Indifferential Privacy: A New Paradigm and Its Applications to Optimal Matching in Dark Pool Auctions

· AAMAS 2025 (aamas25-00189)

mirror found
paperIndifferential Privacy: A New Paradigm and Its Applications to Optimal Matching in Dark Pool Auctions
authors
venueAAMAS 2025
filed underunclassified
judged bygpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1)
judge confidencehigh
authors would recognise ityes

The anchor — algorithmic

Lemma 3.2

Algorithm 1 returns a maximum matching between real buy and sell nodes in the matching graph.

statement extracted from the paper’s text layer

Every anchor argued

The continuous mirror question

Given a finite set \(R=B\cup S\) of complete order types \(t=(\sigma_t,p_t,q_t)\), rational society masses \(\mu\in\Delta_R\), and compatibility \(p_b\ge p_s\), compute a nonnegative flow \(z_{bs}\) maximizing \(\sum_{b,s}z_{bs}\), subject to \(z_{bs}=0\) for incompatible pairs, \(\sum_s z_{bs}\le q_b\mu_b\), and \(\sum_b z_{bs}\le q_s\mu_s\).

The model it lives in

A high-multiplicity dark-pool society is given by type masses \(\mu_t\), where each type records side, public price, and common quantity \(q_t\). Decision variables are compatible matching masses \(z_{bs}\); the objective is total matched volume. On rational repeated-type instances, scaling by the population size recovers the integral discrete matching problem.

The objection that survived

The flow mirror does not preserve owner-specific commitments, fake-node prefixes, or execution transcripts, so it captures maximum matched volume rather than the full protocol behavior.

fatal: False

What the mirror covers

The mirror covers Lemma 3.2's maximum matching correctness over repeated order types, but leaves Lemma 3.3, Corollary 3.4, owner-level privacy, cryptographic commitments, implementation benchmarks, and FHE comparisons uncovered.

Open questions for a prover

The case FOR (proponent)

The strongest honest case is a population mirror of the paper’s exact matching result, with the privacy contribution as a more ambitious secondary mirror.

The lead anchor is Lemma 3.2, proved in this paper: “Algorithm 1 returns a maximum matching between real buy and sell nodes in the matching graph.” The paper does not state a numbered \(P\), NP-hardness, or parameterized-complexity theorem; Lemma 3.2 is a correctness result for a computational problem. That is still a valid anchor, but the distinction matters.

I would call the mirror Continuum Dark-Pool Clearing.

An instance consists of a finite set of complete order types
\[ R=B\cup S,\qquad t=(\sigma_t,p_t,q_t), \]
where \(\sigma_t\in\{\mathrm{Buy},\mathrm{Sell}\}\), \(p_t\in\mathbb Q_{\ge0}\) is the public price, and \(q_t\in\mathbb Z_{>0}\) is the common order quantity of agents of that type. The society is a rational mass vector \(\mu\in\mathbb Q_{\ge0}^{R}\) with \(\sum_t\mu_t=1\). Thus \(\mu_t\) is the fraction of clients of type \(t\), and \(a_t=q_t\mu_t\) is their aggregate order volume.

The solution is a matching-flow vector \(z=(z_{bs})_{b\in B,s\in S}\), where
\[ z_{bs}\ge0,\qquad z_{bs}=0\ \text{if }p_b<p_s, \]
and
\[ \sum_s z_{bs}\le a_b,\qquad \sum_b z_{bs}\le a_s. \]
The objective is to maximize
\[ \sum_{b,s}z_{bs}, \]
the total matched volume. A solution is the flow together with its matched-volume certificate.

This is recognisably the authors’ problem: the same buy–sell compatibility condition, the same objective of maximizing matched units, and the same price ordering that drives their polar-opposite strategy. The only change is that repeated unit nodes are represented by capacity mass. For a finite repeated-agent election with \(n_t\) agents of type \(t\), set \(\mu_t=n_t/N\). Multiplying the continuous solution by \(N\) gives an integral capacitated bipartite matching problem, because the underlying flow polytope is integral. Thus this is not merely a fractional relaxation: on the repeated-type slice it exactly recovers the discrete maximum matching after scaling.

The natural regime is a large dark-pool clearinghouse receiving many institutional child orders, with prices drawn from a moderate number of ticks and order sizes drawn from a moderate number of standard lot classes. There may be millions of orders but only \(\tau\) repeated side–price–quantity types, with \(\tau\ll n\). Agents of the same type are indistinguishable for the matching objective and compatibility relation. If every order has an idiosyncratic price, quantity, or priority rule, this regime is inappropriate; the mirror is explicitly for a repeated-order environment.

I would expect Continuum Dark-Pool Clearing to be in Class A. A sorted sweep, or equivalently a small transportation LP, computes the optimum in time polynomial in \(\tau\) and the rational input length. The paper’s Lemma 3.2 supplies the structural justification: matching the most attractive remaining buy and sell blocks can be made compatible with an optimum. In the continuum this becomes an exchange argument over mass rather than individual nodes.

A secondary anchor is Lemma 3.3, also proved here: using an ideal commitment scheme, the protocol is \((\epsilon,\delta)\)-indifferentially private. Corollary 3.4, proved here as well, gives computational IDP with a real-world commitment scheme. I would treat these as one privacy anchor rather than two separate problems.

The corresponding question is Resolution-\(\eta\) Continuum-IDP Clearing. Let \(u=(\sigma,p)\) be a public order class and let \(q\) be a hidden quantity class. The society is a rational mass distribution \(\mu_{u,q}\). The server observes only the public projection
\[ \nu_u=\sum_q\mu_{u,q}. \]
A neighboring input moves mass \(\eta\) from \((u,q)\) to \((u,q+1)\), leaving every public class unchanged. Here \(\eta\) is an explicit privacy resolution, analogous to one client in a finite population normalized by its population size.

The protocol must output a maximum true matched volume while adding committed fake order mass \(\eta N_u\), where \(N_u\) is sampled from the bounded geometric distribution used in the paper. Its transcript must satisfy
\[ HS^{F_\eta}_{e^\epsilon} \bigl(\mathrm{View}(\mu)\,\|\,\mathrm{View}(\widehat\mu)\bigr) \le\delta, \]
where \(F_\eta\) pairs identical transcripts, except that two transcripts may diverge after the smaller hidden quantity \(q\eta\) has been fully matched. This is exactly the paper’s “privacy may be relaxed after full execution” principle, expressed at mass resolution. The matching part remains the same polar-opposite sweep, with fake mass discarded when opened.

Under a finite type set, rational encoding, and bounded noise support, I would expect this aggregate version also to be Class A: the matching computation is a chain-structured flow problem, and the privacy analysis should follow the paper’s coupling of shifted geometric distributions. The further questions are whether the guarantee composes across many repeated agents, how to handle multiple assets or time priority, and whether a useful finite-population rounding theorem preserves both exact matching and IDP.

The weakest point is the privacy mirror. Literal individual-level privacy does not survive unchanged in an atomless population: a single agent has zero mass, so changing that agent’s quantity is either invisible or requires an external resolution parameter. The \(\eta\)-formulation therefore protects a small cohort or aggregate volume, not literally one named user. Moreover, the paper samples noise per user, whereas the continuum version naturally samples noise per public type or mass block. I would not claim that Corollary 3.4 transfers automatically.

That weakness does not undermine the lead case. Continuum Dark-Pool Clearing is already a precise, faithful mirror of Lemma 3.2, with an exact high-multiplicity bridge and an efficient continuous algorithm. The scope should stop there if necessary: it covers the paper’s maximum-matching result, not its implementation benchmarks, FHE comparisons, or full per-user cryptographic privacy theorem. The paper’s use of “continuous double auction” refers mainly to trading over time; the proposed mirror continuizes the population, which is a separate and legitimate operation.

The case AGAINST (opponent, writing after the proponent)

The strongest negative argument is not that aggregate matching is easy; that would be the wrong objection. The problem is that the proposed mirror changes the object whose correctness the paper proves.

The paper has no complexity theorem in the ChoCo sense: no \(P\)-versus-\(NP\) result, approximation guarantee, or parameterized algorithm. Lemma 3.2 is a correctness lemma for a protocol whose essential structure is per-user: real nodes precede fake nodes, commitments are opened by named owners, and a fake node certifies that that owner’s real order has been exhausted. The paper’s computational object is therefore not merely the value of a maximum matching.

The proponent’s Continuum Dark-Pool Clearing discards precisely this structure. A mass-flow vector \(z_{bs}\) records how much volume moves between public order classes, but not which users’ orders are matched, whose commitments are opened, or which user has been fully executed. That is a different output. Two populations can have the same mass distribution over \((\mathrm{side},p,q)\) and the same maximum matched volume while inducing different owner-level transcripts. If identity is omitted from the type, the mirror cannot reproduce the protocol. If identity is included, then every named trader is effectively its own type and the intended high-multiplicity regime disappears.

One can still study the aggregate volume problem, but it is then a generic capacitated matching problem on a threshold graph. The fake-node argument, commitments, and privacy mechanism have vanished. This is not an objection that the answer is easy; it is an objection that the proposed flow is no longer a mirror of Lemma 3.2 as used in the paper. It mirrors only the stripped-down objective “what total volume can be matched?” The paper’s result is that this volume can be obtained while handling owner-specific hidden quantities through a particular transcript.

The privacy anchor fails more decisively. Lemma 3.3 and Corollary 3.4 protect the quantity of one named user under the neighboring relation \(\lvert x_i-\widehat{x}_i\rvert\le 1\). In a population of \(N\) users, that perturbation has normalized mass \(1/N\), which vanishes in the continuum limit. The proponent’s fixed resolution \(\eta\) avoids this degeneration only by replacing one-user privacy with privacy for a positive-mass cohort or volume block. That is not a harmless rescaling: it changes who is protected and what disclosure is permitted.

The paper’s indifference relation is also intrinsically owner-specific. Its exceptional disclosure occurs after the smaller quantity of user \(i\) has been fully matched. In an atomless mass model there is no canonical “next node” of user \(i\), no finite real-prefix followed by fake nodes, and no owner whose exhaustion generates the maximal common prefix. A flow may split mass across counterpart prices and execution times. Defining a mass-level analogue would require a new event structure, a new transcript model, and a new privacy proof.

Replacing per-user geometric noise by noise on a public type does not repair this. It changes the distribution, the adjacency relation, and the composition behaviour. Independent per-user noise aggregated over a cohort is not the paper’s bounded geometric mechanism, while one noise variable per type protects an aggregate count rather than an individual order. Such a mechanism might be worth studying, but it would be a new aggregate differential-privacy problem, not a continuization of the paper’s IDP theorem. And if all users in a type are genuinely indistinguishable, the individual privacy question largely disappears; the interesting object becomes ordinary histogram privacy, which can be formulated without an atomless population.

Thus every plausible modelling choice seems to force the same tradeoff. Retain identities, owner-specific execution, and the paper’s privacy guarantee, and there is no useful multiplicity collapse. Collapse agents into repeated order types, and one obtains a legitimate high-multiplicity clearing problem, but loses the paper’s distinctive protocol and privacy content.

That said, the negative case is not airtight. A repeated-order dark pool is a plausible high-multiplicity regime, and the aggregate maximum-flow problem is mathematically well-posed. I therefore cannot honestly claim that no continuous question whatsoever can be extracted from this paper. The defensible negative verdict is narrower: the paper offers no worthwhile continuous mirror of its central contribution. Its only surviving mirror is a stripped-down volume-clearing problem, while its genuinely novel owner-level privacy result degenerates or becomes a different problem under continuization.

The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.