Skip to content

[有界计算] complete-component 质量–边界锁定:K/B 联合 Hessian、Palm score 方差与近临界 composition ray #774

Description

@LightChainr

目标

复用 PR #739 907a9d94 已完成的 two-fugacity complete-component activity,不建新 transfer engine、不做 Monte Carlo,直接检验一个由 near-critical intensity + Palm score Ward identity 推出的更强几何猜想。

对 NN complete horizontal-winding component 的 component-Palm law,记

K = occupied-site count,
B = number of distinct external vacant boundary sites,
q=1-p,
S = K/p - B/q.

#739 已有 exact fixed-width identity

d log nu_w / dp = E_Palm[S].

直接 activity 的二阶导数还给出

(d^2/dp^2) log nu_w
  = Var_Palm(S) - E[K]/p^2 - E[B]/q^2.              (1)

如果近临界强度存在

nu_w(p_c + lambda/(a_t w^(3/4)))
  = w^-1 I(lambda) [1+o(1)],

并且 critical complete winding cluster 的 K,B 都是 w^(d_f) 量级、d_f=91/48,则有两层可检验 Ward 预测:

E[K]/p - E[B]/q = O(w^(3/4)),
Var(S) - E[K]/p^2 - E[B]/q^2 = O(w^(3/2)).           (2)

由于 91/48 > 3/2,第二式进一步候选为

Var(S) / [E[K]/p^2 + E[B]/q^2] -> 1.                 (3)

第一式配合 E K ~ w^(91/48)

E[B]/E[K] -> q/p.                                     (4)

更强的 composition-locking 猜想

KB 的 macroscopic shape amplitude 本身有非退化 w^(91/48) 波动,而 (3) 仍成立,则 leading covariance 必须几乎 rank one,并沿

(K,B) proportional to (p,q).

因此提出可杀掉的强猜想

w^(-91/48) (K,B) => A (p_c,q_c)                       (5)

for some nondegenerate random amplitude A, while the transverse score

S = K/p - B/q

has only w^(91/96)-scale fluctuations at leading composition-CLT order (or smaller). (5) is a conjecture, not an acceptance condition;本单首先只测试 exact finite covariance geometry。

Phase A — 复用现有 operator,widths 2..8

Pinned input:PR #739 docs/manuscripts/geometric-balance/tagged-span-resolvent.md / scripts/tagged_winding_span.py 对应当前 head 907a9d94。先确认 branch 最新 blob;不要重建旧 age-cutoff solver。

对 NN graph、每个可闭合 width w=2..8,至少在以下参数评估:

  1. p_ref = 0.59274605079210(仅作 near-critical diagnostic reference,不是本任务新 pc);
  2. 若数值稳定,再加 p_ref +/- c w^(-3/4) 两个预先声明的 lambda 点,c 取一个固定小常数;不要事后扫最漂亮的 c。

从 two-fugacity resolvent 的一、二阶 log derivatives 直接得到

nu,
E K, E B,
Var K, Var B, Cov(K,B),
E S, Var S.

必须 exact / certified 检查

  1. 一阶物理曲线导数:
(d/dp) log nu == E[K]/p - E[B]/q.
  1. 二阶物理曲线导数与 (1) 一致;导数可用 activity 的解析/自动微分,不能用粗 finite difference 当 certificate。

  2. covariance matrix PSD;记录最小 eigenvalue / numerical certificate boundary。

  3. direct activity 与 tagged representation 的 nu 必须在已有重合宽度一致。

主要输出

每个 width 报告机器可读表:

mean_ratio = (E B/E K)/(q/p),
score_mean = E S,
score_var = Var S,
shot_scale = E K/p^2 + E B/q^2,
ward2_ratio = score_var/shot_scale,
Cov(K,B) eigenvalues/eigenvectors,
angle of principal covariance eigenvector to (p,q),
condition number,
Corr(K,B).

额外报告但不要拟合成定理

w^(55/48) * [E B/E K - q/p],
ward2_ratio-1,
small_eigenvalue / trace,
Var(S) / E[K].

若 widths 2..8 已显示强烈不相容(例如 principal direction 不向 (p,q) 靠近、ward2_ratio 远离 1 且恶化),直接记录 negative;不要加更大 width 挽救。

判决

给四项独立 verdict:

  • MEAN_WARD: (4) compatible / incompatible / unresolved;
  • SECOND_WARD: (3) compatible / incompatible / unresolved;
  • COVARIANCE_LOCKING: leading covariance toward rank-one (p,q) / no / unresolved;
  • STRONG_RANDOM_RAY: (5) remains plausible / killed / not testable at these widths。

不做

交付建议:

notes/complete-component-kb-locking-YYYYMMDD.md
results/.../kb-locking.json

相关:#739, #760, #770, draft PR #773

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    priority:P2Deferred research or on-demand support; no default new compute allocation.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions