{"id":"42a51a40-aadb-496c-a31d-7da14a6eae32","arxiv_id":"2410.02260","paper_version":3,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"FedScalar achieves federated learning with constant scalar uploads via random vector inner products, proving O(d/sqrt(K)) convergence to stationary points for smooth non-convex losses while reducing variance with Rademacher vectors.","lead":"FedScalar lets each federated learning client send only two scalar numbers per round instead of the full high-dimensional model update. This fixed communication cost could make collaborative training feasible on networks where sending large vectors is impossible due to bandwidth limits.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest assumption correctly isolates the reconstruction step; that step is valid and directly supports the convergence claim. No load-bearing gap is present, so the UNVERDICTED verdict is unchanged.","tokens_in":1714,"tokens_out":230,"duration_ms":27495,"concrete_test":"Compute the exact second-moment E[‖(<Δ,r>)r − Δ‖²] analytically for Rademacher r (fixed ‖r‖² = d) versus isotropic Gaussian r and confirm the paper's variance-reduction statement matches the derived expressions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The mechanism of transmitting <Δ, r> plus seed and reconstructing the unbiased estimator scalar × r holds for both Gaussian and Rademacher r (E[<Δ, r> r] = Δ when coordinates are independent, zero-mean, unit variance). The O(d/√K) rate is the expected consequence of the estimator variance scaling as O(d ‖Δ‖²). No internal inconsistency appears in the stated claim or weakest assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes FedScalar, a federated learning algorithm for bandwidth-constrained settings in which each client transmits only two scalars per round: the inner product of its local update difference with a locally generated random vector, plus the seed. The server reconstructs an unbiased estimate of the update from this information. The manuscript proves a convergence rate of O(d/√K) to a stationary point for smooth non-convex objectives, shows that Rademacher random vectors yield lower aggregation variance than Gaussian vectors, and reports simulations demonstrating wall-clock and energy gains over FedAvg and QSGD.","tokens_in":1797,"tokens_out":376,"duration_ms":21014,"significance":"If the convergence analysis is correct, the work supplies a theoretically supported method for extreme communication reduction (constant scalars independent of d) while retaining an explicit rate for non-convex FL. The explicit variance comparison between Rademacher and Gaussian vectors, together with the simulation evidence of practical efficiency, strengthens the contribution for resource-limited networks. The mechanism relies only on standard unbiasedness properties of zero-mean unit-variance random vectors and does not appear to introduce circularity or free parameters.","major_comments":[],"minor_comments":[{"comment":"Abstract: the statement that the upload cost is 'dimension-free' is correct for communication volume, but the O(d/√K) rate explicitly depends on dimension through the estimator variance; a brief parenthetical clarification would prevent misreading.","section":"Abstract"},{"comment":"The convergence theorem statement should include the precise dependence on smoothness constant, variance bound, and number of clients to make the O(d/√K) claim fully traceable without re-deriving the variance term.","section":"Convergence Analysis"}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the positive review and recommendation of minor revision. The summary accurately reflects the FedScalar contributions to scalar communication in federated learning.","responses":[],"tokens_in":1249,"tokens_out":49,"duration_ms":15695,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this paper gives a concrete way to make federated learning work when bandwidth is the hard limit: each client sends only a scalar inner product and a seed, letting the server reconstruct an unbiased estimate without shipping the full update vector. Communication cost stays fixed at two numbers no matter how big d gets. They back this with a convergence proof for smooth non-convex losses and show Rademacher vectors cut aggregation variance relative to Gaussian ones, plus some simulation results on time and energy versus FedAvg and QSGD. That combination of scalar encoding, seed reconstruction, and the specific variance comparison is not in the baselines they cite. The unbiasedness follows from basic properties of zero-mean unit-variance random vectors, and the proof is presented separately from the experiments. The rate O(d/sqrt(K)) is the direct result of the estimator variance scaling linearly with dimension, which is standard for this style of compression. The soft spot is exactly that d factor in the bound. It does not disappear, so for very high-dimensional models the number of rounds needed still grows with d even though each round is cheap. The abstract does not give the full derivation or the precise experimental setup for the bandwidth constraints, so it is not possible to verify the variance calculations or how closely the runs track the theory. This work is for people who need to run FL under tight bandwidth budgets and are willing to accept a dimension-dependent iteration count in exchange for fixed per-round cost. A reader looking for practical mechanisms with some supporting analysis would find the reconstruction trick and the Rademacher comparison useful to check. It deserves a serious referee because the central claim is falsifiable, the mechanism is simple to implement, and the problem it targets is real. I would send it out for review.","headline":"FedScalar cuts client uploads to two scalars via random inner products and seeds, with a stated O(d/sqrt(K)) rate and Rademacher variance edge, but the dimension factor in the bound is the expected cost of that estimator.","tokens_in":2270,"tokens_out":447,"would_cite":false,"duration_ms":16261,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"FedScalar scalar-projection upload has no structural overlap with RS distinction-to-physics forcing","alignment":"orthogonal","rationale":"Paper centers on unbiased random-vector estimators (Gaussian/Rademacher) for gradient reconstruction and O(d/√K) non-convex convergence; RS modules (AbsoluteFloorClosure, Cost/FunctionalEquation, AlexanderDuality, ArithmeticFromLogic, etc.) derive J-cost, φ-ladder, 8-tick periodicity and constants from bare distinguishability. No shared machinery or theorems.","tokens_in":50490,"confidence":"high","tokens_out":126,"duration_ms":4873,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"FedScalar achieves federated learning convergence by sending only two scalars per round via random inner products.","keywords":["federated learning","communication efficiency","scalar communication","bandwidth constraints","convergence analysis","random projections","non-convex optimization"],"falsifier":"A controlled test that averages many reconstructed estimates for known local updates and checks whether the bias is zero within sampling error, or an empirical measurement of convergence speed versus the predicted O(d/√K) scaling on a smooth non-convex problem.","tokens_in":2598,"feed_emoji":"📡","tokens_out":671,"duration_ms":31183,"temperature":0.7,"pith_summary":"The paper introduces FedScalar to solve the bandwidth bottleneck in federated learning, where agents cannot afford to upload full high-dimensional model updates repeatedly. Each agent computes the inner product of its local update difference with a random vector generated from a seed and sends only that scalar value plus the seed. The server regenerates the vector from the seed to form an unbiased estimate of the gradient. This yields a convergence rate of O(d/√K) to a stationary point for smooth non-convex losses. The analysis further shows that Rademacher random vectors produce lower aggregation variance than Gaussian vectors, and simulations demonstrate gains in wall-clock time and energy use under tight bandwidth limits.","feed_headline":"FedScalar cuts federated learning uploads to two scalars per round","feed_subtitle":"Agents encode updates as inner products with seeded random vectors to keep gradient estimates unbiased and achieve O(d/√K) convergence.","key_machinery":"The scalar encoding via inner product with a seed-generated random vector, which replaces the full d-dimensional update while preserving an unbiased gradient estimate at the server.","core_discovery":"FedScalar encodes each local update difference as its inner product with a locally generated random vector and transmits the resulting scalar together with the seed. The server reconstructs an unbiased gradient estimate from this information alone. For smooth non-convex loss functions the method converges to a stationary point at rate O(d/√K). Replacing Gaussian random vectors with Rademacher ones reduces the variance of the aggregated estimates.","pith_inferences":["The dimension dependence in the rate implies that reaching a target accuracy in high-dimensional models requires proportionally more communication rounds.","The same scalar projection approach could be applied to other distributed optimization problems that face similar upload limits.","If seeds are public, the reconstruction step is deterministic and could be implemented with minimal additional computation on the server."],"forward_implications":["The convergence rate of O(d/√K) holds for smooth non-convex loss functions.","Rademacher random vectors reduce aggregation variance relative to the Gaussian case.","Upload communication cost remains two scalars independent of model dimension d.","Numerical results show improved wall-clock time and energy efficiency compared with FedAvg and QSGD under bandwidth constraints."],"fun_headline_variants":["FedScalar limits uploads to two scalars per round","FedScalar encodes updates via random vector inner products","Server reconstructs gradients from scalars and seeds in FedScalar","FedScalar achieves O(d/sqrt K) convergence for nonconvex losses"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The inner product of the local update difference with a locally generated random vector, together with the transmitted seed, enables the server to reconstruct an unbiased gradient estimate without any high-dimensional transmission.","fun_headline_variants_meta":{"raw":{"variants":["FedScalar limits uploads to two scalars per round","FedScalar encodes updates via random vector inner products","Server reconstructs gradients from scalars and seeds in FedScalar","FedScalar achieves O(d/sqrt K) convergence for nonconvex losses"]},"model":"grok-4.3","cost_usd":0.007456,"raw_usage":{"total_tokens":3413,"prompt_tokens":646,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":74562000,"prompt_tokens_details":{"text_tokens":646,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2703,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":646,"tokens_out":64,"duration_ms":16005,"temperature":1.0,"reasoning_tokens":2703,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-23T19:37:55.631802+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled test that averages many reconstructed estimates for known local updates and checks whether the bias is zero within sampling error, or an empirical measurement of convergence speed versus the predicted O(d/√K) scaling on a smooth non-convex problem.","supporting_citations":[],"review_version":1}