关于「替换还是累积:Gerstgrasser 反证与 π²/6 上界」,《抗模型崩溃双锚工程》给出的定量结论是什么?
结论:第三章并陈 Gerstgrasser 等人的反方证据,累积而非替换真数据时,高斯孪生版的均值误差被 π²/6 即 1.644934 封顶。
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。
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: 1.644934
| n | replace | accumulate | ratio |
|---|---|---|---|
| 0 | 1.0000 | 1.000000 | 1.0 |
| 1 | 2.0000 | 1.250000 | 1.6 |
| 10 | 11.0000 | 1.558032 | 7.1 |
| 100 | 101.0000 | 1.635082 | 61.8 |
| 1000 | 1001.0000 | 1.643936 | 608.9 |
| 10000 | 10001.0000 | 1.644834 | 6080.2 |
| accumulate limit pi^2/6 = 1.644934 | |||
| accumulate share of the limit at n = 10: 0.9472 | |||
| pool size under accumulate after n = 100: 101 x N |
查看完整验算程序
# Chapter 3. p31: replace versus accumulate, exact mean-squared drift.
# Model: a normal law with known variance s^2 = 1 and unknown mean;
# each generation adds N points drawn from the current fit. Replace:
# the fit uses only the newest N points, so the error e_n of the mean
# is a random walk and E[e_n^2] = (n + 1) / N. Accumulate: the fit
# uses every point drawn so far; then e_n = e_{n-1} + eps_n / (n + 1)
# and E[e_n^2] = (1/N) * sum_{k=1}^{n+1} 1/k^2 < (pi^2/6) / N. This is
# the Gaussian twin of the linear-regression bound of Gerstgrasser et
# al. (2024). Values are N * E[e_n^2], free of N.
import math
GENS = (0, 1, 10, 100, 1000, 10000)
def replace(n):
return float(n + 1)
def accumulate(n):
return math.fsum(1.0 / (k * k) for k in range(1, n + 2))
print("n replace accumulate ratio")
for n in GENS:
r, a = replace(n), accumulate(n)
print("%-7d %-11.4f %-11.6f %.1f" % (n, r, a, r / a))
print("accumulate limit pi^2/6 = %.6f" % (math.pi ** 2 / 6.0))
print("accumulate share of the limit at n = 10: %.4f"
% (accumulate(10) / (math.pi ** 2 / 6.0)))
print("pool size under accumulate after n = 100: %d x N" % 101)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 3 章「替换还是累积:Gerstgrasser 反证与 π²/6 上界」(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}
}