{"id":"fb1c9599-6004-4371-afcb-4430df64bf16","arxiv_id":"2501.14641","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Principal persistence measures compared by MMD yield a scalable, GPU-parallel topological regularizer with claimed continuous gradients, improving GAN image generation and semi-supervised classification.","lead":"This paper introduces PPM-Reg, a topological regularizer that compares small subsamples of point clouds using persistent homology and kernel maximum mean discrepancy, with a GPU implementation that scales to large latent spaces. It proves the regularizer has continuous gradients for smooth densities and shows gains on shape matching, image generation, and semi-supervised learning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3 is unproven as stated: Appendix D assumes the pushforward hθ(μ) inherits a C1 density from μ, which fails for non-injective C1 maps such as the generators/discriminators used in the experiments.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: Appendix D's claim that a C1 map pushes forward a C1 density to a C1 density is false. I would sharpen it further: the failure is not limited to non-injective maps; even an injective C1 map with a vanishing Jacobian, such as x↦x^3, breaks the density claim. This matters because Theorem 3 is the paper's formal basis for claiming that PPM-Reg has continuous gradients and hence stable adversarial training. Theorems 1 and 2 appear to be adaptations of known kernel/measure arguments and are not the problem. The empirical sections provide genuine evidence of practical utility, with ablated hyperparameters, multiple datasets, and a public implementation, so the paper should not be rejected outright. However, the central smoothness theorem is currently unproven as stated: its proof rests on an equation that is not valid under the stated hypotheses, and the empirical loss is computed on empirical measures outside those hypotheses. The concern is addressable by adding a nondegeneracy/submersion assumption on the network maps or by proving a weaker almost-everywhere differentiability statement that covers discrete empirical measures. Since the reader already marked the paper CONDITIONAL for essentially this reason, my stress-test does not move the verdict.","tokens_in":27547,"tokens_out":7795,"duration_ms":79654,"concrete_test":"Independently verify the Appendix D assertion by computing the pushforward density for μ=Uniform(−1,1) under H(θ,x)=θ x^3 using the change-of-variables formula. If the resulting density is not C1 in y for θ≠0 (it has a −2/3 power singularity at y=0), and is absent for θ=0, then the proof of Theorem 7 does not apply to C1 maps with vanishing Jacobian, and Theorem 3 is unsupported for the intended non-injective network maps. As a numerical corroboration, run the released PPM-Reg implementation on this 1D family with hθ(x)=θ x^3 and hθ(x)=θ x, and compare dTq/dθ near θ=0: the non-immersion case should exhibit non-smooth or unbounded gradients if the theorem's conclusion also fails, while the diffeomorphism case should remain smooth.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix D proves Theorem 7 by postulating that F(θ) = f(θ,x)dx with f a C1 density, justified by the statement that this is true if H is C1 and μ has a C1 density. That implication is false. Even an injective C1 map can destroy the C1 density: take μ = Uniform(−1,1) and hθ(x)=θ x^3. For θ≠0, hθ is C1 and injective, but the pushforward density is (1/6)|θ|^{-1/3}|y|^{-2/3} on the appropriate interval, which is unbounded and not C1 at y=0; for θ=0, hθ(μ)=δ_0, which has no density at all. Thus the 'which is true' assertion is not a consequence of the stated hypotheses, and Equation (46) is not justified for the C1 maps appearing in Theorem 3. GAN generators and discriminators with ReLU, max-pooling, or dead/collapsed units are typically non-injective and have lower-rank Jacobians on open regions, so the hypotheses of Theorem 7 are not satisfied in the intended application. In addition, the actual training procedure optimizes Tq on discrete empirical measures, not on measures with C1 densities, so Theorem 3 does not govern the implemented regularizer. The exception 'wherever the PPM is not the trivial measure at the origin' does not repair the gap: the missing density is needed to write the integral representation of Equation (46) at all, independent of whether the PPM is trivial.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PPM-Reg, a topological regularizer for latent-space matching and GAN training. It replaces persistence diagrams with principal persistence measures (PPMs) computed from many small subsamples, compares PPMs via maximum mean discrepancy (MMD) using persistence-weighted kernels, and provides a GPU implementation. The theoretical section proves characteristicness of the kernels (Theorem 1), topological equivalence of MMD and Wasserstein on PPMs (Theorem 2), and claims C^1 regularity of the regularizer as a function of generator and discriminator parameters (Theorem 3). Experiments on shape matching, 32x32 and 64x64 image generation, and semi-supervised learning on MNIST variants and SVHN show consistent improvements over a Cramer-loss baseline.","tokens_in":27965,"tokens_out":4354,"duration_ms":35593,"significance":"If the results are correct, the paper makes a useful practical contribution: a scalable, parallelizable topological regularizer with a principled kernel metric on persistent homology summaries, backed by code release and extensive ablations. The computational speedups in Table 1 and the consistent gains in image-generation metrics and SSL accuracy are credible and well documented. However, Theorem 3, which is the paper's advertised smoothness guarantee, is not proven as stated: the proof in Appendix D relies on a false regularity assumption on pushforwards of densities, and the implemented method optimizes empirical measures rather than C^1 densities. The theoretical claim therefore needs substantial revision before the paper can be accepted in its current form.","major_comments":[{"comment":"The derivation of Theorem 7 (and hence Theorem 3) relies on the assertion that F(θ) has a C^1 density 'which is true if H is C^1, and μ has a C^1 density.' This implication is false for non-injective C^1 maps: for μ uniform on (-1,1) and h_θ(x)=θ x^3, the pushforward has density (1/6)|θ|^{-1/3}|y|^{-2/3} on its support, which is unbounded and not C^1 at y=0, and for θ=0 the pushforward is a Dirac mass. Since GAN generators and discriminators (ReLU networks with possible dead or collapsed units) are generally not injective and may have rank-deficient Jacobians on open regions, the hypotheses of Theorem 7 are not satisfied in the intended application. Thus Eq. (46) is not justified, and the differentiation under the integral in Eq. (50) does not establish Theorem 3 as stated.","section":"Appendix D, Eq. (46)"},{"comment":"The differentiability exception in Theorem 3 is incorrectly stated: the Hilbert-space norm in Eq. (13) is non-differentiable whenever Φ(PPM_q(µ)) = Φ(PPM_q(ν)), i.e., whenever the two PPMs have the same kernel mean embedding, not merely when 'the PPM is not the trivial measure at the origin.' The proof in Appendix D itself uses 'continuously differentiable away from the origin' in the Hilbert space, so the theorem's statement should exclude the entire coincidence set of the two embeddings (or a neighborhood formulation). As written, the statement is misleading about the set of points where gradients are guaranteed continuous.","section":"Theorem 3 (Section 5)"},{"comment":"The implemented PPM-Reg computes PPMs from s subsamples of discrete empirical mini-batches (Section 6.1), and in the GAN setting the measures d_θ(µ) and d_θ(g_ω(ν)) are empirical measures with atoms, not measures with C^1 densities. Theorem 3 therefore does not govern the object that is actually optimized in the experiments. The paper should either provide a separate statement for empirical measures (e.g., differentiability with respect to sample point positions for fixed subsamples, or a stochastic or almost-sure statement for fixed subsample indices), or explicitly limit the smoothness claim to the idealized density setting and explain why the empirical behavior is nevertheless consistent with the theory.","section":"Section 6.1 / applicability of Theorem 3"}],"minor_comments":[{"comment":"The text says 'Table 5 for AnimeFace and Table 4 for CelebA', but Table 4 is the AnimeFace ablation and Table 5 is the CelebA ablation; these references are swapped.","section":"Appendix F.1"},{"comment":"The captions of Figures 10 and 11 contain the typo 'classifiacter', which should be 'classifier'.","section":"Appendix H"},{"comment":"There is a typo 'wiht' in the implementation details; it should be 'with'.","section":"Appendix E.1"},{"comment":"The notation for PPM dimension is inconsistent: Section 3 uses PPM_q, while Appendix D uses PPM_k in Eq. (44); unify the subscript notation throughout.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The core theoretical gap in Theorem 3 is load-bearing, but the empirical work and computational design are strong. I recommend major revision rather than rejection because the gap appears fixable: the authors could weaken Theorem 3 to differentiability almost everywhere (e.g., using the fact that the pushforward of an absolutely continuous measure under a C^1 map is absolutely continuous with L^1 density under a nondegeneracy condition, or restrict to injective diffeomorphisms), or provide a separate treatment of the discrete empirical measures actually used in training. The paper's published status at ICLR 2025 does not change my assessment of the arXiv manuscript under review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper deserves a serious look, but the headline smoothness theorem is not proven as stated, and the gap is load-bearing.\n\nWhat is genuinely new and good: combining principal persistence measures (Gómez–Mémoli) with persistence-weighted MMD kernels (Kusano et al.) is a natural and useful idea; PPM-Reg is a real contribution. The GPU implementation of the simplified PH computation (Equation 4) is valuable, and the speedups in Table 1 are impressive. The three experiment families (shape matching, image generation, semi-supervised learning) show consistent and sometimes large gains. Theorems 1 and 2 are adaptations of known results, but the independent proofs are fine. Citation practice is fair; the minor self-citation is not load-bearing.\n\nThe soft spot is Theorem 3. Appendix D assumes that a C^1 map pushes forward a C^1 density to a C^1 density. That is false. A simple counterexample is H(θ, x) = θ x^3 on [-1,1]: for θ ≠ 0 the pushforward density is (1/6)|θ|^{-1/3}|y|^{-2/3}, which is unbounded at 0 and not C^1; for θ = 0 you get a Dirac mass. GAN generators and discriminators are generally non-injective and often lower-rank on open regions, so the hypothesis fails in exactly the intended application. Equation (46) is therefore unjustified, and the proof of Theorem 3 does not go through. Moreover, even if a density existed, the actual implementation optimizes over discrete empirical measures, so a theorem about measures with C^1 densities would not directly govern the implemented method. The claim of stable gradients is supported empirically, but not by the stated theorem.\n\nThe experimental reporting also has a minor flaw: hyperparameters and early stopping are selected on the test metrics (smallest CMMD), which inflates confidence. There is no comparison to other topological regularizers, only to the base Cramer loss. These are minor relative to the theoretical gap.\n\nProportionate verdict: the core idea and the engineering are solid; the smoothness theorem needs repair. A serious referee should engage with this paper and push for either a corrected theorem with additional assumptions (e.g., submersivity or injectivity with compact domain) or a weaker statement that applies to the discrete case actually used in training. As it stands, the theoretical justification for the stable-gradient claim is missing.\n\nMy recommendation: yes, engage with it, but flag Theorem 3 as the key issue. If the proof can be fixed, this is a solid contribution to topological deep learning.","headline":"A useful combination of PPMs and MMD with a strong GPU implementation, but the C^1-gradients theorem has a proof gap that needs patching before the stability claim is taken as established.","tokens_in":28455,"tokens_out":2828,"would_cite":true,"duration_ms":29350,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","62R40","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes PPM-Reg, a topological regularizer that replaces full persistent-homology computations with principal persistence measures from many small subsamples, compared by maximum mean discrepancy, and proves continuous…","keywords":["persistent homology","principal persistence measures","maximum mean discrepancy","topological regularization","generative adversarial networks","semi-supervised learning","kernel methods","GPU parallelization"],"falsifier":"Train the regularizer on a GAN whose generator is a ReLU network with discrete empirical measures and compute PPM-Reg gradients by finite differences across a parameter where the latent pushforward changes rank; a jump or undefined gradient at that point would show the $C^1$ guarantee does not hold in the implemented setting.","tokens_in":27353,"feed_emoji":"📐","tokens_out":6945,"duration_ms":65389,"temperature":0.7,"pith_summary":"Latent-space matching in deep learning rarely accounts for the multi-scale shape of point clouds, and the usual topological tool, persistent homology, is too slow and too jagged for adversarial training. This paper proposes computing the persistence of many tiny subsamples, exactly $2q+2$ points for homology dimension $q$, packaging the results as a probability measure called a principal persistence measure, and comparing such measures by maximum mean discrepancy with a lifetime-weighted kernel. The paper proves that this comparison is a $C^1$ function of generator and discriminator parameters when the underlying measures have smooth densities, so gradients do not jump. It also reports experiments in which the regularizer improves shape matching, GAN image generation quality, and semi-supervised classification with very few labels. If the smoothness proof carries over to practice, topological regularization becomes both scalable and stable enough for adversarial learning.","feed_headline":"Tiny subsamples make topological loss cheap and stable for GANs","feed_subtitle":"Persistent-homology features from small batches, compared by MMD, improve GAN images and few-label classification.","key_machinery":"The load-bearing object is the principal persistence measure (PPM): for a probability measure $\\mu$, $PPM_q(\\mu)$ is the pushforward of the product measure $\\mu^{\\otimes (2q+2)}$ along the map that sends a subsample to its dimension-$q$ persistence pair $(b,\\ell)$ in the pointed half-plane $\\Omega$. Because a subsample of exactly $2q+2$ points has at most one dimension-$q$ feature, persistence can be computed from nearest-neighbor distances via the explicit formula $t_b = \\max_x d(x,x^{(2)})$, $t_d = \\min_x d(x,x^{(1)})$, which parallelizes on a GPU. The regularizer compares PPMs through MMD using the lifetime-weighted kernel $k_\\Omega(z_1,z_2)=\\ell_1\\ell_2 k(z_1,z_2)$, whose characteristicness (Theorem 1) and topological equivalence to Wasserstein (Theorem 2) are established in the paper; Theorem 3 then gives the $C^1$ gradient guarantee.","core_discovery":"The central claim is that topological regularization need not choose between fidelity and cost: principal persistence measures (PPMs), obtained by pushing a measure through the persistent-homology map on $2q+2$-point subsamples, are rich enough to encode multi-scale topological features, and comparing them with an MMD built from the kernel $k_\\Omega((b_1,\\ell_1),(b_2,\\ell_2)) = \\ell_1 \\ell_2 k((b_1,\\ell_1),(b_2,\\ell_2))$ gives a metric with the same topology as Wasserstein on persistence measures. The main theoretical result, Theorem 3, asserts that the resulting regularizer $T_q$ is $C^1$ in the generator and discriminator parameters whenever the relevant measures have $C^1$ densities and the maps are $C^1$, except at the trivial measure at the origin. The paper's experiments show the regularizer steering point clouds to reference shapes, improving embedding-based image quality metrics on AnimeFace and CelebA, and sharply raising semi-supervised accuracy on MNIST-style datasets with 200 or 400 labels.","pith_inferences":["A testable extension is to smooth the PPM computation, for example by injecting noise or using kernel density estimates, to recover gradient regularity when generators are non-injective, since the paper's $C^1$ guarantee assumes a density condition that practical neural networks may violate.","Because the MMD-PPM metric metrizes the same topology as Wasserstein on persistence measures, the regularizer is a drop-in candidate for other distribution-matching problems, such as domain adaptation or representation alignment, not only GANs.","The near-flat runtime in point-cloud size suggests the same subsample-and-pushforward trick could be applied to other geometric signatures, as long as the small-subsample computation is explicit and parallelizable.","The semi-supervised gains hint that topological regularization acts partly as a manifold-structure prior; if true, similar gains should appear in other low-label regimes beyond the tested datasets."],"forward_implications":["Topological regularization becomes practical at GAN scale: with the PPM-MMD pipeline, per-step cost is nearly constant as the point cloud grows and sublinear in the number of subsamples.","The MMD metric on PPMs can replace Wasserstein or Sinkhorn comparisons of persistence diagrams, since it induces the same topology on persistence measures.","Continuous gradients remove the main obstacle to using topological losses in adversarial training, where gradient discontinuities cause unstable dynamics.","In the reported experiments, adding PPM-Reg improves embedding-based image quality metrics on AnimeFace and CelebA at 32x32 and 64x64 resolutions.","Semi-supervised classification with 200 or 400 labels improves substantially on MNIST-style datasets, indicating that topological latent structure carries label-relevant information."],"supporting_citations":[{"why":"Introduces principal persistence measures and the Wasserstein stability bound that the regularizer builds on.","marker":"Gómez & Mémoli, 2024"},{"why":"Defines persistence weighted kernels for persistence diagrams, from which the lifetime-weighted kernel on PPMs is adapted.","marker":"Kusano et al., 2016"},{"why":"Shows that subsampling for topological function optimization improves smoothness and cost, motivating the subsampling strategy.","marker":"Solomon et al., 2021"},{"why":"Treats persistence diagrams as measures and gives linear representations, the perspective used for Theorem 2.","marker":"Divol & Lacombe, 2021"},{"why":"Provides the Cramer distance used as the base loss in all experiments against which PPM-Reg is added.","marker":"Bellemare et al., 2017"},{"why":"Supplies the result that a continuous bounded characteristic kernel metrizes the weak topology, used to show MMD and Wasserstein induce the same topology.","marker":"Sriperumbudur, 2016"},{"why":"Defines the GAN minimax framework that the regularizer is applied to.","marker":"Goodfellow et al., 2014"}],"fun_headline_variants":["Scalable topological regularizer from tiny subsamples","GPU topological regularizer with continuous gradients","Subsampled persistence makes topo-loss cheap and smooth","PPM regularizer: scalable, GPU, stable for GANs","Topological regularizer gone GPU with smooth gradients"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that the regularizer has continuous gradients assumes that the distribution produced by pushing the noise through the generator has a smoothly varying density; neural-network generators and discriminators are usually non-injective and training operates on discrete samples, so this assumption can fail exactly where the method is used.","fun_headline_variants_meta":{"raw":{"variants":["Scalable topological regularizer from tiny subsamples","GPU topological regularizer with continuous gradients","Subsampled persistence makes topo-loss cheap and smooth","PPM regularizer: scalable, GPU, stable for GANs","Topological regularizer gone GPU with smooth gradients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000439,"raw_usage":{"total_tokens":2249,"prompt_tokens":983,"completion_tokens":1266,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":1190}},"tokens_in":599,"tokens_out":1266,"duration_ms":14723,"temperature":1.0,"reasoning_tokens":1190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:56:57.719500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the regularizer on a GAN whose generator is a ReLU network with discrete empirical measures and compute PPM-Reg gradients by finite differences across a parameter where the latent pushforward changes rank; a jump or undefined gradient at that point would show the $C^1$ guarantee does not hold in the implemented setting.","supporting_citations":[{"cited_title":"A Fast and Robust Method for Global Topological Functional Optimization","cited_arxiv_id":null,"evidence_quote":"Shows that subsampling for topological function optimization improves smoothness and cost, motivating the subsampling strategy."},{"cited_title":"On the optimal estimation of probability measures in weak and strong topologies","cited_arxiv_id":null,"evidence_quote":"Supplies the result that a continuous bounded characteristic kernel metrizes the weak topology, used to show MMD and Wasserstein induce the same topology."}],"review_version":1}