian lattices of bounded level). A proof needs a bound for class numbers of
definite Hermitian lattices in terms of the mass and the level. The source’s statements were not
read for this section, with one exception, the sign of vh . In the record of the source vh is a real
number with no condition on its sign (Lean: the field vv of HtNums, and the same in KerSide); the
logarithm of the Lean is defined for every real number, so the record needs none. Where the source
builds the record, vh = rdE /2 (Lean: NumDatum.v), and vh ≥ 0 is proved from rdE ≥ 1 inside the
proof of the bound with c5 (log_eight_leS). The source’s statement of the analytic bound for
one map (LiftedMapEmbeddingLogBound) has vh > 0 among its hypotheses. So vh ≥ 0 is not a
hypothesis of the record, and it is a consequence of the construction.
(S2), the pairing at the end. The inequality ∥Π∥2v ≤ ehΠ , with hΠ bounded as in the
first line of (S2) with eΠ = 0, is the conclusion of the paragraph “The pairing at the end of the
chain” of §III.4, with its own inputs; there the two terms are bounded by log |ϑ| ≤ ψ d1 log Γ and
log κ ≤ Ψ(Ψ−1) 2 log Γ. The source does not name log |ϑ| and log κ: its numbers mτ , m′τ are these
two majorants with ψ = Nm − 1, Ψ = Nm and 2a in the place of Γ. The identification of d0,τ , d1,τ ,
ℓτ , λ1,τ , bτ with the d0 , d1 , ℓ, λ1 , bt of §III.4 was confirmed in the Lean (lastSide_of_logT_le).
The remaining conditions of (S2) express the exponent ℓτ = λ1,τ − d0,τ and the degrees d0,τ , d1,τ
through the weights of the construction: at a distinguished place the first weight is aτ up to the
shift (Nm − k)/2. They are statements about the weights only; this is the “substitution of the
actual weights” of item (ii) of the list in §III.4. A proof needs the weights of the last two spaces of
the chain as explicit functions of the data. In the source eΠ = 0 and CΠ = 0.
(S3), the eigenvalues. This is the paragraph “The eigenvalues of a lift” of §III.4, with its
hypotheses (E1)–(E7). The source’s evaluation of the scalar is the following statement (Lean:
scalar_bound_lowC; among the declarations the final theorem depends on). Let
− ∥ϕ∥2 Y
λ = D0 · vol(F ) · τ )H(Aτ ) · (III.208)
Y
H(A+ · Jτ · |E0 |,
τ DS c τ
where D0 ≥ 1 is a product of dimensions of representations; vol(F) is the volume of a fundamental
set for the level subgroup of the definite group; ∥ϕ∥2 is the squared L2 -norm of the finite test
function for a measure in which the rational lattice has covolume one; Jτ is the integral, over the
real unitary group at τ , of the squared absolute value of a matrix coefficient of a unit vector of the
124
lowest type for the maximal compact subgroup; E0 is an Euler product over the prime ideals P of
E outside a finite set, with parameters of absolute value at most (NP)(r−1)/2−(1+t0 )/2 , as in (E5);
and DS c is the normalising factor outside that set. Assume the level-index identity (the index of
−1
the level subgroup times ∥ϕ∥2 equals 1/DS ), the lower bound wEϱ ϱ! for the mass, with wE the
√ 2
number of roots of unity of E, and the bound − log Jτ ≤ 2 + r log 2 2 r ϱ(dτ + 2) . Then λ > 0
2
and
−
log λ ≥ τ ) + log H(Aτ ) − 8f ϱ L,
log H(A+ L = log b rdE g0 g1 ϱ . (III.209)
X 2
τ
This is the deduction written in §III.4, with one difference. The mass bound in (III.63) is
−1 −1
(8f 2 )ϱ ϱ! ; here and in the Lean it is wEϱ ϱ! . The second implies the first, because wE ≤ 8f 2 .
The reason: write eul(u) for Euler’s function of an integer u ≥ 1. The field E contains Q(ζwE ),
which has degree eul(wE ) over Q. So eul(wE ) divides [E : Q] = 2f , and eul(wE ) ≤ 2f . Next,
u ≤ 2 eul(u)2 for every u ≥ 1: the quotient eul(u)2 /u is the product of the numbers pe−2 (p − 1)2
over the prime powers pe that divide u exactly, and such a number is at least 43 for p ≥ 3 and
at least 12 for p = 2. Hence wE ≤ 2(2f )2 = 8f 2 . The Lean has this inequality as a theorem,
for the field E of its record Host: host_torsionOrder_le, which states Units.torsionOrder
H.E ≤ 8f 2 . The inner-product formula itself is used in the following form (Lean: the hypothesis
InnerProductFormulaC of term_eval_lowC): for theta lifts θ(ϕi , fi ), to the definite group, of forms
fi on the group G lifted from,
Z
θ(ϕ1 , f1 ), θ(ϕ2 , f2 ) = vol(F) ⟨ω(g)ϕ1 , ϕ2 ⟩ ⟨g · f1 , f2 ⟩ dg, (III.210)
G(A)
the integral being written as a sum over finitely many cosets at the finite places and an integral
over the real points, absolutely convergent. This is the inner-product formula in the convergent
range, as an integral of matrix coefficients; it is not the form with a special value of an L-function,
and no Siegel–Weil formula appears in the statement. For the classical statements see S. Rallis,
L-functions and the oscillator representation, Lecture Notes in Math. 1245 (1987) [Ral87] and
Injectivity properties of liftings associated to Weil representations, Compositio Math. 52 (1984);
J.-S. Li, Non-vanishing theorems for the cohomology of certain arithmetic quotients, J. reine
angew. Math. 428 (1992) (the formula as an integral of matrix coefficients, in the stable range);
W. T. Gan, Y. Qiu, S. Takeda, The regularized Siegel–Weil formula (the second term identity)
and the Rallis inner product formula, Invent. Math. 198 (2014); S. Yamana, L-functions and
theta correspondence for classical groups, Invent. Math. 196 (2014). All from memory; none was
consulted. A proof of (S3) needs: this formula for the actual test functions; the unramified
computation of the local integrals (the factors of E0 ); the mass formula for definite unitary
groups (for instance [GHY01]; its formula is for maximal lattices, and local comparisons would
be needed for the levels used here); the formal degree of the discrete series at the real places
(Harish-Chandra); the transport to the other embeddings (E7); and the description of the target
by Hecke components (E6). The source has a theorem that derives the inner-product formula
(innerProductFormula_of_hypotheses) from four hypotheses. Three are statements named
Published.CuspRapidDecay, WeilSiegelFormulaUnitary and SiegelEisensteinContinuation.
The fourth is the existence of a choice of parameters (Lean: a ParameterChoice’ exists for some
SizeData2 ). Each of the three named statements has a closed theorem in the source. Their
proofs were not read. The dependency computation shows that the final theorem depends on
innerProductFormula_of_hypotheses; the three named statements were not put to it.
(S5), the slope at the lower level. In the source this is a theorem (IndefSide.boundsMu,
from slope_bounds). Its conclusion is the display of (S5), with the sum of the l3,j . Its hypotheses
125
are the following. The first is the covolume bound: the normalised logarithmic covolume of the
lattice of the jth lift, divided by dim Y , is at most l3,j (Lean: the field l3 of IndefSide; in the
source the covolume is minus the normalised slope). The others are all at finite places. For a
finite place y of K write wt(y) = log N(y)/[K : Q], ram(y) for its ramification index and rc(y)
for its residue characteristic. There are a finite set Sfin of places containing the set Stw of places
where a twist is made, and real numbers gd(y), ix(y, j), err(y, j) ≥ 0, jx(y, j), et(y, j), such that:
the valuation at y of a Gram determinant of the quotient lattice is 0 outside Sfin and at most
dim Y · gd(y) on Sfin ; on Stw the sum over j of the two local indices of the lattice of the jth lift (at
y and at ȳ) is at most j (2 ram(y) ix(y, j) + ram(y) err(y, j)) + 2 ram(y)/rc(y) dim Y ; a further
2
P
local index is at most ram(y) jx(y, j) dim Y + et(y, j); outside Stw , for each j, the sum of these
three local indices (the two at y and ȳ and the further one) vanishes, and so does ix(y, j) (Lean:
IndefData.Untwisted; it asks for the vanishing of the sum, not of each index). Then (S5) holds
with F as follows, and with every Eloc that is at least the sum written below:
F= wt(y) ram(y) ix(y, j), (III.211)
X X
j y∈Stw
et(y, j)
Eloc = wt(y) gd(y) + wt(y) ram(y) err(y, j) + jx(y, j) + (III.212)
X X X
y∈Sfin j y∈Stw
dim Y
X 2 wt(y) ram(y)
+ . (III.213)
y∈Sfin
rc(y)2
The display gives Eloc in the case of equality. In the source the four sums of Eloc are not computed:
each is bounded by a free real number (Lean: LG, E1, E2, E3), and the conclusion has the sum of
these majorants. The Lean E3 is the last sum of Eloc . It is not the E3 of (S6), which the Lean calls
El3. It is a statement of the geometry of numbers over K: the covolume of a lattice glued from
local lattices, compared with the covolumes of the lattices of the lifts through the scalar relations
of (S3) (which is where ⟨log κ⟩ enters) and through local indices. No period occurs in it. The term
F is the term F0 /2 of Part II, §6.7. The local index bounds are separate inputs; those at the places
of Stw are of the kind treated in §III.5. The proof in the source was not read.
(S6), the covolumes. In the source l3,j = τ pj wjτ log ΩE + C3 B0 log log x by definition,
P
with a constant C3 (Lean: ClosedScale3.l3m, l3u) and with pj = k + 2j − 1 − m. So (S6) holds
with E3 = m C3 B0 log log x (levelThree_bound_closed). The content is therefore in the first
hypothesis of the theorem quoted under (S5), the covolume bound (IndefSide.l3): the normalised
logarithmic covolume of the lattice of the jth lift, divided by dim Y , is at most τ pj wjτ log ΩE up
P
to C3 B0 log log x. This is the count of forms of Part II, §6.6 in its arithmetic form: the lattice is a
space of algebraic automorphic forms of weight governed by wjτ , its metric is the Petersson norm,
and the period appears because integral forms are normalised by the algebraic differential, with the
same integral lines lτ as in subsection 4. That the constant here is the same ΩE as in (S4) is exactly
what makes the cancellation in Step 4 possible; it is assumed, not proved, in this section. A proof
needs the covolume of the lattice of integral forms on a definite unitary group for the Petersson
metric (mass formula and dimension formula, as in (6.4) of Part II), with the period normalisation
of subsection 4. The source derives it from a statement named CovolumeBoundThirdPlaceSeries
(linked_bundles_l3_of_open_alt), whose hypotheses include statements on discrete series and
on Harish-Chandra modules; it was not read beyond its header. One of its hypotheses is a covolume
bound named CovolumeBoundOfX. The source proves that one (topCovolSlim2_use_alt_alt)
from statements of algebraic and Arakelov geometry, which it takes as hypotheses: among them
resolution of singularities and principalisation in characteristic zero, a Poincaré–Lelong formula,
an arithmetic Hilbert–Samuel statement, a bigness statement named after Moriwaki, toroidal
126
compactifications, and a list named PublishedHodgeBundleTheorems. Only the headers of these
two theorems were read.
(S7), the passage between levels. Two statements: the bound µmax (R0 ) ≤ Hmax − ⟨log κ⟩
with Hmax = xCH b for a constant CH (Lean: IndefExtraR.fun0), and the comparison of µ(R∨ )
with µ(R∨ 0 ) displayed in (S7) (IndefExtraR.passage). In the source la is the logarithm of the
order of a finite group, and de = 8f r/ min(g0 , g1 ) with r = k + m − 1. The number Q0 is max(1, · )
of a majorant of the left side of its defining inequality: any upper bound for j l3,j + 2F + Eloc , plus
P
Hmax + 28 la. The condition 3 de Q0 ≤ 1 then says that the level primes g0 , g1 are at least 24 f r Q0 .
(The source proves an inequality of this form, 24 f r Q ≤ min(g0 , g1 ) for its number QOfR, from
g0 , g1 ≥ (kb)8 and further size conditions: hg_of_sizeG, SizeG. Only the statement was read.)
So the passage costs a bounded amount because the level primes are large. The comparison is
a statement about two lattices of forms of different level and the trace map between them; the
first bound is a bound for the largest slope of a lattice of forms. Proofs were not read; the source
reaches them through statements on Fourier coefficients (fittedFunction_at_zero_close_alt).
(S8), the size of the local errors, and (G). These are inequalities between the parameters.
(S8) contains the only occurrence of Nc . With E3 = m C3 B0 log log x as above, and with the
bounds for wt(y) gd(y), err, jx, et supplied by the local statements, it holds for large x. The
part of Eloc with et(y, j) is where the bound for Galois cohomology of Part II, §6.8 is used (Lean:
divinN_of_hypV); this was read in the notes of another reader and not checked here. (G) is a
list of size conditions on the parameters of the chain, valid for x ≥ x0 in the source; among them
| log ΩE | ≤ La follows from |cd | ≤ 4 log x (Theorem III.1.1(d)) for large x.
(W1), (W2). See remark (1) after the list of inputs: they follow from the conclusions of “Row
by row”, which are conditional on the inputs (o), (a)–(e) there, and from three facts of the source.
6. Assumed here, from the source
The Lean statements corresponding to the inputs. “Dependency: yes” means that the name is
among the declarations the final theorem depends on, by the dependency computation; “reached”
means that the name is not in the lists computed, but is used in the proof or in the definition of a
name that is, either directly or through further declarations that are not in the lists. The route is
given each time.
• divClausesR_of (dependency: yes): the assembly; its third clause is Theorem III.7.1(b) with
the error (50C0 + 10) U1 + 60ce U2 , its second clause is Theorem III.7.1(a) in the form of a
majorant at each place.
• (S1): fields cb, nKcb of DivInputs0R and i2–i10, k2, k3 of HtNums: each matrix of the chain
is a sum over classes of blocks with squared norm at most a bound whose logarithm is at most
Σj (v). chain_coupled_bound (yes) turns this into log ∥Kj ∥2v ≤ Σj (v).
• (S2): field nP of DivInputs0R and t1–t11, k1 of HtNums.
• (S3): fields l1b, l1c, l6 of IndefColsR and eig6 of DivInputs0R; scalar_bound_lowC
(yes). The source has a theorem named lamLow_of_classicalTheorems; among its hy-
potheses are the Siegel–Weil formula for unitary groups with its convergence, the doubling
method, and the unramified doubling computation for GL (weilSiegelFormulaUnitary,
siegelWeilUnitaryConvergent, doublingMethod, unramifiedDoublingGL). By its name it
gives the lower bound for the scalars. That it is what fills the field l6 was not verified.
127
• (S4): field i1 of HtNums; meanPeriod (yes), mean_log_period (yes), chowla_selberg_eta
(yes). Reached: two_log_brN and br_meanLogOm_eq, through meanPeriodAt, which is the
body of meanPeriod and is not in the lists; index_lineR, through br_meanLogOm_eq; lineR_eq,
through index_lineR. The ideal b: principal, Arith.bb_ex, field b_spec of Host (depen-
dency: not queried. The computation was asked for the name KcDiff.principal. It compares
the last component of a name only, so this name cannot match, and the empty answer says
nothing).
• (S5): IndefSide.boundsMu (yes), slope_bounds (yes), with the fields bi, bLG, biv,
borderedMinorImage, twisted_third_sum_le of DivInputs0R, on and l3 of IndefSide and
untw of IndefExtraR.
• (S6): field L3main of DivInputs0R.
• (S7): fields level, passage, fun0 of IndefExtraR; fields de0, deQ, Q1, Qb, la_le,
constant_ge_twenty_eight of DivInputs0R.
• (S8): field errB of DivInputs0R.
• (W1), (W2), (G): fields weight_height_sum_le, weight_sum_eq, sE, sδ, p1–p38, c1–c5 of
RecNums; for remark (1), NumFacts, stepE_num, RecNums.ofNum.
• Step 1: height_chain_atR (yes); height_constant_bound (reached directly: it is
used in the proof of height_chain_atR). Step 2: height_bound_at_set (yes). Step
4: divisor_inequality_explicit (yes), divIneqRecord (yes), error_absorption (yes);
divisor_inequality_assembly (reached through divisor_inequality, which is not in the
lists and is applied by divisor_inequality_explicit).
• The identifications h = H − ⟨log κ⟩, F0 = 2F, µCC = µ(R∨ ) with the quantities of the final state-
ment of the source: hTie_of (yes) and the hypotheses tMu, tFo, tR of wordFacts_of_clauses.
In the source the function h(u) of the final statement is defined by the formula for H − ⟨log κ⟩
(hExpl6); so it already contains −⟨log κ⟩.
7. What this section proves, and what it assumes
Proved: Theorem III.7.1: the inequality (III.75) at the end of §III.4 follows from (S1)–(S8), (W1),
(W2), (G), with the constant C = max(50C0 + 10, 60ce ); the descent lemma; the two lemmas on real
numbers, with every constant; the table of the origin of each term. The Proposition of subsection
4: the mean (III.203) of the logarithms of the periods of the integral lines is log ΩE , given the class
mean formula (I2), (I3) of §III.4, the multiplicativity of the different and the formula for its norm
(cited), and the ideal b; Lemmas B, C, D are proved. The existence of b for the field E of the paper,
with (P2), on which it rests. Remark (1) after the list of inputs: (W1), (W2) from the conclusions
of “Row by row” and three facts of the source.
Assumed: (S1)–(S8), (W1), (W2), (G). The input (S4) is not proved: it would follow from the
Proposition and (P1)–(P3), and (P1), (P3) are assumed from the source. Of the inputs, (S1), (S2),
(S3), (W1), (W2) correspond to conclusions of conditional paragraphs of §III.4, each with inputs
of its own that are named there. For (S1) and (S2) the exact match of the numerical forms (the
constants of (S1), the weights in (S2)) is part of what is assumed. For (W1), (W2) the match is
made in remark (1), given the three facts quoted there.
128
8. What remains
(i) The covolume bound with the period ΩE of subsection 4 (the first hypothesis of the theorem
behind (S5), with the values of l3,j in (S6)): the statement that the lattice of integral forms
of the jth lift has normalised covolume at most τ pj wjτ log ΩE up to the stated error. This
P
is the largest gap: the cancellation of the period is proved here only as an identity between
two inputs.
(ii) The slope bound (S5) and its local index bounds, including the definition of F.
(iii) The passage between levels (S7).
(iv) The inner-product formula, the mass, the formal degrees and the local factors, for (S3).
(v) The statements (P1) and (P3) on the periods of the actual kernels, and item (vi) of the list of
§III.4.
(vi) The verification of (G) and (S8) for the parameters of Part II for large x, and the bounds for
the constants and the class numbers in (S1).
III.8 The images in general position: the trace-formula estimates
This section concerns the two smallness statements that the linear algebra of Part II, §6.9–6.10
consumes, and the identity that connects them with the pairings of the chain of lifts. In the Lean
development these are three statements: LinkS (bounds for the Gram blocks of a family of maps;
the trace statement), LinkIRR (a count; the rank statement), and BesselId (the pairing at the
end of the chain is an inner product with the image of such a map; the pairing statement). The
title needs a word of explanation. The source does not prove that the images of the maps are
nearly orthogonal. The images cannot all be mutually orthogonal: by the trace statement their
dimensions add up to more than the dimension of the target (Remark (1) of subsection 2). What
the source bounds is this: every block Bw† Bw′ with w ̸= w′ has at most (1 + η0 )(1 + η1 ) times
the mean size that it has for subspaces in general position, where η0 and η1 are the numbers of
Statement (S) of subsection 1, at most exp(−xε/8 ). So the operator Q = w Bw Bw† is close to its
P
mean eigenvalue in Hilbert–Schmidt norm (Remark (1) of subsection 2). The operator Q is not a
sum of projections. Proved here: Lemma T1 and Corollary T2. This section deduces the dimension
bound from the three trace, rank and pairing statements and the size and subspace conditions in
T2, with explicit constants. It also proves Lemma T3, a statement about finite sets and weights
which is the mechanism of a second moment in the simplest case (a sum over twists of Bχ† Bχ is
a count of pairs of representations, and its main term is a mass). Quoted from the source: the
three statements themselves, transcribed in subsection 1, and the theorems from which the source
deduces them, listed in subsection 4 with their Lean names. The construction of the spaces and
maps, the arithmetic inputs and the assembly of (M) are quoted. None of these is proved here.
The error term that Part II, §6.10 takes from this section is recalled at the end of subsection 2.
Below, §1, §2, . . . without a part number mean the subsections of this section.
Letters introduced in this section: W, w (words); H, Cw , Mw , E0 , E, Uw (spaces), with
Mw,L , EL (spaces over a number field L); Pw , Bw , Kww′ , s, Bw , βw , Ψ, Ψw , Λ, πE (maps and
pairings); dw , dram w , D∗ , Dr (dimensions); η0 , η1 , e0 , δp , δE , κ, αram , α , εz , εC (small numbers); G, CG ,
′
C0 , CH , Jz , c0 (w), λw (constants); Wg, T (w), ξi , ψi , θj , ej , ∆ (in §2, §4, §5); and, in §3 only,
′
I, I ′ , ω, ω ′ , Gi , Rep, A, inv, Θχ , Bχ , N , N ◦ , N◦ , m, σ, ℓ0 , ςι . The letters Q, D, S, e, η, α, u, H have
the meaning of Part II, §6.9–6.10: Q is the sum of the operators Bw Bw† , D the dimension of the
129
target, S the total dimension of the sources, e, η, α the three constants of the lemma of §6.9, and u,
H the spread and the height bound of §6.10. From Corollary T2 on, the share of directions lost at
the level primes is written αram ; it is one part of the constant α of the lemma, which in Corollary T2
is αram + α′ , and it is not the α = u/(u + t) of §6.10. The target space, called H in §6.9, is written
H here. The letters k, m, t0 , f, r, g0 , g1 , q, d, b have the meaning of §III.4 and §III.5: k is the initial
rank, m the number of lifts, t0 the integer parameter written t in §III.4 and t0 in §III.5 (the field t
of the Lean record of ranks), f = 2n the number of weight places, r = k + m − 1, g0 , g1 are the two
level primes, q, d the auxiliary primes, and b = exp(xε ) with 0 < ε < 1/52. ∥A∥2HS = tr(A† A) is the
squared Hilbert–Schmidt norm, and A† the adjoint. Inner products are conjugate-linear in the first
variable and linear in the second. V ∨ is the space of C-linear forms on a complex space V . The
letter ν denotes a dimension where it occurs, and j, j ′ , l, ι are indices. The exponential function is
written exp in this section, because e is a constant of the lemma of §6.9.
Other letters. Used in more than one place: φ and F (elements of Mw or Mw,L , and of E0 or
EL ); x0 (a threshold for x) and c, c′ , C, C ′ (positive constants), which are not the same at each
occurrence; y (a real number in the proof of Lemma T3, and xε/8 in §4); i (an index in ci , ξi , ψi ,
an element of I in §3, and the imaginary unit in iF in §1); v (a finite place in (R1) and (R2), a
positive real number in (R3)); X (a tuple of vectors in (R2), the variable of a polynomial in (R4));
g (a group element in (R2), a prime in (R4)); M (a rank, in (R2) and (R3)). Used in one place
only and explained there: s (the inner mean of §2, a real number; it is not the map s); X , Z, Y, h
(in the proofs of Lemma T1(d) and Corollary T2); Ξ, ζ, ζ ′ , p (in the Remarks); Y , ρ, χ (in §3); τ ,
K (τ ) , K ′′ , s0 (in (R1)); E/E + , J, Z, Φ, ϱv (in (R2)). The number e′0 of §5 is the first of the e′j ; it
is not the e0 of (S).
The letters are local. No letter of this section was changed to agree with the other sections of
Part III, and several have another meaning there. The clashes most likely to confuse are these. C0
is 3 here; in §III.7 it is a constant with C0 ≥ 28. H is the target space here; in §III.7 it is a height.
G is the number exp(CG m5 f xε ) here and in Part II, §6.10; in §III.7 it is a group. T (w) is a level
prime here, as T is in Part II, §6.7; in §III.6 T is a set. ςι are minors here (§3); in §III.4 and §III.7
ς is a number formed from the ranks. Λ is a map here; in §III.6 the Λw are lattices, and elsewhere
Λ is the von Mangoldt function. Ξ is a random unitary map here; in §III.6 it is a set of primes and
in §III.7 an operator. Ψ, Ψw are unitary maps and ψi weights here; in §III.4 Ψ = ψ + 1 and ψ are
ranks, and in §III.9 Ψ is a character. ∆ is the step of the weights here (§5); in §III.4 it is the width
of the weight interval. Also: α is a constant here and a member of the family in §III.6 (the word of
this section; see §1); B and c0 occur in §III.9, κv in §III.7 and Y in §III.5, with other meanings.
1. The three statements in the source
The data. At a scale x the source works with a record (Lean: D7) which contains, among many
other things, the following. They are fields of the record. Their identification with spaces of
automorphic forms is made in the source by further hypotheses, which were not read (see §6).
• A finite set W of words. Each word w has a sequence of letters in {0, 1}k−2 , and distinct
words have distinct sequences (letter, letter_inj). A word fixes, at each of the f weight
places, a weight ξ = (ξ0 > · · · > ξk−1 ) of a unitary group in k variables; see §5.
• A complex inner product space H of finite dimension D (HH, Dall), and a real number D∗ > 0
(Dstar). The positivity of D∗ is not a field of the record. It follows from two fields of other
records: Dstar_eq and Dstar_pos.
• For each word a complex inner product space Cw of finite dimension dw (Cu, duS), a subspace
130
Mw ⊂ Cw of dimension dram w (Mram, du), a linear map Pw : Cw → H (Pu), and a number
c0 (w) ≥ 0 (c0, a product of two volumes divided by a dimension). When c0 (w) > 0 put
Bw = c0 (w)−1/2 Pw , Kww′ = Bw† Bw′ : Cw′ → Cw . (III.214)
When c0 (w) = 0 the definitions of the source give Kww = 0. In the notation of Part II, §6.9
the maps Bi are the Pw , the spaces Vi are the Cw , and the weights are ci = 1/c0 (w) (the letter
ci is that of §6.9; the number c0 (w) of the record is not one of the ci ); so Ki = Kww and
Q = w Bw Bw† .
P
• A complex inner product space E0 of dimension Dr (EE), and a conjugate-linear map s : E0 → H
which preserves norms (sEE). So s(cF ) = c̄ s(F ) for c ∈ C, and s is injective. Its image E = s(E0 )
is a complex subspace of H of dimension Dr . There is a number δp ≤ 1 with Dr = (1 − δp )D
(Drows, delp, D_eq). The spaces E0 and E have the same elements up to s, with conjugate
actions of the scalars. They are kept apart below.
• For each word a pairing Bw : Mw × E0 → C, linear over C in each variable (Bu, Bu_linL,
Bu_linR). In the source it stands for the pairing at the end of the chain of m lifts of Part II,
§6.2, between the space at the start of the chain and the space at its end (it is tied to the
chain by the fields perCol and PtilCol, which were not read). Both spaces carry structures
over a number field L ⊂ C: there are L-vector spaces Mw,L and EL , with lattices in them, and
maps Mw,L → Mw and EL → E0 , linear over L (jN, jEE), under which an L-basis of Mw,L
is a C-basis of Mw and an L-basis of EL is a C-basis of E0 (jN_basis, jEE_basis: linear
independence and span). The second map goes to E0 and not to E: composed with s it would
be conjugate-semilinear over L, and so linear over L only if L ⊂ R. In particular
dimL EL = Dr (III.215)
(RRV_finrank). There is an L-bilinear pairing Balg
w : Mw,L × EL → L (Bk). The slopes of
Part II, §6.8–6.10 are those of these lattices.
Comparison with the letters of §III.6. The following was checked in the types of the Lean record.
The field L is the coefficient field L0 of §III.6 (LL, with the embedding embLL). The words are
the members α of §III.6 (Word). The space Mw,L is the space Uα that carries the glued lattice
Λgl
α (Nu), and EL is the space U with the lattice R (RR): the maps jN and jEE have the sources
′
(div.Nu u).V and div.RR.V. Not checked in the Lean: that Balg w (Bk) is the pairing ⟨ , ⟩α of §III.6
(perCol), and that the level prime T (w) of §2 is the prime Lvα of §III.6; both sections describe
them in the same words. The letters of §III.7 were not compared. The index of the family is the
word. It is not a Hecke character: in the source the twists by Hecke characters form another index
(Tup), with one twist attached to each place (muAt). By our reading it serves to glue the lattices
(Part II, §6.8). The space E0 and the pairings Bw do not depend on it: Tup does not occur in the
types of EE and Bu, nor in LinkS_at, Frame_at and frame_at_of.
Translates. There is one record for each translate of Part II, §4.2, and each record has its own
set of words (Lean: the definition data in QuadraticLogDerivCongr.lean, whose family of index
sets is fun u0 => (Fm.D x (M.pt u0)).what2 .Word; there u0 is the translate and u the word).
The weights of the table of Part II, §4.2 depend on the translate and not on the word (w u0 j).
The level prime, the height and the rank depend on the translate and on the word (T u0 u, h u0
u, d u0 u).
Statement (S) (Lean: LinkS_at, LinkS, with C0 = 3). Put G = exp(CG m5 f xε ), where
CG > 0 is a constant of the family. There is x0 such that for every x ≥ x0 and every record of the
131
family at the scale x there are real numbers η0 , η1 and e0 ≥ 0 with
√
η0 ≤ exp(−xε/8 ), η1 ≤ exp(−xε/8 ), e0 ≤ exp(−xε/8 ), (III.216)
such that, if
C0−1 ≤ dw G2 /D∗ ≤ C0 for every word w, (III.217)
then
dw dw′
∥Kww′ ∥2HS ≤ (1 + η0 ) (w ̸= w′ ), (III.218)
D∗
∥Kww − 1∥2HS ≤ e0 dw , (III.219)
D ≤ (1 + η1 ) D∗ . (III.220)
In the proof in the source the three numbers are η0 = b−6 , e0 = 2C0 /G + 2 exp(−kf log k) and
η1 = exp(−c rf log r) with a constant c > 0 (linkS_concl). So they are of size exp(−c′ xε ), and
(III.216) is a weaker, uniform way of recording this. When (III.217) fails for a record, that proof
takes η0 = η1 = e0 = 0, and the implication is empty (linkS_of_standing_hypotheses). The
source proves (S) for families satisfying a further condition on the level primes of the words (TuSize).
Statement (I) (Lean: LinkIRR). There is x0 such that for x ≥ x0 and every record of the
family at the scale x
0 ≤ δp ≤ exp − xεz /7 , εz = ε/Jz , (III.221)
where Jz ≥ 1 is a constant of the family. So E is almost all of H.
Statement (P) (Lean: BesselId). For every record: (P1) for φ ∈ Mw,L and F ∈ EL ,
Bw (φ, F ) = Balg
w (φ, F ), the right side being read in C through the fixed embedding of L, and φ, F
on the left being their images in Mw and E0 ; (P2) there are a unitary map Ψ of E0 and, for each
word, a complex number λw ̸= 0 and a unitary map Ψw of Mw , such that
Bw (φ, F ) = λw sΨF, Pw Ψw φ H (φ ∈ Mw , F ∈ E0 ). (III.222)
Both sides are linear in φ and in F . For the right side and F : replace F by iF . Then ΨF is
multiplied by i, since Ψ is linear; sΨF is multiplied by −i, since s is conjugate-linear; and the inner
product is multiplied by i, since it is conjugate-linear in its first variable. The map s cannot be
omitted: for a complex subspace E of H, a pairing that is linear on Mw × E cannot be equal to
λw ⟨ΨF, Pw Ψw φ⟩ with Ψ unitary on E, unless it is zero: the left side would be multiplied by i and
the right side by −i. In the Lean, sEE has the type of a conjugate-linear isometry from EE to the
target, the pairing besS is inner (sEE F) (Pu ...), and BesselId asserts Bu u phi F = a *
besS u (U phi) (V F) with linear isometric equivalences V of EE and U of the space of φ. In the
source (P) is a clause of the statement LinkDIV_at. The name refers to the realisation of ⟨sF, Pw φ⟩
in the source as a Bessel period of F ; in the form (III.222) no integral appears.
2. From the three statements to the hypotheses of Part II, §6.9
The lemma of Part II, §6.9 has three hypotheses, with constants e, η, α: a bound i ∥Ki −1∥2HS ≤ eS
P
for the diagonal blocks; a bound tr(Q2 ) ≤ (1 + η)(tr Q)2 /D for the operator Q; and subspaces of
total codimension at most αS. Its conclusion is that the images span all of the target but a part of
√
dimension at most (η + 4α + 4 e)D. This is the form that the proof in §6.9 gives. The display
of §6.9 has 2η in place of η, which is weaker. This section uses the form with η everywhere, and
132
so does the source (FrameLA1; see below). The next lemma obtains the first two hypotheses from
(III.218)–(III.220).
Lemma T1. Let H be a complex inner product space of dimension D ≥ 1, let W be a finite
nonempty set, and for w ∈ W let Bw : Cw → H be a linear map from an inner product space of
dimension dw ≥ 1. Put Kww′ = Bw† Bw′ , Q = w Bw Bw† and S = w dw .
P P
(a) tr Q = w tr Kww and tr(Q ) =
2
w,w′ ∥Kww′ ∥HS .
2
P P
(b) Let D∗ > 0, η0 ≥ −1, e0 ≥ 0, and assume (III.218) and (III.219). Then
S2 √ √
tr(Q2 ) ≤ (1 + η0 ) + (1 + e0 )2 S, tr Q ≥ (1 − e0 )S, ∥Kww − 1∥2HS ≤ e0 S.
X
D∗ w
(III.223)
(c) Assume in addition (III.220), and let 0 < κ ≤ 41 be such that
√
D∗ ≤ κS, η0 ≤ κ/2, η1 ≤ κ/2, e0 ≤ κ/2. (III.224)
Then
(tr Q)2
tr(Q2 ) ≤ (1 + 5κ) , 1
e0 ≤ 64 . (III.225)
D
(d) Under the hypotheses of (c), let Uw ⊂ Cw be subspaces with w (dw − dim Uw ) ≤ αS, let E ⊂ H
P
be a complex subspace of dimension (1 − δE )D with 0 ≤ δE < 1, and let πE be the orthogonal
projection onto E. Then
√
5κ + 4α + 4 e0
dim E − dim dim E. (III.226)
X
πE Bw Uw ≤
w 1 − δE
†
Proof. (a) The first identity is tr(BB † ) = tr(B † B). For the second, Kww ′ = Kw′ w , so
tr Bw Bw† Bw′ Bw† ′ = tr Bw† ′ Bw Bw† Bw′ = tr Kww
†
′ Kww′ = ∥Kww′ ∥HS ,
2
(III.227)
and tr(Q2 ) is the sum of these over all pairs.
(b) Let K be an operator on a space of dimension ν. By the Cauchy–Schwarz inequality for
√ √
the Hilbert–Schmidt inner product, | tr(K − 1)| ≤ ν ∥K − 1∥HS , so Re tr K ≥ ν − ν ∥K − 1∥HS ;
√ √
and ∥K∥HS ≤ ν + ∥K − 1∥HS . With (III.219) this gives tr Kww ≥ (1 − e0 )dw (the trace is real)
√
and ∥Kww ∥2HS ≤ (1 + e0 )2 dw . Sum over w and use (a). For the pairs w ̸= w′ use (III.218) and
w̸=w′ dw dw′ ≤ S ; here 1 + η0 ≥ 0 is used.
P 2
(c) By (III.220), 1 + η1 ≥ D/D∗ > 0 and 1/D ≥ 1/((1 + κ/2)D∗ ). By (b) and (III.224),
tr Q ≥ (1 − κ/2)S ≥ 0, and, since S ≤ κS 2 /D∗ ,
S2
tr(Q2 ) ≤ (1 + κ2 ) + (1 + κ2 )2 κ , (III.228)
D∗
(tr Q)2 (1 + 5κ)(1 − κ2 )2 S 2
(1 + 5κ) ≥ · . (III.229)
D 1 + κ2 D∗
So it suffices that (1 + κ2 )2 + (1 + κ2 )3 κ ≤ (1 + 5κ)(1 − κ2 )2 . Expanding both sides, the difference of
the right side and the left side is
κ 2 − 13
2 κ + 2κ − 8κ ,
1 2 1 3
(III.230)
133
which is positive for 0 < κ ≤ 14 , since then the bracket is at least 2 − 13
8 − 512 > 0. Finally
1
e0 ≤ κ /4 ≤ 1/64.
2
(d) By (b) and (c) the hypotheses of the lemma of Part II, §6.9 hold for the maps Bw , with all
weights ci equal to 1, with e = e0 ≤ 14 , η = 5κ, and the given α. So, by the form of that lemma
with η which its proof gives, the orthogonal complement Z of X = w Bw Uw in H has dimension
P
√
at most (5κ + 4α + 4 e0 )D. Now πE (X ) = (X + E ⊥ ) ∩ E, and the codimension of this space in E
equals the codimension of X + E ⊥ in H, because X + E ⊥ = πE (X ) ⊕ E ⊥ and H = E ⊕ E ⊥ . This
codimension is at most that of X , which is dim Z. Finally D = dim E/(1 − δE ).
Remarks. (1) The blocks are not small, and the images are not nearly orthogonal. For every
family, tr(Q2 ) ≥ (tr Q)2 /D, by the Cauchy–Schwarz inequality for the eigenvalues of Q. If all Bw
are isometries, then Kww = 1, and this and (a) give
S2
(III.231)
X
∥Kww′ ∥2HS ≥ − S.
w̸=w′
D
The sum of the bounds dw dw′ /D over the same pairs is (S 2 − w d2w )/D, which exceeds the right
P
side by at most S, that is, by the relative amount D/S at most. By (III.220), 1/D∗ ≤ (1 + η1 )/D.
So (III.218) and (III.220) assert that the squared norm of every off-diagonal block is at most
(1 + η0 )(1 + η1 ) times dw dw′ /D. Up to this factor and up to the relative error D/S, that is the least
size that the blocks can have on average. Statement (S) bounds D by D∗ from above only; the bound
in the other direction is in (R3) of §4 and is not used in this section. The number dw dw′ /D is also the
mean size for subspaces in general position: if Ξ is a random unitary map of H for Haar measure and
ζ, ζ ′ are unit vectors, the mean of |⟨ζ, Ξζ ′ ⟩|2 is 1/D, so the mean of ∥B † ΞB ′ ∥2HS for isometries B, B ′
of ranks ν, ν ′ is νν ′ /D. Orthogonal images would mean Kww′ = 0. Mutually orthogonal images
have dimensions that add up to at most D. For injective maps this is impossible when S > D. The
maps of the source are not known to be injective, but (III.219) bounds their ranks from below. The
operator Kww − 1 is −1 on the kernel of Bw , so ∥Kww − 1∥2HS ≥ dw − rank Bw , and (III.219) gives
rank Bw ≥ (1 − e0 ) dw . Under the hypotheses of Corollary T2 below, e0 ≤ 64 1
, S ≥ xCH D∗ ≥ 4D∗
and D ≤ 98 D∗ . So the ranks add up to at least 63 64 S > D, which is more than the dimension of the
target, and the images cannot all be mutually orthogonal. The Gram block estimates control their
total squared overlap and put Q close to its mean eigenvalue in Hilbert–Schmidt norm. Precisely,
for a Hermitian operator Q on H one has tr(Q2 ) − (tr Q)2 /D = ∥Q − (tr Q/D) · 1∥2HS , so under the
hypotheses of Lemma T1(c) the bound (III.225) is
tr Q 2 (tr Q)2
Q− ·1 ≤ 5κ . (III.232)
D HS D
The operator Q = w Bw Bw† , with Bw = c0 (w)−1/2 Pw for the maps of the source, is not the sum of
P
the projections onto the images: Bw Bw† is a projection only if Bw is a partial isometry. (2) The
√
constant 5. The term (1 + e0 )2 S in (III.223) comes from the diagonal blocks. It is the reason why
(III.225) needs D∗ ≤ κS: the total dimension S of the sources must exceed the dimension of the
target by the factor 1/κ. A term of this size is present on average for maps in general position:
for isometries Bw composed with independent random unitary maps of H, the mean of tr(Q2 ) is
S + (S 2 − w d2w )/D by Remark (1), while tr Q = S; so the mean of tr(Q2 )D/(tr Q)2 − 1 is about
P
D/S. This is a statement about a mean. Lemma T1(c) is an upper bound, and (III.218)–(III.220)
do not bound tr(Q2 )D/(tr Q)2 − 1 from below. For example, if every Bw is a unitary map of H,
then (III.218)–(III.220) hold with D∗ = D and η0 = e0 = η1 = 0, and Q = |W| · 1, so this number
is 0, while D∗ /S = 1/|W|. (Another example, with dw = 1: the p(p + 1) lines spanned by the
vectors of p + 1 mutually unbiased bases of Cp , where such bases exist. Here Q = (p + 1) · 1 and
134
D∗ /S = 1/(p + 1).) (3) The archive contains statements with these names. CorBounds is for (b).
TraceSquareAndDiagBounds is for (c), with e0 ≤ 41 where (c) has 64 1
. CompressTraceSquareBound
is a variant of (c), not of (d): the maps are first restricted to subspaces of total codimension at most
√
αS, and the conclusion is the second-moment bound for the restricted maps with η = 5κ + 3 α
(for κ ≤ 20 1
, 0 ≤ α ≤ 14 ). The bound for the diagonal blocks of the restricted maps, with
e = 2e0 , is the conclusion of another statement, CompressDiagDefectBound (for α ≤ 12 ). The
proofs above are ours. The Lean proofs of these statements were not read. By the dependency
computation the final theorem depends on CompressTraceSquareBound and not on CorBounds or
TraceSquareAndDiagBounds. The name CompressDiagDefectBound was not among those queried.
Corollary T2 (the hypotheses of Part II, §6.9 for the maps of the source, and for arbitrary
subspaces). Assume (S) with (III.217), (I) with δp < 1, and (P), for one record. Let CH > 0 and
put κ = x−CH . Let αram , α′ ≥ 0, and for each word let Uw ⊂ Mw be a complex subspace. Assume
further
|W| ≥ C0 G2 xCH , dw −dram
w ≤ αram dw (w ∈ W), κ ≤ 14 , exp(−xε/8 ) ≤ κ/2, (III.233)
w − dim Uw ≤ α′ (III.234)
X X
dram
dw .
w w
For each word let βw : Mw → HomL (EL , L) ⊗L C be the complexification of the map φ 7→ Balgw (φ, ·)
on Mw,L ; here Mw,L ⊗L C is identified with Mw as in §1. The target is identified with the dual E0∨
of E0 by the map EL → E0 of §1. Then the sum of the spaces βw (Uw ) has codimension at most
√
5x−CH + 4(αram + α′ ) + 4 e0
Dr (III.235)
1 − δp
in E0∨ , which has dimension Dr by (III.215). In particular, for Uw = Mw and α′ = 0, the sum of
√
the images of the βw has codimension at most (5x−CH + 4αram + 4 e0 )Dr /(1 − δp ).
Proof. First, c0 (w) > 0 for every word, so that Bw is defined. Indeed dw ≥ 1 by the lower
bound in (III.217), and e0 < 1 by (III.216); if c0 (w) = 0, then Kww = 0 and (III.219) would read
dw ≤ e0 dw . (This is unfolding_constant_pos_of_diagonal in the source.) By the lower bound in
(III.217), dw ≥ D∗ /(C0 G2 ), so S ≥ |W|D∗ /(C0 G2 ) ≥ D∗ /κ. With (III.216) and the last condition
of (III.233), (III.224) holds; and D ≥ 1, because (III.219) with e0 < 1 gives Bw ̸= 0. (If η0 < −1,
(III.218) forces |W| = 1, which the first condition of (III.233) excludes.) By the second condition of
(III.233) and by (III.234), w (dw − dim Uw ) ≤ (αram + α′ )S.
P
The estimate in H. The space E = s(E0 ) is a complex subspace of H, because c s(F ) = s(c̄F ). Its
dimension is Dr = (1 − δp )D, because s is injective and maps a basis of E0 to a basis of E. Let πE be
the orthogonal projection of H onto E. Apply Lemma T1(d) with the subspaces Ψw Uw ⊂ Mw ⊂ Cw ,
which have the dimensions of the Uw , with αram + α′ in place of α, and with δE = δp . Since Pw and
Bw differ by a positive factor, the space Y = w πE Pw Ψw Uw has codimension at most (III.235) in
P
E. This step uses only the subspace E of H, and neither s nor the scalars of E0 .
The passage to the dual. Let Λ : H → E0∨ send h to the form
Λ(h) : F 7−→ ⟨sΨF, h⟩H (F ∈ E0 ). (III.236)
The form Λ(h) is C-linear in F , by the computation after (III.222); so it lies in E0∨ . The map Λ is
C-linear in h. Its kernel is the set of h orthogonal to sΨ(E0 ) = s(E0 ) = E, that is, E ⊥ . So Λ = Λ ◦ πE ,
the restriction of Λ to E is injective, and since dim E = Dr = dim E0∨ it maps E isomorphically onto
E0∨ . Hence Λ maps H onto E0∨ , and Λ(Y) has in E0∨ the codimension of Y in E. By (III.222), the
135
map φ 7→ Bw (φ, ·) from Mw to E0∨ is λw ΛPw Ψw , a C-linear map. The image of Uw under it is
ΛπE Pw Ψw Uw , because λw ̸= 0 and the image is a complex subspace. So the sum of these images is
Λ(Y).
The algebraic model. The map EL → E0 is linear over L and maps an L-basis to a C-basis. So
restriction of forms along it is a C-linear isomorphism of E0∨ with HomL (EL , L) ⊗L C. Under it and
under Mw,L ⊗L C = Mw , the map φ 7→ Bw (φ, ·) is βw : both are C-linear, and by (P1) they agree
on the image of Mw,L , which spans Mw .
On the two scalar actions. Every space in this proof is a complex space, and every map in
it is C-linear, except s. The conjugate-linear map s occurs only inside the inner product, in its
conjugate-linear variable. The constants are those of Lemma T1(d).
The corollary is stated for arbitrary subspaces because Part II, §6.10 needs it so: there the rank
bound is applied after discarding directions, for every level of the slope filtration. The subspaces
that occur are complexifications of L-subspaces of Mw,L .
The side conditions in the source. For the second condition of (III.233): each word has a level
prime T (w) ≥ bm/2 , and dram
w = dw (1 − 1/(T (w) − 1))f (RamRank_at); by Bernoulli’s inequality
this gives αram = 2f /b for b ≥ 3 (ramShare_at_of). The number T (w) is the level T of Part II,
§6.7, which is said there to be attached to one choice of lattice data. In the source it depends
on the translate and on the word (T u0 u). The first condition of (III.233) is ChCode_at. In the
source the number of words is fixed at that size: with C0 = 3,
S
|W| = 3 G2 xCH , so that, under (III.217), xCH ≤ ≤ 9xCH + 3 (III.237)
D∗
(the equality is a clause in DesignS.lean, where the exponent is the one passed to ChCode_at;
the two bounds are ours, from D∗ /(3G2 ) ≤ dw ≤ 3D∗ /G2 ). Since the words have k − 2 letters,
this requires 2k−2 ≥ |W|. The source has log |W| ≤ (k − 3) log 2 (a clause of Win_at) and the size
condition 8CG m5 f xε ≤ k (a clause of SizeKRel), with m ≈ x1/2−ε/16 , n ≈ x2+8ε and k ≈ x9/2+10ε ,
√
each up to a factor x±ε/100 (SizeM, SizeN, SizeK). The rounded sizes m ≈ x, n ≈ x2 , k ≈ x9/2
of §I.8 do not satisfy this condition: they give m5 f xε = 2x9/2+ε . By the dependency computation
the final theorem depends on RamRank_at, ramShare_at_of and ChCode_at. The names Win_at,
SizeKRel, SizeM, SizeN and SizeK were not among those queried.
How the source uses the corollary. The source does not state Corollary T2. It applies
a theorem on lattices (TheoremF). By the names of the declarations, its proof uses the lemma of
Part II, §6.9 in the form FrameLA1, which is stated for arbitrary subspaces and has η and not
2η (the chain of names from frame_at_of to frame_linear_algebra_first was followed; the
proof terms were not read). The conclusion is the mean-slope inequality of Part II, §6.10, for each
translate, with the error term
1 √ √
4u log 1 + H + (η + 4 e)H , η = 5x−CH + 3 αram , e = 2e0 , (III.238)
1 − δp u
and with − log |W| as a separate term of the lower bound, outside the bracket and not multiplied
by (1 − δp )−1 (Frame_at, frame_at_of). Here u and H are the spread and the height bound of
Part II, §6.10, and αram is any number with αram ≤ 14 for which the second condition of (III.233)
holds (de in the Lean); the statement also requires x−CH ≤ 20 1
. The values of η and e in (III.238)
ε/8
are the ones that the proof of frame_at_of supplies. The source then takes αram = e−2x , so that
√ √
αram = exp(−x√
ε/8 ), and uses e0 ≤ exp(−xε/8 ) from (III.216) (matRealF_of). For all sufficiently
√
large x, (3 + 4 2) exp(−x )x H ≤ 1. Since η = 5x−CH + 3 αram and e = 2e0 , this gives
ε/8 C
√ √
(η + 4 e) xCH ≤ 6, so that (η + 4 e)H ≤ 6b for H = xCH b (III.239)
136
√ √ √
(FrameF, frameF_of).
√ Indeed (η + 4 e)xCH = 5 + (3 αram + 4 2e0 )xCH , and the bracket is at
most (3 + 4 2) exp(−xε/8 ). The weaker bound√exp(−xε/8 ) ≤ x−CH /2, which is the last condition
of (III.233), would not suffice:
√ it leaves
√ (3 + 4 2)/2 > 1 for the terms beyond 5. In the Lean the
hypothesis of frameF_of is (3 de+4 eB) xCH ≤ 1 with 0 ≤ de ≤ 14 and 0 ≤ eB ≤ 14 (FrameRealF),
where de is αram and eB is an upper bound for e; matRealF_of proves it for x ≥ 18 + |a0 | + |a1 |.
Here a0 is the threshold in its hypothesis (MatSF), and a1 is a number such that xCH +1 ≤ exp(xε/8 )
for all x ≥ max(a1 , 1). For such x the proof deduces exp(−xε/8 )xCH ≤ 18 1
, and it uses the cruder
√
constant 11 in place of 3 + 4 2. So the threshold is not 18; it depends on ε and CH . The exponent
CH is the one in the bound H ≤ xCH b of the source (a clause of LinkDIV_at).
What is known about the use of these declarations. blockReal_at_of is in the dependency
lists as a declaration that the final theorem depends on. frame_at_of and matRealF_of were
first reached by reading proof terms from blockReal_at_of, and a later dependency compu-
tation confirms that the final theorem depends on both. The chain of proof terms is this.
blockReal_at_of is used only in the proof of matKF_of. matKF_of is used in four places. In
each, its value is an argument of frameRealF_of”, and the value of frameRealF_of” is an argu-
ment of frameF_of. frameRealF_of” uses matRealF_of through frameRealF_of’, and frameF_of
uses frame_at_of. The same computation shows that the final theorem depends on LinkIRR,
TheoremF, FrameLA1, frame_linear_algebra_first, Frame_at, FrameF, matSF_of, ChCode_at
and CodeUpF, and does not depend on MeanBoundSecondForm. The statement FrameLA and its
proof frame_linear_algebra_bound are in the lists as not used by the final theorem, as Part II, §6.9
notes. The final theorem also depends on matKF_of, on unfolding_constant_pos_of_diagonal
and on TheoremFSharp.
The error term and the two means. Part II, §6.9–6.10 states these points; here they are
given with the names of the source. The error. For each translate, the mean-slope inequality of the
source has the form
1 √
(mean slope of the target) ≥ s − log |W| − 4u log 1 + H + (η + 4 (III.240)
e)H ,
1 − δp u
√
with η = 5x−CH + 3 αram and e = 2e0 as in (III.238). Here s is the mean over the words, weighted
by the ranks, of the slopes of the sources minus their heights (sFrame), and the target is the dual
lattice (muDual); this is the statement Frame_at. The factor (1 − δp )−1 multiplies the error bracket
only. It is there because the lattice of the target spans a space of dimension Dr , not D, and the
second moment is taken in H; this is where (I) is used, as in Lemma T1(d). The term − log |W|
is separate. It is at most (k − 3) log 2 in absolute value (Win_at), which is negligible against b.
The archive has a statement with the same hypotheses and no such term (TheoremFSharp). In η
√
only αram is exponentially small. The power of x comes from the number of words (III.237): by
Lemma T1(c) a bound 5κ for η needs D∗ ≤ κS, and D∗ /S ≥ 1/(9xCH + 3). So no bound for η
better than a negative power of x follows from (S) by Lemma T1. This is enough, by (III.239). We
do not assert that the number tr(Q2 )D/(tr Q)2 − 1 of the family is of that size: it was not bounded
from below (Remark (2)). The two means. Each translate of Part II, §4.2 has its own record, with
its own set of words (§1). The mean over the words is an inner mean, weighted by the ranks, inside
(III.240) for one translate: the number s is, in the file MeanBoundSecondForm.lean (source [7] of
Part II),
def s (u0 : U0) : R := wmean (D.d u0) (fun u => D.muN u0 u - D.h u0 u).
That file defines the record and its hypotheses, among them the slope inequality (III.240) for each
translate (Data.SlopeIneq; the final theorem depends on it). It also states a bound for the mean
137
over the translates, with an error ErrX that contains the mean of log |W| (MeanBoundSecondForm).
The final theorem does not depend on that statement. The route of the final theorem keeps the
translate fixed: each_unit_bound_statement, in the file LevelOfCongruence.lean (source [28]
of Part II), has the conclusion
D.ASK u0 ≤ D.Main u0 + ErrX5at D C Xpp u0 for every u0,
and log |W|, for the set of words of the translate u0, is a term of ErrX5at. The final theorem
depends on each_unit_bound_statement; its two uses in the archive, in the files ChTop.lean
(source [5]) and PChain.lean, are again for one translate. So the bound of the source holds for
each translate, and the mean over the translates in (M) is an outer mean, taken afterwards. The
inner mean is not visible in (M). The “two moments” of Part II, §6.9–6.10 are what the lemma
there needs. What the source assumes are the bounds (III.218)–(III.220) for the Gram blocks; the
two moments follow from them by Lemma T1(b),(c).
3. The mechanism in a finite model
Statement (S) is proved in the source by expanding each block Kww′ as a sum of terms and
computing its squared norm as a sum over pairs of terms (see §4). The simplest instance of this kind
of computation is the following, for theta lifts between two definite groups as in Part II, §6.1–6.2.
It is a statement about finite sets.
The setting. Let I and I ′ be nonempty finite sets with positive weights ω and ω ′ (classes
of lattices Li and L′i′ , with ω(i) = 1/|Aut Li |). For i ∈ I let Gi be a finite group of order
1/ω(i). For i ∈ I, i′ ∈ I ′ let Rep(i, i′ ) be a finite set on which Gi acts freely (the isometric
embeddings ρ : Li → L′i′ , with σ ∈ Aut Li acting by ρ 7→ ρσ). Let A be a finite abelian group and
inv : Rep(i, i′ ) → A maps with inv(ρσ) = inv(ρ) for σ ∈ Gi . For a character χ of A put
Θχ (i, i′ ) = (Bχ f )(i′ ) = ω(i) Θχ (i, i′ ) f (i). (III.241)
X X
χ(inv(ρ)),
ρ∈Rep(i,i′ ) i∈I
So Θ1 (i, i′ ) is the number of representations of Li by L′i′ , and Θχ is a representation number twisted
by a character. The map Bχ goes from L2 (I, ω) to L2 (I ′ , ω ′ ), where ⟨f, g⟩ = i ω(i)f (i)g(i),
P
conjugate-linear in the first variable as everywhere in this section. An operator Y on L2 (I, ω) has a
kernel, also written Y , with (Y f )(i1 ) = i2 ω(i2 )Y (i1 , i2 )f (i2 ); the kernel of the identity is 1/ω(i1 )
P
for i1 = i2 and 0 otherwise, and ∥Y ∥2HS = i1 ,i2 ω(i1 )ω(i2 )|Y (i1 , i2 )|2 . For i1 , i2 ∈ I put
P
N (i1 , i2 ) = ω ′ (i′ ) # (ρ1 , ρ2 ) ∈ Rep(i1 , i′ ) × Rep(i2 , i′ ) : inv(ρ1 ) = inv(ρ2 ) , (III.242)
X
i′ ∈I ′
and let N ◦ (i1 , i2 ) be the same sum restricted to the pairs that are not of the form (ρ, ρσ) with
i1 = i2 and σ ∈ Gi1 .
Lemma T3. (a) The operator χ Bχ Bχ on L (I, ω), the sum being over all characters of A,
† 2
P
has the kernel |A| N .
(b) Assume that there is a number m with
ω ′ (i′ ) |Rep(i, i′ )| = m for every i ∈ I. (III.243)
X
i′ ∈I ′
χ Bχ Bχ = |A| (m · 1 + N ), where N is the operator with kernel N ≥ 0, and
Then
P † ◦ ◦ ◦
2
Bχ† Bχ − |A| m · 1 = |A|2 ω(i1 )ω(i2 ) N ◦ (i1 , i2 )2 . (III.244)
X X
χ HS
i1 ,i2 ∈I
138
(c) Under (III.243) and m > 0, the family of the adjoint maps Bχ† : L2 (I ′ , ω ′ ) → L2 (I, ω) satisfies
the second-moment hypothesis of Part II, §6.9: with Q = χ Bχ† Bχ ,
P
(tr Q)2 1 X
tr(Q2 ) ≤ (1 + η) , η= ω(i1 )ω(i2 ) N ◦ (i1 , i2 )2 . (III.245)
|I| m2 |I| i ,i
1 2
Proof. (a) The adjoint of Bχ is (Bχ† g)(i) = i′ ω ′ (i′ )Θχ (i, i′ )g(i′ ), as one sees by writing
P
out ⟨Bχ f, g⟩ = ⟨f, Bχ† g⟩. So Bχ† Bχ has the kernel i′ ω ′ (i′ )Θχ (i1 , i′ )Θχ (i2 , i′ ). Insert (III.241)
P
and sum over χ: by the orthogonality of the characters of A, χ χ(inv(ρ1 ))χ(inv(ρ2 )) is |A| if
P
inv(ρ1 ) = inv(ρ2 ) and 0 otherwise. This gives |A| N (i1 , i2 ).
(b) Let i1 = i2 = i. The pairs (ρ, ρσ) with ρ ∈ Rep(i, i′ ) and σ ∈ Gi are distinct for distinct
(ρ, σ), because the action is free; they satisfy inv(ρ) = inv(ρσ); and their number is |Rep(i, i′ )|/ω(i).
By (III.243) their contribution to N (i, i) is m/ω(i), which is the kernel of m · 1. The rest is
N ◦ , by definition, and N ◦ ≥ 0 because it is a weighted count. (III.244) is the formula for the
Hilbert–Schmidt norm of an operator with a kernel.
(c) For a Hermitian operator Q on a space of dimension ν and any real y, tr(Q2 ) − (tr Q)2 /ν =
∥Q − trνQ ∥2HS ≤ ∥Q − y∥2HS . Take ν = |I| and y = |A| m and use (III.244). Also tr Q = |A|(m|I| +
tr N◦ ) ≥ |A| m|I|, because tr N◦ = i ω(i)N ◦ (i, i) ≥ 0. Divide.
P
What the lemma shows. The sum over the twists is a count of pairs of representations subject to
one condition, inv(ρ1 ) = inv(ρ2 ): this is the form a relative trace formula takes when both groups
are definite. The pairs that are forced, (ρ, ρσ), give a multiple of the identity, and the multiple
is the mean number of representations (III.243). The other pairs are coincidences, and η is their
normalised mean square. Part (b) uses three hypotheses, and each is needed. The first is that the
action of Gi is free: with stabilisers of order 2 the identity of (b) fails. The second is the invariance
inv(ρσ) = inv(ρ): without it the forced pairs do not all satisfy the condition, and the identity of (b)
and (III.244) fail. (The number N ◦ is a restricted count and stays nonnegative; it is the kernel
N − m/ω on the diagonal that can become negative.) The third is (III.243), which is not formal.
For lattices it is Siegel’s theorem: if I ′ is a full genus and the Li belong to one genus, the left side
of (III.243) is the mass of the genus I ′ times a product of local densities, which depend only on the
genus of Li (C. L. Siegel, Über die analytische Theorie der quadratischen Formen, Ann. of Math. 36
(1935); in the adelic form, A. Weil, Sur la formule de Siegel dans la théorie des groupes classiques,
Acta Math. 113 (1965); both cited from memory).
An example. Let L1 , L2 be the binary lattices with Gram matrices 21 12 1 and 41 16 (twice
the forms X1 + X1 X2 + 6X2 and 2X1 + X1 X2 + 3X2 of discriminant −23, which form one genus
2 2 2 2
and represent the two classes under GL2 (Z)), with automorphism groups of orders 4 and 2, and
let I ′ consist of the lattice Z5 , which is alone in its genus, with ω ′ = 1/(25 5!). A count gives
|Rep(L1 , Z5 )| = |Rep(L2 , Z5 )| = 1920, so (III.243) holds with m = 12 ; for Z 6 both numbers are
19200. (For two forms of determinant 60 in different genera, 20 30 0 and 60 10
0 , the numbers for Z5
are 4480 and 13440.) Let ς0 , . . . , ς9 be the ten 2 × 2 minors of the pair of vectors ρ, in lexicographic
order. Take for inv(ρ) the 55 products ςι ςι′ modulo a prime ℓ0 ; they are invariant under Gi , and
A = (Z/ℓ0 )55 . For ℓ0 = 3 one finds N ◦ = 20 03 . Then the true value tr(Q2 )|I|/(tr Q)2 − 1 is 19 , and
the bound (III.245) gives η = 5. For ℓ0 = 5 and ℓ0 = 11 there are no coincidences, N ◦ = 0, and
χ Bχ Bχ = |A| m exactly. With the single invariant inv(ρ) = ς0 + 2ς9 + 3ς1 ς4 modulo 5 (so |A| = 5)
P † 2 2
the coincidences dominate: N ≈ 242◦
173 209 . So the family of twists must be large compared with
173
the number of representations for the count N ◦ to be small. The identities (III.244), (III.245) were
also checked numerically on random finite models.
139
4. The statements behind (S), (I) and (P) in the source
These are assumed here. For each we give the statement that was read, its Lean name, the kind of
theorem it is, and the parameter that produces the saving. References to the literature are from
memory.
(R1) The expansion of a block and its four parts (ClKppOff, ClKppDiag, ClOff, ClDiag;
(τ )
they give (S) by linkS_concl). Each block is an absolutely convergent sum Kww′ = τ Kww′ over
P
a set of terms τ (Term, aTerm). A term has an invariant in the totally real field (Gterm) and may
be the identity term, or generic. Write Kww
′′
′ for the sum of the terms with invariant 0 other than
the identity term. The four statements are: Kww ′′
′ = 0 for w ̸= w ; ∥Kww ∥HS ≤ dw exp(−kf log k);
′ ′′ 2
′′ −6 dw dw′ d2w
∥Kww′ − Kww ′ ∥HS ≤ (1 + b
2
) (w ̸= w′ ), ′′
∥Kww − Kww − 1∥2HS ≤ G . (III.246)
D∗ D∗
With the upper bound in (III.217), G d2w /D∗ ≤ (C0 /G) dw : this is (III.219), and it shows where
e0 comes from. The diagonal bound is small because each Cw is a small part of H, of relative
dimension about G−2 , while the constant in the second bound of (III.246) is only G. This is an
expansion of the kind of Petersson’s formula for the inner products of Poincaré series, here for the
Bessel periods of a unitary group in 2( m 2 − 1) + k + 1 variables along a unitary group in k variables;
this number is k + m − 1, the r of §III.4 (the periods of W. T. Gan, B. Gross and D. Prasad,
Astérisque 346 (2012)); it is a relative trace formula in the sense of Jacquet. In the source the
squared norm of a part of the sum is expressed as a sum, over pairs of terms, of inner products of
vectors attached to the terms (pairSum; compare (III.242)), and the pairs are divided by a rank
s0 ≥ 1 and by being “scalar” or not. The pairs with s0 = 1 that are scalar and generic give the
main term. The statements read are bounds for the classes of pairs: RowNGP (non-generic terms: at
most 2−(k−m−3) exp(Cm5 f log b) dw dw′ /D∗ ) and RowMainEqP (the main pairs of a diagonal block:
at most exp(Cm5 f log b) d2w /D∗ ). Both are deduced in the source from (R2). The saving in η0 = b−6
comes from k and from the level primes: the non-generic and the higher-rank pairs carry factors
2−k and exp(−ckf log k) with k ≥ C ′ m5 f xε , and the factor v (1 + 3/|Fv |) over the places above
Q
g0 , g1 is 1 + O(b−7 ) when g0 , g1 ≥ (kb)8 (NumPrice). How these are combined in the proof of the
first bound of (III.246) was not read.
(R2) The Siegel–Weil formula in the convergent range (SiegelWeilUnitaryConvergent;
the theorem siegelWeilUnitaryConvergent is among the declarations the final theorem depends
on). Let E/E + be a CM extension, J a nondegenerate Hermitian matrix of size M over E, and ν
an integer with 2ν < M . The quotient of the adelic points of the unitary group of J by its rational
points has finite volume, and for every nondegenerate Hermitian Z of size ν and every factorisable
Schwartz function Φ = Φ∞ ⊗ v Φv , the mean over this quotient of X Φ(gX), the sum being over
N P
the ν-tuples X of rational vectors with Gram matrix Z, equals ϱ∞ (Z) v ϱv (Z), where ϱv is the
Q
density at Z of the image of the measure Φv dX under the Gram map. (The description of the
theta sum is our reading of the names; its definition was not read.) This is Siegel’s theorem in
Weil’s form, for a unitary group, in a range of absolute convergence; for unitary groups beyond that
range see A. Ichino, A regularized Siegel–Weil formula for unitary groups, Math. Z. 247 (2004), and
S. Yamana, On the Siegel–Weil formula: the case of singular forms, Compos. Math. 147 (2011). It
plays the role of (III.243): it evaluates the mean number of representations, and so the main pairs,
by local densities. It contains no saving by itself.
(R3) The count of H (CountAllPointsClass; it gives (III.220)). |D − D∗ | ≤
exp(−c rf log r) D∗ , where D is the multiplicity of a discrete series representation, at a fixed
level, in the automorphic spectrum of the unitary group in r variables, and D∗ is the product of
140
its formal degree, the volume of the quotient and the value at 1 of the idempotent defining the
level, provided the Harish-Chandra parameter has all gaps at least Cr19 f 5 log(b rdE ), and under
two conditions on the sizes of f and b. (The description of D and D∗ is our reading of the
definitions Dcount and Dstar.) This is a limit multiplicity statement with a rate (in the tradition
of D. L. de George and N. R. Wallach, Ann. of Math. 107 (1978), and L. Clozel, Invent. Math. 83
(1986)); D∗ is the identity term of the trace formula. The saving comes from the gaps of the weights.
The volume in D∗ is a mass. The mass formula of the source is siegel_mass_ratio: the quotient
of the masses in ranks M and M − 1 is ((M − 1)! (v/π)M )f (2v)−f /2 times a further factor (here
v is a positive real number, not a place, and π is the number), which is the recurrence displayed
before (6.4) in Part II, §6.6 (H. Braun, Abh. Math. Sem. Hamburg 14 (1941); W. T. Gan, J. Hanke
and J.-K. Yu, Duke Math. J. 107 (2001)). Here we read the source’s vInv, fDeg, lValEta as v, f
and that factor; their definitions, and the place where the formula is applied to D∗ , were not read.
(R4) The count behind (I) (closed_linkIRR_of_join_count, delp_of_allRows). The
space H has a basis each of whose D elements carries, at each of the places above g0 and above g1 ,
a label: a monic polynomial of degree r over Fgι (ι = 0, 1) with nonzero constant term. A tuple
of labels at the places above one gι is good if the polynomials are squarefree, their roots satisfy
two conditions of non-coincidence, for every j with 1 ≤ j ≤ r/2 some place has a label with no
factor of degree j, and some place has a label with an irreducible factor of prime degree j ′ with
r/2 < j ′ ≤ r − 3 (IsGood). Dr is the number of basis elements whose tuples are good at both
primes. The proof of delp_of_allRows contains the step
1 + εC
Wg(g0 ) + Wg(g1 ) , εC = exp(−c rf log r), (III.247)
δp ≤
1 − εC
3f 23f log g 1 1 + log r r 7 f f
Wg(g) = + √ + 2f m + + + exp − , (III.248)
g g 4 min(q, d) g−1 2 8 5 log r
where q, d are the auxiliary primes of §III.1. The statement delp_of_allRows has as a hypothesis
that the right side of (III.247) is at most a number of the source (E7), and as its conclusion
0 ≤ δp ≤ exp(−xεz /7 ); (III.247) itself is not a statement of the source. Two inputs of that proof:
the count (R3), applied to the part of H with prescribed labels, whose expected dimension is D∗
times a product over the places of local weights; and the fact that the local weight of a squarefree
label is g −r (1 + O(r/g)), the labels that are not squarefree having total weight at most 3/g (admLq,
lawAt), so that the labels at each place are distributed, up to these errors, like uniformly random
monic polynomials of degree r with nonzero constant term. The latter rests on the parametrisation
of the irreducible characters of GLr (Fg ) by J. A. Green, The characters of the finite general linear
groups, Trans. Amer. Math. Soc. 80 (1955), which the source takes as a hypothesis Green (the
cuspidal characters of GLj (Fg ) correspond to the monic irreducible polynomials of degree j with
nonzero constant term and have degree j−1 i=1 (g − 1); for j = 1 these are the g − 1 polynomials
i
Q
X − a with a = ̸ 0, and the polynomial X is excluded; Lean: IsPrimeLab). The terms of Wg(g)
are then statements about the factorisation of uniformly random monic polynomials with nonzero
constant term, taken at f independent places. The bound (III.221) follows from (III.247) and the
sizes g0 , g1 ≥ exp(8xεz ), f ≥ x2 , f, m, r ≤ xC and min(q, d) ≥ exp(xεz /6 ) (SizesAt). The term that
limits the bound is f m/(2 min(q, d)); the others are much smaller. The good tuples are those for
which the source can prove that an attached residual Galois representation is irreducible; this use
was not read. By the dependency computation the final theorem depends on delp_of_allRows
and on Green.
(R5) Uniqueness of Bessel models, and decay (besselMapUniqGap, wuDecayOfGap,
adm13_of_kernel_join). For a discrete series representation of a real unitary group in r variables
141
whose Harish-Chandra parameter has all gaps at least 120r3 , a representation of the unitary group in
k variables of a given type, and a generic character, any two Bessel maps on the smooth vectors are
proportional. (The source states this for k + 2j + 1 variables with any j ≥ 0; here j = m
2 − 1, which
gives r variables.) Compare Theorem A of D. Jiang, B. Sun and C.-B. Zhu, Uniqueness of Bessel
models: the archimedean case, Geom. Funct. Anal. 20 (2010): it gives multiplicity at most one for
Bessel models at an archimedean place. The statement of the source has in addition a hypothesis
on the gaps. Theorem A does not give the quantitative decay used here. The decay is a further
statement of the source: from the uniqueness it deduces an integrability statement for a function
attached to a word (WuDecay; the name refers to this function), under the condition that r ≥ 1801
and that consecutive weights θj of §5 differ by at least 16r19 (GapHyp); by our reading this serves
the convergence of the expansion in (R1). Statement (P2) is proved by adm13_of_kernel_join,
from hypotheses that were not read.
Where the exponents come from.
number value in the source large parameter
η0 b−6
k, and g0 , g1 ≥ (kb)8
e0 2C0 /G + 2 exp(−kf log k) G = exp(CG m5 f xε ): the ratio D∗ /dw ≈ G2
η1 exp(−c rf log r) the rank r ≥ k and the degree f
αram 2f /b the level primes T (w) ≥ bm/2
δp f m/(2 min(q, d)): leading the auxiliary primes q, d ≥ exp(xεz /6 )
term in the upper bound for
δp
√
bound for η 5x−CH + 3 αram the number of words, |W| = ⌈3G2 xCH ⌉
The exponent xε/8 in (III.216) is not the true size of any of these numbers. It is chosen in
√
the source so that one bound, exp(−y) with y = xε/8 and xε = y 8 , covers η0 , η1 and e0
(eF_nonneg_and_le_exp_neg). The exponent εz /7 in (III.221) is just below the εz /6 of the
auxiliary primes. The last row is the bound for the constant η of Part II, §6.9 proved in the source
(frame_at_of); see the paragraph on the error term at the end of §2 for what it does and does not
say about η.
5. The interlacing of the weights
At each weight place the source fixes rational numbers
ξ0 > ψ0 > ξ1 > ψ1 > · · · > ψk−2 > ξk−1 > −θ0 , θ0 < e′0 < θ1 < e′1 < · · · < θm−1 < e′m−1 ,
(III.249)
with ξi ∈ Z+ 12 and ψi , θj , e′j ∈ Z (HCParam; the condition is named Interlace; we write ξi , ψi , θj , e′j
for the numbers called u, v, b, e there; e′0 is not the e0 of (S)). The first chain is the condition
under which the representation of the unitary group in k − 1 variables with parameter ψ occurs in
the restriction of the one in k variables with parameter ξ: the classical branching rule, here for
parameters shifted by half the sum of the positive roots. In this section it is used in one place,
the definition of the family: the words are the weights ξ that interlace one fixed ψ. Precisely,
consecutive ψl differ by 4∆ for a fixed ∆, the numbers ξ0 and ξk−1 are the same for all words, and
for 0 ≤ l ≤ k − 3
ξl+1 = ψl − 12 − ∆ · (the lth letter of the word), (III.250)
and for both values of the letter this number lies between ψl+1 and ψl , since (III.249) is part of the
data for every word (cloud_gap, u_letter, hcW_v, hcW_first, hcW_last). So two words differ in
142
the weight ξ at the positions where their letters differ. That the bounds (III.246) for w ̸= w′ use
this (through the archimedean factors of the terms) is our reading of the definitions; the proofs were
not read. The second chain of (III.249), with gaps at least (t0 + 3)/2 (gap_be, gap_eb), concerns
the weights along the chain of lifts and belongs to the divisibility and size estimates (§III.4, §III.5).
6. Assumed here, from the source
(i) Statement (S): LinkS with C0 = 3, proved by link_alt_alt_bis from
theoremS_transport_link_zero, for families with TuSize. Its inputs, as hypotheses
of link_alt_alt_bis: the decay WuDecayOfGap; a finiteness statement RowsFinDim; the
bounds TypeP, RowNGP, RowMainEqP, LoopRowS; the off-diagonal bound TheoremSSelected;
the count CountAllPointsClass; and compatibility conditions. Only the statements were
read, and for RowNGP, RowMainEqP the deduction from (R2) by name.
(ii) Statement (I): LinkIRR, proved by closed_linkIRR_of_join_count from Green and a count
hypothesis (CountAllRows).
(iii) Statement (P): BesselId, the conjunction of BuRational and adm13; the latter by
adm13_of_kernel_join.
(iv) The side conditions of Corollary T2 : the proviso (III.217) (ChProviso_at) and the number
of words (ChCode_at), which are conditions on the family; the formula for dram
w (RamRank_at)
and T (w) ≥ bm/2 (TuGe). That c0 (w) > 0 is not assumed: it follows from (S), see the proof of
Corollary T2.
(v) The identification of the data of §1 with spaces of automorphic forms, with Poincaré series
and with the pairing of the chain of lifts; and the passage from Corollary T2 to the mean-slope
inequality (TheoremF, frame_at_of), which uses the slope filtration of Part II, §6.10. For
what is known about the dependence of the final theorem on these two declarations see the
end of §2.
(vi) (R1)–(R5) of §4.
What this section proves, and what it assumes. This section deduces the dimension
bound from the three trace, rank and pairing statements (S), (I), (P) and the size and subspace
conditions (III.233), (III.234) in T2. Proved outright: Lemma T1, including the polynomial
inequality (III.230); Corollary T2, for arbitrary subspaces, as that deduction; Lemma T3; and
the arithmetic of the example of §3 (by computation). Quoted from the source without proof: (S),
(I), (P) and everything in §4 and §6. So the construction of the spaces and maps, the arithmetic
inputs and the assembly of (M) are quoted. Lemma T3 is a model. It is not a step of the
source’s proof: there the family is the set of words, the terms are attached to a Bessel subgroup,
and the larger group is, by our reading, not definite (the parameter sigY0 of the source is the
signature (k + m2 − 1, 2 ); it is the signature of the space Y 0 of the record at every embedding, by
m
the field baseSpace_signature_eq). The statements FrameLA, frame_linear_algebra_bound,
ThmSFree and KernelFree of the archive are not used by the final theorem; the declarations LinkS,
LinkS_at, BesselId, blockReal_at_of, link_alt_alt_bis, theoremS_transport_link_zero,
closed_linkIRR_of_join_count, rowNGP_of_siegelWeil, rowMainEqP_of_siegelWeil,
siegelWeilUnitaryConvergent, adm13_of_kernel_join, wuDecayOfGap, besselMapUniqGap
and siegel_mass_ratio are. So are, by a later run of the same computation, LinkIRR, TheoremF,
FrameLA1, frame_linear_algebra_first, Frame_at, FrameF, frameF_of, frame_at_of,
matRealF_of, matSF_of, ChCode_at, CodeUpF, RamRank_at, ramShare_at_of, delp_of_allRows,
143
Green, matKF_of, unfolding_constant_pos_of_diagonal, TheoremFSharp, RRV_finrank,
jEE_basis (the field of D7) and CompressTraceSquareBound. The final theorem does
not depend on CorBounds, TraceSquareAndDiagBounds or MeanBoundSecondForm. Several
declarations of the archive are named offDiag. The one that concerns this section is
GlobalTheory.CarrierGap.offDiag, which proves the off-diagonal bound TheoremSSelected of
(i) from the Siegel–Weil statements and LoopRowS. The final theorem depends on it, and also on
each_unit_bound_statement and on Data.SlopeIneq (§2).
What remains.
(i) The definition of H, Cw and Pw in classical terms (the group, the level, the discrete series, the
Poincaré series), and of the terms of the expansion (R1).
(ii) A proof of the first bound of (III.246): the main term with constant 1 from (R2), and the
bounds for the other pairs. This is the centre of (S).
(iii) A proof of (P): why the pairing of the chain is an inner product with Pw φ.
(iv) The count (R3) with its rate, and the distribution of the labels in (R4).
(v) A run of the dependency computation for the names that were not queried: Win_at, SizeKRel,
SizeM, SizeN, SizeK, SizesAt, CompressDiagDefectBound.
(vi) Whether the number tr(Q2 )D/(tr Q)2 − 1 of the family of the source is of the size x−CH or
smaller; only the upper bound is known (the paragraph on the error term at the end of §2).
III.9 The bound for Galois cohomology
This section concerns the bound (II.46) of Part II, §6.8 (Part I, §I.7, “Denominators”): a bound for
the lengths of certain Galois cohomology groups of the field E, which the announcement [Alp26]
calls a weakening of Vandiver’s conjecture. It is the part of the proof that uses the most mathematics
from outside analytic number theory, and no proof of it is given here. The section states it exactly,
shows how it is used, describes the structure of its proof in the Lean development with the published
theorems that correspond to each step, and works out the simplest case.
Proved outright. Lemma G1 (lengths, orders and indices); Proposition G2 (how the bound gives
the estimate for the denominators, given the inequality LinkN_at of the source).
Proved from theorems of books that we consulted. Proposition G3: for the field Q the groups
vanish when a class group vanishes; with its corollaries for odd twists (Herbrand’s theorem), for the
twists with (ℓ − 1) | 2e, and under Vandiver’s conjecture. The theorems used are in Washington’s
book [Was82]; their statements
√ and numbers were checked in a copy of the first edition. Lemma G4:
the class number of Q( ℓ) is less than ℓ; it is proved from Minkowski’s theorem on lattice points
[Neu99, Ch. I, (4.4)]. The numerical statements of §5 are computed exactly.
Proved from the standard statements of class field theory. The vanishing of Sha5 (Q(i), 4), at
the end of §5. The proof is for this one field and this one twist.
Taken from the source without proof. Theorem G (the bound itself, Lean: VWinAt for the trivial
character); the inequality LinkN_at; and every statement of §4.
Citations. They are of three kinds. (1) The books [Was82], [Neu99] and [NSW00] were read in
local copies ([Was82] and [NSW00] in the first edition); the test group G19 compared each statement
cited from them with the page. (2) The articles [Sou03], [Kat04], [Kur92], [BC04], [Car12], [Car14],
[CH13], [BS23] and the book [Rub00] were not available to us. Where a theorem or a section of
144
one of them is named, the wording was compared with the publication by an outside reviewer, and
we say no more than his report says. (3) Everything else is cited from memory and is marked so:
[Sou79], the theorem of Voevodsky and Rost, [MW84], [FW79], the work of Harris and Taylor and
of Shin, and the last column of the table of §4.
Letters introduced in this section: ℓ (a prime, as in §III.5), e (the twist; Part II, §6.8 writes a),
t (a number of twists; in the construction it is the parameter written t0 in §III.5, and t in §III.4;
Part II, §6.8 does not name it and writes the bound ⌊x400 ⌋), V0 , Sℓ , GE,ℓ , Gv , We , Shaℓ (E, e), hℓ,e ,
h±ℓ,e , ȷ (complex conjugation), ψ (a Dirichlet character), B, ηj (w), Hj , y = x , V , eO , fO , κ, cψ , x ;
ε ′
in §4: K, R, R, R , c0 , ϑ, ϱ; and in §5: A, A , X∞ , H∞ , Γ, Π, ω, F , h, d, a, α, T , L, j, N , c, δ,
′ (i)
λ, z, ϕ, a1 , a2 , rec, u, ϵ, β. The letters x, ε, b, m, f, q, d, n have their meaning from Parts I and II;
in particular b = exp(xε ).
The letters are local. The letters of this section have the meanings given here in this section
only. The clashes most likely to confuse are these. n is the power of 2 with [E : Q] = 4n; §III.5
takes its n from the record of the construction (ord), and §1 says how the order of the frame is
related to the n of Theorem G; Soulé’s twist, which he writes n, is written n′ in §4. t is a number of
twists, the t0 of §III.5; in §III.5 the letter t is an index. e is the twist; e(w) and eO are ramification
indices (the ew of §III.6); e is never the base of the natural logarithm here. L0 is the base field of a
tower, in §4 only; the coefficient field, which is L0 in §III.6, is called K here, and K is an imaginary
quadratic field. ψ is a Dirichlet character and Ψ = ψχeℓ a character; in §III.2 ψ(u, χ) is a sum over
prime powers, in §III.4 and §III.7 ψ and Ψ are ranks, and in §III.8 they are weights and unitary
maps. C is the constant of (III.257), or an exponent in the table of §4; it is not the C of another
section, and it is not one of c (a homomorphism, in §5), c0 , cψ . Inside the section four letters have
two or more uses. T is a finite group in Lemma G1 and a real number in Lemma G4. N is a level
or a conductor in §3 and §4, the norm N (w), N (a) of a prime or an ideal, and the norm map of
class groups in the proof of the corollaries. j numbers the maps of the chain in §2 and is the map
that extends ideals in §5. ϵ = ±1 is a sign in Step 4 of §5; it is not ε.
1. The statement in the source
The groups. Let E be a number field, ℓ a prime and e an integer. Let Sℓ be the set of primes of
E above ℓ, and GE,ℓ the Galois group over E of the largest extension of E in which no finite prime
outside Sℓ ramifies. Let χℓ : GE,ℓ → Z×
ℓ be the cyclotomic character, and let We = Qℓ /Zℓ (−e): the
discrete group Qℓ /Zℓ on which σ acts by multiplication by χℓ (σ)−e . For v ∈ Sℓ let Gv ⊂ GE be a
decomposition group at v, mapped to GE,ℓ . Put
Shaℓ (E, e) = ker H 1 (GE,ℓ , We ) −→ H 1 (Gv , We ) ,
Y
v∈Sℓ (III.251)
hℓ,e = lengthZℓ Shaℓ (E, e) ∈ {0, 1, 2, . . . } ∪ {∞}.
Cohomology is that of continuous cochains: H 1 is the group of continuous maps ξ with
ξ(σσ ′ ) = ξ(σ) + σ · ξ(σ ′ ), modulo the maps σ 7→ σ · w − w. So Shaℓ (E, e) consists of the
classes that are unramified outside ℓ and trivial at every prime above ℓ, as in Part II, §6.8.
The length of a Zℓ -module is the length of a longest chain of submodules; a finite module of
length h has ℓh elements, and the length is ∞ if there is no finite bound. In the Lean: LData
(a choice of a place of E above each v ∈ Sℓ and of the character χℓ on GE,ℓ ), Sha and shaLen
(Formalism/GaloisCohomology/Statement.lean, lines 54, 75, 78); the value ∞ is written ⊤. The
length does not depend on the choices; this is proved in the source (shaLen_eq_shaInv).
145
The statement VVle(E, t, V0 ), for an integer t ≥ 1 and a real number V0 (Lean: VVle, in the
namespace UpperTheory.Top; ExistsPageN.lean:12). For every algebraic closure of E and every
choice of the data above, for every prime ℓ:
(1 ≤ e < t), and log ℓ for every finite set S of primes.
X X
hℓ,e < ∞ hℓ,e ≤ V0
ℓ∈S 1≤e 1, y = xε , b = exp(y),
f, m ≥ 1, Nc ≥ 0, D0 ≥ 0, C real, B0 = f m b y. Suppose (log x)2 ≤ y, that (III.257) holds with
dj = D0 , and that V = ℓ log ℓ 1≤e 0 the set |x| + |y| ≤ T is convex and symmetric, of area 2T 2 . Minkowski’s lattice √ point theorem
[Neu99, Ch. I, (4.4)] says that it contains a nonzero point of the lattice if 2T 2 > 4 d N (a). The
inequality is strict there. (The case of equality for compact √ sets is the last sentence of Exercise 2 of
§4 there, without proof; we do not use it.) Take T = 2 d N (a). For every T > T the theorem
2 ′
gives a nonzero lattice point with |x| + |y| ≤ T ′ . The lattice has only finitely many points with
|x| + |y| ≤ 2T . So one of them, nonzero, serves for a sequence of numbers T ′ decreasing to T . We
get α ∈ a, α ̸= 0, with |α| + |α′ | ≤ T , hence
2 √
|αα′ | ≤ 14 |α| + |α′ | ≤ 14 T 2 = 12 d N (a). (III.265)
Now let an ideal class be given, and let a be an integral ideal in the inverse class. The√ ideal (α)a
−1
is integral, because α ∈ a. It lies in the given class, and its norm is |αα |/N (a) ≤ 2 d.
′ 1
Counting: for j ≥ 1, the number of integral ideals of F of norm j is at most j. An ideal of
norm j is in one way a product, over the primes p dividing j, of ideals whose norms are the powers
of p in j. So let j = pk . If p splits, (p) = pp′ , the ideals of norm pk are the k + 1 ideals pk1 p′ k2
with k1 + k2 = k. If p is inert, there is one such ideal for even k and none for odd k. If p ramifies,
(p) = p2 , there is one, pk . In each case the number is at most k + 1, the number of divisors of pk .
Hence the number of ideals of norm j is at most the number of positive divisors of j, which is at
most j.
√ √
Conclusion. Every ideal class contains an integral ideal of norm j ≤ 12 d ≤ ℓ. The sum of
√ √ √
the integers j with 1 ≤ j ≤ ℓ is at most 12 ℓ ( ℓ + 1). So
√ √ √
ℓ ( ℓ + 1) ℓ+ ℓ
= (III.266)
X
h ≤ j ≤ < ℓ,
√ 2 2
1≤j≤ ℓ
√
because ℓ < ℓ for ℓ > 1. This holds for every prime ℓ, including 2 and 3.
√
Numerical check. The test group G12 computed the class numbers of Q( ℓ) for
all 17 984 primes ℓ < 2 · 105 by cycles √ of reduced forms (script t04_classno.py in
zeta-paper_tests/displays/G12): h < ℓ in every case. The test group G19 reproduced
this with PARI (the largest class number is 153, at ℓ = 176401; PARI’s values were not certified),
and for the 9592 primes ℓ < 105 it checked every inequality of the proof.
Corollaries. (a) Let e ≥ 1 be odd. If ℓ does not divide the numerator of Be+1 /(e + 1), then
Shaℓ (Q, e) = 0. (b) Let e ≥ 2 be even and (ℓ − 1) | 2e. Then Shaℓ (Q, e) = 0. (c) Let e ≥ 2 be even.
If ℓ does not divide the class number of Q(ζℓ + ζℓ−1 ), then Shaℓ (Q, e) = 0.
Proof. We use this fact: if F ⊂ Q(ζℓ ) is a subfield, then the ℓ-part of the class group of F is the
part of A fixed by Gal(Q(ζℓ )/F ). Indeed, write L = Q(ζℓ ) in this proof, let j : Cl(F ) → Cl(L) be
the map that extends ideals, and N : Cl(L) → Cl(F ) the norm. Then
N ◦ j = [L : F ] on Cl(F ), j◦N = σ on Cl(L). (III.267)
X
σ∈Gal(L/F )
The degree [L : F ] is prime to ℓ. By the first identity j is injective on the ℓ-part. By the second, a
class z of A fixed by Gal(L/F ) satisfies j(N (z)) = [L : F ] z, so it lies in the image of j; and the
image of j is fixed by Gal(L/F ). (a) Here i ≡ −e is odd. If (ℓ − 1) | (e + 1) then i ≡ 1 and A(1) = 0
[Was82, Prop. 6.16]. Otherwise take 3 ≤ i ≤ ℓ − 2. By Herbrand’s theorem [Was82, Thm. 6.17],
A(i) = 0 unless ℓ | Bℓ−i ; and Bℓ−i /(ℓ − i) ≡ Be+1 /(e + 1) (mod ℓ) by Kummer’s congruence [Was82,
155
Cor. 5.14], since ℓ − i ≡ e + 1 ̸≡ 0 (mod ℓ − 1). (b) If (ℓ − 1) | e then i ≡ 0 and A(0) is the ℓ-part of
the class group of Q, by the fact above. Otherwise (ℓ − 1)/2 divides e and is even, so ℓ ≡ 1 (mod 4)
and i ≡ (ℓ − 1)/2. The
√ part of((ℓ−1)/2)
A fixed by the squares in Π is A(0) ⊕ A((ℓ−1)/2)
√ , and the field fixed
by the squares is Q( ℓ). So A is the ℓ-part of the class group of Q( ℓ). It is 0, because the
class number is less than ℓ (Lemma G4). (c) For even e, A(i) lies in the part of A fixed by complex
conjugation, which is the ℓ-part of the class group of the real subfield.
Kummer’s congruence, as used in (a), was also tested numerically by the test group G12:
Br /r ≡ Br′ /r′ (mod ℓ) for the 2543 triples (ℓ, r, r′ ) with ℓ prime, 5 ≤ ℓ < 200, r and r′ even,
2 ≤ r ≤ ℓ − 3, r < r′ ≤ 220 and r ≡ r′ (mod ℓ − 1); in 14 of them ℓ divides Br (script
t03_bernoulli.py, part G, in zeta-paper_tests/displays/G12). The test group G19 dropped
the condition r ≤ ℓ − 3 and required r ̸≡ 0 (mod ℓ − 1) instead: the congruence holds for all 9738
such triples.
What is proved here, and what the Lean states. Everything proved in this subsection is about
the groups Shaℓ (E, e) of (III.251). Two identifications with the Lean are assumed and not proved.
The first is the transcription of §1: that Sha and shaLen of the Lean are the group and the length of
(III.251). It was compared at the level of the definitions (the same set Sℓ , the same twist χ−e
ℓ , discrete
coefficients Qℓ /Zℓ , continuous cocycles, the kernel of the restrictions to chosen decomposition groups;
shaLen_eq_shaInv removes the choices). Under it, the vanishing of Sha5 (Q(i), 4) proved below says
that shaLen is 0 for Q(i), ℓ = 5, e = 4. The second concerns PropZ, which says: for ℓ ≥ 5 and even
e ≥ 2 with (ℓ − 1) | 2e, the number Chr.V of the trivial character is 0. Corollary (b) is, for ℓ ≥ 5,
this statement, if the group Chr.V of the Lean is identified with Shaℓ (Q, e) with enlarged coefficients.
At the level of the definitions this is so: Chr.V is the length of the group of classes that vanish
on the decomposition groups above ℓ and on all inertia groups (ShaFamMono.lean:30–32). The
cohomology of the Lean is continuous cohomology: H1U is defined by restricting the representation
to the subgroup (RestrictionMemInvariantH1.lean:35), and the group H1 is the quotient of the
continuous cocycles by the coboundaries (SeparatedDvrSecondCohomologyMem.lean:42–66). A
full identification of the coefficients and of the local conditions of the Lean with those of (III.251)
was not checked. So Corollary (b) is a proof of the classical statement, and it is a proof of PropZ
only under this identification. The source proves PropZ independently (X.propZ). Corollary (a) is
proved from Herbrand’s theorem and does not use PsiBound. As said in §4, PsiBound with ψ = 1
needs a field with complex multiplication, so it gives a statement over Q only after a descent. The
precise length over Q, vℓ (Be+1 /(e + 1)) for (ℓ − 1) ∤ (e + 1), is the theorem of Mazur and Wiles
([MW84], from memory).
The field Q(i). For odd ℓ, Shaℓ (Q(i), e) is the sum of the part on which complex conjugation
ȷ acts as +1 and the part on which it acts as −1. The first is Shaℓ (Q, e), and the second is the part
of the character χ4 modulo 4 (restriction and corestriction; not proved here). For odd e the first is
the minus part and the second the plus part; for even e it is the other way round. The minus part
Pfχ
is governed by Be+1 (e odd) and by Be+1,χ4 (e even), where Br,χ = fχr−1 α=1 χ(α)Br (α/fχ ), fχ
the conductor and Br (X) the Bernoulli polynomial (Lean: genB). With this range of summation
B1,1 = + 12 , as in the Lean. Exact computation gives the odd primes dividing the numerator of
Be+1,χ /(e + 1):
156
e χ odd prime factors e χ odd prime factors
1, 3, 5, 7, 9 1 none 2 χ4 none
11 1 691 4 χ4 5
13 1 none 6 χ4 61
15 1 3617 8 χ4 5 · 277
17 1 43867 10 χ4 19 · 2659
19 1 283 · 617 12 χ4 5 · 13 · 43 · 967
21 1 131 · 593 14 χ4 47 · 4241723
23 1 103 · 2294797
25 1 657931
27 1 9349 · 362903
29 1 1721 · 1001259881
31 1 37 · 683 · 305065927
By Corollary (a), Shaℓ (Q, e) = 0 for every odd ℓ when e ∈ {1, 3, 5, 7, 9, 13}, and for every odd
ℓ≠ 691 when e = 11. Grant the description of the two parts given above. Then for E = Q(i) and
t = 10 the minus part of the sum in (III.252) has no contribution from odd e; for even e ≤ 8 and odd
ℓ the statement PsiBound of the source bounds the contribution of ℓ by vℓ 4 p≤e+2 p · Be+1,χ4 .
Q
A prime that divides the Bernoulli number while the group is zero. Take e = 4 and
ℓ = 5. The prime 5 divides the numerator of B5,χ4 /5 = − 52 . The reason is an Euler factor. Let
L5 (s, χ4 ω) be the 5-adic L-function. For r ≡ 1 (mod 4) its value at 1 − r is −(1 − χ4 (5)5r−1 )Br,χ4 /r
[Was82, Thm. 5.11]. At r = 1 this is 0, because χ4 (5) = 1. All its values at integers are congruent
modulo 5 [Was82, Cor. 5.13]. So 5 divides the numerator of Be+1,χ4 /(e + 1) for every e ≡ 0 (mod 4),
e ≥ 4. Consistently, the computation finds the factor 5 for every e ≡ 0 (mod 4) with e ≤ 24 and
for no e ≡ 2 (mod 4) in that range.
The group Sha5 (Q(i), 4) is nevertheless 0; in particular its part on which ȷ acts as −1 is 0.
We prove this. The proof uses local and global class field theory in their standard form and
one computation modulo 25. It is for this field and this twist only. Put E = Q(i), G = GE,5 ,
W = W4 = Q5 /Z5 (−4) and χ = χ5 . For a subgroup H write Hom(H, F5 ) for the continuous
homomorphisms.
Step 1: the invariants. The prime 5 splits in E. So S5 = {v, v ′ } and Ev = Ev′ = Q5 . The
character χ maps G onto Z× 5 , because E is ramified only at 2 and every subfield ̸= Q of the field of
5-power roots of unity is ramified at 5. It maps each of Gv , Gv′ onto Z×5 , because the cyclotomic
character of Q5 is onto. An element w ∈ W is fixed if and only if (u − 1)w = 0 for all u ∈ Z×
−4
5.
Now u4 ≡ 1 (mod 5) for all u, and 24 − 1 = 15 is not divisible by 25. So, for H = G, Gv or Gv′ ,
H 0 (H, W ) = W [5], (III.268)
the subgroup of order 5, and H acts trivially on it. We identify W [5] with F5 by 51 7→ 1.
5
Step 2: the connecting homomorphism. The sequence 0 → W [5] → W −
→ W → 0 gives, for the
same H, the exact sequence
5 δ
H 0 (H, W ) −−→ H 0 (H, W ) −−→ Hom(H, F5 ) −→ H 1 (H, W ). (III.269)
The first map is 0 by Step 1. So the kernel of Hom(H, F5 ) → H 1 (H, W ) is δ(W [5]), a group of order 5
generated by c = δ( 15 ). To compute c, lift 15 to 25
1
∈ W : c(g) is the class of g· 25
1 1
− 25 = (χ(g)−4 −1)/25,
that is,
χ(g)−4 − 1
c(g) = mod 5. (III.270)
5
157
The same formula holds on G and on Gv , Gv′ , so the restriction of c is the generator there. For
u ∈ Z×
5 put λ(u) = (u − 1)/5 mod 5. It depends only on u modulo 25. It is a homomorphism
4
×
Z5 → F5 , because (1 + 5x)(1 + 5y) ≡ 1 + 5(x + y) (mod 25). And c(g) = −λ(χ(g)), because
u−4 − 1 = −(u4 − 1)u−4 and u−4 ≡ 1 (mod 5).
Step 3: a class killed by 5. Let z ∈ Sha5 (E, 4) with 5z = 0. By the sequence of Step 2 continued
one term, z is the image of some ϕ ∈ Hom(G, F5 ). The restriction of z to Gv is 0, so the restriction
of ϕ to Gv lies in the kernel described in Step 2. The same holds at v ′ . So there are a1 , a2 ∈ F5 with
ϕ|Gv = a1 c|Gv , ϕ|Gv′ = a2 c|Gv′ . (III.271)
Step 4: reciprocity gives a1 = a2 . Let rec be the reciprocity map of local class field theory for
Q5 , from Q×5 to the Galois group of the largest abelian extension. In its action on the roots of
unity of 5-power order, rec(5) is trivial, and rec(u) acts as uϵ for u ∈ Z×
5 , where ϵ = ±1 depends
only on the normalisation of rec. So at each of the two places, for u ∈ Z× 5 and an integer k,
c rec(5k u) = −λ(uϵ ) = −ϵ λ(u). (III.272)
Global class field theory says: the homomorphism ϕ factors through the largest abelian quotient of
G, and for every β ∈ E × the sum over all places w of E of the values of ϕ on the local reciprocity
image of β at w is 0. Take β = 2 + i, of norm 5. At a finite place w ∤ 5, β is a unit and ϕ is
unramified, so the term is 0. At the infinite place the term is 0, because C× is connected. There
remain v and v ′ . They correspond to the two square roots of −1 in Z5 . Let v be the place at which
i ≡ 7 (mod 25); indeed 72 = 49 ≡ −1. At v, β ≡ 9 (mod 25) is a unit, and
94 = 6561 ≡ 11 (mod 25), λ(9) = 2. (III.273)
At v ′ , i ≡ −7 and β ≡ −5 (mod 25) is not a unit. Write β = 5/β̄ with β̄ = 2 − i. At v ′ , β̄ ≡ 9
(mod 25), so β = 5u with λ(u) = −λ(9) = −2. The relation is
0 = a1 · (−ϵ) · 2 + a2 · (−ϵ) · (−2) = −2ϵ (a1 − a2 ) in F5 . (III.274)
Hence a1 = a2 . (The unit i gives no condition, as it must: λ(7) = 0, because 74 ≡ 1 (mod 25).)
Step 5: z = 0. The homomorphism ϕ − a1 c : G → F5 is trivial on Gv and on Gv′ . So it is
unramified at the primes above 5. It is unramified at all other primes, by the definition of G. The
field E has no real place. So the extension of E cut out by ϕ − a1 c is abelian and unramified
everywhere. The class number of Q(i) is 1, so this extension is E and ϕ = a1 c. By Step 2 the image
of c in H 1 (G, W ) is 0. Hence z = 0.
Step 6: the whole group. A continuous cocycle G → W has finite image, because G is compact
and W is discrete. So its values lie in the subgroup killed by some power of 5, and that power kills
its class. If z ̸= 0 were a class in Sha5 (E, 4) of order 5k , then 5k−1 z would be a nonzero class in
Sha5 (E, 4) killed by 5. Step 5 excludes this. So Sha5 (Q(i), 4) = 0.
The congruences of Step 4 were recomputed by a script; the element 1 + 2i in place of 2 + i
gives the same relation. The result agrees with Corollary (b): if the description of the two parts
given above is granted, the part with ȷ = +1 is Sha5 (Q, 4), which is 0 because (5 − 1) | 2 · 4.
Why the Bernoulli number is divisible by 5 and the group is 0 (from memory; not proved
here and not used). By the duality of Poitou and Tate, Sha5 (E, 4) is dual to the kernel of
H 2 (G, Z5 (5)) → v|5 H 2 (Gv , Z5 (5)), and by Tate’s local duality and Step 1 each local group has
Q
order 5. A consequence of the main conjecture [MW84] should give the order 5 v5 (B5,χ4 /5) = 5 for
158
the part of H 2 (G, Z5 (5)) with ȷ = −1. So the Bernoulli number would measure a global group
that maps injectively to the local groups. We have not compared this specialisation of the main
conjecture with [MW84]. The proof above does not use it. The class number of Q(ζ20 ), which is 1
(from memory), does not give the result by the method of Proposition G3: two primes of Q(ζ20 ) lie
above 5 and both ramify in the cyclotomic Z5 -tower, so [Was82, Prop. 13.22] does not apply.
Three tests of the normalisations of the source, for the characters of conductor 1, 4, 8 and r ≤ 42:
Br,χ ̸= 0 exactly when χ(−1) = (−1)r (except χ = 1, r = 1, where B1,1 = 12 in the normalisation
above); the number N p≤r+1 p · Br,χ is an integer for N ∈ {fχ , 8} (and N Br,χ alone is not, in 42
Q
cases); and its absolute value is at most 16N r+1 r! (the largest ratio found is 1/32). For 1 ≤ e ≤ 40
the sum ℓ vℓ log ℓ over the odd primes in the numerators above is at most 2 log 16 · 4e+2 (e + 1)! ,
P
the bound that PrimeBound gives for Q(i).
Sizes. For the field of Theorem G the bound for the minus part in the table is 3√· 4n · t2 log(tN )
with n ≤ 2x2+8ε and t ≤ x401 : a power of x, about x804 . This upper bound is below b = exp(xε /2)
only for very large x: for ε = 1/53, from about log x = 742 on (script). The crossing illustrates a
sufficient scale, obtained from an upper bound. It gives no lower bound on the scale at which the
actual cohomology satisfies Theorem G.
6. Assumed here, from the source
(i) Theorem G. The statement VWin (ExistsPageN.lean:49). It is proved by vwin_of_row
from Row0b, and Row0b by galoisLocalGlobalStrong_recParameter. Both are among the
declarations the final theorem depends on, and vwin_of_row depends only on the three
standard axioms.
(ii) The inequality (III.257). The statement LinkN. Its proof in the source is in the theorems
final_alt_alt and linkN_of_twGood_galoisPreconditions. Only the statements were
read. The proof names bounds at the primes 2 and 3 (boundAtTwoThree_linked_of_recT)
and a family of “Galois preconditions”.
(iii) The fact that dj = D0 (SmallLevelRankEq). The meaning of ηj (w) is not in this list of things
taken from the source, because it is not established: the lattice definition (etaW, locLen,
eta_units) is for the number of a family, and the declarations TieEta and etaN_eq_etaJ,
cited earlier for the tie with the field etaJ of the record, are not among those the final theorem
depends on (§2). The identification of ηj (w) with the defect of the index of lattices of Part II,
§6.8 is open. Nothing proved here uses it.
(iv) The identification of the data. That the numbers shaLen in the record of the construction are
the lengths hℓ,e of the field E (nd.sha_eq); and the passage from Theorem G to HypV described
in §1: the change of scale (winPairs), the bound for the number of twists (RankVal and the
theorem eventually_rank_le_rpow_of_lt_one_div_fiftytwo) and supplyOfAdmOf. The
statements and the proof of winPairs were read; the other proofs were not.
(v) Every row of the table of §4, the hypotheses of the row for Case I, and the inputs listed after
it.
What this section proves, and what it assumes. Proved here outright: Lemma G1 and
Proposition G2. Proved here from theorems of [Was82], checked in the first edition (Propositions
6.16 and 13.22, Lemma 13.16, Theorem 6.17, Corollary 5.14): Proposition G3 and the corollaries (a),
(b), (c); corollary (b) also uses Lemma G4. Proposition 13.22 does not state the compatibility with
Π; the proof of Proposition G3 supplies it. The divisibility of Be+1,χ4 /(e + 1) by 5 for e ≡ 0 (mod 4)
159
follows from Theorem 5.11 and Corollary 5.13 of the same book. Proved here from Minkowski’s
lattice point theorem ([Neu99], Ch. I, (4.4) with its strict inequality, and (5.2); checked in a copy):
Lemma G4. Proved here from the standard statements of local and global class field theory (the
reciprocity law for principal ideles, the action of the local reciprocity map of Q5 on roots of unity,
and the triviality of the Hilbert class field of Q(i)), for this single field and twist: Sha5 (Q(i), 4) = 0.
Not proved here: the description of the two parts of Shaℓ (Q(i), e) by characters; the remark on the
order of H 2 (GE,5 , Z5 (5)) after the proof for Q(i); the two identifications with the Lean stated in
§5; and the meaning of ηj (w) as a length of a quotient of lattices (§2). Computed exactly: the table
of §5 and the three tests. Taken from the source: Theorem G, (III.257), items (i)–(v). Cited as
published theorems: the correspondences in the last column of the table of §4, and the theorems of
Soulé, Voevodsky and Rost, Mazur and Wiles, and Kurihara quoted in §3. They are cited from
memory, except where a theorem or a section is named; a named place in an article was compared
by a reviewer, not by us (see “Citations” at the head of the section).
What is in the literature and what is not. We sort the statements of the source named in
this section into three lists. The source proves all of them again.
(A) Published, and cited with the place. FerreroWashington: [FW79] (from memory), in the
form with bounded rank by [Was82, Prop. 13.23] (read). The duality theorems in LitCoh: Poitou
and Tate [NSW00, §8.6]; for Tate’s local theorems, the formula of Greenberg and Wiles and the
local theorems of Bloch and Kato the attribution is from memory. BhattScholzeIntegral: [BS23,
Thms. 1.2 and 5.6] (compared by a reviewer). We did not compare the Lean statements with the
published ones word by word. PropZ does not belong to this list: it is not cited from a publication.
Its classical form is Corollary (b), proved in §5 from class numbers; that Corollary (b) is PropZ
depends on the identification of Chr.V stated in §5, which was not checked in full.
(B) A published counterpart exists, and it covers less or has not been matched. PsiBound:
an upper bound for a cohomology length in terms of the valuation of a Bernoulli number with
explicit integral factors. It corresponds to a consequence of [MW84] (from memory); general
bounds from Euler systems are in [Rub00, §2.2], and the specialisation is not supplied. Row0b: a
continuous semisimple representation, with predicted weights and the local comparisons in PropsB;
compare [CH13], [Car12], [Car14, Thm. 1.1], with normalisations and hypotheses still to be matched.
KatoTwoW: the two rational Selmer spaces of Bloch and Kato vanish under the attachment, Used
and WitAQ hypotheses; Kato [Kat04, §14.1, Thm. 14.2(1)] proves finiteness of the discrete Selmer
groups at the critical twists other than the central one. LemmaK: Soulé gives finiteness ([Sou79], from
memory) and a quantitative bound away from small primes ([Sou03, Thms. 2 and 3], compared by
a reviewer). We have not identified a published estimate for LemmaK in its full small-prime range.
OneCharU_s: the method of congruences is illustrated by [BC04, Thm. 1.1], which constructs a
Selmer extension; the upper bound is in list (C).
(C) Appears to be new, or not identified in the literature. Theorem G itself (VWin); the
bounds for the plus part, CountI, OneCharU_s, PropGK, CountII, CountIII, CountV, SumIn,
SumOut; the uniform statement ABound with the window, ThmBab; LinkN; the statement
SymplecticFourImageDense on GSp4 ; and LemmaK in its range of small primes. Each of them is a
theorem of the archive that the final theorem depends on. Three of these names are not explained
in the body. SumOut says: for the odd primes ℓ ∈ / B together, the sum of the plus parts over E,
weighted by log ℓ, for the twists e < t, is at most (N t)C , with the level N of the source. PropGK
says: there is a nonzero integer, of logarithm bounded by a power of N (e + 1) times log p for an
auxiliary prime p (and a factor that depends on the orbit), such that for every ℓ = ̸ p twice its
ℓ-adic valuation bounds the sum of the normalised lengths over an orbit of characters. CountV is
described after the table of §4. For ThmBab and SumIn the table of §4 names related work (Ferrero
160
and Washington, Iwasawa, Soulé); that work is not a counterpart of the statements. An upper
bound for the plus part of the groups Shaℓ (E, e), e ≥ 1 (groups of twisted Galois cohomology; they
are not ideal class groups), by a congruence number of automorphic forms, uniform in the field and
the twist, is not a published theorem as far as we know. Every statement of novelty in this list is
provisional: it is a judgment made without a search of the literature, and it was not tested.
What remains.
(i) To read the proof of one row of the table completely. The minus part (PsiBound) is the
natural first choice: it concerns only abelian fields and Bernoulli numbers.
(ii) To state OneCharU_s with its hypotheses on ρ (Adm3U_s, Adm2A) in classical terms, to compare
them with Adm3 and Adm2, and to find how the nonvanishing integer of PropGK is produced.
(iii) To identify the published counterpart of the rows for GSp4 .
(iv) To establish the meaning of ηj (w): to find in the proof of LinkN (linkN_of_twGood_
galoisPreconditions) the statement that ties the field etaJ of the record to the lattice
length etaW at the twisted primes, and to check that the final theorem depends on it. The
declarations TieEta and etaN_eq_etaJ do not serve: the final theorem does not depend on
them. Until then the identification of ηj (w) with the defect of the index of lattices of Part II,
§6.8 is not established.
√
(v) To check the constants in the passage from the three parts to b (vwin_of_joint), and to
evaluate the constant cψ of §1 for the trivial character.
(vi) To prove the description of the parts of Shaℓ (Q(i), e) by characters, and to compare the
remark on H 2 (GE,5 , Z5 (5)) in §5 with [MW84].
(vii) To state PsiBound over Q by descent, and to match the normalisations of Row0b with
[CH13, Car12, Car14].
Computations
The numerical statements in this paper were computed with the following scripts (Python, with
mpmath and sympy); each prints the numbers quoted.
e1_running_example.py G, Gχ , Rq,d , the table for q = 3, d = 23, P and M (Part I, §I.1,
I.5, I.6; Part II, §4.1–4.3, §5.2–5.3, §6.7)
e1b_check_cells.py the eight cell constants, by a second method (Part II, §5.2)
e2_lemniscate.py, e2b_size.py the numbers An : divisibility and size (Part I, §I.3; Part II, §6.3)
e4_sigma.py the Weierstrass σ-function: Gaussian size without the divisibil-
ity (Part II, §6.5)
e5_kronecker.py the two-variable function σ(z + w)/(σ(z)σ(w)) of the lemnis-
catic lattice: divisibility at inert primes and size of its mixed
derivatives (an experiment; its numbers are not quoted in the
text)
e6_hecke_theta.py Tn , the period, the rational numbers rk (Part I, §I.4.1)
References
[Alp26] L. Alpöge, [Remark on a proof of the Riemann hypothesis], with a formal proof in Lean 4, 20
November 2026. https://alpo.ge/remark.pdf.
161
[Apo76] T. M. Apostol, Introduction to Analytic Number Theory, Springer, 1976.
[Bar61] V. Bargmann, On a Hilbert space of analytic functions and an associated integral transform,
Comm. Pure Appl. Math. 14 (1961), 187–214.
[BC04] J. Bellaïche and G. Chenevier, Formes non tempérées pour U (3) et conjectures de Bloch–Kato,
Ann. Sci. École Norm. Sup. 37 (2004), 611–662.
[Beh46] F. A. Behrend, On sets of integers which contain no three terms in arithmetical progression, Proc.
Nat. Acad. Sci. U.S.A. 32 (1946), 331–332.
[Bom65] E. Bombieri, On the large sieve, Mathematika 12 (1965), 201–225.
[Bom74] E. Bombieri, Counting points on curves over finite fields (d’après S. A. Stepanov), Séminaire
Bourbaki, exp. 430 (1973), Lecture Notes in Math. 383, Springer, 1974.
[Bom87] E. Bombieri, Le grand crible dans la théorie analytique des nombres, 2nd ed., Astérisque 18, Soc.
Math. France, 1987.
[Bos96] J.-B. Bost, Périodes et isogénies des variétés abéliennes sur les corps de nombres (d’après D. Masser
et G. Wüstholz), Séminaire Bourbaki, exp. 795, Astérisque 237 (1996), 115–161.
[BS23] B. Bhatt and P. Scholze, Prismatic F -crystals and crystalline Galois representations, Camb. J.
Math. 11 (2023), 507–562.
[Car12] A. Caraiani, Local-global compatibility and the action of monodromy on nearby cycles, Duke
Math. J. 161 (2012), 2311–2413.
[Car14] A. Caraiani, Monodromy and local-global compatibility for l = p, Algebra Number Theory 8
(2014), 1597–1646.
[CH13] G. Chenevier and M. Harris, Construction of automorphic Galois representations, II, Camb. J.
Math. 1 (2013), 53–73.
[Che52] P. L. Tchebichef, Mémoire sur les nombres premiers, J. Math. Pures Appl. 17 (1852), 366–390.
[CS67] S. Chowla and A. Selberg, On Epstein’s zeta-function, J. reine angew. Math. 227 (1967), 86–110.
[Dav00] H. Davenport, Multiplicative Number Theory, 3rd ed., revised by H. L. Montgomery, Springer,
2000.
[Dam70] R. M. Damerell, L-functions of elliptic curves with complex multiplication. I, Acta Arith. 17
(1970), 287–301.
[Deu41] M. Deuring, Die Typen der Multiplikatorenringe elliptischer Funktionenkörper, Abh. Math. Sem.
Hansischen Univ. 14 (1941), 197–272.
[DH66] H. Davenport and H. Halberstam, The values of a trigonometrical polynomial at well spaced
points, Mathematika 13 (1966), 91–96.
[Erd32] P. Erdős, Beweis eines Satzes von Tschebyschef, Acta Litt. Sci. Szeged 5 (1932), 194–198.
[FW79] B. Ferrero and L. C. Washington, The Iwasawa invariant µp vanishes for abelian number fields,
Ann. of Math. 109 (1979), 377–395.
[Gal70] P. X. Gallagher, A large sieve density estimate near σ = 1, Invent. Math. 11 (1970), 329–339.
[GGP12] W. T. Gan, B. H. Gross and D. Prasad, Symplectic local root numbers, central critical L-values,
and restriction problems in the representation theory of classical groups, Astérisque 346 (2012),
1–109.
162
[GHY01] W. T. Gan, J. P. Hanke and J.-K. Yu, On an exact mass formula of Shimura, Duke Math. J. 107
(2001), 103–133.
[Gra84] D. R. Grayson, Reduction theory using semistability, Comment. Math. Helv. 59 (1984), 600–634.
[HB95] D. R. Heath-Brown, A mean value estimate for real character sums, Acta Arith. 72 (1995),
235–275.
[Hon70] T. Honda, On the theory of commutative formal groups, J. Math. Soc. Japan 22 (1970), 213–246.
[How79] R. Howe, θ-series and invariant theory, in Automorphic Forms, Representations and L-functions,
Proc. Sympos. Pure Math. 33, part 1, Amer. Math. Soc., 1979, 275–285.
[Hur99] A. Hurwitz, Über die Entwicklungskoeffizienten der lemniskatischen Funktionen, Math. Ann. 51
(1899), 196–226.
[Jac86] H. Jacquet, Sur un résultat de Waldspurger, Ann. Sci. École Norm. Sup. 19 (1986), 185–229.
[Kat04] K. Kato, p-adic Hodge theory and values of zeta functions of modular forms, Astérisque 295
(2004), 117–290.
[Kud84] S. S. Kudla, Seesaw dual reductive pairs, in Automorphic Forms of Several Variables (Katata,
1983), Progr. Math. 46, Birkhäuser, 1984.
[Kur92] M. Kurihara, Some remarks on conjectures about cyclotomic fields and K-groups of Z, Compositio
Math. 81 (1992), 223–236.
[Lan05] E. Landau, Über einen Satz von Tschebyschef, Math. Ann. 61 (1905), 527–550.
[Leg08] A.-M. Legendre, Essai sur la théorie des nombres, 2nd ed., Paris, 1808.
[LT65] J. Lubin and J. Tate, Formal complex multiplication in local fields, Ann. of Math. 81 (1965),
380–387.
[Mon71] H. L. Montgomery, Topics in Multiplicative Number Theory, Lecture Notes in Math. 227, Springer,
1971.
[MV07] H. L. Montgomery and R. C. Vaughan, Multiplicative Number Theory I. Classical Theory,
Cambridge Univ. Press, 2007.
[MW84] B. Mazur and A. Wiles, Class fields of abelian extensions of Q, Invent. Math. 76 (1984), 179–330.
[Neu99] J. Neukirch, Algebraic Number Theory, Grundlehren Math. Wiss. 322, Springer, 1999.
[NSW00] J. Neukirch, A. Schmidt and K. Wingberg, Cohomology of Number Fields, 1st ed., Grundlehren
Math. Wiss. 323, Springer, 2000.
[Pag35] A. Page, On the number of primes in an arithmetic progression, Proc. London Math. Soc. 39
(1935), 116–141.
[Ral87] S. Rallis, L-functions and the Oscillator Representation, Lecture Notes in Math. 1245, Springer,
1987.
[Rub00] K. Rubin, Euler Systems, Ann. of Math. Studies 147, Princeton Univ. Press, 2000.
[Sie35] C. L. Siegel, Über die analytische Theorie der quadratischen Formen, Ann. of Math. 36 (1935),
527–606.
[Sou03] C. Soulé, A bound for the torsion in the K-theory of algebraic integers, Doc. Math. Extra Vol.
Kato (2003), 761–788.
163
[Sou79] C. Soulé, K-théorie des anneaux d’entiers de corps de nombres et cohomologie étale, Invent. Math.
55 (1979), 251–295.
[Ste69] S. A. Stepanov, On the number of points of a hyperelliptic curve over a finite prime field, Izv.
Akad. Nauk SSSR Ser. Mat. 33 (1969), 1171–1181.
[Stu76] U. Stuhler, Eine Bemerkung zur Reduktionstheorie quadratischer Formen, Arch. Math. 27 (1976),
604–610.
[Tit86] E. C. Titchmarsh, The Theory of the Riemann Zeta-function, 2nd ed., revised by D. R. Heath-
Brown, Oxford University Press, 1986.
[TZ21] J. Thorner and A. Zaman, Refinements to the prime number theorem for arithmetic progressions,
Math. Z. 306 (2024), article 54; arXiv:2108.10878.
[Was82] L. C. Washington, Introduction to Cyclotomic Fields, Graduate Texts in Math. 83, Springer, 1982.
[Wei52] A. Weil, Sur les “formules explicites” de la théorie des nombres premiers, Comm. Sém. Math.
Univ. Lund, tome supplémentaire (1952), 252–265.
[Wei64] A. Weil, Sur certains groupes d’opérateurs unitaires, Acta Math. 111 (1964), 143–211.
[Wei76] A. Weil, Elliptic Functions according to Eisenstein and Kronecker, Springer, 1976.
[Wey25] H. Weyl, Theorie der Darstellung kontinuierlicher halb-einfacher Gruppen durch lineare Transfor-
mationen. I, Math. Z. 23 (1925), 271–309.
164