Reading the formal proof

The Riemann Hypothesis via a chain of theta lifts: a human-readable account of the Lean proof

In short

  • The Lean theorem proves GRH for every Dirichlet character; RH is the trivial-character case. Our independent build is clean: 175,187 modules, Mathlib's own statement, only the three standard axioms.
  • The whole square-root saving is one balance: relative to the main term the error is about x/m + m for a chain of m theta lifts, smallest at m ≈ √x.
  • Both sides of that balance are paid by one inequality, the divisibility bound. Of the charges of the same nominal size, the energy and slope budgets are slack, and the "signature drift" is an artefact smaller than its bound by a factor of x1131.
  • The real cost is one m used twice: m lifts, each charged a Bloch–Kato rank of m. Charging each lift for the whole chain below it would give x/m + m² and stop at x2/3, the barrier in the author's outline; the code avoids it by charging every local piece to exactly one lift.
  • The last step is a windowed Landau theorem: a one-sided bound on windows [Y, Y²] rules out zeros. The Tate–Shafarevich step rests on formal proofs of Ferrero–Washington, Poitou–Tate and CM theory, all proved in Lean.

How to read the labels. Exact checked against the list of declarations the final theorem uses. Verified read directly in the source. Reading our interpretation; a specialist should confirm it. Ours a finding from this trace that we had not seen posted elsewhere as of 20 November 2026.

1 Introduction

1.1 What is proved

Theorem 1 (Main theorem, UpperTheory.EveryCharacter). For every modulus \(N\ge 1\), every Dirichlet character \(\chi \bmod N\) and every \(s\in\mathbb{C}\) with \(\operatorname{Re}s > 1/2\), if \(\chi\neq 1\) or \(s\neq 1\) then \(L(s,\chi)\neq 0\).

The Riemann Hypothesis is the trivial-character case. In the Lean source: \[\begin{gathered} \texttt{riemannHypothesis} := \texttt{rhOfAll},\qquad \texttt{rhOfAll} := \texttt{rh}\_\texttt{of}\_\texttt{every allChi},\\ \texttt{allChi} := \texttt{FinalTheorems.everyCharacter}. \end{gathered}\] Exact (files Main.lean, MainTheorems.lean and MainConclusion.lean).

1.2 What this document is, and is not

This is an account of the formal proof. Every statement below is either a paraphrase of a Lean statement, a summary of a Lean proof, or our interpretation, and we say which. It is not an independent proof: we have not re-derived the classical inputs (Ferrero–Washington [4], Poitou–Tate duality [13], CM theory [14] and others), which are proved inside the development and checked by the Lean kernel, and some proofs are summarised rather than reproduced line by line. Where we have not traced something, we say so (Section 11).

1.2.0.1 Evidence labels.

1.3 The idea in one paragraph

Fix a character \(\chi\) and a large \(x\). The proof attaches to \(\chi\), and to two auxiliary primes \(q\equiv 3 \pmod 8\), \(d\equiv 7\pmod{16}\), a family of arithmetic objects: Hermitian lattices indexed by words \(u\), carried along a chain of \(m\) theta lifts between unitary groups whose signature moves from \((k-1,m)\) to \((k+m-1,0)\), one step per lift. One quantity, \(\mathrm{ASK}\), is computed twice. At the finite places it is a sum over primes of Legendre-type counts weighted by \(\chi\); at the infinite places it is controlled by Chowla–Selberg periods, giving a main term \(\mathrm{MAIN}\). A slope inequality for one common lattice joins the two and yields \(\mathrm{ASK}\le \mathrm{MAIN}+ \mathrm{Err}\), with \(\mathrm{Err}/(\text{main}) \approx x/m + m\). With \(m\approx\sqrt{x}\) this gives a bound of size \(x^{1/2+O(\varepsilon)}\) for a weighted prime sum. Fourier analysis on a small group isolates the prime sum of \(\chi\), and a windowed Landau theorem converts the one-sided bound into zero-freeness.

The author’s outline [1] describes the chain of lifts as playing the role of multiplicity in Stepanov’s method [5], and says that charging the lifts one by one would stop at exponent \(2/3\). Section 9 reconstructs that \(2/3\) from the code’s own quantities and shows how the code avoids it.

1.4 Structure of the proof

Figure 1
Figure 1. The six steps. The square root is paid once, in the balance of step C, by the divisibility error of step D. Steps E and F have room to spare.
  1. Windowed Landau (Section 3). A one-sided bound on windows \([Y,Y^{1+\delta}]\) rules out zeros.

  2. Fourier isolation (Section 4). Averaging over a thick set isolates the prime sum of the target character; the auxiliary characters are made harmless by choosing \(q,d\) with a zero-density estimate.

  3. The size bound (Section 5). The error relative to the main term is about \(x/m+m\).

  4. The divisibility layer (Section 6). The degree identity, the common lattice, the rank bound and local divisibility.

  5. Near-orthogonality (Section 7). The words are orthogonal up to \(\exp(-x^{\varepsilon/8})\).

  6. The \(\text{Ш}\) window (Section 8). Tate–Shafarevich losses over cyclotomic layers of degree about \(x^2\), resting on formal Iwasawa theory.

2 Notation and parameters

Throughout, \(\varepsilon\in(0,1/52)\) is fixed and \(x\) is large.

Symbol Value or window Meaning
\(b\) \(\exp(x^{\varepsilon})\) common weight; every polynomial loss is negligible against it
\(a\) \(x\cdot b\) top weights (Scale.a)
\(\delta_1\) \(x^{-\varepsilon/8}\)
\(m\) \(\sqrt{x\delta_1}\le m\le 2\sqrt{x\delta_1}\), even number of lifts; \(m\approx x^{1/2-\varepsilon/16}\)
\(n=|H|\) \(x^{2+8\varepsilon}\le n\le 2x^{2+8\varepsilon}\) order of the character group of the host
\(f\) \(2n\) number of archimedean places (CM degree)
\(k\) \(x^{9/2+10\varepsilon}\le k\le 2x^{9/2+10\varepsilon}\), even second size parameter; word length
\(t\) odd, \(3m<t\), \(t\ge x^{377}\) size of the cyclotomic twist block
\(r\) \(k+m-1\) starting rank
\(B_0\) \(f\cdot m\cdot b\cdot x^{\varepsilon}\) per-lift budget (sizeBudget)
\(n_T,\ n_{T2}\) \(n_T\ge x^{-\varepsilon/8}n/2\), \(n_{T2}/n\ge x^{-\varepsilon}\) numbers of top and thick points
\(\eta=\eta_{Sz}\) \(\le b/(t^3 f k^3)\) spread of the weights across lifts

Exact The windows for \(m,n,k\) are the literal content of RankVal (.../GlobalTwisting/PairRowIffQdRows.lean:28); \(m\) and \(k\) are picked as even numbers in their windows by exists_ranks_rankVal_frame. Concretely exists_even_between takes \(m=2\lceil\tfrac12\sqrt{x\delta_1}\rceil=2\lceil\tfrac12x^{1/2-\varepsilon/16}\rceil\), so for fixed \(\varepsilon>0\) the chain is slightly shorter than \(\sqrt x\) (\(m/\sqrt x\to0\)); this was pointed out by hhh2210 in the community repository’s issue #5. \(B_0\) is a definition, not an estimate: SizeData_2.B0 := sizeBudget D.R x eps, and the hypothesis BNBounds is proved by equality (\(B_0 = 2|H|\cdot m\cdot b\cdot x^\varepsilon\)).

Figure 2
Figure 2. The sizes of the family as powers of \(x\) (up to \(x^{O(\varepsilon)}\)). Only \(m\approx\sqrt x\) is chosen; the exponents \(2\) and \(9/2\) of \(n\) and \(k\) follow from it (the text). The weight \(b=\exp(x^\varepsilon)\) is larger than every power of \(x\), so every polynomial loss is negligible against it.

2.0.0.1 Ranks and signatures.

Verified (Formalism/LocalRingCardinality/Params.lean) \[N(j) = k + j(t+1),\qquad \operatorname{rankY}(j) = r + jt,\qquad \operatorname{sig}(j) = (k+j-1,\ m-j),\qquad j=0,\dots,m.\] So the chain starts at signature \((k-1,m)\) and ends at the definite signature \((k+m-1,0)\); each lift moves one unit across and adds \(t+1\) rows.

Figure 3
Figure 3. The chain of theta lifts. Each lift moves the signature one step towards the definite group and adjoins one block of \(t+1\) rows. The indefinite part \(m\) of the start is both the number of lifts and, in the text, the Bloch–Kato rank of every piece.

3 Step A: from a bound on windows to the absence of zeros

3.1 The quantity bounded

Definition 2 (Deduction.Gchi, Formalism.Gz). For \(y\ge 1\) let \(\operatorname{ramp}_\chi(y)=\sum_{n\le y}\Lambda(n)\chi(n)\,(y/n-1)\). Put \[G_\chi(y) = \begin{cases} \operatorname{ramp}_\chi(y) - y\log y + (1+\gamma)y, & \chi = 1,\\[2pt] \operatorname{ramp}_\chi(y) + y\,\dfrac{L'(1,\chi)}{L(1,\chi)}, & \chi\neq 1. \end{cases}\] For \(\chi=1\) this is \(G_\zeta(x) = \sum_{n\le x}\Lambda(n)(x/n-1) - x\log x + (1+\gamma)x\) (Formalism/StatementOC.lean:13).

\(\operatorname{ramp}_\chi(y) = \int_1^y \psi(t,\chi)\,dt/t\) is the logarithmic Riesz mean of the Chebyshev function, and \(G_\chi\) is its remainder after the main term. RH is equivalent to \(G_\zeta(x) = O(x^{1/2+\varepsilon})\).

Remark 3 (Verified). In one place in the development the von Mangoldt function appears under the name abelian_2, defined as ArithmeticFunction.vonMangoldt. Searches for the usual names miss it; Step 5 of the author’s timeline (“definitively switch to \(\Lambda\)”) is exactly the expression \(G_\zeta\) above.

3.2 The windowed Landau theorem

Definition 4 (CharFamily). A character family is a finite set of Dirichlet characters \(\chi_i \bmod N_i\) with weights \(a_i>0\), the trivial character occurring only with modulus \(1\). Put \(G_m(y) = \sum_i a_i \operatorname{Re}G_{\chi_i}(y)\).

Theorem 5 (Windowed Landau, zeroFree_of_rpow_window, Exact). Let \(0<\theta<1\), \(\delta>0\), \(C\) real. Suppose that for every \(Y_0\) there is \(Y\ge Y_0\) with \[G_m(y)\le C\,y^{\theta}\qquad\text{for all } y\in[Y,\,Y^{1+\delta}].\] Then \(L(s,\chi_i)\neq 0\) for every \(i\) and every \(\operatorname{Re}s>\theta\).

Proof sketch (from the Lean proof). The proof goes through lemmaL and landau_window. poleData builds a function \(\Phi\) (the Mellin transform of \(G_m\), assembled from the characters and their inverses) whose poles in \(\operatorname{Re}s>0\) are exactly the zeros of the \(L(s,\chi_i)\) and of their conjugates, and shows the zeros are isolated and have real part at most \(2\) (landau_pole_re_le). Suppose a zero \(\rho_0\) has \(\operatorname{Re}\rho_0>\theta\). landau_window_key produces, from the window hypothesis (in the integrated form WindowHypU), a real \(\sigma\) and a growth bound for the Mellin integral restricted to windows. landau_dichotomy then splits into two cases: either there is a zero-free disc around a real point with a zero on its boundary, or there is an isolated real zero \(\beta\) in the region; landau_nearest picks the nearest pole. In both cases the window bound contradicts the growth forced by the pole (landau_growth_contra, landau_window_contra). ◻

Figure 4
Figure 4. The hypothesis of the windowed Landau theorem (on a \(\log\log\) scale each window \([Y,Y^2]\) has the same length \(\log 2\)). A sequence of windows reaching to infinity is enough to rule out zeros with \(\operatorname{Re}s>\theta\).

Remark 6 (Why one sign is enough, Reading checked by arithmetic). In the levels actually used, Level.cyc psi is the cyclic group generated by a primitive \(\psi\), each of its \(o-1\) nontrivial members weighted \(1/(3o+1)\). Every Dirichlet coefficient of the combined series is then at least \(\Lambda(n)\bigl(1-\tfrac{o-1}{3o+1}\bigr)>0\). Positive coefficients are exactly what Landau’s argument needs, so an upper bound on \(G_m\) alone forces zero-freeness. The classical theorem needs the bound for all large \(y\); here a sequence of windows reaching to infinity suffices.

Remark 7 (Verified). A different Landau theorem, zero_free_of_Chebyshev_growth in the namespace Formalism.HC.Landau, is proved in the archive but not used.

3.3 The square-root bound and the assembly

Theorem 8 (UpperTheory.Comparison_2.core_gm, Exact). For every level \(\Lambda\) and \(\varepsilon>0\) there is \(x_1\) such that for all \(x\ge x_1\), every frame and host satisfying the hypotheses EAtH and CoreAtH, \[\Lambda.G_m(x)\le x^{1/2+(e_D+e_L+7)\varepsilon}.\]

CoreAtH packages the data record of Section 6 together with the size windows, in particular RestSize: \(x^{1/2-\varepsilon}\le m\) and \(k\le n^2x^{1/2-\varepsilon}\). The chain from Theorem 8 to Theorem 1 is (Verified): formB \(\to\) zeroFree_of_rpow_window \(\to\) level_of_tiedWin, zeroFreeOn_of_tiedWin_cyc \(\to\) zeroFreeAll_of_topG \(\to\) imprimitive \(\to\) EveryCharacter.

Corollary 9. Every later step only has to hold for \(x\) in windows. Exact The \(\text{Ш}\) window of Section 8 is stated on \([Y,Y^2]\) with \(Y\ge X^{1/J}\), and supplyChi turns it into host fields on windows with exponent \(3/2\) (HostExW), so the Landau step is applied with \(\delta=1/2\), that is, on windows \([Y,Y^{3/2}]\). We thank Manvir Jaswal, whose guide [6] states the windows this way, for the correction.

4 Step B: isolating the target character

4.1 The group and its characters

The host comes with a finite abelian group \(H\) and a group \(E_0\) with four characters \(0,\psi,\chi,\chi\psi\) (psC, chC, chpsC). Exact (CharacterSignsAtBaseClass) At the base class \(g_0\): \[\chi(g_0)=-1,\qquad \psi(g_0)=+1,\qquad \chi\psi(g_0)=-1.\] Define \[G_{\mathrm{sig}}= G_0 + G_\psi - G_\chi - G_{\chi\psi}.\] This is the Fourier projection onto the class of \(g_0\); its trivial part \(G_0\) carries the prime sum of the target.

Figure 5
Figure 5. Fourier isolation on \(E_0\). The signs of \(G_{\mathrm{sig}}\) are the conjugate character values at the base class, so averaging keeps exactly the class of \(g_0\). The three nontrivial terms are bounded by \(x^{1/2+\varepsilon}\) each, because the auxiliary primes were chosen to make them so.

4.2 An exact identity

Proposition 10 (Deduction.ch_balance, Exact). For every \(a>0\), \(t\in H\), weights \(w>0\) on the admissible set and \(N\) large enough, \[\mathrm{ASK}(D,N,a,t,w) - \mathrm{MAIN}(\ell,a,w) = \operatorname{De}(D,\ell,a,t,w),\] where \(\operatorname{De}\) is an explicit combination of the \(G_\psi\) evaluated at \(a/w\).

This is an identity, not an inequality: it expands the counted quantity \(\mathrm{ASK}\) over all characters of \(E_0\times H\) (askExp). Averaging over the thick set collapses the character sum.

Proposition 11 (ThickPartTranslateMean, Exact). For a thick set \(T_2\) of density \(\delta_2 = |T_2|/|H|\), \[\frac{1}{|H|}\sum_{u_0\in H}\frac{1}{|\Phi|}\sum_{\substack{g\notin\Phi\\ g \text{ thick at } u_0}} \sum_{\psi}\overline{\psi(g)}\, b\,G_\psi(A/b) \;=\; \frac{\delta_2\, b}{2}\,G_{\mathrm{sig}}(A/b).\]

4.3 The growth inequality

Theorem 12 (GrowthInequalityExplicit, Exact). Under GrowthInequalityHypotheses (which include \(\mathrm{topExcess}\le \mathrm{Err}\) and ValuesAtX), \[|T_T|\, m\, b\,\frac{\delta_2}{2}\Bigl(G_\zeta(x) - 3x^{1/2+\varepsilon} - \mathrm{moveCost}\Bigr) \le \mathrm{Err}+ |T_T|\, m\cdot 2f_{Pl}(x+1)\,b\,\theta + |T_T|\, m\cdot 6f_{Pl}(x+1)(\eta+1),\] where \(\mathrm{topExcess} = \frac{1}{|H|}\sum_{u_0}\sum_{t\in T_T}\sum_{j}\bigl(\mathrm{ASK}- \operatorname{Re}\mathrm{MAIN}\bigr)\) and \(\mathrm{moveCost} = C_d + 3(\log(x+1)+C_d+C_L\log x)\).

The term \(3x^{1/2+\varepsilon}\) comes from signal_term: for \(|y-x|\le 1\), \[G_\zeta(x) - 3x^{1/2+\varepsilon} - \mathrm{moveCost} \le \operatorname{Re}G_{\mathrm{sig}}(y),\] using ValuesAtX: \(\|G_\psi(x)\|\le x^{1/2+\varepsilon}\) and \(\|\ell(\psi)\|\le C_L\log x\) for the three nontrivial characters of \(E_0\), and a Lipschitz bound (DerivInt).

Remark 13 (Reading). Dividing Theorem 12 by \(|T_T|\,m\,b\,\delta_2/2\) gives \[G_\zeta(x)\le 3x^{1/2+\varepsilon} + \frac{\mathrm{Err}}{|T_T|\,m\,b\,\delta_2/2}+\dots\] So the whole problem is to prove \(\mathrm{Err}\lesssim |T_T|\,m\,b\,x^{1/2+O(\varepsilon)}\). Sections 5 and 6 do exactly that.

4.4 The auxiliary characters are chosen, not controlled

The characters \(\psi,\chi,\chi\psi\) are twists by the quadratic characters of the auxiliary primes \(q\equiv 3\pmod 8\) and \(d\equiv 7\pmod{16}\) (PrimePair). Their bounds in ValuesAtX come from the following chain, all Exact: QH.pairEx \(\leftarrow\) pair_exists \(\leftarrow\cdots\leftarrow\) count_bad_core \(\leftarrow\) ZD.zero_density_box.

Definition 14 (Box). For a character \(\psi\), \(\mathrm{Box}(Q,C,\tau,Y,X,\psi)\) means: \(\psi\) has a good conductor, \(\|L'/L(1,\psi)\|\le C\log Y\), and \(\|G_\psi(y)\|\le y^{1/2+\tau}\) for all \(y\in[Y,X]\).

Theorem 15 (count_bad_core, Exact). There is \(C\ge 0\) such that for \(Q\ge 3\) and any set \(S\) of primes \(q>F\) each carrying a primitive character of conductor \(fq\le Q\), \(f\le F\), without a zero-free box, \(|S|\le C\,Q^{1/4}\).

The zero-density estimate zero_density_box (compare [5, Ch. 10]) is assembled from a coefficient square bound, zero detection by a mollifier identity, a duality inequality and a pair-twist bound.

Remark 16 (Reading). The auxiliary \(L\)-functions are not proved to satisfy GRH. The proof picks \(q\) and \(d\) outside the few bad moduli. The target character is the only one whose \(L\)-function is not controlled in advance.

4.5 The thick set

Theorem 17 (hostThickExists, Exact). There is \(C>0\) such that for all \(n\) and \(L\ge 2\) there is \(T\subset\mathbb{N}\) with \(4Ls<n\) for \(s\in T\), no relation \(j(s_2-s_1)=i(s_3-s_1)\) with distinct \(s_1,s_2,s_3\in T\), \(0<|i|,|j|\le L\), \(i\neq j\), and \[|T|\ \ge\ \frac{n\,e^{-C\sqrt{\log n\,\log L}}}{L}.\]

Reading The density \(\delta_2\) enters only through the division above and costs \(x^{o(1)}\).

5 Step C: the size bound, where \(\sqrt{x}\) is paid

5.1 The balance

Theorem 18 (proposition_size_bound, Exact). In the abstract setting of GrowthSizeBound, \[n_T\,m\,b\,\frac{\delta_2}{2}\bigl(G(x)-4x^{1/2+\varepsilon}\bigr)\le \mathrm{Err}+ \frac{2n_T m b}{x},\] \[\mathrm{Err}\le x^{C\varepsilon}\Bigl[n_T a_E\Bigl(1+\frac{m^2}{k}\Bigr) + m\cdot B_0 + m\cdot\frac{n_{T2}\,k\,b\,x^{\varepsilon}}{n^2x^{-\varepsilon}}\Bigr],\] and under the windows below, \(G(x)\le x^{1/2+(C+3)\varepsilon}\).

Dividing each error term by the size \(n_T\,m\,b\) of the main term (lemmas SZ.T1–SZ.T4, Verified):

Term After dividing Bound Needs
\(n_T a_E\) \(a_E/(mb)\approx \mathbf{x/m}\) \(\le x^{1/2+\varepsilon}\) \(m\ge \sqrt{x}\) (T1)
\(n_T a_E m^2/k\) \(xm/k\) \(\le 2\) \(k\ge xm\) (T2)
\(m\cdot B_0\) \(2nm/n_T\approx \mathbf{m}\) \(\le x^{1/2+\dots}\) \(m\le\sqrt{x}\) (T3)
\(m\cdot(k\text{-term})\) \(k/(n\,n_T)\) \(\le x^{1/2+\dots}\) \(k\le n^2\sqrt{x}\) (T4)

So, relative to the main term, the error is about \(x/m+m\), and both pieces are \(\sqrt{x}\) exactly when \(m\asymp\sqrt{x}\).

Figure 6
Figure 6. The balance. Lengthening the chain lowers \(x/m\) but raises the divisibility cost. With the code’s accounting (one budget \(B_0\) per lift) the optimum is \(m=\sqrt x\) and the bound is \(x^{1/2}\). Charging each lift for the whole chain below it would make the cost \(m^2\) and stop at \(x^{2/3}\) (the text).

5.1.0.1 Credit.

The balance \(x/m+m\) at \(m\asymp\sqrt{X}\) was stated independently by Jonas Whidden (Lean Zulip) and in Bartosz Naskręcki’s proof guide (nasqret/riemann-rev, §4). What follows in this section and the next (where each side is paid, which charges are slack, how the exponents of \(n\) and \(k\) arise) is Ours.

5.2 The windows are chosen, not forced

Verified The windows used by the size layer are proved from RankVal:

Window Lemma Source Slack in the family
\(m\ge x^{1/2-\varepsilon}\) winL SizeM \(\approx x^{0.93\varepsilon}\)
\(m\le x^{1/2+\varepsilon/2}\) winA SizeM \(\approx x^{0.55\varepsilon}\)
\(xm\le k\) winB SizeM, SizeK \(\approx x^3\)
\(k\le n^2x^{1/2-\varepsilon}\) winM k_up \(\approx x^{5\varepsilon}\)

Verified The constructor exists_ranks_of_frame accepts any even \(k,m\) with \(2\le m\), \(m+3\le k\), \(m\le 2n\le k\); the host construction admissible_ranks_of_CMDatum works for every rank datum, needing only \(q,d>4tr+1\). So \(m\approx\sqrt{x}\) is a literal in RankVal: nothing in the arithmetic construction asks for it.

5.2.0.1 Why \(n\approx x^2\) and \(k\approx x^{9/2}\): the word count. Ours

The one arithmetic link between \(m\) and \(k\) is the number of words. Verified: \(\log\Gamma = C_\Gamma\, m^5 f L\) (gamB, RowsConc.lean:52), the number of words is \(\lceil 3\Gamma^2x^{C_F}\rceil\), words are coded injectively by \(k-2\) bits (exists_injective_words), and Win_at requires \(\log|W|\le(k-3)\log 2\). Hence \(k\gtrsim m^5 f L\). Writing \(m=x^\mu\), \(n=x^\nu\), \(k=x^\kappa\): \[\kappa\ge 5\mu+\nu\ \ (\text{word count}),\qquad \kappa\le 2\nu+\tfrac12\ \ (\text{T4}),\qquad \Rightarrow\ \nu\ge 5\mu-\tfrac12.\] With \(\mu=1/2\): \(\nu=2\) and \(\kappa=9/2\). So the exponents of \(n\) and \(k\) are derived from \(m\approx\sqrt{x}\) and the \(m^5\) word count. m_pow_five_mul_f_mul_L_le proves \(m^5 fL\le 128\,x^{9/2+9\varepsilon}\) against \(k\ge x^{9/2+10\varepsilon}\): the tightest window in the family.

Figure 7
Figure 7. Exponent bookkeeping. The word count (\(\kappa\ge 5\mu+\nu\)) and term T4 (\(\kappa\le 2\nu+\tfrac12\)) only require \(\nu\ge 5\mu-\tfrac12\); any point of the shaded region is a valid family. Only the balance of Figure 6 picks \(\mu=\tfrac12\).

Remark 19 (Reading). Since \(G_\zeta(x)=\Omega(\sqrt{x})\), no valid argument can push \(m\) beyond \(\sqrt{x}\) and win. As far as the constraints above go, the family could be built with \(m=x^{1/2+\delta}\); so the barrier must be paid on the error side, in T3. Section 6 locates it.

6 Step D: the divisibility layer

6.1 The abstract data and the pointwise bound

The size layer works with a record Data of real numbers and a bundle of hypotheses SizeData_2: DIV, EComp5, HComp, Count5, Slopes, QuotientSlopeBoundsWith, LossBounds, SlopeIneq, HHBound, FParams. No automorphic objects appear there. ChainTied (TieGlobal.lean:97) attaches the record to an arithmetic family FamP through TiePts, TieGlobal, TieLocal, TieArith; it is produced by linkCore from fourteen family hypotheses, and only the caller linkGLOBAL7_items is used (Exact).

Theorem 20 (each_unit_bound_statement, Exact). For every point \(u_0\): \(\mathrm{ASK}(u_0)\le\mathrm{MAIN}(u_0)+\mathrm{ErrX5at}(u_0)\).

Figure 8
Figure 8. What each per-lift charge really costs (the text). Only \(E_b\) uses its budget; the drift’s full size is an artefact of taking absolute values, and the energy and slope budgets are almost empty. Bars are schematic: the energy bar would be invisible at true scale.

The terms of \(\mathrm{ErrX5at}\) that matter (Verified, QuotientSlopeBoundsWith.lean:45):

Term From After dividing by \(n_Tmb\) Balance
\(c_{Div}\,n_T\,a\,L_a\) DIV \(\approx x/m\) T1
\(c_{Div}\,m(B_0+N_c)x^{e_{Div}\varepsilon}\) DIV \(\approx m\) T3
\(m\,c_E\,B_0\) EComp5 same order (slack)
\(c_{Sl1}\,m\,B_0\) QuotientSlopeBoundsWith same order (slack)
\(c_{Slope}\,n_T a L_a(1+m/k)\) bookkeeping \(x/m(1+m/k)\) T1, T2

6.2 The slope layer and the cancellation of the main terms

Verified EComp5 says \(|E_e(u_0,u)-m\cdot\mathrm{sumA}\cdot\mathrm{lg}|\le c_E\, m\, B_0\): the chain energy is \(m\) copies of one step’s slope, \(\mathrm{lg}=\log(g_0g_1)\) for the two host primes. Ga_bound5 combines the upper bound \(E_e\le m(\mathrm{sumA}\cdot\mathrm{lg}+c_EB_0)\) with key_identity5, which contains \(+m\cdot\mathrm{sumA}\cdot\mathrm{lg}\); the main terms cancel exactly and the proof ends in linarith. This is the “cancellation of order-\(m\) terms” of the outline (entry 9).

6.3 The local energy and why its budget is slack

Verified The energy is pinned by SizeData_2.EDef: \(E(u,j,w)=\mathrm{Eval}(u,j,w)\), a case split on the residue characteristic of the place \(w\):

\(w\) lies over \(E(u,j,w)\) depends on \(j\)?
\(T_u\) (auxiliary prime of \(u\)) twice the weight deviation \(+\ f\cdot N_m + \dots\) no
\(g_0\) or \(g_1\) (host primes) \(v_p(D_\Gamma)+\mathrm{FTypeG}(j)+v_{NN}+c_g f(N_m^2+m^2\eta_{Sz})\) yes
ramified class at \(d\) \(v_p(D_\Gamma)+\mathrm{FRamD}(j)+\dots\) yes
small \(p\le X''\) \(v_p(D_\Gamma)+\) log terms yes
other \(v_p(D_\Gamma)\) no

At \(g_i\), with \(\pi_{g_i}\bar\pi_{g_i}=g_i\) in the CM field, \(v_{NN}=\bigl(\sum_\tau \mathrm{nuAbs}(u,j,\tau)\,v_w(\tau(\bar\pi_{g_i}))\bigr)/e_w\).

Proposition 21 (Per-lift energy, Exact chain). EComp per lift (adm11) follows from

  • NormDecompositionCond: \(\sum_{w\mid g_i} sz(w,v_{NN}) = \bigl(\sum_\tau\mathrm{nuAbs}(u,j,\tau)\bigr)\log g_i\) (an exact equality);

  • factEtaMean: \(\sum_{w\mid \ell}\mathrm{lw}(w)e_w = \log \ell\) (product formula);

  • DegreeCountBound: \(\bigl|\sum_\tau\mathrm{nuAbs} - \mathrm{fockDeg}(j) - \sum_t a_T(t)\bigr|\le \mathrm{errG}(j)\);

  • TypePrimeSizeBound and the off-\(g\) bounds.

6.3.0.1 The row structure.

Verified (nuAbs_place_fine) At lift \(j=i+1\) and an archimedean place \(\tau\), nuAbs sums \(N(j-1)=k+i(t+1)\) rows \(\mathrm{nuRow}(u,j,\tau,r)=\mathrm{srcPar}(u,\tau,r)+w_{Sz}(j,\tau)-\tfrac12(N(j-1)+N(j)-1-2r)\): the two end rows carry top and off-top weights, the middle rows are \(\approx w_j(j,\tau)\), and the \(i(t+1)\) block rows are \(\approx 0\) up to \(2\eta_{Sz}\). Reading A parameter plus a weight minus a \(\rho\)-type half-sum looks like the Harish-Chandra parameters at lift \(j\); kerPlace sums \(\phi(\nu)=\nu\log\nu-\nu\) over the rows, a Stirling shape.

Proposition 22 (Ours). The real per-lift energy error is \(\mathrm{errG}(j)\approx 2\sum_\tau w_j(j,\tau)\approx 2fb\), with no factor \(m\), against a budget \(B_0=f\,m\,b\,x^\varepsilon\). The energy charge \(m\,c_E B_0\) therefore has a spare factor of about \(m x^\varepsilon\). The slope upper bound is ClaimHigh: \(\mu_{\max}(N)\le\mu_{\max}(M)+G_a+F_o-E_e\) with no error term; slopesN_at_of adds the \(m B_0\) allowance from \(0\le C m B_0\) alone.

6.4 The divisibility inequality

Theorem 23 (Data.DIV, Exact). \[\begin{gathered} \mu_{CC}+h(u)-F_o \le \sum_t\Bigl(m\bigl(\phi(a_T(t))-a_T(t)\log\Omega\bigr)-a_T(t)\log T(u)\Bigr)\\ + c_{Div}\,n_T\,a\,L_a + c_{Div}\,m\,(B_0+N_c)\,x^{e_{Div}\varepsilon}. \end{gathered}\]

6.4.0.1 The height of a chain.

Verified (height_chain_atR) The pairing map \(B_u\) factors through a chain of \(m\) maps \(K_j\), one per lift, with \(\|K_j\|_v^2\le\exp(h_K(j,v))\), and \[\log\|B_u\|_v^2\le h_P+\sum_{j=1}^m h_K(j,v)-\log\kappa_v-\min_p\sum_j\log\lambda_j(p,v).\] This is submultiplicativity of operator norms along the chain. Reading It is where “a chain of theta lifts” becomes an inequality.

6.4.0.2 The per-lift charge is built in.

Verified (ZscOfNum, errB_of_units) The scale data sets, per lift, an error unit \(E_{1u}=x^\varepsilon(B_0+N_c)\), a main term \(l_{3m}(j)=\sum_\rho[(k+j-1)-(m-j)]\,w(j,\rho)\log\Omega\) (the signature difference \(p_j-q_j\) times the weights) and \(l_{3u}=B_0\log\log x\). Summing over \(m\) lifts gives the \(m\,x^\varepsilon(B_0+N_c)\) term.

6.4.0.3 Absorption.

Verified (divisor_inequality_assembly, error_absorption) Every error term inside DIV is bounded against \(m(B_0+N_c)x^\varepsilon\):

Error term Bound used Versus the budget \(mB_0x^\varepsilon\)
\(\sum e_E\) (place heights) \(15C_0 f m b L\) spare factor \(\approx m\)
\(\sum|\delta||\log\Omega|\) \(15C_0fmbL + C_0n_TaL_a\) spare factor \(\approx m\)
\(\sum e_K\) (per-lift polynomial) \(5C_0 m f b L\) spare factor \(\approx m\)
\(e_6,e_P,L_{pass}\) \(\le B_0\) spare factor \(\ge m\)
\(\sum|pq_j-k||w||\log\Omega|\) (drift) \(6mB_0\) full budget
\(E_b\) (rank charges) hypothesis \(C_0mx^\varepsilon(B_0+N_c)\) full budget

6.5 The signature drift is an artefact

Proposition 24 (Ours). The drift term is smaller than the bound used for it by a factor of at least \(x^{1131}\).

Proof. Verified: \(pq_j=(k+j-1)-(m-j)\), so \(pq_j-k=2j-1-m\) (NumbersLink.lean:574). The proof bounds each term \((pq_j-k)\,w(j,\rho)\log\Omega\) by its absolute value before summing (step hA of divisor_inequality_assembly). By SizeData_2.be_close, \(|w(j,\rho)-w(j',\rho)|\le 2\eta\) for all lifts, and szRow_eta_bounds gives \(\eta\le b/(t^3fk^3)\). Since \(\sum_{j=1}^m(2j-1-m)=0\), \[\sum_j\sum_\rho(2j-1-m)\,w(j,\rho)\log\Omega=\log\Omega\sum_\rho\sum_j(2j-1-m)\bigl(w(j,\rho)-w(1,\rho)\bigr),\] so the signed drift is at most \(L_a f(m^2/2)2\eta=fm^2\eta L_a\), while the absolute bound used is \(2fm^2bL_a\). The ratio is at most \(\eta/(2b)\le 1/(2t^3fk^3)\le x^{-1131}\) since \(t\ge x^{377}\). ◻

Figure 9
Figure 9. The signature drift. Its coefficients sum to zero and the weights \(w(j,\rho)\) agree across lifts up to \(2\eta\), so the signed drift nearly cancels. The proof bounds each bar by its absolute value instead, which is where its nominal \(m^2\) comes from.

6.6 The rank bound \(E_b\): where the \(m^2\) really lives

Theorem 25 (Ours, each step Verified). \(E_b\approx m\cdot B_0\), and both factors of \(m\) are structural.

Proof sketch (from the Lean proof).

  1. The outer \(m\) is the number of lifts. errB_of_units sums over \(J\) with \(|J|=m\) (RecNums.p12).

  2. The real obligation. \(E_{1u}\) is set by definition; the content is the field IndefData.ON: for each lift \(j\), \(\sum_{w\in T_w}\log w\,(\mathrm{ram}_w\,\mathrm{err}_0(w,j)+\mathrm{ram}_w J_0(w,j))\le E_1(j)\), supplied on the family side by DIVinN_at and LinkN_at: \(\mathrm{etaSum}(j)\le d_{Z0}(j)\bigl(C(fmbL^2+\log^2x\,N_c)+\sum_\ell\mathrm{vvTerm}(\ell)\bigr)\). The last sum is the \(\text{Ш}\) loss of Section 8, added once and bounded by \(\sqrt b\).

  3. No double counting. linkN_of_gd bounds \(\mathrm{etaSum}(j)\) through \(\mathrm{etaJ}(j,w)=\sum_{i\in\mathrm{locs}(j,w)}\mathrm{etaLoc}(i)\), and by NData.locs_mem, \(i\in\mathrm{locs}(j,w)\iff j(i)=j\wedge w(i)=w\): every local piece belongs to exactly one lift.

  4. The inner \(m\) is a Bloch–Kato rank. The allowance at a small prime \(\ell\) is \(\mathrm{costC}(\ell)=6000\,f\,m\,h_B\,\mathrm{sl}(\ell)+30000f^2t^2n_V\,(\dots)\) with \(h_B=b+S_c+\eta_{Sz}+t+2\in[b,6b]\). The colength bound behind it calls rank_le_row with \(m:=D.R.m\): \[\dim H^1_f(K_w,T)\le\mathrm{locDeg}(w)\cdot m,\] the Bloch–Kato count [2] \(\dim H^1_f=\sum_\tau\#\{\text{non-negative Hodge--Tate weights}\}\), and the family supplies the hypothesis \(|L_o|\le m\) by equality (le_of_eq), from its weight pattern ShapeJumps k m t.

 ◻

Factor of \(E_b\) Source Size
number of lifts \(|J|=m\) \(m\)
rank per piece Bloch–Kato, fixed by equality \(m\cdot\mathrm{locDeg}\)
exponent per piece lift exponent \(e'\) \(\approx b\)
primes \(\sum\log\ell\) over small primes \(\approx x^\varepsilon\)
\(\text{Ш}\) loss \(\sum\mathrm{vvTerm}\le\sqrt b\), once per lift negligible

6.6.0.1 Why the two \(m\)’s are one \(m\). Ours

Verified ShapeJumps k m t (EmbeddingTypes.lean:99), required by admG2 at every small prime, place and embedding, says the Hodge–Tate multiplicity of the twisted lifted representation is the indicator of \(C\cup L\cup[0,t)\): \(k-1\) weights in a far cloud \(C\), exactly \(m\) weights \(L\) near \(0\), and the \(t\)-block \([0,t)\) from cyclotomic twists \(\chi^c\), \(c=1,\dots,t\). With \(r=k+m-1\) and \(n_V=r+t\), the weights \(C\cup L\) have sizes \((k-1,m)\): the starting signature. Reading The \(m\) low weights are the indefinite side of the signature, and the chain has exactly \(m\) lifts because it walks that side down to \(0\). So \(E_b\)’s \(m^2\) is one \(m\) used twice, by construction.

Figure 10
Figure 10. The weight pattern ShapeJumps k m t (schematic, not to scale). The \(m\) low weights are both the indefinite part of the starting signature and the Bloch–Kato rank of each piece.

6.6.0.2 Where \(b\) enters. Verified

The lift exponent is \(e'=e_L(\ell,N,h)=32(2h\,s_N(\ell,N)+1)+\lambda_N(\ell,N,h)\) with \(h=\lfloor h_B\rfloor\), linear in a height \(h_B\in[b,6b]\). Reading A lifting exponent growing linearly with the Hodge–Tate height, as in Fontaine–Laffaille-type bounds. Since both the main term \(n_Tmb\) and the cost \(mB_0\) are linear in \(b\), \(b\) cancels in the balance: \(e'\approx b\) sets the absolute size of \(E_b\), not its dependence on \(m\).

6.7 The \(+m\) side of the balance is the divisibility error

Theorem 26 (Ours, pieces Verified). Both sides of the balance \(x/m+m\) come from DIV: the \(x/m\) term from \(c_{Div}n_TaL_a\) (top weights \(a=xb\)), the \(m\) term from \(c_{Div}m(B_0+N_c)x^{e_{Div}\varepsilon}\), which on the lattice side is \(\mathrm{errB}=\text{drift}+E_b\).

Remark 27 (Verified, unused route). exists_data_completion (PChain.lean:306), proved but not used, builds from the single bound \(\mathrm{ASK}\le\mathrm{MAIN}+c_{Div}n_TaL_a+c_{Div}m(B_0+N_c)x^{e_{Div}\varepsilon}+L_{oP}\) a degenerate data record (one word per point, zero energy) satisfying all of SizeData_2. So even with zero energy and slope charges, the divisibility error alone gives T3.

Corollary 28 (Reading). With the drift an artefact (Proposition 24) and the energy and slope budgets slack, the reason the chain is cut at \(m\approx\sqrt{x}\) rests on one term: \(E_b\), \(m\) lifts times a Bloch–Kato rank of \(m\) per piece. This is consistent with \(G_\zeta(x)=\Omega(\sqrt{x})\).

6.8 The common lattice

Proposition 29 (exists_divStage_of_pk, Exact; Ours as the answer to Whidden’s question). The single lattice \(RR\) shared by all words is \[x\in RR.L\iff \forall w\ \exists d\in\mathcal{O}_{LL},\ d\notin\mathfrak p_w,\ \ d\,x\in\mathcal{O}_{LL}\text{-span}\{y : j_E y\in \mathrm{RRset}(\mu_w)\},\] where \(\mathrm{RRset}(\mu)=\iota_\mu(\Lambda_\mu)\), \(\iota_\mu=\mathrm{multZero}\circ\mathrm{downTo}_{\mu,m}\circ\mathrm{multTop}^{-1}\) is the composite of the \(m\) lift steps, and \(\mu_w=\texttt{muAt }w\) is the \(\xi\)-twist data at \(w\) (prescribedRowLat).

So \(RR\) is the intersection over places of the localisations built from each place’s own character twist: a lattice gcd. Every family carries it as the field FamP.rrDef : DV.RRDef.

Figure 11
Figure 11. The common lattice \(RR\) (prescribedRowLat). Each place contributes the image of its own integral points under the composite of the \(m\) lift steps, built from its own character twist; \(RR\) is their intersection.

Reading This matches the outline’s entry 15 (“glue an integral structure across Hecke characters for a gcd”). Its only visible cost is \(\log\#W\) in the frame inequality.

6.9 The degree identity: how a slope becomes \(\mathrm{ASK}-\mathrm{MAIN}\)

There is no single lemma; it is a chain of four used statements.

Figure 12
Figure 12. The degree identity. The Arakelov degree of the divisible lattice is computed at the finite places (where the prime counts \(\ASK\) enter) and at the infinite places (where the periods \(\MAIN\) enter); the slope inequality on the one common lattice joins them.
  1. Slope \(=\) reference \(+\) local indices (NprimeMuR, supplied through LinkMu \(\Leftarrow\) J.linkMu_of_linkDIV \(\Leftarrow\) LinkDIV): \[\begin{gathered} \Bigl|\mu(Nu\,u)-\mu(NuH\,u)-\sum_{w\in\mathrm{supp}} sz(w,sg(u,w))\Bigr|\le C X'',\\ sg(u,w)=\sum_j\bigl(\mathrm{gain}_j(w)+F_j(w)-E(u,j,w)\bigr), \end{gathered}\] with \(sz(w,y)=2\,\mathrm{lw}(w)\,e_w\,y\). Summed: \(\mu(Nu\,u)\approx\mu(NuH\,u)+G_a+F_o-E_e(u)\).

  2. Gain \(\ge\) prime count (LinkG_at, where \(\mathrm{ASK}\) enters). At an inert place \(w\) over \(p\le X''\): \[\sum_j\sum_t\bigl(\mathrm{vMCS}_{j,t}(p)-\mathrm{loss}\bigr)-m\,\mathrm{fl}(p)\le\sum_j\mathrm{gain}_j(w)-\tau(w),\] \[\mathrm{vMCS}_{j,t}(p)=\sum_{i\text{ odd}}\max\bigl(0,\,a_t/p^i-\mathrm{wPart}(j,t,p,i)\bigr).\] This is a Legendre-type count, like \(v_p(a!)=\sum_i\lfloor a/p^i\rfloor\) but over odd \(i\) with weights removed. Since \(\mathrm{ASK}=\sum_j\sum_t\bigl(\sum_{p\text{ inert}}2\log p\cdot\mathrm{vMCS}_{j,t}(p)+\dots\bigr)\), summing over \(w\mid p\) with the product formula gives \(\mathrm{ASK}-G_a\le L_{oP}\) (loss_bound_at).

  3. Dual slope \(\le\) period term (LinkDIV_at, where \(\mathrm{MAIN}\) enters): \[\mu_{dual}(RR)+h(u)-F_o\le\mathrm{MAIN}(u)+C|T_T|aL_a+Cm(B_0+N_c)x^{C_{abs}\varepsilon},\] \[\mathrm{MAIN}(u)=\sum_t\bigl[m(\phi(a_t)-a_t\log\Omega)-a_t\log T_u\bigr],\qquad \log\Omega=\gamma+\log(\pi/v)-\frac{L'}{L}(1,\chi_d),\] the Chowla–Selberg period [3] (meanPeriod), and \(\phi(a)=a\log a-a\).

  4. Slope inequality (Frame_at): the rank-weighted mean \(s\) of \(\mu(Nu\,u)-h_u\) satisfies \(s-\log\#W-(\text{small})\le\mu_{dual}(RR)\), for the one lattice \(RR\) of Section 6.8.

Theorem 30 (Assembled, Reading with each step Verified). \[\mathrm{ASK}-L_{oP}\le G_a,\qquad \mu(Nu\,u)\approx\mu(NuH\,u)+G_a+F_o-E_e(u),\] \[\operatorname{mean}_u[\mu(Nu\,u)-h_u]\le\mu_{dual}(RR)+\log\#W,\] \[\mu_{dual}(RR)+h(u)-F_o\le\mathrm{MAIN}(u)+\text{errors} \quad\Longrightarrow\quad \mathrm{ASK}-\mathrm{MAIN}\le\text{errors},\] after the \(m\cdot\mathrm{sumA}\cdot\mathrm{lg}\) main terms cancel in Ga_bound5. The finite part of the Arakelov degree of the divisible lattice is a sum of local indices bounded below by the prime counts; the archimedean part is bounded by the CM period main term.

6.10 Local divisibility: the \(\sqrt p\) trick

Lemma 31 (key_minor_dvd, Exact). Let \(A\) be a commutative ring, \(p\) prime, \(w\in A\) with \(w^2=p\). Let \(M,N'\) be \((k+1)\times(k+1)\) matrices over \(A\) with \(M_{ba}^{p^e}\equiv N'_{ba}\pmod p\) for all \(b,a\), and let \(u,u'\in A^{k+1}\) with \(w\mid u_b\) and \(w\mid u'_b\) for \(b\neq0\). Then for every column \(c\), \[w\ \Big|\ \Bigl(\sum_b u'_bN'_{bc}\Bigr)\Bigl(\sum_b u_bM_{b0}\Bigr)^{p^e}-\Bigl(\sum_b u'_bN'_{b0}\Bigr)\Bigl(\sum_bu_bM_{bc}\Bigr)^{p^e}.\]

Exact Used chain: key_minor_dvd \(\to\) kW_ord \(\to\) key3Ord (CoarseMultiplicityOrderBound: for \(g\in GL\) over the Witt vectors \(W(\mathbb F_{p^{n'}})\), the key polynomials after the lift substitution lie in \(W_1\)) \(\to\) third_key_of_order \(\to\) restRawTail1_of \(\to\) stmtEExt_of_15c (ExtendedClosedForm) \(\to\) extendedClosedForm. LinkG is assembled by link_of_local_closed from the first and second closed-form theorems and the local model results, all instantiated from one structure Loc, proved by loc_proved from TraceSpectra.stmtE.

6.10.0.1 The gain, orbit by orbit.

Exact The gain clause of LinkG_at is AuxBlock8.Conclusion1. It is proved from St.starAll_of, one inequality per prime \(p\), lift \(j\) and orbit representative \(t_0\), \[\sum_{s\in\mathrm{tops}(t_0)}(\mathrm{askP}-\mathrm{lo})\le g_U+d_l-P_r-\lceil t_v\rceil-D_p-1, \qquad g_U=\gamma_p\sum_s\alpha(s)-\sum_s\sum_{e\text{ odd}}\min\bigl(\alpha(s)/p^e,\mathrm{cap}\bigr)-d_l,\] then summed (gain_bound_all_of_star, gainBoundAll_add_one_of_gainBoundAllRequired).

Remark 32 (Reading). A congruence mod \(p\) gives whole units of \(p\). Working where \(p\) has a square root extracts half a unit per place, and the only valuation corrections in the gain are sums over odd exponents, the ones where a half unit shows. We did not trace the direct call from key3Ord into \(g_U\).

7 Step E: near-orthogonality of the words

Theorem 33 (LinkS_at, LinkS, Exact). If every word \(u\) is comparable, \(C_0^{-1}\le d_u(u)\Gamma^2/D^*\le C_0\), then

  1. \(\|K(u,u')\|_{HS}^2\le(1+\eta_0)\,d_u(u)d_u(u')/D^*\) for \(u\neq u'\);

  2. \(\|K(u,u)-\mathrm{id}\|_{HS}^2\le e_0\,d_u(u)\);

  3. \(D_{all}\le(1+\eta_1)D^*\);

with \(\eta_0,\eta_1,\sqrt{e_0}\le\exp(-x^{\varepsilon/8})\) and \(\Gamma=\exp(C_\Gamma m^5fL)\).

Verified The comparability premise is discharged where LinkS is consumed: matKF_of feeds it ChosenProviso, proved by chosenProviso_min from the class number formula (ClassNumberFormula_2) and RankTiesF. The bound itself comes from theoremS_link_concrete, from ResidEv, RowsTieP, CountAllPointsClass and four hypotheses ClKppOff (the correction \(K_{pp}\) vanishes off the diagonal), ClKppDiag, ClOff, ClDiag.

Reading The normalised Gram matrix of the words is the identity up to \(\exp(-x^{\varepsilon/8})\). The \(m^5\) in \(\Gamma\) is the same \(m^5\) as the word count of Section 5.2. This is the “approximate orthogonality of columns” of the outline (entry 16), and it does not touch the balance.

8 Step F: the \(\text{Ш}\) window and the classical inputs

Definition 34 (IsChiField, VVle). For \(e,q,d\) and a character \(\psi\), a field \(E\subset\mathbb{Q}(\zeta_N)\) is a \(\chi\)-field if it is fixed exactly by the \(a\) with \(a\equiv\pm1\pmod{2^{e+2}}\), \(\psi(a)=1\), \((a/q)=(a/d)=1\). \(\mathrm{VVle}(E,t,B)\) means: for every prime \(\ell\) and \(0<e<t\) the Tate–Shafarevich group \(\text{Ш}_\ell(e)\) has finite length, and \[\sum_{\ell\in S}\log\ell\sum_{1\le e<t}\operatorname{len}\text{Ш}_\ell(e)\le B\qquad\text{for every finite set } S.\]

Theorem 35 (VWinAt psi 401, Exact). For every \(\varepsilon\in(0,1/52)\) and page choice \(e(x)\) with \(x^{2+8\varepsilon}\le 2^{e}\le 2x^{2+8\varepsilon}\), there are \(J\ge1\) and \(X_0\) such that for \(X\ge X_0\) and every window pair \((q,d)\) there is \(Y\) with \(X^{1/J}\le Y\), \(Y^2\le X\), and for every \(x\in[Y,Y^2]\) every \(\chi\)-field \(E\) satisfies \(\mathrm{VVle}\bigl(E,\lfloor x^{401}\rfloor,\sqrt{\exp(x^\varepsilon)}\bigr)\).

Reading \(E\) is the real \(2^{e+2}\)-cyclotomic layer cut by \(\psi\) and the two auxiliary quadratic characters, of degree about \(2^e\approx x^{2+8\varepsilon}\approx n\), matching the host. supplyChi turns the window into host existence for each character.

8.0.0.1 The proof chain.

Exact One chain, all used: vwin_of_row \(\to\) automorphic_window_of_inputs \(\to\) v_window_of_depth_zero_residual_bc_cm \(\to\) v_window_of_literal_automorphy_n \(\to\) v_window_of_automorphy_csd_ss_mono_odd \(\to\) vwin_of_hypotheses_H_E_alt \(\to\) of_four_alt \(\to\) vwin_of_counts \(\to\) vwin_of_joint \(\to\) vwinAt_of_joint.

Figure 13
Figure 13. What the \(\text{Ш}\) window rests on. All yellow and grey inputs are proved in Lean from one OpenImageCondition; none is assumed.
Input Rests on
bounded-height part (ThmBab) thmBab_of_lemmaK with ferreroWashington
minus part (ThmBcFree) a Stickelberger-type bound \(\operatorname{len}\text{Ш}^-\log\ell\le C\ell[E:\mathbb{Q}]^2(e+1)\log(e\ell N)^2\), via ML.iv (Poitou–Tate and unramified orthogonality)
plus part (ThmHV) counts I, II, III, V, Z and LemmaK: Kato-type bounds, \(\mathrm{Sp}_4\) inputs, unitary lifts
duality (LitCoh) Poitou–Tate, local Tate duality, local and global Euler characteristics, local and global finiteness, Greenberg–Wiles
CM inputs (LiteratureCMHypothesis) Grothendieck deformation, de Rham comparison, Serre–Tate, CM main theorem, Hilbert class field via \(j\), Siegel–Klingen and Deligne–Ribet, functional equation of partial zeta

Theorem 36 (Ferrero–Washington [4). , UpperTheory.V.FW.ferreroWashington, Exact] For a prime \(\ell\) and an abelian number field \(K\) containing \(\mu_\ell\) (\(\mu_4\) if \(\ell=2\)) there is \(r\) such that every layer \(K'\) of the cyclotomic \(\mathbb{Z}_\ell\)-tower over \(K\) satisfies \(|\operatorname{Cl}(K')/\operatorname{Cl}(K')^\ell|\le\ell^r\).

Reading Bounded \(\ell\)-rank up the tower is the statement \(\mu=0\) [16]. The proof occupies a 62-file folder (GaloisLibrary/FerreroWashingtonIndex).

8.0.0.2 Provenance.

The literature structures are filled by Lean proofs (..._from_literature, ..._of_open) from one OpenImageCondition, which comes from Row0b. That in turn is proved by the theorem galoisLocalGlobalStrong_recParameter. The clean axiom check means none is assumed.

8.0.0.3 Budget. Reading

The minus part is polynomial in \(t=x^{401}\), \([E:\mathbb{Q}]\approx x^{2+8\varepsilon}\) and logarithms; the plus part is at most \(\exp(x^{4\varepsilon/5})\) times a polynomial. Both are far below \(\sqrt b=\exp(x^\varepsilon/2)\), so the window does not touch the square root.

9 Why the \(2/3\) barrier does not appear

The outline (entry 14) says a calculation of the cost of each lift shows one would “stop at \(2/3\) a priori”. That calculation is not in the archive (Section 9.4). Here it is reconstructed from the code’s own quantities. Ours.

9.1 What the code charges

Verified \[\mathrm{charge}(j)=\sum_{\ell\text{ small}}\log\ell\sum_{\text{pieces at }j}\mathrm{colength} \approx f\,m\,b\,x^\varepsilon=B_0,\qquad E_b=\sum_{j=1}^m\mathrm{charge}(j)\approx m B_0.\] Each piece belongs to one lift (NData.locs_mem), and each lift is charged on its own block: the family is DefFam with \(L:J\to T\to\mathrm{BlockDat}\) (RowLat.lean:213), and the exponents and lengths of lift \(j\) are read from \(L(j,t)\) alone (LiftLat.lean:147--151).

9.2 Accounting 1: compounding gives exactly \(2/3\)

If lift \(j\) were charged for the whole composite \(\Theta_j\circ\dots\circ\Theta_1\), then \[\mathrm{charge}'(j)=\sum_{j'\le j}\mathrm{charge}(j')\approx jB_0,\qquad E_b'=\sum_{j=1}^m jB_0\approx\frac{m^2}{2}B_0,\qquad \frac{E_b'}{n_Tmb}\approx m^2x^\varepsilon,\] and the balance becomes \(x/m+m^2\), minimised at \(m=x^{1/3}\), giving \(x^{2/3}\). This is the outline’s \(2/3\) exactly. The two facts above remove the extra factor of \(m\).

Figure 14
Figure 14. Two accountings of the same chain. The code assigns every local piece to exactly one lift (NData.locs_mem) and reads each lift’s charge from its own block (DefFam.L); the compounding accounting adds a factor of \(m\) and gives the outline’s \(2/3\).

9.3 Accounting 2: no gcd does not give \(2/3\)

Without the common lattice, the averaging over words in Frame_at would be lost. Its only visible cost is \(\log\#W\le(k-3)\log2\approx x^{9/2+10\varepsilon}\), negligible against \(n_Tmb\). So in this bookkeeping the gcd does not save a factor of \(m\); its role is to make one divisibility statement hold for all words.

9.4 No \(2/3\) route in the archive

Exact About twenty unused variants of the final growth bound exist, among them:

Every one concludes \(G(x)\le x^{1/2+O(\varepsilon)}\). No \(m^2B_0\) budget, no per-word or per-character dual lattice, and no balance with exponent \(2/3\) or \(1/3\) appears anywhere. The only “\(2/3\)” constants are factors like \((2/3)^f\) in volume bounds.

9.5 A planned test

We designed a perturbation experiment (mechanism/PERTURB.md): loosen only the upper bound on \(m\) in RankVal by \(x^{1/40}\), rebuild the roughly 13,000 dependent modules, and record which proofs fail. Our prediction, written before running it: the failures appear in the window lemmas (szRow_rankSizes, winA) and in the \(m^5\) word-count lemma, not in the arithmetic layers, because \(m\) and \(k\) are locked with only about \(x^{0.26\varepsilon}\) of headroom. The results will be posted as an update to this document.

10 The verification record

Check Result (20–21 November 2026)
Source scan 0 sorry, admit, axiom declarations, native_decide, reduceBool; 5 build-time code hits (informational)
Full build 175,187 modules, 183,893 jobs, OK, 401 minutes on 94 cores
Statement identical to Mathlib’s RiemannHypothesis (Mathlib.NumberTheory.LSeries.RiemannZeta)
Axioms propext, Quot.sound, Classical.choice only; OVERALL: CLEAN
Olean comparison 76,968 byte-identical to Sutherland’s manifest, including Main.olean; 98,219 differ, every one embedding the absolute build path; 0 differ without a path. Recompiled at Sutherland’s build path, a random sample of 300 of the 98,219 is byte-identical to the manifest, 300 of 300
Export hash byte-identical match to the published export: 61,570,595,788 bytes, sha256 5ea74b91…244e (lean4export v4.33.0, commit 15f6055e, 89 minutes)
Kernel replay Main with imports trusted: done; full –fresh replay not run by us (others report full replays: Sutherland’s leanchecker, 175,187 of 175,187 modules; Simon DeDeo, likewise)
Figure 15
Figure 15. Our build against Sutherland’s manifest of compiled files.

Remark 37. The differing oleans store the source path (for example /home/.../rh/lean/src/.../OrbitMult.lean, once per file), which differs between machines. The kernel export holds only declarations, and our re-export matches the published 61.6 GB export byte for byte, confirming that the differences cannot touch the proof.

Verified The path is the only difference. Sutherland’s build passed sources to lean as /home/ubuntu/grh/src/lean/src/<Module>.lean (his comment in the community repository’s issue #6, where patunz first reported the path dependence). We replayed, for a random sample of 300 of the 98,219 differing modules (seed 20261121), the exact lean command that Lake recorded for each module in its .trace file, changing only the output location and the source prefix to his path. All 300 outputs are byte-identical to his manifest. Before the sample, the same replay at our own path reproduced our own .olean files exactly, so the replay is faithful to the build (scripts/11_path_sample.sh).

11 Open questions, and what we could settle

Each question below is followed by what we could settle from the source and the outline. Status: three answered within the code (2, 3, 4), one narrowed (5), one still needs the author (1); the call-level gap in step D4, the role of the constants and a static reading of SizesAt are settled below.

Question 1. Is the compounding counterfactual of Section 9 the calculation behind the outline’s \(2/3\)?

Still open. The outline (entry 14, dated 15 November 2026) reports a “precise computation of the cost of each lift” and the figure \(2/3\), but does not show the calculation. Our reconstruction is the simplest accounting in the code’s own quantities that gives exactly \(2/3\); only the author can say whether it is his.

Question 2. Could a family with a smaller indefinite part per lift lower the \(m^2\) cost of \(E_b\), or does the Bloch–Kato count force it?

Answered within this chain. Verified ShapeJumps k m t is required with the same \(m\) at every local piece, whatever lift the piece belongs to (admG2, EmbeddingTypes.lean:122), so every lift is charged rank \(m\). Reading checked by arithmetic: even a family whose rank tracked the shrinking indefinite part, \(m-j\) at lift \(j\), would pay \(\sum_{j=1}^{m}(m-j)=m(m-1)/2\), still of order \(m^2\). So no refinement of the per-lift rank along this chain changes the exponent; it saves at most a factor of \(2\). Lowering it would need a rank of \(o(m)\) per lift, that is a chain with a smaller indefinite part, hence fewer lifts, which costs on the \(x/m\) side.

Question 3. Why must the lift exponent \(e'\) be as large as \(b=\exp(x^\varepsilon)\)? A polynomial \(e'\) would shrink \(B_0\), though not its dependence on \(m\).

Answered within this design. Verified \(e'\) is linear in a Hodge–Tate height \(h_B\in[b,6b]\), and the heights are of size \(b\) because the weights on thick places lie in \([b,b+\eta+1]\). Exact The weight scale cannot be polynomial: the \(\text{Ш}\)-window losses are themselves exponential. The bounded-height condition is stated with \(\exp(x^{4\varepsilon/5})\) (BaAt) and the plus part with \((\mathrm{cond}\cdot t)^C+C\exp(x^{4\varepsilon/5})\) (HVAt), both to be absorbed by \(\sqrt b=\exp(x^\varepsilon/2)\). Reading So \(b\) is exponential to swallow those losses, \(e'\approx b\) follows, and since \(b\) cancels in the balance this sets the absolute size of \(E_b\) but not the exponent \(1/2\).

Question 4. Do the formal versions of Ferrero–Washington, Poitou–Tate, the Kato-type bounds and the CM inputs match the published theorems at full strength?

Answered: statements read for all four.

Since every one of these is proved in Lean and the axiom check is clean, Question 4 is about exposition: whether the names are honest descriptions. For the ones we read they are, with the caveat on the Kato-type statement above.

11.0.0.1 Comparison with Jaswal’s guide.

Manvir Jaswal’s reader’s guide [6] compares every classical input on the chain with its archive form (his Table 2) and reaches the same verdicts on Ferrero–Washington and Poitou–Tate. It adds two inputs we did not read, which it classes as weaker than the published theorems their names recall, both proved in the archive.

Verified by name against the export of Main: Row0b, galoisLocalGlobalStrong_recParameter, RecTExists, recTExists, GoodCusp, cuspEx and JunkEx are used; LocalLanglandsGL is not. So the proof does not depend on the local Langlands correspondence. Jaswal also classes the Kato-type input as “Different” in shape, since it is stated with its own hypotheses (Used, a newform attachment, WitAQ), which is compatible with the reading above. His Table 1 lists the statements on the chain for which he knows no reference; several of them (ch_balance, TheoremSSelected, LinkN, lemmaK) are the steps we trace in Sections 4–8.

Question 5. Is the windowed Landau theorem (Theorem 5) new as a stated result?

Narrowed. Deducing zero-free regions from one-sided or localised information on prime sums is a classical line going back to Landau [9] and developed by Pintz (see the recent survey [7]), and effective versions of Landau’s theorem already give quantitative lower bounds for averages over ranges \([1,Y]\). The windowed statement is therefore close to known methods in substance; whether this exact form appears in the literature is for specialists to say.

11.0.0.2 The call-level gap in step D4 is closed.

Exact extendedClosedForm is the field p15 of local_flatness_results, which feeds loc8All_of, first_theorem_threshold_functions and first_theorem_closed_form; the family bound from it is Adm8F, a hypothesis of admissible_family_no_kernel, whose output is what the gain statement (AuxBlock8.Conclusion1, via starAll_of) is proved from. So the \(\sqrt p\) lemma reaches the gain inequality through the first closed-form theorem. Not checked: which lines of the gain proof consume p15 specifically.

11.0.0.3 The constants do not matter for the theorem.

Exact ChebyshevInequalityAtWindowScale asks for some constants \(e_D,e_L\) and \(\varepsilon_1>0\) such that for every \(\varepsilon<\varepsilon_1\) there is a family with \(G_m(x)\le x^{1/2+(e_D+e_L+7)\varepsilon}\) on its windows. Since \(e_D,e_L\) are fixed while \(\varepsilon\to0\), the Landau step gives zero-freeness for every \(\operatorname{Re}s>1/2\) whatever their values. Explicit constants would only matter for an effective (numerical) version, which the development does not attempt.

11.0.0.4 SizesAt under other exponents, read statically.

Exact Write \(m=x^\mu\), \(n=x^\nu\), \(k=x^\kappa\). In the abstract size layer the exponents enter only through \[\tfrac12-\varepsilon\le\mu\le\tfrac12+\tfrac\varepsilon 2,\qquad \kappa\ge1+\mu,\qquad \kappa\le 2\nu+\tfrac12-\varepsilon,\qquad \kappa\le e_K\] (the bounds on \(\mu\) are in RestSize and ErrWindows0), together with the family’s word count \(\kappa\ge5\mu+\nu\). Every other window (CostWindows, GrowthInequalitySizes, the bounds on \(L\), \(L_a\), \(b'\)) compares polynomials in \(x\) with \(b=\exp(x^\varepsilon)\) and holds for any fixed exponents. So statically, \(\mu\) is pinned only by the two windows that encode the balance, exactly as predicted. What reading cannot rule out is a hidden dependence elsewhere in the 13,000 modules downstream of the size choices; the perturbation run (Section 9) will check that mechanically.

11.0.0.5 Still not traced.

The Lean proofs of the classical inputs of Section 8, as opposed to their statements; numerical values of the constants; the perturbation run.

Acknowledgements

The formal proof is Levent Alpöge’s. This account builds on public work by others on 20 and 21 November 2026, and overlaps with some of it:

Errors of interpretation are ours.

11.0.0.6 Use of AI tools.

Claude (Anthropic) was used to help read the Lean source, run searches over the declaration list, and draft parts of this text and its figures. Statements marked Exact or Verified were checked against the Lean source or the build outputs in the accompanying repository, and the references against their published sources. The author has read the whole text and takes full responsibility for its content.

References

[1] L. Alpöge, note accompanying the Lean development, 20 November 2026, https://alpo.ge/remark.pdf.

[2] S. Bloch and K. Kato, \(L\)-functions and Tamagawa numbers of motives, in: The Grothendieck Festschrift, Vol. I, Progr. Math. 86, Birkhäuser, 1990, 333–400.

[3] S. Chowla and A. Selberg, On Epstein’s zeta-function, J. reine angew. Math. 227 (1967), 86–110.

[4] B. Ferrero and L. C. Washington, The Iwasawa invariant \(\mu_p\) vanishes for abelian number fields, Ann. of Math. (2) 109 (1979), 377–395.

[5] H. Iwaniec and E. Kowalski, Analytic Number Theory, AMS Colloquium Publications 53, American Mathematical Society, 2004.

[6] M. Jaswal, The Riemann hypothesis for Dirichlet \(L\)-functions [after Levent Alpöge], preprint, 21 November 2026, https://doi.org/10.5281/zenodo.24410057.

[7] D. R. Johnston and T. Trudgian, A round of Pintz to celebrate oscillations in sums, arXiv:2511.09978.

[8] K. Kato, \(p\)-adic Hodge theory and values of zeta functions of modular forms, Astérisque 295 (2004), 117–290.

[9] E. Landau, Handbuch der Lehre von der Verteilung der Primzahlen, Teubner, Leipzig, 1909.

[10] L. de Moura and S. Ullrich, The Lean 4 theorem prover and programming language, in: Automated Deduction (CADE-28), LNCS 12699, Springer, 2021, 625–635.

[11] The mathlib Community, The Lean mathematical library, in: Proceedings of the 9th ACM SIGPLAN International Conference on Certified Programs and Proofs (CPP 2020), 367–381.

[12] B. Naskręcki, riemann-rev (status dashboard, proof guide and book), https://github.com/nasqret/riemann-rev, November 2026.

[13] J. Neukirch, A. Schmidt and K. Wingberg, Cohomology of Number Fields, 2nd ed., Grundlehren der mathematischen Wissenschaften 323, Springer, 2008.

[14] G. Shimura, Introduction to the Arithmetic Theory of Automorphic Functions, Princeton University Press, 1971.

[15] A. V. Sutherland, build and verification page for the Lean development, https://math.mit.edu/~drew/grh/, November 2026.

[16] L. C. Washington, Introduction to Cyclotomic Fields, 2nd ed., Graduate Texts in Mathematics 83, Springer, 1997.

12 Map from steps to Lean declarations

Step TRACE.md Key declarations
A. Windowed Landau 1, 25.1 zeroFree_of_rpow_window, lemmaL, landau_window, poleData, core_gm, rh_of_every
B. Isolation 23 ch_balance, ThickPartTranslateMean, GrowthInequalityExplicit, signal_term, QH.pairEx, zero_density_box, hostThickExists
C. Size bound 2, 5, 12 proposition_size_bound, SZ.T1–T4, RankVal, exists_ranks_rankVal_frame, m_pow_five_mul_f_mul_L_le
D. Pointwise bound 6, 7, 14 each_unit_bound_statement, Ga_bound5, ChainTied, linkCore, linkGLOBAL7_items
D. Energy 8, 9, 15 SizeData_2.EDef, NormDecompositionCond, factEtaMean, nuAbs_place_fine, ClaimHigh
D. DIV 10, 11, 16 Data.DIV, height_chain_atR, errB_of_units, divisor_inequality_assembly, error_absorption
D. Rank \(E_b\) 13, 18 rank_le_row, ShapeJumps, NData.locs_mem, linkN_of_gd
D. Common lattice 17 prescribedRowLat, exists_divStage_of_pk, FamP.rrDef
D. Degree identity 20 NprimeMuR, LinkG_at, LinkDIV_at, Frame_at, loss_bound_at
D. Local divisibility 21, 25.2 key_minor_dvd, key3Ord, link_of_local_closed, starAll_of, gain_bound_all_of_star
E. Orthogonality 22 LinkS_at, theoremS_link_concrete, matKF_of, chosenProviso_min
F. \(\text{Ш}\) window 24, 25.3 VWinAt, vwinAt_of_joint, of_four_alt, ferreroWashington, ML.iv
\(2/3\) barrier 15, 18, 19 NData.locs_mem, DefFam, Frame_at