{"id":"bcca3a60-a52d-4c84-be92-3d59997cbb1f","arxiv_id":"2605.25306","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"With a fractional-power nonlinearity applied to coordinate zero-order estimates and primal-only communication, ZOOM-PB keeps the O(√(p/(nT))) nonconvex stationarity rate and O(p/(nT)) PL rate.","lead":"ZOOM-PB is a distributed black-box optimization algorithm that applies a fractional-power \"powerball\" map to each noisy zero-order estimate while agents exchange only their current decision vectors. Its claims: the standard nonconvex convergence rate is preserved, and weak-signal regimes converge faster in practice.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's second-moment bound E||ε||²≤Cp (Eq. 22) is false for the coordinate-subset estimator when n_c<p; the honest variance is Θ(p²/n_c), so Theorem 1's advertised O(√(p/(nT))) is not established for the advertised small-subset regime.","rationale":"The paper proposes a sensible engineering idea—inserting a fractional-power gain into a primal-only distributed zeroth-order recursion—but the proof of the headline rate is internally inconsistent. The central moment bound (Eq. 22) does not follow for the coordinate-subset estimator with n_c<p; standard coordinate-sampling algebra gives Θ(p²/n_c). This is not a loose-constant issue: it changes the p-dependence of the final rate. The paper explicitly advertises small coordinate subsets and distributed exploration of coordinate subspaces, which is precisely the regime where the proof fails. The same flaw affects the γ=1 endpoint, so Proposition 2 does not rescue the analysis. With a fix—either proving a variance bound that honestly accounts for n_c or specializing to n_c=Θ(p)/full-coordinate queries—the rate claim may be recoverable in a revised form, but as written the central theorem is not established. The numerical comparisons lack code/seeds/error bars, but that is secondary; the proof gap is decisive.","tokens_in":11254,"tokens_out":9406,"duration_ms":104926,"concrete_test":"Analytically re-derive E[||g̃_i,k||²] for the two-point estimator (5) with n_c as an explicit parameter, under Assumption 3, and compare with Eq. (22). For a concrete check, take p=100, n_c=1, f(x)=0.5||x||², x=(1,...,1), δ small, and no oracle noise; the estimator's second moment is exactly p·||∇f(x)||² = p² = 10^4, whereas Eq. (22) would predict O(p)=O(10^2). Running the estimator with 10^4 independent subset samples should reproduce this p² scaling. If the computation gives Θ(p²/n_c) rather than Cp, Lemma 1 is false and the dimension dependence in Theorems 1–2 collapses for small n_c.","verdict_should_be":"REJECT","load_bearing_attack":"The reader's concern is the controlling one. For the estimator in (4)/(5), condition on F_k and let a_i denote the vector of smoothed local coordinate derivatives (including stochastic-oracle noise). Uniform sampling of n_c coordinates with scaling p/n_c gives the exact identity E[||g̃_i||² | F_k] = (p/n_c) E[||a_i||²]. Under Assumption 3, E[||a_i||²] is Θ(p) in the worst case (each of p coordinates has constant-scale variance), so E[||ε_i,k||²] ≤ C p²/n_c, not C p. This invalidates Eq. (22). The agent-averaged variance in Eq. (24) then becomes C p²/(n_c n), the Lyapunov noise term in Eq. (34) is C p²η²/n_c instead of C pη², and the stationarity bound in Theorem 1 degrades to O((p/n_c)√(p/(nT))) plus the n/T transient. For the advertised 'small coordinate subsets' n_c≪p this is a large extra dimension factor; only n_c=Θ(p) recovers the claimed order, which defeats the coordinate-sampling contribution. Because γ cancels in the second-moment bound, the same flaw invalidates Proposition 2/endpoint ZOOM.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ZOOM-PB, a distributed stochastic zeroth-order method in which each agent samples n_c coordinates, forms a coordinate-wise finite-difference estimator scaled by p/n_c, applies a componentwise powerball nonlinearity sigma(g,gamma)=sgn(g)|g|^gamma, and runs consensus with only primal states. The central theoretical claim is that, under smoothness, gradient-noise, and bounded-gradient assumptions, the method attains O(sqrt(p/(nT))) stationarity in nonconvex problems and O(p/(nT)) PL objective residual, with an O(n/T) consensus transient. The proof rests on Lemma 1, which asserts a per-agent ZO estimator second moment O(p) and an averaged variance O(p/n). The numerical section isolates the powerball gain from the linear endpoint ZOOM.","tokens_in":11576,"tokens_out":14984,"duration_ms":158842,"significance":"The intended contribution — a primal-only nonlinear-gain ZO method that preserves known distributed ZO rates — is interesting and the problem is well motivated. The ablation against the linear endpoint is a good design, and the communication/query budgets are stated cleanly. If the rate claim held, the method would be a useful addition. However, the proof of the central rate does not withstand scrutiny: the coordinate-subset estimator's variance is Theta(p^2/n_c), so the advertised small-subset regime has a much worse dimension dependence. I do not see a circularity problem; the failure is a concrete moment estimate, not a fitted parameter.","major_comments":[{"comment":"For the coordinate estimator (4)/(5), conditioning on F_k and writing a_i as the vector of smoothed local coordinate derivatives, uniform sampling of n_c coordinates with scaling p/n_c gives E[||g_tilde_i||^2 | F_k] = (p/n_c) * sum_{j=1}^p E[a_{i,j}^2 | F_k]. Under Assumption 3 each E[a_{i,j}^2 | F_k] can be of order 1 in the worst case, so the second moment is Theta(p^2/n_c), not O(p). Eq. (22) is therefore false when n_c < p; the correct averaged variance in Eq. (24) is at best C p^2/(n_c n). The constants in Lemma 1 are allowed to depend on n_c, but the explicit p-dependence cannot be hidden. This invalidates the Lyapunov recursion (8b) and the convergence theorems that depend on it.","section":"Appendix A, Eq. (22); Lemma 1"},{"comment":"With the corrected variance, the stationarity bound degrades to O((p/n_c)*sqrt(p/(nT))) + O(n/T), and only n_c = Theta(p) recovers the advertised O(sqrt(p/(nT))). Similarly, the PL residual in Theorem 2 becomes O(p^2/(n_c n T)) rather than O(p/(nT)). Since the paper motivates 'small coordinate subsets' and assigns different agents to different subspaces, this is a load-bearing change: the central dimension-rate claim fails in the advertised regime. Proposition 2 inherits the same issue, because gamma=1 does not alter the estimator variance.","section":"Theorems 1-2; Proposition 2"},{"comment":"There is a second, independent problem with the smoothing schedule. With delta_k^2 <= C/(sqrt(p)*sqrt(n)*sqrt(k+1)), summing (13) gives p^2 * eta * sum delta_k^2 = Theta(p^{3/2} sqrt(T)/sqrt(n)), while the statistical term is p T eta^2 = n. After dividing by n T eta, the bias contribution is (p/n)*sqrt(p/(nT)), not O(sqrt(p/(nT))), unless p = O(n). Theorem 1 is not conditioned on p = O(n), so the stated rate is not established even if the estimator-variance issue is set aside. A different delta schedule or an explicit p/n restriction is needed.","section":"Appendix C, Eq. (13) and Theorem 1"}],"minor_comments":[{"comment":"The displayed rate appears as 'O(\\n p\\n p/(nT))' in the full text; this should be O(sqrt(p/(nT))).","section":"Abstract"},{"comment":"The text uses 'UA V' in several places (e.g., 'UA V swarm', 'UA V source seeking'); this should be 'UAV'.","section":"Section IV"},{"comment":"The ratio |sigma(g,gamma)|/|g| is undefined at g=0; Proposition 1 states g != 0, but Eq. (3) should state this restriction as well.","section":"Eq. (3) and Proposition 1"},{"comment":"The derivation of the consensus bound E[||x_T||_K^2/n] = O(p/T^2) is sketched as 'applying the disagreement part once more'; this step is not transparent and should be expanded, especially because the driver contains gradient terms that are not obviously bounded.","section":"Theorem 2 proof"}],"recommendation":"reject","confidential_remarks":"The central rate claim is not supported because of the estimator-variance miscalculation and, separately, the smoothing-bias schedule. A correction would require restated rates (with an explicit n_c and p/n dependence) or a restriction to n_c = Theta(p), which would defeat the paper's stated coordinate-subset contribution. I therefore recommend rejection, although the general research direction is sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Neat idea, careful writing, and a clean ablation against the linear endpoint — but the central rate proof is not supported as written. The coordinate-sampling variance bound in Lemma 1 is simply wrong for n_c<p, and that invalidates the dimension dependence in Theorems 1 and 2 for the advertised small-subset regime.\n\nThe genuinely new thing here is applying the fractional-power powerball map to coordinate ZO estimates inside a primal-only distributed consensus recursion. That combination isn't in the cited literature, and the paper is appropriately modest: it claims rate preservation, not a new minimax exponent. The empirical design is also good — comparing γ<1 against γ=1 under identical oracle and communication budgets isolates the gain effect. The authors state assumptions clearly and don't engage in any circular argument; the self-citations are used only as baselines.\n\nThe problem is Lemma 1. Eq. (22) asserts E||ε_i,k||^2 ≤ Cp after 'the usual coordinate-sampling scaling.' But for the estimator in (4)/(5), uniform sampling of n_c coordinates with scaling p/n_c gives, conditional on the history, E||g̃_i||^2 = (p/n_c) E||a_i||^2 where a_i is the smoothed local direction. Under Assumption 3, E||a_i||^2 is Θ(p) in the worst case, so the honest variance is Θ(p^2/n_c), not Θ(p). That changes the Lyapunov noise term from pη_k^2 to (p^2/n_c)η_k^2, and the stationarity bound degrades to O((p/n_c)√(p/(nT))). Only n_c=Θ(p) recovers the claimed order — which defeats the coordinate-sampling contribution the paper advertises. The same flaw carries into Proposition 2.\n\nA second, related soft spot: the nonlinearity's Jensen bias isn't controlled. Even if the variance bound were fixed, replacing E[σ(g̃_i,k)] with σ(∇f_i(x_i,k)) in the descent step needs a bias term from the noise distribution; the proof doesn't bound it.\n\nThese are fixable. One path is to prove the rate for the honest n_c scaling and show how small n_c degrades the bound; another is to restrict to n_c=p and present the powerball gain as the only novelty. Either way the current claim as written doesn't hold. The paper deserves a serious referee — the idea is worth engaging, and the fix is nontrivial but plausible. I'd send it to review with a specific request to check the variance computation and the Jensen bias, not desk-reject it.","headline":"Neat powerball-in-coordinate-ZO idea, but the headline rate is not proven: Lemma 1's variance bound misses the p^2/n_c scaling, so the advertised small-subset speedup collapses.","tokens_in":12096,"tokens_out":3355,"would_cite":false,"duration_ms":34098,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A distributed zeroth-order optimizer with a fractional-power feedback gain reaches the leading nonconvex rate O(sqrt(p/(nT))) while exchanging only primal states.","keywords":["distributed zeroth-order optimization","powerball map","coordinate sampling","primal-only consensus","nonconvex stationarity","Polyak–Łojasiewicz condition","black-box control","multi-agent systems"],"falsifier":"Compute the empirical second moment of the one-point coordinate estimator (4) with n_c=1 on a smooth quadratic in p=100 dimensions: if it grows like p^2 rather than p, then Lemma 1's bound (22) fails for small n_c and the theorem rates would need to be re-derived with an explicit n_c dependence.","tokens_in":11063,"feed_emoji":"⚡","tokens_out":4108,"duration_ms":41687,"temperature":0.7,"pith_summary":"The paper aims to establish that a coordinate-sampled distributed zeroth-order method can incorporate a fractional-power nonlinear gain—the powerball map—without losing the leading convergence rate. In nonconvex stochastic settings, the network-average squared gradient decays as O(sqrt(p/(nT))), and under the Polyak–Łojasiewicz condition the objective residual decays as O(p/(nT)). The update transmits only primal decision vectors, so the nonlinearity is a local feedback gain rather than a change in communication or oracle schedule. A reader would care because black-box networked control problems, such as multi-vehicle source seeking, often operate in flat regions where the estimated descent signal is weak and the gain accelerates the transient.","feed_headline":"Primal-only ZO optimizer hits leading rate with powerball gain","feed_subtitle":"A fractional-power feedback on coordinate estimates keeps O(sqrt(p/nT)) stationarity and quickens weak-signal transients.","key_machinery":"The load-bearing object is the powerball map sigma(g,gamma)=sgn(g)|g|^gamma, applied componentwise to each agent's coordinate-sampled ZO estimate. It amplifies small components, attenuates large ones, and preserves sign; combined with a primal-only consensus step and a Lyapunov function that couples average descent with disagreement, it allows the proof to bound the nonlinear residual directly from the raw oracle assumptions.","core_discovery":"The paper's central claim is that a componentwise fractional-power nonlinearity, sigma(g,gamma)=sgn(g)|g|^gamma for gamma in [1/2,1], can be inserted into a coordinate-sampled distributed zeroth-order recursion without losing the leading convergence order. When each agent builds a coordinate ZO estimate from a small random subset of coordinates, applies the powerball map locally, and exchanges only primal decisions with neighbors, the network-average gradient norm decays as O(sqrt(p/(nT))) in nonconvex settings and the objective residual decays as O(p/(nT)) under the Polyak–Łojasiewicz condition. The nonlinearity is thus a state-dependent feedback gain that reshapes weak versus large local e","pith_inferences":["The per-agent variance bound in Lemma 1 assumes E[||eps||^2] ≤ C p independent of subset size n_c; for the coordinate estimator in (4)-(5) the natural variance scales like p^2/n_c, so the stated p-dependence may only hold when n_c is proportional to p. If small n_c is used, the dimension term in the rates likely becomes worse.","If the analysis were reworked with an explicit n_c, one could trade off function-query budget against dimension dependence, which the current theorem statements leave implicit.","The powerball map may be interpreted as a per-coordinate adaptive step size; in that view, the paper shows such state-dependent gains can be inserted into distributed consensus without breaking the Lyapunov recursion, which may extend to event-triggered or compressed communication variants."],"forward_implications":["Adding the powerball gain costs nothing asymptotically: the same O(sqrt(p/(nT))) stationarity guarantee holds as for the linear endpoint, so practitioners can use it without sacrificing the order of convergence.","Under the PL condition, the objective residual O(p/(nT)) gives linear speedup in the number of agents for black-box objectives with a unique global optimum.","Because only primal states are exchanged, communication payload per round is one p-vector, independent of gamma and of the number of sampled coordinates.","The nonlinear gain is the sole difference from the linear endpoint in the experiments, so observed speed-ups are attributable to gain shaping rather than extra oracle calls."],"fun_headline_variants":["Powerball trick speeds distributed black-box optimization","Distributed ZO with powerball hits leading rate","Fractional-power mixing keeps ZO convergence order","Powerball gain sharpens networked ZO convergence","ZOOM-PB: powerball for faster distributed black-box"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof relies on each agent's coordinate-sampled ZO estimate having per-agent mean squared error O(p) independent of how few coordinates are sampled; for small subsets the natural variance is larger, roughly p^2/n_c, so the dimension dependence in the stated rates may be optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Powerball trick speeds distributed black-box optimization","Distributed ZO with powerball hits leading rate","Fractional-power mixing keeps ZO convergence order","Powerball gain sharpens networked ZO convergence","ZOOM-PB: powerball for faster distributed black-box"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000531,"raw_usage":{"total_tokens":2384,"prompt_tokens":727,"completion_tokens":1657,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":1583}},"tokens_in":471,"tokens_out":1657,"duration_ms":13857,"temperature":1.0,"reasoning_tokens":1583,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:12:41.590941+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical second moment of the one-point coordinate estimator (4) with n_c=1 on a smooth quadratic in p=100 dimensions: if it grows like p^2 rather than p, then Lemma 1's bound (22) fails for small n_c and the theorem rates would need to be re-derived with an explicit n_c dependence.","supporting_citations":[],"review_version":2}