{"id":"28ad39f3-5501-4173-9c39-353988c82ac4","arxiv_id":"2505.04168","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Introduces a consistent estimator for curves of probability measures in Wasserstein space, based on a length-penalized principal curve objective, and proves it recovers the ground-truth curve up to time reversal.","lead":"This paper introduces principal curves in metric spaces, including the Wasserstein space of probability distributions, and proves that fitting them to empirical samples recovers a ground-truth curve of measures in the limit. The result could let biologists collect many unlabeled time-course samples in parallel and order them afterward, which is useful for single-cell RNA-sequencing studies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised probability-1 convergence in Theorem 1.1 is not supported by the proofs, which establish only subsequential convergence; the missing M-growth condition and the non-unique, time-reversal-symmetric minimizers make the central consistency claim overreach.","rationale":"The paper's mathematical core appears sound: the existence, stability, discretization, and iterated Glivenko-Cantelli arguments are careful, and the range-recovery and order-up-to-reversal results are supported by the proofs. My stress-test does not find an internal contradiction in the main theorems as rigorously stated in Section 4 and the appendices. The load-bearing concern is instead a gap between the informal Theorem 1.1 advertised in the introduction and the actual subsequential theorems. This matters because the paper's practical selling point is a consistent seriation estimator, and the guarantees that can be quoted from the proofs are weaker than the abstract suggests. The reader's weakest-assumption flag about injectivity and uniform sampling is related but not identical; injectivity is an explicit identifiability assumption, and uniformity is used mainly for time-label estimation rather than for order recovery. My concern focuses on the proved-vs-advertised convergence mode and the missing M-growth condition, which is not flagged by the reader. Since the reader already assigned CONDITIONAL and noted that Theorem 1.1 is informal, I recommend no change to the verdict. A revised version should state the subsequential character explicitly, add the M ≥ C(log N)^q condition where needed, and clarify that convergence is for every subsequential limit of minimizers (or for the range and order-up-to-reversal), not for an arbitrary selected minimizer sequence.","tokens_in":44645,"tokens_out":19755,"duration_ms":224392,"concrete_test":"Formally re-derive Theorem 1.1 from Corollary 4.3 and Theorem 3.1, tracking the subsequence extractions. Then check whether the full probability-1 convergence of minimizers in the sense of Theorem 1.1 follows when M grows sub-logarithmically, e.g., M_N = floor(log log N). Separately, exhibit the non-uniqueness failure in the minimal example X=[0,1], ρ(t)=t, Λ=Unif[0,1], β_n→0: show both γ(t)=t and γ(t)=1−t minimize PPC(Λ;β_n), and that a sequence of minimizers alternating between them fails to converge, although every subsequential limit has the same range. This settles that the literal limit statement requires a selection rule or a range-only interpretation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, as stated in Theorem 1.1, asserts that with probability 1, minimizers of the discrete objective (3) converge, in the double limit, to a curve with the same range as the ground truth, and that the ordering is recovered up to total reversal. The proofs, however, deliver only subsequential convergence. Corollary 4.3 is explicit: convergence is 'up to passage to a subsequence twice' in N,K and then in β. Theorem 2.5 supplies subsequential convergence in K given weak* convergence of the data measures; Proposition 2.3 supplies subsequential convergence in N; Theorem 4.2 supplies subsequential convergence in β. The informal theorem drops all three qualifiers and also omits the condition from Theorem 3.1(2), namely M ≥ C(log N)^q, that would give full almost-sure weak* convergence of the doubly empirical measure rather than only convergence along a subsequence. This is not purely cosmetic. The functional PPC is invariant under time reversal, and Proposition 2.2 shows minimizers can be non-unique. In the simple case X=[0,1], ρ(t)=t, Λ=Unif[0,1], both γ(t)=t and γ(t)=1−t are minimizers for every β>0. A sequence of arbitrarily chosen minimizers alternating between these two curves does not converge as β→0, even though every subsequential limit has the correct range. Thus the literal 'minimizers converge with probability 1' statement is false without specifying a selection rule or explicitly restricting the claim to the range and to the order-up-to-reversal. The paper's own 'precise statement' in Corollary 4.3 is weaker than the introductory Theorem 1.1, so the headline consistency guarantee is not what is actually proved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a theory of length-penalized principal curves in compact metric spaces and specializes it to Wasserstein space over a compact convex domain. The main results are: existence of minimizers for the penalized functional (Proposition 2.1), stability of minimizer sets under weak-* convergence of the data distribution (Proposition 2.3), a discrete-to-continuum convergence theorem for a piecewise-geodesic discretization (Theorem 2.5), an iterated Glivenko-Cantelli theorem for doubly empirical measures over distributions (Theorem 3.1), and a seriation consistency theorem stating that, as the length penalty tends to zero, minimizers recover the range and the time-ordering (up to reversal) of an injective ground-truth curve (Theorem 4.2 and Corollary 4.3). The paper also proposes a coupled Lloyd-type algorithm, a nonlocal kernel variant, and numerical experiments on simulated branching curves in Wasserstein space. The advertised flagship result, Theorem 1.1, claims probability-1 convergence of empirical minimizers to the ground-truth curve in the joint limit over the number of time points, samples, and knots and then in the length penalty.","tokens_in":44908,"tokens_out":4294,"duration_ms":48875,"significance":"If the consistency claims are stated precisely, this is a valuable contribution. It provides a rigorous variational framework for principal curves in metric spaces, and the discrete-to-continuum convergence (Theorem 2.5) appears to be new even in Euclidean settings. The iterated Glivenko-Cantelli theorem for doubly empirical measures is a useful standalone tool. The application to seriation for Wasserstein-space-valued data is well motivated by single-cell trajectory inference and the experiments demonstrate that the method is competitive with spectral seriation and TSP-based seriation. The proofs are detailed and largely self-contained, with an extensive appendix covering compactness, lower semicontinuity of length, RKHS background, and deferred arguments. However, as discussed below, the informal Theorem 1.1 overstates what the proofs establish, and the fixed-endpoint variant used in the experiments is not fully justified.","major_comments":[{"comment":"The advertised probability-1 convergence in Theorem 1.1 is not supported by the proofs. Corollary 4.3 explicitly gives convergence only \"up to passage to a subsequence twice,\" and Theorem 3.1(1) gives weak-* convergence of the doubly empirical measure only in probability, with almost-sure convergence along a subsequence; full almost-sure convergence requires the growth condition M >= C (log N)^q from Theorem 3.1(2), which is absent from Theorem 1.1. Moreover, literal convergence of arbitrarily selected minimizers is false because of time-reversal symmetry and non-uniqueness: for X=[0,1], rho(t)=t and Lambda=Unif[0,1], both gamma(t)=t and gamma(t)=1-t minimize PPC for every beta>0, so a sequence of minimizers alternating between the two has no limit as beta tends to 0. The theorem should be restated as a subsequential consistency result, or should explicitly impose a selection rule and the M-growth condition if an almost-sure statement is intended.","section":"Section 1, Theorem 1.1; Section 4, Corollary 4.3; Section 3, Theorem 3.1"},{"comment":"The experimental estimator is the fixed-endpoint, nonlocal objective PPC_K^w run with Algorithm 3. Appendix D.2 states that consistency for the fixed-endpoint variant \"can be shown ... by an identical argument,\" but no proof is given, and the nonlocal discretization proof in Proposition D.1 does not address fixed endpoints. Since the experiments and the claimed practical seriation performance rely on this specific variant, the paper should either supply the missing consistency proof or explicitly identify this variant as heuristic and outside the proven theory.","section":"Section 4.2 and Appendix D.2, Algorithm 3"}],"minor_comments":[{"comment":"The displayed sum in Theorem 1.1 uses the index T in the upper limit while the surrounding text uses N; the notation should be harmonized.","section":"Section 1, Theorem 1.1 display"},{"comment":"The notation for the doubly empirical measure alternates between \\hat{\\Lambda}_{N,M} and \\hat{\\Lambda}_{M,N}; please standardize to avoid confusion.","section":"Section 3, Theorem 3.1"},{"comment":"There are several typographical errors, including \"principle curve\" in the captions of Figures 3 and 4 and \"V oronoi\" spacing artifacts; these should be corrected in the final version.","section":"Section 4.2 and Appendix E"},{"comment":"The proof of Lemma A.9 is somewhat involved and its role in Theorem 4.2 is central; consider adding a short intuitive explanation before the formal casework.","section":"Appendix A, Lemma A.9"}],"recommendation":"major_revision","confidential_remarks":"The core variational and convergence results are substantial and appear sound, but the paper's headline theorem overclaims almost-sure convergence without the necessary subsequence qualifiers or growth conditions. Since this is a load-bearing statement for the paper's central contribution, the authors should revise it to match the proven subsequential statements and also address the missing fixed-endpoint consistency proof. With those changes, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper genuinely does something new: principal curves in Wasserstein space, with a discrete-to-continuum consistency proof that is new even for Euclidean principal curves. The mathematical core is strong and the iterated Glivenko-Cantelli theorem is a real contribution. The second thing: the advertised Theorem 1.1 overstates what the proofs deliver. The precise statements give convergence only along subsequences—Corollary 4.3 says 'up to passage to a subsequence twice'—and the missing M≥C(logN)^q condition from Theorem 3.1(2) means you don't even get full almost-sure weak* convergence of the doubly empirical measure in general.\n\nThe body of the paper is careful and self-contained: existence (Prop 2.1), stability (Prop 2.3), discrete-to-continuum (Thm 2.5), curve recovery as β→0 (Thm 4.2), and the iterated Glivenko-Cantelli result (Thm 3.1) are all proved in detail, and the supporting lemmas in the appendix look right to me. The authors also acknowledge non-uniqueness (Prop 2.2) and time-reversal symmetry, so the overreach in the introduction is not hidden. The citation pattern is honest; prior algorithms are credited and the genuinely new contribution is clearly separated.\n\nWhere the paper is soft: Theorem 1.1 literally claims probability-1 convergence of minimizers, but with non-unique minimizers and time-reversal symmetry that statement is false without a selection rule. The simple example ρ(t)=t on [0,1] gives two minimizers for every β>0, and alternating between them as β→0 produces a sequence that does not converge even though every subsequential limit has the correct range. That is a real mismatch between the headline and the proofs. The fix is to state the result set-valued, or match Corollary 4.3 explicitly and add the M growth condition if full a.s. convergence is wanted.\n\nSecond gap: the fixed-endpoint variant used in all the experiments is asserted to be consistent 'by an identical argument' in Appendix D.2, but the proof is not written down. In a paper whose selling point is rigorous guarantees, that should be supplied. The experiments themselves are only two synthetic branching datasets, with no code or data released, so the practical payoff is unvalidated. The injectivity and uniform-time assumptions are load-bearing but clearly stated; they are not flaws, since the seriation problem is genuinely unidentifiable without them.\n\nBottom line: this deserves a serious referee. Send it out, and ask for the theorem statement to be corrected, the fixed-endpoint proof to be written out, and ideally code/data for the experiments.","headline":"A genuinely new consistency result for metric-space principal curves, but the headline theorem promises probability-1 convergence that the proofs don't deliver—subsequential convergence is the actual content.","tokens_in":45596,"tokens_out":3662,"would_cite":true,"duration_ms":36456,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","49Q20","62P10","62R20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A length-penalized curve fitted to unlabeled population snapshots provably recovers their true developmental ordering.","keywords":["principal curves","Wasserstein space","optimal transport","seriation","manifold learning","trajectory inference","Glivenko-Cantelli","consistency"],"falsifier":"Take a compact convex domain $V$, an explicit injective Lipschitz curve $\\rho:[0,1]\\to\\mathcal{P}(V)$, draw $N$ i.i.d. uniform times with $M$ samples per time, and solve the discrete objective (3) at increasing $N,M,K$ with $\\beta$ shrinking; if the minimizer's projection pseudotime does not drive the Kendall-tau ordering error to zero, or if the limiting curve's range strictly contains the range of $\\rho$, the theorem is false.","tokens_in":44337,"feed_emoji":"🧬","tokens_out":6700,"duration_ms":63902,"temperature":0.7,"pith_summary":"This paper establishes that principal curves—nonlinear one-dimensional summaries of a data distribution—can be defined and computed in any compact metric space, and that in the Wasserstein space of probability measures they give a statistically consistent way to recover an unlabeled developmental trajectory. The central setting is a curve $t\\mapsto \\rho_t$ of cell-population distributions, of which we see only unordered, noisy empirical snapshots; the paper proves that minimizers of a length-penalized curve-fitting objective converge, as data grow and the penalty shrinks, to a curve with the same range as $\\rho_t$. When $\\rho_t$ is injective, the ordering of the observed snapshots is recovered up to total reversal, which turns the method into a seriation algorithm. This matters because it makes high-density time-course experiments feasible: embryos or other samples can be processed in parallel without recorded collection times, and the times can be inferred afterward.","feed_headline":"Principal curves recover hidden order in Wasserstein space","feed_subtitle":"A discretized principal-curve objective provably sorts unlabeled developmental snapshots up to total reversal.","key_machinery":"The load-bearing object is the penalized principal-curve functional $\\mathrm{PPC}(\\Lambda)(\\gamma)=\\int_X d^2(x,\\Gamma)\\,d\\Lambda(x)+\\beta\\,\\mathrm{Length}(\\gamma)$, minimized over absolutely continuous curves; the length penalty prevents space-filling solutions that would otherwise drive the fit to zero. Its discrete analogue $\\mathrm{PPC}_K(\\Lambda_N)$ replaces the curve by $K$ knots and the length by a sum of pairwise distances, and is minimized by a coupled Lloyd's algorithm alternating TSP reordering, Voronoi assignment, and Wasserstein-barycenter updates. The argument is carried by three linked convergence facts: minimizers of $\\mathrm{PPC}_K(\\Lambda_N)$ converge piecewise-geodesically to minimizers of $\\mathrm{PPC}(\\Lambda)$ as $N,K\\to\\infty$ (Theorem 2.5); the doubly empirical distribution over empirical measures converges to the true distribution over distributions (Theorem 3.1); and, when the data distribution is supported on an injective curve, sending $\\beta\\to 0$ forces minimizers to have exactly the curve's range, with the only remaining freedom a monotone or reverse-monotone reparametrization (Theorem 4.2, via Lemma A.9).","core_discovery":"The paper's main theorem (Theorem 1.1) says the following. Let $V$ be compact and convex, let $\\rho_t$ be an injective Lipschitz curve in the Wasserstein space $\\mathcal{P}(V)$, and form the empirical measure $\\hat\\Lambda$ from $N$ i.i.d. uniform times, with $M$ cells sampled at each time. Then, with probability 1, minimizers of the discretized penalized objective (3) converge, as $N,M,K$ grow and $\\beta\\to 0$, to a curve $\\gamma^*$ with the same range as $\\rho_t$; if $\\rho_t$ is injective, the projection pseudotime recovers the time ordering up to total reversal. The proof runs through a general theory: existence and stability of minimizers of $\\mathrm{PPC}(\\Lambda)(\\gamma)=\\int d^2(x,\\Gamma)\\,d\\Lambda(x)+\\beta\\,\\mathrm{Length}(\\gamma)$ in compact metric spaces, a discrete-to-continuum convergence theorem for the $K$-knot discretization, an iterated Glivenko-Cantelli theorem for doubly (and triply) empirical measures, and a characterization of length-minimizing curves through the range of an injective curve. A finite-read noise version also holds when sequencing depth grows fast enough.","pith_inferences":["A natural next step the paper leaves implicit is non-uniform sampling: if collection times are denser in fast-developing phases, the projection pseudotime would need to be corrected for the sampling density; the consistency theorem assumes uniform times.","The same discrete-to-continuum argument appears extendable to the loop and multiple-curve variants sketched in Section 5, which would give consistent cell-cycle ordering and branching analyses, though these extensions are not proven in the paper.","For fixed sequencing budgets, the theorem suggests a quantitative trade-off: $M$ must grow for each empirical measure to be accurate, while $N$ must grow to sample the curve densely, so an optimal allocation balancing both limits is a testable design question.","In branching data, ordering snapshots in Wasserstein space yields a single comparable order across branches, unlike feature-space pseudotimes, but the consistency theory currently assumes one injective curve and does not cover branch points."],"forward_implications":["Wasserstein principal curves give a consistent seriation method: fit the discrete penalized curve to empirical distributions, project each snapshot to the curve, and read off a pseudotime that matches the true order up to reversal.","Developmental biologists can collect time courses in parallel and infer collection times afterward, because the estimator requires neither known time labels nor a manual per-time-point pipeline.","The discrete-to-continuum theorem applies in any compact geodesic metric space, so the numerical discretization scheme is rigorously justified for Euclidean principal curves as well, where such schemes had previously been justified only heuristically.","With single-cell sequencing noise, consistency is retained provided read depth grows fast enough relative to the number of cells per time point, so shallow but well-spread sequencing can still recover the trajectory curve.","The framework yields a form of trajectory inference without time labels: a corollary is recovering latent one-dimensional structure from unlabeled marginal samples, something earlier optimal-transport trajectory inference did not address."],"supporting_citations":[{"why":"Defines principal curves and the nonlocal smoothing scheme the paper generalizes.","marker":"[42]"},{"why":"Provides the Wasserstein-space geometry: geodesic structure, compactness, and metric-speed facts.","marker":"[3]"},{"why":"Formulates the optimal-transport trajectory-inference problem the paper extends to unlabeled time points.","marker":"[60]"},{"why":"Supplies the spectral seriation baseline used in the experiments.","marker":"[6]"},{"why":"Gives the Wasserstein convergence rates for empirical measures used in the iterated Glivenko-Cantelli theorem.","marker":"[12]"},{"why":"Provides the finite-read sequencing error model and concentration inequalities for the noisy version.","marker":"[54]"},{"why":"Supplies existence and injectivity results for the average-distance problem in Euclidean space that support Proposition 2.1.","marker":"[63]"},{"why":"Introduces the coupled Lloyd-style algorithm and multiple-curve penalization the discretization borrows.","marker":"[56]"}],"fun_headline_variants":["Principal curves recover time order up to reversal in Wasserstein space","Provable seriation via Wasserstein principal curves","Consistent curve recovery from unordered probability measures","Principal curves give order to unlabeled developmental snapshots","Wasserstein principal curves: a provable seriation method"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ground-truth curve is injective and the observed time labels are i.i.d. uniform on $[0,1]$, so the data distribution is exactly the pushforward of uniform time along the curve; without that, order is not identifiable and the proof of Theorem 4.2 does not go through.","fun_headline_variants_meta":{"raw":{"variants":["Principal curves recover time order up to reversal in Wasserstein space","Provable seriation via Wasserstein principal curves","Consistent curve recovery from unordered probability measures","Principal curves give order to unlabeled developmental snapshots","Wasserstein principal curves: a provable seriation method"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000471,"raw_usage":{"total_tokens":2358,"prompt_tokens":974,"completion_tokens":1384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":1307}},"tokens_in":590,"tokens_out":1384,"duration_ms":12359,"temperature":1.0,"reasoning_tokens":1307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:36:41.838516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a compact convex domain $V$, an explicit injective Lipschitz curve $\\rho:[0,1]\\to\\mathcal{P}(V)$, draw $N$ i.i.d. uniform times with $M$ samples per time, and solve the discrete objective (3) at increasing $N,M,K$ with $\\beta$ shrinking; if the minimizer's projection pseudotime does not drive the Kendall-tau ordering error to zero, or if the limiting curve's range strictly contains the range of $\\rho$, the theorem is false.","supporting_citations":[{"cited_title":"Principal Curves","cited_arxiv_id":null,"evidence_quote":"Defines principal curves and the nonlocal smoothing scheme the paper generalizes."},{"cited_title":"Toward a mathematical theory of trajectory inference","cited_arxiv_id":null,"evidence_quote":"Formulates the optimal-transport trajectory-inference problem the paper extends to unlabeled time points."},{"cited_title":"Optimal sequencing depth for single-cell rna-sequencing in wasserstein space","cited_arxiv_id":null,"evidence_quote":"Provides the finite-read sequencing error model and concentration inequalities for the noisy version."},{"cited_title":"Average-distance problem for parameterized curves","cited_arxiv_id":null,"evidence_quote":"Supplies existence and injectivity results for the average-distance problem in Euclidean space that support Proposition 2.1."},{"cited_title":"Multiple penalized principal curves: Analysis and computation","cited_arxiv_id":null,"evidence_quote":"Introduces the coupled Lloyd-style algorithm and multiple-curve penalization the discretization borrows."}],"review_version":1}