关于「方差逐代坍缩:高斯回喂链的期望公式与几乎必然归零」,《抗模型崩溃双锚工程》给出的定量结论是什么?

结论:第一、二章推导高斯回喂链与离散吸收链:每代 100 个样本时第 100 代期望方差只剩 0.3660,典型值约以两倍速度坍缩;一个类别最终丢失的概率等于 1 减去初始份额。

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。 程序只用 Python 标准库与结论吻合的打印值: 0.3660, 100

E[v_n] under the MLE refit, true variance 1
N n1 n=10 n=100 n=1000 half-life
100.9000 0.3487 0.0000 0.0000 6.6
1000.9900 0.9044 0.3660 0.0000 69.0
10000.9990 0.9900 0.9048 0.3677 692.8
100000.9999 0.9990 0.9900 0.9048 6931.1
unbiased refit: E[v_n]1.0000 for every N and n
half-life / N at N10000: 0.6931 (ln 2 = 0.6931)
E[v_n] after n = N generations, N1000: 0.3677 (1/e = 0.3679)
查看完整验算程序
# Chapter 1. p11: expected variance after n refits, sigma^2 (1 - 1/N)^n.
# Model: generation 0 is N draws from a normal law with variance 1.
# Each generation fits mean and variance by maximum likelihood (the
# sum of squares divided by N) and draws N new points from the fit.
# Then E[v_n] = (1 - 1/N)^n exactly. With the unbiased estimator (sum
# of squares divided by N - 1) the expectation stays at 1 for ever.
# The sample sizes N and generation counts n are illustrative.
import math

NS = (10, 100, 1000, 10000)
GENS = (1, 10, 100, 1000)


def expect_mle(n_samp, gen):
    return (1.0 - 1.0 / n_samp) ** gen


def half_life(n_samp):
    return math.log(0.5) / math.log(1.0 - 1.0 / n_samp)


print("E[v_n] under the MLE refit, true variance 1")
print("N       " + "".join("n=%-8d" % g for g in GENS) + "half-life")
for n_samp in NS:
    cells = "".join("%-10.4f" % expect_mle(n_samp, g) for g in GENS)
    print("%-7d %s%.1f" % (n_samp, cells, half_life(n_samp)))
print("unbiased refit: E[v_n] = %.4f for every N and n" % 1.0)
ratio = half_life(10000) / 10000
print("half-life / N at N=10000: %.4f (ln 2 = %.4f)" % (ratio,
                                                      math.log(2.0)))
print("E[v_n] after n = N generations, N=1000: %.4f (1/e = %.4f)"
      % (expect_mle(1000, 1000), math.exp(-1.0)))

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 1 章「方差逐代坍缩:高斯回喂链的期望公式与几乎必然归零」(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