5D Library › 超临界病毒级联、信息缺口风险率与多层付费墙优化 · English

第二章 有限的受众、聚集的网络与没有归零的获客成本

受众是有限的

分支过程假设每个新用户都面对一个无穷大的、从未被邀请过的人群。真实的受众是有限的:随着越来越多的人已经被触达,邀请越来越多地落在已经是用户或已经拒绝过的人身上,有效的 K 值不断下降。这是传染病学中最古老的结论之一:克马克与麦肯德里克的「最终规模」关系 z=1−e−Kzz = 1 – e^{-Kz} 给出一次流行最终触达的人群比例 (Kermack and McKendrick 1927)。

第三段程序在一个 100 万人的受众中逐代推演,并计算口号所说的「获客成本归零」实际是多少。

推导

# Chapter 2. A finite audience, and what "zero acquisition cost" hides.
# Part 1. Users can only be recruited from a finite reachable audience of
# 1,000,000 people. Generation by generation (100 seed users), each new
# user's invitations land on people already reached in proportion to the
# share already reached, so the effective number is K_eff = K x (unreached
# share). The program reports the final share reached (it solves the
# classical final-size relation z = 1 - exp(-K z) as a check), the
# generation at which K_eff falls below 1, and the peak generation size.
# Part 2. Every new user receives a free reading that costs c = 0.15 to
# generate and deliver; a 5 % share pay the price p = 9.9; a paying user
# whose referral link brought the payer earns a commission of 50 %. With
# the funnel of p02 at K = 1.15, a share 0.05 / 1.15 of joiners comes from
# paid links. Acquisition cost per paying user =
#   c / 0.05 + (share of payers from paid links) x 0.5 x 9.9.
import math

AUDIENCE, SEEDS = 1_000_000, 100

def run(k):
    reached, new, gen = SEEDS, SEEDS, 0
    below, peak = None, (0, SEEDS)
    while new >= 1:
        unreached = 1 - reached / AUDIENCE
        if below is None and k * unreached < 1:
            below = gen
        new = k * unreached * new
        reached += new
        gen += 1
        if new > peak[1]:
            peak = (gen, new)
    return reached / AUDIENCE, below, peak

def final_size(k):
    z = 0.5
    for _ in range(500):
        z = 1 - math.exp(-k * z)
    return z

if __name__ == "__main__":
    print("   K      reached   z = 1 - exp(-Kz)   K_eff < 1 from gen"
          "   peak gen (size)")
    for k in (1.15, 1.5, 2.0, 3.0):
        share, below, (pg, ps) = run(k)
        print("   {:4.2f}    {:5.3f}        {:5.3f}                {:3d}"
              "             {:3d} ({:,.0f})".format(k, share, final_size(k),
                                                    below, pg, ps))
    c, conv, price, alpha, k = 0.15, 0.05, 9.9, 0.5, 1.15
    via_link = 0.05 / k
    cac = c / conv + via_link * alpha * price
    print("\n   cost per paying user: free readings {:.2f} + commissions"
          " {:.2f} = {:.2f}".format(c / conv, via_link * alpha * price, cac))
    print("   as a share of the price 9.9: {:.3f}".format(cac / price))
    print("\nReading: in a finite audience K_eff falls below 1 once a share")
    print("1 - 1/K has been reached; a K just above 1 reaches only about a")
    print("quarter of the audience. (The generation recurrence reaches a")
    print("little more than the classical relation, which also counts")
    print("repeated invitations to the same person.) Referral growth is not")
    print("free: the free readings given to everyone who never pays are the")
    print("main acquisition cost.")
#    K      reached   z = 1 - exp(-Kz)   K_eff < 1 from gen   peak gen (size)
#    1.15    0.255        0.249                 42              42 (9,421)
#    1.50    0.618        0.583                 19              19 (72,388)
#    2.00    0.869        0.797                 12              12 (187,722)
#    3.00    0.993        0.940                  8               8 (394,652)
#
#    cost per paying user: free readings 3.00 + commissions 0.22 = 3.22
#    as a share of the price 9.9: 0.325
#
# Reading: in a finite audience K_eff falls below 1 once a share
# 1 - 1/K has been reached; a K just above 1 reaches only about a
# quarter of the audience. (The generation recurrence reaches a
# little more than the classical relation, which also counts
# repeated invitations to the same person.) Referral growth is not
# free: the free readings given to everyone who never pays are the
# main acquisition cost.

K = 1.15 时,级联最终只触达受众的 0.255;K = 1.50 时为 0.618;K = 3.00 时几乎覆盖全部(0.993)。逐代推演的结果略高于经典关系(0.249、0.583……),因为经典关系还计入了对同一个人的重复邀请。有效 K 在受众被触达 1−1/K1 – 1/K 的比例时跌破 1,此后每一代都比上一代小;K = 1.15 时这发生在第 42 代,单代新增的峰值约为 9,421 人。

有限受众中,最终触达比例随 K 的变化。实线为逐代推演,虚线为经典的最终规模关系。

下半部分是口号省略的账单。每个新用户都得到一份免费测算,生成与送达的成本为 0.15;付费率为 5% 时,每个付费用户背后有 20 份送出去的免费测算,即 3.00;再加上推荐分佣 0.22,每个付费用户的获客成本为 3.22,占 9.9 元价格的 0.325。获客成本没有归零,它只是从广告账单转到了算力账单与分佣账单上。

朋友的朋友

分支过程还假设一个用户的朋友彼此不认识。真实的社交网络是聚集的:你的朋友往往也是彼此的朋友。第四段程序在一个两万人的环形网络上比较三种结构:完全聚集(每人连接左右最近的十人)、少量随机连接、几乎完全随机。

推导

# Chapter 2. Friends of friends: why real networks give less than K.
# 20,000 people on a ring, each tied to the 10 nearest (5 each side);
# each tie is rewired to a random person with probability beta (beta = 0:
# tightly clustered, everyone's friends know each other; beta = 1: almost
# random). Each new user invites every friend not yet reached; each
# invitation is accepted with probability q = 0.15. On a tree a user has
# 9 friends besides the inviter, so the naive estimate is K = 9 x 0.15 =
# 1.35. From 5 seeds, 20 runs per beta: the measured mean number of new
# users per user in generations 1-3, and the final share reached.
import random

N, HALF, Q = 20000, 5, 0.15

def network(beta, rng):
    nb = [set() for _ in range(N)]
    for i in range(N):
        for d in range(1, HALF + 1):
            j = (i + d) % N
            if rng.random() < beta:
                j = rng.randrange(N)
                while j == i or j in nb[i]:
                    j = rng.randrange(N)
            nb[i].add(j)
            nb[j].add(i)
    return [list(s) for s in nb]

def spread(nb, rng):
    reached = set(rng.sample(range(N), 5))
    frontier = list(reached)
    early_kids = early_parents = 0
    gen = 0
    while frontier:
        nxt = []
        for u in frontier:
            kids = 0
            for v in nb[u]:
                if v not in reached and rng.random() < Q:
                    reached.add(v)
                    nxt.append(v)
                    kids += 1
            if 1 <= gen <= 3:
                early_kids += kids
                early_parents += 1
        frontier = nxt
        gen += 1
    return early_kids / max(early_parents, 1), len(reached) / N

if __name__ == "__main__":
    rng = random.Random(63004)
    print("   beta    early K (gens 1-3)    final share reached")
    for beta in (0.0, 0.1, 1.0):
        ks, shares = [], []
        for _ in range(20):
            k, sh = spread(network(beta, rng), rng)
            ks.append(k)
            shares.append(sh)
        print("   {:3.1f}          {:5.3f}                {:5.3f}".format(
            beta, sum(ks) / len(ks), sum(shares) / len(shares)))
    print("\n   naive tree estimate: K = 9 x 0.15 = {:.2f}".format(9 * Q))
    print("\nReading: in a clustered network a new user's friends have")
    print("mostly been invited already, so the measured K falls below 1")
    print("although the tree estimate is 1.35, and the cascade fills a local")
    print("pocket and stalls. Random ties help, but only a nearly random")
    print("network gets close to the tree estimate and a large reach.")
#    beta    early K (gens 1-3)    final share reached
#    0.0          0.934                0.003
#    0.1          1.042                0.081
#    1.0          1.402                0.570
#
#    naive tree estimate: K = 9 x 0.15 = 1.35
#
# Reading: in a clustered network a new user's friends have
# mostly been invited already, so the measured K falls below 1
# although the tree estimate is 1.35, and the cascade fills a local
# pocket and stalls. Random ties help, but only a nearly random
# network gets close to the tree estimate and a large reach.

树状估计给出 K = 1.35。在完全聚集的网络中,前几代实测的 K 只有 0.934:一个新用户的朋友大多已经被同一批人邀请过,邀请落空。级联填满一个局部之后就停住,最终只触达 0.003。十分之一的连接随机化后,实测的 K 为 1.042、触达 0.081;只有在几乎完全随机的网络中,实测的 K(1.402)才接近树状估计,触达达到 0.570。

证据

聚集对不同的传播作用相反。 小世界网络的研究说明,少量的长程连接就能大大缩短网络中的距离 (Watts and Strogatz 1998)。但对于需要多次接触才会采纳的「复杂传染」,聚集反而有利:森托拉的在线实验发现,在聚集的网络中,一项健康行为传播得更远、更快,因为人们从多个邻居那里得到了强化 (Centola 2010)。拉人解锁的邀请更接近只需一次接触的「简单传染」,本章的模型就是按简单传染写的;如果用户需要被多个朋友劝说才会加入,结论可能相反。

注

对「流量自发指数级爆炸」的含义。 一个按树状结构估算的 K 值,在真实的社交网络中会被高估;在有限的受众中会随时间下降;而它所带来的每一个新用户都有成本。增长计划应当按逐代实测的 K、触达比例与每个付费用户的完整获客成本来写,而不是按一个起始 K 值外推。

小结

在有限的受众中,有效 K 随触达比例下降,K = 1.15 的级联最终只触达约四分之一的受众。聚集的社交网络让实测的 K 低于树状估计:完全聚集时为 0.934,对比估计的 1.35,级联在局部停住。获客成本并未归零:送给不付费者的免费测算是主要成本,在例子中每个付费用户的获客成本为价格的 0.325。

参考文献

Centola, Damon. 2010. “The Spread of Behavior in an Online Social Network Experiment.” Science 329 (5996): 1194–97.
Kermack, W. O., and A. G. McKendrick. 1927. “A Contribution to the Mathematical Theory of Epidemics.” Proceedings of the Royal Society of London A 115 (772): 700–721.
Watts, Duncan J., and Steven H. Strogatz. 1998. “Collective Dynamics of ’Small-World’ Networks.” Nature 393: 440–42.

Fan, Y. (2026). Supercritical Viral Cascades, Information-Gap Hazard Rates and Multi-Tier Paywall Optimization: Galton-Watson Referral Thresholds, Share-to-Unlock Incentives and Second-Degree Price Discrimination [超临界病毒级联、信息缺口风险率与多层付费墙优化] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090174 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).