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

第一章 拉人解锁的分支过程

两条通道

设想一个测算网站的漏斗。新用户先看到一段免费的测算,然后面临三种选择:邀请朋友,凑够 mm 人加入就免费解锁完整结果(分享通道);直接付费,并得到一个可以分佣的推荐链接(付费通道);或者离开。每一个接受邀请的朋友又是一个新用户,重新走一遍同样的漏斗。

这正是高尔顿—沃森分支过程:每个「个体」独立地产生随机数目的「后代」,后代再产生后代 (Harris 1963; Athreya and Ney 1972)。口号里的 K 值就是每个用户平均带来的新用户数,即后代分布的均值。分支过程的基本定理说:K ≤ 1 时,从一个种子出发的级联以概率 1 熄灭;K > 1 时,它以正的概率无限增长。

推导

两通道的 K 值。 记分享率为 s(m)s(m)(愿意为解锁去拉人的用户比例,随所需人数 mm 下降),每份邀请被接受的概率为 qq,分享者最多发出 N=20N = 20 份邀请、凑够 mm 人即停。付费率为 bb,付费者发出 10 份推荐链接、各以 0.10 的概率被接受。则 K=s(m)E[min⁡(Bin(N,q),m)]+b×10×0.10.K = s(m)\, \mathrm{E}\big[\min(\mathrm{Bin}(N, q),\, m)\big] + b \times 10 \times 0.10 . 因为分享者凑够 mm 人就停,第一项不可能超过 s(m)ms(m)\,m:无论 qq 多高,K 都被分享率封顶。

第一段程序取 s(m)=s0e−(m−1)/3s(m) = s_0\, e^{-(m-1)/3},对三个 s0s_0 计算 K、它的上限,以及使 K 恰为 1 的接受率 q*q^{*}。

推导

# Chapter 1. The reproduction number of a two-channel referral funnel.
# A new user sees a free teaser and then takes one of three paths:
#   share path  (prob s(m)): to unlock the full reading they invite contacts
#               one by one, each accepting with prob q, until m have joined
#               or N = 20 invitations are used up; offspring min(Bin(N,q),m)
#   pay path    (prob b = 0.05): they pay and may share a commission link;
#               they send 10 invitations, each accepted with prob 0.10
#   leave       (the rest): no offspring
# Willingness to take the share path falls with the number demanded:
#   s(m) = s0 exp(-(m - 1) / 3),  s0 = the share rate when m = 1.
# Every accepted invitee is a new user who starts the same funnel, so the
# users form a Galton-Watson branching process with mean offspring
#   K = s(m) E[min(Bin(N, q), m)] + b 10 0.10.
# K > 1 is the supercritical regime. Because a sharer stops at m joiners,
# K can never exceed s(m) m + 0.05, whatever q is. For three values of s0
# the program tabulates K at q = 0.15, that ceiling, and the acceptance
# rate q* at which K = 1.
import math

N, B, N2, Q2 = 20, 0.05, 10, 0.10

def s_of(m, s0):
    return s0 * math.exp(-(m - 1) / 3)

def mean_min(n, q, m):
    return sum(min(k, m) * math.comb(n, k) * q ** k * (1 - q) ** (n - k)
               for k in range(n + 1))

def k_value(m, q, s0):
    return s_of(m, s0) * mean_min(N, q, m) + B * N2 * Q2

def q_star(m, s0):
    lo, hi = 0.0, 1.0
    if k_value(m, hi, s0) < 1:
        return None
    for _ in range(60):
        mid = (lo + hi) / 2
        if k_value(m, mid, s0) < 1:
            lo = mid
        else:
            hi = mid
    return hi

if __name__ == "__main__":
    for s0 in (0.30, 0.50, 0.80):
        print("   s0 = {:.2f}   m   K at q = 0.15   ceiling   q* for K = 1"
              .format(s0))
        for m in range(1, 7):
            qs = q_star(m, s0)
            print("               {}       {:5.3f}        {:5.3f}       {}"
                  .format(m, k_value(m, 0.15, s0),
                          s_of(m, s0) * m + B * N2 * Q2,
                          "never" if qs is None else "{:.3f}".format(qs)))
    print("\n   pay channel alone contributes {:.3f}".format(B * N2 * Q2))
    print("\nReading: a sharer brings at most m joiners, so the share rate")
    print("caps K. With 30 % or 50 % of users willing to share, no unlock")
    print("threshold and no acceptance rate reaches K = 1. Only when most")
    print("users share does K cross 1, at a moderate threshold (m = 2 to 3)")
    print("and an acceptance rate that the table states exactly.")
#    s0 = 0.30   m   K at q = 0.15   ceiling   q* for K = 1
#                1       0.338        0.350       never
#                2       0.434        0.480       never
#                3       0.417        0.512       never
#                4       0.352        0.491       never
#                5       0.280        0.445       never
#                6       0.218        0.390       never
#    s0 = 0.50   m   K at q = 0.15   ceiling   q* for K = 1
#                1       0.531        0.550       never
#                2       0.690        0.767       never
#                3       0.661        0.820       never
#                4       0.553        0.786       never
#                5       0.433        0.709       never
#                6       0.331        0.617       never
#    s0 = 0.80   m   K at q = 0.15   ceiling   q* for K = 1
#                1       0.819        0.850       never
#                2       1.074        1.196       0.124
#                3       1.028        1.282       0.143
#                4       0.854        1.227       0.192
#                5       0.662        1.104       0.283
#                6       0.499        0.957       never
#
#    pay channel alone contributes 0.050
#
# Reading: a sharer brings at most m joiners, so the share rate
# caps K. With 30 % or 50 % of users willing to share, no unlock
# threshold and no acceptance rate reaches K = 1. Only when most
# users share does K cross 1, at a moderate threshold (m = 2 to 3)
# and an acceptance rate that the table states exactly.

结果的第一行就推翻了口号的前半句。只有 30% 的用户愿意分享时,K 在任何门槛下都不超过 0.512,没有任何接受率能让它越过 1;50% 时上限是 0.820,同样不行。拉人门槛的选择存在一个内部最优:要求的人越多,每个分享者带来的人越多,但愿意分享的人越少,K 在 m=2m = 2 或 3 时最大。只有当 80% 的用户愿意分享时,K 才能越过 1——门槛为 2 人时需要每份邀请有 0.124 的接受率,门槛为 5 人时需要 0.283。付费通道本身只贡献 0.050。

超临界意味着什么

即使 K 大于 1,口号的后半句「指数爆炸」也只对一部分种子成立。第二段程序在三个 K 值下各模拟 20,000 个从单个种子开始的级联。

推导

# Chapter 1. What a supercritical K does - and does not - promise.
# The funnel of p01 with s0 = 0.80 and unlock threshold m = 2. The
# invitation acceptance rate q is tuned so that K = 0.8, 1.0 and 1.15
# (at m = 2 no acceptance rate gives more than 1.196, the ceiling of p01).
# Offspring distribution: with prob s(2) the user brings min(Bin(20, q), 2),
# with prob 0.05 Bin(10, 0.10), otherwise 0. The extinction probability is
# the smallest root of eta = G(eta), G the offspring generating function
# (found by iterating from 0). 20,000 cascades, each started by one seed
# user and stopped at 10,000 users, give the size distribution.
import math
import random

S, B = 0.80 * math.exp(-1 / 3), 0.05

def dist(q):
    p = [0.0] * 11
    p[0] += 1 - S - B
    for k in range(21):
        pk = math.comb(20, k) * q ** k * (1 - q) ** (20 - k)
        p[min(k, 2)] += S * pk
    for k in range(11):
        p[k] += B * math.comb(10, k) * 0.1 ** k * 0.9 ** (10 - k)
    return p

def mean(p):
    return sum(k * v for k, v in enumerate(p))

def q_for(k_target):
    lo, hi = 0.0, 1.0
    for _ in range(60):
        mid = (lo + hi) / 2
        lo, hi = (mid, hi) if mean(dist(mid)) < k_target else (lo, mid)
    return hi

def extinction(p):
    eta = 0.0
    for _ in range(100000):
        eta = sum(v * eta ** k for k, v in enumerate(p))
    return eta

def cascade(p, rng, cap=10000):
    cum, acc = [], 0.0
    for v in p:
        acc += v
        cum.append(acc)
    total = alive = 1
    while alive and total < cap:
        u = rng.random()
        kids = next(i for i, c in enumerate(cum) if u <= c)
        alive += kids - 1
        total += kids
    return min(total, cap)

if __name__ == "__main__":
    rng = random.Random(63002)
    print("   K     q      dies out   reaches 1,000   median   mean size")
    for kt in (0.8, 1.0, 1.15):
        q = q_for(kt)
        p = dist(q)
        sizes = sorted(cascade(p, rng) for _ in range(20000))
        big = sum(x >= 1000 for x in sizes) / len(sizes)
        print("   {:4.2f}  {:5.3f}    {:5.3f}       {:5.3f}"
              "         {:4d}     {:7.1f}".format(
                  kt, q, extinction(p), big, sizes[len(sizes) // 2],
                  sum(sizes) / len(sizes)))
    print("\n   (subcritical theory: mean size 1 / (1 - K) = {:.1f}"
          " at K = 0.8)".format(1 / (1 - 0.8)))
    print("\nReading: even at K = 1.15 most single seeds die out within a")
    print("few generations; a supercritical K promises explosion only for")
    print("the minority of seeds that survive the first steps. At K = 1 the")
    print("mean is large but the median cascade is a handful of users.")
#    K     q      dies out   reaches 1,000   median   mean size
#    0.80  0.081    1.000       0.000            1         5.0
#    1.00  0.124    1.000       0.028            3       173.8
#    1.15  0.200    0.729       0.273            3      2736.9
#
#    (subcritical theory: mean size 1 / (1 - K) = 5.0 at K = 0.8)
#
# Reading: even at K = 1.15 most single seeds die out within a
# few generations; a supercritical K promises explosion only for
# the minority of seeds that survive the first steps. At K = 1 the
# mean is large but the median cascade is a handful of users.

K = 0.80 时,每个级联都熄灭,平均规模 5.0,正好是理论值 1/(1−K)1/(1-K)。K = 1.00 时,级联的平均规模很大(173.8),但中位数只有 3 人:平均值由极少数巨大的级联撑起。K = 1.15 时,0.729 的种子仍然熄灭——这正是分支过程的熄灭概率,即方程 η=G(η)\eta = G(\eta) 的最小根——只有 0.273 的种子长到 1,000 人以上。

级联规模的分布:至少达到某一规模的种子比例(双对数坐标)。K = 1.00 附近,规模分布呈重尾;K = 1.15 时,大约四分之一的种子突破上限,其余的在几代内熄灭。

证据

真实的级联大多很小。 对数十亿次在线传播事件的研究发现,绝大多数分享级联在第一层之内就停止,大规模的「病毒式」传播极其罕见,而且大规模事件往往主要来自一次广播式的大量曝光,而不是一代代的人传人 (Goel et al. 2012, 2016)。对推荐营销网络的早期研究同样发现,推荐链条通常很短,平均每次推荐带来的购买很少 (Leskovec et al. 2007)。

假说 · 尚未验证

作者的假说(尚未验证): 在一个真实的分享解锁漏斗中,按推荐链条直接测得的 K 低于 1,网站的增长主要来自少数大型种子(广告或公众号推送),而不是一代代的裂变。什么结果会证明它错:用完整的推荐日志测得的逐代 K 在多个独立时期都显著大于 1,并且增长主要来自第三代及以后的用户。

小结

把「拉人解锁 + 付费分佣」写成分支过程后,K 等于分享率乘以每个分享者带来的平均人数,再加上付费通道的贡献;因为分享者凑够门槛就停,K 被分享率封顶。分享率不高时,任何门槛与接受率都不能让 K 越过 1;门槛存在内部最优。即使 K 超过 1,大多数单个种子仍会熄灭:K = 1.15 时熄灭概率为 0.729,爆炸只属于撑过最初几代的少数种子。

参考文献

Athreya, Krishna B., and Peter E. Ney. 1972. Branching Processes. Springer.
Goel, Sharad, Ashton Anderson, Jake Hofman, and Duncan J. Watts. 2016. “The Structural Virality of Online Diffusion.” Management Science 62 (1): 180–96.
Goel, Sharad, Duncan J. Watts, and Daniel G. Goldstein. 2012. “The Structure of Online Diffusion Networks.” Proceedings of the 13th ACM Conference on Electronic Commerce, 623–38.
Harris, Theodore E. 1963. The Theory of Branching Processes. Springer.
Leskovec, Jure, Lada A. Adamic, and Bernardo A. Huberman. 2007. “The Dynamics of Viral Marketing.” ACM Transactions on the Web 1 (1): 5.

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).