关于「一行递推:累积和就是在线的对数似然比」,《累积和告警》给出的定量结论是什么?

结论:双侧要守住同样的误报预算,告警限必须从 4.096 提高到 4.775,因为现在有两条累积和都可能误报。代价落在检测速度上:漂移半个标准差时,平均运行长度从 27.7 天增加到 35.3 天,多等 7.6 天;漂移四分之一个标准差时,从 81.9 天增加到 121.6 天,多等 39.7 天;漂移越大,代价越小,到一个标准差只多 1.4 天。

数据口径:书中参数为示意;全部计算由书中程序生成;本书属于非个人化经营建议,第八章只做离线模型,不访问任何服务器。 书中以标注框区分已发表的证据、可运行的 Python 计算(23 段程序)、原典与作者的假说。

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 标准库与结论吻合的打印值: 1.4, 121.6, 27.7, 35.3, 39.7, 4.096

one-sided h = 4.096, ARL0 = 370.0
two-sided h = 4.775, ARL0 = 370.0
shiftone-sidedtwo-sidedextra days
0.2581.9121.639.7
0.5027.735.37.6
1.008.69.91.4
2.003.43.90.5
查看完整验算程序
# Chapter 2. p23: watching both directions costs delay in the one that
# matters.
# Model: an upper and a lower CUSUM run side by side (k = 0.5). With
# ARL+ and ARL- from the Brook and Evans chain, the two-sided ARL is
# taken from 1 / ARL = 1 / ARL+ + 1 / ARL- (exact enough when k is not
# small). The two-sided h is found so that ARL0 = 370, and compared with
# the one-sided h of p12 for upward shifts (illustrative).
import math


def cdf(x):
    return 0.5 * math.erfc(-x / math.sqrt(2.0))


def upper(x):
    return 0.5 * math.erfc(x / math.sqrt(2.0))


def solve(a, b):
    # Gaussian elimination with partial pivoting; back substitution
    # adds its terms with math.fsum
    n = len(b)
    m = [list(row) + [b[i]] for i, row in enumerate(a)]
    for p in range(n):
        q = max(range(p, n), key=lambda r: abs(m[r][p]))
        m[p], m[q] = m[q], m[p]
        piv = m[p]
        for r in range(p + 1, n):
            f = m[r][p] / piv[p]
            if f != 0.0:
                m[r] = m[r][:p] + [x - f * y for x, y in
                                   zip(m[r][p:], piv[p:])]
    x = [0.0] * n
    for i in range(n - 1, -1, -1):
        s = math.fsum(m[i][j] * x[j] for j in range(i + 1, n))
        x[i] = (m[i][n] - s) / m[i][i]
    return x


def arl(k, h, delta, m=80):
    w = 2.0 * h / (2 * m - 1)
    a = []
    for i in range(m):
        c = i * w - k + delta
        row = [-cdf(w / 2.0 - c)]
        for j in range(1, m):
            row.append(cdf(j * w - w / 2.0 - c)
                       - cdf(j * w + w / 2.0 - c))
        row[i] = row[i] + 1.0
        a.append(row)
    return solve(a, [1.0] * m)[0]


def limit_for(arl0):
    lo, hi = 0.0, 10.0
    for _ in range(200):
        mid = 0.5 * (lo + hi)
        if upper(mid) > 1.0 / arl0:
            lo = mid
        else:
            hi = mid
    return 0.5 * (lo + hi)


def h_for(k, target, m=80):
    lo, hi = 0.5, 12.0
    for _ in range(40):
        mid = 0.5 * (lo + hi)
        if arl(k, mid, 0.0, m) < target:
            lo = mid
        else:
            hi = mid
    return 0.5 * (lo + hi)


def two_sided(k, h, d):
    up, down = arl(k, h, d), arl(k, h, -d)
    return 1.0 / (1.0 / up + 1.0 / down)


if __name__ == "__main__":
    k = 0.5
    h1 = h_for(k, 370.0)
    h2 = h_for(k, 740.0)
    print("one-sided h = {:.3f}, ARL0 = {:.1f}".format(h1, arl(k, h1, 0.0)))
    print("two-sided h = {:.3f}, ARL0 = {:.1f}".format(
        h2, two_sided(k, h2, 0.0)))
    print(" shift  one-sided  two-sided  extra days")
    for d in (0.25, 0.5, 1.0, 2.0):
        a1 = arl(k, h1, d)
        a2 = two_sided(k, h2, d)
        print(" {:5.2f}  {:9.1f}  {:9.1f}  {:10.1f}".format(
            d, a1, a2, a2 - a1))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《累积和控制图与平均运行长度、延迟漂移与转化暗降的最快无误报告警母典》第 2 章「一行递推:累积和就是在线的对数似然比」(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-305 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026305,
  author    = {Fan, Yafu},
  title     = {CUSUM Control Charts and Average Run Length: the Fastest Alarm within a False-Alarm Budget for Latency Drift and Silent Conversion Drops},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论