{"id":"70ced36a-c323-4086-9212-5a0fa02fec45","arxiv_id":"2506.04003","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Principal Observable Analysis defines stable mean and covariance statistics for metric measure spaces using 1-Lipschitz functions and proves Wasserstein and Kantorovich-Sturm stability bounds for them.","lead":"This paper introduces a way to summarize and reduce data that lives in any metric space by projecting it onto all 1-Lipschitz scalar fields, then applying a PCA-like variance-maximizing analysis called Principal Observable Analysis. It proves that the resulting mean and covariance statistics are stable under small perturbations of the data distribution, which matters for real-world metric data like networks, shapes, and images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stability of the covariance functional does not transfer to the principal observables that POA actually outputs; the practical method's reliability is therefore not established by the paper's main bounds.","rationale":"The reader's weakest assumption identifies the DCCP optimization gap and the lack of principal-observable stability; I agree partially. My stress-test focuses on the structural transfer gap, because it persists even with exact global optimization. The theorem proofs themselves are clean: I checked the decompositions in Lemmas 5.3, 6.8, and 6.9, the choice of centered counterparts in Theorem 5.4, and the correspondence constructions in Lemmas 6.5 and 6.10, and found no algebraic error. The homogeneous case handles the change of centering correctly via Theorem 5.1, and the heterogeneous case uses McShane-Whitney extensions to build relations that are genuinely surjective, with distortion estimates matching the stated constants. So the object whose stability is proven is indeed stable. The gap is that POA's output is not that object: it is a sequence of maximizers. Convex maximization over a convex set has no general sensitivity theory in the Hausdorff topology used here, and the authors explicitly defer PO stability. Consequently, the abstract's promise of a vectorization, dimension reduction, and visualization method is supported only heuristically and by examples without code or error bars. This justifies the CONDITIONAL verdict but does not warrant rejection: the observable mean and covariance theory is novel and correct, and the missing transfer is clearly identified by the authors. The recommended action is to keep CONDITIONAL and require either a structural stability statement for PO sets under small perturbations, or a clear caveat that POA as implemented is heuristic and not covered by the stability theorems.","tokens_in":17994,"tokens_out":16833,"duration_ms":169089,"concrete_test":"Use a certified global optimizer, or vertex enumeration of the feasible centered-Lipschitz polytope, to compute PO1 exactly on small finite metric spaces (n approximately 6-10, with random distance matrices and the counting measure). For a sequence mu_t -> mu_0 with w1(mu_t,mu_0) -> 0 on a symmetric space such as a cycle graph, compute the set PO1(mu_t) and its signed, permutation-invariant Hausdorff distance to PO1(mu_0). If the set distance does not vanish while dH(Sigma_mu_t,Sigma_mu_0) -> 0, then covariance stability does not transfer to the principal observables that POA outputs. Running the DCCP solver on the same instances also provides a direct check of whether its outputs match the certified global maxima.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proofs of Theorems 5.1, 5.4, 6.3, and 6.6 appear correct, and the paper is explicit (Section 1: 'Stability of principal observables is a more delicate question that requires further investigation') that principal observables are not covered. That qualifier is the load-bearing gap. POA's outputs--the principal observables and the embedding ik(x)=(phi1(x),...,phik(x)) of Section 3.2--are argmaxes of a convex function over a compact convex set, not direct functionals of Sigma_mu. A graph-Hausdorff bound on Sigma_mu and Sigma_nu is a statement about values of the bilinear form over all pairs of centered observables; it does not by itself control the maximizers of variance or the orthogonally constrained maximizers in (14). Thus even if every variance maximization were solved exactly and globally, POA embeddings could jump under arbitrarily small Wasserstein perturbations, and the practical method would not inherit the proven stability. The DCCP solver (Section 3.1) adds a separate gap: it is a heuristic for convex maximization with no optimality certificate. Neither gap is a flaw in the stability theorems, but together they mean that the central applied claim of the paper is weaker than the headline theorems suggest.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces observable mean and covariance operators for probability measures on compact metric spaces, based on 1-Lipschitz observables. It proposes principal observable analysis (POA), a variance-maximization analogue of PCA, and proves Lipschitz stability of the mean and covariance with respect to the Wasserstein distance on a fixed metric space and the Kantorovich-Sturm distance for heterogeneous data. Consistency of empirical estimates is given as a corollary. The paper also presents applications of POA to network and image data and proposes principal observables as basis functions for signal representation.","tokens_in":18207,"tokens_out":8673,"duration_ms":78090,"significance":"The stability theorems are a useful contribution, and the proofs are self-contained, using standard tools such as Arzela-Ascoli, McShane-Whitney extensions, and Wasserstein metrization. Explicit constants are provided. The paper is honest about the open stability question for principal observables. However, the theoretical results do not, as they stand, establish the reliability of the POA method's output, since the argmax definition of principal observables is not stable under the proven covariance bounds, and the optimization is performed with a heuristic.","major_comments":[{"comment":"The paper defines principal observables as global maximizers in (13)-(14), but proves no stability for these maximizers. Theorems 5.4 and 6.6 bound the covariance operator in graph-Hausdorff distance; this is a statement about values of the bilinear form over all centered observables and does not control the argmax of the variance functional, nor the constrained argmax in (14). Consequently, the POA embedding i_k(x)=(phi_1(x),...,phi_k(x)) of Section 3.2 may be discontinuous under arbitrarily small Wasserstein perturbations of mu, even if every maximization is solved exactly. The paper acknowledges this in Section 1 ('Stability of principal observables is a more delicate question that requires further investigation'), but the abstract and introduction present POA as the main application of the stable statistics. To make the central applied claim supported, the authors should either prove a stability result for principal observables under additional hypotheses (e.g., a spectral gap that makes the maximizer unique up to sign) or explicitly restrict the paper's stability claims to the mean and covariance operators and state that POA embeddings are not proven stable.","section":"Sections 3.1-3.2, Theorems 5.4 and 6.6"},{"comment":"The computation of principal observables is described as employing the disciplined convex-concave programming (DCCP) algorithm of [27]. Since (13)-(14) are convex maximization problems, DCCP is a heuristic with no optimality certificate; the paper provides no analysis of the optimization gap. Thus the functions produced by the algorithm and used in the experiments of Sections 3 and 4 are not known to satisfy the defining equations (13)-(14). The paper should state this limitation explicitly and, ideally, provide a bound on the suboptimality (e.g., via a duality gap) or report the achieved gap on the examples. Without this, the term 'principal observables' for the computed outputs is not fully justified.","section":"Section 3.1"}],"minor_comments":[{"comment":"In the definition of T_n, the last component should be <f, phi_{n-1}>, not <f, phi_1> repeated.","section":"Appendix A.1"},{"comment":"The integrand is missing the exponent p; it should read |f(x)-g(x)|^p dmu(x).","section":"Proof of Theorem 6.3, Eq. (36)"},{"comment":"The last term inside the max should be r_{p,mu'}(F'_2,G'_2), not r_{p,mu}(F'_2,G'_2).","section":"Proof of Lemma 6.10, Eq. (47)"},{"comment":"'The maximal and minimal 1-Lipschitz extension' should be 'extensions'.","section":"Section 3.3"},{"comment":"The claim that the number of principal observables is infinite whenever |X| is infinite is stated without proof; a justification or reference would be helpful.","section":"Introduction"},{"comment":"The name 'Mémoli' is typeset with a spurious accent; use the standard spelling.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The stability theorems appear correct and are a solid mathematical contribution. The main concern is the mismatch between the title/abstract emphasis on POA and the lack of stability or optimality guarantees for the principal observables themselves. The authors explicitly acknowledge the stability gap, so a major revision that reframes the claims and addresses the optimization gap could make the paper publishable. The paper is suitable for math.ST if the theory is foregrounded; the empirical part is illustrative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth reading if you work on metric measure spaces or topological data analysis. The core objects are genuinely new: the observable mean and observable covariance as functionals on centered 1-Lipschitz functions, and Principal Observable Analysis as a variance-maximization analogue of PCA. The stability theorems are the real contribution. Theorem 5.4 gives a clean Wasserstein bound on the Hausdorff distance between covariance operators, and Theorem 6.6 extends this to heterogeneous domains via the Kantorovich-Sturm distance. The proofs are self-contained and correct, using standard tools like Arzela-Ascoli and McShane-Whitney extensions. The authors also handle the changing domain of centered observables honestly by comparing graphs of the bilinear forms. This is good mathematics.\n\nThe soft spots are real but not fatal. The practical POA pipeline solves a convex maximization problem using DCCP, which is a heuristic with no optimality certificate. So the computed principal observables are not guaranteed to be the true global maximizers. And even if they were, stability of the covariance operator does not automatically transfer to its maximizers. Principal observables are argmaxes of a convex function over a convex set, and a tiny Wasserstein perturbation can in principle make the argmax jump. The paper explicitly says 'stability of principal observables is a more delicate question' — that is not a hidden flaw, but it does mean the advertised dimension-reduction tool does not inherit the proven stability. That gap should be stated prominently, and the empirical sections should either provide code/data or avoid strong practical claims.\n\nThe citation pattern looks fine; the self-citations are for terminology and auxiliary constructs, not for the central theorems. The empirical illustrations lack error bars and code, which is a minor issue at this stage.\n\nMy take: this deserves a serious referee. The stability bounds on the covariance functional are a genuine contribution, and the paper is honest about the open question. A good referee report should ask for a clearer separation between what is proven (stability of the covariance) and what is conjectured (stability of principal observables), and perhaps for an empirical check of whether the DCCP solutions are close to optimal on small examples. Send it out.","headline":"A solid, honest paper that introduces observable mean/covariance and proves stability bounds for the covariance functional, but the practical POA method rests on an unproven optimization gap and on stability of principal observables that the paper itself flags as open.","tokens_in":18747,"tokens_out":1777,"would_cite":true,"duration_ms":19588,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62R20","51F30","55N31"],"pacs":[],"model":"deepseek-v4-flash","headline":"Observable mean and covariance on metric spaces are proven stable under measure changes.","keywords":["metric measure spaces","metric observables","observable covariance","principal observable analysis","Wasserstein stability","Kantorovich-Sturm distance","1-Lipschitz functions","dimension reduction"],"falsifier":"One could settle the stability claims by constructing compact metric spaces and measures with $w_1(\\mu,\\nu)$ arbitrarily small but $d_H(\\Sigma_\\mu, \\Sigma_\\nu) > \\max\\{1,4D_X\\} w_1(\\mu,\\nu)$, or, for the implemented method, run the Section 3.1 routine on a fixed finite metric space from many random starting points and check whether different local optima produce visibly different POA embeddings while the covariance operators stay close.","tokens_in":17777,"feed_emoji":"📐","tokens_out":8049,"duration_ms":72143,"temperature":0.7,"pith_summary":"On any compact metric measure space, this paper proposes to summarize the shape of the data by projecting it through all 1-Lipschitz scalar functions—metric observables—and recording the mean and covariance of the projected measures. The observable mean and observable covariance are shown to be stable: if two distributions on the same space are close in the Wasserstein metric, the covariance operators are close in a functional Hausdorff metric, with $d_H(\\Sigma_\\mu, \\Sigma_\\nu) \\le \\max\\{1, 4D_X\\}\\, w_1(\\mu,\\nu)$, and an analogous bound holds across different metric spaces using the Kantorovich–Sturm distance. Because these summaries move continuously with the data, the principal observable analysis built from them—maximizing variance over centered 1-Lipschitz functions, in the spirit of PCA—becomes a trustworthy route to vectorizing, reducing, visualizing, and analyzing metric or networked data.","feed_headline":"Observable covariance proven stable on metric spaces","feed_subtitle":"Mean and covariance of 1-Lipschitz projections move at most linearly with Wasserstein distance.","key_machinery":"The load-bearing object is the observable covariance $\\Sigma_\\mu$, the bilinear form $(f,g)\\mapsto \\int_X fg\\, d\\mu$ restricted to centered 1-Lipschitz functions. The 1-Lipschitz condition bounds every centered observable by $D_X$ and gives uniform equicontinuity, so the Arzelà–Ascoli theorem makes the admissible set compact and variance maximization well-posed; the same bounds turn coupling integrals into transport distances. To compare operators with different domains of definition, the paper uses a graph-Hausdorff distance on their domains, and for heterogeneous spaces it builds correspondences between observable spaces by McShane–Whitney extension of observables across a metric coupling. Those correspondences convert an optimal metric and probabilistic coupling into the $d_{KS,p}$ stability bound.","core_discovery":"The paper's central claim, stated on its own terms, is that observable mean and observable covariance are stable statistics for metric measure spaces. Theorem 5.4 gives $d_H(\\Sigma_\\mu, \\Sigma_\\nu) \\le \\max\\{1, 4D_X\\} \\, w_1(\\mu,\\nu)$ for probability measures $\\mu,\\nu$ on a compact metric space $(X,d)$, and Theorem 6.6 gives $d_{GH}(\\Sigma_{p,\\mu}, \\Sigma_{p,\\mu'}) \\le \\max\\{2, 2(D_X + D_{X'})\\}\\, d_{KS,p}(\\mu,\\mu')$ for measures on different compact spaces. A corollary is that empirical means and covariance operators converge almost surely to their population versions as sample size grows. On the practical side, the paper defines principal observable analysis, proves the existence of principal observables as variance maximizers, shows that the POA embedding into $\\mathbb{R}^k$ is 1-Lipschitz when $\\mathbb{R}^k$ is given the $\\ell^\\infty$ metric, and provides empirical evidence that the resulting embeddings and signal representations are useful.","pith_inferences":["The paper establishes stability of the covariance operator, not of the principal observable functions themselves; a user of the implemented algorithm is implicitly assuming the convex-concave programming step finds global maximizers and that those maximizers move continuously with the data, which is not proven.","A natural testable extension is to replace the balanced McShane–Whitney extension for out-of-sample points with other 1-Lipschitz extensions and measure how much downstream classification changes, since all extensions agree on the training data but can differ off it.","The $\\ell^\\infty$ geometry of POA embeddings suggests that standard Euclidean-distance classifiers are mismatched with the method; adapting neighbor searches to $\\ell^\\infty$ or to the push-forward measure might preserve the stability guarantees."],"forward_implications":["Empirical observable means and covariance operators converge almost surely to their population versions; the cited Wasserstein convergence estimates give finite-sample rates.","Small perturbations of a distribution, measured in Wasserstein distance, cannot change the observable covariance operator by more than a constant times the perturbation, so second-order summaries do not jump.","Datasets on unrelated metric spaces can be compared through their covariance operators using the Kantorovich–Sturm distance, with a bound depending only on the diameters.","POA embeddings are 1-Lipschitz into $\\mathbb{R}^k$ with the $\\ell^\\infty$ metric, so far-apart input points cannot be mapped artificially close together.","Principal observables form a 1-Lipschitz basis for signal representation, and on weighted graphs the Lipschitz constraints reduce to checking edges, which keeps computations feasible."],"supporting_citations":[{"why":"Defines the Kantorovich–Sturm $L^p$ transportation distance used in the heterogeneous stability theorems.","marker":"[28]"},{"why":"Supplies the McShane extension used to build 1-Lipschitz observables on a metric coupling.","marker":"[14]"},{"why":"Supplies the McShane–Whitney extension construction, alongside [14], used in correspondences.","marker":"[23]"},{"why":"Provides the Arzelà–Ascoli compactness and weak-convergence facts behind existence and consistency.","marker":"[9]"},{"why":"Provides Wasserstein convergence rates for empirical measures invoked in the rates remark.","marker":"[30]"},{"why":"Gives the disciplined convex-concave programming algorithm used to compute principal observables in practice.","marker":"[27]"}],"fun_headline_variants":["Stable covariance for metric observables","Principal observable analysis: PCA on metric spaces","Observable covariance stable under Wasserstein","Metric spaces get stable covariance via observables"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical method assumes that the optimization routine really finds the globally most variable centered 1-Lipschitz functions; the paper does not prove this or quantify the gap, and it explicitly leaves stability of the principal observables themselves for future work.","fun_headline_variants_meta":{"raw":{"variants":["Stable covariance for metric observables","Principal observable analysis: PCA on metric spaces","Observable covariance stable under Wasserstein","Metric spaces get stable covariance via observables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00041,"raw_usage":{"total_tokens":2159,"prompt_tokens":1012,"completion_tokens":1147,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1094}},"tokens_in":628,"tokens_out":1147,"duration_ms":11042,"temperature":1.0,"reasoning_tokens":1094,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:49:02.171963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One could settle the stability claims by constructing compact metric spaces and measures with $w_1(\\mu,\\nu)$ arbitrarily small but $d_H(\\Sigma_\\mu, \\Sigma_\\nu) > \\max\\{1,4D_X\\} w_1(\\mu,\\nu)$, or, for the implemented method, run the Section 3.1 routine on a fixed finite metric space from many random starting points and check whether different local optima produce visibly different POA embeddings while the covariance operators stay close.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Kantorovich–Sturm $L^p$ transportation distance used in the heterogeneous stability theorems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the McShane extension used to build 1-Lipschitz observables on a metric coupling."},{"cited_title":"Petrakis","cited_arxiv_id":null,"evidence_quote":"Supplies the McShane–Whitney extension construction, alongside [14], used in correspondences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Arzelà–Ascoli compactness and weak-convergence facts behind existence and consistency."},{"cited_title":"Weed and F","cited_arxiv_id":null,"evidence_quote":"Provides Wasserstein convergence rates for empirical measures invoked in the rates remark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the disciplined convex-concave programming algorithm used to compute principal observables in practice."}],"review_version":1}