{"id":"f8b87a7a-8fcd-40bf-a9ad-dcb9a6230164","arxiv_id":"2412.03825","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A deep hyperbolic graph network with residual connections, multi-origin product manifolds, and Gaussian dropout reduces over-smoothing and improves node classification accuracy on four standard graphs.","lead":"Residual connections, product-manifold embeddings, and Gaussian dropout are added to hyperbolic graph networks to stop deep layers from collapsing into identical node features. The paper reports accuracy gains on four benchmark graphs, though the anti-smoothing proof is a sketch with several unproven steps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of the central over-smoothing claim is absent and the in-text sketch is internally inconsistent, as Eq. (18) concludes a lower bound via a coefficient described as 'negligible'.","rationale":"The reader's weakest_assumption concerned the product-manifold construction in Eq. (25), and that is a genuine issue: [0, X_i] is tangent at the standard origin o = [1,0,...,0] only; for a random origin o_j with non-zero spatial part, ⟨o_j, [0, X_i]⟩_L = o_j^spatial · X_i ≠ 0 in general, so exp_{o_j}([0, X_i]) is undefined unless parallel transport is specified. However, the more load-bearing issue is the central theoretical claim. The over-smoothing proof is the paper's primary contribution, and it is both omitted (deferred to a non-existent supplement) and internally inconsistent at Eq. (18). If the proof cannot be supplied, the paper's scientific claim is not established. This aligns with the reader's CONDITIONAL verdict, which already requests a complete proof; hence no change to the verdict is needed. Agreement is partial because the reader identified a different weakest assumption, though they did note the missing proof in the rationale.","tokens_in":14057,"tokens_out":6495,"duration_ms":65286,"concrete_test":"Independently re-derive Eq. (18) from Eqs. (14)-(17) by expanding log_o(exp_{z1}(α P_{o→z1}(log_o(H(0))))) symbolically, retaining all terms. If the result is not exactly of the form ~θ_l log_o(H(0)) + ~φ_l ˜P H(l-1) + ~ψ_l o with ~θ_l bounded away from 0 as a function of α_l and H(0), the claimed lower bound E(H(l)) ≥ c > 0 does not follow. This can be checked with a computer algebra system on small random graphs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline theoretical claim—that the hyperbolic residual connection keeps Dirichlet energy bounded away from zero—is not proven as written. The section 'Effectiveness of Hyperbolic Residual Connection' says 'For details of formulas and proofs in this section, see the supplementary material,' but no supplementary material is included in the arXiv submission. The in-text sketch is internally inconsistent: after Eq. (17), the text says θ_l is 'not negligible,' but Eq. (18) introduces ~θ_l as 'negligible similarly' and then concludes E(H(l)) = ~θ_l^2 E(H(0)) + ... is bounded away from zero. If ~θ_l is negligible, this term is small and supplies no positive lower bound; no other mechanism bounding the remainder is given. Moreover, Eq. (17) asserts z = θ_l log_o(H(0)) + φ_l ˜P H(l-1) + ψ_l o as an exact linear combination in the ambient space, which does not follow from the nonlinear exponential and logarithmic maps in Eqs. (2)-(3) and (15)-(16). The conclusion therefore rests on an unsupported algebraic step. This is load-bearing because the paper's stated contribution is the theoretical proof; absent that proof, the empirical results alone do not establish the mechanism claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes R-HGCN, a deep hyperbolic graph convolutional network that combines three components: a hyperbolic residual connection with hyperbolic identity mapping, a product manifold of Lorentz models with randomly chosen origin points per component, and HyperDrop, a multiplicative Gaussian-noise regularization method. The central claim is that the hyperbolic residual connection provably mitigates over-smoothing, in the sense that the Dirichlet energy of the node features remains bounded away from zero even at large depth. The empirical section reports node-classification accuracies on PubMed, CiteSeer, Cora, and Airport under different layer counts and product-manifold configurations, together with ablations showing that removing the residual connection leads to severe degradation at 16 layers.","tokens_in":14378,"tokens_out":5402,"duration_ms":57568,"significance":"If the theoretical claim were rigorously established, the paper would make a useful contribution to hyperbolic graph representation learning by giving a principled way to train deep HGCNs without feature collapse. The empirical study is reasonably broad: it covers multiple datasets, layer depths, and product-manifold structures, and the ablation in Table 4 clearly shows that the proposed residual connection prevents the catastrophic performance drop observed without it at 16 layers. HyperDrop also yields small but fairly consistent gains across datasets. The main weakness is that the theoretical section is only a sketch, with an internal inconsistency and a deferral to supplementary material that is not present in the arXiv submission. The product-manifold construction also contains a definitional gap. As a result, the central theoretical contribution is not currently established, and the empirical gains, while positive, are modest and do not by themselves support the paper's main claim without a valid proof.","major_comments":[{"comment":"The paper's central theoretical claim is not proven in the submitted manuscript. After Eq. (18), the text concludes that E(H^{(l)}) = \\tilde\\theta_l^2 E(H^{(0)}) + \\cdots is bounded away from zero, but it explicitly states that \\tilde\\theta_l is \"negligible\"; a negligible coefficient cannot supply a positive lower bound unless the remaining terms are controlled, and no such control is provided. In addition, Eq. (17) is asserted as an exact ambient-space linear combination z = \\theta_l \\log_o(H^{(0)}) + \\varphi_l \\tilde{P} H^{(l-1)} + \\psi_l o, which does not follow from the nonlinear exponential and logarithmic maps in Eqs. (2)–(3) and (15)–(16). The section also says \"For details of formulas and proofs in this section, see the supplementary material,\" but no supplementary material is included. Since the abstract and the introduction list the theoretical proof as a main contribution, this gap is load-bearing for the paper's central claim.","section":"Effectiveness of Hyperbolic Residual Connection, Eqs. (17)–(18)"},{"comment":"The initial embedding on the product manifold is not well-defined as written. Eq. (25) defines H_{j,i}^{(0)} = \\exp_{\\vec{o}}([0, X_i]) for each Lorentz component, where \\vec{o} = [\\vec{o}_1, \\ldots, \\vec{o}_k] and the \\vec{o}_j are randomly prescribed origin points. For the standard origin \\vec{o} = [1, 0, \\ldots, 0], the vector [0, X_i] is tangent because \\langle \\vec{o}, [0, X_i] \\rangle_L = 0, as the authors note. For a random origin \\vec{o}_j, however, this tangency condition fails unless the spatial part of \\vec{o}_j is orthogonal to X_i. The paper does not specify a parallel transport or projection to make the construction valid. This affects the model definition used in all experiments and needs to be corrected or clarified.","section":"Product Manifold, Eq. (25)"},{"comment":"The proof assumes that \"the features in process all have positive entries so that ReLU does not affect the evaluation of Dirichlet energy.\" This assumption is stated without justification. The Lorentz operations in Definitions 1–4 include matrix-vector multiplication by arbitrary learned weight matrices, Lorentz scalar multiplication, parallel transport, and vector addition; there is no evident reason why all entries of the resulting features remain positive. If the assumption fails, the identity in Eq. (14) does not hold after applying the ReLU-based Lorentz activation function σ_L from Eq. (8). The authors should either prove positivity under the model's operations or provide a version of the argument that handles the activation function.","section":"Effectiveness of Hyperbolic Residual Connection, assumption before Eq. (14)"}],"minor_comments":[{"comment":"Tables 2 and 3 are referenced in the text as \"Table ??\", and the ablation section refers to \"Tabel 4\"; there are also blank section references of the form \"Section \". These should be fixed.","section":"Throughout"},{"comment":"The proposed model is called R-HGCN in the text and abstract, but Tables 2–5 use the name P-HGCN[d×m]. This naming inconsistency should be resolved to avoid confusion about what is being evaluated.","section":"Throughout"},{"comment":"The Bayesian interpretation of HyperDrop is not quite accurate as written: Eq. (23) multiplies the entire matrix θ by a single scalar ξ, so the entries m_{r,c} are perfectly correlated. The posterior q_φ(m_{r,c}) = N(θ_{r,c}, σ²θ_{r,c}²) describes independent Gaussians, which is not what the single-ξ reparameterization induces. The interpretation should be corrected or qualified.","section":"Hyperbolic Dropout, Eqs. (22)–(23)"},{"comment":"The notation E(F) = tr(log_o(F)^T \\tilde\\Delta log_o(F)) for a vector field F ∈ R^{(d+1)×c} is imprecise: log_o(F) is defined for points on the Lorentz manifold, but F is a matrix of features. The intended columnwise or node-wise application of the logarithmic map should be spelled out.","section":"Definition 5"},{"comment":"The derivation leading from Eq. (12) to the exponential over-smoothing bound of HGCNs is compressed: the factor d in E(H^{(l)}) ≤ d(1−λ)²E(H^{(l−1)}) is not derived in the main text, and the use of Lemma 1 with \\sqrt{n} appears to be a loose bound. Since this is not the main claim, it could be cleaned up, but it should be made checkable.","section":"Effectiveness of Hyperbolic Residual Connection, Eqs. (12)–(13)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is formatted as an AAAI submission and explicitly defers proofs to supplementary material that is not included in the arXiv version; the editor should request that the supplementary material be provided before any further evaluation. The naming inconsistency between R-HGCN and P-HGCN and the unresolved table/section references suggest that the paper may have been assembled from earlier drafts; this is a manuscript-quality issue rather than an integrity concern, but it warrants checking that the reported experiments correspond to the described method. The central theoretical claim is currently unsupported, and the revision should either supply a rigorous proof or substantially weaken the claimed contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2412.03825. The paper adapts GCNII's residual and identity mappings to hyperbolic graph convolution, adds a product manifold with per-component origins, and proposes a hyperbolic dropout (HyperDrop). That combination is new, and the empirical work is honest and reasonably thorough: ablations show the residual connection matters at depth, and HyperDrop gives consistent small gains on standard benchmarks. The experiments appear internally consistent, and the comparisons are fair (they even note that HGCN uses link-pretraining, so they don't claim to beat it on PUBMED).\n\nThe soft spots are real. The central theoretical claim—that the residual connection keeps Dirichlet energy bounded away from zero—is not proven in the paper. The proof section explicitly says 'For details of formulas and proofs in this section, see the supplementary material,' but no supplementary material is included. The in-text sketch is not just incomplete; it's inconsistent. After Eq. (17) we're told θ_l is 'not negligible,' then Eq. (18) introduces ~θ_l as 'negligible similarly' and concludes E(H(l)) = ~θ_l^2 E(H(0)) + ... is bounded away from zero. If ~θ_l is negligible, that term supplies no positive lower bound. The assertion that Eq. (17) is an exact linear combination in ambient space also doesn't follow from the nonlinear exp/log maps. This is load-bearing: the paper's stated contribution is the proof.\n\nThere's also a technical gap in the product-manifold construction. Eq. (25) uses [0, X_i] as a tangent vector at each component's origin o_j, but for arbitrary random origins that vector isn't tangent unless o_j's spatial part is orthogonal to X_i. No parallel transport or projection is specified. This might be fixable, but as written it's not well-defined.\n\nWhat's good: the paper is clearly written, the ablations support the residual connection's practical value, and HyperDrop is a sensible idea with stable gains. The authors cite the right prior work (GCNII, Cai & Wang, H2HGCN) and don't overclaim when comparing to pretrained baselines.\n\nVerdict: this deserves a serious referee, but not acceptance as-is. The authors need to provide the missing proof, fix the negligible-coefficient contradiction, clarify the origin/tangent-space construction, and release code. If they do that, the paper could be a solid contribution to the hyperbolic-GNN subfield. I wouldn't cite it in its current form.","headline":"A useful empirical adaptation of GCNII to hyperbolic space, but the central over-smoothing proof is absent and the product-manifold construction has a tangent-space gap.","tokens_in":14873,"tokens_out":3908,"would_cite":false,"duration_ms":33949,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding a hyperbolic residual connection prevents over-smoothing in deep hyperbolic graph convolutional networks, and proves the Dirichlet energy stays bounded away from zero.","keywords":["hyperbolic graph convolutional networks","residual connections","over-smoothing","Lorentz model","product manifolds","hyperbolic dropout","Dirichlet energy","node classification"],"falsifier":"Compute the Lorentz inner product $\\langle o_j, [0, X_i]\\rangle_L$ for a randomly chosen origin $o_j$ and a nonzero feature vector $X_i$. It is almost surely nonzero, meaning the vector $[0, X_i]$ is not tangent at $o_j$ and the exponential-map embedding used in Eq. (25) is not defined as stated; this single calculation would settle whether the product-manifold component construction is valid.","tokens_in":13831,"feed_emoji":"🌐","tokens_out":9670,"duration_ms":90666,"temperature":0.7,"pith_summary":"Deep hyperbolic graph convolutional networks suffer from over-smoothing: after many layers, node features converge together and become useless for classification. This paper tries to show that a hyperbolic residual connection, which carries the initial node features into every layer and adds a hyperbolic identity mapping, prevents that collapse. The paper proves that the Dirichlet energy of the node features stays bounded away from zero in the residual network, even when the same network without the residual decays exponentially. It then adds two supporting mechanisms: a product manifold whose Lorentz components have different randomly chosen origin points, and HyperDrop, a dropout that multiplies hyperbolic representations by Gaussian noise. If correct, this would allow hyperbolic graph models to be built much deeper while preserving long-range hierarchical information.","feed_headline":"Residual connections stop hyperbolic graph nets from over-smoothing","feed_subtitle":"Deep hyperbolic nets stay distinguishable when initial node features are re-injected at every layer.","key_machinery":"The load-bearing object is the hyperbolic residual connection, built from four Lorentz-space operations: matrix-vector multiplication, scalar multiplication, vector addition, and activation, each defined through exponential and logarithmic maps and parallel transport. The residual connection combines two paths with scalar weights $\\alpha$ and $\\beta$: an initial-feature injection $H^{(0)}$ and an identity-mixed weight $(1-\\beta)I+\\beta W$, which is the hyperbolic analogue of the Euclidean residual idea that inspired it. The theoretical argument runs through a Dirichlet energy pulled back to the tangent space at the origin, and the product manifold $(L^d)_{o_1} \\times \\cdots \\times (L^d)_{o_k}$ with randomly prescribed origin points is what lets different components register different geometric perspectives. HyperDrop multiplies each Lorentz component by Gaussian noise $\\xi_j \\sim N(1, \\sigma^2)$ via Lorentz scalar multiplication, regularizing without breaking the manifold constraint.","core_discovery":"On its own terms, the paper's central claim is that the residual hyperbolic graph convolution operator $hgc(H) = \\sigma_L(((1-\\beta)I+\\beta W) \\otimes \\bar H)$ with $\\bar H = ((1-\\alpha) \\odot (\\tilde A \\otimes H)) \\oplus (\\alpha \\odot H^{(0)})$ keeps node representations distinguishable at any depth. Using a hyperbolic Dirichlet energy $E(H) = \\operatorname{tr}(\\log_o(H)^T \\tilde \\Delta \\log_o(H))$, the paper shows that a hyperbolic GCN without the initial-input term contracts its energy by at least a factor $(1-\\lambda)^2\\|(1-\\beta_l)I+\\beta_l W_l\\|^2$ each layer and can over-smooth exponentially, whereas the residual version carries a contribution $\\tilde\\theta_l^2 E(H^{(0)})$ that remains non-negligible. The conclusion is stated as: in R-HGCNs with initial input, $E(H^{(l)})$ is bounded away from zero even when the no-residual energy decays to zero. The product manifold with different origins and HyperDrop are presented as complementary ways to improve representation quality and generalization.","pith_inferences":["The same initial-feature injection argument should transfer to spherical or mixed-curvature product manifolds, where the mechanism does not depend on the specific Lorentz parametrization; testing it there would show whether the energy bound is a general residual effect.","The proof assumes features have positive entries so ReLU does not change the Dirichlet energy; applying the analysis to signed or centered features is a natural test of how broadly the bound holds.","Random origins could be replaced by learned or data-dependent origins, which might increase the diversity of perspectives more deliberately than random sampling; the paper does not explore this.","The layer-wise energy bound suggests a principled schedule for $\\alpha$ and $\\beta$: keep the initial-feature term strong early and let the identity mapping dominate later, which could be validated by measuring Dirichlet energy during training."],"forward_implications":["A hyperbolic GCN with the residual connection can be trained with many layers without node features collapsing, so depth becomes a usable resource rather than a source of degradation.","Because the initial node representation is re-injected at every layer, long-range dependencies in hierarchical graphs remain representable even in very deep models.","Different Lorentz origins in the product manifold produce multiple embeddings of the same input, giving the network several perspectives on the graph structure and stabilizing the Dirichlet energy across components.","HyperDrop improves generalization on the tested citation datasets and can be combined with parameter-space DropConnect, since it acts on hyperbolic representations rather than Euclidean weights."],"supporting_citations":[{"why":"This citation supplies the Dirichlet-energy decay analysis that the paper adapts to Lorentz space.","marker":"[Cai and Wang 2020]"},{"why":"This citation provides the Euclidean residual and identity-mapping design that the hyperbolic residual connection generalizes.","marker":"[Chen et al. 2020]"},{"why":"This citation defines the hyperbolic graph convolution framework and the HGCN baseline that R-HGCN extends.","marker":"[Chami et al. 2019]"},{"why":"This citation establishes that graph convolution acts as Laplacian smoothing, which is the over-smoothing mechanism the paper addresses.","marker":"[Li, Han, and Wu 2018]"},{"why":"This citation shows that tangent-space convolution distorts hyperbolic structure, motivating manifold-native operations.","marker":"[Dai et al. 2021]"},{"why":"This citation supplies the dropout-with-noise principle that HyperDrop adapts to Gaussian multiplicative noise.","marker":"[Srivastava et al. 2014]"},{"why":"This citation provides DropConnect, the parameter-space regularization that HyperDrop is compared against and combined with.","marker":"[Wan et al. 2013]"},{"why":"This citation provides the citation-network datasets used in the node classification experiments.","marker":"[Sen et al. 2008]"}],"fun_headline_variants":["Residual links keep hyperbolic GCNs from over-smoothing","Hyperbolic residual nets beat over-smoothing at any depth","R-HGCN: residual connections curb over-smoothing in hyperbolic GCNs","Residual hyperbolic GCNs stay distinct layer after layer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes that the same Euclidean feature vector $[0, X_i]$ can be embedded by the exponential map at every randomly chosen Lorentz origin $o_j$, but that vector is tangent at $o_j$ only when $o_j$'s spatial part is orthogonal to $X_i$; the paper provides no projection or parallel transport for the general case.","fun_headline_variants_meta":{"raw":{"variants":["Residual links keep hyperbolic GCNs from over-smoothing","Hyperbolic residual nets beat over-smoothing at any depth","R-HGCN: residual connections curb over-smoothing in hyperbolic GCNs","Residual hyperbolic GCNs stay distinct layer after layer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000636,"raw_usage":{"total_tokens":2975,"prompt_tokens":1028,"completion_tokens":1947,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":1871}},"tokens_in":644,"tokens_out":1947,"duration_ms":13282,"temperature":1.0,"reasoning_tokens":1871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:03:11.566269+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the Lorentz inner product $\\langle o_j, [0, X_i]\\rangle_L$ for a randomly chosen origin $o_j$ and a nonzero feature vector $X_i$. It is almost surely nonzero, meaning the vector $[0, X_i]$ is not tangent at $o_j$ and the exponential-map embedding used in Eq. (25) is not defined as stated; this single calculation would settle whether the product-manifold component construction is valid.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation provides the Euclidean residual and identity-mapping design that the hyperbolic residual connection generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation defines the hyperbolic graph convolution framework and the HGCN baseline that R-HGCN extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation shows that tangent-space convolution distorts hyperbolic structure, motivating manifold-native operations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation supplies the dropout-with-noise principle that HyperDrop adapts to Gaussian multiplicative noise."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This citation provides DropConnect, the parameter-space regularization that HyperDrop is compared against and combined with."}],"review_version":1}