# The Privacy Subsidy: Kyle's $λ$ under Noise-Perturbed Order-Flow Observation

Crypto.world

> Privacy-preserving exchanges change what a pricing mechanism is allowed to see about order flow — through MPC, batching, sealed bids, or explicit noise injection.

**arXiv**: 2605.15746

**Published**: 2026-05-15

**Authors**: Yuki Nakamura

## AI summary

## Overview

**Research area:** Market microstructure and game theory, applied to privacy-preserving cryptocurrency exchange design (DeFi, ZK/shielded AMMs, differential privacy).

**Technical level:** Advanced. The paper is short and the prose is readable, but the derivations rely on Gaussian projection formulas, the classical Kyle (1985) framework, and σ-algebra/coarse-signal reasoning.

**Scope:** A single-period Kyle model in which the market maker prices on a privacy-noised order-flow signal, yielding a closed-form linear equilibrium and a closed-form per-period welfare transfer (the "privacy subsidy") from the protocol's liquidity pool to traders. Author: Yuki Nakamura (The Open University of Japan), arXiv:2605.15746v5 [cs.GT], 31 May 2026, CC BY 4.0.

## What This Paper Is About

Privacy-preserving exchanges change what a pricing mechanism is allowed to see about order flow — through MPC, batching, sealed bids, or explicit noise injection. The paper asks what happens to the classic Kyle equilibrium when a committed, Bayesian market maker prices on order flow that has been perturbed by independent Gaussian privacy noise rather than the true flow. Its goal is to quantify, in closed form, who pays for that privacy and how much an exchange of this kind must charge in fees to break even.

## Key Contributions

1. **A closed-form linear Kyle equilibrium under a noisy signal.** For a market maker observing ỹ = y + ε with ε ~ N(0, σ_ε²) independent of (v, u), the unique linear equilibrium is λ = σ_v/(2√(σ_u² + σ_ε²)) and β = √(σ_u² + σ_ε²)/σ_v, recovering textbook Kyle exactly at σ_ε = 0.
2. **A welfare decomposition identifying the "privacy subsidy."** A per-period transfer |π_M| = σ_v σ_ε²/(2√(σ_u² + σ_ε²)) from the protocol/LP pool to traders, identically zero in textbook Kyle, which the paper frames as the break-even fee any privacy-aggregated exchange must charge.
3. **A mapping to shielded AMMs with explicit additive (differential-privacy) noise injection**, plus an explicit boundary of applicability: which neighbouring ZK exchange designs the framework does and does not cover.
4. **A stated positioning against Loss-Versus-Rebalancing (LVR).** The privacy subsidy is presented as the single-period, closed-form analog of LVR (Milionis et al.), isolating a cost from noise-obscured flow observation rather than from stale prices.

## Main Findings

- **Price impact falls, informed trading rises, and their product is invariant.** ∂λ/∂σ_ε < 0 and ∂β/∂σ_ε > 0, but λβ = 1/2 for all σ_ε ≥ 0. The paper calls this the persistence of Kyle's "half-revealing" property under privacy noise, and stresses it is a robustness statement about a known result rather than a new quantity.
- **The price distribution conditional on v is unchanged by privacy.** Substituting the equilibrium gives p = (p₀ + v)/2 + λ(u + ε), so E[p | v] = (p₀ + v)/2 and Var(p | v) = λ²(σ_u² + σ_ε²) = σ_v²/4, both independent of σ_ε. The cost of privacy is therefore invisible at the level of the price's conditional distribution.
- **The subsidy is intrinsic to coarse-signal pricing, not to the modeling label.** Remark 2 separates two notions of zero profit: pricing at E[v | ỹ] is Bertrand-competitive given the signal, but since y = x + u is not measurable with respect to σ(ỹ) when σ_ε > 0, it loses E[(p − v)y] = −λσ_ε² = −|π_M| on the real flow. A rule that instead breaks even unconditionally requires λ̃ = σ_v/(2σ_u), independent of σ_ε, but over-reacts to ỹ and is not implementable by a price-taking maker that observes only ỹ. Informational efficiency and zero profit against real flow are incompatible whenever σ_ε > 0.
- **Trader P&L moves in opposite directions from the protocol's.** π_I = +½σ_v√(σ_u² + σ_ε²), π_N = −σ_v σ_u²/(2√(σ_u² + σ_ε²)), π_M = −σ_v σ_ε²/(2√(σ_u² + σ_ε²)), with π_I + π_N + π_M = 0.
- **The subsidy is increasing and has a single inflection point.** It is strictly increasing in σ_ε on [0, ∞); for small σ_ε it behaves as (σ_v/2σ_u)σ_ε² + O(σ_ε⁴), so doubling the privacy parameter quadruples the LP cost; for large σ_ε it grows as ½σ_vσ_ε + O(σ_ε⁻¹). It is convex up to σ_ε* = √2 σ_u and concave beyond it, so past that threshold additional privacy is less expensive at the margin.
- **Noise traders also gain, not at the informed trader's expense.** ∂π_N/∂σ_ε > 0, so uninformed traders lose less as privacy rises. To leading order around σ_ε = 0, both trader types gain equally: π_I(σ_ε) − π_I(0) = π_N(σ_ε) − π_N(0) = (σ_v/4σ_u)σ_ε² + O(σ_ε⁴). In the high-privacy limit the informed trader captures essentially all of it (π_I ~ σ_vσ_ε/2) while the noise trader's gain saturates at σ_vσ_u/2.
- **A volume-proportional break-even fee is exactly welfare-neutral at the partial-equilibrium level.** Using expected absolute volumes E|x| = (σ_v/2λ)√(2/π) and E|u| = σ_u√(2/π), total volume is Q = √(2/π)(√(σ_u² + σ_ε²) + σ_u), and f = |π_M|/Q. Each side's fee exactly equals its incremental gain over the σ_ε = 0 baseline, and net-of-fees both traders revert to classical Kyle values (π_I − f·E|x| = σ_vσ_u/2 and π_N − f·E|u| = −σ_vσ_u/2) while the market maker is exactly compensated.
- **Batching produces no subsidy under this framework.** For Penumbra-style batched swaps, one batch of τ pooled noise-trader draws gives λ = σ_v/(2σ_u√τ) and β = σ_u√τ/σ_v — textbook Kyle with σ_u → σ_u√τ — and π_M = 0 exactly, because the maker prices on and trades against the same observable Y_τ.
- **Three nearby designs fall outside the model.** Sealed-bid order-flow auctions (Suave-style) create temporal rather than additive observation asymmetry and sit closer to LVR; oracle-pegged crossings (Renegade-style) consume no flow signal to compute price and fall outside any Kyle-style analysis; batched swaps reduce to textbook Kyle with rescaling rather than to the noisy-signal model.
- **An illustrative calibration is reported.** Appendix 0.A tabulates a BTC/USDT fee floor in USD per day; at σ_ε = σ_u the LP/protocol must collect approximately $1M/day, described as comparable to the entire revenue from a 0.1% fee on $1B of daily volume.

## Methodology in Plain English

The paper builds a deliberately stripped-down three-player model. An asset has a true terminal value v that only an informed trader can see. Uninformed "noise" traders submit random order flow u. The exchange's market maker does not see total flow y = x + u; it sees y plus Gaussian privacy noise ε, as would happen if a differential-privacy layer jittered the observable reserve state before the pricing rule runs. The maker is modeled as a smart contract that mechanically quotes the posterior mean E[v | ỹ] — an optimal price given what it can see, but one that cannot break even against the real flow it never sees.

Solving is then standard equilibrium algebra: conjecture linear strategies (x = β(v − p₀) and p = p₀ + λỹ), impose that the quote equals the Bayesian posterior (a Gaussian projection formula), impose that the insider's order maximizes expected profit given the price rule, and solve the resulting system. Two equations in two unknowns collapse to λ² = σ_v²/(4(σ_u² + σ_ε²)), which has a unique positive root — hence uniqueness within the linear class. The welfare results come from substituting the equilibrium back into three expected P&L expressions that sum to zero, then differentiating in σ_ε to get comparative statics. A fee variant is checked by computing expected absolute volumes and showing the fee bill matches each trader's incremental gain.

The paper is explicit about what it does not do: it treats the ZK/DP primitive as a black box that outputs ỹ, does no protocol-level soundness analysis, restricts to linear strategies and Gaussian noise, assumes ε is independent of (v, u), and does not solve a full fee equilibrium in which the fee distorts the linear strategy structure. The conclusion (Section 8) is referenced in the roadmap but its text is not included in the supplied content, and the discussion in Section 7 is cut off mid-sentence at "A vanishing of cross-".

## Why This Matters

**Impact on research.** The equilibrium itself is textbook Kyle under the substitution σ_u² ↦ σ_u² + σ_ε², and the paper says so plainly — the novelty is the welfare object, not the equilibrium. That framing is useful: it gives DeFi microstructure a closed-form, single-period cost measure for privacy that sits in the same family as LVR, and it draws a sharp theoretical line (unimplementability of the unbiased rule, via σ(ỹ)-measurability) showing why efficient pricing and break-even-against-real-flow cannot coexist once the signal is strictly coarser than the executed flow.

**Real-world applications:**

- **Shielded AMMs with differential-privacy noise injection** — the primary application, where σ_ε is literally the DP noise scale and Theorem 5.1 gives a per-period fee floor for the LP pool.
- **Fee design and LP compensation in privacy-aggregated exchanges** — translating the subsidy into a per-trade or volume-proportional levy, with the caveat that a per-trade fee f breaks the linear-strategy structure because informed traders stop trading when |v − p₀| < f.
- **Protocol treasury and reserve sizing** — treating |π_M| as a recurring liability that must be budgeted, with the growth rate (quadratic in σ_ε for small noise, linear for large) informing how much privacy a protocol can afford.
- **Comparative design evaluation** — the Penumbra result (π_M = 0 because the maker prices on and trades against the same observable) gives designers a concrete criterion: coarsening between signal and settlement, not the mere presence of cryptography, is what generates the subsidy.

**Industry relevance.** For teams shipping ZK or DP-based exchanges, the paper converts a qualitative worry ("privacy costs LPs something") into a specific number with a specific scaling law and a specific break-even constraint. The reported calibration — roughly $1M/day at σ_ε = σ_u under a BTC/USDT setup, comparable to all revenue from a 0.1% fee on $1B daily volume — is the kind of figure that shapes whether a privacy feature is economically viable at a given volume.

## Future Directions

- **Full fee equilibrium with endogenous volume response.** The break-even bound is computed against no-fee equilibrium volumes; a per-trade fee breaks the linear-strategy structure, so the paper leaves a complete fee-equilibrium analysis — potentially with additional deadweight loss — open.
- **Combining privacy noise with LVR in continuous time.** The paper conjectures, explicitly labelled speculative, that the break-even fee decomposes to leading order as LVR + |π_M|, while noting that cross-terms between time-lag and privacy noise need not vanish at higher orders. Section 7 is truncated mid-sentence in the supplied content.
- **Adaptive and non-Gaussian privacy mechanisms.** Noise correlated with the flow it perturbs is flagged as a separate problem, and the closed-form subsidy is Gaussian-specific; the qualitative existence of a positive privacy-induced LP loss is claimed to generalize to other tractable distributions.
- **Endogenous σ_ε and separate frameworks for uncovered designs.** A privacy-utility tradeoff that optimizes σ_ε under additional constraints is not pursued, and sealed-bid auctions (temporal asymmetry, continuous-time lag) plus oracle-pegged crossings require frameworks the paper does not develop.

## Target Audience

Graduate students and researchers in market microstructure and algorithmic game theory who know the Kyle model and want a clean extension; DeFi protocol designers and mechanism engineers working on shielded AMMs, differential privacy in CFMMs, or ZK exchange architectures; and quantitatively minded tokenomics or risk analysts who need a defensible fee floor for a liquidity pool operating under a privacy layer. Readers primarily interested in ZK cryptography itself will find it out of scope — the paper treats the ZK primitive as a black box and performs no protocol-level soundness or completeness analysis.

## Authors’ abstract

Privacy-preserving cryptocurrency exchanges alter what the pricing mechanism observes about order flow. We derive the unique linear Kyle equilibrium when a committed Bayesian market maker observes order flow perturbed by independent Gaussian privacy noise. The price-impact coefficient and informed-trader strategy rescale by reciprocal factors of the privacy parameter (one down, one up), so their product is invariant. A welfare decomposition then identifies a closed-form…

[The full abstract on arXiv](https://arxiv.org/abs/2605.15746)

[Read the original paper](https://arxiv.org/abs/2605.15746)

---

Canonical: https://crypto.world/paper/the-privacy-subsidy-kyle-s-under-noise-perturbed-order-flow-observation-2605.15746
