关于「一万亿种气味?重算 Bushdid 估计量及其两份反驳」,《沉香挥发与嗅觉适应》给出的定量结论是什么?

结论:第一章逐位复现 Bushdid 等人 2014 年的估计 1.72 × 1012,再并陈 Gerkin 与 Castro、Meister 的反驳与原作者一方的理由:同一公式可以给出从 4.56 × 103 到 1.54 × 10^29 的任何数;固定一个鼻子、只增加测试次数,估计从 10 的 15.27 次方爬到 26.15 次方;只能区分 9 种气味的一维模型也被说成 10 的 18.19 次方。结论是推算不成立、数目仍未知。

数据口径:书中的房间、香材、香炉与适应参数均为示意;全部计算由书中程序生成;本书是非个人化经营建议,不讨论健康影响,不是医疗建议。 第五章把嗅觉适应写成带时间常数的过程:示意参数下,点香后第 15 分钟主人只剩新鲜鼻子 0.250 的灵敏度,用加量追赶需要 12.4 倍浓度;「主人几乎闻不到」依赖适应深度,须实测。

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: 1.54, 1.72, 10, 12, 29, 4.56

N overlap swaps r z = discriminable mixtures
100.5117 2.441 9.37*10^7
100.5743 2.128 4.01*10^8
200.5117 4.883 1.12*10^11
200.5743 4.257 1.55*10^12
300.5117 7.324 1.72*10^12
300.5743 6.385 5.58*10^13
N30, r = 0 (any difference noticed): 1.54*10^29
N30, r = 15 (only disjoint pairs): 4.56*10^3
library 128256 at the same overlap: x 1.43*10^7
Crocker-Henderson code space 9^46561
least sensitive subject 7.84*10^7: implied overlap 0.2965
most sensitive subject 1.03*10^28: implied overlap 0.9774
查看完整验算程序
# Chapter 1. p11: re-deriving the "one trillion odours" estimate.
# Model (Bushdid et al. 2014, Science): C = 128 molecules, mixtures of
# N components. If two mixtures that differ in a fraction f of their
# components are just discriminable, each discriminable mixture owns a
# Hamming-type ball of radius r = N f / 2 (r swaps), and the number of
# discriminable mixtures is z = binom(C, N) / ball(r). The ball is
# summed over integer r and interpolated log-linearly in between.
# Overlap thresholds 51.17 % and 57.43 % are the published values; the
# two extreme rows (r = 0, r = N / 2) are the bounds quoted by Gerkin &
# Castro (2015, eLife). Crocker-Henderson: 4 qualities x 9 levels.
# The published per-subject extremes for N = 30 (7.84*10^7 and
# 1.03*10^28) are inverted to the overlap each subject implies.
import math

C = 128


def ball(n, c, r):
    terms = [math.comb(n, k) * math.comb(c - n, k) for k in range(r + 1)]
    return math.fsum(terms)


def log10_z(n, c, r_real):
    total = math.log10(math.comb(c, n))
    lo = int(math.floor(r_real))
    hi = min(lo + 1, n // 2)
    a = total - math.log10(ball(n, c, lo))
    b = total - math.log10(ball(n, c, hi))
    w = r_real - lo
    return a + (b - a) * w


def sci(lg):
    e = int(math.floor(lg))
    m = 10 ** (lg - e)
    if m >= 9.995:
        m, e = m / 10, e + 1
    return "{:.2f}*10^{}".format(m, e)


print("N   overlap   swaps r   z = discriminable mixtures")
for n in (10, 20, 30):
    for ov in (0.5117, 0.5743):
        r = n * (1 - ov) / 2
        print("{:2d}  {:7.4f}  {:7.3f}   {}".format(
            n, ov, r, sci(log10_z(n, C, r))))
print("N = 30, r = 0 (any difference noticed): {}".format(
    sci(log10_z(30, C, 0))))
print("N = 30, r = 15 (only disjoint pairs):   {}".format(
    sci(log10_z(30, C, 15))))
r30 = 30 * (1 - 0.5117) / 2
gain = log10_z(30, 2 * C, r30) - log10_z(30, C, r30)
print("library 128 -> 256 at the same overlap: x {}".format(sci(gain)))
print("Crocker-Henderson code space 9^4 = {}".format(9 ** 4))


def overlap_for(lg_target, n=30):
    lo, hi = 0.0, n / 2
    for _ in range(60):
        mid = (lo + hi) / 2
        if log10_z(n, C, mid) > lg_target:
            lo = mid
        else:
            hi = mid
    return 1 - 2 * lo / n


for label, lg in (("least sensitive", math.log10(7.84e7)),
                  ("most sensitive", math.log10(1.03e28))):
    print("{} subject {}: implied overlap {:.4f}".format(
        label, sci(lg), overlap_for(lg)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《沉香倍半萜挥发动力学、嗅觉受体组合编码与嗅觉适应时间常数母典》第 1 章「一万亿种气味?重算 Bushdid 估计量及其两份反驳」(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-205 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026205,
  author    = {Fan, Yafu},
  title     = {Agarwood Sesquiterpene Volatilisation, Combinatorial Olfactory-Receptor Coding and the Time Constant of Olfactory Adaptation},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论