{"id":"5a0e59bf-9e8b-4473-93fb-e94fc4278a37","arxiv_id":"2411.10128","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A hyperbolic 1-D expansive deep CNN is claimed to be strongly universally consistent for regression, extending the Euclidean proof of Lin et al. (2022).","lead":"The paper proposes a hyperbolic (Poincaré-disc) version of one-dimensional convolutional networks and claims it is universally consistent for regression. The proof adapts the Euclidean consistency argument, while experiments on synthetic and real data suggest faster error reduction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 11 is false as stated: every network output lies in the Poincaré disc |f|<1/√c, while f_ρ can be any real number; a deterministic target with log_0(y)=2 gives E(f_m)-E(f_ρ)≥1 for all m.","rationale":"The reader rejected the paper because Lemmas 16 and 19 are asserted without proof, and also noted the output/loss mismatch. I agree that the paper should be rejected, but the more load-bearing problem is stronger: Theorem 11 is false as stated, because the final exponential map confines every estimator to the Poincaré disc while universal consistency over distributions with log_c^0(Y) ∈ L^2 requires approximating real-valued regression functions that can lie outside that disc. The counterexample above is elementary and requires no capacity or approximation theory. It shows that even if the missing lemmas were supplied, the claimed theorem would still fail. This does not diminish the reader's other concerns; it reinforces the rejection verdict, but for a reason the reader did not list as the weakest assumption.","tokens_in":18710,"tokens_out":12189,"duration_ms":130599,"concrete_test":"Run the stated theorem on this one distribution: c=1, d=2, s=2, X={0}, and ρ = δ_{(0, tanh(2))}. Verify that f_ρ(0)=2, that every f ∈ H_{L,s} satisfies |f(0)| < 1, and therefore E(f_m) − E(f_ρ) ≥ 1 for all m, contradicting Theorem 11. If the authors reply that the final exp in (V.5) is a typo and the intended output is the real scalar a_L·log_0(h_L(x)), then the theorem must be restated for that modified class and Lemmas 16 and 19 must be proved for it; the current statement and proof do not support that correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim fails even before the capacity and approximation transfers in Lemmas 16 and 19 are examined. By Definition 5, specifically (V.5), every f ∈ H_{L,s} is the final output h_L(x) = exp_0[a_L·log_0(h_L(x))] ∈ D^1_c, so |f(x)| < 1/√c for all x. The regression target is log_c^0(y), which can be any real number. Choose c=1, d=2, s=2, X={0} as a compact subset of D^2_1, and the deterministic distribution ρ with y = exp_0(2) = tanh(2) ∈ D^1_1. Then log_0(y)=2 ∈ L^2, and the hyperbolic regression function is f_ρ ≡ 2. For every m and every parameter choice, |f_m(0)| < 1, so E(f_m) − E(f_ρ) = (f_m(0) − 2)^2 ≥ 1. The truncation π_{M_m} does not help because M_m → 1/√c = 1 and π_M f_m remains bounded by 1. Consequently, the almost-sure limit asserted in Theorem 11 cannot hold for this distribution. Lemma 19 is also directly false for the constant function 2 on the compact set Ω={0}: no function in H_{L,s} has sup-norm distance less than 1 from it. This is an internal inconsistency in the stated result, not merely a missing proof of the Euclidean-to-hyperbolic transfer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hyperbolic deep convolutional network (eHDCNN) on the Poincaré disc, defines hyperbolic expansive convolution via log/exp maps and Möbius operations, and claims strong universal consistency of the truncated empirical risk minimizer (Theorem 11). The proof follows the three-part strategy of Lin et al. [5]: a covering-number bound for the hyperbolic class (Lemma 16), a concentration inequality (Lemmas 17-18), and a universal approximation lemma (Lemma 19). Experiments on synthetic and real regression/classification data are reported. The central theorem is not established: the architecture outputs points in the Poincaré disc, while the regression targets are real-valued and unbounded, so the risk minimizer cannot converge to arbitrary regression functions.","tokens_in":19022,"tokens_out":8824,"duration_ms":80225,"significance":"If Theorem 11 were correct, the paper would provide the first statistical consistency result for a fully hyperbolic convolutional architecture, a meaningful contribution to the theoretical understanding of hyperbolic neural networks. The paper also ships experiments and an anonymous implementation, which is a strength. However, the central claim is false as stated because of the range mismatch between the network output and the regression target, and the two transfer lemmas (capacity and approximation) are unproved, with Lemma 19 actually false. The claimed contribution therefore is not established; the paper would require a substantially different architecture or a restricted target class to make the theorem viable.","major_comments":[{"comment":"Theorem 11 is false as stated because every f in the class H_{L,s} takes values in the Poincaré disc D^1_c. By Eq. (V.5), the final output is h_L(x)=exp_0[a_L·log_0(h_L(x))], so |f(x)|<1/sqrt(c) for all x. The risk is defined against targets log_0(y) that are arbitrary real numbers, and the regression function f_ρ(x)=E[log_0(Y)|x] can exceed 1/sqrt(c). For a concrete counterexample, take c=1, d=2, s=2, X={0} subset of D^2_1, and the deterministic distribution ρ with Y=exp_0(2) almost surely. Then log_0(Y)=2 is in L^2 and f_ρ is identically 2. For every f in H_{L,s}, |f(0)|<1, so E(f)-E(f_ρ)=(f(0)-2)^2≥1. The truncation π_{M_m} does not help because M_m→1/sqrt(c)=1, so π_{M_m} f_m remains bounded by 1. Consequently, the almost-sure limit asserted in Theorem 11 cannot hold for this distribution. This is an internal inconsistency of the stated result, not merely a missing proof detail.","section":"Theorem 11, Definition 5 (V.5)"},{"comment":"The covering-number bound for the hyperbolic class is asserted without proof. The text states that Lemmas 2-4 of [5] are results on general metric spaces and then simply states Lemma 4 in the hyperbolic context. This does not justify the transfer: the class H_{L,s} is defined through Möbius addition, layer-wise exp/log maps, and ReLU in the tangent/log domain, none of which appears in the Euclidean derivation of [5]. The proof of Lemma 18 uses this exact entropy bound to obtain the exponential decay after the covering-number term, so if the hyperbolic entropy differs from the Euclidean one, the proof of Theorem 11 collapses. The bounded range of H_{L,s} makes the standard Euclidean comparison particularly delicate, since truncation interacts with the range in a different way.","section":"Lemma 16"},{"comment":"The universal approximation lemma is false in the stated form. Its proof defines g(y)=f(exp_0(y)), takes a Euclidean DCNN approximator g^{w,b}_L from Theorem 1 of [8], and sets f^{w,b}_L(x)=g^{w,b}_L(log_0(x)). The authors do not prove that this composed function can be realized by the recursive hyperbolic architecture (V.4)-(V.5); in particular, the Möbius operations and layer-wise exp/log mappings are not eliminated by a single conjugacy with log_0. Moreover, the claim for arbitrary f in C(Ω) is contradicted by the range constraint: with c=1 and Ω={0}, the constant function f≡2 cannot be approximated in sup norm by any element of H_{L,s}, since all such elements satisfy |f(0)|<1. Lemma 19 is used in the proof of Theorem 11 through the term S_7, so the proof fails at this point as well.","section":"Lemma 19"}],"minor_comments":[{"comment":"The statement of Lemma 9 omits a square: the proof gives E(f)-E(f_ρ)=∫(f-f_ρ)^2 dρ_X, which is ∥f-f_ρ∥^2_{L^2_{ρ_X}}, not the norm itself as written.","section":"Lemma 9"},{"comment":"Definition 10 quantifies over 'every Borel probability distribution λ' but the rest of the paper uses ρ for the joint distribution; the condition log_0(Y)∈L^2 should be stated for the joint distribution on X×Y to be unambiguous.","section":"Definition 10"},{"comment":"The experiments do not specify how real-valued inputs are mapped into the Poincaré disc before applying the hyperbolic architecture, which makes it difficult to connect the empirical results to the theoretical setup where X is a compact subset of D^d_c.","section":"Section VII"},{"comment":"Remark 20 claims the proof shows faster convergence than the Euclidean equivalent, but the displayed inequalities only bound errors by different expressions involving M_m and tanh^{-1}; no rate comparison between the hyperbolic and Euclidean estimators is actually derived.","section":"Remark 20"}],"recommendation":"reject","confidential_remarks":"For the editor: the range-mismatch counterexample is decisive and makes the main theorem false as stated. The issue is not a matter of interpretation or a minor technical gap: the final layer of the architecture is defined as a point in the Poincaré disc by Eq. (V.5), while the regression target log_0(y) is an unbounded real value. I would recommend rejecting the paper in its current form; a viable version would need to reformulate the output layer (for example, in the tangent space) and then establish the capacity and approximation transfers from scratch."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the headline is that Theorem 11 is false as stated, and the counterexample is not subtle. Every network output in the proposed architecture lies in the Poincaré disc D^1_c (Definition 5, Eq. V.5), so |f(x)| < 1/√c. But the loss compares f(x) to log_c^0(y), which is a real number and is unbounded when y approaches the boundary of the disc. Take c=1, X={0}, and the deterministic distribution with y = tanh(2) ∈ D^1_1. Then log_0(y) = 2, so f_ρ ≡ 2. For every m, |f_m(0)| < 1, and truncation does not help because M_m → 1. Thus E(f_m) − E(f_ρ) ≥ 1 for all m, contradicting the almost-sure convergence asserted in Theorem 11. Lemma 19 is also false: the constant function 2 on {0} cannot be approximated in sup-norm by functions bounded by 1.\n\nWhat the paper does well: it is a serious attempt to extend the Euclidean universal-consistency framework of Lin et al. to a fully hyperbolic convolutional architecture. The hyperbolic convolution definition via log/exp maps is sensible, and the authors are transparent about the restriction to 1-D convolutions and the Poincaré disc. The proof plan follows the standard route: capacity bound, concentration, approximation. But the two key lemmas (16 and 19) are asserted rather than proved, and the transfer from Euclidean classes to the hyperbolic class is not justified. Even if those gaps were repaired, the range mismatch is a separate, fatal problem.\n\nThe experiments are illustrative at best: no error bars, no seeds, no proper comparison, and the claim of faster convergence is not backed by statistical evidence. The ablation study is also informal.\n\nWho this is for: a reader working on hyperbolic neural network theory might want to see the architecture and the proof strategy, but not the theorem. The main result is not salvageable without changing the network output to the tangent space or modifying the loss. I would not send this to peer review in its current form; it needs a fundamental reformulation before it is ready for a referee.","headline":"Theorem 11 is false: the network's bounded Poincaré-disc output cannot approximate unbounded real regression targets, so the main consistency claim collapses.","tokens_in":19556,"tokens_out":4623,"would_cite":false,"duration_ms":40309,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","68T07","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a deep convolutional network built on the Poincaré disc is strongly universally consistent: its truncated empirical risk minimizer converges almost surely to the Bayes-optimal regression function as sample size grows.","keywords":["hyperbolic deep convolutional neural networks","Poincaré disc","expansive convolution","universal consistency","empirical risk minimization","regression","covering number","pseudo-dimension"],"falsifier":"Compute the metric-entropy covering number of the truncated hyperbolic class $\\pi_M \\mathcal{H}_{L,s}$ at positive curvature $c$ and compare it with the bound claimed in Lemma 16: if the bound is exceeded for any allowed $L$, $s$, $d$, and $M$, the proof collapses. Alternatively, simulate the empirical risk minimizer on a regression problem with a known hyperbolic regression function and check whether $E(\\pi_{M_m} f_{D,L_m,s}) - E(f_\\rho)$ fails to converge to zero when the growth conditions of Theorem 11 are satisfied, which would disprove the theorem.","tokens_in":18408,"feed_emoji":"🧠","tokens_out":5014,"duration_ms":49960,"temperature":0.7,"pith_summary":"The paper proposes an expansive hyperbolic deep convolutional network (eHDCNN) on the Poincaré disc and argues that its empirical risk minimizer is strongly universally consistent: as the sample size grows, the expected squared error converges almost surely to the Bayes-optimal regression error for every distribution whose target is square-integrable after logarithmic projection. The proof extends the Euclidean consistency framework for deep convolutional networks by replacing linear convolution with a hyperbolic operation that maps inputs to the tangent space, convolves there, and maps back through the exponential map. If the proof holds, hyperbolic convolutional architectures inherit the statistical guarantees of Euclidean ones while using curved geometry to represent hierarchical data. The authors also report experiments on synthetic and real datasets showing faster error reduction than Euclidean counterparts, with the best performance at low positive curvature.","feed_headline":"Hyperbolic CNNs proven to reach Bayes-optimal regression","feed_subtitle":"Expansive convolution on a curved disc stays statistically consistent as data grows, and beats flat-space nets in tests.","key_machinery":"The argument is carried by three pieces. First, hyperbolic convolution is defined by transporting inputs to the tangent space at $0$ via $\\log_c^0$, applying ordinary expansive convolution there, and returning via $\\exp_c^0$; layer outputs are combined with Möbius addition and ReLU, so the network lives in nested Poincaré discs of increasing dimension. Second, a capacity bound (Lemma 16) controls the metric entropy of the truncated function class $\\pi_M \\mathcal{H}_{L,s}$ in terms of depth $L$, filter length $s$, dimension $d$, and truncation level $M$, reused from the Euclidean setting. Third, a concentration inequality (Lemma 17) together with a hyperbolic universal approximation lemma (Lemma 19), which approximates continuous functions by composing Euclidean approximators with $\\log_c^0$, yields almost-sure convergence of the truncated empirical error to the truncated generalization error. The truncation operator $\\pi_M$ is needed because the projected target $\\log_c^0(y)$ grows like $\\tanh^{-1}(M\\sqrt{c})$ as samples approach the boundary of the disc.","core_discovery":"The central claim is Theorem 11: for filter length $2 \\le s \\le d$ and depth $L_m \\to \\infty$, the truncated empirical risk minimizer $\\pi_{M_m} f_{D,L_m,s}$ of the eHDCNN class $\\mathcal{H}_{L,s}$ satisfies $\\lim_{m\\to\\infty} E(f_m) - E(f_\\rho) = 0$ almost surely, provided the truncation level $M_m \\to 1/\\sqrt{c}$, the sample-size constraint $m^{-\\theta} M_m^2 [1 + 1/(M_m \\sqrt{c} \\tanh^{-1}(M_m \\sqrt{c}))]^2 \\to 0$, and the growth condition in equation (VI.1) hold for $\\theta \\in (0,1/2)$. Here $f_\\rho$ is the hyperbolic regression function, the conditional expectation of $\\log_c^0(Y)$ given $X$, which minimizes the hyperbolic generalization error. In words, the architecture asymptotically recovers the optimal regression function under any Borel distribution with $\\log_c^0(Y)$ square-integrable, and the result reduces to the Euclidean theorem when $c \\to 0$. The paper also introduces a fully hyperbolic 1-D expansive convolution operation on the Poincaré disc as a generalization of Euclidean convolution.","pith_inferences":["Extending beyond the paper: the unproved transfer of the covering-number bound from Euclidean to hyperbolic layers could be tested numerically, and if it fails, a corrected bound with curvature-dependent factors might still preserve consistency under modified growth conditions on depth and truncation.","Extending beyond the paper: because all model hyperbolic spaces are isometric via the Killing-Hopf theorem, the same consistency result should carry over to hyperboloid or Klein-Beltrami models if the log/exp maps are replaced accordingly.","Extending beyond the paper: the faster-convergence claim suggests that hyperbolic convolution could be most advantageous for regression targets with hierarchical or tree-structured dependencies, where curved geometry captures the structure in lower dimensions.","Extending beyond the paper: the open higher-dimensional case may be approachable by factorizing 2-D convolution into sequences of 1-D expansive convolutions, but the paper does not establish this."],"forward_implications":["When the curvature $c$ tends to $0$, Theorem 11 reduces to the Euclidean universal consistency result, so the hyperbolic theorem is a strict generalization of the flat-space statement.","The eHDCNN architecture is claimed to asymptotically match the Bayes-optimal regression function without any prior on the distribution beyond square-integrability of the projected target.","Lemma 18 states that the truncated empirical error converges to the truncated generalization error faster for hyperbolic convolution than for Euclidean convolution, which the paper ties to requiring fewer training iterations.","The result is currently limited to 1-D expansive convolution with one input channel and ReLU activation; the paper states that higher-dimensional and contractive cases remain open.","Experiments across synthetic and four real-world datasets indicate that low positive curvature improves error convergence, while very high curvature shrinks the Poincaré disc and degrades performance."],"supporting_citations":[{"why":"Supplies the Euclidean universal consistency framework, including the metric entropy bound and the empirical-risk decomposition that the hyperbolic proof mirrors.","marker":"[5]"},{"why":"Provides the Euclidean deep convolutional universal approximation theorem used inside the proof of Lemma 19.","marker":"[8]"},{"why":"Provides the concentration inequality (Theorem 11.4) and the covering-packing number relations used in Lemmas 15 and 17.","marker":"[44]"},{"why":"Supplies the gyrovector space axioms and Möbius addition and scalar multiplication that define the hyperbolic network layers.","marker":"[42]"},{"why":"Motivates the hyperbolic neural network construction on the Poincaré disc that the eHDCNN architecture generalizes.","marker":"[12]"}],"fun_headline_variants":["Hyperbolic CNNs match Bayes-optimal on curved space","Curved CNNs prove universal consistency in regression","Expansive hyperbolic layers reach optimal regression","Poincaré nets: Bayes-optimal via expansive convolution","Hyperbolic DCNNs beat flat rivals with Bayes-optimal fit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that the complexity and approximation estimates proved for Euclidean convolutional networks transfer unchanged to the hyperbolic function class, even though the hyperbolic layers use nonlinear Möbius addition and are not simple conjugates of Euclidean layers; if that transfer fails, Theorem 11 is not established.","fun_headline_variants_meta":{"raw":{"variants":["Hyperbolic CNNs match Bayes-optimal on curved space","Curved CNNs prove universal consistency in regression","Expansive hyperbolic layers reach optimal regression","Poincaré nets: Bayes-optimal via expansive convolution","Hyperbolic DCNNs beat flat rivals with Bayes-optimal fit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1286,"prompt_tokens":994,"completion_tokens":292,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":214}},"tokens_in":610,"tokens_out":292,"duration_ms":3673,"temperature":1.0,"reasoning_tokens":214,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:56:48.533771+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the metric-entropy covering number of the truncated hyperbolic class $\\pi_M \\mathcal{H}_{L,s}$ at positive curvature $c$ and compare it with the bound claimed in Lemma 16: if the bound is exceeded for any allowed $L$, $s$, $d$, and $M$, the proof collapses. Alternatively, simulate the empirical risk minimizer on a regression problem with a known hyperbolic regression function and check whether $E(\\pi_{M_m} f_{D,L_m,s}) - E(f_\\rho)$ fails to converge to zero when the growth conditions of Theorem 11 are satisfied, which would disprove the theorem.","supporting_citations":[{"cited_title":"Universal consistency of deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Euclidean universal consistency framework, including the metric entropy bound and the empirical-risk decomposition that the hyperbolic proof mirrors."},{"cited_title":"Universality of deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Euclidean deep convolutional universal approximation theorem used inside the proof of Lemma 19."},{"cited_title":"Gy ¨orfi, M","cited_arxiv_id":null,"evidence_quote":"Provides the concentration inequality (Theorem 11.4) and the covering-packing number relations used in Lemmas 15 and 17."},{"cited_title":"Ungar, A gyrovector space approach to hyperbolic geometry","cited_arxiv_id":null,"evidence_quote":"Supplies the gyrovector space axioms and Möbius addition and scalar multiplication that define the hyperbolic network layers."},{"cited_title":"Hyperbolic neural networks,","cited_arxiv_id":null,"evidence_quote":"Motivates the hyperbolic neural network construction on the Poincaré disc that the eHDCNN architecture generalizes."}],"review_version":1}