5D Library › Supercritical Viral Cascades, Information-Gap Hazard Rates and Multi-Tier Paywall Optimization · 中文

Chapter 5. Pareto Willingness to Pay, One Price and the Revenue Ceiling

The heavy tail of willingness to pay

Among people willing to pay for a reading, most will pay little and a few will pay a great deal. The Pareto distribution is the standard model of such a heavy tail: the share whose willingness to pay exceeds xx is (xmin/x)a(x_{\min}/x)^{a}, and the smaller the tail index aa, the heavier the tail. The marginal cost of a digital good is close to zero, so the seller’s problem is purely how to collect as much as possible from this distribution.

Economics divides price discrimination into three degrees (Pigou 1920; Varian 1989): first degree charges each person the most they will pay (perfect discrimination); second degree offers a menu and lets people select themselves (quantities, versions, tiers); third degree prices separately for observable groups. First-degree discrimination sets the ceiling on revenue: the sum of all willingness to pay.

Calculation

One price and the ceiling. At a price p≥xminp \ge x_{\min} revenue is p(xmin/p)ap\,(x_{\min}/p)^{a}. For a>1a > 1 it falls with pp, so the best single price is xminx_{\min} itself, with revenue xminx_{\min}; the ceiling is the mean willingness to pay axmin/(a−1)a\,x_{\min}/(a-1). Their ratio is a/(a−1)a/(a-1): without limit as aa tends to 1, and for a≤1a \le 1 the mean willingness to pay is infinite.

The ninth program computes the best single price, its revenue and the ceiling at several tail indices, and tests a classical theorem.

Calculation

# Chapter 5. Pareto willingness to pay, one price, and the ceiling.
# Among visitors who would pay anything, willingness to pay v is Pareto:
# P(v > x) = (3 / x)^a for x >= 3 (minimum 3, tail index a). Marginal cost
# of a digital reading is taken as 0.
# Part 1. Revenue per buyer-candidate from the best single price, against
# the full surplus E[v] that perfect (first-degree) discrimination would
# take. For a > 1 every price above 3 loses more buyers than it gains in
# price, so the best single price is 3; E[v] = 3a / (a - 1). For a <= 1
# E[v] is infinite; here values are capped at 1,000 to show the size.
# Part 2. Riley and Zeckhauser: for ONE good no menu - not even a menu of
# lotteries "get the reading with probability x for price t" - beats the
# best posted price. A grid search over two-option menus (x, t; 1, p)
# with a = 1.5 checks this, with revenue computed in closed form.
def tail(x, a, cap=None):
    if x <= 3:
        return 1.0
    if cap is not None and x >= cap:
        return 0.0
    return (3 / x) ** a

def mean_v(a, cap=None):
    if cap is None:
        return 3 * a / (a - 1)
    n, lo = 200000, 3.0
    h = (cap - lo) / n
    return lo + sum(tail(lo + (i + 0.5) * h, a, cap) * h for i in range(n))

def best_price(a, cap=None):
    top = cap if cap else 1000.0
    grid = [3 * (top / 3) ** (i / 4000) for i in range(4000)]
    return max((p * tail(p, a, cap), p) for p in grid)

def menu_revenue(x, t, p, a):
    v_low = t / x                  # indifferent between lottery and nothing
    v_high = (p - t) / (1 - x)     # indifferent between lottery and good
    if v_low >= v_high:            # lottery never chosen: posted price p
        return p * tail(p, a)
    return t * (tail(v_low, a) - tail(v_high, a)) + p * tail(v_high, a)

if __name__ == "__main__":
    print("   tail a   best price   revenue   full surplus   ceiling ratio")
    for a, cap in ((3.0, None), (2.0, None), (1.5, None), (1.1, None),
                   (0.8, 1000.0)):
        rev, p = best_price(a, cap)
        ev = mean_v(a, cap)
        print("   {:4.1f}{}    {:7.2f}    {:6.3f}     {:8.3f}       {:6.2f}"
              .format(a, "*" if cap else " ", p, rev, ev, ev / rev))
    print("   (* values capped at 1,000)")
    best = (0.0, None)
    for i in range(1, 40):
        x = i / 40
        for j in range(1, 60):
            t = 0.25 * j * x
            for k in range(1, 60):
                p = 3 + 0.25 * k
                r = menu_revenue(x, t, p, 1.5)
                if r > best[0]:
                    best = (r, (x, t, p))
    print("\n   a = 1.5: best two-option menu earns {:.3f} against {:.3f}"
          " for one price".format(best[0], best_price(1.5)[0]))
    print("\nReading: with a heavy tail the gap between one price and full")
    print("extraction is large - three times at a = 1.5, and without bound")
    print("as a falls to 1; below 1 the best single price chases the few")
    print("buyers at the very top. For a single reading no menu closes the")
    print("gap: splitting one good into tiers or chances cannot beat one")
    print("price.")
#    tail a   best price   revenue   full surplus   ceiling ratio
#     3.0        3.00     3.000        4.500         1.50
#     2.0        3.00     3.000        6.000         2.00
#     1.5        3.00     3.000        9.000         3.00
#     1.1        3.00     3.000       33.000        11.00
#     0.8*     998.55     9.585       35.937         3.75
#    (* values capped at 1,000)
#
#    a = 1.5: best two-option menu earns 2.961 against 3.000 for one price
#
# Reading: with a heavy tail the gap between one price and full
# extraction is large - three times at a = 1.5, and without bound
# as a falls to 1; below 1 the best single price chases the few
# buyers at the very top. For a single reading no menu closes the
# gap: splitting one good into tiers or chances cannot beat one
# price.

The heavier the tail, the wider the gap between one price and the ceiling: at a=3.0a = 3.0 the ceiling is 1.50 times the revenue of one price, at a=1.5a = 1.5 it is 3.00 times, at a=1.1a = 1.1 it is 11.00 times. At a=0.8a = 0.8 (willingness to pay capped at 1,000) the best single price jumps to 998.55 — selling only to the very few at the top.

The last line tests the theorem of Riley and Zeckhauser: for one good, the seller’s optimal mechanism is a posted price, and no menu — not even a lottery menu of the form “receive the reading with probability xx for a payment tt” — does better (Riley and Zeckhauser 1983; Myerson 1981). The best two-option menu in the grid search earns 2.961, below the 3.000 of one price. Splitting one good into tiers cannot, by itself, take more money.

Revenue against a single price under Pareto willingness to pay (logarithmic horizontal axis). For a > 1 revenue is highest at the lowest price; at a = 0.8 it rises all the way to the cap.

Then why does a multi-tier paywall take more?

Because it does not sell one good. The tenth program imagines four paid products after the free hook — a check of past highs and lows, a ten-year overview, a two-person match and a single question — each valued at a common Pareto scale times a product weight times a personal random factor with log-standard-deviation σ\sigma.

Calculation

# Chapter 5. When do several tiers beat one price? When they are
# different goods.
# Four paid products after the free teaser, with weights w:
#   past-events check 1, ten-year overview 3, two-person match 2,
#   one question answered 0.8.
# Buyer i values product k at v_ik = theta_i w_k eps_ik: theta is a common
# Pareto scale (minimum 1, tail 1.5, capped at 200), eps an idiosyncratic
# log-normal factor with log-sd sigma (sigma = 0: everyone ranks the four
# products the same way, in fixed proportions). 20,000 buyers, cost 0.
# Strategies, each with prices chosen by grid search on the sample:
#   one product     sell only the overview, one price
#   four prices     each product at its own price
#   mixed bundle    the four single prices held at their separate optimum,
#                   plus a bundle price f * (sum of the four); only the
#                   bundle discount factor f is searched
# The ceiling is the total value (perfect discrimination).
import math
import random

W = (1.0, 3.0, 2.0, 0.8)

def sample(sigma, rng, n=20000):
    out = []
    for _ in range(n):
        th = min((1 - rng.random()) ** (-1 / 1.5), 200.0)
        out.append([th * w * math.exp(rng.gauss(0, sigma)) for w in W])
    return out

def best_single(vals):
    vs = sorted(vals, reverse=True)
    best = (0.0, 0.0)
    for i, v in enumerate(vs):
        best = max(best, (v * (i + 1), v))
    return best

def mixed(buyers, prices):
    # single prices stay at their separate optimum; search f only
    tot = [sum(b) for b in buyers]
    best = (0.0, None)
    for f in [0.5 + 0.02 * i for i in range(26)]:
        pb = f * sum(prices)
        rev = 0.0
        for b, t in zip(buyers, tot):
            sep = sum(max(v - p, 0.0) for v, p in zip(b, prices))
            if t - pb >= sep and t - pb >= 0:
                rev += pb
            else:
                rev += sum(p for v, p in zip(b, prices) if v >= p)
        best = max(best, (rev, pb))
    return best

if __name__ == "__main__":
    rng = random.Random(63010)
    print("   sigma   one product   four prices   mixed bundle   ceiling"
          "   (per buyer)")
    for sigma in (0.0, 0.5, 1.0):
        buyers = sample(sigma, rng)
        n = len(buyers)
        one = best_single([b[1] for b in buyers])[0] / n
        four, prices = 0.0, []
        for k in range(4):
            r, p = best_single([b[k] for b in buyers])
            four += r / n
            prices.append(p)
        mix = mixed(buyers, prices)[0] / n
        ceil = sum(sum(b) for b in buyers) / n
        print("   {:3.1f}       {:6.3f}        {:6.3f}        {:6.3f}"
              "       {:6.3f}".format(sigma, one, four, mix, ceil))
        print("               ({:.3f})       ({:.3f})       ({:.3f})"
              "       share of ceiling".format(one / ceil, four / ceil,
                                               mix / ceil))
    print("\nReading: more tiers earn more because they are more goods, not")
    print("because they screen one good. When the four values move together")
    print("(sigma = 0) a bundle adds nothing to four separate prices; the")
    print("more buyers differ in which product they want, the more a bundle")
    print("adds. Even then, well over half of the value stays with buyers.")
#    sigma   one product   four prices   mixed bundle   ceiling   (per buyer)
#    0.0        3.000         6.800         6.800       20.040
#                (0.150)       (0.339)       (0.339)       share of ceiling
#    0.5        2.649         5.995         6.321       21.458
#                (0.123)       (0.279)       (0.295)       share of ceiling
#    1.0        3.163         7.177         7.808       32.611
#                (0.097)       (0.220)       (0.239)       share of ceiling
#
# Reading: more tiers earn more because they are more goods, not
# because they screen one good. When the four values move together
# (sigma = 0) a bundle adds nothing to four separate prices; the
# more buyers differ in which product they want, the more a bundle
# adds. Even then, well over half of the value stays with buyers.

Selling only the ten-year overview takes 0.150 of the value ceiling. Pricing the four products separately raises this to 0.339 of the ceiling — the extra money comes from selling three more goods, not from screening the same good. The mixed bundle keeps each single price at its separate optimum and searches only the bundle discount factor ff (the bundle price is ff times the sum of the four single prices). When everyone’s values for the four products are exactly proportional (σ=0\sigma = 0), adding a price for all four together adds nothing (6.800 against 6.800); the more people’s preferences across products differ, the more a bundle helps (from 7.177 to 7.808 at σ=1.0\sigma = 1.0). Even so, more than three quarters of the value stays with the buyers.

Note

The legal line between second- and first-degree discrimination (to be checked locally). Letting users choose a tier from the same public menu is second-degree discrimination; quoting different prices for the same good to different people on the basis of their personal data comes close to first degree. Article 24 of China’s Personal Information Protection Law requires automated decisions that use personal information to be fair and impartial, and forbids unreasonable differential treatment of individuals in transaction terms such as price (Standing Committee of the National People’s Congress 2021). The book’s models discuss only public menus.

Summary

Under Pareto willingness to pay, the gap between one price and perfect discrimination widens as the tail grows heavier — three times at a tail index of 1.5, without limit as it tends to 1. But for one good no menu beats one price: the best two-option menu earns 2.961 against 3.000. A multi-tier paywall takes more because each tier is a different good; a bundle adds revenue only when people’s preferences across products differ, and most of the value stays with the buyers.

References

Myerson, Roger B. 1981. “Optimal Auction Design.” Mathematics of Operations Research 6 (1): 58–73.
Pigou, A. C. 1920. The Economics of Welfare. Macmillan.
Riley, John, and Richard Zeckhauser. 1983. “Optimal Selling Strategies: When to Haggle, When to Hold Firm.” Quarterly Journal of Economics 98 (2): 267–89.
Standing Committee of the National People’s Congress. 2021. Personal Information Protection Law of the People’s Republic of China, Article 24. In Chinese.
Varian, Hal R. 1989. “Price Discrimination.” In Handbook of Industrial Organization, edited by Richard Schmalensee and Robert Willig, vol. 1. North-Holland.

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).