{"id":"308bbe10-527f-4175-b2bd-33ee6d386051","arxiv_id":"2501.16174","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Energy distance with a Taylor approximation is proposed as a scalable feature-heterogeneity measure for federated learning, but the core approximation is underived, its multivariate form is inaccurate, and the proposed penalty-weighting benefit is untested.","lead":"This paper proposes using energy distance, a known statistical measure of distributional difference, to quantify feature heterogeneity across nodes in federated and distributed learning, and introduces a Taylor approximation that reduces its computational cost. It reports simulations where larger measured heterogeneity coincides with worse federated accuracy, and suggests the measure could guide penalty weights for model alignment, though that alignment step is not tested.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The multivariate Taylor approximation (Eq. 4) has a dimension-dependent leading constant that is wrong for d>1: for iid d-dimensional Gaussians it gives 0.75·sqrt(2d) instead of the exact 2·Γ((d+1)/2)/Γ(d/2), so every H value in §3.2 inherits a large relative error in the high-dimensional regime…","rationale":"The reader's weakest assumption identifies the multivariate Taylor approximation as the core defect, and the paper's text confirms this: Section 2.3 defers the derivation, and Eq. (4) is the only formula used to compute H values in Section 3.2. My independent check shows the problem is even more concrete than 'univariate thinking': the leading coefficient 3/4 in Eq. (4) is not the correct high-dimensional constant. For iid standard d-dimensional Gaussians, the exact E||X−Y|| is 2Γ((d+1)/2)/Γ(d/2), which asymptotically equals √(2d), while Eq. (4) gives 0.75√(2d). This 25% asymptotic error directly undermines the abstract's claim that the Taylor approximation 'preserves key theoretical quantitative properties.' The scale-invariance claim in Section 2.1 is also false—energy distance scales linearly under scaling—and the penalty-weight application in Section 4 is never tested. Because the primary quantitative contribution is unsupported and demonstrably inaccurate in the intended regime, the rejection verdict is appropriate; no verdict change is needed.","tokens_in":8550,"tokens_out":5507,"duration_ms":54346,"concrete_test":"Set d=10 and take X, Y iid N(0, I_d). Compute the exact expected distance E||X−Y|| = 2Γ(5.5)/Γ(5) ≈ 4.36 and compare with Eq. (4), which yields (3/4)√20 ≈ 3.35. If the discrepancy exceeds 20%, the approximation fails in the high-dimensional regime. Repeat with d=100 (exact ≈ √200, Eq. (4) ≈ 0.75√200) and with correlated features, e.g., Σ having off-diagonal 0.8, to confirm that the missing covariance term causes additional error. A Monte Carlo estimate with 10^7 samples can serve as a numerical check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Eqs. (2)–(4) approximate energy distance accurately enough to measure feature heterogeneity and to set penalty weights in distributed learning. The multivariate approximation in Eq. (4) is the load-bearing object: it is used to compute all H values in §3.2 and is the proposed basis for the untested penalty-weighting scheme in §4. The derivation is explicitly deferred (Section 2.3: 'Detailed derivations will be provided in the extended version'), but the formula as written already fails a basic consistency check. For X, Y iid N(0, I_d), ν_XY = 2d, C_4, C_3, and δ terms vanish, so Eq. (4) predicts E||X−Y|| ≈ (3/4)√(2d). The exact value is E||X−Y|| = 2Γ((d+1)/2)/Γ(d/2), which for large d behaves as √(2d). Thus the leading coefficient is consistently 0.75 instead of 1, giving about a 25% relative error for large d. The error is not negligible at the dimensions relevant to MNIST features either: for d=10, Eq. (4) gives 3.35 while the exact value is about 4.36. The formula also depends only on marginal variances, marginal kurtoses, and the scalar sums δμ1 and δμ2, so it is blind to correlations between feature dimensions and is not invariant to rotations, even though energy distance is rotation-invariant. This means the multivariate approximation is not a valid Taylor expansion of E||X−Y|| for d>1, and the heterogeneity measurements built on it are unreliable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using energy distance to measure feature heterogeneity across clients in federated and distributed learning, summarized by a normalized Energy Coefficient H. To avoid the quadratic cost of empirical energy distance, it introduces Taylor approximations based on summary statistics (Eqs. 2–4), reports simulations comparing approximate and exact formulas, and uses the Taylor approximation to compute H for MNIST clients in two feature-distribution regimes. It also proposes, without experiments, that H can be used to set penalty coefficients between guest and host models for better coordination.","tokens_in":8872,"tokens_out":9823,"duration_ms":89532,"significance":"If the Taylor approximations were sound, the O(nd) estimator and the normalized H coefficient would be useful practical tools for distributed learning; the univariate Gaussian formulas (5) and (7) appear correct, and the computational-time comparisons are a reasonable sanity check. However, the multivariate approximation used to compute every H value in the federated experiment has a leading-order error in high dimensions and ignores covariance structure, and the proposed penalty-weighting scheme is untested. The paper's central quantitative claims are therefore not established.","major_comments":[{"comment":"The multivariate Taylor approximation is not a valid approximation to E||X−Y|| for d>1. For X,Y iid N(0,I_d), ν_XY=2d and all correction terms vanish, so Eq. (4) gives E||X−Y||≈0.75√(2d). The exact value is E||X−Y||=2Γ((d+1)/2)/Γ(d/2), which tends to √(2d) as d grows; the formula therefore has a relative error of about 25% in high dimensions and is already substantially off at d=10 (3.35 vs 4.36). The formula also uses only marginal moments and the scalar sums δμ1 and δμ2, so it ignores all covariances and is not invariant to orthogonal rotations, contradicting the invariance claimed in Section 2.1. Since Section 2.3 defers the derivation to an extended version, the approximation cannot be verified, and every H value in Section 3.2 inherits this error.","section":"Section 2.3, Eq. (4)"},{"comment":"The federated-learning experiment computes H with the Taylor approximate method, i.e., with Eq. (4), but provides no validation against the empirical energy statistic (1) or exact expected distances on the MNIST feature space. The conclusion that the mean H of 0.657 'indicates the feature distributions between clients are different' is therefore not supported by independent evidence; at minimum, a comparison with the exact estimator would be needed before drawing any conclusion about heterogeneity or its effect on accuracy.","section":"Section 3.2"},{"comment":"The proposed use of H to set penalty coefficients between guest and host models is stated only as a possibility ('can be used', 'enables dynamic adjustments'). No algorithm, simulation, or theoretical analysis tests this mechanism, yet the abstract lists 'enhancing coordination' as a contribution. This claim is currently unsupported.","section":"Section 4"}],"minor_comments":[{"comment":"D^2 is stated to be 'invariant to scaling'; in fact, D^2(cX,cY)=c D^2(X,Y) for common scaling c>0. Only the normalized coefficient H is scale-invariant.","section":"Section 2.1"},{"comment":"The experiment does not specify whether H is computed on raw pixel vectors, what preprocessing is used before computing moments, or how client-pair H values are aggregated; these details are needed for reproducibility.","section":"Section 3.2"},{"comment":"The caption says 'MINST' instead of 'MNIST'.","section":"Figure 5 caption"},{"comment":"The text discusses Bernoulli distributions with p=0.05 and p=0.1, but Bernoulli is not listed in the experimental setup and is not identified in any panel of Figure 2; please clarify whether these results appear in the figure and how they were generated.","section":"Section 3.1"}],"recommendation":"reject","confidential_remarks":"The error in Eq. (4) is load-bearing and elementary, and the experiments are built on it, so I do not see a minor-revision path. If the authors rederive the multivariate approximation (including covariance terms or a correct Gaussian normalization), validate it on high-dimensional data, and test the H-based penalty scheme, the revised work could be considered as a new submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read. The novel piece is applying energy distance to feature heterogeneity in federated learning and proposing the normalized coefficient H. That is a sensible diagnostic, and the MNIST experiment does show H separates the mixed from the block-allocated clients. The computational speedup claim is also real: replacing pairwise distances with moment-based estimates is O(nd) instead of O(n^2 d).\n\nThe trouble is the paper's load-bearing math. The multivariate Taylor approximation, Eq. (4), is not a valid expansion of E||X-Y|| for d>1. For iid standard Gaussians it predicts (3/4)√(2d) while the exact value is 2Γ((d+1)/2)/Γ(d/2) ≈ √(2d). That's a ~25% error that persists in high dimensions, exactly the regime the paper targets. The formula uses only marginal moments, so it is blind to correlations and not rotation-invariant, though energy distance itself is. This is not a minor gap: every H value in Section 3.2 is computed with Eq. (4), so the heterogeneity numbers are unreliable.\n\nThere are other soft spots. The derivation is explicitly deferred to an 'extended version' (Section 2.3), which is honest but means the central claim is unverified. The paper also states energy distance is scale-invariant, which it is not; it scales linearly with the data. The penalty-weighting scheme in Section 4 is proposed but never tested. The single learning experiment shows the known effect that feature heterogeneity hurts FedAvg, not that H drives better aggregation.\n\nWhat's good: the univariate approximations (2)-(3) are standard delta-method results and are likely fine for near-Gaussian data with small kurtosis. The comparison of H across distributions in Figure 2 is a reasonable sanity check. If the authors correct the multivariate formula, or restrict the method to univariate or independent features, the paper could become useful.\n\nMy bottom line: I would not send this to peer review as it stands. The central approximation is demonstrably wrong in the multivariate case, and the authors have not supplied the derivation. The idea is worth a rewrite, but this version doesn't support its claims.","headline":"The FL heterogeneity idea is sensible, but the paper's load-bearing multivariate Taylor approximation is wrong for d>1 and the derivation is deferred, so the current version doesn't support its claims.","tokens_in":9437,"tokens_out":2810,"would_cite":false,"duration_ms":24981,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a Taylor-approximated energy distance, normalized into an Energy Coefficient H, can quantify feature heterogeneity across distributed-learning nodes at linear cost, and that H can set penalty weights for aligning…","keywords":["energy distance","feature heterogeneity","distributed learning","federated learning","Taylor approximation","Energy Coefficient H","non-IID data","distance measure"],"falsifier":"Compute the empirical energy distance between two independent samples from a d-dimensional standard normal, with d = 100 and n = $10^{4}$, and compare it with the value predicted by Eq (4); the exact expected distance is about $\\sqrt$(2d), while the approximation's leading term is (3/4)$\\sqrt$(2d), so a ratio near 1.33 would show a systematic 25 percent gap in the high-dimensional regime the paper targets.","tokens_in":8301,"feed_emoji":"📊","tokens_out":6046,"duration_ms":56020,"temperature":0.7,"pith_summary":"The paper tackles feature heterogeneity in distributed and federated learning, the setting where different nodes see different marginal distributions of predictors. It proposes the energy distance, normalized as the Energy Coefficient H, as a sensitive measure of these discrepancies, and shows that H is near zero for identically distributed clients and large for deliberately shifted feature distributions. Because the exact energy distance costs O($n^{2}$ d) pairwise operations, the paper derives Taylor approximations using means, variances, skewnesses, and kurtoses that reduce the cost to O(nd), and it reports simulations where these approximations track empirical values while speeding computation. The authors argue that H can act as a penalty weight between guest and host models, allowing heterogeneous nodes to align predictions and improve convergence. The central claim is that moment-based summaries can preserve the quantitative behavior of energy distance at scale.","feed_headline":"Energy coefficient H measures client data drift cheaply","feed_subtitle":"Taylor-approximated energy distance cuts pairwise cost to linear, giving federated nodes a cheap heterogeneity gauge.","key_machinery":"The central object is the Energy Coefficient H = $D^{2}$(X,Y)/(2E||X−Y||), a normalized energy distance in [0,1] that is zero only when the two feature distributions are identical. The load-bearing mechanism is the Taylor expansion of g(z) = $\\sqrt$(z) around the mean of (X−Y)^2: formula (2) approximates E||X−X'|| as $\\sqrt$(2)σ(1−(γ4+4)/16), formula (3) handles one-dimensional X versus Y, and formula (4) extends the idea to d dimensions by summing per-feature moments. This machinery is what converts a quadratic pairwise-distance computation into a linear-cost summary of marginal moments, and every H value reported in the MNIST experiments is computed through it.","core_discovery":"On its own terms, the paper claims that the energy distance, normalized as H = $D^{2}$(X,Y)/(2E||X−Y||), is a sensitive, non-parametric measure of feature heterogeneity among nodes in distributed and federated learning, and that this measure can be made scalable with a moment-based Taylor approximation. The one-dimensional formulas (2) and (3) and the d-dimensional formula (4) express expected Euclidean distances through means, variances, skewnesses, and kurtoses, reducing the empirical energy statistic's O($n^{2}$ d) pairwise computation to O(nd). The paper reports that H computed this way is near zero for identically distributed MNIST clients and large (mean 0.657) when clients receive different feature types, and that the high-H regime coincides with slower, less stable FedAvg convergence. It then argues that H can set the penalty coefficient between guest and host models, making heterogeneity itself a tunable input to coordination.","pith_inferences":["The approximation's 3/4 Gaussian baseline is univariate in origin: for d-dimensional Gaussians the exact expected distance coefficient tends to 1 as d grows, so Eq (4) may systematically understate distances in high dimension; a Monte Carlo check against d = 100 standard normals would settle this directly.","The paper proposes H-based penalty weights but does not run an experiment that uses them; the convergence results only correlate high H with poor FedAvg performance, leaving open whether H-weighted penalties actually improve alignment and against what baseline.","If H is computed from marginal moments only, the denominator 2E||X−Y|| and numerator D^2 may inherit correlated moment-estimation errors; whether the normalized H is more robust than raw energy distance to such errors is an untested consequence of its construction.","The treatment of representation heterogeneity via one energy distance per response value works for discrete labels but is acknowledged as open for continuous responses; a natural extension would be a functional or kernel-smoothed H over response values."],"forward_implications":["If the approximation is accurate enough, every pair of nodes in a federated system can be assigned an H value in [0,1] at linear cost, turning heterogeneity from an unmeasured nuisance into a monitored quantity.","MNIST experiments indicate that low H tracks stable FedAvg convergence, so H could flag clients whose feature distributions are drifting before accuracy collapses.","Because H is zero only when feature distributions match, it can support hypothesis tests for whether two nodes are sampling from the same distribution without permutation tests.","The proposed penalty-weighting scheme makes H a tunable input to distributed collaborative learning, allowing guest models to align with a host model proportionally to measured feature discrepancy."],"supporting_citations":[{"why":"Introduces the energy distance that the paper adopts as its heterogeneity measure.","marker":"Székely (2003)"},{"why":"Defines the empirical energy statistic and its asymptotic testing properties, which the paper relies on for hypothesis testing.","marker":"Székely and Rizzo (2005)"},{"why":"Motivates the robust skewness-and-kurtosis correction used to adjust the Taylor approximation toward exact Gaussian expectations.","marker":"Hampel et al. (2005)"}],"fun_headline_variants":["Energy coefficient H: cheap drift detection for federated nodes","Taylor-approximated energy distance makes heterogeneity measurable","Fast and sensitive: energy distance for feature drift","Cut energy distance cost to linear with Taylor","Heterogeneity gauge: energy distance for federated learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the multivariate squared Euclidean norm can be approximated from marginal moments alone: Eq (4) ignores all covariances between feature dimensions and fixes the Gaussian baseline at 3/4 of sqrt(nu), a coefficient calibrated in one dimension that does not match high-dimensional Gaussians, while the derivation is deferred to an extended version of the manuscript.","fun_headline_variants_meta":{"raw":{"variants":["Energy coefficient H: cheap drift detection for federated nodes","Taylor-approximated energy distance makes heterogeneity measurable","Fast and sensitive: energy distance for feature drift","Cut energy distance cost to linear with Taylor","Heterogeneity gauge: energy distance for federated learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1758,"prompt_tokens":837,"completion_tokens":921,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":846}},"tokens_in":453,"tokens_out":921,"duration_ms":8301,"temperature":1.0,"reasoning_tokens":846,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:39:28.573645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical energy distance between two independent samples from a d-dimensional standard normal, with d = 100 and n = $10^{4}$, and compare it with the value predicted by Eq (4); the exact expected distance is about $\\sqrt$(2d), while the approximation's leading term is (3/4)$\\sqrt$(2d), so a ratio near 1.33 would show a systematic 25 percent gap in the high-dimensional regime the paper targets.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the robust skewness-and-kurtosis correction used to adjust the Taylor approximation toward exact Gaussian expectations."}],"review_version":1}