5D Library › 分销网络幂律激活、网红矩阵组合优化与激励稳定边界 · English

第一章 分销网络的重尾产出与普赖斯定律

「根号 N 个人产出一半」在什么条件下成立

科学计量学家普赖斯曾提出:在 NN 位作者中,最多产的 N\sqrt{N} 位贡献约一半的论文 (Price 1963)。这条经验法则常被搬到销售团队与分销网络上:一万个分销员里,最好的一百人卖出一半。它能不能成立,取决于产出分布的尾部有多重。下面让每位分销员的产出服从尾指数为 α\alpha 的 Lomax(第二类帕累托)分布,α\alpha 越小尾部越重。

推导

# Chapter 1. p01: Does the top square root of N really produce half? Price's
# law says that in a group of N producers, the top sqrt(N) produce about
# half of all output. Here output per affiliate follows a Lomax (Pareto
# type II) law with tail index alpha: x = U^(-1/alpha) - 1. For several
# alpha and N, the median share of the top sqrt(N) over repeats is shown,
# seed 87. A smaller alpha means a heavier tail (alpha <= 1: infinite mean).
import random
from math import isqrt

def lomax(rng, alpha):
    return (1.0 - rng.random()) ** (-1.0 / alpha) - 1.0

def top_share(rng, n, alpha):
    xs = sorted((lomax(rng, alpha) for _ in range(n)), reverse=True)
    k = isqrt(n)
    return sum(xs[:k]) / sum(xs)

if __name__ == "__main__":
    rng = random.Random(87)
    sizes = (100, 1000, 10000)
    print("   share of output from the top sqrt(N) (median of repeats)")
    print("   alpha    N = 100   N = 1,000   N = 10,000")
    for alpha in (1.1, 1.5, 2.0, 3.0):
        row = []
        for n in sizes:
            reps = 300 if n < 10000 else 60
            s = sorted(top_share(rng, n, alpha) for _ in range(reps))
            row.append(s[reps // 2])
        print("   {:5.1f}    {:7.3f}   {:9.3f}   {:10.3f}"
              .format(alpha, *row))
    print("   top sqrt(N) as a share of people: {:.3f} / {:.3f} / {:.3f}"
          .format(10 / 100, 31 / 1000, 100 / 10000))
#    share of output from the top sqrt(N) (median of repeats)
#    alpha    N = 100   N = 1,000   N = 10,000
#      1.1      0.691       0.603        0.545
#      1.5      0.584       0.422        0.299
#      2.0      0.492       0.306        0.189
#      3.0      0.430       0.228        0.119
#    top sqrt(N) as a share of people: 0.100 / 0.031 / 0.010

只有在两种组合下,前 N\sqrt{N} 人的份额才接近一半:NN = 100 而 α\alpha = 2.0(0.492),或尾部极重的 α\alpha = 1.1(NN = 10,000 时 0.545)。α\alpha = 1.5 时,份额从 100 人的 0.584 降到一万人的 0.299。原因很简单:N\sqrt{N} 占总人数的比例随 NN 变小,一万人时只有 0.010;除非尾部重到少数人的产出可以压过整体,这一小撮人不可能一直扛起一半。

前 \sqrt{N} 人的产出份额随网络规模变化(p01)。

沉睡的分销员与尾指数的估计

真实的分销网络里,大多数注册者从未卖出一单。p02 设定一万名注册者中 75% 沉睡,活跃的四分之一按 α\alpha = 1.5 的 Lomax 分布销售,再用 Hill 估计量从最大的 kk 个销售额估计尾指数 (Hill 1975; Clauset et al. 2009)。

推导

# Chapter 1. p02: Dormant affiliates and the tail index. Of 10,000
# registered affiliates, 75% never make a sale; the active quarter sell
# 10 x Lomax(alpha = 1.5) units a year (see p01). The top sqrt(N) share
# depends on whether N counts everyone or only the active. The tail
# index is then estimated with Hill's estimator from the k largest
# sellers, alpha_hat = k / sum(log(x_i / x_(k+1))), for several k: too
# large a k reaches into the body of the law, where Lomax is not a pure
# power law, and biases alpha_hat downward: the tail looks heavier than
# it is. 200 repeats, seed 87.
import random
from math import isqrt, log

def network(rng, n=10000, dormant=0.75, alpha=1.5):
    out = []
    for _ in range(n):
        if rng.random() < dormant:
            out.append(0.0)
        else:
            out.append(10.0 * ((1.0 - rng.random()) ** (-1.0 / alpha) - 1))
    return out

def hill(xs_desc, k):
    ref = xs_desc[k]
    return k / sum(log(x / ref) for x in xs_desc[:k])

if __name__ == "__main__":
    rng = random.Random(87)
    ks = (25, 50, 100, 250, 500, 1000)
    est = {k: [] for k in ks}
    s_all, s_act, med, zero = [], [], [], []
    for _ in range(200):
        xs = sorted(network(rng), reverse=True)
        act = [x for x in xs if x > 0]
        s_all.append(sum(xs[:isqrt(len(xs))]) / sum(xs))
        s_act.append(sum(act[:isqrt(len(act))]) / sum(act))
        med.append(xs[len(xs) // 2])
        zero.append(1 - len(act) / len(xs))
        for k in ks:
            est[k].append(hill(xs, k))
    mid = lambda v: sorted(v)[len(v) // 2]
    print("   top sqrt(N) share, N = all registered (top 100): {:.3f}"
          .format(mid(s_all)))
    print("   top sqrt(N) share, N = active only (top {}):      {:.3f}"
          .format(isqrt(2500), mid(s_act)))
    print("   median registered affiliate sells {:.1f}; share never"
          " selling {:.3f}".format(mid(med), mid(zero)))
    print("   Hill estimate of alpha (true 1.5)")
    print("       k   median   5% to 95%")
    for k in ks:
        v = sorted(est[k])
        print("   {:5,d}   {:6.2f}   {:.2f} to {:.2f}".format(
            k, v[100], v[10], v[189]))
#    top sqrt(N) share, N = all registered (top 100): 0.460
#    top sqrt(N) share, N = active only (top 50):      0.356
#    median registered affiliate sells 0.0; share never selling 0.750
#    Hill estimate of alpha (true 1.5)
#        k   median   5% to 95%
#       25     1.50   1.10 to 2.22
#       50     1.45   1.16 to 1.87
#      100     1.39   1.17 to 1.63
#      250     1.30   1.17 to 1.44
#      500     1.17   1.09 to 1.25
#    1,000     0.97   0.92 to 1.01

把沉睡者算进 NN 时,前一百人的份额是 0.460,看起来很像普赖斯定律;只算活跃者,前五十人的份额是 0.356。同一个网络,结论取决于分母。注册者的中位销售额是零。尾指数的估计同样对 kk 很敏感:kk = 25 时中位估计 1.50 但区间宽达 1.10 至 2.22;kk 取到 500 时估计降到 1.17,到 1,000 时只有 0.97,误把尾部估得比实际重,甚至越过均值无穷的界线。

注

报告集中度时的三条规矩: 写明 NN 是注册者还是活跃者;同时报告中位数与平均数;估计尾指数时给出 kk 的选择方法与区间,不要只报一个数。

小结

普赖斯定律不是常数,而是尾指数与网络规模的函数;在 α\alpha = 1.5 这类常见的重尾下,网络越大,前 N\sqrt{N} 人的份额越低。沉睡者会让集中度看起来更高,Hill 估计在 kk 取得太大时会把尾部估得过重。

参考文献

Clauset, Aaron, Cosma Rohilla Shalizi, and M. E. J. Newman. 2009. “Power-Law Distributions in Empirical Data.” SIAM Review 51 (4): 661 to 703.
Hill, Bruce M. 1975. “A Simple General Approach to Inference about the Tail of a Distribution.” The Annals of Statistics 3 (5): 1163 to 1174.
Price, Derek J. de Solla. 1963. Little Science, Big Science. Columbia University Press.

Fan, Y. (2026). Affiliate Network Power Laws, Influencer Portfolio Optimization and Incentive Stability Bounds: Heavy-Tailed Productivity, Markowitz Traffic Allocation and Single-Tier Commission Equilibria [分销网络幂律激活、网红矩阵组合优化与激励稳定边界] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090229 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).