{"id":"6678a26c-9744-49b8-b88f-2e1542e47210","arxiv_id":"2501.19247","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"K-means and EM clustering for points in Poincaré hyperbolic balls are defined via conformal barycenters and Möbius distributions, with synthetic experiments in 2D and 3D.","lead":"This paper develops two standard machine learning tools, k-means clustering and expectation-maximization mixture learning, for data placed inside curved \"hyperbolic ball\" spaces that can encode tree-like hierarchies. It tests them on synthetic point sets in two and three dimensions. Generalists may care because hyperbolic geometry is increasingly used for embeddings of words, networks, and biological data, yet it has lacked basic clustering tooling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"K-means updates minimize the conformal-barycenter objective (8), not the stated squared-hyperbolic-distance objective; no proof links them, so the central clustering algorithm is internally inconsistent.","rationale":"I read the paper as claiming a rigorous, self-contained foundation for unsupervised learning in hyperbolic balls. For that claim, each update must be justified against the objective it is supposed to optimize. The most fragile point is not the imported conformal barycenter itself (which is cited to [17,18]) but the paper's own Section 5.1 k-means update: it substitutes the minimizer of H_2 into a squared-hyperbolic-distance objective. This is a concrete, checkable internal inconsistency rather than a disagreement with community consensus, and it affects the paper's headline contribution even if every cited result is correct. The reader's weakest_assumption emphasized the unproven weighted-barycenter routine; that is a serious missing proof, but the k-means objective mismatch is more directly falsifiable and is independently sufficient to reject. The stationarity sign error in Remark 3 corroborates the same failure mode: printed updates do not match the stated objectives. I therefore keep the reader's REJECT verdict unchanged.","tokens_in":12910,"tokens_out":14324,"duration_ms":126760,"concrete_test":"Take the B^2 configuration y1=0.6, y2=-0.5+0.3i, y3=0.1+0.7i. Compute a_C by integrating the complex ODE (12) with K<0 until z1+z2+z3 is near 0 and applying the inverse Moebius map, exactly as in Section 5.1 step iv. Compute a_LS = argmin_a sum_i d_hyp^2(a,y_i) by direct numerical optimization. If d_hyp(a_C,a_LS) > 1e-3, the center update is not the minimizer of J; record the value of J before and after one Lloyd update. Repeating on random configurations settles whether the mismatch is generic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Load-bearing concern: the k-means algorithm does not optimize the objective it defines. In Section 5.1 the objective is J = sum_i sum_j kappa_ij d_hyp^2(xi_i, a_j). Step iv then replaces each cluster center a_j with the conformal barycenter of the assigned points, i.e. the unique minimizer of H_2(a) = - sum_{i in cluster} log((1-|a|^2)(1-|xi_i|^2)/rho(xi_i,a)) from Eq. (8). No theorem in the paper proves that the minimizer of (8) also minimizes sum of squared hyperbolic distances; the two functions are different, and the concrete check below shows their minimizers differ for a generic configuration. Hence one Lloyd iteration can increase J, and the algorithm has no stated convergence guarantee for J. Since the paper's first contribution is 'k-means clustering in hyperbolic balls, based on the novel definition of barycenter,' this mismatch is not a technicality: the algorithm as printed does not implement the stated clustering objective. The same failure pattern appears in the B3 EM update, where the printed stationarity equation for s has the digamma difference with the wrong sign, so for n>=3 and H_tilde>0 no solution exists.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a mathematical framework for unsupervised learning in Poincaré balls. It defines conformal barycenters and weighted barycenters, introduces a family of Möbius distributions, and then presents two algorithms: k-means clustering whose centers are conformal barycenters, and an EM algorithm for mixtures of Möbius distributions. Experiments are reported for two- and three-dimensional balls. The central claim is that these algorithms provide rigorous, well-founded clustering tools for hyperbolic representations.","tokens_in":13127,"tokens_out":13761,"duration_ms":139175,"significance":"If correct, the paper would address a genuine gap: hyperbolic machine learning lacks standard unsupervised tools, and the conformal-invariance properties of the proposed barycenter and distribution family are appealing. The exposition is clear, and the experiments demonstrate that the proposed updates terminate and recover parameters in some simple settings. However, the paper relies on two unreviewed preprints by the same first author for the two central mathematical objects, and, more importantly, the algorithms as printed do not optimize the objectives they state. The k-means update minimizes a log-ratio barycenter objective rather than squared hyperbolic distance, and the n-dimensional EM M-step for concentration parameters has a sign error in the printed stationarity equation. These are load-bearing inconsistencies, not presentation issues.","major_comments":[{"comment":"The k-means objective is stated as J = sum_i sum_j kappa_ij d_hyp^2(xi_i, a_j), but step iv replaces each center by the conformal barycenter, i.e. the minimizer of H_2 or H_tilde in Eq. (8)/(11). No theorem in the paper shows that the minimizer of (8) minimizes the sum of squared hyperbolic distances. In the one-dimensional real slice, stationarity for J gives artanh(a) = N^{-1} sum artanh(y_i), while stationarity for (8) gives N a/(1-a^2) = sum y_i/(1 - a y_i); these are different for generic configurations, so a Lloyd iteration can increase J. Section 6.1 inherits the same mismatch in all dimensions. The paper therefore does not implement the k-means objective it defines.","section":"Section 5.1, Eq. (8), and Section 6.1"},{"comment":"The stationarity equation printed in Remark 3 has the digamma difference reversed. Differentiating (16) gives ψ(1+s-n/2) - ψ(1+s-n) = H_tilde(hat a) / (N sum_i w_i). The printed equation has the opposite order. Since ψ is increasing, for s > n-1 the printed left-hand side is negative while the right-hand side is nonnegative (and positive except in degenerate cases), so no solution exists. Thus the M-step for concentration parameters in the n-dimensional EM algorithm is inconsistent with the objective, and the B1/B2 results are not produced by the stated EM update.","section":"Section 6.2, Remark 3, Eq. (16)"},{"comment":"The weighted-barycenter routine is asserted without proof or citation. It is not shown that solutions of (9) with f = (K/N) sum_j w_j x_j drive sum_j w_j x_j(T) to zero, nor that the inverse Möbius preimage of zero minimizes the weighted objective (11). The unweighted case is attributed to [18], but the weighted case is used in every cluster-center update in both k-means and EM. The text also assumes strictly positive weights, while EM responsibilities can vanish, so the routine's range of applicability to the E-step is not stated.","section":"Section 3.2, steps ii-iv"},{"comment":"The statement that convergence 'can be established in an analogous way' to Euclidean counterparts is not supported. Euclidean Lloyd convergence relies on the centroid update minimizing the cluster objective, which is exactly the property missing here, and EM monotonicity relies on an exact M-step, which the current n-dimensional s-update does not provide. The paper cannot claim an analogous convergence argument without addressing the objective/update mismatch and the stationarity error.","section":"Section 7"}],"minor_comments":[{"comment":"Definition 4 refers to the unique minimum of 'the function (8)', but the weighted objective should be (11).","section":"Section 3.2, Definition 4"},{"comment":"The weights are denoted w_i in the text and omega_j in the algorithm; the notation should be unified.","section":"Section 3.2"},{"comment":"The mixing-probability update writes sum_{i=1}^n gamma_im; the index should be N, not n.","section":"Section 6.2, iii-1"},{"comment":"The central objects in Definition 3 and Eq. (13) are imported from the preprints [17] and [18] without restating or proving the required convexity, uniqueness, normalization, and MLE properties; the paper should either make these results self-contained or clearly mark them as imported assumptions.","section":"Sections 4 and 5.3"}],"recommendation":"reject","confidential_remarks":"The k-means objective/update mismatch and the sign error in the n-dimensional EM M-step are central, not local, and would require substantial reframing or new proofs to repair. The heavy reliance on refs [17] and [18], both by the same first author, also makes the claimed 'rigorous framework' less self-contained than the abstract suggests. I recommend rejection. If the authors revise, they should in particular re-examine whether their k-means should be defined by the conformal-barycenter objective instead of squared hyperbolic distance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper proposes k-means and EM for clustering in Poincaré balls, but the two main algorithms each have a math problem that breaks the theory as printed. Still worth a referee's time because the underlying idea is useful and the flaws look fixable.\n\nWhat's new: the clustering loops around the conformal barycenter of [17] and the Möbius family of [18]. Those objects come from the same authors' prior work; the contribution here is the k-means iteration and the EM for mixtures. The exposition is clear, and the experiments in the Poincaré disc and B3 give a first look at behavior.\n\nWhere it falls down. First, the k-means objective (Section 5.1) is the sum of squared hyperbolic distances, but the update (step iv) uses the conformal barycenter, which minimizes a different function. No theorem shows these minimizers coincide, and they generally don't. So a Lloyd iteration can increase J, and there is no convergence guarantee for the stated objective. That's not a technicality—it's the paper's first contribution. Second, the EM stationarity equation in Remark 3 has the digamma difference with the wrong sign. Differentiating (16) gives ψ(1+s−n/2) − ψ(1+s−n) = H̃/(N∑w_i), not the printed ψ(1+s−n) − ψ(1+s−n/2). With that sign, the left side is negative for n≥3 while the right side is nonnegative, so the reported B3 fits can't be reproduced. There's also a gap in Section 3.2: the weighted barycenter routine asserts that after sufficiently large T the weighted Euclidean sum w1 x1(T)+...+wN xN(T) goes to zero, but that convergence is not proven or cited. Every center update depends on it.\n\nWhat's good: the paper is honest about not discussing convergence, and the experiments include a failure case (A2/A3, where EM collapses or merges) rather than only cherry-picked successes. But there are no baselines, metrics, error bars, or code, so the empirical section is illustrative at best.\n\nWho it's for: someone working on hyperbolic embeddings who wants a first pass at clustering tools and is willing to do the repair work. This shouldn't be published as-is. I'd recommend reject or major revision, but I would send it to review rather than desk reject, because the core ideas are plausible and the main flaws are identifiable and likely fixable.\n\nBottom line: serious referee, yes; accept, no.","headline":"A useful clustering proposal for Poincaré balls, but the k-means update doesn't optimize its stated objective and the EM stationarity equation has a sign error; still worth a referee's time because the flaws are identifiable and likely fixable.","tokens_in":13767,"tokens_out":4530,"would_cite":false,"duration_ms":36199,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","51M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes a mathematical foundation for k-means and EM clustering inside hyperbolic balls, using a Möbius-invariant barycenter and Möbius mixture distributions.","keywords":["k-means","expectation-maximization","Möbius transformation","hyperbolic machine learning","clustering","Poincaré ball","barycenter","mixture model"],"falsifier":"Take three points in the Poincaré disc with strongly unequal weights (for example, weights 100, 1, 1), compute the weighted barycenter by the paper's gradient-flow routine, and compare it with the numerical minimizer of (11) found by direct optimization; if they disagree, the center update in both algorithms can compute the wrong centers.","tokens_in":12588,"feed_emoji":"📐","tokens_out":6195,"duration_ms":56654,"temperature":0.7,"pith_summary":"This paper proposes that clustering in hyperbolic space can be put on the same footing as Euclidean clustering by replacing the arithmetic mean with a conformal barycenter and the Gaussian family with Möbius distributions. It develops two standard unsupervised-learning algorithms, k-means and expectation-maximization, for data living in the Poincaré ball $\\mathbb{B}^n$ of any dimension. If correct, this gives practitioners principled clustering tools for embedded hierarchical data, because the barycenter and the Möbius family respect the geometry's symmetries. The paper also reports experiments on synthetic mixtures in the disc and in the three-dimensional ball, where the algorithms recover parameters in well-separated cases and fail on overlapping ones. The load-bearing assertion is that the weighted gradient-flow routine computes true weighted barycenters; the paper states this for weighted data without proof.","feed_headline":"Hyperbolic k-means and EM now have a mathematical foundation","feed_subtitle":"A conformal barycenter and Möbius distributions bring clustering to curved data spaces","key_machinery":"The central object is the conformal barycenter, defined as the unique minimizer of the geodesically convex energy $\\widetilde H_n(a) = -\\sum_i w_i \\log\\big((1-|a|^2)(1-|y_i|^2)/\\rho(y_i,a)\\big)$, computed by the gradient-flow system (9)-(10) that transports the weighted points to a configuration whose weighted Euclidean sum is zero and then maps the origin back through the inverse Möbius transformation. The second pillar is the Möbius family of distributions with densities $p(x;a,s)\\propto\\big((1-|x|^2)(1-|a|^2)/\\rho(a,x)\\big)^s$, which is conformally invariant and transitively acted upon by the group of Möbius isometries of the Poincaré ball.","core_discovery":"The paper's central claim is that a hyperbolic analogue of k-means and of Gaussian-mixture EM exists and is computable. For k-means the center of each cluster is the conformal barycenter, defined as the unique minimizer of a geodesically convex function; for mixtures the cluster model is the Möbius family of probability densities, and maximum likelihood estimates the location parameter as the weighted conformal barycenter. Both algorithmic steps — moving points by a hyperbolic gradient flow until their weighted sum vanishes and then pulling the origin back by the resulting Möbius transformation — are inherited from this geometry, so the cluster centers transform equivariantly under isometries of the ball. The paper further presents the EM equations, including a closed-form update for the concentration parameter in the disc and a digamma-function equation in higher dimensions, and reports experiments in $\\mathbb{B}^2$ and $\\mathbb{B}^3$.","pith_inferences":["Beyond the paper, the same conformal-barycenter machinery could define other Euclidean algorithms on hyperbolic balls, such as k-medians or principal geodesic analysis, provided the corresponding objectives are geodesically convex; the paper does not address these.","The unproved weighted-convergence step suggests a testable mathematical conjecture: for any positive weights and initial points in the ball, the flow (9) with $f=(K/N)\\sum_j \\omega_j x_j$ drives the weighted sum to zero, and the limiting Möbius transformation pulls the origin to the weighted barycenter.","The reported failures on overlapping mixtures imply that, as with Euclidean EM, initialization and identifiability matter; a natural extension is an initialization strategy based on the unweighted barycenter or on k-means++ in hyperbolic distance, but the paper does not propose one."],"forward_implications":["Data embedded in a Poincaré ball of any dimension can be clustered by k-means with cluster centers updated by the weighted gradient-flow routine instead of the Euclidean average.","Mixtures of Möbius distributions can be learned by EM, with responsibilities computed from the Möbius density and parameters updated by weighted-barycenter and concentration-parameter estimation.","Because both the barycenter and the Möbius family are conformally invariant, the clustering outcome is equivariant under Möbius transformations: applying an isometry of the ball to the data applies the same isometry to the inferred cluster centers.","The reported experiments in the disc and the three-dimensional ball show that the algorithms recover mixture parameters close to ground truth when clusters are well separated; the same experiments show overlapping components are not recovered, so practical use requires diagnostics such as monitoring responsibilities and log-likelihood.","The M-step is numerically tractable: the concentration parameter has a closed-form update in the Poincaré disc and a digamma-equation solution in higher dimensions."],"supporting_citations":[{"why":"Supplies the geodesically convex potential (8) and the definition of the conformal barycenter as its unique minimizer.","marker":"[17]"},{"why":"Supplies the gradient-flow system (9)-(10), the Möbius distributions (13), their sampling procedure, and the unweighted barycenter computation used in both algorithms.","marker":"[18]"},{"why":"Provides the expectation-maximization method that the mixture-learning algorithm adapts to the Möbius family.","marker":"[14]"}],"fun_headline_variants":["Conformal barycenter makes hyperbolic k-means rigorous","Möbius distributions bring EM to curved data spaces","Hyperbolic clustering gets a mathematical foundation","K-means and EM now work in negatively curved balls","Rigorous clustering in hyperbolic space via barycenter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest load-bearing step is the weighted barycenter computation: the paper assumes that running the gradient-flow system (9) with weighted points drives the weighted Euclidean sum to zero and that the Möbius preimage of zero is exactly the unique minimizer of the weighted objective (11), but it gives no proof or citation for the weighted case.","fun_headline_variants_meta":{"raw":{"variants":["Conformal barycenter makes hyperbolic k-means rigorous","Möbius distributions bring EM to curved data spaces","Hyperbolic clustering gets a mathematical foundation","K-means and EM now work in negatively curved balls","Rigorous clustering in hyperbolic space via barycenter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1225,"prompt_tokens":826,"completion_tokens":399,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":324}},"tokens_in":442,"tokens_out":399,"duration_ms":4053,"temperature":1.0,"reasoning_tokens":324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:50:59.562229+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take three points in the Poincaré disc with strongly unequal weights (for example, weights 100, 1, 1), compute the weighted barycenter by the paper's gradient-flow routine, and compare it with the numerical minimizer of (11) found by direct optimization; if they disagree, the center update in both algorithms can compute the wrong centers.","supporting_citations":[{"cited_title":"Conformal and holomorphic barycenters in hyperbolic balls","cited_arxiv_id":"2410.02257","evidence_quote":"Supplies the geodesically convex potential (8) and the definition of the conformal barycenter as its unique minimizer."},{"cited_title":"& Rubin, D.B","cited_arxiv_id":null,"evidence_quote":"Provides the expectation-maximization method that the mixture-learning algorithm adapts to the Möbius family."}],"review_version":1}