{"id":"28b2d955-f0b5-4294-8dcb-0ea4de329afc","arxiv_id":"2412.18992","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For federated private functional mean estimation, the paper derives minimax rates under common and independent designs, showing independent designs reduce the privacy cost when privacy binds.","lead":"This paper finds the fastest possible rate at which a federated learning system can estimate an average curve from noisy samples while each server protects its users with differential privacy. It shows that when privacy is the bottleneck, letting each person be measured at their own random time points gives more accurate estimates than using one shared set of time points.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lower-bound lemma's Van Trees prior term is off by a factor 2^{2L}; as displayed, Lemma 3 yields rate D^{-2α-2}, not D^{-2α}, so Theorems 1, 2, 5, and 6 do not follow from the stated lemma.","rationale":"The central claim is a matching minimax rate, so the lower bound is as load-bearing as the upper bound. The reader's weakest_assumption concerned design generality (non-uniform or clustered design points), but the main theorems explicitly restrict to equispaced common design and i.i.d. Uniform independent design, so that concern is secondary. A more fundamental issue is that the lower-bound engine, Lemma 3, has a displayed Van Trees prior term that is off by a factor 2^{2L}; if taken literally, the lower bound would be D^{-2α-2}, contradicting the upper bound. The supplement's derivation indicates the intended exponent is 2^{2L(α+1)}, so the result is likely a fixable typo rather than a fatal error, but it must be corrected for the proof to be sound. The reader did list this discrepancy as issue (i), but did not treat it as the weakest assumption; my read is that it is the most load-bearing concern because it directly affects whether Theorems 1, 2, 5, and 6 are proven. The other issues (small-δ restrictiveness, missing simulation details) do not threaten the central claim, so the verdict remains CONDITIONAL, unchanged from the reader.","tokens_in":40157,"tokens_out":23330,"duration_ms":188138,"concrete_test":"Independently re-derive Section 8.1.1's van Trees lower bound: compute J(π)=∑_{k=1}^{2^L} ∫(π'_k)^2/π_k for the product prior on f_k, and verify whether J(π) equals C·2^{2L(α+1)} or C·2^{2L(α+2)}. Then plug the result into Lemma 3 and check that the final minimax lower bound is D^{-2α}. If J(π)=2^{2L(α+1)}, the displayed 2^{2L(α+2)} in Lemma 3 must be corrected and the proof of Theorem 5 harmonized with the supplement. As a secondary check within the same re-derivation, confirm that the constructed X_i^{(s)} in (16) lies in H^α(R) when f_k ranges over the stated submodel; if not, adjust the coefficient range as in main-text equation (13).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 3 (main text, expression (18)) lower-bounds the minimax risk by 2^{2L} / (sup_f ∑_s min{Tr(I^{Y(s)|T(s)}), ...} + 2^{2L(α+2)}). The multivariate van Trees prior used in the supplement (Section 8.1.1) places independent densities on the 2^L coefficients f_k, supported on an interval of half-length a = 2^{-L(α+1/2)}R. For the cosine-type prior used there, the prior Fisher information per coordinate is π²/a², so J(π) = 2^L π²/a² = π² 2^{2L(α+1)}/R². Thus the denominator's last term should be of order 2^{2L(α+1)}, not 2^{2L(α+2)}. If the displayed 2^{2L(α+2)} were correct, the bound would scale as 2^{2L}/2^{2L(α+2)} = D^{-2α-2}, strictly slower than the claimed upper-bound rate D^{-2α}. Since Lemma 3 is the engine for Theorems 5 and 6, and hence for the central minimax claims in Theorems 1 and 2, this is a load-bearing inconsistency in the written proof. The supplement's own later algebra repeatedly reduces to denominators of order D^{2α}, confirming the displayed exponent is wrong, but the main-text lemma and theorem statements must be corrected before the lower bound can be accepted as stated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies federated functional mean estimation under heterogeneous federated differential privacy. For two designs, a common equispaced design and an independent uniform design, it states minimax rates governed by the fixed points D* defined in Eqs. (2) and (3), proposes two algorithms (a wavelet thresholding estimator for the independent design and a bagged local-polynomial estimator for the common design), and claims matching upper and lower bounds up to polylogarithmic factors. Theorems 1 and 2 give the minimax characterizations, Theorems 3 and 4 give the algorithmic upper bounds, and Theorems 5 and 6 give the lower bounds. The paper also compares the two designs in a homogeneous setting and illustrates the rates with simulations.","tokens_in":40421,"tokens_out":5600,"duration_ms":59304,"significance":"If the stated results hold, this is a substantial contribution: it extends federated-DP nonparametric theory to functional mean estimation, handles server-level heterogeneity in sample sizes, measurement counts, and privacy budgets, and reveals a structural advantage of independent designs over common designs in the privacy-dominated regime. A notable strength is that the rate-determining quantity D* is an implicit function of model parameters only, not fitted to data, and the homogeneous rates reduce to the known non-private rates of Cai and Yuan when the privacy constraints are removed. The paper also provides explicit, implementable algorithms and a detailed supplementary proof package. These features make the paper a useful step toward understanding the fundamental cost of privacy in functional data analysis.","major_comments":[{"comment":"The Van Trees prior term in Lemma 3 is displayed as 2^{2L(α+2)} in the main text, but the prior described in Supplement §8.1.1 is supported on [−2^{-L(α+1/2)}R, 2^{-L(α+1/2)}R] with per-coordinate Fisher information π²/a² = π² 2^{2L(α+1)}/R². The correct prior term is therefore of order 2^{2L(α+1)}, and as written Eq. (18) yields a lower bound of order D^{-2α-2}, not D^{-2α}. This is a load-bearing inconsistency because Lemma 3 is the stated engine for Theorems 1, 2, 5, and 6. The supplement's own algebra after Lemma 3 reduces to a denominator of order 2^{2αL}, which is consistent with the corrected exponent, suggesting the displayed value is a typo; nevertheless, the main-text lemma must be corrected and the proofs must use the corrected statement explicitly. The same issue appears in §4.2 and Supplement §8.2, where the prior term is written as π²(2L)^{2α+2}.","section":"§4.1, Eq. (18); Supplement §8.1.1"},{"comment":"The fixed point defining D* in Eq. (3) is not shown to have a unique solution. Unlike Eq. (2), for which the text gives a monotonicity argument, Eq. (3) defines D* as 'the largest number such that' an equality involving an infimum holds, but no existence or uniqueness proof is supplied. Since Theorems 2 and 5 depend on this quantity, the authors should either prove that a unique solution exists under the stated assumptions or add explicit conditions under which it does.","section":"§2.2, Eq. (3)"}],"minor_comments":[{"comment":"The text 'Similarly to the dependent design case' should read 'Similarly to the independent design case'; this is a typographical error in the caption discussion.","section":"§5.3"},{"comment":"The notation for the privacy parameter in the Gaussian mechanism is inconsistent: Algorithm 1 uses δ in the noise variance while Algorithm 2 uses δ_s. Since the servers are heterogeneous, δ_s should be used in both places.","section":"Algorithms 1 and 2"},{"comment":"The claim that the uniform-design assumption can be relaxed to non-uniform distributions 'bounded above and below by a constant' is only asserted, not demonstrated. The stated theorems are for uniform or equispaced designs; please either add the non-uniform derivation or explicitly state the restricted scope of the theorems.","section":"§1.1"},{"comment":"In the proof of Lemma 3, the sentence 'Combining this with the earlier identities of Lemmas 3 and 4' refers to Lemma 3 itself; this should be rephrased to refer to the relevant trace bounds and the main-text Lemma 4.","section":"Supplement §8.1.1"}],"recommendation":"major_revision","confidential_remarks":"The Van Trees exponent issue is a genuine written inconsistency in a load-bearing proof step, but it appears to be a fixable typo rather than a fatal flaw: the supplement's own algebra is consistent with the corrected exponent. The fixed-point uniqueness question for Eq. (3) is also fixable with a short lemma. I recommend major revision rather than reject, and I would be willing to review a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: this is a solid paper that nails the minimax rates for federated differentially private functional mean estimation under heterogeneous server parameters, and the common-vs-independent design comparison is genuinely new and worth publishing. The main theorems are plausible, and the supplement is substantive. But I checked the stress-test concern about Lemma 3, and it's real: the displayed denominator term is 2^{2L(α+2)}, while the prior Fisher information computed in Supplement 8.1.1 is of order 2^{2L(α+1)}. Taken literally, Lemma 3 gives a D^{-2α-2} lower bound, not the D^{-2α} claimed in Theorems 1, 2, 5, and 6. The supplement's own later algebra uses the correct order, so this looks like a typo rather than a fatal gap, but it is load-bearing as displayed and must be corrected.\n\nWhat's new and good: the paper extends the FDP framework of Cai-Chakraborty-Vuursteen to functional mean estimation, handles full heterogeneity in n_s, m_s, and ε_s, proposes two algorithms that provably match the lower bounds up to polylogs, and the homogeneous special cases reduce correctly to the classical non-private rates of Cai-Yuan. The four-bottleneck Fisher information decomposition in the lower-bound proof is a genuine technical advance. The acknowledgment of concurrent work [33] is honest, and the claimed improvement over that work is specific.\n\nSoft spots: (1) the Lemma 3 exponent issue; (2) the small-δ condition is used throughout but its restrictiveness (e.g., whether it rules out pure DP) is never discussed; (3) the simulations have no error bars, no run counts, the value of α is not reported, and no code is shipped, so the empirical claims cannot be independently verified. The design assumptions are clearly stated; the claimed relaxation to non-uniform designs with bounded densities is only sketched, which is fine for a main-text presentation but should be flagged.\n\nWho this is for: statisticians and theorists working on private functional data analysis and federated learning. It deserves a serious referee. My recommendation: send it to peer review, and require a revision that fixes the Lemma 3 display, adds a remark on the δ condition, and makes the simulations reproducible.","headline":"Strong minimax theory for private functional mean estimation, but the lower-bound lemma has a real typo that must be fixed before the theorems are accepted as stated.","tokens_in":675,"tokens_out":709,"would_cite":true,"duration_ms":49321,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G20","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that the minimax risk of federated differentially private mean-curve estimation is set by a single fixed-point number $D_{\\ast}$, giving exact rates for both shared and per-user measurement grids.","keywords":["Federated learning","Differential privacy","Functional mean estimation","Minimax rates","Heterogeneous privacy constraints","Wavelet projection","Local polynomial regression","Privacy-accuracy trade-off"],"falsifier":"In the homogeneous common-design setting with fixed $n$, small $\\varepsilon$, and $\\alpha>1/2$, the theory predicts the privacy term dominates and the MSE should be independent of $m$; run Algorithm 2 with $n=200$, $\\varepsilon=0.5$, $\\alpha=1.5$, $\\delta=1/n^2$, and $m=32,64,128$ on curves with Hölder regularity $1.5$ and compare MSE. If the MSE keeps decreasing as $m$ grows in that regime, the fixed-point rate is not the true limit.","tokens_in":39875,"feed_emoji":"🔒","tokens_out":10495,"duration_ms":96159,"temperature":0.7,"pith_summary":"This paper establishes the minimax risk for estimating a population mean function from discretely sampled curves held across several servers, each with its own number of subjects, number of measurements, and differential-privacy budget. The central claim is that the risk is, up to logarithmic factors, $\\eta_N D_{\\ast}^{-2\\alpha}$, where $D_{\\ast}$ is the solution to a fixed-point equation that balances measurement density, sample size, and privacy noise across servers. The paper also shows that when each individual has private, random design points (independent design), the privacy-dominated rate is faster than when all individuals share one grid (common design). Two one-shot algorithms—a clipped wavelet projection with anisotropic Gaussian noise and a bagged local-polynomial smoother—attain these rates. If correct, this gives a sharp, server-by-server account of the privacy-accuracy trade-off in federated functional data analysis.","feed_headline":"One fixed-point equation sets the cost of private federated curves","feed_subtitle":"A single number sets the exact privacy limit, and two practical algorithms attain it.","key_machinery":"The rate-determining object is $D_{\\ast}$, the 'effective dimension' solving the fixed-point equations; $D_{\\ast}$ determines the optimal truncation level $L=\\log_2 D_{\\ast}$ and can be read as the point where design complexity and statistical/privacy information balance. The upper bound for independent design is a truncated wavelet projection estimator: per-subject contributions are clipped at level $\\tau_l^{(s)}$, rescaled coordinate-wise by $(\\tau_l^{(s)}\\sqrt{2^l\\wedge m_s})^{-1}$, and perturbed with Gaussian noise whose variance is set by the $\\ell^2$ sensitivity $c_A\\sqrt{L}/n_s$, after which servers' coefficients are aggregated with inverse-variance weights $u_l^{(s)}=n_sm_s\\wedge m_s2^{l(2\\alpha+1)}\\wedge 2^{-l}m_sn_s^2\\varepsilon_s^2\\wedge 2^{2l\\alpha}n_s^2\\varepsilon_s^2$. The common-design upper bound uses privatized per-grid-point means with Gaussian noise of variance proportional to $m/(n_s^2\\varepsilon_s^2)$, followed by bagging (grouping the $m$ points into $m_0=D_{\\ast}$ sized groups) and local polynomial regression. The lower bounds are built on a multivariate van Trees inequality: the minimax risk is bounded below by $2^{2L}/(\\sup_f \\sum_s \\min\\{\\mathrm{Tr}(I^{Y|T}), \\mathrm{Tr}(I^Y), \\mathrm{Tr}(I^{X|T}), \\mathrm{Tr}(I^X)\\} + 2^{2L(\\alpha+2)})$, and each Fisher-information trace is controlled through differential-privacy data-processing inequalities, with the random function $X_i$ constructed via $\\mathrm{Beta}(2,2)$ coefficients to isolate the privacy bottleneck.","core_discovery":"The paper's central discovery is that the minimax integrated squared error for federated differentially private estimation of a mean function $f$ in the Hölder class $H^\\alpha(R)$ is $\\eta_N D_{\\ast}^{-2\\alpha}$, with $D_{\\ast}$ defined by the fixed-point equation $D^{2\\alpha}=m^{2\\alpha}\\wedge\\sum_s \\min(n_s, D^{-1}n_s^2\\varepsilon_s^2)$ in the common-design case and by $D_{\\ast}^{2\\alpha}=\\inf_{1\\le D\\le D_{\\ast}}\\sum_s \\min(D^{-1}n_sm_s, D^{-2}m_sn_s^2\\varepsilon_s^2, D^{2\\alpha}n_s, D^{2\\alpha-1}n_s^2\\varepsilon_s^2)$ in the independent-design case. The four terms in the independent-design equation correspond to four information bottlenecks: sampling error, measurement error, and their privacy-constrained counterparts. In the homogeneous setting these formulas yield explicit rates, and a comparison shows that whenever privacy terms dominate, independent designs converge strictly faster than common designs, so protecting the design points does not come at a statistical cost. The paper proves matching upper and lower bounds up to polylogarithmic factors, and the lower-bound argument introduces Fisher-information data-processing inequalities that handle server heterogeneity.","pith_inferences":["Solving the fixed-point equation in advance could be used by a system designer to allocate privacy budgets across servers: a server contributes to the rate only through its summand in the fixed-point equation, so budget should be spent where the summand is currently smallest.","The common-versus-independent comparison suggests a testable policy implication: adding random jitter to otherwise shared measurement grids may lower privacy cost despite making the design points sensitive information, because the diversity of design points dominates the extra noise needed to protect them.","The Fisher-information data-processing technique for federated differential privacy likely extends to other functional targets—covariance operators, regression functions, or derivative curves—where similar fixed-point rates with four-way bottlenecks would appear.","For non-uniform designs the paper only sketches an adaptation; an explicit construction with density-whitening and privacy-preserving histogram estimation would make the independent-design result fully general."],"forward_implications":["The privacy-accuracy trade-off in federated functional mean estimation is fully quantified: for any heterogeneous configuration of subjects, measurements, and budgets, the minimax rate is read off from a single fixed-point equation.","In the common-design regime, extra measurements per subject do not improve the rate once privacy constraints bind, whereas in the independent-design regime they do; privacy-conscious studies should prefer independent measurement schedules.","The proposed algorithms are one-shot—each server transmits a single transcript—so they fit existing horizontal federated architectures without interactive rounds.","Sending $\\varepsilon_s\\to\\infty$ recovers the classical non-private minimax rates for discretely sampled functional mean estimation, and the $S=1$ and $n_s=1$ limits recover central and local differential privacy as special cases.","The independent-design result provides new minimax rates even in the non-private heterogeneous case where different subjects have different numbers of measurements."],"supporting_citations":[{"why":"Supplies the federated differential privacy framework, the van Trees-based lower bound technique for heterogeneous servers, and the privacy-preserving adaptation to non-uniform designs.","marker":"[8]"},{"why":"Establishes the classical non-private minimax rates for discretely sampled functional mean estimation that this paper recovers as $\\varepsilon_s\\to\\infty$ and benchmarks against.","marker":"[9]"},{"why":"Develops the Fisher-information-based van Trees lower bounds for local differential privacy that the paper extends to federated settings.","marker":"[5]"},{"why":"Provides the multivariate van Trees inequality used in both lower-bound proofs.","marker":"[17]"},{"why":"Gives the Daubechies wavelet construction that forms the basis for the truncated projection estimator.","marker":"[12]"},{"why":"Provides the Gaussian mechanism and its privacy guarantees used to privatize clipped coefficients.","marker":"[15]"},{"why":"Supplies the local-polynomial weight bounds and polynomial interpolation property used in the common-design estimator analysis.","marker":"[29]"},{"why":"Gives the data processing inequality for Fisher information that controls the privacy-induced information loss in the lower bounds.","marker":"[34]"}],"fun_headline_variants":["Independent designs beat common ones under privacy limits","Fixed-point equation sets privacy-accuracy tradeoff for curves","Matching bounds for private functional means in federated learning","Private federated curves: one equation fixes the minimax rate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fixed-point rate $D_{\\ast}^{-2\\alpha}$ is proven only for equispaced grids in the common design and i.i.d. uniform design points in the independent design; the paper only sketches the extension to general densities, so clustered or server-dependent sampling designs could break the claimed rate.","fun_headline_variants_meta":{"raw":{"variants":["Independent designs beat common ones under privacy limits","Fixed-point equation sets privacy-accuracy tradeoff for curves","Matching bounds for private functional means in federated learning","Private federated curves: one equation fixes the minimax rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2923,"prompt_tokens":999,"completion_tokens":1924,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":1868}},"tokens_in":615,"tokens_out":1924,"duration_ms":14290,"temperature":1.0,"reasoning_tokens":1868,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:59:17.007699+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the homogeneous common-design setting with fixed $n$, small $\\varepsilon$, and $\\alpha>1/2$, the theory predicts the privacy term dominates and the MSE should be independent of $m$; run Algorithm 2 with $n=200$, $\\varepsilon=0.5$, $\\alpha=1.5$, $\\delta=1/n^2$, and $m=32,64,128$ on curves with Hölder regularity $1.5$ and compare MSE. If the MSE keeps decreasing as $m$ grows in that regime, the fixed-point rate is not the true limit.","supporting_citations":[{"cited_title":"Tony Cai and Ming Yuan","cited_arxiv_id":null,"evidence_quote":"Establishes the classical non-private minimax rates for discretely sampled functional mean estimation that this paper recovers as $\\varepsilon_s\\to\\infty$ and benchmarks against."},{"cited_title":"Fisher information under local differential privacy","cited_arxiv_id":null,"evidence_quote":"Develops the Fisher-information-based van Trees lower bounds for local differential privacy that the paper extends to federated settings."},{"cited_title":"Applications of the van trees inequality: a bayesian cram´ er-rao bound.Bernoulli, pages 59–79, 1995","cited_arxiv_id":null,"evidence_quote":"Provides the multivariate van Trees inequality used in both lower-bound proofs."},{"cited_title":"Tsybakov","cited_arxiv_id":null,"evidence_quote":"Supplies the local-polynomial weight bounds and polynomial interpolation property used in the common-design estimator analysis."},{"cited_title":"∇fL log dPT f dµ′ !#","cited_arxiv_id":null,"evidence_quote":"Gives the data processing inequality for Fisher information that controls the privacy-induced information loss in the lower bounds."}],"review_version":1}