关于「自产回喂的方差坍缩:尾部为什么先消失」,《人机互哺与证伪棘轮》给出的定量结论是什么?

结论:自产回喂让期望方差每代乘以 (n-1)/n,典型路径缩得更快,中位半衰期约为期望半衰期的一半。尾部比方差先走:方差缩两成,千分之一尾部就丢掉将近八成。每代掺进真源,方差就有稳态;不扔掉旧的真实数据,误差就有上界。坍缩是条件命题,不是宿命。

数据口径:书中参数为示意,全部计算由书中程序生成。 本书不是软件工程意见,也不是投资意见,不对任何具体商业模型的内部机制下断言。

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

Refit-own-samples recursion, v01
n shrink c mean half-life log drift median half-life
100.9000 6.6 -0.22057 3.1
200.9500 13.5 -0.10485 6.6
500.9800 34.3 -0.04075 17.0
1000.9900 69.0 -0.02019 34.3
10000.9990 692.8 -0.00200 346.3
Simulation: n10, 30 generations, 2,000 paths
exact mean variance0.0424
simulated mean0.0336
simulated median0.0013
share of paths v < 0.10.952
查看完整验算程序
# Chapter 1. p11: variance collapse when a model refits its own samples.
# Each generation draws n points from the previous fitted normal and
# refits mean and variance by maximum likelihood (divide by n).
# Exact: E[v_t] = c^t v_0 with c = (n - 1) / n.
# Typical path: ln v_t is a random walk with drift
#   d = E ln(chi2_{n-1} / n) = psi((n - 1) / 2) + ln 2 - ln n < 0,
# so the median variance halves after ln 2 / |d| generations, far
# sooner than the mean. A fixed-seed simulation checks n = 10.
# All parameters are illustrative.
import math
import random
import statistics


def digamma(x):
    acc = []
    while x < 8.0:
        acc.append(-1.0 / x)
        x += 1.0
    inv = 1.0 / x
    inv2 = inv * inv
    tail = (math.log(x) - 0.5 * inv - inv2 / 12.0 + inv2 * inv2 / 120.0
            - inv2 ** 3 / 252.0)
    acc.append(tail)
    return math.fsum(acc)


def drift(n):
    return digamma((n - 1) / 2.0) + math.log(2.0) - math.log(n)


def simulate(rng, n, gens, reps):
    finals = []
    for _ in range(reps):
        mu, sd = 0.0, 1.0
        for _ in range(gens):
            xs = [rng.gauss(mu, sd) for _ in range(n)]
            mu = statistics.fmean(xs)
            sd = math.sqrt(statistics.pvariance(xs, mu))
        finals.append(sd * sd)
    return finals


if __name__ == "__main__":
    print("Refit-own-samples recursion, v0 = 1")
    print("   n  shrink c  mean half-life  log drift  median half-life")
    for n in (10, 20, 50, 100, 1000):
        c = (n - 1) / n
        mean_half = math.log(2.0) / -math.log(c)
        d = drift(n)
        med_half = math.log(2.0) / -d
        print("{:>4}  {:>8.4f}  {:>14.1f}  {:>9.5f}  {:>16.1f}".format(
            n, c, mean_half, d, med_half))
    rng = random.Random(37411)
    gens = 30
    finals = simulate(rng, 10, gens, 2000)
    print("Simulation: n = 10, {} generations, 2,000 paths".format(gens))
    print("  exact mean variance   {:.4f}".format(0.9 ** gens))
    print("  simulated mean        {:.4f}".format(statistics.fmean(finals)))
    print("  simulated median      {:.4f}".format(
        statistics.median(finals)))
    low = 0
    for v in finals:
        if v < 0.1:
            low += 1
    print("  share of paths v < 0.1  {:.3f}".format(low / len(finals)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《人机互哺双螺旋进化、实证门禁稳健棘轮与共生创作复利母典》第 1 章「自产回喂的方差坍缩:尾部为什么先消失」(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-374 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026374,
  author    = {Fan, Yafu},
  title     = {Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论