{"id":"e538a5fe-6186-4c90-a558-e7d6fce58505","arxiv_id":"2501.06934","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Introduces conformal and holomorphic barycenters and Möbius-type probability families on hyperbolic balls, together with hyperbolic gradient and maximum likelihood estimation algorithms.","lead":"This paper develops a group-theoretic framework for machine learning in hyperbolic spaces, defining a conformal barycenter, a family of Möbius distributions, and gradient and maximum likelihood algorithms. It is a candidate foundation for making hyperbolic latent-space models more principled, though several mathematical errors need correction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Infinitesimal generator v2 is miscomputed, so Eq. (17) is not a combination of the stated G2 generators and Theorem 5 is unsupported.","rationale":"The strongest claim is that the paper gives detailed and rigorous answers to the four questions, with the conformal barycenter, its computation via Poincaré swarms, and the resulting MLE as central deliverables. The reader's weakest assumption targets exactly this computational core: the assertion that the swarm (17) evolves by Möbius group action. My own check confirms the reader's diagnosis. The derivative of m2(t) at t=0 is 1-z^2, not z^2-1, and the coefficients stated in the paper do not make (17) a linear combination of the listed generators. This is not a harmless sign convention: Theorem 5 is the bridge between the ODE (17) and the group action (18), and Theorem 7 uses (18) to derive the hyperbolic gradient flow (19). Without Theorem 5, the algorithm that computes the conformal barycenter is unsupported, and the MLE for the Möbius family inherits the problem because it minimizes H(a) using that same algorithm. The paper also contains other concrete errors, such as the normalizing constant in (42) contradicting the disc case and the CDF in (27) being wrong, but the generator/sign mismatch is the most load-bearing because it breaks the central computational method. I agree with the reader's verdict and see no reason to change it; the paper as written does not substantiate its central claim.","tokens_in":21904,"tokens_out":8809,"duration_ms":78969,"concrete_test":"Recompute v2 by differentiating m2(t) at t=0 and check whether the right-hand side of (17) can be written as c1(-i z) + c2(1-z^2) + c3(i z^2+i) with real c1,c2,c3. Then independently simulate (17) for a small configuration, e.g. N=2, K<0, and compare ζ_j(t) with g_{a(t)}(ζ_j(0)) where a(t) solves (19); if the trajectories diverge, Theorem 5 is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2's proof that Poincaré swarms evolve by Möbius group action is algebraically wrong. Differentiating m2(t) = (t-z)/(1-tz) at t=0 yields 1-z^2, not z^2-1. Moreover, with the paper's h1 = (K/2N) Σ Re ζ_k and h2 = (K/2N) Σ Im ζ_k, the combination h1(z^2-1) + h2(i z^2+i) equals (K/2N)(Σ ζ_k) z^2 - (K/2N)(Σ \\bar ζ_k), which is not the right-hand side of (17); the z^2 and constant terms are interchanged and the signs do not match. Thus (17) is not a linear combination of the claimed generators, Theorem 6 does not apply, and Theorem 5's conclusion that ζ_j(t) evolves by a Möbius transformation is unsupported. Since Theorem 7 and the barycenter/MLE algorithms depend on this swarm dynamics, the computational core of the paper fails as written. A separate but compounding error: the CDF in (27) is miscalculated, so the inversion (28) samples the wrong distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a group-theoretic framework for machine learning on hyperbolic balls. In the Poincaré disc it defines a conformal barycenter as the point whose associated Möbius transformation balances a configuration, introduces 'Poincaré swarms' whose dynamics are claimed to evolve by the Möbius group and to implement hyperbolic gradient descent, and defines a conformally invariant family of 'Möbius' probability distributions with exact sampling and maximum likelihood estimation. These constructions are then extended to higher-dimensional Poincaré and Bergman balls. The stated contribution is to give rigorous answers to four methodological questions: definition and computation of a mean, sampling, and parameter estimation.","tokens_in":22179,"tokens_out":13670,"duration_ms":115809,"significance":"If the mathematical claims were correct, the framework would offer a principled and computationally useful alternative to current hyperbolic ML tools: a group-invariant notion of mean, a gradient-flow algorithm, and a tractable statistical family with exact sampling and closed-form MLE. The paper also usefully connects hyperbolic ML to the Kuramoto oscillator literature and to complex hyperbolic geometry. However, the manuscript as written contains several algebraic errors in the central derivations: the metric tensor in Definition 1 is not the standard Poincaré metric, an infinitesimal generator in Section 3.2 has the wrong sign, the sampling inversion in Eq. (28) is incorrect, and the normalizing constant in Eq. (42) has the wrong power of π. In addition, core existence and uniqueness results are imported from the author's own unpublished preprints rather than proved. The intended contributions are therefore not established in this version.","major_comments":[{"comment":"The metric defined in Definition 1, g_x(u,v)=⟨u,v⟩/(1-|x|^2), is not the standard Poincaré ball metric. The standard metric is 4⟨u,v⟩/(1-|x|^2)^2 (up to a constant factor) and has constant sectional curvature -1; the metric as written is conformal to Euclidean with conformal factor (1-|x|^2)^{-1} and does not have constant negative curvature. This is not a harmless convention issue: the distance formula (5)-(6), the hyperbolic measure (4), and the hyperbolic gradient formula (23) used throughout the paper are all compatible with the square-denominator metric, not with the displayed one. The definition must be corrected and the curvature- and gradient-dependent claims rechecked.","section":"Section 2.2, Definition 1"},{"comment":"The infinitesimal generator v2 is miscomputed. For m2(t)=(t-z)/(1-tz), one obtains m2'(0)=1-z^2, not z^2-1. With the printed v2=z^2-1 and the stated h1,h2, the combination h1(z^2-1)+h2(iz^2+i) equals (K/2N)(Σ_k ζ_k) z^2 - (K/2N)(Σ_k \\bar ζ_k), whereas the right-hand side of Eq. (17) is -(K/2N)(Σ_k \\bar ζ_k) z^2 + (K/2N)Σ_k ζ_k. The coefficients of z^2 and the constant term are interchanged, so Eq. (17) is not a linear combination of the claimed generators. Replacing v2 by 1-z^2 would make the identity correct, but as written the proof of Theorem 5 and hence the swarm-based algorithm are unsupported.","section":"Section 3.2, Eq. (20)"},{"comment":"The cumulative distribution function for the radial coordinate is incorrect. From the integral in Eq. (27), the CDF is ∫0^{2π}∫0^b (s-1)/π (1-r^2)^{s-2} r dr dφ = 1-(1-b^2)^{s-1}, not 1-(1-√b)^{s-1}. The inversion in Eq. (28) should therefore be |z| = sqrt(1-(1-κ)^{1/(s-1)}), not (1-(1-κ)^{1/(s-1)})^2. As printed, the sampling scheme generates points from a different radial distribution, so the answer to question iii) is not valid.","section":"Section 4.2, Eqs. (27)-(28)"},{"comment":"The normalizing constant for the Möbius family on the Poincaré ball has the wrong power of π. The correct constant for the density with respect to dΛ(x) in Eq. (4) is π^{-d/2} Γ(1+s-d/2)/Γ(1+s-d) (up to a metric normalization), not π^{d/2} times that ratio. As a check, at d=2 Eq. (42) gives π(s-1)(1-|z|^2)^s, while the two-dimensional family in Eq. (26) has density (s-1)/π (1-|z|^2)^s; the two disagree by a factor of π^2. This error propagates into the likelihood (45) and the MLE formulas.","section":"Section 6.1, Eq. (42)"},{"comment":"Load-bearing statements are imported from the author's own preprints [47] and [57] without proof: the existence and uniqueness of the Möbius transformation balancing a configuration (Theorem 1), geodesic convexity and uniqueness of the minimum of H (Theorem 4), and the conformal invariance of the family (Proposition 9). The manuscript says 'We will omit the proofs, as they are provided therein.' Since these results constitute the definition and existence of the barycenter and the definition of the statistical family, the paper is not self-contained. For a paper whose stated aim is to give 'detailed and rigorous answers,' these results should either be proved here or the manuscript should state clearly that it builds on and assumes them.","section":"Sections 3.1 and 4.1 (Theorems 1, 4, Proposition 9)"},{"comment":"The gradient computation in the proof of Theorem 13 contains sign errors. Between the second and third displayed lines, the numerator changes from a|y_i-a|^2 - a(1-|a|^2) + y_i(1-|a|^2) to a|y_i-a|^2 + (y_i-a)(1-|a|^2), and the Möbius transformation h_a in Eq. (7) has numerator proportional to (a-x) rather than (y_i-a). The subsequent substitution that replaces h_a(x_i(0)) by x_i in the hyperbolic gradient is also not justified, so the identification of Eq. (36) as the gradient flow for H_d is not established.","section":"Section 5.1, proof of Theorem 13"}],"minor_comments":[{"comment":"The index n is used interchangeably with the complex dimension m in the definitions of b(z)_ij and K(z,w); the notation should be made consistent.","section":"Section 2.3, Definition 2"},{"comment":"The Jacobian formula uses n for the dimension, which is inconsistent with the rest of the section using d; it should be d throughout.","section":"Section 2.2, Eq. (10)"},{"comment":"Theorem 1 asserts uniqueness 'up to a rotation,' while Definition 4 asserts a unique point a; the role of the residual rotation in the definition of the barycenter should be clarified.","section":"Section 3.1, Theorem 1 vs Definition 4"},{"comment":"The transformation ga(z)=(z-a)/(1-\\bar a z) in Eq. (19) differs from the convention in Eq. (1) by a minus sign; this is likely absorbed by the eiθ factor, but the convention should be stated explicitly to avoid confusion.","section":"Section 3.2, Eq. (19)"},{"comment":"The transitivity statement says 'h(a1) = h(a2)'; this should read 'h(a1) = a2' for a transformation h mapping μ1 to μ2.","section":"Section 6.1, after Eq. (42)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central definitions and existence results are substantially drawn from the author's own recent preprints [47] and [57]; the incremental contribution over those works consists of the higher-dimensional extension, the swarm algorithm as stated, and the MLE formulas. The algebraic errors in the core proofs should be corrected before further review. If [47] and [57] are not yet peer-reviewed, the editor may wish to require that the relevant statements be proved in this paper rather than merely cited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou should know this one: it has a coherent high-level idea but several concrete algebraic errors that break the main proofs as written. The errors look fixable, so I'd treat it as a paper that needs a careful revision, not as a hopeless submission.\n\nThe genuinely new parts are the higher-dimensional extensions. The conformal barycenter and the disc-level Möbius distribution come from the author's earlier work [47] and [57], but the d-dimensional families Moeb_d and HolNat_m, their MLE equations (48) and (50), and the swarm ODEs (35) and (39) are, as far as I can tell, not in those papers. That is a meaningful consolidation, and the group-theoretic viewpoint is well chosen.\n\nThe soft spots are specific and real. Definition 1 gives a metric with denominator (1 - |x|^2) instead of the square; that is not the standard Poincaré metric. The proof of Theorem 5 contains a sign error: differentiating (t-z)/(1-tz) at t=0 gives 1-z^2, not z^2-1. With the paper's h1 and h2, the claimed linear combination of generators does not reproduce Eq (17) unless v2 is corrected. So the proof as written is invalid. The radial CDF in Eq (27) is also miscalculated: the integral gives 1 - (1-b^2)^{s-1}, not 1 - (1-√b)^{s-1}; the inversion in Eq (28) is therefore wrong. The normalizing constant for Moeb_d disagrees with the disc case by a factor of π^2 (plug in d=2). And the paper leans heavily on [47] and [57] for existence, uniqueness, and convexity, without reproving those facts. That reliance would be acceptable if the prior work were rock-solid, but combined with these errors it undermines the 'detailed and rigorous' claim.\n\nDespite all that, the framework is plausible and the new formulas are useful once corrected. A reader in hyperbolic machine learning would want to see this after a fix, but not as is.\n\nFor peer review: I would send it to a referee, not desk-reject, because the contribution is substantive and the errors are the kind a good referee can identify clearly. But the recommendation would be major revision, not acceptance. I would not cite the current version.","headline":"A coherent group-theoretic framework with several correctable algebraic errors; the current version is not rigorous enough to accept, but the higher-dimensional extensions are worth a revision.","tokens_in":22679,"tokens_out":11728,"would_cite":false,"duration_ms":92853,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["51M10","53B20","62F10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims hyperbolic ML can get rigorous answers to its four basic questions—mean, computation, sampling, and parameter estimation—in hyperbolic balls.","keywords":["hyperbolic machine learning","conformal barycenter","Möbius distributions","Poincaré ball","Bergman ball","hyperbolic gradient descent","maximum likelihood","Möbius group"],"falsifier":"Compute $\\frac{d}{dt}\\big|_{t=0} \\frac{t-z}{1-tz}$; it equals $1-z^2$, whereas Section 3.2 states the infinitesimal generator $v_2 = z^2 - 1$. Substituting the paper's $h_1,h_2$ into (20) and comparing with (17) then shows directly whether the linear-combination claim used to prove Theorem 5 is correct.","tokens_in":21709,"feed_emoji":"📐","tokens_out":7361,"duration_ms":65621,"temperature":0.7,"pith_summary":"This paper tries to give hyperbolic machine learning the same basic toolkit that Euclidean machine learning takes for granted. It claims that for hyperbolic balls one can define a principled mean (a conformal barycenter), a family of probability distributions (Möbius distributions) that is invariant under isometries, and efficient algorithms for computing both the barycenter and maximum-likelihood estimates. If the framework is right, hyperbolic latent spaces could use these objects the way Euclidean models use the arithmetic mean and the Gaussian, replacing ad-hoc exponential-map and projected-gradient recipes. The payoff would be compact hyperbolic models whose geometry, statistics, and optimization are mutually consistent.","feed_headline":"Hyperbolic ML gains a group-theoretic mean and statistical models","feed_subtitle":"Conformal barycenters and Möbius distributions answer how to average, sample, and fit parameters in hyperbolic balls.","key_machinery":"The load-bearing object is the group of Möbius (disc-preserving linear-fractional) transformations of the hyperbolic ball and its infinitesimal generators. The barycenter is characterized through this group: a configuration is balanced when its center is zero, and the conformal barycenter is the point whose generating Möbius transformation sends the configuration to a balanced one. The computational machinery is the Poincaré swarm, an ODE whose vector field is asserted to be a time-dependent linear combination of the three infinitesimal generators of the Möbius group; a Lie-group theorem then implies the swarm evolves by Möbius transformations, and a hyperbolic-gradient check identifies the evolving parameter with gradient descent on the potential $H$. The same pattern—generators of the isometry group, a swarm ODE, and a gradient-flow identity—is repeated for Poincaré balls using Lorentz boosts and rotations and for Bergman balls using holomorphic automorphisms.","core_discovery":"The central discovery is that the four basic questions facing any metric-space machine-learning setup—how to average points, how to compute that average, how to sample random points, and how to fit parameters—have rigorous answers when the metric space is a hyperbolic ball. For a configuration of points in the Poincaré disc, the paper defines the conformal barycenter as the unique point $a$ such that the Möbius-transformed configuration is balanced, equivalently the unique minimizer of the potential $H(a) = -\\sum_i \\log \\frac{(1-|a|^2)(1-|z_i|^2)}{|1-\\bar{a}z_i|^2}$; this barycenter is conformally invariant. It then introduces Poincaré swarms, dynamical systems whose trajectories evolve by Möbius group actions and which are hyperbolic gradient flows for $H$, so that running the swarm computes the barycenter. The paper also introduces the Möbius family of densities $p(z;a,s) \\propto ((1-|z|^2)(1-|a|^2)/|1-\\bar{a}z|^2)^s$, shows the family is invariant under the isometry group, gives an explicit inversion-sampling scheme, and proves that maximum likelihood separates into a barycenter problem for $a$ and a one-dimensional convex problem for the concentration $s$. These constructions extend to higher-dimensional Poincaré balls and to the non-equivalent Bergman balls in even dimensions.","pith_inferences":["Editorial extension: if the swarm proof holds, the same Lie-group-plus-gradient-flow template should produce barycenter algorithms on other homogeneous spaces, such as spheres or the hyperboloid model, by writing their isometry generators explicitly.","Editorial extension: one concrete test of the framework's practical value is to benchmark the Möbius family against the wrapped normal distribution in a hyperbolic variational autoencoder; the paper does not report such experiments, but they would directly probe whether the sampling and MLE properties translate into better latent-space modeling.","Editorial extension: conformal invariance suggests that equivariant neural network layers could be designed around the Möbius group rather than around tangent-space operations, but the paper only sketches this direction."],"forward_implications":["Barycenters in hyperbolic balls are conformally invariant: applying any isometry to a configuration moves the mean by that same isometry, which is the equivariance property required of a mean in metric learning.","The Möbius family gives hyperbolic ML a parametric probability model with a closed-form sampling recipe and a maximum-likelihood procedure that reduces to computing the barycenter and solving a one-dimensional convex problem.","Poincaré swarms give a gradient-descent algorithm that respects the hyperbolic metric, so barycenter computation does not require projecting Euclidean updates back into the ball.","In even dimensions, the same framework works for both Poincaré and Bergman balls, and the paper argues the choice between them makes no qualitative difference for statistics or computation.","The framework can support more elaborate pipelines, such as expectation-maximization, variational inference, and normalizing flows in hyperbolic latent spaces."],"supporting_citations":[{"why":"Supplies the definition of conformal and holomorphic barycenters as minimizers of the potential functions used throughout the paper.","marker":"[47]"},{"why":"Introduces the conformally natural family of probability densities on the hyperbolic disc that the paper extends to balls.","marker":"[57]"},{"why":"Provides the template result that globally coupled oscillators evolve by Möbius group action, which the swarm proof adapts.","marker":"[50]"},{"why":"Establishes that Kuramoto dynamics induce hyperbolic gradient flows in the disc, the model for Theorem 7.","marker":"[52]"},{"why":"States the Lie-group theorem used to conclude that a linear combination of infinitesimal generators yields evolution by group transformations.","marker":"[48]"},{"why":"Derives infinitesimal generators for Möbius transformations on spheres and balls, used in the higher-dimensional swarm construction.","marker":"[55]"},{"why":"Provides the rigorous background on hyperbolic balls, Bergman balls, and their symmetry groups.","marker":"[39]"}],"fun_headline_variants":["Group-theoretic toolbox solves hyperbolic ML core tasks","Conformal barycenter and Möbius distributions for hyperbolic ML","Hyperbolic ML gets rigorous averaging via Möbius swarms","Group-theoretic mean and Möbius distributions answer hyperbolic ML's core questions","Poincaré balls get a group-theoretic barycenter and Möbius densities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every swarm algorithm for computing barycenters depends on the assertion that the right-hand side of the swarm ODE is exactly a linear combination of the Möbius group's infinitesimal generators with the stated coefficients; if that assertion fails, the proof that the swarm computes the barycenter collapses.","fun_headline_variants_meta":{"raw":{"variants":["Group-theoretic toolbox solves hyperbolic ML core tasks","Conformal barycenter and Möbius distributions for hyperbolic ML","Hyperbolic ML gets rigorous averaging via Möbius swarms","Group-theoretic mean and Möbius distributions answer hyperbolic ML's core questions","Poincaré balls get a group-theoretic barycenter and Möbius densities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3195,"prompt_tokens":989,"completion_tokens":2206,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":2112}},"tokens_in":605,"tokens_out":2206,"duration_ms":15457,"temperature":1.0,"reasoning_tokens":2112,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:50:21.849405+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $\\frac{d}{dt}\\big|_{t=0} \\frac{t-z}{1-tz}$; it equals $1-z^2$, whereas Section 3.2 states the infinitesimal generator $v_2 = z^2 - 1$. Substituting the paper's $h_1,h_2$ into (20) and comparing with (17) then shows directly whether the linear-combination claim used to prove Theorem 5 is correct.","supporting_citations":[{"cited_title":"Harmonic and subharmonic function theory on t he hyperbolic ball","cited_arxiv_id":null,"evidence_quote":"Provides the rigorous background on hyperbolic balls, Bergman balls, and their symmetry groups."},{"cited_title":"& Strogatz, S","cited_arxiv_id":null,"evidence_quote":"Provides the template result that globally coupled oscillators evolve by Möbius group action, which the swarm proof adapts."},{"cited_title":"& Mirollo, R","cited_arxiv_id":null,"evidence_quote":"Establishes that Kuramoto dynamics induce hyperbolic gradient flows in the disc, the model for Theorem 7."},{"cited_title":"Applications of Lie groups to differential eq uations","cited_arxiv_id":null,"evidence_quote":"States the Lie-group theorem used to conclude that a linear combination of infinitesimal generators yields evolution by group transformations."},{"cited_title":"& Strogatz, S","cited_arxiv_id":null,"evidence_quote":"Derives infinitesimal generators for Möbius transformations on spheres and balls, used in the higher-dimensional swarm construction."}],"review_version":1}