5D Library › 订阅制单位经济学、净收入留存与人工智能推理标度律 · English

第六章 网络效应:n²、n log n,以及用户越多越糟的那一点

两条定律

梅特卡夫定律说,网络的价值随用户数的平方增长,因为可能的连接数就是这样增长的;梅特卡夫本人在以太网问世四十年后重新审视了它 (Metcalfe 2013)。Briscoe、Odlyzko 与 Tilly 认为它严重高估了价值:并非所有连接都同样有价值;如果一个用户第 kk 有价值的联系人约值她最有价值联系人的 1/k1/k(齐普夫律),那么每个用户的价值只按 ln⁡n\ln n 增长,整个网络按 nln⁡nn \ln n 增长 (Odlyzko and Tilly 2005; Briscoe et al. 2006)。

推导

# Chapter 6. How network value grows with the number of users n.
#   Metcalfe:        every pair of users is worth the same -> V ~ n (n - 1) /
#   2 ~ n^2 Odlyzko-Tilly:   a user values her k-th most valuable contact at
#   1/k (Zipf), so each
#                    user is worth H(n-1) = 1 + 1/2 + ... + 1/(n-1) ~ ln n ->
#                    V ~ n ln n
# A. Ratio of value when the network grows tenfold.
# B. Two networks of sizes n1 > n2 interconnect. Under each law, how much does
#    each side's value rise? (Briscoe, Odlyzko and Tilly's argument about why
#    the larger network is reluctant.)
import math

def H(m):
    if m < 1:
        return 0.0
    if m < 1000:
        return sum(1.0 / k for k in range(1, m + 1))
    return math.log(m) + 0.5772156649 + 1 / (2 * m)

def v_metcalfe(n):
    return n * (n - 1) / 2

def v_zipf(n):
    return n * H(n - 1)

if __name__ == "__main__":
    print("A. Value multiple when the network grows tenfold")
    print("   from n        Metcalfe   Zipf (n ln n)")
    for n in (100, 10000, 1000000):
        print(
            "   {:>9,}    x {:6.1f}    x {:5.2f}".format(
                n,
                v_metcalfe(10 * n) / v_metcalfe(n),
                v_zipf(10 * n) / v_zipf(n),
            )
        )
    print(
        "\nB. Interconnecting networks of 1,000,000 and 100,000 users: gain"
        " in value"
    )
    print("   law          larger network's   smaller network's   ratio")
    print(
        "                gain               gain                small/large"
    )
    n1, n2 = 1000000, 100000
    # each side's users gain n_other pairs / 2
    g1m = n1 * n2 / 2.0
    g2m = n1 * n2 / 2.0
    g1z = n1 * (H(n1 + n2 - 1) - H(n1 - 1))
    g2z = n2 * (H(n1 + n2 - 1) - H(n2 - 1))
    print(
        "   Metcalfe     {:14,.0f}     {:14,.0f}      {:6.2f}".format(
            g1m, g2m, g2m / g1m
        )
    )
    print(
        "   Zipf         {:14,.0f}     {:14,.0f}      {:6.2f}".format(
            g1z, g2z, g2z / g1z
        )
    )
    print("\nReading: under Metcalfe, tenfold users means"
          " a hundredfold value and both")
    print("sides gain equally from joining. Under the"
          " Zipf law, tenfold users means a")
    print("little over tenfold value, and the smaller"
          " network gains about two and a")
    print("half times as much as the larger one -"
          " consistent with large networks'")
    print("reluctance to interconnect on equal terms.")
# A. Value multiple when the network grows tenfold
#    from n        Metcalfe   Zipf (n ln n)
#          100    x  100.9    x 14.46
#       10,000    x  100.0    x 12.35
#    1,000,000    x  100.0    x 11.60
#
# B. Interconnecting networks of 1,000,000 and 100,000 users: gain in value
#    law          larger network's   smaller network's   ratio
#                 gain               gain                small/large
#    Metcalfe     50,000,000,000     50,000,000,000        1.00
#    Zipf                 95,310            239,790        2.52
#
# Reading: under Metcalfe, tenfold users means a hundredfold value and both
# sides gain equally from joining. Under the Zipf law, tenfold users means a
# little over tenfold value, and the smaller network gains about two and a
# half times as much as the larger one - consistent with large networks'
# reluctance to interconnect on equal terms.

A 段说明两条定律的标度何等不同。网络增长十倍,在梅特卡夫定律下价值乘以 100.0,在齐普夫律下只乘以 11.60 到 14.46。引用「网络效应」的商业计划,说的通常是前者,过的却是后者的日子。

B 段是 Odlyzko 与 Tilly 最犀利的一点。把一个 1,000,000 用户的网络与一个 100,000 用户的网络互联,在梅特卡夫定律下双方受益相等。在齐普夫律下,小网络增益 239,790,大网络增益 95,310——小网络的增益是大网络的 2.52 倍。这种不对称与大网络的实际做法相符:它们抗拒对等互联,或要求为此收费。

当用户越多越糟

两条定律都没有下行面。真实的网络有:更多用户带来更多噪声、垃圾信息、拥堵与不想要的接触,而每个用户的暴露程度随能触达她的其他用户数而增长。第十三段程序把最简单的这类成本——每个用户 cnc\,n——加到齐普夫收益上。

推导

含拥堵的价值。 V(n)=n(bln⁡n−cn)V(n) = n\,(b \ln n – c\,n),在 b(ln⁡n+1)=2cnb(\ln n + 1) = 2\,c\,n 处取最大值。超过这个规模,每增加一个用户都会降低总价值。

推导

# Chapter 6. When more users make a network worse. Each user gets benefit b ln
# n from the network (the Zipf law of p12) and bears a cost c n from
# congestion, noise and spam that grows with the number of other users who can
# reach her. Total value
#     V(n) = n (b ln n - c n),     maximised where b (ln n + 1) = 2 c n.
# b = 1; c is the cost per reachable user. Moderation, filtering and ranking
# lower c.
import math

def V(n, c, b=1.0):
    return n * (b * math.log(n) - c * n)

def n_star(c, b=1.0):
    lo, hi = 2.0, 1e12
    for _ in range(200):
        mid = math.sqrt(lo * hi)
        if b * (math.log(mid) + 1) > 2 * c * mid:
            lo = mid
        else:
            hi = mid
    return lo

if __name__ == "__main__":
    print(
        "   cost per reachable   value-maximising   peak value      value at"
    )
    print("   user c               n                                  3 x n*")
    for c in (1e-4, 1e-5, 1e-6, 1e-7):
        ns = n_star(c)
        print(
            "   {:>10.0e}       {:>14,.0f}   {:>14,.0f}   {:>14,.0f}".format(
                c, ns, V(ns, c), V(3 * ns, c)
            )
        )
    print("\nReading: with any congestion cost that grows"
          " with reach, value peaks and")
    print("then falls; past the peak, adding users"
          " destroys value. Cutting the cost")
    print("per reachable user tenfold raises the peak"
          " size by more than tenfold,")
    print("which is why filtering and ranking are part"
          " of the network effect, not an")
    print("extra.")
#    cost per reachable   value-maximising   peak value      value at
#    user c               n                                  3 x n*
#         1e-04               60,011          300,126       -1,062,659
#         1e-05              724,674        4,526,848      -15,540,180
#         1e-06            8,476,398       63,372,926     -213,040,342
#         1e-07           96,948,450      842,951,741   -2,791,019,647
#
# Reading: with any congestion cost that grows with reach, value peaks and
# then falls; past the peak, adding users destroys value. Cutting the cost
# per reachable user tenfold raises the peak size by more than tenfold,
# which is why filtering and ranking are part of the network effect, not an
# extra.

每个可触达用户的成本为 10−510^{-5} 时,价值在 724,674 个用户处达到峰值;规模扩大到三倍时价值为负。把这一成本降低十倍,到 10−610^{-6},峰值移到 8,476,398——超过十倍,因为收益只按对数增长。所以过滤、排序与内容治理并不是网络效应上的额外开销:它们决定了网络能有用地长到多大。

左:梅特卡夫 n^2 与齐普夫律 n \ln n 下的网络价值。右:每个可触达用户的拥堵成本为 10^{-5} 与 10^{-6} 时的价值:先升后降。

数据如何「证实」梅特卡夫

有几项研究把收入或市值对用户数拟合,得到接近 2 的指数;其中一项用腾讯与 Facebook 的数据拟合,报告它们证实了梅特卡夫定律 (Zhang et al. 2015)。第十四段程序说明在定律并不成立时,这种结果如何出现。真实价值服从 nln⁡nn \ln n,但公司每单位价值的变现能力随时间提高——它不断增加广告、定价与产品——与此同时用户也在增长。

推导

# Chapter 6. How growth data can "confirm" Metcalfe's law when it is false. A
# company grows from 100,000 to 3,000,000 users over 8 years (logistic
# adoption, 32 quarters). Its TRUE network value per user follows the Zipf law
# of p12 (V ~ n ln n), but the revenue it earns per unit of value - its
# monetisation - rises g a year as it adds advertising, pricing and products.
# Quarterly revenue = monetisation x n ln n x noise (lognormal, sd 0.10). An
# analyst regresses log revenue on log users:
#     log Rev = a + beta log n.
# Metcalfe predicts beta = 2; n ln n predicts beta a little above 1.
import math
import random

def ols(x, y):
    n = len(x)
    mx, my = sum(x) / n, sum(y) / n
    b = sum((a - mx) * (c - my) for a, c in zip(x, y)) / sum(
        (a - mx) ** 2 for a in x
    )
    return b

if __name__ == "__main__":
    rng = random.Random(58014)
    print("   monetisation    fitted beta (mean    share with")
    print("   growth / year   of 500 histories)    beta > 1.8")
    for g in (0.0, 0.25, 0.50, 0.75):
        bs = []
        for _ in range(500):
            xs, ys = [], []
            for q in range(32):
                t = q / 4
                n = 100000 + 2900000 / (1 + math.exp(-(t - 4) * 1.1))
                rev = (
                    (1 + g) ** t
                    * n
                    * math.log(n)
                    * math.exp(rng.gauss(0, 0.10))
                )
                xs.append(math.log(n))
                ys.append(math.log(rev))
            bs.append(ols(xs, ys))
        print("   {:>13.0%}   {:12.2f}         {:10.3f}".format(
            g, sum(bs) / len(bs), sum(b > 1.8 for b in bs) / len(bs)))
    print("\nReading: with flat monetisation the regression"
          " recovers the true law (beta")
    print("near 1.1). Let monetisation grow while users"
          " grow and the slope absorbs")
    print("it: revenue-versus-users plots then look like"
          " Metcalfe's law although the")
    print("network itself follows n ln n. Users and time are confounded.")
#    monetisation    fitted beta (mean    share with
#    growth / year   of 500 histories)    beta > 1.8
#               0%           1.08              0.000
#              25%           1.53              0.000
#              50%           1.91              1.000
#              75%           2.23              1.000
#
# Reading: with flat monetisation the regression recovers the true law (beta
# near 1.1). Let monetisation grow while users grow and the slope absorbs
# it: revenue-versus-users plots then look like Metcalfe's law although the
# network itself follows n ln n. Users and time are confounded.

变现能力不变时,把对数收入对对数用户回归,恢复出真实的定律:指数 1.08。变现能力每年增长 50% 时,指数为 1.91,并且在每一次模拟历史中都超过 1.8;75% 时为 2.23。用户与时间相互混杂:任何随时间增长的东西都会加载到用户系数上。除非测量并控制变现能力,否则收入对用户的拟合无法区分梅特卡夫定律与「nln⁡nn \ln n 加上不断增长的变现」。

注

这并没有说什么。 它没有证明所引用的拟合是错的,只说明它们本身无法区分两条定律。检验需要用户数的变化不同时是时间的变化——例如在同一时刻比较规模不同的市场。

小结

在梅特卡夫定律下,用户增长十倍意味着价值增长百倍;在齐普夫(Odlyzko-Tilly)律下只略多于十倍,而且小网络从互联中获得的比大网络更多。若拥堵成本随触达范围增长,价值先升后降,而降低这一成本会使峰值超比例地移动。看似证实梅特卡夫定律的收入-用户拟合,可以产生于 nln⁡nn \ln n 的价值加上不断增长的变现,因为用户与时间同步变动。

参考文献

Briscoe, Bob, Andrew Odlyzko, and Benjamin Tilly. 2006. “Metcalfe’s Law Is Wrong.” IEEE Spectrum 43 (7): 34–39.
Metcalfe, Bob. 2013. “Metcalfe’s Law after 40 Years of Ethernet.” Computer 46 (12): 26–31.
Odlyzko, Andrew, and Benjamin Tilly. 2005. A Refutation of Metcalfe’s Law and a Better Estimate for the Value of Networks and Network Interconnections. Working paper. Digital Technology Center, University of Minnesota.
Zhang, Xing-Zhou, Jing-Jie Liu, and Zhi-Wei Xu. 2015. “Tencent and Facebook Data Validate Metcalfe’s Law.” Journal of Computer Science and Technology 30 (2): 246–51.

Fan, Y. (2026). Subscription Unit Economics, Net Revenue Retention and AI Inference Scaling Laws: Cohort Survival Mixtures, Negative-Churn Differential Equations, Token Compute Margins and Metcalfe Congestion Limits [订阅制单位经济学、净收入留存与人工智能推理标度律] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090164 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).