Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2 — impact of imposing a little fairness on social welfare

Disproved
CohenPriceFairness.LocalWelfare.theorem_2

by Shuze Chen · Oct 29, 2026 · Mathlib 0df444a (Lean v4.33.1)

fairnessoperations-researchp2o-gran-per-chapterp2o-plan-paperp2o-v1pricing

Consider a seller with unlimited supply, two groups of positive real population masses did_idi​, common marginal cost c≥0c\ge0c≥0, and all nonnegative real prices. Both groups have either uniform valuations on [0,bi][0,b_i][0,bi​] with bi>0b_i>0bi​>0, b0≠b1b_0\ne b_1b0​=b1​, or exponential valuations of rates λi>0\lambda_i>0λi​>0, λ0≠λ1\lambda_0\ne\lambda_1λ0​=λ1​. Purchase occurs at valuation at least price. Profit, consumer surplus and nonbuyer valuation are derived from these laws.

There exists an unconstrained global optimum for each group. For every pair of baseline optima, and separately for each of the four metrics, local selections of global profit-maximizing fair price pairs anchored at that baseline exist. For every such selection, the finite right derivative at zero of social welfare exists and obeys the printed strict signs:

Dprice>0,Dnonbuyer>0,Ddemand<0,Dsurplus<0.D_{\rm price}>0,\qquad D_{\rm nonbuyer}>0,\qquad D_{\rm demand}<0,\qquad D_{\rm surplus}<0.Dprice​>0,Dnonbuyer​>0,Ddemand​<0,Dsurplus​<0.

Here fairness means ∣M0(p0)−M1(p1)∣≤(1−α)∣M0(p0∗)−M1(p1∗)∣|M_0(p_0)-M_1(p_1)|\le(1-\alpha)|M_0(p_0^*)-M_1(p_1^*)|∣M0​(p0​)−M1​(p1​)∣≤(1−α)∣M0​(p0∗​)−M1​(p1∗​)∣ with α∈[0,1]\alpha\in[0,1]α∈[0,1], and welfare is total profit plus d0S0(p0)+d1S1(p1)d_0S_0(p_0)+d_1S_1(p_1)d0​S0​(p0​)+d1​S1​(p1​). All eight family/metric cases are included, without excluding zero cost, high cost, zero baseline gaps, below-cost prices, or absent sales.

This is the source's claim, retained as a statement to resolve. The counterexamples in the accompanying insights are not removed by extra hypotheses or by weakening a strict sign.

Differences from the source:

  • encoding: two customer populations → two positive real masses indexed by Fin 2, without normalization; reason: two nonempty groups are implicit on pp. 6–7; this retains unequal and noninteger populations
  • reading: the total-surplus displays on pp. 8 and 12 repeat S₁ → total surplus is d₀S₀(p₀)+d₁S₁(p₁); reason: the group welfare definition on p. 7 and proof on p. 31 establish the intended index; insight I2
  • reading: conditional mean Nᵢ is undefined when there are no nonbuyers → Nᵢ is explicitly zero when the conditioning event has probability zero; reason: Eq. (1) admits zero prices; this is the right-continuous extension for the goal families and changes no ordinary conditional mean; insight I5
  • reading: optimal prices are named without a tie rule → every baseline optimum and every local optimal selection anchored there are covered, with constrained-selection existence explicit; reason: the tie-independent reading prevents a convenient optimizer choice or an absent selection from hiding a case
  • encoding: the right derivative is written as a finite right limit → existence of a real D and convergence of the difference quotient from strictly positive alpha; reason: the footnote 7 limit is preserved without using a default-valued derivative at a nonexistent limit
  • encoding: unconstrained optimal prices are denoted by argmax → existence of a pair of global baseline maximizers is an explicit conclusion; reason: this exposes the existence implicit in the source instead of assuming it or letting universal baseline quantification be vacuous
Preamble
import Mathlib
import Definitions.Def_CohenPriceFairness_LocalWelfare_Sensitivity
import Definitions.Def_CohenPriceFairness_LocalWelfare_Families
Formal statement
namespace CohenPriceFairness.LocalWelfare

/-- The full printed Theorem 2 (p. 12), including its strict signs at zero and high costs. -/
theorem theorem_2 (m : Market) (hd : ∀ i, 0 < m.population i)
    (hc : 0 ≤ m.cost) (hf : Families m) :
    (∃ b : Prices, BaselineOptimal m b) ∧
    ∀ b : Prices, BaselineOptimal m b →
      SensitivityClaim m b .price (fun D => 0 < D) ∧
      SensitivityClaim m b .nonbuyer (fun D => 0 < D) ∧
      SensitivityClaim m b .demand (fun D => D < 0) ∧
      SensitivityClaim m b .surplus (fun D => D < 0) := by sorry

end CohenPriceFairness.LocalWelfare

/- Imported definition source (verbatim dependency snapshot for standalone inspection).

import Mathlib

open MeasureTheory

namespace CohenPriceFairness.LocalWelfare

/-- A nonnegative valuation law with a finite first moment (pp. 6–7). -/
def ValuationLaw (μ : Measure ℝ) : Prop :=
  IsProbabilityMeasure μ ∧ (∀ᵐ v ∂μ, 0 ≤ v) ∧ Integrable (fun v : ℝ => v) μ

end CohenPriceFairness.LocalWelfare

import Mathlib
import Definitions.Def_CohenPriceFairness_LocalWelfare_ValuationLaw

open MeasureTheory

namespace CohenPriceFairness.LocalWelfare

/-- Two marginal valuation laws, real population masses, and common marginal cost. -/
structure Market where
  law : Fin 2 → Measure ℝ
  population : Fin 2 → ℝ
  cost : ℝ

/-- The general distribution-based market conventions, pp. 6–7. -/
def Market.Admissible (m : Market) : Prop :=
  0 ≤ m.cost ∧ ∀ i, 0 < m.population i ∧ ValuationLaw (m.law i)

noncomputable def demand (m : Market) (i : Fin 2) (p : ℝ) : ℝ :=
  (m.law i (Set.Ici p)).toReal

noncomputable def surplus (m : Market) (i : Fin 2) (p : ℝ) : ℝ :=
  ∫ v, max (v - p) 0 ∂m.law i

/-- Conditional mean of nonbuyers; explicitly zero on a null conditioning event. -/
noncomputable def nonbuyer (m : Market) (i : Fin 2) (p : ℝ) : ℝ :=
  if m.law i (Set.Iio p) = 0 then 0
  else (∫ v in Set.Iio p, v ∂m.law i) / (m.law i (Set.Iio p)).toReal

noncomputable def profit (m : Market) (i : Fin 2) (p : ℝ) : ℝ :=
  m.population i * (p - m.cost) * demand m i p

abbrev Prices := Fin 2 → ℝ

noncomputable def totalProfit (m : Market) (p : Prices) : ℝ :=
  profit m 0 (p 0) + profit m 1 (p 1)

/-- Group-specific surplus indices, as on pp. 7 and 31. -/
noncomputable def welfare (m : Market) (p : Prices) : ℝ :=
  totalProfit m p + m.population 0 * surplus m 0 (p 0) +
    m.population 1 * surplus m 1 (p 1)

end CohenPriceFairness.LocalWelfare

import Mathlib
import Definitions.Def_CohenPriceFairness_LocalWelfare_Market

namespace CohenPriceFairness.LocalWelfare

inductive Metric where
  | price | demand | surplus | nonbuyer
  deriving DecidableEq

noncomputable def metricValue (m : Market) : Metric → Fin 2 → ℝ → ℝ
  | .price => fun _ p => p
  | .demand => demand m
  | .surplus => surplus m
  | .nonbuyer => nonbuyer m

/-- All individual global maximizers on nonnegative prices. -/
def BaselineOptimal (m : Market) (b : Prices) : Prop :=
  ∀ i, 0 ≤ b i ∧ ∀ p : ℝ, 0 ≤ p → profit m i p ≤ profit m i (b i)

noncomputable def gap (m : Market) (k : Metric) (p : Prices) : ℝ :=
  |metricValue m k 0 (p 0) - metricValue m k 1 (p 1)|

/-- Equation (1), including its full parameter and price domains. -/
def FairFeasible (m : Market) (b : Prices) (k : Metric) (α : ℝ) (p : Prices) : Prop :=
  α ∈ Set.Icc 0 1 ∧ (∀ i, 0 ≤ p i) ∧ gap m k p ≤ (1 - α) * gap m k b

/-- The optimum set is defined by comparison against every feasible price pair. -/
def FairOptimal (m : Market) (b : Prices) (k : Metric) (α : ℝ) (p : Prices) : Prop :=
  FairFeasible m b k α p ∧
    ∀ q : Prices, FairFeasible m b k α q → totalProfit m q ≤ totalProfit m p

/-- A selection of actual optima on some right neighborhood, anchored at the baseline.
No continuity or differentiability, and no restriction to active price neighborhoods. -/
def LocalOptimalSelection (m : Market) (b : Prices) (k : Metric)
    (P : ℝ → Prices) : Prop :=
  P 0 = b ∧ ∃ ε : ℝ, 0 < ε ∧ ε ≤ 1 ∧
    ∀ α ∈ Set.Icc 0 ε, FairOptimal m b k α (P α)

end CohenPriceFairness.LocalWelfare

import Mathlib
import Definitions.Def_CohenPriceFairness_LocalWelfare_FairPricing

open Filter Topology

namespace CohenPriceFairness.LocalWelfare

/-- The finite right derivative in footnote 7, p. 12. -/
def RightDerivative (f : ℝ → ℝ) (D : ℝ) : Prop :=
  Tendsto (fun α : ℝ => (f α - f 0) / α) (nhdsWithin 0 (Set.Ioi 0)) (𝓝 D)

/-- Differentiability in a neighborhood and differentiability of the derivative at x.
Continuity at x follows; continuity of the second derivative is not assumed. -/
def TwiceDifferentiableAt (f : ℝ → ℝ) (x : ℝ) : Prop :=
  (∀ᶠ y in 𝓝 x, DifferentiableAt ℝ f y) ∧ DifferentiableAt ℝ (deriv f) x

/-- Monotonicity and an inverse on the image of an open active interval at x.
No derivative condition on the inverse or nonzero derivative is imposed. -/
def ActiveInvertible (f : ℝ → ℝ) (x : ℝ) : Prop :=
  ∃ l u : ℝ, l < x ∧ x < u ∧
    (MonotoneOn f (Set.Ioo l u) ∨ AntitoneOn f (Set.Ioo l u)) ∧
    ∃ g : ℝ → ℝ,
      (∀ p ∈ Set.Ioo l u, g (f p) = p) ∧
      (∀ y ∈ f '' Set.Ioo l u, g y ∈ Set.Ioo l u ∧ f (g y) = y)

/-- Exactly the local regularity of Lemma 1 for all four source functions. -/
def LemmaRegularity (m : Market) (b : Prices) : Prop :=
  ∀ i,
    TwiceDifferentiableAt (demand m i) (b i) ∧
    TwiceDifferentiableAt (surplus m i) (b i) ∧
    TwiceDifferentiableAt (nonbuyer m i) (b i) ∧
    TwiceDifferentiableAt (profit m i) (b i) ∧
    ActiveInvertible (demand m i) (b i) ∧
    ActiveInvertible (surplus m i) (b i) ∧
    ActiveInvertible (nonbuyer m i) (b i)

/-- Existence of a policy, and the specified finite-derivative assertion for every policy. -/
def SensitivityClaim (m : Market) (b : Prices) (k : Metric) (C : ℝ → Prop) : Prop :=
  (∃ P : ℝ → Prices, LocalOptimalSelection m b k P) ∧
    ∀ P : ℝ → Prices, LocalOptimalSelection m b k P →
      ∃ D : ℝ, RightDerivative (fun α => welfare m (P α)) D ∧ C D

end CohenPriceFairness.LocalWelfare

import Mathlib
import Definitions.Def_CohenPriceFairness_LocalWelfare_Market

open MeasureTheory ProbabilityTheory

namespace CohenPriceFairness.LocalWelfare

/-- Uniform probability on [0,b] for positive b: normalized restricted Lebesgue measure. -/
noncomputable def uniformLaw (b : ℝ) : Measure ℝ :=
  (ENNReal.ofReal b)⁻¹ • volume.restrict (Set.Icc 0 b)

/-- Both groups use the same one of the two continuous families, with unequal parameters. -/
def Families (m : Market) : Prop :=
  (∃ b : Fin 2 → ℝ, (∀ i, 0 < b i) ∧ b 0 ≠ b 1 ∧
    ∀ i, m.law i = uniformLaw (b i)) ∨
  (∃ rate : Fin 2 → ℝ, (∀ i, 0 < rate i) ∧ rate 0 ≠ rate 1 ∧
    ∀ i, m.law i = expMeasure (rate i))

end CohenPriceFairness.LocalWelfare
-/
Source
Cohen, Elmachtoub and Lei (2022), Price Discrimination with Fairness Constraints, Management Science, https://doi.org/10.1287/mnsc.2022.4317; published artifact, PDF pages = printed pages; SHA-256 b3f8804f1610eb36d8a2ef32e836f4a65285e6b8ad427f7c82172e7a3fd73147; §2.3, Theorem 2, p. 12; model §§2–2.1, pp. 6–8; proof p. 33.
Read-back

What the Lean code literally says, in plain math · claude-opus-5-5

The statement concerns a two-group market mmm, with groups indexed by i∈{0,1}i \in \{0,1\}i∈{0,1}. Each group has:

  • a Borel measure μi\mu_iμi​ on R\mathbb{R}R, its valuation law;
  • a real number NiN_iNi​, its population mass.

There is also a single real number ccc, the common cost.

Derived quantities. For a group iii and a real price ppp:

  • Demand is di(p)=μi([p,∞))d_i(p) = \mu_i([p,\infty))di​(p)=μi​([p,∞)), converted to a real number. The interval includes the endpoint ppp.
  • Surplus is Si(p)=∫max⁡(v−p, 0) dμi(v)S_i(p) = \int \max(v-p,\,0)\,d\mu_i(v)Si​(p)=∫max(v−p,0)dμi​(v), taken as a Bochner integral. This means it is 000 whenever the integrand is not integrable.
  • Nonbuyer mean is
Mi(p)={0if μi((−∞,p))=0,∫(−∞,p)v dμi(v)μi((−∞,p))otherwise.M_i(p) = \begin{cases} 0 & \text{if } \mu_i((-\infty,p)) = 0,\\[2pt] \dfrac{\int_{(-\infty,p)} v\,d\mu_i(v)}{\mu_i((-\infty,p))} & \text{otherwise.}\end{cases}Mi​(p)=⎩⎨⎧​0μi​((−∞,p))∫(−∞,p)​vdμi​(v)​​if μi​((−∞,p))=0,otherwise.​

The interval here is open at ppp, and the denominator is converted to a real number.

  • Group profit is πi(p)=Ni (p−c) di(p)\pi_i(p) = N_i\,(p-c)\,d_i(p)πi​(p)=Ni​(p−c)di​(p).

For a price pair p=(p0,p1)p=(p_0,p_1)p=(p0​,p1​):

  • Total profit is Π(p)=π0(p0)+π1(p1)\Pi(p) = \pi_0(p_0)+\pi_1(p_1)Π(p)=π0​(p0​)+π1​(p1​).
  • Welfare is
W(p)=Π(p)+N0 S0(p0)+N1 S1(p1).W(p) = \Pi(p) + N_0\,S_0(p_0) + N_1\,S_1(p_1).W(p)=Π(p)+N0​S0​(p0​)+N1​S1​(p1​).

Metrics and gaps. There are four metrics kkk, each with a per-group value xik(p)x^k_i(p)xik​(p):

  • price: xiprice(p)=px^{\mathrm{price}}_i(p) = pxiprice​(p)=p;
  • demand: di(p)d_i(p)di​(p);
  • surplus: Si(p)S_i(p)Si​(p);
  • nonbuyer: Mi(p)M_i(p)Mi​(p).

The gap of a price pair is Gk(p)=∣x0k(p0)−x1k(p1)∣G_k(p) = |x^k_0(p_0) - x^k_1(p_1)|Gk​(p)=∣x0k​(p0​)−x1k​(p1​)∣.

Hypotheses. The theorem assumes exactly the following, and nothing else:

  1. N0>0N_0 > 0N0​>0 and N1>0N_1 > 0N1​>0.

  2. c≥0c \ge 0c≥0. No upper bound is placed on ccc.

  3. One of two family conditions holds:

    • Uniform: there are β0,β1>0\beta_0, \beta_1 > 0β0​,β1​>0 with β0≠β1\beta_0 \ne \beta_1β0​=β1​ such that, for each iii, μi\mu_iμi​ is βi−1\beta_i^{-1}βi−1​ times Lebesgue measure restricted to [0,βi][0,\beta_i][0,βi​]. This is the uniform probability distribution on [0,βi][0,\beta_i][0,βi​].
    • Exponential: there are rates r0,r1>0r_0, r_1 > 0r0​,r1​>0 with r0≠r1r_0 \ne r_1r0​=r1​ such that, for each iii, μi\mu_iμi​ is the exponential distribution with rate rir_iri​. This is Lebesgue measure with density rie−rivr_i e^{-r_i v}ri​e−ri​v for v≥0v \ge 0v≥0 and 000 for v<0v<0v<0.

    Both groups belong to the same family, with unequal parameters.

Baseline optimality. A price pair bbb is baseline-optimal if, for each iii:

  • bi≥0b_i \ge 0bi​≥0, and
  • πi(p)≤πi(bi)\pi_i(p) \le \pi_i(b_i)πi​(p)≤πi​(bi​) for every real p≥0p \ge 0p≥0.

In words, each bib_ibi​ is a global maximizer of its own group's profit over nonnegative prices.

Conclusion. The theorem asserts two things:

  • (a) At least one baseline-optimal bbb exists.
  • (b) For every baseline-optimal bbb, a sensitivity claim holds for each of four metric/sign combinations:
    • price, with D>0D>0D>0;
    • nonbuyer, with D>0D>0D>0;
    • demand, with D<0D<0D<0;
    • surplus, with D<0D<0D<0.

Feasibility and fair optimality. Fix the baseline bbb and a metric kkk. Call a price pair ppp feasible at level α\alphaα if all three of the following hold:

  • α∈[0,1]\alpha \in [0,1]α∈[0,1];
  • p0≥0p_0 \ge 0p0​≥0 and p1≥0p_1 \ge 0p1​≥0;
  • Gk(p)≤(1−α) Gk(b)G_k(p) \le (1-\alpha)\,G_k(b)Gk​(p)≤(1−α)Gk​(b).

Call ppp fair-optimal at level α\alphaα if it is feasible at level α\alphaα and Π(q)≤Π(p)\Pi(q) \le \Pi(p)Π(q)≤Π(p) for every qqq feasible at level α\alphaα.

Local optimal selection. A map α↦P(α)\alpha \mapsto P(\alpha)α↦P(α) from reals to price pairs is a local optimal selection if:

  • P(0)=bP(0) = bP(0)=b, and
  • there exists ε\varepsilonε with 0<ε≤10<\varepsilon\le 10<ε≤1 such that P(α)P(\alpha)P(α) is fair-optimal at level α\alphaα for every α∈[0,ε]\alpha\in[0,\varepsilon]α∈[0,ε].

No continuity of PPP is required. Values of PPP outside [0,ε][0,\varepsilon][0,ε] are unconstrained.

Sensitivity claim. The sensitivity claim for (k,sign)(k, \text{sign})(k,sign) asserts two things:

  • (i) At least one local optimal selection exists.
  • (ii) For every local optimal selection PPP, there is a finite real DDD with the stated sign such that
lim⁡α→0+W(P(α))−W(P(0))α=D.\lim_{\alpha\to 0^+} \frac{W(P(\alpha)) - W(P(0))}{\alpha} = D.α→0+lim​αW(P(α))−W(P(0))​=D.

The limit is taken over α>0\alpha>0α>0 only, and W(P(0))=W(b)W(P(0)) = W(b)W(P(0))=W(b).

So the theorem asserts all of the following:

  • the existence of fair-optimal prices for every α\alphaα in some interval [0,ε][0,\varepsilon][0,ε], for each of the four metrics and each baseline bbb;
  • that the right derivative of welfare exists along every such selection;
  • that this right derivative is strictly positive for the price and nonbuyer metrics;
  • that it is strictly negative for the demand and surplus metrics.

Degenerate cases. The index set always has exactly two elements, so there are no empty or one-element cases. The parameter α\alphaα only ranges over [0,ε]⊆[0,1][0,\varepsilon]\subseteq[0,1][0,ε]⊆[0,1]. Division by α\alphaα in the difference quotient never occurs at α=0\alpha=0α=0, because the limit is taken only over α>0\alpha>0α>0.

In both families, μi((−∞,p))=0\mu_i((-\infty,p))=0μi​((−∞,p))=0 for every p≤0p\le 0p≤0. So the nonbuyer mean takes its default value Mi(p)=0M_i(p)=0Mi​(p)=0 at all nonpositive prices, including p=0p=0p=0.

The surplus integrals are of integrable functions in both families, so the Bochner default of 000 does not arise. The measures are finite, so the conversions of measures to real numbers are exact.

The cost c=0c=0c=0 is permitted. Arbitrarily large ccc is also permitted. For example, in the uniform family with c≥βic \ge \beta_ic≥βi​, the profit πi\pi_iπi​ is at most 000 on all of [0,∞)[0,\infty)[0,∞) and equals 000 at every p≥βip\ge\beta_ip≥βi​. Every such price is then a valid bib_ibi​, and the conclusion must hold for each such choice of bbb.

If bbb has Gk(b)=0G_k(b)=0Gk​(b)=0, the constraint at every level α\alphaα reduces to Gk(p)=0G_k(p)=0Gk​(p)=0. This happens, for instance, when b0=b1b_0=b_1b0​=b1​ under the price metric. In that case bbb itself is fair-optimal at every α∈[0,1]\alpha\in[0,1]α∈[0,1], because it maximizes each πi\pi_iπi​ separately. Hence the constant map P(α)=bP(\alpha)=bP(α)=b is a local optimal selection, and its welfare difference quotient is identically 000. The statement nonetheless requires the strict sign condition on DDD for that selection too.

Human review
  • Endorsed by Community (Bot) · Oct 29, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Shuze Chen · Oct 29, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me