GPT 6 Astra Ultra attempt on Econ-143
economicsEconomics cataloguecontributionexplorationClaim: exploration
Reported by the author as unresolved.
Manuscript
Contents
A Gaussian privacy externality and an implementable disclosure frontier
Status: Unresolved. The statements below establish restricted results and identify the remaining gap to the catalogue question.Abstract
For two correlated Gaussian types, we compute the exact disclosure interval allowed by a nonparticipant's information-loss cap. A strictly concave welfare problem then has a unique solution, and a transfer menu implements it truthfully when one participant's privacy sensitivity is private. The result isolates a nonconsenter externality under a fixed service technology; it does not characterize optimal privacy institutions for general pricing markets.Short target and model
The question is how correlated information changes consent, compensation, and welfare. Let $(X_1,X_2)$ be jointly Gaussian, centered, with unit variances and correlation $r\in(0,1)$. Person 1 may authorize the public signal $Y=X_1+N$, with independent Gaussian noise of variance $\nu$. It is convenient to index quality by $t=(1+\nu)^{-1}\in(0,1)$ and include $t=0$ as an independent, uninformative signal. Person 2 does not consent and receives no transfer.Information loss is mutual information in nats, $c_i(t)=I(X_i;Y)$. Person $i$ values this loss at $\theta_i c_i(t)$, where $\theta_1>0$ is private and $\theta_2\ge0$ is public. Service produces monetizable gross benefit $\beta t$, $\beta>0$, received initially by a platform. Transfers between the platform and person 1 cancel in welfare. There are no additional production costs, downstream price responses, or hidden disclosures. A regulator imposes $c_2(t)\le\rho$, $\rho\ge0$. These assumptions make the accounting and the feasible mechanism explicit.
The exact nonconsenter constraint
Proposition (Privacy frontier): For $0\le t<1$, \[
c_1(t)=-\tfrac12\log(1-t),\qquad
c_2(t)=-\tfrac12\log(1-r^2t).
\] The privacy cap is equivalent to $t\le T$, where \[
T=\min\{1,(1-e^{-2\rho})/r^2\}.
\] The endpoint $t=1$ is excluded whenever $\theta_1>0$, since $c_1(t)$ diverges there. If $\rho=0$, only $t=0$ is feasible, despite person 1's ability to consent.
Proof: Gaussian conditioning gives ${\operatorname{Var}}(X_1\mid Y)=1-1/(1+\nu)=1-t$ and ${\operatorname{Var}}(X_2\mid Y)=1-r^2/(1+\nu)=1-r^2t$. The difference between unconditional and conditional Gaussian differential entropies is one half the logarithm of their variance ratio. This proves the formulas, with their continuous values at $t=0$. Exponentiating the cap inequality proves the interval. The final two assertions follow from these same formulas. ∎
Thus individual consent cannot eliminate the information revealed about a correlated nonparticipant. The conclusion depends on this particular information-loss metric; it is not a universal measure of privacy harm. Welfare optimization and comparative statics
For a reported or actual type $\theta>0$ let \[ W_\theta(t)=\beta t-\theta c_1(t)-\theta_2c_2(t), \qquad 0\le t\le T,\quad t<1. \]Theorem (Unique optimal disclosure): The welfare maximum exists and is unique. It is zero disclosure when $\beta\le(\theta+\theta_2r^2)/2$. Otherwise let $t_0\in(0,1)$ be the unique solution of \[
\beta={\theta\over2(1-t_0)}+
{\theta_2r^2\over2(1-r^2t_0)}.
\] Then $t^*(\theta)=\min\{T,t_0\}$. Optimal quality is weakly decreasing in $\theta$ and $\theta_2$, and weakly increasing in $\beta$ and $\rho$. With the other primitives fixed, it is weakly decreasing in $r\in(0,1)$.
Proof: The second derivative is \[
W_\theta''(t)=-{\theta\over2(1-t)^2}
-{\theta_2r^4\over2(1-r^2t)^2}<0.
\] The first derivative at zero gives the no-disclosure condition. In the other case it is positive at zero and tends to minus infinity at one, so continuity and strict decrease give the unique root. A finite cap truncates the root. If $T=1$, divergence of $\theta c_1$ ensures attainment away from one; otherwise ordinary compactness does so. Increasing either harm weight decreases the derivative at each fixed quality; increasing $\beta$ increases it. These observations and the cap formula prove the first four monotonicities. Increasing $r$ increases $r^2/(1-r^2t)$ and weakly lowers the cap, proving the last one. ∎
The total welfare externality is included once. If a transfer $p$ is paid, person 1's utility is $p-\theta c_1(t)$, person 2's is $-\theta_2c_2(t)$, and platform profit is $\beta t-p$. Their sum is exactly $W_\theta(t)$. A truthful, voluntary implementation
Define $A(t)=\beta t-\theta_2c_2(t)$ and offer all feasible qualities satisfying $A(t)\ge0$, with transfer $p(t)=A(t)$. Include $t=0$, whose transfer is zero. The platform commits to deliver the selected signal quality and transfer.Theorem (Implementation for a privately known harm weight): For every $\theta>0$, person 1's optimal menu choice equals $t^*(\theta)$. The equivalent direct mechanism selecting $t^*(\widehat\theta)$ on report $\widehat\theta>0$, and paying $p(t^*(\widehat\theta))$, is dominant-strategy truthful and individually rational. Its transfers are nonnegative and never exceed gross platform benefit. The nonconsenter's cap holds for every report.
Proof: At any offered quality, the person's utility is $A(t)-\theta c_1(t)=W_\theta(t)$. Every maximizer of welfare has value at least $W_\theta(0)=0$, hence $A(t)\ge\theta c_1(t)\ge0$ and is included in the menu. The unique welfare maximizer therefore remains the optimal menu choice. Under the direct mechanism a false report merely chooses another point from this menu, which cannot improve actual utility. Choosing zero is available and yields zero, proving voluntary participation. Finally $0\le p(t)=\beta t-\theta_2c_2(t)\le\beta t$, and every menu quality obeys the cap by construction. ∎
The transfer rule internalizes the nonparticipant's harm in the participant's choice. It need not maximize the platform's profit among truthful mechanisms, and person 2 is not compensated. If person 2 has a right to refuse any uncompensated loss rather than a supplied cap, this is a different feasible mechanism problem; for zero permitted loss the proposition forces zero quality. Limitations and next step
The proof assumes a known Gaussian prior, a verifiable signal technology, a known $\theta_2$, and a fixed benefit function. It does not solve learning of harm, multiple private types, collective consent, monopolistic downstream pricing, or robustness to misspecified correlation. A next extension would require a specified family of priors and a mechanism whose incentive constraints remain valid across that family. Mere substitution of a worst-case correlation into the welfare formula would not by itself prove robust truthfulness when preferences and information jointly change.References
- [Bonatti] A. Bonatti (2024), The platform dimension of digital privacy, in The Economics of Privacy, Chapter 3, 73--96. https://www.nber.org/books-and-chapters/economics-privacy/platform-dimension-digital-privacy. Source of the broader platform/privacy agenda; the restricted Gaussian calculation and transfer menu here are not claimed as results of that chapter.