{"id":"aa4e6335-cb6b-4a22-b953-dbaa041a18b6","arxiv_id":"2501.07879","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The minimax L2 rate for distributed nonparametric estimation is (Ness)^(-2r/(2r+1)) up to logarithmic factors, where Ness is a single phase-transition formula of m, n, l, and r.","lead":"Many distributed machines, each holding n private samples and allowed l communication bits, cooperate to estimate an unknown smooth function. This paper derives the optimal error rate for every regime, including the phase transitions as samples per machine go from sparse to dense.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Upper bound relies on unproved Lemma 2 from the authors' prior preprint; Theorem 1's claimed rate is contingent on that external result.","rationale":"The paper contributes a plausible and detailed framework: the outer wavelet layer converts nonparametric estimation into parametric distribution estimation, and the five-case characterization of Ness is coherent. The appendices give substantial proofs for the lower bound and for verification of Assumptions 1-3 on five models. However, the upper bound is not self-contained: every invocation of Lemma 2 in the four cases of Theorem 4 replaces a proof with a citation to the authors' own preprint [8]. The reader's weakest-assumption analysis identified exactly this point, and my review agrees. I did not find an independent internal contradiction that would force a stronger verdict; the concern is about verification, not about a specific mathematical step that is demonstrably wrong within the text. A CONDITIONAL verdict is therefore appropriate: the paper should either prove Lemma 2 fully or make the dependence on [8] precise and verifiable. The concrete test above would resolve whether the concern actually lands.","tokens_in":36138,"tokens_out":42208,"duration_ms":362127,"concrete_test":"Obtain the full statement and proof of Theorem 1 of [8] and verify that it implies exactly Lemma 2 items 1-3, including the regime n < k ≤ (2^l − 1)n with error O(log(k/n + 1)/(ml)) and its transition to item 3. If [8] uses a different loss, a different protocol, or has a different regime threshold, the inner-layer rates used in Theorem 4 are unsupported and the upper bound in Theorem 1 fails for the affected cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central upper bound, Theorem 4, is built around Lemma 2, which supplies the inner-layer protocol for parametric distribution estimation. Lemma 2 is not proved in this paper: items 1-3 are cited to the authors' own prior preprint [8], and only item 1' is proved in Appendix A-C. Because Lemma 3 multiplies the inner-layer error by K log^2 N, any error in Lemma 2's rates—especially at the regime boundaries between items 2 and 3, where the error changes from logarithmic to exponential in l—propagates directly into the claimed minimax rate. The OCR text renders the thresholds as '(2l − 1)·n', but the matching error rates suggest the intended threshold is (2^l − 1)·n; if the true threshold differs, the application of items 2 and 3 in Cases 1 and 2 may fall outside their domains. Thus the main theorem's upper bound, and hence the claimed full characterization, is not self-contained at its most load-bearing point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper characterizes the minimax L2 rate of distributed nonparametric estimation when each of m terminals holds n i.i.d. samples and sends l bits, under Sobolev regularity of order r. The main result, Theorem 1, states that the rate is (Ness)^{-2r/(2r+1)} up to polylogarithmic factors, where Ness is an effective sample size depending on m, n, l, and r. The upper bound is obtained by a two-layer protocol: a wavelet outer layer converts the problem into parametric distribution estimation, and an inner layer invokes a protocol for that parametric problem. The lower bound uses a wavelet sieve, information-theoretic inequalities based on strong data processing inequalities, and balls-and-bins estimates. The paper also verifies the required assumptions for density estimation and for Gaussian, binary, Poisson, and heteroskedastic regression models, yielding rate results for those special cases.","tokens_in":36306,"tokens_out":19442,"duration_ms":191529,"significance":"If the proof is correct, this resolves a question left open by prior work that treated only dense samples or only n=1: it covers all relative sizes of m and n and identifies phase transitions in the dependence on the communication budget l. The construction is elegant in principle, and the lower-bound machinery, in particular the extension of strong data processing inequalities to irregular events, is potentially reusable. The paper is also commendably explicit about its assumptions and verifies them on several concrete models. However, the central upper bound depends on an unproved external lemma from the authors' own preprint, and there is a suspected gap in the proof of the main lower-bound lemma for the polynomial regimes; both points are load-bearing and need to be addressed before the characterization can be considered established.","major_comments":[{"comment":"Lemma 2 is the core of the upper bound: items 1)-3) are used in Cases 1-4 of the error analysis, yet only item 1') is proved in this manuscript; items 1)-3) are cited to the authors' own preprint [8]. The claimed minimax characterization is therefore not self-contained at its most load-bearing point. I ask the authors to include a proof of Lemma 2, or at least to state it with a complete proof in an appendix, and to make explicit which regimes depend on [8]. If the paper is meant to rely on a published version of [8], that should be stated clearly.","section":"§III-B, Lemma 2; §IV-B"},{"comment":"The notation '2l' in Lemma 2 items 2)-3), in equation (2), and in the error analyses of Section IV-B is ambiguous; it appears to mean 2^l (e.g., the error term k/(2lmn) in item 3 and the Case 1 verification rely on an exponential factor), but the text as written could be read as 2l. The distinction is load-bearing because it determines whether item 2 or item 3 applies at the boundary nn^{2r}$, and $(mn)^{1/(2r+1)}$ for $m\\le n^{2r}$.","pith_inferences":["An extension the authors leave implicit: the same balls-and-bins tensorization should yield lower bounds for heterogeneous per-terminal sample sizes (different $n_i$ per terminal), with $\\sum_i n_i$ replacing $mn$ in the occupancy bounds; the paper does not state this.","Another consequence of the same machinery: the layered protocol can be read as a design rule — pick the bit budget $l$ by inverting the rate formula for a target error — even though the paper presents the protocol only as an achievability proof.","A testable extension: run the protocol with unequal $n_i$ and check whether the rate is governed by the same effective-sample-size formula with $mn$ replaced by $\\sum_i n_i$; the paper's fixed-$n$ analysis does not cover this."],"forward_implications":["For all parameter regimes covered by Assumptions 1 through 3, the minimax $L^2$ error is $N_{\\mathrm{ess}}^{-2r/(2r+1)}$ up to logarithmic factors, so the communication-constrained problem is equivalent in rate to a centralized problem with $N_{\\mathrm{ess}}$ samples.","In the sparsest regime $m\\ge n^{2r+1}$ with $1\\le l\\le \\frac{1}{2r+1}\\log(m/n^{2r+1})$, the optimal rate decays exponentially in $l$; in all other regimes it depends on $l$ only polynomially.","To match the centralized rate, terminals need about $n$ bits each when $m>n^{2r}$, and about $(mn)^{1/(2r+1)}$ bits each when $m\\le n^{2r}$.","The same theorem supplies minimax rates for nonparametric density estimation and for Gaussian, binary, Poisson, and heteroskedastic regression as direct corollaries.","Previously solved special cases, such as dense samples with $m