5D Library › 玄空飞星置换代数、二十四山罗盘离散化与骨相比例计量 · English
第五章 特征点误差与巴纳姆效应
几个像素的误差
从照片读比例,要先在照片上标出特征点。设比例 , 为 120 像素, 为 200 像素(),四个特征点在两个坐标上各有标准差为 像素的独立误差。按一阶近似(delta 方法), 的相对标准差为 。若一种解读以「 落在 0.58 至 0.62 之间」为理想,两位标注者对同一张照片的判断会有多少分歧?
推导
# Chapter 5. p09: Landmark error and how often a reading flips.
# A ratio R = a / b of two distances on a photo, a = 120 px and
# b = 200 px (R = 0.6). Each of the four landmarks is placed with an
# independent normal error of sigma px in each coordinate. To first
# order (delta method) the error of a distance along its own line has
# sd sigma sqrt(2), so the relative sd of R is
# sqrt(2 sigma^2 / a^2 + 2 sigma^2 / b^2).
# A reading calls a face "ideal" if R lies in [0.58, 0.62]. For a face
# whose true R is 0.600 (centre of the band) or 0.615 (near its edge),
# we estimate how often two independent annotations of the same photo
# disagree on that call. 100,000 simulated pairs, fixed seed.
import math
import random
A, B = 120.0, 200.0
def measured(rng, sigma, ra):
a = A * ra / 0.6
pts = [(0.0, 0.0), (a, 0.0), (0.0, 50.0), (B, 50.0)]
q = [(x + rng.gauss(0, sigma), y + rng.gauss(0, sigma)) for x, y in pts]
da = math.dist(q[0], q[1])
db = math.dist(q[2], q[3])
return da / db
if __name__ == "__main__":
rng = random.Random(9)
print(" sigma px delta-method sd of R simulated sd of R")
for s in (1.0, 2.0, 4.0):
sd = 0.6 * math.sqrt(2 * s * s / A ** 2 + 2 * s * s / B ** 2)
xs = [measured(rng, s, 0.6) for _ in range(20000)]
m = sum(xs) / len(xs)
sim = math.sqrt(sum((x - m) ** 2 for x in xs) / (len(xs) - 1))
print(" {:8.1f} {:22.4f} {:19.4f}".format(s, sd, sim))
print("\n two annotations disagree on 'ideal' (band 0.58 to 0.62):")
print(" sigma px true R 0.600 true R 0.615")
for s in (1.0, 2.0, 4.0):
row = []
for ra in (0.600, 0.615):
dis = 0
for _ in range(50000):
r1 = measured(rng, s, ra)
r2 = measured(rng, s, ra)
dis += (0.58 <= r1 <= 0.62) != (0.58 <= r2 <= 0.62)
row.append(dis / 50000)
print(" {:8.1f} {:14.4f} {:14.4f}".format(s, *row))
print("\nReading: one pixel of placement error already makes two")
print("annotators disagree on four faces in ten near the edge of the")
print("band; at two pixels they disagree on a third of the faces in")
print("its very middle. A reading that turns on such a band reports")
print("the annotator's hand as much as the face.")
# sigma px delta-method sd of R simulated sd of R
# 1.0 0.0082 0.0082
# 2.0 0.0165 0.0165
# 4.0 0.0330 0.0330
#
# two annotations disagree on 'ideal' (band 0.58 to 0.62):
# sigma px true R 0.600 true R 0.615
# 1.0 0.0316 0.3984
# 2.0 0.3513 0.4790
# 4.0 0.4992 0.4871
#
# Reading: one pixel of placement error already makes two
# annotators disagree on four faces in ten near the edge of the
# band; at two pixels they disagree on a third of the faces in
# its very middle. A reading that turns on such a band reports
# the annotator's hand as much as the face.delta 方法与模拟完全一致: 为 1、2、4 像素时, 的标准差分别为 0.0082、0.0165、0.0330。对真实 恰在区间中点(0.600)的脸,1 像素的误差只造成 0.0316 的分歧,2 像素时就达到 0.3513;对接近区间边缘(0.615)的脸,1 像素时已有 0.3984 的分歧,接近抛硬币。依赖这种区间的解读,报告的既是脸,也是标注者的手。
测得越多,人人越特别
Forer 给学生做了一个人格测验,然后发给每人同一份笼统的描述,学生们普遍认为它准确地描述了自己 (Forer 1949)。面相解读还有一个更算术的版本:测量的比例越多,每个人越必然「有特别之处」。程序设每个比例的中央 30% 为「普通」,其余 70% 为「显著」;中央 90% 以外为「罕见」,各比例相互独立。
推导
# Chapter 5. p10: Many ratios, and everybody has something special.
# A face reading measures K proportions. For each, call a face
# "notable" if it falls outside a central band holding 30 % of people
# (so 0.7 of faces are notable on that ratio), and "rare" if it falls
# outside the central 90 % (0.1 of faces). With independent ratios:
# P(no notable ratio) = 0.3^K; P(at least one rare) = 1 - 0.9^K;
# expected notable ratios = 0.7 K.
# Correlated ratios behave like fewer independent ones; we also show
# K_eff = K / 2. A check by simulation for K = 12 (100,000 faces).
import random
if __name__ == "__main__":
print(" K nothing notable, 1 in P(some rare ratio) mean notable")
for k in (5, 10, 12, 20, 30):
print(" {:2d} {:23,.0f} {:20.4f} {:14.1f}".format(
k, 0.3 ** -k, 1 - 0.9 ** k, 0.7 * k))
print("\n with K_eff = K / 2:")
for k in (10, 20, 30):
print(" K = {:2d}: P(some rare ratio) {:.4f}".format(
k, 1 - 0.9 ** (k / 2)))
rng = random.Random(1949)
n, rare = 100000, 0
for _ in range(n):
rare += any(rng.random() < 0.1 for _ in range(12))
print("\n simulation, K = 12: share with a rare ratio {:.4f}".format(
rare / n))
print("\nReading: with a dozen measured ratios, almost every face has")
print("several notable ones and most have at least one rare one. A")
print("reading that finds 'something special' in each client finds")
print("what arithmetic guarantees, which is why it feels accurate.")
# K nothing notable, 1 in P(some rare ratio) mean notable
# 5 412 0.4095 3.5
# 10 169,351 0.6513 7.0
# 12 1,881,676 0.7176 8.4
# 20 28,679,719,908 0.8784 14.0
# 30 4,856,935,749,618,867 0.9576 21.0
#
# with K_eff = K / 2:
# K = 10: P(some rare ratio) 0.4095
# K = 20: P(some rare ratio) 0.6513
# K = 30: P(some rare ratio) 0.7941
#
# simulation, K = 12: share with a rare ratio 0.7183
#
# Reading: with a dozen measured ratios, almost every face has
# several notable ones and most have at least one rare one. A
# reading that finds 'something special' in each client finds
# what arithmetic guarantees, which is why it feels accurate.测量 12 个比例时,一个比例都不显著的人只有 1,881,676 分之一;平均每人有 8.4 个显著的比例;有至少一个罕见比例的人占 0.7176(模拟值 0.7183)。测量 30 个比例时,这一比例为 0.9576。即使比例之间相关、有效个数只有一半,测 30 个比例仍有 0.7941 的人至少有一个罕见比例。一种在每位客户身上都找到「特别之处」的解读,找到的是算术所保证的东西;这正是它让人觉得准确的原因。
假说 · 尚未验证
可检验的推论。 若面相解读包含关于本人的信息,那么在盲选测验中,客户应能从两份解读中选出根据自己照片做出的那一份,命中率显著高于一半。什么会证明这一推论错:若命中率与一半无差别,解读的「准确感」来自巴纳姆效应,而不是来自面部。
小结
几个像素的标点误差就足以让两位标注者对同一张脸的「理想」判断产生大量分歧。测量的比例越多,每个人越必然显得特别;这使任何解读都容易被感受为准确。
参考文献
Fan, Y. (2026). Xuankong Flying-Star Permutations, Twenty-Four-Mountain Compass Discretization and Craniofacial Ratio Metrology: Lo-Shu Orbit Cycles, Magnetic Declination Flip Rates and Landmark Measurement Error Propagation [玄空飞星置换代数、二十四山罗盘离散化与骨相比例计量] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090215 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).