5D Library › 多模态历法张量融合、模九十谐波拨盘与跨系统共振诊断 · English

第六章 集成的上限:孔多塞、高斯 Copula 与「共振」的基率

孔多塞陪审团定理

若 KK 个评判者各自以略高于一半的概率给出正确答案,并且相互独立,多数票的正确率会随 KK 增加而趋于 1:这是孔多塞陪审团定理 (Condorcet, Marquis de 1785)。它的关键前提是独立;评判者相关时,多数票的增益会大幅缩小 (Ladha 1992)。

程序设每套系统以 0.6 的概率给出正确的是非判断(这是为了说明而设的假设,不是本书的发现),用单因子高斯 Copula 引入相关系数 ρ\rho (Nelsen 2006),模拟 KK 套系统的多数票正确率。KK 个等相关信号平均之后的方差,相对于单个信号是 Var(X‾)σ2=ρ+1−ρK,\frac{\mathrm{Var}(\bar X)}{\sigma^2} = \rho + \frac{1-\rho}{K} , 它随 KK 增大只能降到 ρ\rho,而不能降到零。

推导

# Chapter 6. [p11] Condorcet's jury theorem with correlated systems.
# K systems each give the right yes/no answer with probability p = 0.6
# (an assumption for illustration, not a finding). Correlation enters
# through a Gaussian copula: system i is right when Z_i < q, with
# q = Phi^-1(p) and corr(Z_i, Z_j) = rho (one common factor). The
# majority's accuracy for K = 1, 3, 5, 9 and rho = 0, 0.3, 0.6, by
# simulation (100,000 draws, seed 70), and the variance floor of an
# average of K equally correlated signals: rho + (1 - rho) / K.
import math
import random
from statistics import NormalDist

P = 0.6
Q = NormalDist().inv_cdf(P)

def majority(k, rho, rng, n=100000):
    a, b = math.sqrt(rho), math.sqrt(1 - rho)
    good = 0
    for _ in range(n):
        f = rng.gauss(0, 1)
        right = sum(a * f + b * rng.gauss(0, 1) < Q for _ in range(k))
        good += right * 2 > k
    return good / n

if __name__ == "__main__":
    rng = random.Random(70)
    print("    rho    K=1     K=3     K=5     K=9   variance ratio K=9")
    for rho in (0.0, 0.3, 0.6):
        acc = [majority(k, rho, rng) for k in (1, 3, 5, 9)]
        print("   {:4.1f} {:7.4f} {:7.4f} {:7.4f} {:7.4f} {:12.3f}".format(
            rho, *acc, rho + (1 - rho) / 9))
    print("\nReading: independent systems that are each a little better")
    print("than chance add up fast; correlated ones barely add up. At")
    print("rho = 0.6 nine systems are hardly better than one.")
#     rho    K=1     K=3     K=5     K=9   variance ratio K=9
#     0.0  0.6014  0.6464  0.6814  0.7348        0.111
#     0.3  0.6001  0.6275  0.6404  0.6537        0.378
#     0.6  0.5962  0.6144  0.6203  0.6231        0.644
#
# Reading: independent systems that are each a little better
# than chance add up fast; correlated ones barely add up. At
# rho = 0.6 nine systems are hardly better than one.
多数票正确率与系统个数(单个系统正确率 0.6)。

独立时,9 套系统的多数票正确率从 0.6014 升到 0.7348;相关系数为 0.3 时只升到 0.6537;为 0.6 时只有 0.6231,9 套系统几乎不比 1 套更好。方差比在 K=9K = 9 时分别是 0.111、0.378 与 0.644。第五章显示,各系统之间的相关取决于所问的问题;对于与季节有关的问题,集成的增益可能很小。

双系统共振与正交分歧

作者提纲中的「双系统同向共振」,指东方的年度飞星与西方的太阳弧推运在同一年指向同一生活领域。程序设东方系统对某一领域「触发」的基率为 1/12(年度化忌落入该宫),西方系统的基率为 0.35,用高斯 Copula 引入相关。

推导

# Chapter 6. [p12] Two systems in one yearly window: co-resonance and
# orthogonal divergence under a Gaussian copula. System E (an eastern
# annual cycle) fires for a given life domain with base rate a = 1/12;
# system W (western solar-arc directions) fires with base rate b =
# 0.35. Latent normals with correlation rho decide firing. For rho = 0,
# 0.3 and 0.6: P(both fire), P(exactly one fires) and P(W fires | E
# fires), from 400,000 draws (seed 70).
import math
import random
from statistics import NormalDist

A, B = 1 / 12, 0.35
QA, QB = NormalDist().inv_cdf(A), NormalDist().inv_cdf(B)

def joint(rho, rng, n=400000):
    both = one = e = 0
    s = math.sqrt(1 - rho * rho)
    for _ in range(n):
        x = rng.gauss(0, 1)
        y = rho * x + s * rng.gauss(0, 1)
        fe, fw = x < QA, y < QB
        both += fe and fw
        one += fe != fw
        e += fe
    return both / n, one / n, both / e

if __name__ == "__main__":
    rng = random.Random(70)
    print("   independent product a x b = {:.4f}".format(A * B))
    print("    rho   P(both)  P(exactly one)  P(W | E)  lift")
    for rho in (0.0, 0.3, 0.6):
        both, one, cond = joint(rho, rng)
        print("   {:4.1f} {:9.4f} {:15.4f} {:9.4f} {:5.2f}".format(
            rho, both, one, cond, both / (A * B)))
    print("\nReading: 'double resonance' is the rarer event only because")
    print("E is rare; when E fires, W agrees a third of the time even if")
    print("the two are independent. Divergence, not agreement, is the")
    print("normal state, and it carries no information either.")
#    independent product a x b = 0.0292
#     rho   P(both)  P(exactly one)  P(W | E)  lift
#     0.0    0.0294          0.3751    0.3504  1.01
#     0.3    0.0469          0.3403    0.5673  1.61
#     0.6    0.0669          0.2999    0.8064  2.29
#
# Reading: 'double resonance' is the rarer event only because
# E is rare; when E fires, W agrees a third of the time even if
# the two are independent. Divergence, not agreement, is the
# normal state, and it carries no information either.

两者独立时,同时触发的概率是 0.0294,等于两个基率之积;「恰有一方触发」(正交分歧)的概率是 0.3751,远比共振常见。值得注意的是条件概率:即使两套系统完全独立,在东方触发的年份,西方也有 0.3504 的概率同时触发。相关系数为 0.3 时,这一条件概率升到 0.5673,共振的概率是独立时的 1.61 倍。「共振」看起来稀有,只是因为东方的基率低;一旦东方触发,西方的附和非常常见。

一生中的偶然一致

把时间拉长到 40 个成年年份:东方系统每年标出一个领域,西方系统每年标出 kk 个领域,两者毫无关联。

推导

# Chapter 6. [p13] How often do two unrelated systems "agree" over a
# life? Over 40 adult years, an eastern system marks one life domain
# of 12 each year (the palace an annual ji flies into); a western
# system marks k domains each year (houses touched by major solar-arc
# contacts). With no connection between them, the number of years in
# which the eastern domain is among the western ones is binomial with
# p = k / 12. Also: if the person names one year with a major event,
# the chance that both systems agree that year. Exact binomial sums.
from math import comb

YEARS = 40

def at_least(n, k, p):
    return sum(comb(n, j) * p ** j * (1 - p) ** (n - j)
               for j in range(k, n + 1))

if __name__ == "__main__":
    print("    k   p = k/12   mean agreeing years   P(>= 8 years)"
          "   P(>= 12 years)")
    for k in (1, 2, 3, 4, 6):
        p = k / 12
        print("   {:2d} {:9.4f} {:20.2f} {:15.4f} {:16.4f}".format(
            k, p, YEARS * p, at_least(YEARS, 8, p),
            at_least(YEARS, 12, p)))
    print("   event year agreement by chance, k = 3: {:.4f}".format(3 / 12))
    print("\nReading: with a western system marking a quarter of all")
    print("domains each year, the two agree in about ten years of forty")
    print("by chance alone. A list of 'resonance years' that matched")
    print("life events would need to beat this baseline to mean anything.")
#     k   p = k/12   mean agreeing years   P(>= 8 years)   P(>= 12 years)
#     1    0.0833                 3.33          0.0159           0.0001
#     2    0.1667                 6.67          0.3466           0.0261
#     3    0.2500                10.00          0.8180           0.2849
#     4    0.3333                13.33          0.9789           0.7265
#     6    0.5000                20.00          1.0000           0.9968
#    event year agreement by chance, k = 3: 0.2500
#
# Reading: with a western system marking a quarter of all
# domains each year, the two agree in about ten years of forty
# by chance alone. A list of 'resonance years' that matched
# life events would need to beat this baseline to mean anything.

西方每年标出 3 个领域(四分之一)时,两套系统平均有 10.00 年「一致」,至少 8 年一致的概率是 0.8180,至少 12 年的概率是 0.2849。若当事人指出一个重大事件的年份,两套系统在那一年偶然一致的概率是 0.2500。任何一份把「共振年份」与人生事件对照的报告,都必须以这一基率为参照。

假说 · 尚未验证

作者的诠释(尚未验证): 多系统集成的价值上限由系统之间的相关决定;在相关未经测量的情况下,「几套系统都这么说」不应当被当作比单套系统更强的证据。「共振年份」的命中率,必须与本章的偶然一致基率比较。什么会证明这一判断错:若在第八章的预注册检验中,共振评分最高的年份与事先封存的重大事件年份的关联显著高于偶然,说明集成确实提供了单套系统之外的信息。

小结

孔多塞定理的增益依赖独立:相关系数为 0.6 时,9 套系统几乎不比 1 套更好。「双系统共振」在独立假设下的概率就是基率之积,而一旦稀有的一方触发,另一方的附和很常见。一生 40 年中,两套毫无关联的系统平均有 10.00 年偶然一致。

参考文献

Condorcet, Marquis de. 1785. Essai Sur l’application de l’analyse à La Probabilité Des décisions Rendues à La Pluralité Des Voix. Imprimerie Royale.
Ladha, Krishna K. 1992. “The Condorcet Jury Theorem, Free Speech, and Correlated Votes.” American Journal of Political Science 36 (3): 617–34.
Nelsen, Roger B. 2006. An Introduction to Copulas. 2nd ed. Springer.

Fan, Y. (2026). Multimodal Calendrical Tensor Fusion, Modulo-90 Harmonic Dials and Cross-System Resonance Diagnostics: Climate-Balance Vectors, Flying-Transformation Matrices and Solar-Arc Midpoint Trees [多模态历法张量融合、模九十谐波拨盘与跨系统共振诊断] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090189 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).