关于「按价值放大采样:重要性采样与置信上界」,《量子启发与猜想证伪》给出的定量结论是什么?

结论:按价值放大采样,经典统计有两种成熟做法。重要性采样换一个对准目标的分布、用权重纠偏:估计正态尾部概率时,平移量取 4.5,单次相对标准差从 542.5 降到 2.3;平移不足或过头,效果都打折。多臂老虎机边抽边学:十个方向两千次检验,ε 贪心的遗憾 77.9,UCB1 为 188.0,平均分配为 300.0。两者都依赖对价值的知识,先验与学习速度才是放大的真正来源。

数据口径:书中参数为示意,全部计算由书中程序生成。 本书不构成投资、学术评审或量子硬件方面的建议。

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: 2.3, 4.5, 542.5

A. P(Z > 4.5) exact3.3977e-06; 100,000 draws per batch
proposal mean batch estimates (x 1e-6) rel.sd theory
mu0.0 0.000 0.000 0.000 0.000 0.000 542.5
mu2.0 3.717 3.518 3.660 3.344 3.617 13.7
mu4.5 3.358 3.351 3.391 3.381 3.435 2.3
mu6.5 3.459 3.286 3.407 3.350 3.287 6.0
B. draws to see one marked item, marked fraction p = exact
classical 1/p294,319; sqrt-law (pi/4)/sqrt(p) = 426
proposal mu2.0: P(draw lands in event) = 0.006
proposal mu4.5: P(draw lands in event) = 0.500
查看完整验算程序
# Chapter 4. p41: importance sampling as classical "amplification".
# Target: the rare-event probability P(Z > 4.5), Z standard normal,
# exact value 0.5 * erfc(4.5 / sqrt 2). Plain Monte Carlo with 100,000
# draws almost never sees the event. Drawing from a shifted normal
# N(mu, 1) and reweighting by exp(-mu x + mu^2 / 2) is unbiased for any
# mu; a well-placed mu concentrates the draws where the value is, a
# badly placed one is worse than doing nothing. Part B compares the draw
# counts with the square-root law for amplitude amplification of a marked
# fraction p (a calculation of query counts only, no hardware). Seeded.
import math
import random
import statistics

A, DRAWS, BATCHES = 4.5, 100000, 5


def q(z):
    return 0.5 * math.erfc(z / math.sqrt(2))


def rel_sd(mu, p):
    # second moment of the weighted indicator: exp(mu^2) Q(A + mu)
    return math.sqrt(math.exp(mu * mu) * q(A + mu) - p * p) / p


def estimate(rng, mu):
    vals = []
    for _ in range(DRAWS):
        x = rng.gauss(mu, 1.0)
        vals.append(math.exp(-mu * x + mu * mu / 2) if x > A else 0.0)
    return statistics.fmean(vals), statistics.pstdev(vals)


if __name__ == "__main__":
    exact = q(A)
    print("A. P(Z > 4.5) exact = {:.4e}; {:,} draws per batch".format(
        exact, DRAWS))
    print("   proposal mean   batch estimates (x 1e-6)     rel.sd theory")
    rng = random.Random(20261120)
    for mu in (0.0, 2.0, 4.5, 6.5):
        ests = [estimate(rng, mu)[0] for _ in range(BATCHES)]
        cells = " ".join("{:6.3f}".format(e * 1e6) for e in ests)
        print("   mu = {:<4}      {}   {:>10.1f}".format(
            mu, cells, rel_sd(mu, exact)))
    print("B. draws to see one marked item, marked fraction p = exact")
    p = exact
    print("   classical 1/p = {:,.0f}; sqrt-law (pi/4)/sqrt(p) = {:,.0f}"
          .format(1 / p, math.pi / 4 / math.sqrt(p)))
    for mu in (2.0, 4.5):
        print("   proposal mu = {}: P(draw lands in event) = {:.3f}".format(
            mu, q(A - mu)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《量子启发式组合寻优、拉马努金式直觉猜想与突破性著作推演母典》第 4 章「按价值放大采样:重要性采样与置信上界」(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-364 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026364,
  author    = {Fan, Yafu},
  title     = {Quantum-Inspired Combinatorial Search, Abductive Intuition and Computationally Verified Breakthrough Monograph Engineering},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论