关于「尾部先死:离散分布的逐代灭绝与 Wright–Fisher 吸收态」,《抗模型崩溃双锚工程》给出的定量结论是什么?
结论:程序 p21 说明,一代有限抽样就会删掉一大片尾部:每代 1000 个样本时,1000 类的齐普夫名单期望消失 666.3 类,尾部自身的质量丢掉 0.8254,头部几乎无损。程序 p22 用精确的转移矩阵说明,单个类别最终丢失的概率等于 1 减去初始份额,回喂链的终点是单点分布。程序 p23 展示了整张名单从尾部先死到头部互吞的两阶段衰亡。既然坍缩源于每一代都把上一代的产出当作全部训练数据,一个自然的问题是:如果不丢旧数据,而是把每一代的产出累积起来呢?下一章检验 Gerstgrasser 等人提出的这个反方证据。
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: 0.8254, 1000, 666.3
| K | 1000, harmonic H = 7.4855, p_1 = 0.13359, p_K = 0.0001336 |
| N lost_cats lost_mass head_lost tail_lost tail_mass | |
| 1000 | 666.3 0.2091 0.000002 415.8 0.8254 |
| 3000 | 389.1 0.0937 0.000000 288.7 0.5648 |
| 10000 | 91.8 0.0171 0.000000 83.8 0.1564 |
| 100000 | 0.0 0.0000 0.000000 0.0 0.0000 |
| tail share of mass (ranks 501-1000) | 0.0925 |
| N | 1000: expected count drops below 1 at rank 134 |
| N | 3000: expected count drops below 1 at rank 401 |
查看完整验算程序
# Chapter 2. p21: what one resampling generation deletes from a tail.
# Model: a Zipf law over K = 1000 categories, p_i proportional to 1/i
# (an illustrative stand-in for a vocabulary or a set of rare facts).
# One generation draws N items; category i is absent with probability
# (1 - p_i)^N, exactly. We report the expected number of categories
# that vanish, the probability mass they carry, and the same split
# into the head (ranks 1-10) and the tail (ranks 501-1000).
import math
K = 1000
H = math.fsum(1.0 / i for i in range(1, K + 1))
P = [1.0 / (i * H) for i in range(1, K + 1)]
def lost(n_draw, lo, hi):
ps = P[lo - 1:hi]
cats = math.fsum((1.0 - p) ** n_draw for p in ps)
mass = math.fsum(p * (1.0 - p) ** n_draw for p in ps)
return cats, mass
print("K = %d, harmonic H = %.4f, p_1 = %.5f, p_K = %.7f"
% (K, H, P[0], P[-1]))
print("N lost_cats lost_mass head_lost tail_lost tail_mass")
tail_total = math.fsum(P[500:])
for n_draw in (1000, 3000, 10000, 100000):
cats, mass = lost(n_draw, 1, K)
head, _ = lost(n_draw, 1, 10)
tail, tmass = lost(n_draw, 501, K)
print("%-8d %-10.1f %-10.4f %-10.6f %-10.1f %.4f" % (
n_draw, cats, mass, head, tail, tmass / tail_total))
print("tail share of mass (ranks 501-1000): %.4f" % tail_total)
for n_draw in (1000, 3000):
rank = min(i for i in range(1, K + 1) if n_draw * P[i - 1] < 1.0)
print("N = %d: expected count drops below 1 at rank %d"
% (n_draw, rank))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 2 章「尾部先死:离散分布的逐代灭绝与 Wright–Fisher 吸收态」(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}
}