5D Library › 梦境意象二分图、确定性哈希牌卡与潜意识先验激活 · English

第五章 利特尔伍德定律与选择性记忆

每天一张牌,与一天里的事件

设用户每天早上抽一张「行动指令牌」,比如一扇门的图像,在之后 24 小时里经历 MM 件小事:消息、偶遇、无意中听到的话。每件小事以很小的概率 qq「吻合」牌面的意象,与牌本身无关。一天之内至少吻合一次的概率是 1−(1−q)M1 – (1-q)^M,非常接近泊松形式 1−e−qM1 – e^{-qM}。数学家利特尔伍德对「奇迹」做过同样的计数:如果一个人在清醒的八小时里大约每秒注意到一件事,那么百万分之一的事件大约每个月就会发生在一个普通人身上 (Littlewood 1953);戴森后来把这称为利特尔伍德定律 (Dyson 2004)。迪亚科尼斯与莫斯特勒给出了一般方法:在惊叹巧合之前,先数一数机会有多少 (Diaconis and Mosteller 1989)。

推导

# Chapter 5. p09: Littlewood's law and the daily card. A card is drawn
# in the morning; over the next 24 hours the user meets M small events,
# each of which "fits the card's image" with probability q, independent
# of the card. P(at least one fit) = 1 - (1 - q)**M, close to the
# Poisson form 1 - exp(-qM). A three-card spread gives each event three
# chances. Littlewood's count: one event a second for 8 waking hours.
import math

def p_fit(q, M, cards=1):
    q3 = 1 - (1 - q) ** cards
    return 1 - (1 - q3) ** M

if __name__ == "__main__":
    qs = (0.005, 0.01, 0.02, 0.05)
    print("   P(at least one fit in a day), one card")
    print(("      M  " + "".join("  q={:<6}".format(q) for q in qs))
          .rstrip())
    for M in (20, 50, 100, 200):
        print("   {:4d}  ".format(M) +
              "".join("  {:8.3f}".format(p_fit(q, M)) for q in qs))
    print("   Poisson form at q=0.01, M=50: {:.3f}".format(
        1 - math.exp(-0.5)))
    print("   three-card spread, q=0.01, M=50: {:.3f}".format(
        p_fit(0.01, 50, 3)))
    print("   days in a 30-day month with no fit at q=0.01, M=50: {:.1f}"
          .format(30 * (1 - p_fit(0.01, 50))))
    per_day = 8 * 3600
    print("   Littlewood: {:,} events a day;".format(per_day))
    print("   a one-in-a-million event every {:.1f} days".format(
        1e6 / per_day))
#    P(at least one fit in a day), one card
#       M    q=0.005   q=0.01    q=0.02    q=0.05
#      20       0.095     0.182     0.332     0.642
#      50       0.222     0.395     0.636     0.923
#     100       0.394     0.634     0.867     0.994
#     200       0.633     0.866     0.982     1.000
#    Poisson form at q=0.01, M=50: 0.393
#    three-card spread, q=0.01, M=50: 0.779
#    days in a 30-day month with no fit at q=0.01, M=50: 18.2
#    Littlewood: 28,800 events a day;
#    a one-in-a-million event every 34.7 days

q=0.01q = 0.01、M=50M = 50 时,单张牌在 0.395 的日子里至少吻合一件事(泊松形式 0.393);q=0.02q = 0.02 时是 0.636;200 件小事、q=0.05q = 0.05 时是 1.000。三张牌的牌阵让每件小事有三次吻合机会,把 0.395 提高到 0.779。即使 q=0.01q = 0.01、M=50M = 50,一个 30 天的月份里仍约有 18.2 天毫无吻合。按利特尔伍德的计数,一天有 28,800 件事,百万分之一的事件每 34.7 天发生一次。

每日一张牌至少吻合当天一件事的概率(p09)。

选择性记忆:从四成到「每次都准」

人们记住牌面吻合的日子,胜过记住不吻合的日子。设吻合的日子被记住的概率是 0.90,不吻合的是 0.20。

推导

# Chapter 5. p10: Selective memory turns a 40% hit rate into "it works".
# Days on which the card fitted an event are remembered with probability
# 0.90, days on which it did not with probability 0.20. The remembered
# hit rate is r = 0.9 p / (0.9 p + 0.2 (1 - p)), with p from p09
# (q = 0.01 and 0.02, M = 50). 10,000 simulated users keep no diary for
# 30 days (seed 89); a diary written every evening removes the bias.
import random

def remembered(p, a=0.90, b=0.20):
    return a * p / (a * p + b * (1 - p))

def p_fit(q, M):
    return 1 - (1 - q) ** M

if __name__ == "__main__":
    for q in (0.01, 0.02):
        p = p_fit(q, 50)
        print("   q={:.2f}: true hit rate {:.3f}, remembered {:.3f}".format(
            q, p, remembered(p)))
    rng = random.Random(89)
    p = p_fit(0.01, 50)
    rates, sure = [], 0
    for _ in range(10000):
        hit = miss = 0
        for _ in range(30):
            if rng.random() < p:
                hit += rng.random() < 0.90
            else:
                miss += rng.random() < 0.20
        r = hit / (hit + miss) if hit + miss else 0.0
        rates.append(r)
        sure += r >= 0.70
    rates.sort()
    print("   30 days, no diary: median remembered rate {:.3f}; "
          "{:.3f} of users".format(rates[5000], sure / 10000))
    print("     remember it working on at least 70% of days")
    print("   with a daily diary the recorded rate is the true {:.3f}"
          .format(p))
#    q=0.01: true hit rate 0.395, remembered 0.746
#    q=0.02: true hit rate 0.636, remembered 0.887
#    30 days, no diary: median remembered rate 0.750; 0.671 of users
#      remember it working on at least 70% of days
#    with a daily diary the recorded rate is the true 0.395

真实命中率 0.395 被记成 0.746,0.636 被记成 0.887。模拟 10,000 名不记日记的用户 30 天:记忆中命中率的中位数是 0.750,有 0.671 的用户记得牌在至少 70% 的日子里应验。每天晚上写日记,在记忆筛选之前记下吻合与否,就能找回真实的 0.395。

吻合的日子以 0.90、其余日子以 0.20 的概率被记住时,记忆中的命中率与真实命中率(p10)。

假说 · 尚未验证

作者的诠释(尚未验证): 连续四周每天记录「吻合与否」的用户,对每日牌准确度的感受会低于不记录的用户,而他们报告的「帮助自己开始工作」的程度不会下降。什么会证明它错: 记日记者感到的准确度并不更低,或者他们报告的开工帮助与感到的准确度同样下降。

小结

每天 50 件小事、每件百分之一的吻合机会,每日一张牌单凭偶然就在 0.395 的日子里「应验」,三张牌的牌阵是 0.779。记住命中、忘掉落空,会把 0.395 变成 0.746。每天记日记是最便宜的校正。

参考文献

Diaconis, Persi, and Frederick Mosteller. 1989. “Methods for Studying Coincidences.” Journal of the American Statistical Association 84 (408): 853 to 861.
Dyson, Freeman. 2004. “One in a Million.” The New York Review of Books 51 (5).
Littlewood, John E. 1953. A Mathematician’s Miscellany. Methuen.

Fan, Y. (2026). Dream Motif Bipartite Graphs, Deterministic Oracle Decks and Subconscious Prior Activation: REM Semantic Spreading, SHA-256 Fisher-Yates Permutations and Retroactive Coincidence Rates [梦境意象二分图、确定性哈希牌卡与潜意识先验激活] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090233 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).