{"id":"6abbe6d6-57b5-40a8-a862-a5d74aea0c62","arxiv_id":"2608.06267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A provably convergent gradient descent that estimates a data manifold from a point cloud and stays near it during optimization, applied to learned operator correction in inverse problems.","lead":"This paper introduces a gradient-descent method that keeps optimization steps on an unknown data surface estimated only from sample points. It proves convergence when samples and step sizes are chosen well, and demonstrates the method on a photoacoustic imaging task with a learned correction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Remark 2.15's 'r→0' is circular: Lemmas 2.9–2.10 give small β,γ only near M, but Theorem 2.14 needs them uniformly on the whole basin; the resulting inequality for d(ξ_k,M) need not vanish.","rationale":"The reader's weakest assumption concerns the unverified consistency bounds and the non-manifold nature of the PAT data. My read sharpens this: even in the smooth-manifold setting, the consistency bounds derived from local PCA degrade linearly with d(ξ_k,M), so they cannot be uniformly small on the ball B(ξ*,R) required by Theorem 2.14. Thus the convergence claim for the fully learned scheme is not established independently of the PAT application. This is a real soft spot, but it is a repairable gap rather than a contradiction in the conditional theorem. Since the paper's main theorem is still valid under its explicit assumptions, and the numerical section is already presented as a demonstration, the conditional verdict remains appropriate. No change to the reader's verdict is needed.","tokens_in":34609,"tokens_out":13690,"duration_ms":154810,"concrete_test":"Set M={0}⊂R, E(x)=(x−1)^2, n(x)=−x, and choose learned errors matching Lemmas 2.9–2.10: β_k=C(|x_k|+ε), γ_k=C(|eξ_{k+1}|+τ+ε)(|eξ_{k+1}|+τ+ε+α). Write the one-step recursion from Theorem 2.14 with δτ=Mβ_k+λγ_k+λL_nMτ and solve for the fixed point ρ=limsup |x_k| as τ,ε,α→0 over a range of C,M,μ. If ρ has a positive limit for any admissible parameter range, the vanishing-error conclusion in Remark 2.15 fails; if ρ→0 for all C, the gap is repairable by a bootstrapping argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2.14 itself is a sound conditional perturbation statement. The unsupported step is Remark 2.15's claim that combining the theorem with Lemmas 2.9 and 2.10 makes the error in (2.39) vanish. Lemma 2.9 gives β ≲ α+ε+d(ξ_k,M), and Lemma 2.10 gives γ ≲ (d(eξ_{k+1},M)+τM+ε)(d(eξ_{k+1},M)+τM+ε+α). Substituting these into (2.39) for ρ=limsup d(ξ_k,M) yields an inequality of the form ρ ≤ (2/μ)[M(α+ε+ρ)+λC(ρ+τM+ε)(ρ+τM+ε+α)+λL_nMτ]. This does not force ρ→0 as τ,ε,α→0 for generic constants; it has a nonzero root. Taking uniform bounds over B(ξ*,R) instead gives β,γ ~ R, which also does not vanish. Remark 2.15 lists r→0 as a condition, but r=d(ξ_k,M) is exactly the quantity (2.39) is meant to bound. A bootstrapping argument (first track the unperturbed flow near ξ*, then use the smallness of the tube) might close the gap, but it is not supplied. The PAT experiment does not fill this gap: the disc-image set is not a smooth manifold, and (2.32)–(2.34) are never checked for the U-Net corrected normal operator.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a projected gradient descent scheme for minimizing an energy over an unknown manifold represented by a point cloud. Section 2 develops continuum and discrete formulations for known manifolds, then extends to unknown manifolds using local PCA tangent estimates and an estimated normal correction. The main theoretical results are Theorem 2.13, a fixed-time convergence result for the learned discrete scheme to the exact projected flow under consistency estimates, and Theorem 2.14, a perturbative bound on convergence to a neighbourhood of a locally exponentially stable minimizer of size O(Mβ+λγ+λL_nMτ). Section 3 contains a synthetic surface experiment, and Section 4 applies the scheme to learned operator correction in photoacoustic tomography with a U-Net correction of the normal operator.","tokens_in":34948,"tokens_out":8157,"duration_ms":79492,"significance":"If Theorem 2.14 is read as a conditional perturbation statement, it is a useful, clean result: it quantifies how tangent and normal estimation errors propagate into the final iterate, and the PCA consistency analysis in Lemmas 2.7–2.10 provides explicit rates. The paper is also transparent about the local exponential stability condition (2.32) being an assumption. However, the claimed vanishing of the asymptotic error as τ, ε, α tend to zero is not justified by the supplied arguments, and the PAT application does not verify the structural conditions needed to apply the theorem. The significance is therefore conditional, and the manuscript needs revision before its central claim can be accepted.","major_comments":[{"comment":"The statement that the asymptotic error in (2.39) vanishes provided τ→0, ε→0, α→0, r→0 is circular, because r=d(ξ_k,M) is exactly the quantity the bound is supposed to control. Substituting the bounds β ≲ α+ε+r and γ ≲ (r+τM+ε)(r+τM+ε+α) into (2.39) and writing ρ = limsup d(ξ_k,M) yields an inequality of the form ρ ≤ (2/μ)[M(α+ε+ρ)+λC(ρ+τM+ε)(ρ+τM+ε+α)+λL_nMτ], which need not force ρ→0 and typically has a nonzero root. Taking uniform bounds over B(ξ*,R) instead gives β,γ ~ R, which also does not vanish. A bootstrapping argument might close the gap, but it is not supplied. This is a load-bearing gap for the abstract claim that the scheme converges as the time step and sampling errors vanish.","section":"§2.3, Remark 2.15"},{"comment":"The update in (4.3) is described as gradient descent for a corrected model, but no energy is displayed whose gradient equals NΘ(eA*eAx_k)−A*y+α∂R(x_k). The network NΘ is learned by regression and is not guaranteed to be the gradient of any potential, so the update field need not be conservative. Consequently the hypotheses of Theorem 2.14—an energy E with a Lipschitz gradient and a locally exponentially stable gradient flow—are not verified in the application. The manuscript should either impose and justify a conservative-structure assumption on NΘ, or explicitly present the operator-correction experiments as a heuristic demonstration outside the scope of the convergence theorems.","section":"§4.2, Eq. (4.3)"},{"comment":"The assumptions behind the point-cloud theory are not checked for the PAT application. The data set consists of random indicator discs in a 64×64 pixel grid; this set is not a smooth manifold with positive reach, so Theorem 2.6 and Lemmas 2.9–2.10 do not apply as stated. Moreover, no evidence is provided that the U-Net corrected normal operator satisfies the consistency bounds (2.33)–(2.34) or that the exact projected flow satisfies the local exponential stability condition (2.32). The numerical comparisons in Figures 7–9 and Table 1 are informative, but they do not fill this gap.","section":"§4.4–4.5"}],"minor_comments":[{"comment":"The statement contains a typo: \"d(eξ0,M) =≤ r0/(1+τλ)\" should presumably read \"d(eξ0,M) ≤ r0/(1+τλ)\".","section":"§2.2, Proposition 2.4"},{"comment":"The notation ∂R(x_k) is used in (4.3), and then R is omitted when optimizing over the manifold; the relation between the regularizer and the normal projection could be stated more explicitly.","section":"§4.2"},{"comment":"The phrase \"The first possible construction for a minimising sequence ξ(t) would be to consider the gradient flow\" is grammatically awkward and should be reworded.","section":"§2.1"},{"comment":"Reference [13] contains a typo (\"opological\" for \"Topological\"); several other references have formatting errors, for example \"V ol. 2. 3\" in [11].","section":"References"},{"comment":"The distinction between \"Corrected (early stopping)\" and \"Corrected (converged)\" deserves a sentence explaining how early stopping is determined, since Figure 7 shows the corrected curve diverging near the minimizer.","section":"§4.5.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has useful conditional theory but overclaims in Remark 2.15 and in the application section. I would be willing to review a revision that either supplies the missing bootstrap argument or scales the claims back to match the proven statements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: this is a useful paper for anyone doing optimization over point-cloud manifolds, and the convergence framework deserves engagement. The catch is that the advertised convergence story in the learned case has a circular step in Remark 2.15, and the PAT section runs an update that is not covered by the theory.\n\nWhat is new: the fully discrete scheme (2.4)–(2.7) combines local PCA tangent estimates and a learned normal correction into one provably stable iteration. Theorem 2.14 is a clean conditional perturbation result: if the tangent and normal estimators satisfy (2.33)–(2.34), the iterates converge to a neighbourhood of ξ* whose size is (2/μ)(Mβ+λγ+λL_nMτ). That is a real contribution. Lemma 2.10's quadratic normal-error estimate is also useful. The normal-operator correction idea in Section 4—learning a correction for A*A rather than correcting A and its adjoint separately—is sensible and under-explored, and the paper explains clearly why range mismatch matters.\n\nSoft spots. First, Remark 2.15 overclaims. Lemma 2.9 bounds β by α+ε+d(ξ_k,M), and Lemma 2.10 bounds γ by something quadratic in r+τM+ε+α. Feeding these into (2.39) gives a limsup inequality whose right-hand side contains a term proportional to r. Taking r→0 is exactly what you are trying to prove, so the error need not vanish for generic constants. A bootstrapping argument might close this, but it is not supplied. This does not invalidate Theorem 2.14 as a conditional statement, but it does invalidate the asymptotic claim attached to it.\n\nSecond, equation (4.3) is presented as gradient descent for the corrected normal operator, but no energy is displayed whose gradient is NΘ(...)−b. The Jacobian of NΘ is missing, so the iteration is not the gradient of the shown objective. The authors should either derive the exact gradient and state the needed Lipschitz bounds, or explicitly label Section 4 a heuristic demonstration. The numerical section also tunes hyperparameters on the test image and gives no error bars or code; for a theory paper that is tolerable, but only if the heuristic status is acknowledged. The disc-image set is not a smooth manifold, so (2.32)–(2.34) are not checked there.\n\nVerdict: this deserves a serious referee. A revision that fixes Remark 2.15 and repositions Section 4 would be publishable; even as is, it is one of the few papers that attacks the fully discrete learned-manifold setting head-on.","headline":"A genuinely useful convergence framework for gradient descent over point-cloud manifolds, with a circular asymptotic claim in Remark 2.15 and an application section that runs ahead of the theory.","tokens_in":35464,"tokens_out":2198,"would_cite":true,"duration_ms":26067,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K10","65M32","68T07","49M15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Point-cloud gradient descent converges to manifold minimisers","keywords":["gradient descent on manifolds","point clouds","tangent space estimation","local PCA","learned operator correction","inverse problems","photoacoustic tomography","convergence analysis"],"falsifier":"On a known manifold with a dense point cloud and known minimiser $x^*$, run the algorithm while sending $\\tau$, the PCA radius $\\varepsilon$, and the tangent estimation error $\\alpha$ to zero and measure $\\limsup_k\\|\\xi_k-x^*\\|$; if this limit does not approach zero, either condition (2.32) or the consistency bounds (2.33)-(2.34) fail for that instance. Alternatively, compute the empirical $\\beta$ and $\\gamma$ along the PAT trajectory and check whether they shrink with denser sampling and smaller steps.","tokens_in":34347,"feed_emoji":"🎯","tokens_out":8041,"duration_ms":89055,"temperature":0.7,"pith_summary":"Gradient descent usually needs a known ambient space; this paper treats the case where the objective is defined on a manifold that is known only through a scattered point cloud. The proposed scheme takes a tangential step along a locally estimated tangent space and then a normal step that pulls the iterate back toward the manifold. The central result is that if the ideal projected flow is locally exponentially stable at a minimiser, and if the learned tangent and normal estimates satisfy consistency bounds, the iterates converge to a neighbourhood of that minimiser whose size vanishes with the step size and the estimation errors. The motivating application is learned operator correction in inverse problems, where gradients computed from a corrected model are trustworthy only near the training data; staying near the point cloud stabilises reconstruction, as demonstrated for photoacoustic tomography.","feed_headline":"Point-cloud gradient descent converges to manifold minimisers","feed_subtitle":"A two-step tangent-and-normal update keeps optimisation on the data manifold when gradients fail away from it.","key_machinery":"The load-bearing object is the two-step projected update: a tangential step using the averaged local-PCA projection $\\hat\\Pi_k=\\sum_i w_i^k\\Pi_{x_i,n,\\varepsilon}$ onto tangent planes estimated at nearby samples, followed by a normal step using the estimated normal vector $\\hat n(\\tilde\\xi_{k+1};\\xi_k)$, which moves the iterate back toward the point-cloud manifold. Its role is to keep all iterates inside a tubular neighbourhood where the learned gradient field is reliable. Convergence is carried by the local exponential stability inequality $\\langle \\xi-\\xi^*,F(\\xi)\\rangle\\le-\\mu\\|\\xi-\\xi^*\\|^2$ for the exact flow, together with the consistency Lemmas 2.9 and 2.10 that bound the tangent and normal estimation errors $\\beta$ and $\\gamma$ in terms of the PCA bandwidth, the point-cloud density, and the distance to the manifold.","core_discovery":"The paper's central claim is Theorem 2.14: for a local minimiser $x^*\\in M$ that is a locally exponentially stable equilibrium of the exact projected flow $\\dot\\xi = -\\Pi_{T_\\xi M}\\nabla E(\\xi)+\\lambda n(\\xi)$, the fully discrete scheme obtained from learned tangent projections and learned normal vectors satisfies $\\limsup_{k\\to\\infty}\\|\\xi_k-x^*\\|\\le (2/\\mu)(M\\beta+\\lambda\\gamma+\\lambda L_n M\\tau)$, provided the consistency bounds $\\|\\hat\\Pi_k-\\Pi_{T_{\\xi_k}M}\\|_{\\mathrm{op}}\\le\\beta$ and $\\|\\hat n(\\tilde\\xi_{k+1};\\xi_k)-n(\\tilde\\xi_{k+1})\\|\\le\\gamma$ hold along the iteration. In words, the learned gradient descent converges to a neighbourhood of the local minimiser, and that neighbourhood shrinks to nothing as the step size $\\tau$, the tangent estimation error $\\beta$, and the normal estimation error $\\gamma$ all tend to zero. The same mechanism resolves a practical difficulty in learned operator correction: because the iterates never leave the data manifold, the corrected normal operator is only ever evaluated where it was trained, so the scheme converges without retraining along the trajectory.","pith_inferences":["Beyond the paper: the theory only needs bounds $\\beta$ and $\\gamma$, so local PCA could be swapped for any other tangent and normal estimator with the same consistency guarantees without changing the convergence statement.","Beyond the paper: the observed divergence of the corrected scheme near the minimiser suggests a practical design principle, early stopping when the gradient norm regrows, which the theorem's neighbourhood bound rationalises as the size of the residual ball $2\\delta_\\tau/\\mu$.","Beyond the paper: a testable extension is adaptive sampling, refining the point cloud near the minimiser instead of globally, which the experiments suggest is necessary because fine global sampling is the main bottleneck.","Beyond the paper: if the data do not lie on a smooth manifold with positive reach, condition (2.32) may fail; running the algorithm on a noisy or non-smooth data set would show whether the practical convergence persists outside the theorem's assumptions."],"forward_implications":["If the consistency bounds hold with $\\tau,\\beta,\\gamma\\to 0$, the iterates converge to the local minimiser without any geodesic convexity or curvature information about the manifold.","In the photoacoustic example, optimising over the point-cloud manifold with the corrected normal operator achieves reconstructions comparable to the accurate model, while unconstrained optimisation with the same corrected operator fails.","A single correction of the normal operator $\\tilde A^*\\tilde A$ suffices, avoiding the separate forward and adjoint corrections needed in earlier learned-correction schemes, because the projection keeps iterates where the correction is valid.","The manifold projection replaces the explicit regulariser: the authors omit $R$ when optimising on the manifold, and the projection itself acts as the regulariser, with total variation still giving the best quantitative result for the accurate model."],"supporting_citations":[{"why":"Supplies the projection and normal-vector properties on tubular neighbourhoods used throughout the continuum and discrete analyses.","marker":"[27]"},{"why":"Defines learned operator correction, the motivating application and the baseline that required retraining along the trajectory.","marker":"[39]"},{"why":"Provides the compact embedding used to turn weak convergence of the interpolated discrete flow into uniform convergence.","marker":"[11]"},{"why":"Gives the known tangent-space estimation result for constant density that Theorem 2.6 generalises to nonuniform densities.","marker":"[47]"},{"why":"Supplies the Davis-Kahan variant used to control the PCA tangent-projection error in Theorem 2.6.","marker":"[49]"},{"why":"Provides the Jacobian bound used in Lemma 2.8's expansion of the local covariance near the manifold.","marker":"[18]"}],"fun_headline_variants":["Point-cloud GD converges to minimisers despite unknown manifold","Learned tangent gradients steer point-cloud descent to minimisers","Data-driven gradient descent on point clouds reaches local minima","Point-cloud GD avoids retraining, converges to manifold minimisers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence rests on the assumption that the ideal projected flow is locally exponentially stable at the minimiser and that the learned tangent projection and normal vector obey the deterministic consistency bounds (2.33)-(2.34), which are assumed rather than verified in the photoacoustic application.","fun_headline_variants_meta":{"raw":{"variants":["Point-cloud GD converges to minimisers despite unknown manifold","Learned tangent gradients steer point-cloud descent to minimisers","Data-driven gradient descent on point clouds reaches local minima","Point-cloud GD avoids retraining, converges to manifold minimisers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000864,"raw_usage":{"total_tokens":3724,"prompt_tokens":898,"completion_tokens":2826,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":2760}},"tokens_in":514,"tokens_out":2826,"duration_ms":23212,"temperature":1.0,"reasoning_tokens":2760,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:01:00.586392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a known manifold with a dense point cloud and known minimiser $x^*$, run the algorithm while sending $\\tau$, the PCA radius $\\varepsilon$, and the tangent estimation error $\\alpha$ to zero and measure $\\limsup_k\\|\\xi_k-x^*\\|$; if this limit does not approach zero, either condition (2.32) or the consistency bounds (2.33)-(2.34) fail for that instance. Alternatively, compute the empirical $\\beta$ and $\\gamma$ along the PAT trajectory and check whether they shrink with denser sampling and smaller steps.","supporting_citations":[{"cited_title":"Curvature measures","cited_arxiv_id":null,"evidence_quote":"Supplies the projection and normal-vector properties on tubular neighbourhoods used throughout the continuum and discrete analyses."},{"cited_title":"On Learned Operator Correction in Inverse Problems","cited_arxiv_id":null,"evidence_quote":"Defines learned operator correction, the motivating application and the baseline that required retraining along the trajectory."},{"cited_title":"Brezis and H","cited_arxiv_id":null,"evidence_quote":"Provides the compact embedding used to turn weak convergence of the interpolated discrete flow into uniform convergence."},{"cited_title":"Tangent space estimation for smooth embeddings of riemannian manifolds®","cited_arxiv_id":null,"evidence_quote":"Gives the known tangent-space estimation result for constant density that Theorem 2.6 generalises to nonuniform densities."},{"cited_title":"A useful variant of the Davis–Kahan theorem for statisticians","cited_arxiv_id":null,"evidence_quote":"Supplies the Davis-Kahan variant used to control the PCA tangent-projection error in Theorem 2.6."},{"cited_title":"A graph discretization of the Laplace–Beltrami operator","cited_arxiv_id":null,"evidence_quote":"Provides the Jacobian bound used in Lemma 2.8's expansion of the local covariance near the manifold."}],"review_version":1}