ZFCρ Paper LXVI: Gradual Chiseling and the Spectral Bridge — Two-Regime Structure of the Centered Residual ACF
ZFCρ 第LXVI篇：渐进凿除与谱桥补偿——Centered Residual的两阶段结构

§1. The Sole Open Core After Paper 65

Paper 65 established the conditional architecture for 59.1: B_j = B_count + B_resid + B_gauge, with count term controlled by Branch-Orthogonality (Paper 64), gauge defect controlled by j-smooth adjacent reference (Paper 65), and cross-term small. The sole remaining open core is:

Ψ_η(m_j) = O(j)

This paper investigates the scale structure of Ψ_η(W): why it does not grow as a simple power law, and what mechanism returns full-block Ψ to O(j).

§2. Via Negativa: The 13 Kills

Organized by layer, not chronology. Each kill eliminates a candidate shortcut mechanism.

Layer 1 — Coordinate and Low-Rank Illusions (Papers 64–65): Kill 1: β-pinning is restating (Null 1 permutes h(p)/p within each gap class, preserving Γ(g) and B_j identically). Kill 2: rank-one is not IC-specific (Null 1 gives higher s₁/s₂; rank-one energy comes from count marginal baseline, not IC recursion).

Layer 2 — Mean-Mode Artifacts (Paper 65): Kill 3: cross-branch anti-correlation is algebraic artifact of μ_A·μ_M < 0. Kill 4: global reference centering fails (μ_A ∼ C_A/2^j, scale mismatch). Kill 5: Haar-Carleson multiscale fails (geometric energy decay).

Layer 3 — State-Space and Pointwise ACF Shortcuts (Paper 66): Kill 6: gap-state transfer operator (τ ≈ 1, instant mixing, R² = 0.26). Kill 7: increment curvature/Abel (45% monotone violations, 50/50 convexity).

Layer 4 — Screening Models (Paper 66): Kill 8: naive Thermo layer hierarchy (ρ_ret × τ_0 = 2.77 > 1, diverges). Kills 9+10: exponential cutoff and pointwise tail compensation (initially interpreted in working notes as "negative ACF tail → Ψ_full = O(1)"; revealed as noise artifact from interpolated ACF; partially reversed at integrated level in §6).

Layer 5 — Ancestry Shortcuts (Paper 66): Kill 11: chain boundary = screening reset (mean chain = 1.3, too fragmented). Kill 12: GCD factor-sharing (constant across all lags). Kill 13: per-boundary constant attenuation (ACF × boundaries grows as lag^{0.7}).

After all five layers: simple positive exponential cutoff, pointwise negative-tail interpolation, and clean Thermo hierarchy are excluded; what remains is signed shell-level spectral bridge. Low-dimensional ancestry shortcuts are excluded; influence-cone overlap and renewal remain Paper 67 candidates. The surviving bare mechanism is distributed cumulative micro-erosion from min operations combined with block-scale spectral bridge compensation.

§3. Doukhan Confirmation: ACF Suffices for the Second-Order Target

θ_k/γ_k ≈ 2.5 universally across j = 25–32 and lag 1–2000. R²_nonlinear/R²_linear ≈ 0.94–1.01 at all lags. Under the conditional-mean and quantile-bin diagnostics tested, nonlinear predictors do not outperform linear predictors. For the second-order target Ψ, ACF is the sufficient empirical object.

§4. Regime 1: Gradual Chiseling and Power-Law Growth

Theorem 66.A (Local Micro-Erosion Law). If per-step erosion satisfies ε_k ∼ C/k + o(1/k), C > 0, then γ(k) = ∏(1 − ε_s) ∼ k^{−C}. Proof: ln γ(k) ≈ −C·H_k ≈ −C·ln k.

Empirically C ≈ 0.3. The mechanism: each logarithmic shell contributes constant erosion budget. Individual lag steps contribute ≈ C/k.

However, 66.A alone cannot yield 59.1. Power-law extrapolation to full block gives Ψ ∼ m_j^{1−C}/((1−C)(2−C)) ∼ (2^j/j)^{0.7}/1.19, which grows exponentially in j.

Power-law prediction vs reality:

j	α_fit	Predicted Ψ/j	Actual Ψ_full/j	Overshoot
27	0.47	137	0.015	9000×
29	0.40	593	0.76	780×
31	0.34	2674	0.91	2900×
32	0.31	5551	15.66	350×

Note: Actual Ψ_full/j values fluctuate wildly (0.015 to 15.66) because each j has only one block; the full-block identity Ψ = S²/(m·σ²) follows a χ²(1)-type distribution. The pattern shows no systematic super-linear growth. Sub-block ensemble statistics (§5) provide far more reliable estimates.

§5. Regime 2: Sub-Block Ψ Peak-Then-Crash

Sub-block Ψ(W) = Var(sub-block sums)/(W·σ²), computed with K = m/W sub-blocks. Confidence zones: K ≥ 30 (high), 10 ≤ K < 30 (medium), K < 10 (exploratory).

Universal pattern across j = 22–32: Ψ(W) grows as power law for W < 0.5% m_j, peaks, then crashes 5–40×.

j	Peak W	Peak Ψ/j	Crash W	Crash Ψ/j
25	10K	2.53	100K	1.20
29	100K	13.36	5M	1.02
31	1M	35.08	10M	2.73
32	1M	59.28	10M	10.50

§6. Integrated Tail Compensation

Kill 10 (pointwise negative tail) partially reversed: individual ACF values at large lag are at noise floor (|γ| < 0.003), but their integrated shell contribution produces genuine negative mass in the sub-block variance. Pointwise negative tail was not measurable; shell-integrated negative mass is real.

§7. Spectral Bridge Identity and Circularity Warning

The finite-block identity Ψ_full = S²/(m·σ²) diagnoses the spectral bridge but cannot prove 59.1, because small S = Ση is the target itself. Full-block identity serves as consistency check; sub-block Ψ is the independent empirical capstone.

Crucially, the sub-block Ψ measurement (§5) does not depend on the size of the full-block sum Ση. It is computed directly from the ensemble of K distinct sub-block sums, independent of whether Ση is small. This is Paper 66's primary empirical claim. The full-block identity serves only as a consistency cross-check and does not participate in the evidence chain.

§8. Conditional Theorems

Theorem 66.B (Spectral Bridge Criterion). Let M_+(j) = Σ_r S_j^+(r) be the total positive shell mass and M_-(j) = −Σ_r S_j^-(r) the total negative shell mass. Regime 1 (§4) implies M_+(j) grows as a power of m_j. If the negative shell mass nearly matches the positive:

M_-(j) = M_+(j) − g(j), where g(j) = O(j),

then Ψ_η(m_j) = M_+(j) − M_-(j) + O(1) = g(j) + O(1) = O(j), and Paper 65's conditional theorem yields the L² analogue of 59.1. The "spectral bridge" is this near-exact cancellation between positive and negative shell masses, with only an O(j) deficit.

Observation 66.C. The sub-block Ψ peak-then-crash is universal across j = 22–32, constituting direct empirical evidence that M_-(j) ≈ M_+(j) with small deficit.

§9. Two-Regime Unified Picture

Local chiseling explains why Ψ grows (Regime 1). Global spectral bridge explains why it does not keep growing (Regime 2). Both are needed: chiseling alone gives Ψ ∼ m^{0.7} (too large); constraint alone gives Ψ = 0 (tautological). Together: Ψ peaks then crashes to O(j).

§10. SAE Reading

The 13 kills and the surviving mechanism exhibit structural resonance: hypothesis elimination at the research level mirrors min-operation erosion at the mathematical level. Both are instances of chiseling (凿). The research process and the mathematical object share the same structural operation.

§11. Remaining Proof Task (Paper 67 Target)

Derive the spectral bridge compensation from IC recursion. Candidate frameworks: influence-cone overlap theory, branch-normal Ward identity, j-smooth reference flow, finite-block gauge pinning.

§12. Status Map

Content	Status
Theorem 66.A: micro-erosion → power law	Conditional theorem
Theorem 66.B: spectral bridge → Ψ = O(j)	Conditional theorem
Observation 66.C: sub-block Ψ crash	Empirically confirmed
Power-law overshoot 100–10,000×	Empirically confirmed
13 hypotheses killed	Empirical/analytic
Spectral bridge from IC recursion	Open (Paper 67)
Unconditional proof of 59.1	Open (Paper 67–68)

§1. Paper 65后的唯一Open Core

1.1. Paper 65的条件定理

Paper 65建立了59.1的条件架构：

B_j = B_count + B_resid + B_gauge

其中count term由Branch-Orthogonality控制（|B_count|/√(jD) < 0.5，Paper 64），gauge defect由j-smooth adjacent reference控制（Paper 65 §4），cross-term小。剩余的唯一open core是centered residual η的spectral bound：

Ψ_η(m_j) = O(j)

其中Ψ_η(W) = 1 + 2·Σ_{k=1}^{W-1} (1 - k/W)·γ_η(k)是标准partial-sum variance放大因子。

1.2. Paper 66的目标

理解Ψ_η(W)的尺度结构：它为什么不是简单的幂律增长？什么机制使full-block Ψ回到O(j)？

---

§2. Via Negativa：13次否决

2.1. 否决的方法论地位

每一次否决都排除了一个看似合理的shortcut explanation，缩小了可能机制的搜索空间。否决按层级组织，不按时间顺序。

2.2. Layer 1：坐标与低秩幻象（Papers 64-65）

Kill 1. β-pinning是restating。 Paper 64的Null 1在每个gap class内置换h(p)/p，保持Γ(g)和B_j不变，给出identical β structure。β-pinning不是IC-specific的reduction，只是B_j的gap-coordinate restatement。

Kill 2. Rank-one不是IC-specific。 Null 1 surrogate给出更高的s₁/s₂ ratio（347 vs 321）。Rank-one energy主要来自count marginal N(u,g)的极端rank-one与gap-conditioned Γ(g) profile，属marginal baseline structure，不属IC-specific mechanism。

2.3. Layer 2：均值模artifact（Paper 65）

Kill 3. Cross-branch anti-correlation是代数artifact。 Branch centering实验：减去branch-specific mean后，AA-ACF平台从0.25掉到0.025，cross-ACF从-0.06翻正到+0.01。Raw cross-branch负相关完全来自μ_A·μ_M/σ² < 0。λ calibration使α·μ_A + (1-α)·μ_M = 0，branch mean mode的贡献代数消去。

Kill 4. Global reference centering失败。 μ_A(j) ~ C_A/2^j，相邻j之间差异2倍。j = 32时global reference使variance增大3.6倍。Adjacent-block centering在所有j稳定work。

Kill 5. Haar-Carleson多尺度失败。 能量沿scale geometric decay，不是O(1)/scale。

2.4. Layer 3：状态空间与逐点ACF shortcut（Paper 66）

Kill 6. Gap-state transfer operator。 Gap-state Markov chain的spectral gap ≈ 0.93，relaxation time τ ≈ 1.08（near-instant mixing）。但R² = 0.26：gap-conditional mean只解释26%方差。Within-state residual ξ的ACF和Ψ几乎未改善。Long-range memory在within-state ρ-landscape drift里，不在gap transitions里。

Kill 7. Increment curvature/Abel。 Fine-resolution ACF有45% monotone violations，convexity 50/50。Abel summation identity成立但不给bound（因为interior sum可正可负）。

2.5. Layer 4：Screening模型（Paper 66）

Kill 8. Naive Thermo层级hierarchy。 Thermo VII-VIII的renewal encapsulation framework预测geometric hierarchy with ρ_ret × τ_0 = 0.76 × 3.65 = 2.77 > 1。Naive geometric hierarchy diverges。Multi-exponential peeling只找到1-2个massive components（τ ~ 30000-57000），不是clean layered decomposition。

Kill 9+10. Exponential screening cutoff与tail compensation。 初始Large-W ACF数据（W = 100K at j = 25）显示SumGamma变负，本系列在Paper 66 working notes阶段曾将此解读为"ACF negative tail → tail compensation → Ψ_full = O(1)"。但full-block identity Ψ_full = S²/(m·σ²) = 8.87与interpolated ACF积分Ψ(W = 947K) = -440矛盾。Diagnosis：large-lag ACF在noise floor（|γ| < 0.003），log-spaced points之间的linear interpolation放大noise创造artifactual negative structure。逐点negative tail不可测。

但Kill 10后来在§5被部分overturn：逐点不可测的tail在integrated level（sub-block variance）产生genuine补偿效应。个体不可测，集体有效应。

2.6. Layer 5：祖先shortcut模型（Paper 66）

Kill 11. Chain boundary = screening reset。 Add-chain mean length = 1.3，76%是单素数链。Lag ≥ 10时same-chain pairs = 0。所有long-lag correlation都是cross-chain。ACF在1000个chain boundaries后仍为0.02。Screening不是discrete reset。

Kill 12. GCD factor-sharing = ACF source。 Pr(GCD(p-1, p+k-1) > 1) = 1.000在所有lag（所有偶数share factor 2）。Pr(GCD > 10)和E[log₂ GCD]在lag > 10后几乎constant。GCD不track ACF。Small prime factors ubiquitous，不carry correlation information。

Kill 13. Per-boundary constant attenuation。 ACF × boundaries grows as lag^{0.7}，not constant。每个boundary attenuates by ~0.3%平均，但ACF衰减速度比boundary累积慢得多。Attenuation per step不是constant fraction。

2.7. 否决后的搜索空间

所有macro-cancellation被排除（Kills 1-5）。所有state-space shortcuts被排除（Kill 6）。所有pointwise ACF structural assumptions被排除（Kill 7）。简单的positive exponential cutoff/pointwise negative-tail interpolation/clean Thermo hierarchy被排除（Kills 8-10）；剩下的是signed shell-level spectral bridge。低维ancestry shortcuts（chain/GCD/constant attenuation）被排除（Kills 11-13）；influence-cone overlap/renewal仍是Paper 67候选。

剩余的bare mechanism：min操作的分布式累计微侵蚀，加上block尺度的谱桥补偿。

---

§3. Doukhan确认：二阶目标下ACF是充分对象

在进入surviving mechanism之前，Doukhan conditional dependence test确认了一个关键structural fact。

θ_k/γ_k ≈ 2.5是universal constant。 跨j = 25-32，跨lag 1-2000。Doukhan θ-coefficient和ACF严格成比例。

Linearity test： R²_nonlinear/R²_linear ≈ 0.94-1.01在所有lag和所有j。在本文测试的conditional-mean和quantile-bin diagnostics中，nonlinear predictor未优于linear predictor。

含义： 对Ψ_η这个二阶目标，ACF是充分的经验对象。问题reduce到integrated ACF bound。

---

§4. Regime 1：渐进凿除与幂律增长

4.1. 物理图景

IC递推ρ(n) = min(ρ(n-1)+1, min_{ab=n}[ρ(a)+ρ(b)+2])中，每一次min操作都对correlation产生微小侵蚀。不是"reset"（erase all memory），是"erode"（remove tiny fraction）。

1000步侵蚀后correlation从0.18降到0.02：total reduction ~9×。平均per-step attenuation约0.2%，局部值在0.2%-0.5%间波动。

4.2. 幂律ACF的微侵蚀推导

Theorem 66.A（Local Micro-Erosion Law）。 若per-step erosion fraction ε_k满足

ε_k ~ C/k + o(1/k)，C > 0

则

γ(k) = ∏_{s=1}^{k} (1 - ε_s) = exp(-Σ ε_s) ≈ exp(-C·ln k) = k^{-C}

推导。 ln γ(k) = Σ_{s=1}^{k} ln(1 - ε_s) ≈ -Σ ε_s = -C·H_k ≈ -C·ln k。其中H_k是harmonic series。因此γ(k) ~ k^{-C}。

状态： Conditional theorem。条件是ε_k ~ C/k的specific form。C的值由IC combinatorics决定，实测C ≈ 0.3。

4.3. 为什么ε_k ~ 1/k

候选机制：每个logarithmic shell（从k到2k）提供常数级erosion budget。即：在尺度k附近，IC min递推产生的新"凿击"总量约为常数C。因此individual lag step的平均erosion ≈ C/k。

更深层来说：距离k的两个素数p和p+k，它们的IC construction tree共享intermediate node的概率随k的倒数衰减。这来自于factorization树的几何截面随距离扩散。

形式化推导留待Paper 67。

4.4. Regime 1的Ψ行为

Ψ(W) = 1 + 2·Σ_{k

Riemann sum近似给出Ψ(W) ~ W^{1-C}·∫₀¹(1-x)·x^{-C}dx = W^{1-C}·B(1-C, 2) = W^{1-C}/((1-C)(2-C))。

对C ≈ 0.3：Ψ(W) ~ W^{0.7}/(0.7 × 1.7) ≈ 0.84·W^{0.7}。

4.5. 66.A不足以给59.1

如果幂律严格延伸到full block m_j ~ 2^j/j：

Ψ(m_j) ~ (2^j/j)^{0.7} / (0.7 × 1.7)

这是j的指数函数，远超O(j)。局部幂律解释Regime 1的ACF形状，但不控制full-block Ψ。

---

§5. Regime 2：Sub-Block Ψ的Peak-Then-Crash

5.1. 实验设计

对j = 22-32的每个block I_j，divide into K = m/W个complete sub-blocks of size W。计算sub-block sums S_b = Σ_{primes in b} η_p。

Ψ(W) = Var_b(S_b) / (W·σ²_η)

这是proper ensemble statistic，precision ~ 1/√K_eff。

5.2. 核心发现：Ψ(W)先增后降

表2. Sub-block Ψ(W)/j（representative j values）。

j = 29（m = 26.2M）：

W	K	Ψ/j	SE/j	Confidence
100	262K	0.28	0.001	High
1000	26K	1.32	0.01	High
10000	2620	5.24	0.14	High
100K	262	13.4	1.17	High
500K	52	10.5	2.08	Medium
1M	26	8.72	2.47	Medium
5M	5	1.02	0.72	Exploratory

Confidence分级：K ≥ 30 = High，10 ≤ K < 30 = Medium，K < 10 = Exploratory。

Pattern在所有j一致： Ψ(W)先按幂律增长，在W ≈ 0.5% m_j处到达峰值，然后crash 5-40倍。

表3. Peak-crash summary across j。

j	Peak W	Peak Ψ/j	Crash W	Crash Ψ/j
25	10K	2.53	100K	1.20
27	50K	6.28	1M	2.02
29	100K	13.36	5M	1.02
31	1M	35.08	10M	2.73
32	1M	59.28	10M	10.50

表中最大W的crash点常处于Medium/Exploratory区（K < 30），但peak后下降的趋势已经在K ≥ 30的区间出现。最大W点用于显示趋势延伸，不单独承担证明负担。

5.3. 幂律外推的failure

表4. 幂律预测vs实际full-block Ψ。

j	α_fit	Predicted Ψ/j	Actual Ψ_full/j (identity)	Overshoot
27	0.47	137	0.015	9000×
29	0.40	593	0.76	780×
31	0.34	2674	0.91	2900×
32	0.31	5551	15.66	350×

α_fit从W = 1000和W = 10000两点拟合。幂律外推超出实测100-10000倍。

注：Actual Ψ_full/j（identity列）在j之间剧烈波动（0.015到15.66），因为每个j只有一个block，Ψ_full = S²/(m·σ²)服从χ²(1)-type分布（单次采样的variance极大）。这些数值的pattern是不随j systematic grow，但个体j值高度noisy。Sub-block ensemble statistics（§5.2 表2-3）提供far more reliable的Ψ estimates。

这是Paper 66最强的单一证据： 幂律ACF在full-block尺度上fundamentally wrong。Power-law是Regime 1的local description，不是global truth。

---

§6. Integrated Tail Compensation

6.1. Kill 10的partial reversal

§2中Kill 10杀死了"逐点ACF negative tail"：individual ACF values at lag > 3000在noise floor（|γ| < 0.003），无法reliable determine sign。

但sub-block Ψ(W)的crash是genuine ensemble statistic（K ≥ 30的zone有SE << crash magnitude）。

Resolution： 逐点negative tail不可测。但millions of noise-floor ACF values的integrated shell contribution产生genuine negative mass。

个体不可测，集体有效应。

6.2. Sub-block Ψ crash = integrated compensation

Ψ(W)是ACF的Abel加权积分。Ψ(W)从峰值下降意味着：新增的large-lag shell（从W_peak到W_current）的net contribution为负。

不需要每个individual γ(k)都measurably negative。只需要它们的sum为负——而这正是sub-block variance直接测量的。

6.3. 逐点vs积分的关系

层级	状态	证据
逐点large-lag ACF sign	Not reliably measurable	Individual γ at noise floor (	γ	< 0.003)
Shell-integrated compensation (inferred from sub-block Ψ)	Supported	Sub-block Ψ crash with K ≥ 30

Kill 10杀死的是"从sparse/noisy large-lag ACF点线性插值得到稳定negative tail"。复活的是"sub-block Ψ(W)显示large-lag shells的净积分贡献为负"。两者不矛盾。

---

§7. 谱桥恒等式与循环警告

7.1. Finite-block identity

对block内m个centered residuals η₁, ..., η_m，定义σ² = (1/m)·Σ η_i²，以及lag-normalized ACF r(k) = [1/((m-k)·σ²)]·Σ_{i=1}^{m-k} η_i·η_{i+k}（注意η_i不再额外减去sample mean）：

Ψ_full(m) = 1 + 2·Σ_{k=1}^{m-1} (1-k/m)·r(k) = (Σ η_i)² / (m·σ²)

这是exact identity。右边 = S²/(m·σ²)。

7.2. Full-block identity作为cross-check

j	Ψ_full/j (identity)
22	0.27
25	0.36
27	0.015
29	0.76
31	0.91
32	15.66

Single-block values wildly fluctuate（χ²(1)-type noise from ONE block per j）。但不systematically grow。

7.3. 循环警告

Full-block identity说 Ψ_full = S²/(m·σ²)。如果用"S small → Ψ small"作为proof of 59.1，循环：因为"S = Ση small"本身就是residual版的59.1 core。

Adjacent centering去除了wrong branch mean gauge，但它本身不prove residual zero-mode small。Paper 65的分解B_j = B_count + B_resid + B_gauge中，gauge和count已安全，真正的open core是B_resid = Ση。

Full-block identity是diagnostic tool和consistency check，不是proof。Sub-block Ψ才是independent empirical evidence。

Sub-block Ψ measurement（§5）不依赖于full-block Ση的size。它直接从K个distinct sub-block sums的ensemble computed，与Ση是否small无关。这是Paper 66的primary empirical claim。Full-block identity仅作为consistency cross-check，不参与evidence chain。

---

§8. 条件定理

8.1. Theorem 66.A：Local Micro-Erosion Law

若adjacent-centered residual η_p的ACF满足recursive attenuation

γ(k+1) = γ(k)·(1 - C/k + o(1/k))

则γ(k) ~ k^{-C}。

状态： Conditional theorem。条件是per-step erosion form ε_k ~ C/k。实测C ≈ 0.3。

66.A解释Regime 1，但幂律外推到full block给Ψ ~ m^{0.7}，远超O(j)。66.A alone不给59.1。

8.2. Theorem 66.B：Spectral Bridge Criterion

设γ_η(k)在lag区间[1, m_j]内的正shell mass和负shell mass分别为

M_+(j) = Σ_r S_j^+(r)，其中 S_j^+(r) = Σ_{2^r ≤ k < 2^{r+1}} max(γ(k), 0)

M_-(j) = -Σ_r S_j^-(r)，其中 S_j^-(r) = Σ_{2^r ≤ k < 2^{r+1}} min(γ(k), 0)

Regime 1（§4）给出M_+(j)随m_j呈power-law增长。这是"spectral bridge"的一侧。

Theorem 66.B（Spectral Bridge Criterion）。 若negative shell mass nearly matches positive shell mass：

M_-(j) = M_+(j) - g(j)，其中g(j) = O(j)

则Ψ_η(m_j) = M_+(j) - M_-(j) + O(1) = g(j) + O(1) = O(j)，结合Paper 65条件定理即得59.1的L²类比。

状态： Conditional theorem。条件是正负shell mass之间的deficit g(j)为O(j)。这是非平凡条件：M_+(j)可以很大（power-law growth），但M_-(j)必须几乎精确match它，只留O(j)的gap。Sub-block Ψ crash（66.C）是M_-(j) ≈ M_+(j)的empirical evidence。

66.B解释Regime 2： 正shell mass（来自Regime 1 power-law）被负shell mass（来自block-scale补偿）几乎精确抵消。"Spectral bridge"指的就是这种near-exact cancellation。

8.3. Observation 66.C：Sub-Block Peak-Then-Crash

对j = 22-32，sub-block Ψ(W)在W ≈ 0.5% m_j处到达峰值后crash 5-40倍。

Power-law外推预测的full-block Ψ超出实测100-10000倍。

该发现是spectral bridge（66.B）的直接经验证据，在K ≥ 30的高置信度zone可robust复制。

---

§9. 两阶段谱结构的统一图景

9.1. Local chiseling解释增长

每步min操作chiseling away a small fraction of correlation。Cumulative product给power-law ACF。Ψ(W)在pre-compensation window（W < 0.5% m_j）按W^{0.7}增长。

9.2. Global bridge解释抑制

某种block-scale机制（可能来自λ-calibration/branch gauge/IC递推的spectral structure）使f(0)远小于power-law预测。Ψ(W)在post-peak window crash。

9.3. 图景的honest bounds

Local chiseling是identified mechanism with heuristic basis（每个log-shell常数erosion budget）。

Global bridge是observed phenomenon（sub-block crash），其IC递推来源尚未推导。

两者结合解释59.1：增长+抑制 = O(j)。但formal proof requires从IC axiom推出both components。

---

§10. SAE四分形读法

10.1. 研究过程与数学对象的结构共振

13次否决在research level的运作和IC min操作在mathematical level的运作share同一结构：凿。

层面	操作	效果
Research	Hypothesis killed	搜索空间缩小
Mathematics	min(A,B)	Correlation eroded

研究方法和被研究对象的structural operation coincide。这不是单纯修辞，而是一个candidate operational correspondence。但本文不把该对应作为数学证明输入。

10.2. 四分形对应

Paper	Phase	Content
62	标而不构	双河坐标系
63	加法路径	Increment正交性
64	乘法路径	Branch-Orthogonality, 2D anatomy
65	闭合路径开始	Branch centering, 条件定理
66	闭合路径继续	13 kills → two-regime mechanism

10.3. 负结果的方法论地位

传统科研把negative results视为waste。SAE methodology将其提升为first-class output。Paper 66的13 kills不是"13次失败"，是"13步搜索空间缩小"。没有这13步，surviving mechanism（gradual chiseling + spectral bridge）不可能被identify。

Paper 60 §9.5的教训"wrong hypotheses have value because each kill sharpens understanding"在Paper 66得到最完整的实践。

---

§11. 剩余证明任务与Paper 67 Target

11.1. Paper 66完成了什么

1. 精确识别了Ψ_η(W)的两阶段谱结构
2. 通过13次否决排除了shortcut explanations
3. 识别了surviving local empirical law（power-law ACF），并提出gradual chiseling作为解释机制
4. 发现了global phenomenon（sub-block Ψ crash = spectral bridge）
5. 给出了两个conditional theorems（66.A和66.B）

11.2. Paper 66没有完成什么

1. 没有从IC axiom推导ε_k ~ C/k
2. 没有从IC axiom推导spectral bridge
3. 没有unconditional proof of Ψ = O(j)

11.3. Paper 67 Target

从IC递推推出谱桥补偿。

候选框架：

• Influence-cone overlap theory（直接递归协方差界）
• Branch-normal Ward identity（从branch gauge structure推spectral constraint）
• j-smooth reference flow（μ_A(j)的smoothness properties → gauge defect bound → indirect Ψ control）
• Finite-block gauge pinning（dyadic block geometry → zero-frequency notch）

每个候选框架在Paper 67中将被tested（可能部分被killed）。

---

§12. Status Map

内容	层级
B_j = B_count + B_resid + B_gauge	Decomposition（Paper 65）
Count term: Branch-Orthogonality controlled	Empirical confirmed（Paper 64）
Gauge term: adjacent reference controlled	Empirical confirmed（Paper 65）
Doukhan: θ_k/γ_k ≈ 2.5, linear dependence	Empirical confirmed（Paper 66）
Theorem 66.A: ε_k ~ C/k → γ ~ k^{-C}	Conditional theorem
Theorem 66.B: spectral bridge → Ψ = O(j)	Conditional theorem
Observation 66.C: sub-block Ψ peak-then-crash	Empirical confirmed
Power-law overshoot 100-10000×	Empirical confirmed
13 hypotheses killed	Empirical/analytic
ε_k ~ C/k from IC recursion	Open（Paper 67）
Spectral bridge from IC recursion	Open（Paper 67）
59.1 unconditional proof	Open（Paper 67-68）

---

Methods

A. 数据

IC function ρ(n)预计算到N = 10¹⁰（int16 binary，~20GB），包含455,052,511个素数。所有实验使用该数据集。

B. Centered residual

先定义raw residual：

x_p = [ρ(p-1) + 1 - λ·ln p - μ_12(p mod 12)] / p

其中λ = 3.8568（OLS），μ_12是residual class mean。

Adjacent-centered residual定义为：

η_p = x_p - μ_{τ(p)}^{adj}(j)

其中τ(p) ∈ {A, M}是p的branch label（add-path或mult-path），μ_τ^{adj}(j)是由相邻blocks j-1和j+1的同branch primes给出的reference mean。本文所有ACF和Ψ均基于η_p计算。

C. Sub-block Ψ estimation

对block I_j divide into K = m/W complete sub-blocks。S_b = Σ_{primes in b} η_p。Ψ(W) = Var_b(S_b) / (W·σ²_η)。Standard error: SE ≈ Ψ·√(2/(K-1))（assuming approximate independence of sub-blocks）。由于相邻sub-block sums可能仍相关，该SE是optimistic baseline。严格置信区间应使用block-jackknife或K_eff校正。本文的"High/Medium/Exploratory"分级按raw K给出，主要用于标注样本量，不作为严格独立性断言。

D. Full-block identity

Ψ_full = S²/(m·σ²)其中S = Σ η_p over entire block。这是O(m)计算。

E. Doukhan test

10 quantile bins，conditional mean E[η_{i+k} | η_i in Q_q]。θ_k = max_q |E[η_{i+k}|Q_q]| / σ_η。R²_nonlinear via 10-bin piecewise-constant predictor vs linear predictor。

F. Transfer operator

Gap classes g ∈ {2,4,6,...}，transition matrix P from consecutive pairs。|λ₂| = second eigenvalue。f(g) = E[x|gap=g]。ξ = x - f(g)。

G. Ancestry experiments

Add-path: ρ(p-1) = ρ(p-2)+1。Chain: consecutive add-path primes。GCD: gcd(p_i - 1, p_{i+k} - 1)。Chain boundaries = mult-path primes count in lag interval。

---

参考文献

[1] H. Qin, "ZFCρ Paper LXV," Zenodo (2026). DOI: 10.5281/zenodo.19687343.

[2] H. Qin, "ZFCρ Paper LXIV," Zenodo (2026). DOI: 10.5281/zenodo.19674521.

[3] H. Qin, "ZFCρ Paper LXIII," Zenodo (2026). DOI: 10.5281/zenodo.19659860.

[4] H. Qin, "ZFCρ Paper LXII," Zenodo (2026). DOI: 10.5281/zenodo.19656365.

[5] H. Qin, "ZFCρ Paper LIX," Zenodo (2026). DOI: 10.5281/zenodo.19480518.

[6] H. Qin, "ZFCρ Paper LX," Zenodo (2026). DOI: 10.5281/zenodo.19480563.

[7] H. Qin, "ZFCρ Thermodynamics Papers VI-VIII," Zenodo (2026). DOIs: 10.5281/zenodo.19658595, .19673078, .19688303.

[8] H. Qin, "SAE Methodology Paper 00 (Via Rho)," Zenodo (2026). DOI: 10.5281/zenodo.19657439.