{"id":"57c5f54f-0150-41f0-a0c9-6e6df3037b82","arxiv_id":"1908.03776","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A finite-element-based lifting framework gives sublabel-accurate convex relaxations for manifold-valued variational problems with general convex regularizers.","lead":"This paper extends functional lifting, a technique that converts hard image-processing problems into convex ones, to images whose pixel values lie on curved spaces such as circles, spheres, or rotation groups. It uses finite-element meshes on those curved spaces to reduce the number of sample points needed and to support regularizers beyond total variation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central transfer of exactness from vectorial lifting [51] to manifolds is asserted, not proved: absent a tightness result for the relaxation (11)-(12), a global minimizer of the lifted problem need not correspond to any near-minimizer of (10).","rationale":"The formal lifting is load-bearing because every advertised advantage—global optimality, sublabel accuracy, reduced label counts—depends on the relaxed problem's optimizer being projectable to a near-optimal solution of (10). The paper explicitly defers the analysis that would establish this, and it concedes the general difficulty of projection in §1.1. This is not an objection to the empirical framework; the code, reproducible experiments, and qualitative results are real evidence that the method works on the tested instances. But they do not establish the central transfer for arbitrary convex regularizers and coarse triangulations. The reader's weakest assumption points to the same formal step; my concern sharpens it by identifying the concrete mechanism (loose relaxation, non-Dirac minimizers, non-convex projection) and by proposing a quantitative test. Because no new evidence contradicts the paper's claims and the issue is exactly the acknowledged formal gap, the conditional verdict stands.","tokens_in":19314,"tokens_out":8088,"duration_ms":95407,"concrete_test":"Take S^1 with Ω=[0,1], a piecewise constant noisy signal concentrated at two antipodal values, a quadratic data term, and a small regularization weight λ. Solve the discretized lifted problem (24) on a coarse 8-vertex triangulation and on a fine reference triangulation. Compare (a) the original energy (10) evaluated at the projected lifted solution, (b) the known global minimum computed by exhaustive one-dimensional search, and (c) the mean integrality ∫ max_k v_k dx / |Ω|. If the projected energy exceeds the global minimum by more than a small tolerance (e.g., 1% of the data term scale) or the solution is consistently non-Dirac, the claimed near-global-optimality transfer is not realized for manifold ranges. Repeating on S^2 and SO(3) with the paper's own datasets would determine whether the failure is systematic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline promise is that the lifted convex problem can be solved globally and that projection recovers an approximate global minimizer of (10). The only support for this on manifolds is the sentence in §2.1 that the vectorial construction of [51] 'can be generalized to this setting by replacing the range Γ with M,' immediately followed by the admission that 'most derivations will be formal' and that function spaces and well-posedness are left for future work. In the scalar and vectorial cases, exactness and calibration results justify the relaxation; for manifold ranges, no analogue is supplied. The relaxation from M to P(M) is known to be loose in general: the paper itself notes in §1.1 that lifted minimizers need not lie in the image of the embedding and that projection is not guaranteed to minimize the original functional, and §2.4's Riemannian center-of-mass projection is non-convex and lacks any optimality bound. Thus the global optimum of the discretized saddle-point problem (24)-(27) may, even in the continuum limit, project to a value with original energy substantially above the true minimum. The discrete geometry adds a second gap: the replacement of P_z by per-simplex flat maps P_T in (18)-(19) changes the problem to one on a piecewise-flat manifold, with only an S^1 scalar correction and no Γ-convergence or error analysis. The numerical sections are qualitative and compare against the authors' own earlier method at different mesh resolutions, so they do not close either gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey-style chapter on functional lifting for first-order variational problems u: Ω → M, where M ⊂ R^N is a Riemannian submanifold. The authors embed M into the space of probability measures via Dirac measures, define a lifted functional (11) through the convex conjugate of the integrand, and discretize the range by a triangulated manifold M_h with a first-order finite element space (15)-(16). After replacing the tangent-space projection P_z by per-simplex projections P_T in (17)-(19), they arrive at the saddle-point formulation (24)-(27). The paper claims that this framework extends sublabel-accurate lifting to manifolds and generalizes the previous TV-only lifting [47] to arbitrary convex regularizers such as Huber and Tikhonov. Several qualitative experiments are reported (Klein bottle, SO(3), S^2 normals, InSAR S^1 data). The paper explicitly declares the derivations in Section 2 formal and leaves function-space and well-posedness questions for future work.","tokens_in":19625,"tokens_out":8541,"duration_ms":85625,"significance":"If the central claims are correct, the paper offers a useful unifying presentation and a practical algorithmic extension: first-order finite elements on the range reduce label bias, the framework handles non-orientable manifolds and general convex regularizers, and the released code and real-data experiments make the approach reproducible. The treatment of non-orientable manifolds and the extension beyond TV are genuine contributions. However, because no tightness theorem, Γ-convergence result, or error estimate is proved, the significance is conditional: the paper does not establish that the global minimizer of the lifted problem yields a near-global minimizer of the original manifold-valued problem. The numerical evidence is visual and compares mainly with the authors' earlier methods, so the 'sublabel accuracy' advantage is not yet quantitatively validated. The paper is honest about these limitations, which is a strength.","major_comments":[{"comment":"The central promise that the lifted convex problem can be solved globally and that its projection approximates a global minimizer of (10) rests on an unproved transfer of exactness from the vectorial theory [51]. The text states that the construction of [51] can be generalized 'by replacing the range Γ with M' and immediately adds that 'most derivations will be formal' and that well-posedness is left for future work. No analogue of the calibration or exactness arguments available for scalar/vectorial ranges is supplied for manifold ranges, and Section 1.1 explicitly concedes that no projection guarantee is known in the manifold case. Since the relaxation from M to P(M) is not tight in general, the global minimizer of (11) may project to a value substantially above the infimum of (10). This gap is load-bearing for the headline claim and should be addressed, at minimum by formulating the exactness claim as a conjecture with a precise statement and by providing numerical evidence of the energy gap on a simple test problem.","section":"Section 2.1, Eqs. (10)-(12)"},{"comment":"The substitution of the pointwise tangent-space projection P_z by per-simplex flat projections P_T changes the model to a piecewise-flat surrogate manifold. The paper derives a scalar correction only for S^1 in Section 2.3 and otherwise asserts empirical negligibility. No Γ-convergence, interpolation error, or mesh-refinement analysis is given for M_h → M. Without such a consistency result, the claim that sublabel accuracy allows much coarser range discretizations than [47] is not quantitatively established; the authors should either add a convergence analysis or report a numerical mesh-refinement study showing that the solution stabilizes as the triangulation is refined.","section":"Section 2.2, Eqs. (17)-(19)"},{"comment":"The final projection step to M via the Riemannian center of mass (32) is itself a non-convex problem with no optimality bound, as the paper acknowledges. Because this projection is an essential part of the pipeline, the lack of any guarantee for it compounds the tightness gap in the first major comment and directly affects the claim that the method finds approximate global minimizers of (10). Even an assumption-dependent bound, for example for data terms supported in a geodesically convex ball, would clarify the status of the claim.","section":"Section 2.4, Eq. (32)"},{"comment":"The numerical evaluation is qualitative and compares mainly with the authors' own earlier methods at different mesh resolutions (12 vs 162 vertices on S^2 in Section 3.3; 60 vs 720 on SO(3) in Section 3.2). No energy values, ground-truth errors, or comparisons with local optimization baselines are reported. Since sublabel accuracy and label-bias reduction are central claims, a quantitative experiment, such as projected energy versus the true minimum on a problem with known solution, or error versus number of labels, is needed to substantiate the visual improvement.","section":"Section 3"}],"minor_comments":[{"comment":"The sentence 'The resolution of the signal (250 one-dimensional data points) is far below the resolution of the triangulation' appears to say the opposite of what is meant, since 250 data points are many more than the 25 vertices of the 5×5 triangulation; please rephrase.","section":"Section 3.1"},{"comment":"The range of the dual variable p is not consistently stated: below (17) the text says p∈ S_h^d, while (24) writes p∈ C^1(Ω, S_h^{d+1}); please align the notation and state the dimension of p explicitly.","section":"Equation (24)"},{"comment":"The stopping criterion 'relative gap between primal and dual objective fell below 10^{-5}' is not defined; please specify how the primal and dual objectives are evaluated for the discretized saddle-point problem.","section":"Section 2.4"},{"comment":"The caption contains an apparent typesetting artifact ('/tieaccentlowercase'); please correct it.","section":"Figure 5 caption"},{"comment":"The sentence 'This is very desirable for discrete Γ' appears right after a discussion of label bias; it would be clearer to state explicitly that label bias is desirable only in the discrete-label setting, not in the continuous-range setting.","section":"Section 1.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a survey-style chapter rather than a full research article, and its central manifold extension is explicitly formal. The main gap is the absence of any rigorous consistency statement for the relaxation, compounded by the non-convex projection step. The numerical section would benefit from an independent baseline and quantitative energy comparisons. I recommend major revision rather than rejection because the framework is coherent, the code is available, and the scientific claims could be made precise with reasonable additional work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a survey with a real extension: it takes functional lifting from TV-only on manifolds to general convex regularizers, and it reframes the discretization as finite elements on the range manifold, which reduces label bias. The Klein bottle and SO(3) experiments, plus the released code, make the claims concrete. I'd use it as a reference for sublabel-accurate lifting for manifold-valued problems.\n\nThe paper is also honest: it says in Section 2.1 that the derivations are formal and that function spaces and well-posedness are left for future work. That is exactly where the main risk sits. The exactness of the lift for vectorial problems, which this generalizes, is justified by calibration/tightness results. No analogue is proved here. Without a tightness result, the relaxed problem's global minimizer need not project to a near-minimizer of the original manifold-valued problem. The paper's own projection via Riemannian center of mass is non-convex and lacks optimality bounds, so the gap is real.\n\nThe second soft spot is the piecewise-flat approximation M_h. Replacing the manifold by a simplicial complex changes the geometry; the authors give only an S^1 curvature correction and rely on empirical negligibility. No Gamma-convergence or error analysis is provided. If the paper is read as a survey with heuristic improvements, that is fine; but the abstract promises approximate global minimizers, which needs more support.\n\nThe numerics are qualitative and compare against the authors' earlier method at different mesh resolutions. That does not close the theoretical gaps, though it does demonstrate practical benefit: 12 vertices on the icosahedron versus 162 previously.\n\nI would send it to peer review. It deserves referee time, but a careful referee should ask for either a rigorous tightness/convergence analysis or a clear statement that the method is heuristic, plus some quantitative error evaluation on a problem with a known solution. The paper is worth reading for the survey alone; the extension is plausible but unproven.","headline":"A useful survey-and-extension of functional lifting for manifold-valued problems, with the key exactness claim formally unproven but honestly flagged.","tokens_in":20182,"tokens_out":2089,"would_cite":true,"duration_ms":22873,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K10","65N30","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that first-order finite-element lifting makes manifold-valued variational problems solvable to a global optimum with far fewer range labels than earlier methods.","keywords":["functional lifting","manifold-valued image processing","convex relaxation","sublabel-accurate discretization","finite element methods","total variation","Huber regularization","primal-dual hybrid gradient"],"falsifier":"Take a manifold-valued denoising problem with a known exact global minimizer, for instance the $S^1$ two-point mean from the paper's motivating example, solve the lifted saddle-point problem with the coarsest reported triangulation, and compare the projected solution's original energy with the exact global energy. If the projected solution lands on the local minimum that gradient-style methods find, or if the energy gap does not shrink toward zero as the triangulation is refined, the central claim of near-global, label-bias-free optimization fails for that instance.","tokens_in":19127,"feed_emoji":"🌐","tokens_out":10477,"duration_ms":103285,"temperature":0.7,"pith_summary":"This paper tries to establish that functional lifting, the technique of turning a hard variational problem into a convex one by embedding the value range into a higher-dimensional space, works for images whose values live on a curved manifold when the range is discretized with first-order finite elements. The authors argue that this finite-element discretization gives sublabel-accurate lifting on manifolds: the data term is represented more faithfully between the chosen sample points, so far fewer sample points are needed than in earlier manifold lifting based on total variation. The paper also drops the total-variation-only restriction, so regularizers such as Tikhonov and Huber fit the same convex saddle-point formulation. If the construction is right, one convex optimization routine can find approximate global minimizers for manifold-valued denoising, inpainting, and related problems, and the back-projected solutions stay close to the true optimum even on coarse range triangulations.","feed_headline":"Manifold-valued imaging goes convex with far fewer labels","feed_subtitle":"Lifting image values to probability measures turns nonconvex denoising on spheres and rotation groups into a solvable saddle-point problem.","key_machinery":"The central object is the first-order finite element space $S_h$ on a triangulated manifold $M_h\\subset\\mathbb{R}^N$, with nodal basis $\\chi_1,\\ldots,\\chi_L$. The space of probability measures $P(M)$ is replaced by the convex set of nonnegative measures $\\mu_h$ on the vertices satisfying $\\sum_k \\langle\\mu_h,\\chi_k\\rangle=1$, i.e. the probability simplex over the $L$ labels. The integrand is split as $\\rho(x,z)+\\eta(P_z\\xi)$, where $P_z$ maps ambient Jacobians into the tangent subspace; on each simplex $T$ this becomes two epigraphical constraints involving the convex conjugates $\\eta^*$ and $\\rho^*_T$, linked by a linear equation. These constraints are what make the optimization Euclidean despite the curved range, and the resulting convex-concave saddle-point problem is solved by a primal-dual hybrid gradient iteration. On $S^1$ the construction also yields a curvature correction factor $\\alpha_T=d_{S^1}(Z_1^T,Z_2^T)/\\|Z_1^T-Z_2^T\\|_2$, the ratio of geodesic to Euclidean distance between the two vertices of a simplex, which rescales the finite element gradient.","core_discovery":"On its own terms, the paper claims to unify and extend manifold-valued functional lifting through a finite element interpretation. For a variational problem of the form $\\min_{u:\\Omega\\to M} \\int_\\Omega (\\rho(x,u(x)) + \\eta(P_{u(x)}Du(x)))\\,dx$ with convex regularizer $\\eta$, the range $M$ is approximated by a triangulated manifold $M_h$, and each solution is lifted to a probability measure over the vertices of the triangulation. The lifted energy is written in saddle-point form: the convex conjugate of $\\eta$ enters through the constraint $\\eta^*(P_T\\nabla_T p(x)) \\le a_T(x)$, and the data term through $\\rho^*_T(q_{T,1}(x)) \\le b_T(x)$, with the linear coupling $a_T(x)+b_T(x)=-q_{T,2}(x)$. Because $p$ is piecewise linear on each simplex, the sublabel-accurate data approximation of Euclidean lifting carries over to the manifold. Numerically, the paper reports that TV, Huber, and Tikhonov regularizers all run in this framework on orientable and non-orientable manifolds, and that good results need only 60 vertices for $SO(3)$ instead of 720, 12 vertices for $S^2$ instead of 162, and a $5\\times 5$ grid on the Klein bottle.","pith_inferences":["The finite-element viewpoint suggests trying higher-order elements or adaptive refinement on the range manifold to push the required number of labels below the counts reported here; the paper itself only demonstrates first-order elements.","If the equivalence between the lifted convex problem and the original nonconvex problem can be proven rather than assumed, it would provide a priori optimality bounds for manifold-valued regularizers, which the paper notes are currently missing for the manifold case.","The per-simplex correction factor on $S^1$ hints that on more curved manifolds a per-simplex linear transformation between the geometric and the finite-element gradient could systematically correct coarse-triangulation bias; the paper finds the effect negligible in its experiments, so this is a testable prediction rather than a claim.","Because the lifted problem is a Euclidean convex-concave saddle-point problem, it inherits the fast GPU-friendly solvers of scalar lifting, which could make global manifold optimization practical at the resolution of real InSAR and elevation data."],"forward_implications":["The same lifted saddle-point formulation handles total variation, Huber, and quadratic (Tikhonov) regularization for manifold-valued images, generalizing earlier methods that were limited to total variation.","Because the data term is approximated convexly between labels rather than only at labels, coarse range triangulations suffice: 60 vertices on $SO(3)$ and 12 on $S^2$ are reported to match earlier results that needed hundreds of labels.","The method applies to non-orientable manifolds such as the Klein bottle, because the tangent-basis map $P_z$ is not required to be continuous.","After the lifted problem is solved to its globally optimal saddle point, the solution is projected back to the manifold by a Riemannian center-of-mass computation, yielding a concrete denoising and inpainting pipeline."],"supporting_citations":[{"why":"Baseline manifold lifting restricted to total variation; this paper extends it to general convex regularizers.","marker":"[47]"},{"why":"Supplies the formal lifting construction via currents that the manifold version assumes carries over.","marker":"[51]"},{"why":"Introduces the finite-element interpretation of sublabel-accurate lifting whose constraint splitting is reused here.","marker":"[50]"},{"why":"Provides the convex sublabel-accurate data-term approximation and the epigraph projections used for the data term.","marker":"[38]"},{"why":"Supplies surface finite element machinery for defining and discretizing the range manifold.","marker":"[25]"},{"why":"Provides the epigraph projection formulas and solver components for TV, quadratic, and Huber regularizers.","marker":"[54]"},{"why":"Gives the Riemannian center-of-mass iteration used to project lifted solutions back to the manifold.","marker":"[35]"}],"fun_headline_variants":["Manifold lifting goes convex with sublabel accuracy","Fewer vertices for convex manifold imaging","Sublabel-accurate lifting solves nonconvex manifold problems","Finite elements cut vertex count in manifold lifting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the formal lifting step, carried over from vector-valued problems without a complete proof, truly preserves the manifold problem's global minimizers; if the convex relaxation is not tight, the global solution of the lifted problem need not correspond to a global or near-global minimizer of the original problem.","fun_headline_variants_meta":{"raw":{"variants":["Manifold lifting goes convex with sublabel accuracy","Fewer vertices for convex manifold imaging","Sublabel-accurate lifting solves nonconvex manifold problems","Finite elements cut vertex count in manifold lifting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2315,"prompt_tokens":928,"completion_tokens":1387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1325}},"tokens_in":544,"tokens_out":1387,"duration_ms":10310,"temperature":1.0,"reasoning_tokens":1325,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:02:19.224477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a manifold-valued denoising problem with a known exact global minimizer, for instance the $S^1$ two-point mean from the paper's motivating example, solve the lifted saddle-point problem with the coarsest reported triangulation, and compare the projected solution's original energy with the exact global energy. If the projected solution lands on the local minimum that gradient-style methods find, or if the energy gap does not shrink toward zero as the triangulation is refined, the central claim of near-global, label-bias-free optimization fails for that instance.","supporting_citations":[{"cited_title":"In: Proc ICCV 2013, pp","cited_arxiv_id":null,"evidence_quote":"Baseline manifold lifting restricted to total variation; this paper extends it to general convex regularizers."},{"cited_title":"In: Proc CVPR 2019 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the formal lifting construction via currents that the manifold version assumes carries over."},{"cited_title":"In: Proc ICCV 2017, pp","cited_arxiv_id":null,"evidence_quote":"Introduces the finite-element interpretation of sublabel-accurate lifting whose constraint splitting is reused here."},{"cited_title":"In: Proc ECCV 2016, pp","cited_arxiv_id":null,"evidence_quote":"Provides the convex sublabel-accurate data-term approximation and the epigraph projections used for the data term."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies surface finite element machinery for defining and discretizing the range manifold."},{"cited_title":"SIAM J Imaging Sci3(4), 1122–1145 (2010)","cited_arxiv_id":null,"evidence_quote":"Provides the epigraph projection formulas and solver components for TV, quadratic, and Huber regularizers."},{"cited_title":"Commun Pure Appl Math 30, 509–541 (1977)","cited_arxiv_id":null,"evidence_quote":"Gives the Riemannian center-of-mass iteration used to project lifted solutions back to the manifold."}],"review_version":1}