关于「加入第二矩:截断正态族、凸指数与一条标准差能买到多少信息」,《最大熵需求重建》给出的定量结论是什么?

结论:标准差 60 与 75 时,最大熵答案变成凸形,第 90 百分位分别推到 166.1 与 236.8 元,中位数却降到 57.0 与 47.9 元:两极分化让中间变空。不截断的正态在这里完全失效,它的第 10 百分位分别是 3.1 元和 \(-16.1\) 元,越界份额 0.202 与 0.252,上尾第 90 百分位 156.9 与 176.1 元,都远低于最大熵答案。熵在标准差 75 时下降了 0.497 比特,是表中最多的,因为一个远大于隐含值的标准差是一条很「意外」的信息。

数据口径:本书推翻这一常识:正态捷径私加了对称形状与标准差,示意参数下均值 80、下限 30 时它把 0.133 的概率放在不可能的区间,均值 40 时升到 0.412;均匀捷径隐含均值在区间中点,直接违背已知均值。 只承认上下限与均值时,杰恩斯 1957 年提出的最大熵原则给出截断指数分布,示意甜品店的第 90 百分位客单为 146.0 元。

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: -16.1, 0.202, 0.252, 0.497, 10, 156.9

mean only: P10 35.4 P50 65.4 P90 146.0 H 7.080
sd max-ent P10 P50 P90 bits gained | normal P10 P90 <30
3042.477.0 121.00.235 |41.6 118.4 0.048
4037.170.7 136.30.036 |28.7 131.3 0.106
6033.757.0 166.10.094 |3.1 156.9 0.202
7532.447.9 236.80.497 |-16.1 176.1 0.252
查看完整验算程序
# Chapter 4. p42: what a stated spread buys, in bits and in quantiles.
# Model: as p41 (bounds [30, 300], mean 80, illustrative). For each
# stated standard deviation we print the two-moment maximum-entropy
# quantiles, the entropy it gives up against the mean-only answer
# (the information the spread adds), and the quantiles an uncut
# normal(80, sd) would claim, with the share it puts below 30.
import math
from statistics import NormalDist

A, B, M = 30.0, 300.0, 80.0
L = B - A
NODES = 2000


def grid():
    h = 1.0 / NODES
    us = [i * h for i in range(NODES + 1)]
    cs = [1.0 if i in (0, NODES) else (4.0 if i % 2 else 2.0)
          for i in range(NODES + 1)]
    return us, cs


US, CS = grid()
UP = [[u ** k for u in US] for k in range(1, 5)]


def moments(l1, l2, kmax=4):
    # Simpson weights times exp(-(l1 u + l2 u^2)) on u in [0, 1]
    ex = [-(l1 * u + l2 * u * u) for u in US]
    top = max(ex)
    w = [c * math.exp(e - top) for c, e in zip(CS, ex)]
    z = math.fsum(w)
    mom = [math.fsum(wi * uk for wi, uk in zip(w, UP[k - 1])) / z
           for k in range(1, kmax + 1)]
    logz = math.log(z / (3.0 * NODES)) + top
    return mom, logz


def solve2(mean, sd, l1=0.0, l2=0.0):
    # Newton on (l1, l2) so that E[x] = mean and Var[x] = sd^2
    t1 = (mean - A) / L
    t2 = t1 * t1 + (sd / L) ** 2
    for _ in range(60):
        (m1, m2, m3, m4), _ = moments(l1, l2)
        g1, g2 = m1 - t1, m2 - t2
        if abs(g1) + abs(g2) < 1e-13:
            break
        c11 = m2 - m1 * m1
        c12 = m3 - m1 * m2
        c22 = m4 - m2 * m2
        det = c11 * c22 - c12 * c12
        d1 = (c22 * g1 - c12 * g2) / det
        d2 = (c11 * g2 - c12 * g1) / det
        step = 1.0
        while step > 1e-4:
            n1, n2 = l1 + step * d1, l2 + step * d2
            (k1, k2, _, _), _ = moments(n1, n2)
            if abs(k1 - t1) + abs(k2 - t2) < abs(g1) + abs(g2):
                break
            step /= 2.0
        l1, l2 = n1, n2
    return l1, l2


def entropy_bits(l1, l2):
    (m1, m2, _, _), logz = moments(l1, l2)
    return (logz + l1 * m1 + l2 * m2 + math.log(L)) / math.log(2.0)


def part(l1, l2, hi, n=400):
    # integral of exp(-(l1 u + l2 u^2)) over [0, hi], Simpson rule
    h = hi / n
    f = lambda u: math.exp(-(l1 * u + l2 * u * u))
    terms = [f(0.0), f(hi)]
    terms += [4.0 * f(i * h) for i in range(1, n, 2)]
    terms += [2.0 * f(i * h) for i in range(2, n, 2)]
    return math.fsum(terms) * h / 3.0


def quantile2(l1, l2, p):
    z = part(l1, l2, 1.0, NODES)
    lo, hi = 0.0, 1.0
    for _ in range(40):
        mid = (lo + hi) / 2.0
        if part(l1, l2, mid) / z < p:
            lo = mid
        else:
            hi = mid
    return A + L * (lo + hi) / 2.0


def solve1():
    lo, hi = -50.0, 50.0
    for _ in range(80):
        mid = (lo + hi) / 2.0
        (m1, _, _, _), _ = moments(mid, 0.0)
        lo, hi = (mid, hi) if m1 > (M - A) / L else (lo, mid)
    return (lo + hi) / 2.0


if __name__ == "__main__":
    l0 = solve1()
    h0 = entropy_bits(l0, 0.0)
    q0 = [quantile2(l0, 0.0, p) for p in (0.1, 0.5, 0.9)]
    print("mean only:  P10 {:5.1f}  P50 {:5.1f}  P90 {:5.1f}  H {:.3f}"
          .format(q0[0], q0[1], q0[2], h0))
    print("    sd  max-ent P10  P50   P90   bits gained |"
          " normal P10  P90  <30")
    for sd in (30.0, 40.0, 60.0, 75.0):
        l1, l2 = solve2(M, sd)
        q = [quantile2(l1, l2, p) for p in (0.1, 0.5, 0.9)]
        nd = NormalDist(M, sd)
        print("  {:>4.0f}  {:>11.1f} {:>5.1f} {:>5.1f}   {:>11.3f} |"
              " {:>10.1f} {:>5.1f} {:.3f}".format(
                  sd, q[0], q[1], q[2], h0 - entropy_bits(l1, l2),
                  nd.inv_cdf(0.1), nd.inv_cdf(0.9), nd.cdf(A)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《最大熵原理与缺数据需求重建、仅知均价与上下限时最不偏的客单分布推断母典》第 4 章「加入第二矩:截断正态族、凸指数与一条标准差能买到多少信息」(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-301 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026301,
  author    = {Fan, Yafu},
  title     = {The Maximum-Entropy Principle and Demand Reconstruction from Missing Data: the Least-Biased Order-Value Distribution Given Only a Mean and Bounds},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论