关于「去相关策略比较:重采样、异提示、异模型与标准库验算」,《相关误差下的多模型投票》给出的定量结论是什么?
结论:第四章比较去相关策略:同一模型重采样的上限只有 0.7366,五家不同模型的上限是 0.8257,让标准库验算在能覆盖的题上直接拍板,四个模型加验算可达 0.8887。
数据口径:书中参数为示意;全部计算由书中程序生成;本书是非个人化的经营分析,不构成经营建议。 书中以标注框区分已发表的证据、可运行的 Python 计算(19 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: 0.7366, 0.8257, 0.8887
| A. majority accuracy by pool and number of votes | |
| pool p rho n | 1 n=3 n=5 n=9 ceiling |
| resample | 0.72 0.60 0.7200 0.7289 0.7316 0.7337 0.7366 |
| prompts | 0.70 0.35 0.7000 0.7263 0.7362 0.7446 0.7578 |
| models | 0.68 0.15 0.6800 0.7292 0.7533 0.7773 0.8257 |
| B. four different models plus a standard-library check | |
| four models alone, coin on 2-2 | 0.7292 |
| check as one ordinary fifth vote | 0.7969 |
| check decides where it applies | 0.8887 |
| C. decider design by check coverage | |
| coverage 0.2 | 0.7824 |
| coverage 0.4 | 0.8355 |
| coverage 0.6 | 0.8887 |
| coverage 0.8 | 0.9418 |
查看完整验算程序
# Chapter 4. p42: comparing ways to decorrelate a multi-model verdict.
# Model: each pool is exchangeable Beta-binomial (p21) with its own
# mean accuracy p and correlation rho. Illustrative pools:
# resample one model, one prompt, sampled again p 0.72 rho 0.60
# prompts one model, different prompts p 0.70 rho 0.35
# models different models from other makers p 0.68 rho 0.15
# A standard-library check (a deterministic recomputation) applies to
# a share cov = 0.60 of questions, is right with prob 0.995 there and
# is independent of the models. Used as one ordinary vote it joins
# four models; used as a decider it settles the questions it covers.
# Even splits are broken by a coin. Exact sums with math.fsum.
import math
def lbeta(a, b):
return math.lgamma(a) + math.lgamma(b) - math.lgamma(a + b)
def ab(p, rho):
s = 1 / rho - 1
return p * s, (1 - p) * s
def bb_major(n, p, rho):
if rho == 0:
lp, lq = math.log(p), math.log(1 - p)
terms = [math.exp(math.lgamma(n + 1) - math.lgamma(k + 1)
- math.lgamma(n - k + 1) + k * lp
+ (n - k) * lq)
for k in range(n // 2 + 1, n + 1)]
return math.fsum(terms)
a, b = ab(p, rho)
terms = [math.exp(math.lgamma(n + 1) - math.lgamma(k + 1)
- math.lgamma(n - k + 1) + lbeta(k + a, n - k + b)
- lbeta(a, b))
for k in range(n // 2 + 1, n + 1)]
return math.fsum(terms)
def betacf(a, b, x):
tiny = 1e-300
c, d = 1.0, 1.0 - (a + b) * x / (a + 1)
d = 1.0 / (d if abs(d) > tiny else tiny)
h = d
for m in range(1, 400):
m2 = 2 * m
for num in (m * (b - m) * x / ((a + m2 - 1) * (a + m2)),
-(a + m) * (a + b + m) * x / ((a + m2) * (a + m2 + 1))):
d = 1.0 + num * d
d = 1.0 / (d if abs(d) > tiny else tiny)
c = 1.0 + num / c
c = c if abs(c) > tiny else tiny
h *= d * c
if abs(d * c - 1.0) < 1e-15:
break
return h
def betai(a, b, x):
if x <= 0:
return 0.0
if x >= 1:
return 1.0
lead = math.exp(a * math.log(x) + b * math.log(1 - x) - lbeta(a, b))
if x < (a + 1) / (a + b + 2):
return lead * betacf(a, b, x) / a
return 1.0 - lead * betacf(b, a, 1 - x) / b
def limit(p, rho):
if rho == 0:
return 1.0 if p > 0.5 else 0.0
a, b = ab(p, rho)
return 1.0 - betai(a, b, 0.5)
def pmf(n, p, rho):
if rho == 0:
return [math.comb(n, k) * p ** k * (1 - p) ** (n - k)
for k in range(n + 1)]
a, b = ab(p, rho)
return [math.exp(math.lgamma(n + 1) - math.lgamma(k + 1)
- math.lgamma(n - k + 1) + lbeta(k + a, n - k + b)
- lbeta(a, b)) for k in range(n + 1)]
def maj(dist, extra=None):
# extra: None = no extra voter; 1 / 0 = one more voter, right / wrong
n = len(dist) - 1
tot = n if extra is None else n + 1
terms = []
for k, w in enumerate(dist):
right = k if extra is None else k + extra
if 2 * right > tot:
terms.append(w)
elif 2 * right == tot:
terms.append(w / 2)
return math.fsum(terms)
def plain(n, p, rho):
return maj(pmf(n, p, rho), None)
def with_check(p, rho, cov, pc, decider):
d4 = pmf(4, p, rho)
alone = maj(d4, None)
if decider:
return cov * pc + (1 - cov) * alone
joined = pc * maj(d4, 1) + (1 - pc) * maj(d4, 0)
return cov * joined + (1 - cov) * alone
if __name__ == "__main__":
pools = (("resample", 0.72, 0.6), ("prompts", 0.70, 0.35),
("models", 0.68, 0.15))
print("A. majority accuracy by pool and number of votes")
print(" pool p rho n=1 n=3 n=5 n=9 ceiling")
for name, p, rho in pools:
cells = " ".join("{:.4f}".format(plain(n, p, rho))
for n in (1, 3, 5, 9))
print(" {:<9} {:.2f} {:.2f} {} {:.4f}".format(
name, p, rho, cells, limit(p, rho)))
print("B. four different models plus a standard-library check")
p, rho = 0.68, 0.15
print(" four models alone, coin on 2-2 {:.4f}".format(
plain(4, p, rho)))
print(" check as one ordinary fifth vote {:.4f}".format(
with_check(p, rho, 0.6, 0.995, False)))
print(" check decides where it applies {:.4f}".format(
with_check(p, rho, 0.6, 0.995, True)))
print("C. decider design by check coverage")
for cov in (0.2, 0.4, 0.6, 0.8):
print(" coverage {:.1f}: {:.4f}".format(
cov, with_check(p, rho, cov, 0.995, True)))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《孔多塞陪审团定理与相关误差、多模型投票何时越投越准何时同源共错母典》第 4 章「去相关策略比较:重采样、异提示、异模型与标准库验算」(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-309 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026309,
author = {Fan, Yafu},
title = {Condorcet's Jury Theorem with Correlated Errors: When Multi-Model Voting Gets More Accurate and When Shared Origins Make Everyone Wrong Together},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}