{"id":"f340d3d2-89ae-4677-9a30-9c7013cfe24d","arxiv_id":"2507.21769","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Under alpha-local differential privacy, the maximal Fisher information equals roughly alpha^2/4 times the squared mean absolute score as alpha goes to zero, and an optimal extremal mechanism exists.","lead":"This paper shows that, under local differential privacy, the best possible Fisher information in a parametric model is asymptotically alpha squared over four times the squared mean absolute score, and that an optimal privacy mechanism always exists. It gives a general factorization result and applies it to estimate the endpoint of a uniform distribution near-optimally.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniform-application efficiency claim in §5.2 depends on a preliminary estimator θ̂p≤θ0 that the paper never constructs; the abstract's unqualified claim is unsupported.","rationale":"I checked the central theoretical argument in Theorems 2 and 3 with particular attention to the continuous Choquet-based factorization, the compactness and convexity argument in Proposition 2, and the approximation of q(μ) by finite-valued mechanisms in Theorem 3. The apparent uniform-continuity gap noted by the reader is not fatal: tθ0(r)=(∫sθ0 r pθ0)/(∫r pθ0) is explicitly defined on all of C, where the denominator is at least 1, and is weak-* continuous, so the approximation step can be made rigorous. Lemma 8's convexity of i is also valid by the quadratic-over-linear inequality. The most load-bearing weakness I find is in the uniform-distribution application advertised in the abstract. Proposition 4 itself exposes the issue: the estimator is inconsistent when θ̂p>θ0, and the paper provides no α-LDP mechanism producing θ̂p≤θ0 with probability tending to one while θ̂p/θ0→1. The remark that a preliminary consistent estimator on a small subset can be used is an assertion, not a proof, and the numerical experiments explicitly simulate the overestimation failure. Since the abstract claims a consistent and asymptotically efficient estimator without stating the θ̂p≤θ0 condition or supplying a private construction, the claim as stated outruns the proof. This supports the reader's conditional verdict rather than requiring rejection: the main theorems appear sound, but the uniform application needs either a concrete preliminary estimator or a qualified statement.","tokens_in":42707,"tokens_out":16008,"duration_ms":219771,"concrete_test":"Construct an explicit α-LDP preliminary estimator for Uniform[0,θ]—for example, a private lower confidence bound from a disjoint subsample of size m=n^ρ using a one-bit or RAPPOR-style channel—and verify both P(θ̂p≤θ0)→1 and θ̂p/θ0→P1. Then recompute the unconditional asymptotic variance of θ̂n under the two-stage scheme and compare it with θ0²/α². If the construction works, include it and the efficiency claim is restored; if no such estimator can be built with these properties, or if the recomputed variance exceeds the Fisher-information bound by a nonvanishing factor, the abstract's efficiency claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2's 'consistent and asymptotically efficient estimator' is conditional on a preliminary estimator θ̂p satisfying θ̂p≤θ0 with probability tending to 1, but no private construction is supplied. Proposition 4(1) shows that if θ̂p>θ0, then θ̂n→θ0∨θ̂p a.s., so any overestimate causes inconsistency. After Remark 11 the paper only asserts that 'it is possible to replace it with some preliminary consistent estimator based on a small subset of the data' without giving the estimator or proving the needed one-sided consistency under α-LDP. Under noninteractive local differential privacy, a downward-biased consistent estimator of a support endpoint is not automatic: constructing it requires a separate private procedure, and the variance of the final estimator depends on how close θ̂p is to θ0. Thus the abstract's unqualified efficiency statement for the uniform model is not established by the submitted proof. This concern is load-bearing for the advertised application, although it does not invalidate the core Theorems 2–3 for regular models.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the maximal Fisher information achievable by an alpha-locally differentially private mechanism for a one-dimensional regular parametric model. The authors prove a factorization lemma showing that any alpha-LDP channel can be written as an extremal 'staircase' channel followed by post-processing; in the continuous case the factorization is obtained through Choquet's theorem on the set C={1<=v<=e^alpha} in L^infty. For discrete models they obtain an exact small-alpha formula for the maximal Fisher information (Theorem 1); for continuous models they obtain matching upper and lower bounds with asymptotic constant alpha^2/4 (integral |s_{theta0}| p_{theta0} dx)^2 as alpha->0 (Theorem 2) and prove existence of an extremal mechanism attaining the supremum for every alpha (Theorem 3). The final section applies the framework to estimating the endpoint theta of a uniform distribution on [0,theta], proposing a two-point extremal mechanism and a two-stage estimator whose asymptotic variance matches the derived Fisher information bound when a preliminary estimate theta_hat_p <= theta0 is available. Numerical experiments illustrate the behavior for theta_hat_p above and below theta0.","tokens_in":42971,"tokens_out":13623,"duration_ms":158655,"significance":"The conceptual contribution is valuable: the factorization viewpoint unifies the discrete staircase mechanisms of Kairouz et al. with a continuous Choquet-based representation, and Theorem 2 gives a clean, explicit asymptotic benchmark that matches the Gaussian case in [30]. The proof of Theorem 3 is an interesting application of convex analysis (maximal measures and Zorn's lemma) to a problem where the extremal set E is not closed. The discrete theorem is carefully worked out, with an explicit two-point optimal mechanism. The uniform application is attractive and the numerical experiments support the conditional claims. However, two load-bearing points need repair before the advertised results are fully established: the uniform-continuity step in the proof of Theorem 3 and the missing private preliminary estimator in Section 5.2.","major_comments":[{"comment":"The proof asserts that 't_theta0 and p_tilde_theta0 are uniformly continuous on the compact set C', but the preceding results only establish continuity on E: Lemma 1(3) states continuity on E, and E is not compact in the d_star topology (Remark 12). The functions t_theta0 and p_tilde_theta0 are not defined on all of C in the manuscript. Since the claim that the approximation error in Eq. (69) can be made arbitrarily small depends on this uniform continuity, the existence proof is incomplete as written. The fix is straightforward if formulas (32)-(33) are used to extend p_tilde_theta0 and t_theta0 to all of C and their weak-* continuity is proved there; please do so or replace the argument.","section":"7.2.3, proof of Theorem 3"},{"comment":"The abstract's claim that the proposed mechanism 'yields a consistent and asymptotically efficient estimator in high privacy regime' is not established by the submitted proof. Proposition 4(1) shows that if the preliminary value theta_hat_p exceeds theta0, then theta_hat_n converges almost surely to theta0 vee theta_hat_p, i.e. it is inconsistent. The analysis in Proposition 4(2) is conditional on theta_hat_p <= theta0, and Remark 11 only asserts, without proof or construction, that theta_hat_p can be replaced by 'some preliminary consistent estimator based on a small subset of the data'. No alpha-LDP procedure producing a downward-consistent estimator of the support endpoint theta0 is given, and no argument shows that such an estimator can satisfy theta_hat_p/theta0 -> 1 fast enough for the stated variance equivalence. The efficiency statement for the uniform model should either be removed from the abstract or supplied with the missing preliminary estimator and its one-sided consistency proof.","section":"5.2, Remark 11, and abstract"},{"comment":"The proof derives a pointwise derivative of theta -> p_tilde_theta(r) for mu-almost every r and then computes the integral of (partial_theta p_tilde)^2 / p_tilde, but it does not verify the differentiability-in-quadratic-mean condition of Definition 2, which is what the statement 'the model is DQM' asserts. Since the subsequent efficiency comparison in Remark 11 uses this Fisher information as the asymptotic variance lower bound, Proposition 3 needs a proof of DQM, or else the statement should be weakened to the property actually established.","section":"5.1, Proposition 3"}],"minor_comments":[{"comment":"There are two results numbered Lemma 1: the discrete factorization lemma in Section 3.1 and the regularity lemma for continuous extremal mechanisms in Section 4.2; renumbering would prevent confusion.","section":"3.1 and 4.2"},{"comment":"The definition q(z):=ess inf_x q_x(z) is asserted to satisfy q(z)>0 without argument; as in the discrete Lemma 1, outputs with q(z)=0 have q_x(z)=0 almost everywhere and should be discarded before dividing by q(z).","section":"4.1, Proposition 1"},{"comment":"The homogeneity identity i(lambda r)=|lambda| i(r) is written for lambda in R, but i is only defined on C where the denominator is positive; the convexity argument should be restricted to the convex-combination parameters actually needed.","section":"7.2.3, proof of Lemma 8"},{"comment":"The comparison with [30] states the Gaussian optimal Fisher information as (2/pi)(e^alpha-1)^2/(e^alpha+1)^2; since the derivation is only sketched, a precise reference to the corresponding equation in [30] would help the reader verify the match.","section":"Remark 7"},{"comment":"The captions of Figures 1 and 2 are very terse; in particular, the dashed line in Figure 2 should be identified as the standard-deviation lower bound derived from Proposition 3.","section":"Figures 1-2"}],"recommendation":"major_revision","confidential_remarks":"The two major gaps are fixable within the manuscript's scope, so I would not reject. The uniform-application overclaim in the abstract should be softened or the missing private preliminary estimator should be constructed. The paper is long and the functional-analytic apparatus is heavy, but the material is largely self-contained and the central discrete and asymptotic bounds appear sound. I suggest asking the authors to address the Theorem 3 continuity gap and to either construct the preliminary estimator or qualify the abstract's efficiency claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core Theorems 2 and the asymptotic characterization look right and are the real news: exact alpha->0 constant for maximal Fisher information under alpha-LDP, with matching bounds from explicit two-point staircase mechanisms. The continuous factorization via Choquet's theorem is substantial new machinery, not just a restatement of Kairouz et al. The paper earns credit for the Bochner integral and point-evaluation setup.\n\nSoft spots, in proportion. First, Theorem 3's proof has a genuine gap. It invokes uniform continuity of the score t_theta0 and the density ptilde_theta0 on the compact set C, but those functions are only defined on E, which is neither closed nor compact in the weak-* topology. The approximation by finite-valued mechanisms needs uniform continuity on E, and the paper does not provide it. This looks repairable, since many such functions extend continuously to the closure of E, but as written the existence proof for every alpha is incomplete. Second, the uniform application in Section 5.2 is oversold. The consistent and asymptotically efficient estimator requires a preliminary estimate below the true parameter; Proposition 4 shows overestimates produce a biased limit. The paper only asserts that such a preliminary estimator can be obtained from a small subset of the data, without giving a private construction or proving one-sided consistency under alpha-LDP. The abstract's unqualified efficiency claim is therefore not supported. This does not damage Theorems 2 and 3, but it should be fixed or qualified.\n\nThe discrete factorization lemma is essentially a known restatement of Kairouz et al., and the authors acknowledge that lineage. The citation pattern is fine; self-citations are not load-bearing.\n\nWho this is for: anyone working on private parametric efficiency, especially high-privacy asymptotics. It deserves a serious referee. My recommendation: engage, but require the authors to repair the uniform-continuity step in Theorem 3 and either construct the preliminary estimator or qualify the uniform claim.","headline":"Solid high-privacy Fisher information characterization with a real gap in the existence proof and an overclaimed uniform application; worth refereeing but needs fixing.","tokens_in":43422,"tokens_out":1468,"would_cite":true,"duration_ms":19155,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","68P27","62B15","46A55"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that under $\\alpha$-local differential privacy, the maximal Fisher information of a regular one-dimensional parametric model behaves like $\\frac{\\alpha^2}{4}(\\int |s_{\\theta_0}(x)|p_{\\theta_0}(x)dx)^2$ as $\\alpha\\to 0$…","keywords":["local differential privacy","Fisher information","extremal mechanisms","staircase mechanisms","Choquet theorem","asymptotic efficiency","uniform distribution estimation","factorization lemma"],"falsifier":"For the central theorem, one could numerically search for an $\\alpha$-LDP mechanism on a standard normal location model whose Fisher information exceeds $(e^\\alpha-1)^2/(2\\pi)$ at a fixed $\\alpha$; any such mechanism would refute the upper bound in Theorem 2. For the uniform application, the paper's own Figure 1 provides the disconfirming observation: with $\\hat{\\theta}_p=1.3\\theta_0$ and $\\alpha=0.3$, the estimator's empirical mean stays near $1.3\\theta_0$ rather than $\\theta_0$, exactly the predicted failure mode.","tokens_in":42496,"feed_emoji":"🔒","tokens_out":7525,"duration_ms":82571,"temperature":0.7,"pith_summary":"This paper identifies the statistical efficiency ceiling for locally differentially private estimation. It proves that any $\\alpha$-local differential privacy mechanism can be factorized as an extremal 'staircase' mechanism followed by arbitrary extra randomization, and because the extra step can only destroy Fisher information, only the extremal mechanisms matter. For a regular one-dimensional parametric model the maximal Fisher information under $\\alpha$-LDP is asymptotically $\\frac{\\alpha^2}{4}(\\int |s_{\\theta_0}(x)|p_{\\theta_0}(x)dx)^2$ as $\\alpha\\to 0$, with matching upper and lower bounds, and for every $\\alpha>0$ an extremal mechanism attaining the maximum exists. These results give the exact high-privacy benchmark for efficient estimation and a construction principle that the authors then apply to estimating the range of a uniform distribution.","feed_headline":"High-privacy Fisher information tops out at a simple score integral","feed_subtitle":"A factorization lemma shows every local-DP mechanism is an optimal staircase plus extra noise, which pins the efficiency ceiling.","key_machinery":"The central object is the factorization lemma: any $\\alpha$-LDP channel $q$ factorizes as $q=q^{(2)}\\circ q^{(\\mu)}$, where $q^{(\\mu)}$ is extremal, meaning its log-likelihood ratios take only the values $0$ and $\\alpha$, and $q^{(2)}$ is arbitrary extra randomization. In finite spaces this follows by writing each ratio vector in the hyperrectangle $[1,e^\\alpha]^d$ as a convex combination of its vertices via Carathéodory's theorem; in continuous spaces the same decomposition is obtained by Choquet's theorem on the compact convex set $C=\\{1\\le v\\le e^\\alpha\\}$ in $L^\\infty$ with the weak-$*$ topology, whose extreme points are the measurable functions with values in $\\{1,e^\\alpha\\}$. A measurable point-evaluation operator and Bochner integrals make the pointwise evaluation of these integrals rigorous. Because the extra randomization step can only decrease Fisher information, the optimization problem reduces to choosing the sub-probability measure $\\mu$ on the extreme set $E$.","core_discovery":"The central claim is that the optimization problem $\\sup_{q\\in\\mathcal{Q}_\\alpha} I_{\\theta_0}(q\\circ P)$ has a complete answer in one dimension. For any regular model, $J^{\\max,\\alpha}_{\\theta_0}$ is squeezed between $\\frac{(e^\\alpha-1)^2}{2e^\\alpha(1+e^\\alpha)}(\\int |s_{\\theta_0}|p_{\\theta_0})^2$ and $\\frac{(e^\\alpha-1)^2}{4}(\\int |s_{\\theta_0}|p_{\\theta_0})^2$, so as $\\alpha\\to 0$ the maximal Fisher information is equivalent to $\\frac{\\alpha^2}{4}(\\int |s_{\\theta_0}|p_{\\theta_0})^2$. Theorems 2 and 3 further show that the supremum is always attained, by an extremal mechanism $q^{(\\mu)}_x(dr)=e_x(r)\\mu(dr)$ for some Radon sub-probability measure $\\mu$ on the extreme points $E=\\{r\\text{ measurable}: r\\in\\{1,e^\\alpha\\}\\text{ a.e.}\\}$. In the finite case, Theorem 1 gives the exact value $\\frac{(e^\\alpha-1)^2}{4}\\cdot \\frac{E[|s_{\\theta_0}(X)|]^2}{[(1-n_{\\max})+e^\\alpha n_{\\max}][n_{\\max}+(1-n_{\\max})e^\\alpha]}$, achieved by a mechanism that randomizes between the two score-sign regions. The authors apply the same two-point mechanism to the uniform model $U[0,\\theta]$, proving a consistent, asymptotically Gaussian estimator whose variance matches the bound as $\\alpha\\to 0$ whenever the preliminary estimate satisfies $\\hat{\\theta}_p\\le \\theta_0$.","pith_inferences":["The same factorization argument should apply to other utility criteria, since Proposition 2 maximizes any bounded continuous convex functional $j$ over measures on $E$; one could define efficient mechanisms for risks other than Fisher information.","The multidimensional extension is left open in the paper, but the factorization itself is dimension-free; a natural next step is to maximize a trace or directional version of the Fisher information matrix, where the scalar score-sign split becomes a hyperplane choice.","The uniform example exposes a general two-stage recipe: use a cheap preliminary estimate to locate the score-sign boundary, then apply the extremal Bernoulli mechanism; making the preliminary step itself private without losing the below-truth guarantee is the practical bottleneck.","If the conjecture in Remark 13 holds, the continuous staircase mechanism would transfer to metric spaces with a non-atomic measure, opening the way to locally private inference for stochastic processes."],"forward_implications":["In the high-privacy regime, no $\\alpha$-LDP estimator for a regular one-dimensional model can beat asymptotic variance $4/(\\alpha^2(\\int |s_{\\theta_0}|p_{\\theta_0}dx)^2)$ per observation, and the score-sign staircase mechanism achieves it.","Any optimal $\\alpha$-LDP mechanism can be taken to be extremal, so the search over all privacy channels collapses to the choice of a sub-probability measure on the extreme set $E$.","For every fixed $\\alpha>0$, the Fisher-information maximization over $\\alpha$-LDP mechanisms has a solution, not merely a supremum.","For the uniform range model, the proposed two-point mechanism yields a $\\sqrt{n}$-consistent estimator with variance equivalent to $\\theta_0^2/\\alpha^2$ when the preliminary estimate equals $\\theta_0$, matching the Fisher-information upper bound."],"supporting_citations":[{"why":"Supplies the finite-alphabet extremal-mechanism framework and the privacy-utility linear program that Theorem 1 parallels.","marker":"[29]"},{"why":"Provides the regularity-preservation result under LDP, the score formula, and the Fisher information inequality used throughout the upper bounds.","marker":"[45]"},{"why":"Choquet's theorem gives the integral representation of points of a compact convex set as barycenters of measures on its extreme points, the continuous analogue of Carathéodory's theorem.","marker":"[16]"},{"why":"Earlier upper bounds on Fisher information under LDP that this paper sharpens and shows to be attainable.","marker":"[7]"},{"why":"Efficient Gaussian mean estimation under LDP is the benchmark compared in Remark 7, where the small-$\\alpha$ limits coincide.","marker":"[30]"},{"why":"The one-dimensional staircase mechanism introduced here is a precursor of the extremal mechanisms studied in this paper.","marker":"[25]"}],"fun_headline_variants":["Tight Fisher bound for local privacy: score integral is optimal","Extremal mechanisms achieve Fisher info ceiling in high privacy","Alpha-LDP: Fisher info max equals score integral squared","High-privacy Fisher info peaks at score integral, not more","Optimal local DP estimator: Fisher bound reached by extremal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The uniform-law efficiency result assumes a preliminary estimate $\\hat{\\theta}_p$ that is no larger than the true $\\theta_0$, yet the paper gives no private construction of such an estimate; if $\\hat{\\theta}_p$ exceeds $\\theta_0$, Proposition 4 shows the estimator converges to $\\theta_0\\vee\\hat{\\theta}_p$ and stays biased.","fun_headline_variants_meta":{"raw":{"variants":["Tight Fisher bound for local privacy: score integral is optimal","Extremal mechanisms achieve Fisher info ceiling in high privacy","Alpha-LDP: Fisher info max equals score integral squared","High-privacy Fisher info peaks at score integral, not more","Optimal local DP estimator: Fisher bound reached by extremal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1495,"prompt_tokens":1107,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":723,"completion_tokens_details":{"reasoning_tokens":318}},"tokens_in":723,"tokens_out":388,"duration_ms":5325,"temperature":1.0,"reasoning_tokens":318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:23:27.054457+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the central theorem, one could numerically search for an $\\alpha$-LDP mechanism on a standard normal location model whose Fisher information exceeds $(e^\\alpha-1)^2/(2\\pi)$ at a fixed $\\alpha$; any such mechanism would refute the upper bound in Theorem 2. For the uniform application, the paper's own Figure 1 provides the disconfirming observation: with $\\hat{\\theta}_p=1.3\\theta_0$ and $\\alpha=0.3$, the estimator's empirical mean stays near $1.3\\theta_0$ rather than $\\theta_0$, exactly the predicted failure mode.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the finite-alphabet extremal-mechanism framework and the privacy-utility linear program that Theorem 1 parallels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the regularity-preservation result under LDP, the score formula, and the Fisher information inequality used throughout the upper bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Choquet's theorem gives the integral representation of points of a compact convex set as barycenters of measures on its extreme points, the continuous analogue of Carathéodory's theorem."},{"cited_title":"P., Chen, W","cited_arxiv_id":null,"evidence_quote":"Earlier upper bounds on Fisher information under LDP that this paper sharpens and shows to be attainable."},{"cited_title":"Efficient Estimation of a Gaussian Mean with Local Differential Privacy","cited_arxiv_id":"2402.04840","evidence_quote":"Efficient Gaussian mean estimation under LDP is the benchmark compared in Remark 7, where the small-$\\alpha$ limits coincide."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The one-dimensional staircase mechanism introduced here is a precursor of the extremal mechanisms studied in this paper."}],"review_version":1}