Skip to content

[理论/有界计算] 独立热正规坐标与 Harris 正控制:用质量 κ 校准 M''/(2M') #770

Description

@LightChainr

动机

#737 已经给出完整有限 root-source Hessian 与 annealed/quenched root 恒等式;Round-4 又发现一个需要独立校准的现象:

c2_M(L) := M''_L(p_c)/(2 M'_L(p_c))

在 L=3,4 的 exact torus 上约为 0.2911、0.2942,变化很小。一个自然解释是:M'' ~ L^(3/4) 的 leading 部分主要来自裸 occupation coordinate p 相对于真正 thermal scaling field 的二阶非线性,而不是新的 even primary。

这个解释不能用同一 M 数据自证。需要一个独立 thermal coordinate

同时,#737 的 quenched-root 公式提供了一个非常干净的 Harris-criterion 正控制,可以把 source Hessian 与 near-critical scaling 接起来,而不改 original-U 合同。

Part A — mass-defined thermal coordinate

对 subcritical square-site 连接质量 / inverse correlation length kappa(p),假设标准 near-critical form

kappa(p) = C (p_c-p)^(4/3) [1 + c1 (p_c-p) + ...].

定义物理 thermal coordinate(只到整体线性归一化)

t_mass(p) = - kappa(p)^(3/4),   p<p_c.

则其坐标曲率可在不知道振幅 C 的情况下写成

c2_mass(p)
 = t_mass''/(2 t_mass')
 = 1/2 [ kappa''/kappa' - (1/4) kappa'/kappa ].

单独两项在 p->p_c 时发散,但 leading 4/3 singularity 精确抵消;若上式趋有限极限,就是 a2/a1 的独立估计。

目标 A1

判断

lim_{p->p_c^-} c2_mass(p)

是否存在,以及是否与 torus matching-function 的

lim_{L->inf} M''/(2M')

同一个 microscopic p coordinate 下相容。

数据/方法优先级

  1. 先检索已有 square-site correlation-length / connection-mass 高精度数据或 series;
  2. 若仓库 [路线2026-09 / 理论] 圆柱尾质量的定量局部性:γ_w−κ 是否小于 1/w #760 / first-exit mass 工具有足够靠近 p_c 且 width-locality 可认证的数据,复用,不新建平行 mass engine;
  3. gamma_w-kappa 误差在近临界范围尚不可控,明确写 BLOCKED_ON_MASS_LOCALITY,不要拿同一 nu_w 反拟合 kappa

必须报告导数估计对 interpolation / fit ansatz 的敏感性。两个大项相消,数值病态是本题核心风险之一。

Part B — exact matching-pair parity

利用有限恒等式

M_G(p) = -M_Ghat(1-p)

记录

M_G' = +M_Ghat',
M_G'' = -M_Ghat'',
M_G''' = +M_Ghat''', ...

因此

c2_G(p) = - c2_Ghat(1-p)

exactly at every finite size。解释这与 pair-symmetric thermal coordinate

t_G(p) = - t_Ghat(1-p)

的相容性。这个 parity test 是免费 exact control,不需要新采样。

Part C — Harris 正控制(由 #737 exact formulas 推出)

#737 对 independent additive-p quenched offsets p_i=p+sigma xi_i 已有

E q = p_* + sigma^2 M''/(2 N M') + O(sigma^4),
Var q = sigma^2/N + O(sigma^4)

(centrally symmetric iid noise)。取 square torus N=L^2,并使用 clean thermal window L^-3/4

Std[ L^(3/4)(q-p_*) ] ~ sigma L^-1/4,
Var[ scaled root ] ~ sigma^2 L^-1/2,
Mean scaled bias ~ sigma^2 c2 L^-5/4.

目标 C1

把这写成一个 exact-finite -> scaling 的 Harris positive control:variance crossover exponent -1/2 正好对应 uncorrelated disorder coupling y_Delta = 2/nu - d = -1/2;标准差 exponent 是 -1/4

这里不是声称发现新 disorder fixed point,而是检查 source/root machinery 是否与已知 Harris irrelevance 一致。

目标 C2 — centered environment

若每个 realization 先减去自己的 spatial mean,则 #737 已证明 O(sigma) root fluctuation 消失,而 leading mean curvature bias保留。明确比较 centered / uncentered 两个 ensemble;不能把“mean-zero marginals”和“每个样本 mean-zero”混为一谈。

Part D — 长程相关 source(设计,不默认生产)

Cov xi(x) xi(0) ~ |x|^-a,则 spatial average 的方差在 a<2 时约 L^-a。因此 thermal-scaled root first-order fluctuation预测

Std ~ L^(3/4-a/2),

relevance boundary a=3/2,与 Weinrib–Halperin extended Harris criterion a nu = 2(nu=4/3)一致。

这里只需给数学 derivation 与 source-Hessian Fourier weighting formula。除非现有 #737 Fourier modes 足够做 tiny control,否则不请求新 correlated-disorder Monte Carlo

判决

最终给四项:

1. c2_mass exists / unresolved / contradicted;
2. c2_mass vs c2_M compatible / incompatible / not yet comparable;
3. Harris exact-finite scaling control PASS / FAIL;
4. whether a mass-normalized thermal coordinate should replace raw p in subsequent derivative/CFT interpretation.

若 A 被近临界 mass locality 卡住,B/C 仍应完成;不要因为一个数值导数难算而停止整个交付。

不做

相关:#737, #760, #768, #739

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

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions