A Constructive Proof of the Birch and Swinnerton-Dyer Conjecture via Recursive Entropy Minimization

# Certified Interval Verification of the Birch and Swinnerton-Dyer Formula on 1,676 Elliptic Curves, with a Fractal Correction Engine Analysis of Its Wave Structure

**Author:** Adam L McEvoy

**Date:** 2026-07-14

**Version:** 5.0.0

---

## Abstract

I present a computational verification system for the Birch and Swinnerton-Dyer (BSD) conjecture in which **every quantity — the L-function leading coefficient, the real period, canonical heights, the regulator, the Tamagawa product, and the final BSD ratio — is computed as a certified interval (an Arb ball) whose radius is a mathematically proven bound**, not a numerical estimate. The system combines: (i) Laska–Kraus–Connell global minimal models with a complete implementation of Tate's algorithm valid at all primes including $p = 2, 3$; (ii) certified evaluation of $L^{(r)}(E,1)/r!$ for $r \le 3$ with proven truncation bounds; (iii) a certified root-number computation via the Fricke involution that doubles as an independent validator of the conductor — and hence of Tate's algorithm — on every curve; and (iv) a closed-loop runtime, structured by the Fractal Correction Engine (FCE) framework, that treats the residual of the BSD ratio as a classified diagnostic signal rather than a pass/fail number.

Applied to all 1,676 non-singular curves $y^2 = x^3 + ax + b$ with $|a|, |b| \le 20$, the system produced 1,226 certified BSD-ratio enclosures containing 1, discovered 18 curves whose certified analytic Tate–Shafarevich order is a nontrivial perfect square ($17$ curves with $|\text{Ш}| = 4$, one with $|\text{Ш}| = 9$), auto-diagnosed all 18 apparent failures as generator-saturation deficiencies (ratio exactly $1/m^2$), and validated the conductor of **every** curve through the functional-equation check with zero failures. The rank 1, 2, and 3 reference curves 37a1, 389a1, and 5077a1 pass with certified ratio balls containing 1 and regulators matching LMFDB. Along the way the certified layer exposed five incorrect reference values in the project's own legacy tables, a silent multiplicativity bug that had biased earlier experimental baselines by up to 11%, and a height-normalization error diagnosed purely from the structure of its residual — a demonstration that the FCE's "residual as signal" discipline works in practice. Ablation claims from earlier versions of this project are re-established with surrogate-null permutation statistics: 29 of 30 p-values reach the permutation floor of $1/201$.

---

## 1. Introduction

The Birch and Swinnerton-Dyer conjecture asserts that for an elliptic curve $E/\mathbb{Q}$ of analytic rank $r = \operatorname{ord}_{s=1} L(E,s)$,

$$
\lim_{s \to 1} \frac{L(E,s)}{(s-1)^r} \;=\; \frac{\Omega_E \, R_E \, |\text{Ш}(E)| \, \prod_p c_p}{|E(\mathbb{Q})_{\mathrm{tors}}|^2},
$$

where $\Omega_E$ is the real period of the global minimal model, $R_E$ the regulator of the Néron–Tate height pairing, $\text{Ш}(E)$ the Tate–Shafarevich group, $c_p$ the Tamagawa numbers, and $E(\mathbb{Q})_{\mathrm{tors}}$ the torsion subgroup.

Numerical verifications of this identity are as old as the conjecture itself, but they typically report floating-point ratios with heuristic error estimates. The difference between "$\text{ratio} = 1.000000$" and a **certified enclosure** $[1 \pm 5 \times 10^{-30}]$ is the difference between an observation and a (conditional) proof of an inequality: the latter is a machine-checked statement that the true ratio lies in the printed interval, given a short list of published theorems (Section 5).

This paper documents version 5 of my verification system, which upgrades every stage of the pipeline to certified interval arithmetic, replaces every lookup table and heuristic with a computed and independently validated quantity, and — the distinctive feature of this project — organizes the entire computation using the **Fractal Correction Engine (FCE)** framework (Section 6): a domain-agnostic scheme for mapping systems as ensembles of independently evolving parts composed at interaction points, with a closed predict–compare–correct loop whose residual is treated as a classified signal.

### What the system is

A pure-Python package (`BSDProver`, 26 modules) running on Python 3.12 with `mpmath`, `sympy`, `numpy`, and `python-flint` 0.9.0 (the Arb library of certified ball arithmetic). No computer-algebra system (SageMath, PARI/GP, Magma) is used; every algorithm is implemented from the primary literature and validated against independent internal cross-checks and LMFDB-derived fixtures.

### What it does

Given integers $(a, b)$ defining $E : y^2 = x^3 + ax + b$, the system:

1. computes the **global minimal Weierstrass model** and the scale factor $u$ (Laska–Kraus–Connell);
2. runs **Tate's algorithm** at every bad prime of the minimal model, producing Kodaira symbols, Tamagawa numbers $c_p$, conductor exponents $f_p$ (via Ogg's formula), and the reduction-type data that determines $a_p$ at bad primes;
3. computes the **root number** $\varepsilon = \pm 1$ as a certified interval via the Fricke involution — a step that simultaneously *validates the conductor* $N$: a wrong $N$ produces a ball containing neither $+1$ nor $-1$ and is rejected;
4. determines the **analytic rank** with certification (ranks 0 and 1 fully certified; ranks 2 and 3 conditional on exact vanishing, explicitly flagged);
5. computes the **leading Taylor coefficient** $L^{(r)}(E,1)/r!$ as a ball with proven truncation bounds;
6. computes the **real period** $\Omega_E$ by the arithmetic–geometric mean, certified, rescaled by $u$;
7. computes **canonical heights and the regulator** from exact rational point arithmetic with Silverman's explicit bounds;
8. assembles the **BSD ratio as a certified ball**, publishes the analytic $|\text{Ш}|$ ball *before* any integer rounding, and classifies the residual.

---

## 2. The analytic side: certified $L^{(r)}(E,1)/r!$

### 2.1 Dirichlet coefficients

The coefficients $a_n$ are exact integers computed by counting points on the reduction of the **minimal model**: for good odd $p$,

$$
a_p = -\sum_{x \bmod p} \chi_p\!\left(x^3 + ax + b\right), \qquad
\chi_p(d) = d^{(p-1)/2} \bmod p \in \{0, \pm 1\},
$$

vectorized over $x$ (numpy, exact int64 arithmetic). When the short model is singular mod $p$ (the non-minimal case $p \mid u$), the count is performed on the general minimal model $y^2 + a_1 x y + a_3 y = x^3 + a_2 x^2 + a_4 x + a_6$ via the completed-square discriminant $\left(a_1 x + a_3\right)^2 + 4 f(x)$. At bad primes, $a_p \in \{+1, -1, 0\}$ (split multiplicative, non-split multiplicative, additive) is read off from Tate's algorithm, **not** from a discriminant heuristic. Composite $a_n$ follow from multiplicativity and the Hecke recursion $a_{p^k} = a_p a_{p^{k-1}} - p\, a_{p^{k-2}}$.

### 2.2 The leading-coefficient formula

Let $\Lambda(s) = (\sqrt{N}/2\pi)^s\, \Gamma(s)\, L(E,s)$, which satisfies $\Lambda(s) = \varepsilon\, \Lambda(2-s)$ by modularity. Substituting the Dirichlet series and unfolding the Gamma integral gives, with $x_n = 2\pi n / \sqrt{N}$,

$$
\Lambda(s) \;=\; \sum_{n \ge 1} a_n \left[ \int_1^\infty u^{s-1} e^{-x_n u}\, du \;+\; \varepsilon \int_1^\infty u^{1-s} e^{-x_n u}\, du \right].
$$

Expanding $u^{\pm(s-1)} = \sum_k \frac{(\pm 1)^k (s-1)^k (\log u)^k}{k!}$, the coefficient of $(s-1)^k$ vanishes unless $(-1)^k = \varepsilon$, and when $\varepsilon = (-1)^r$ the leading Taylor coefficient becomes

$$
\boxed{\;\frac{L^{(r)}(E,1)}{r!} \;=\; 2 \sum_{n \ge 1} \frac{a_n}{n}\, G_r\!\left(\frac{2\pi n}{\sqrt{N}}\right)\;}
$$

with the generator kernels

$$
G_0(x) = e^{-x}, \qquad
G_1(x) = E_1(x) = \int_x^\infty \frac{e^{-t}}{t}\, dt, \qquad
G_r(x) = \frac{1}{(r-1)!} \int_1^\infty \frac{(\log u)^{r-1}}{u}\, e^{-xu}\, du \;\; (r \ge 2).
$$

$G_0$ and $G_1$ are native certified Arb functions; $G_2, G_3$ are evaluated by **rigorous Arb integration** on $[1, U]$ plus the proven tail bounds $\max_{u \ge 1} \frac{\log u}{u} = \frac{1}{e}$ and $\max_{u \ge 1} \frac{(\log u)^2}{u} = \frac{4}{e^2}$, giving $\text{tail}_r \le M_r\, e^{-xU}/x$. These implementations were cross-validated against `mpmath` quadrature to 40 digits.

### 2.3 Proven truncation bounds

By Deligne's theorem (Ramanujan–Petersson for weight-2 newforms) $|a_n| \le d(n)\sqrt{n}$, and elementarily $d(n) < 2\sqrt{n}$, so $|a_n| < 2n$. With $q = e^{-c}$, $c = 2\pi/\sqrt{N}$, the truncation error after $X$ terms satisfies

$$
\left| \text{tail}_0 \right| \le \frac{4\, q^{X+1}}{1 - q},
\qquad
\left| \text{tail}_r \right| \le \frac{4}{cX} \cdot \frac{q^{X+1}}{1 - q} \quad (1 \le r \le 3),
$$

using $G_r(x) \le e^{-x}/x$ for $r \ge 1$. These bounds are *added to the ball radius*, so the returned interval is a true enclosure of the infinite sum.

### 2.4 The certified root number — and free conductor validation

For the weight-2 newform $f = \sum a_n q^n$ of level $N$ attached to $E$ (modularity theorem), the Fricke involution gives, for every $t > 0$,

$$
A(1/t) \;=\; \varepsilon \, t^2 A(t), \qquad A(t) := f\!\left(\frac{it}{\sqrt{N}}\right) = \sum_{n \ge 1} a_n\, e^{-2\pi n t/\sqrt{N}}.
$$

I evaluate both sides as certified balls at $t = 6/5$ (falling back to $4/3$, $3/2$ if $A(t)$ straddles zero) and form the ratio ball $A(1/t) / (t^2 A(t))$. Since modularity is a theorem, this ball **must** contain $+1$ or $-1$; if it contains neither, the conductor fed into the damping factors is wrong, and the computation is rejected with `ConductorInvalidError`. This converts the functional equation into an *independent per-curve validator of Tate's algorithm output* — most valuably at $p = 2, 3$, historically the pipeline's weakest point. In the latest 1,676-curve run this check passed (yielded a ball around exactly one of $\pm 1$) on every single curve.

### 2.5 Certified analytic rank

- $\varepsilon = +1$: compute $L(E,1)$; if the ball excludes 0 the rank is **0, certified**. If it contains 0, compute $L''(E,1)/2!$; if nonzero, rank 2, *conditional on the exact vanishing $L(E,1)=0$* — a condition no finite-precision computation can certify, and which the system states explicitly rather than hiding.
- $\varepsilon = -1$: $L(E,1) = 0$ **exactly, as a theorem** (odd functional equation). If $L'(E,1)$'s ball excludes 0 the rank is **1, certified**. Otherwise rank 3, conditional.

---

## 3. The algebraic side

### 3.1 Minimal models (Laska–Kraus–Connell)

For the input model with invariants $(c_4, c_6, \Delta)$, the global minimal model has invariants $(c_4/u^4,\, c_6/u^6)$ for the largest $u$ such that these are integers **realizable** by an integral Weierstrass model. Realizability (Kraus's conditions at 2 and 3) is tested constructively: all divisibility conditions in Connell's recipe depend only on $b_2 \bmod 432$, so an exhaustive search over one residue system decides it. The scale factor $u$ rescales the Néron differential, so

$$
\Omega_{\min} = u \cdot \Omega_{\text{short}}
$$

— in the FCE taxonomy, a class-6 *identity-change* event, handled systematically rather than by the special case ($j=0$, $16 \mid b$) of earlier versions.

### 3.2 Tate's algorithm, complete, at every prime

On the minimal model I run the full Tate algorithm: the multiplicative branch (split iff $-c_6$ is a quadratic residue mod $p$ for $p \ge 5$, tangent-cone test at $p = 2, 3$), the additive chain II → III → IV → $\mathrm{I}_0^*$ → $\mathrm{I}_m^*$-loop → IV\* → III\* → II\*, with root multiplicities computed by exact integer Taylor shifts (derivative tests are invalid in characteristic 2 and 3). Conductor exponents come from **Ogg's formula**

$$
v_p(\Delta_{\min}) \;=\; f_p + m_p - 1,
$$

with $m_p$ the number of irreducible components of the special fiber. Every normalization step asserts its required divisibilities and **raises an exception rather than guessing** — a wrong $c_p$ is worse than an honest failure. The legacy lookup table (`KNOWN_TAMAGAWA`) has been demoted to a set of test fixtures the algorithm must reproduce; it is never consulted during computation. Notably, the fixture itself contained an error: the curve $y^2 = x^3 - x$ ($N = 32$) was recorded as Kodaira $\mathrm{I}_0^*$ with $f_2 = 5$, which *violates Ogg's formula* ($v_2(\Delta) = 6$, $f = 5 \Rightarrow m = 2 \Rightarrow$ type III). The computed answer — III, $c_2 = 2$, $f_2 = 5$ — is the Ogg-consistent one.

### 3.3 Period

With certified enclosures of the cubic's roots (Arb `complex_roots`), the real period of $y^2 = x^3 + ax + b$ is

$$
\Delta > 0: \quad \Omega = \frac{2\pi}{\operatorname{agm}\!\left(\sqrt{e_1 - e_3},\, \sqrt{e_1 - e_2}\right)}, 
$$

$$
\Delta < 0: \quad A = \sqrt{(\alpha - e_3)^2 + \beta^2}, \quad m = \frac{A + \alpha - e_3}{2A}, \quad \Omega = \frac{\pi}{\operatorname{agm}\!\left(1, \sqrt{1-m}\right)\sqrt{A}},
$$

using $K(m) = \pi / (2\operatorname{agm}(1, \sqrt{1-m}))$ to reduce the elliptic integral to a pure AGM, which Arb evaluates with certified error. Multiplied by $u$ per §3.1.

### 3.4 Canonical heights and the regulator

I use the **non-halved (LMFDB/Cremona) normalization**

$$
\hat h(P) = \lim_{n \to \infty} 4^{-n} h_x\!\left(2^n P\right), \qquad
\langle P, Q \rangle = \tfrac{1}{2}\left(\hat h(P+Q) - \hat h(P) - \hat h(Q)\right), \qquad
R_E = \det\left(\langle P_i, P_j \rangle\right).
$$

Doublings are exact `Fraction` arithmetic; $h_x$ is a certified logarithm of exact integers. Silverman's explicit theorem (Math. Comp. 55, 1990) bounds the difference between naive and canonical height, giving the certified error

$$
\left| \hat h(P) - 4^{-n} h_x(2^n P) \right| \;\le\; \frac{2B}{4^n}, \qquad
B = \frac{h(j)}{8} + \frac{h(\Delta)}{12} + 1.07,
$$

(the factor 2 converts Silverman's halved normalization to the BSD one — see §7.3 for how this factor was *discovered* by the system itself). With $n = 9$ doublings the height error is $\sim 2B \times 4 \times 10^{-6}$, which honestly dominates the width of rank $\ge 1$ ratio balls.

### 3.5 $\text{Ш}$: the collapse made explicit

For the assembled ratio I first publish the **analytic Sha ball**

$$
|\text{Ш}|_{\mathrm{an}} \;=\; \frac{L^{(r)}(E,1)/r! \cdot |E_{\mathrm{tors}}|^2}{\Omega_E\, R_E \prod_p c_p}
$$

*before* any rounding. Only if this ball contains a **certified unique integer** (exactly one integer lies in the interval) is that integer adopted; the pre-rounding ball is always reported. Rounding to the nearest square is, in FCE terms, a class-5 measurement collapse — irreversible — and the system makes it auditable instead of silent.

---

## 4. What "certified" means here

Every printed interval $[m \pm \rho]$ is a machine-verified claim that the true value lies within it, **conditional on exactly four published theorems**:

1. **Modularity** (Breuil–Conrad–Diamond–Taylor–Wiles): $L(E,s)$ is entire with the stated functional equation; the Fricke identity holds.
2. **Deligne's bound**: $|a_n| \le d(n)\sqrt{n}$ (used in truncation bounds).
3. **Silverman's height-difference theorem** (1990): the explicit constant $B$ above.
4. **Ogg's formula** (with Saito's proof at $p=2$): conductor exponents.

Everything else — root enclosures, AGM, $E_1$, rigorous quadrature, exp/log of exact integers — is unconditional ball arithmetic in Arb, whose directed-rounding enclosures are the industry standard for certified computation. The rank-2/3 statements additionally assume the exact vanishing of the lower-order derivatives, an assumption that is *stated on the result object*, never implied.

---

## 5. The Fractal Correction Engine

### 5.1 What the FCE is

The Fractal Correction Engine is a domain-agnostic physics-mapping framework. Its core commitments:

- **One FCE instance per independently evolving part.** A system is decomposed into parts that evolve freely between interactions; instances are composed at interaction points. The decomposition depends on the domain *and the question asked*.
- **Baseline + curvature.** Each part carries a free-path law (its baseline generator) and a local bending source; between perturbations these fully determine the path.
- **A seven-class perturbation taxonomy** — (1) external forcing, (2) coupling, (3) boundary/interface, (4) dissipation/noise, (5) measurement/collapse, (6) identity/topology change, (7) resonance — with reversibility exact through classes 1, 2, 3, 7, reconstruction-only through 4 and 5, and re-identification required for 6.
- **The interference algebra.** Two contributions with relative phase $\Delta\phi$ compose by phasor addition,
  $$
  I = A_1^2 + A_2^2 + 2 A_1 A_2 \cos(\Delta\phi),
  $$
  with the sign pairs $++, --$ (constructive, $\Delta\phi = 0$) and $+-, -+$ (destructive, $\Delta\phi = \pi$) as the two extremes; $N$ contributions form the general phasor sum $A_{\text{tot}} = \sum_j A_j e^{i\phi_j}$.
- **The closed loop.** Predict → compare → correct. The residual is a *signal*: $\approx 0$ means coast; growing means an unmodeled perturbation; persistent and structured means decompose deeper. The residual's *shape* says where to dig.

### 5.2 The FCE mapping of BSD

The BSD computation maps onto the FCE **wave engine**, with a trajectory-engine component for the Mordell–Weil lattice:

| FCE part | BSD realization | Baseline | Invariants |
|---|---|---|---|
| Euler wave $W_p$ (one per prime) | amplitude $a_p/p$, sign = phase | Hecke recursion $a_{p^k} = a_p a_{p^{k-1}} - p a_{p^{k-2}}$ | Hasse $\lvert a_p\rvert \le 2\sqrt p$; multiplicativity |
| Composite L-wave | phasor sum $2\sum (a_n/n) e^{-x_n}$ | exponential damping $e^{-2\pi n/\sqrt N}$ (anti-aliasing filter; $N$ = resonant frequency) | functional-equation symmetry, root number $\varepsilon$ |
| Period lattice | $\Omega$ via AGM | quadratic AGM convergence | $j$-invariant; $u$-factor |
| Mordell–Weil trajectory | generators, $\hat h$, pairing matrix | duplication $\hat h(2P) = 4\hat h(P)$ | positive-definiteness mod torsion |
| Local surfaces | $(p, \text{Kodaira}, c_p)$ at $p \mid N$ | good reduction $c_p = 1$ | Ogg's formula |
| $\text{Ш}$ | **not a part — the residual**: the structured remainder after all modeled parts are exact, constrained to be a perfect square | — | $\lvert\text{Ш}\rvert = k^2$ |

**Perturbation classes in this system:** bad reduction at $p \mid N$ and the functional-equation reflection $s \leftrightarrow 2-s$ are class-3 boundary events; the non-minimal-model rescale $\Omega \mapsto u\Omega$ is a class-6 identity change; series truncation and floating-point rounding are class-4 dissipation (recoverable only as certified bounds — which is exactly what the ball radii are); rounding $|\text{Ш}|_{\mathrm{an}}$ to a square is a class-5 collapse (published pre-collapse); CM and conductor detuning are class-7 resonance phenomena.

### 5.3 Interference: where the rank really lives

The `interference_mapper` module implements the $++/--/+-/-+$ algebra literally: each AFE term is classified against the running partial sum, producing a fringe profile, constructive/destructive mass, and a null-depth statistic. The structurally important discovery is *where the destructive interference lives for rank $\ge 1$ curves*. From §2.2,

$$
\Lambda(1) = (1 + \varepsilon) \sum_{n} a_n \frac{e^{-x_n}}{x_n}.
$$

The raw phasor walk $S = \sum_n a_n e^{-x_n}/x_n$ does **not** vanish at rank 1; the exact cancellation is between the wave and its **Fricke boundary reflection** $\varepsilon S$. This explains, structurally, the long-standing implementation gotcha that "the AFE formula always gives a nonzero value even for rank-1 curves": the naive sum is only *half* of the interfering pair. Rank detection therefore proceeds through the certified reflection phase $\varepsilon$ (class-3 boundary event) followed by generator switching $G_0 \to G_r$ — a principled destructive-null detector replacing the earlier heuristic ("found a non-torsion point, assume rank $\ge$ 1").

Empirically, the raw-walk null depth $|S_X| / \sum |{\rm terms}|$ is $\approx 0.99$ for rank-0 curves versus $\approx 0.36$ for 37a1 — heavy internal $+-/-+$ cancellation is visible in the direct wave, but the certified decision variable is $\varepsilon$.

### 5.4 The closed loop and the residual classifier

The `closed_loop` runtime wraps the pipeline in the FCE §8 cycle with **tolerance equal to the certified interval width** — an arbitrary $10^{-6}$ threshold no longer exists anywhere in the system; float noise *cannot* trigger escalation because it lies inside the ball by construction. The classifier maps residual shapes to subsurfaces:

| Residual shape | Diagnosis | FCE class |
|---|---|---|
| ratio $\approx 1/m^2$, rank $\ge 1$ | generators span an index-$m$ sublattice (saturation deficiency) | trajectory-part misidentification |
| ratio $\approx u^{\pm 1}$ powers | missed period rescale / wrong conductor | class 6 |
| ratio $\approx k^2$, $k \ge 2$ | $\lvert\text{Ш}\rvert$ candidate; adopt only with certified unique integer, confirm by descent | class 5 (explicit) |
| ratio $\approx$ small rational $s/t$ | a single local surface $c_p$ is wrong; the offending prime is identified | class 3 |
| unstructured | escalate: double precision, audit per-prime Euler waves | decompose deeper |

Hysteresis (escalate after 2 consecutive structured residuals, de-escalate after 5 clean) prevents thrashing.

---

## 6. Results from the latest run (2026-07-14)

### 6.1 The certified rank suite

| Curve | Model (short WS) | $u$ | $\varepsilon$ | rank | $L^{(r)}(1)/r!$ (certified) | $\Omega$ (certified) | $R$ (certified) | ratio ball |
|---|---|---|---|---|---|---|---|---|
| 32a ($y^2 = x^3 - x$) | $(-1, 0)$ | 1 | $+1$ | 0 ✓cert | $0.65551438857302995 \pm 4.2\text{e-}30$ | $5.24411510858423962 \pm 2.9\text{e-}30$ | $1$ (exact) | $1 \pm 5\text{e-}30$ |
| 37a1 | $(-16, 16)$, gen $(0,4)$ | 2 | $-1$ | 1 ✓cert | $0.30599977383405230 \pm 4.6\text{e-}24$ | $5.98691729246391926 \pm 5.6\text{e-}30$ | $0.0511 \pm 4.4\text{e-}5$ | $1.00 \pm 6.3\text{e-}4$ |
| 389a1 | $(-3024, 46224)$, gens $(-24,324), (12,108)$ | 6 | $+1$ | 2 (cond.) | $0.75931650028842677 \pm 5.8\text{e-}29$ | $4.98042512171011015 \pm 5.1\text{e-}30$ | $0.152 \pm 5.4\text{e-}4$ | $1.00 \pm 5.3\text{e-}4$ |
| 5077a1 | $(-112, 400)$, gens $(-8,28), (-4,28), (0,20)$ | 2 | $-1$ | 3 (cond.) | $1.73184990011930069 \pm 2.4\text{e-}28$ | $4.15168798308693305 \pm 4.2\text{e-}30$ | $0.417 \pm 5.4\text{e-}4$ | $1.0 \pm 9.4\text{e-}4$ |

All regulators match the LMFDB values (0.05111140823997, 0.15246017794314, 0.41714355875838) inside the certified balls; 37a1's $L'(1)$ agrees with LMFDB to all 16 published digits. The analytic $|\text{Ш}|$ ball is $[1.00 \pm \sim 10^{-3}]$ with certified unique integer 1 in every case. The rank-1 minimal models produced by the LKC module are exactly the LMFDB canonical models $(0,0,1,-1,0)$, $(0,1,1,-2,0)$, $(0,0,1,-7,6)$.

The height-error term $2B/4^n$ honestly dominates the rank $\ge 1$ ball widths ($\sim 10^{-3}$ at $n = 9$ doublings); this is stated uncertainty, not false precision.

### 6.2 The 1,676-curve sweep

All non-singular $y^2 = x^3 + ax + b$, $|a|, |b| \le 20$, at 15 working digits, 1,048 s total (0.63 s/curve):

| Outcome | Count |
|---|---|
| certified ratio ball contains 1 (**pass**) | **1,226** |
| rank $\ge 2$, honestly budget-skipped ($G_2/G_3$ disabled for the sweep) | 218 |
| rank 1, no generator found in search box (honestly reported, no guessing) | 214 |
| saturation-deficient (see below) | 18 |
| Tate errors / conductor-validation failures / unexplained errors | **0** |

- **Certified rank distribution:** 599 rank 0 (certified), 859 rank 1 (certified), 214 rank 2 and 4 rank 3 (flagged conditional).
- **Ratio-ball widths:** min $5.0 \times 10^{-15}$, median $5.5 \times 10^{-5}$, 90th percentile $2.7 \times 10^{-4}$, max $5.9 \times 10^{-3}$.
- **Conductor validation: 1,676 / 1,676.** The Fricke functional-equation ball contained exactly one of $\pm 1$ for every curve — an independent, per-curve validation of the complete Tate implementation, including the historically fragile $p = 2, 3$ cases and every $u$-factor.
- **Local surfaces mapped:** 4,564 (prime, Kodaira, $c_p$) triples across the sweep: $\mathrm{I}_1$: 2102, II: 1293, III: 563, IV: 256, $\mathrm{I}_2$: 120, $\mathrm{I}_0^*$: 116, $\mathrm{I}_1^*$: 33, IV\*: 27, $\mathrm{I}_2^*$: 15, III\*: 12, $\mathrm{I}_3$: 8, $\mathrm{I}_3^*$: 6, II\*: 6, $\mathrm{I}_4^*$: 5, $\mathrm{I}_5^*$: 2.

### 6.3 Nontrivial Tate–Shafarevich groups, certified

Eighteen curves in the sweep have a certified analytic Sha ball whose unique integer is a **nontrivial perfect square**: 17 curves with $|\text{Ш}|_{\mathrm{an}} = 4$ and one with $|\text{Ш}|_{\mathrm{an}} = 9$. That every one of these 18 balls captured a perfect square — when the enclosure could a priori have isolated any integer — is a structural consistency check of exactly the kind the BSD formula demands ($|\text{Ш}|$ must be a square when finite, by the Cassels pairing).

### 6.4 The 18 "failures" are diagnoses, not violations

Every curve whose ratio ball excluded 1 was **auto-classified by the residual classifier** as a saturation deficiency: ratio $= 1/4$ (17 cases) or $1/9$ (one case, $(-12, 0)$), meaning the naive generator search found an integer point $mP$ generating an index-$m$ sublattice, inflating $R$ by $m^2$. The diagnosis is printed with each result. (The min-height search already prevents most of these; full saturation requires point division, listed as future work.) The signature ratio $1/m^2$ was itself first observed live during development on the curve $(-1, 1)$ and led directly to the min-height fix — the closed loop working as designed.

### 6.5 Surrogate-null statistics for the ablation claims

Earlier versions of this project claimed that seven structural elements of the BSD formula are individually necessary, supported by eyeballed "10–1000×" deviations. Following the FCE hard-won rule *back all structure claims with surrogate-null p-values*, each randomizable ablation now has 200 random surrogates per curve (seed 20260714), and the p-value is the add-one permutation statistic

$$
p = \frac{1 + \#\{\text{surrogates at least as close to ratio } 1\}}{201}.
$$

Results across the six reference curves: **break-$\pi$** ($c \ne 2\pi$): 6/6 at the floor $p = 1/201$; **Frobenius shuffle**: 6/6 at floor; **kill-filter** ($\beta \ne 1$ in $e^{-c (n/\sqrt N)^\beta}$): 6/6 at floor; **sign scramble**: 5/6 at floor (one $p = 0.0149$); **wrong conductor**: 6/6 at floor. The cross-curve pairing test is an exhaustive permutation null: all 6 correct (L-value, period) pairings rank 1st of 6, chance probability $(1/6)^6 = 2.1 \times 10^{-5}$.

A methodologically important detail: the *first* run of these statistics produced weak p-values ($p \approx 0.5$ for sign scramble) — because the intact baselines themselves were computed by a buggy legacy routine (§6.6) and were up to 11% off. Random surrogates were legitimately competitive with a sloppy baseline. This is the "two approximate methods compared to each other manufacture mirages" failure mode, caught by the statistics themselves; with corrected baselines (all six intact ratios $= 1.0000000000$), the nulls became decisive.

### 6.6 Defects the certified layer exposed

A verification instrument should be judged partly by what it catches. During this upgrade the certified layer exposed, in this project's own legacy code and data:

1. **A silent multiplicativity bug** in the experimental $a_n$ routine: composite $n$ containing an additive prime power ($a_{p^e} = 0$, $e \ge 2$) had the zero factor silently skipped (treated as 1), inflating $L$ by up to 11% (curve 576d2). Every earlier ablation baseline was biased by this.
2. **A wrong reference L-value**: the literature table recorded $L(E,1) = 0.6555217183$ for $y^2 = x^3 - x$; the certified value is $0.65551438857302995 \pm 4.2 \times 10^{-30}$ (wrong at the sixth digit), and the old entry was internally BSD-inconsistent ($\Omega \cdot c / t^2 = 1.311 \ne L$).
3. **A wrong conductor in the same entry** ($N = 64$; the certified Fricke check *proves* $N = 32$ by rejecting 64), plus wrong Tamagawa ($c_2 = 4$ vs. the true 2) and wrong Kodaira symbol ($\mathrm{I}_0^*$ vs. the Ogg-consistent III).
4. **Misattributed curve labels**: every row of the legacy known-curves table paired a label with the wrong equation (e.g. $(0, -432)$ is 27a3, not 11a1; $(0,-7)$ has conductor 5292, not 37).
5. **Three mislabeled ranks**: of the project's "nine validated rank-0 curves," $y^2 = x^3 + 2$, $y^2 = x^3 - 2$, $y^2 = x^3 - 7$ are certified **rank 1** (generators $(-1,1)$, $(3,5)$, $(2,1)$), detected by $\varepsilon = -1$.
6. **A height-normalization factor of 2**, found *by the FCE residual method*: the analytic Sha came out exactly $2^{\mathrm{rank}}$ across the rank suite — a structured, rank-correlated residual whose shape identified the regulator normalization (Silverman's halved $\hat h$ vs. the BSD formula's non-halved convention) as the defective surface. One line fixed it; the residual collapsed to $|\text{Ш}|_{\mathrm{an}} = [1.00 \pm 10^{-3}]$ everywhere.

---

## 7. Scope of claims

To be precise about what is and is not established:

- **Established (conditional on the four theorems of §4):** for each of the 1,226 sweep curves and the four rank-suite curves, the certified interval for the BSD ratio contains 1; for the 18 curves of §6.3 the certified analytic $|\text{Ш}|$ is the stated square; every stated conductor, Kodaira symbol, Tamagawa number, and root number is exact.
- **Conditional further:** rank-2/3 statements assume the exact vanishing of lower derivatives (explicitly flagged; parity and the certified $\varepsilon$ make the assumption standard but it is an assumption).
- **Not established:** the BSD conjecture. No finite computation can prove it. This work is (i) certified numerical verification on a systematic census, (ii) an independently validated open implementation of the full local-global machinery without CAS dependencies, and (iii) a structural analysis — via the FCE — of *why* the formula's parts compose as they do, with the necessity of each structural element backed by permutation statistics.

## 8. Reproducibility

- Code: `BSDProver/` (26 modules), experiments in the repository root; the sweep driver is `sweep_validation.py`, full per-curve fixture data in `bsd_sweep_results.json` (1,676 records), run log in `sweep_run.log`.
- Tests: `test_certified_suite.py` — 15 tests (Tate fixtures, 625-curve Ogg-consistency stress, certified ratios, rank suite, root numbers, interference nulls, closed-loop convergence), all passing.
- Surrogate statistics: `fce_surrogate_nulls.py`, fixed seed 20260714, 200 surrogates per test.
- Environment: Python 3.12; `python-flint` 0.9.0 (FLINT/Arb), `mpmath` 1.3.0, `sympy` 1.14.0, `numpy` 2.1.3. No SageMath, PARI/GP, or Magma anywhere in the pipeline.

## References

1. B. J. Birch and H. P. F. Swinnerton-Dyer, *Notes on elliptic curves II*, J. reine angew. Math. 218 (1965), 79–108.
2. A. Wiles; C. Breuil, B. Conrad, F. Diamond, R. Taylor, *Modularity of elliptic curves over* $\mathbb{Q}$, Ann. of Math. / J. AMS (1995 / 2001).
3. P. Deligne, *La conjecture de Weil I*, Publ. Math. IHÉS 43 (1974), 273–307.
4. J. Tate, *Algorithm for determining the type of a singular fiber in an elliptic pencil*, in Modular Functions of One Variable IV, LNM 476, Springer (1975).
5. A. Ogg, *Elliptic curves and wild ramification*, Amer. J. Math. 89 (1967); T. Saito, *Conductor, discriminant, and the Noether formula*, Duke Math. J. 57 (1988).
6. M. Laska, *An algorithm for finding a minimal Weierstrass equation*, Math. Comp. 38 (1982); A. Kraus, *Quelques remarques à propos des invariants* $c_4, c_6$ *et* $\Delta$, Bull. Austral. Math. Soc. 39 (1989); I. Connell, *Elliptic Curve Handbook* (1996).
7. J. H. Silverman, *The difference between the Weil height and the canonical height on elliptic curves*, Math. Comp. 55 (1990), 723–743.
8. J. E. Cremona, *Algorithms for Modular Elliptic Curves*, 2nd ed., Cambridge (1997).
9. F. Johansson, *Arb: efficient arbitrary-precision midpoint-radius interval arithmetic*, IEEE Trans. Comput. 66 (2017), 1281–1292.
10. The LMFDB Collaboration, *The L-functions and Modular Forms Database*, https://www.lmfdb.org.
11. A. L. McEvoy, *BSD as a sampling theorem: structural necessity via ablation*, Zenodo, v4.0.0 (2026) — superseded in its statistical methodology by §6.5 of the present work.