| paper | d4-conf-aaai-2011-0372 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | no |
Theorem 12 is a named protocol-existence impossibility, so the paper clears the computational-result gate. The proposed mirror gives a census-level threshold oracle over \(\mu\), making a weighted-median decision tree possible but replacing the paper’s protocol among privately reporting individuals and its privacy notion. I do not think this is a continuous version the authors would clearly recognize.
fails bit b-changes — the continuum limit changes the question
The threshold oracle assumes access to aggregate answers about private locations, replacing the paper’s individual-input protocol model and its question about what a protocol must elicit from agents.
fatal: True
The proposed mirror covered only Theorem 12’s impossibility for exact privacy in median facility location; it left the second-price auction results and the other facility-location privacy-ratio results alone.
The strongest positive case is for one result: Theorem 12, stated as the paper’s own result and proved there. It says that for every \(n\ge2\), no protocol for the \(n\)-agent facility-location problem perfectly preserves privacy for the median mechanism. This is a protocol-existence impossibility, rather than a standard \(P\)-versus-\(NP\) classification. If “computational result” here means only complexity-class claims, the paper has no qualifying anchor: its other numbered results give privacy-ratio formulae and bounds.
A continuous question inspired by Theorem 12 is Perfectly Private Mass-Median Protocol. Let \(Q=\{q_1<\cdots<q_\tau\}\) be a finite grid of possible ideal facility locations. A society is a distribution \(\mu\) over \(Q\), where \(\mu_q\) is the fraction of residents whose complete type is ideal location \(q\), with cost \(|y-q|\) for facility location \(y\). The objective is to choose the leftmost weighted median
\[
f(\mu)=q_j,\qquad j=\min\left\{i:\sum_{h\le i}\mu_{q_h}\ge\tfrac12\right\}.
\]
A protocol may ask whether at least half the population has an ideal location at or below a threshold \(r\), and receives only the aggregate yes/no answer. Its solution is a finite decision tree that outputs \(f(\mu)\) for every \(\mu\), with each transcript class exactly equal to the set of societies having the same \(f(\mu)\). That equality is the continuous analogue of revealing only the information needed to determine the outcome.
This mirror has a plausible regime: a large municipality with many residents but relatively few distinct ideal sites and identical distance costs within each site type, so \(n\gg\tau\). A binary search over \(Q\) constructs the requested protocol in \(O(\log |Q|)\) aggregate queries. I expect this version to be tractable. It covers the median facility-location problem and the privacy impossibility in Theorem 12; it does not mirror the paper’s SPA results.
The weak point is the protocol model. The paper studies transcripts over individually supplied types; this mirror lets the mechanism see aggregate threshold answers, and privacy is defined over mass distributions. Theorem 12’s obstruction—revealing the identity of a median agent—disappears when the society is represented by type masses. That is a natural way to ask the question for a continuous society, but the authors might reject the aggregate query interface as an added assumption. Without it, the paper’s privacy result does not directly yield a continuous computational problem.
Theorem 12 is the paper’s strongest anchor, but its impossibility is about labeled agents and their private inputs. Its protocol partitions the product space of individual type profiles into rectangles. For facility location, the paper attributes the obstruction to having to reveal the identity of an agent with a median value. That identity is exactly what a society represented only by type masses discards.
A continuous version therefore faces a substantive fork. If the algorithm receives the mass vector \(\mu\), it already has the population’s aggregate information, so the paper’s question about what a protocol must learn from private agents is gone. If citizens still hold and report their individual locations, the protocol remains a finite-agent problem; repeating types does not turn its input or privacy notion into a distribution over societies. And in an atomless society there are no individual agents whose identities a transcript could reveal.
The proponent’s aggregate threshold queries give a coherent way to compute the weighted median, but they add a global aggregation operation absent from the paper’s protocol model. A stronger alternative would give private type masses to \(\tau\) type representatives and study communication or leakage when they jointly compute the weighted median. That could be a useful problem, but its private object is the composition of the population, and its answer depends on choices the paper does not supply: who holds each mass, what messages they may send, and how leakage over the simplex is measured. It is not a continuation of the median-agent identity result.
The paper’s privacy-ratio theorems do not provide a clearer computational anchor: they give ratios and bounds for finite-agent protocols, rather than complexity classifications for a mass-input problem. So the negative case is that no canonical continuous mirror of Theorem 12 emerges from the paper: preserving its privacy question keeps the individuals, while making the population the input changes what privacy means. That does not rule out a worthwhile new project on private computation over population masses, and the case is weakest if ChoCo wants to include such new privacy models. But that would need its own motivation and model; this theorem does not establish it as a continuous mirror.
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.