关于「真源混合阈值:固定比例保住了多少方差,又漏掉了多少尾部」,《抗模型崩溃双锚工程》给出的定量结论是什么?

结论:第四章算出真源混合的不动点:同样每代混入 0.01 的真源,N 从 100 增到 10000,期望方差从 0.6644 升到 0.9950;尾部覆盖只认个数,每代 1000 个真样本时不论比例如何都只覆盖 0.0256 的尾部类别。

数据口径:书中参数为示意;全部计算由书中程序生成,不调用任何模型接口,不使用任何读者或客户的数据;本书是非个人化的经营与工程参考。 书中以标注框区分已发表的证据、可运行的 Python 计算(22 段程序)、原典与作者的假说。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 与结论吻合的打印值: 0.01, 0.6644, 0.9950, 100, 1000, 10000

N = 100; V = expected fitted variance (truth 1)
sharefresh_mV(n=10)V(n=100)V(n=1000)V*(1-V*)*m
0.0000.90440.36600.00000.00000.0000
0.0110.91240.68070.66440.66440.3356
0.0220.91950.79880.79760.79760.4049
0.0550.93670.90650.90650.90650.4673
0.10100.95520.94970.94970.94970.5025
0.20200.97310.97280.97280.97280.5435
0.50500.98680.98680.98680.98680.6623
N = 100share 0.01: V* = 0.6644
N = 1000share 0.01: V* = 0.9521
N = 10000share 0.01: V* = 0.9950
查看完整验算程序
# Chapter 4. p41: exact expected variance with a fixed fresh share a.
# Model: truth N(0, 1); each generation trains on N points, m = a*N of
# them fresh from the truth and N - m synthetic from N(mu_n, v_n); the
# fit is maximum likelihood. With V = E[v_n] and M = E[mu_n^2]:
#   V' = (1 - 1/N)(a + (1 - a) V) + a (1 - a) M
#   M' = (1 - a)^2 M + (a + (1 - a) V) / N
# This linear pair is exact. Its fixed point is V* = a(A + b) /
# (1 - (1 - a)(A + b)) with A = 1 - 1/N and b = a(1 - a) /
# (N (1 - (1 - a)^2)). N and the shares are illustrative.

N = 100
SHARES = (0.0, 0.01, 0.02, 0.05, 0.1, 0.2, 0.5)


def step(v, m, a):
    v2 = (1.0 - 1.0 / N) * (a + (1.0 - a) * v) + a * (1.0 - a) * m
    m2 = (1.0 - a) ** 2 * m + (a + (1.0 - a) * v) / N
    return v2, m2


def fixed(a):
    if a == 0.0:
        return 0.0
    big_a = 1.0 - 1.0 / N
    b = a * (1.0 - a) / (N * (1.0 - (1.0 - a) ** 2))
    return a * (big_a + b) / (1.0 - (1.0 - a) * (big_a + b))


print("N = %d; V = expected fitted variance (truth 1)" % N)
print("share  fresh_m  V(n=10)  V(n=100)  V(n=1000)  V*      (1-V*)*m")
for a in SHARES:
    v, m = 1.0, 0.0
    got = {}
    for gen in range(1, 1001):
        v, m = step(v, m, a)
        if gen in (10, 100, 1000):
            got[gen] = v
    vs = fixed(a)
    fresh = round(a * N)
    print("%-6.2f %-8d %-8.4f %-9.4f %-10.4f %-7.4f %.4f" % (
        a, fresh, got[10], got[100], got[1000], vs, (1.0 - vs) * fresh))
for big_n in (100, 1000, 10000):
    N = big_n
    print("N = %-6d share 0.01: V* = %.4f" % (big_n, fixed(0.01)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《抗模型崩溃双锚真源工程、自产语料回喂方差坍缩速率与标准库验算校勘免疫母典》第 4 章「真源混合阈值:固定比例保住了多少方差,又漏掉了多少尾部」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-265 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026265,
  author    = {Fan, Yafu},
  title     = {Anti-Collapse Engineering with Two Anchors of Truth: the Variance-Collapse Rate of Self-Generated Training Data and Immunity through Standard-Library Recomputation and Textual Collation},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论