W1. The text is really hard to read. It lacks specificity at some moments and consistency at others, it repeats excerpts throughout the main body (e.g., l.191-192 and 198-199, or the two consecutive sentences on graph structure learning in l.124-128), and the preliminaries are too short and not self-contained, i.e., readers have to resort to the appendix multiple times when reading and reasoning about the following sections. At the very least, Section 3 should give the minimum information needed to understand the Cholesky manifold and gyrovector spaces, and why they are needed here. For instance, An appears in Section 3.2 before being defined in Section 4.1, and the parallel transport used in Lemma 1 is never defined, its expression only appearing without derivation in Eq. (109). Regarding specificity, it is not clear what a "complete graph topology" (l.234-235) is, nor what the "global evolution patterns" of the abstract refer to. Regarding consistency, Section 4.1 introduces G=(V,E,A) with a standard adjacency matrix, whose diagonal is zero, but the entries of L are then constrained to [0,1] with a strictly positive diagonal (l.217-220), so weighted graphs with self-loops should be defined from the start. Some citations are also misplaced. The adjacency matrix of an undirected graph and its symmetry are standard and need neither a citation nor a dedicated sentence (l.199-201), the role of Boumal (2020) in l.244 is unclear, and Riemannian Adam is due to Bécigneul and Ganea (2019), not to Kingma and Ba (2014). There are also typos, e.g., "positived iagonal" (l.218) and an unbalanced parenthesis (l.223). Finally, the proof of Lemma 1 contains wrong intermediate steps, since Eq. (112) adds a diagonal term outside the exponential and Eq. (113) assigns the diagonal D(K)−D(LE) to [K]−[LE], which is strictly lower triangular. Both errors disappear without justification in Eqs. (114) and (115), though the final expression is nevertheless correct.
W2. The Cholesky manifold and the map Φ restrict the graph topology to have self-loops, i.e., every point of An has a strictly positive diagonal, which might be an issue especially for heterophilic datasets such as Amazon and Minesweeper. That weakens the claims that Φ recovers the complete topology of the undirected graph (l.229-230) and that An is the space of graph topologies over n nodes (l.234-235), and with them the method's flexibility and generality when compared to fixed graph-structure ones. Similarly, the constraint that the lower-triangular entries lie in [0,1] is not preserved by the operations of Sections 4 and 5. Geodesics and the exponential map, as well as the Gaussian sampling, move these entries along straight lines without bounds, and the gyro inverse of Lemma 2 has off-diagonal part 2[LE]−[L], which is negative wherever L has an edge that the reference graph lacks. As a consequence, An is closed under these operations, and Theorem 1 holds, only if the constraint is dropped, in which case An contains matrices with negative edge weights. Furthermore, Propositions 1 to 5 follow from the Log-Cholesky geometry of Lin (2019) through a linear change of variables, and the gyro addition of Lemma 1 is the abelian group operation that Lin (2019) already defines on the Cholesky space, recentered at E. Finally, the claim of being the first investigation of graph learning from the perspective of Riemannian geometry (l.93-94 and l.110) overlooks Riemannian GNNs, e.g., Chami et al. (2019) and Bachmann et al. (2020), as well as continuous spaces of graph structures such as the graphon space of Lovász and Szegedy (2006) – though this last one is not geometrically motivated.
W3. It is very difficult to understand how the proposed method is even trained. The main text never describes the procedure used in the experiments, and Appendix G.3 only states that 2-hop subgraphs are extracted, that a "posterior" with learnable mean and variance is placed on the tangent space, that several samples are mapped back through the exponential map during training, and that predictions are averaged over samples at inference. Do we need to compute tangent vectors or even construct the tangent space? How are the parameters μA and ΣA estimated? Appendix G.3 calls this distribution a posterior, but no prior, likelihood, or divergence term is ever written, so the training loss remains unknown. Many other choices are left unspecified, such as the parameterization of μA and ΣA (shared or per subgraph, full or diagonal), the number of samples, which entries of L are learned (observed edges only or all node pairs of the subgraph), how L is initialized from the observed graph, how the [0,1] constraint is enforced, and what GraphEuclidean is exactly. None of the gyro operations of Section 5 appears in this pipeline either, which leaves the need for that extension unexplained. What is the time complexity of the method compared with the baselines? If L is dense, its number of entries grows quadratically with the size of each subgraph, which can be large around high-degree nodes, and the Monte Carlo averaging multiplies the inference cost by the number of samples, so wall-clock time and memory against the baselines should be reported.
W4. The reported improvement is also less significant than the authors claim it to be. Given the reported standard deviations over the five seeds, the margin over the strongest baseline in Table 1 is smaller than the standard deviation of GraphManifold⋆ in every setting except SimGRACE on Cora, and the margin is negative in three settings. In the ablation of Table 2, the full model is indeed better given some pre-training methods and for homophilic datasets (SimGRACE/Cora and LP-GPPT/PubMed), but there is no statistically significant improvement in the other scenarios. The manifold alone never differs significantly from GraphEuclidean and is even worse on Cora under all four pre-training strategies, which does not support the "clear advantages" claimed in the main body. Moreover, GraphEuclidean already outperforms the strongest baseline of Table 1 in several settings, so the gains over the baselines seem to come from the prompting pipeline and the sampling rather than from the Riemannian structure. The two controls that would settle this are missing, namely a Euclidean counterpart of the distribution component and a version without the Monte Carlo averaging at inference, which acts as a test-time ensemble that the baselines do not use.
Dear Reviewers,
Please review the authors' rebuttal and participate in the discussion to address any remaining questions before the period closes on Nov 18 AoE.
Thank you!
AC