关于「卡方检验:临界值 15.507、九次 z 检验的虚警与大样本下的必然拒绝」,《本福特定律与流水报销审计》给出的定量结论是什么?
结论:第三章说明九次逐个数字的 z 检验让一千笔诚实记录有 0.3242 的机会被标红,几乎服从定律的数据在一万笔上被卡方拒绝的概率是 0.999。
数据口径:书中参数为示意;全部计算由书中程序生成;本书不是法律意见,也不是个人化经营建议,检验结果只是复核信号。 书中以标注框区分已发表的证据、可运行的 Python 计算(20 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: 0.3242
| chi-square df 8: 5% critical 15.507, 1% critical 20.090 | |||
| check: P(chi2 > 15.507) = 0.0500 | |||
| N | min E cnt | chi2 size | any-z size |
| 30 | 1.37 | 0.0510 | 0.1918 |
| 50 | 2.29 | 0.0508 | 0.2067 |
| 100 | 4.58 | 0.0490 | 0.2525 |
| 300 | 13.73 | 0.0515 | 0.2875 |
| 1000 | 45.76 | 0.0512 | 0.3242 |
| independent-test bound 1 - 0.95^9 = 0.3698 |
查看完整验算程序
# Chapter 3. p31: chi-square critical values and the real false-alarm rate.
# Exact: the chi-square CDF with 8 degrees of freedom through the
# regularized lower incomplete gamma function (series and continued
# fraction), critical values by bisection. Monte Carlo (fixed seed):
# samples drawn from Benford itself, N first digits each, 4000 samples
# per N; we count how often (a) the chi-square test at 5 percent and
# (b) at least one of nine per-digit z tests at 5 percent cries fraud.
# z uses the continuity correction 1/(2N) of the auditing literature.
import math
import random
P1 = [math.log10(1 + 1 / d) for d in range(1, 10)]
CUM = [math.fsum(P1[:i + 1]) for i in range(9)]
rng = random.Random(20261112)
def gamma_p(a, x):
if x <= 0:
return 0.0
lead = math.exp(-x + a * math.log(x) - math.lgamma(a))
if x < a + 1:
terms, t, k = [1.0 / a], 1.0 / a, 0
while t > 1e-17:
k += 1
t *= x / (a + k)
terms.append(t)
return lead * math.fsum(terms)
b, c, d = x + 1 - a, 1e300, 1 / (x + 1 - a)
h = d
for i in range(1, 400):
an = -i * (i - a)
b += 2
d = an * d + b
d = 1 / (d if abs(d) > 1e-300 else 1e-300)
c = b + an / c
c = c if abs(c) > 1e-300 else 1e-300
h *= d * c
if abs(d * c - 1) < 1e-16:
break
return 1 - lead * h
def chi2_cdf(x, df):
return gamma_p(df / 2, x / 2)
def chi2_crit(alpha, df):
lo, hi = 0.0, 200.0
for _ in range(200):
mid = (lo + hi) / 2
if 1 - chi2_cdf(mid, df) > alpha:
lo = mid
else:
hi = mid
return (lo + hi) / 2
c05, c01 = chi2_crit(0.05, 8), chi2_crit(0.01, 8)
print("chi-square df 8: 5%% critical %.3f, 1%% critical %.3f" % (c05, c01))
print("check: P(chi2 > %.3f) = %.4f" % (c05, 1 - chi2_cdf(c05, 8)))
print("%6s %10s %10s %12s" % ("N", "min E cnt", "chi2 size", "any-z size"))
for n in (30, 50, 100, 300, 1000):
hit_c = hit_z = 0
for _ in range(4000):
cnt = [0] * 9
for d in rng.choices(range(9), cum_weights=CUM, k=n):
cnt[d] += 1
chi = math.fsum((cnt[i] - n * P1[i]) ** 2 / (n * P1[i])
for i in range(9))
hit_c += chi > c05
for i in range(9):
se = math.sqrt(P1[i] * (1 - P1[i]) / n)
z = (abs(cnt[i] / n - P1[i]) - 1 / (2 * n)) / se
if z > 1.96:
hit_z += 1
break
print("%6d %10.2f %10.4f %12.4f"
% (n, n * P1[8], hit_c / 4000, hit_z / 4000))
print("independent-test bound 1 - 0.95^9 = %.4f" % (1 - 0.95 ** 9))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《本福特首位数定律、账目舞弊检测与代购流水报销卡方审计母典》第 3 章「卡方检验:临界值 15.507、九次 z 检验的虚警与大样本下的必然拒绝」(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-216 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026216,
author = {Fan, Yafu},
title = {Benford's First-Digit Law, Ledger Fraud Detection and Chi-Square Audits of Buying-Agent Cash Flows and Expense Claims},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}