{"id":"b99311a8-3292-4429-b2f8-4036003ec2b3","arxiv_id":"2411.16589","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors define the Grassmann Distance Complexity for subanalytic sets in Grassmannians, prove uniform bounds for algebraic hypersurfaces, and derive a nonlinear Eckart-Young theorem for simple Schubert varieties.","lead":"This paper introduces a new complexity measure, the Grassmann Distance Complexity, which quantifies how hard it is to find the closest point on a curved shape inside a space of planes. The authors prove the measure is finite for a wide class of shapes and use it to solve a natural optimization problem for Schubert varieties, a nonlinear analogue of the classical Eckart-Young low-rank approximation.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 42 rests on an asserted but unproved o-minimal definability of the critical-pair set A; if that assertion fails, the uniform finiteness bound for GDC collapses.","rationale":"The reader's weakest assumption correctly identifies the unproved definability of the critical-pair set A as the main gap in the proof of Theorem 42. This is indeed the load-bearing point for the central claim that GDC(X) is generically finite and uniformly bounded: without definability, the o-minimal trivialization argument cannot start, and the paper provides no alternative argument. The assertion is very likely true (the inverse exponential map and the gradient of a definable C^1 function are definable in a globally subanalytic structure), but the paper gives neither a proof nor a citation. I therefore agree with the CONDITIONAL verdict: the result probably holds, but the proof is incomplete pending this justification. I also checked the Schubert-variety arguments (Theorems 54, 61, 64); they appear sound and do not introduce a more serious concern. The Pfaffian bound proof of Theorem 49 contains a separate, smaller gap: the SVD parametrization can fail to be a submersion at repeated singular values, so some solutions of the Lagrange equations may be spurious; nevertheless, since every true critical point lifts to a solution and each connected component of the solution set maps to at most one critical point, the inequality GDC(X) ≤ b0(S) can be repaired. That issue is not load-bearing for the central finiteness claim, which is why I focus on the definability assertion.","tokens_in":38055,"tokens_out":33522,"duration_ms":330926,"concrete_test":"Produce an explicit first-order definition of A in the o-minimal structure of globally subanalytic sets. Concretely: let v(L,S) be the unique vector in int(R_{π/2}) ⊂ T_L G(k,n) with exp_L(v)=S, defined on the definable set {(L,S): S ∉ cut(L)} and shown definable because its graph is the swap of the graph of the definable exponential map. Then express 'S is critical for δ_L|_X' as the vanishing of g_S(D_v exp_L(v/||v||), w) for all w in T_S X, where the metric and the derivative D_v exp_L are definable. If such a formula cannot be written, Theorem 42 is unproved; if it can, insert it as a lemma before Theorem 42 and explicitly invoke the definability of derivatives in o-minimal structures.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 42 defines A = {(S,L) in X×G(k,n) | S notin cut(L), S is critical for δ_L|_X} and states: 'By Proposition 23 and the definability of X, it follows that also the set A is definable.' This is the load-bearing step: the o-minimal trivialization argument that produces the uniform bound C_X applies only if A is definable in an o-minimal structure, and the conclusion that GDC(X) is finite (rather than infinite) depends on it. Proposition 23, however, only proves that the principal angles and the distance function δ are globally subanalytic. Criticality of S for δ_L|_X involves the gradient of δ_L at S, i.e. the tangent vector obtained from the unique length-minimizing geodesic from L to S, and the condition that this vector is orthogonal to T_S X. The paper does not prove that this gradient map, or the orthogonality condition written through the normal bundle, is definable. The claim is likely true: away from the cut locus the inverse exponential map v(L,S) = exp_L^{-1}(S) is the inverse of a definable diffeomorphism, and derivatives of definable functions in o-minimal structures are definable. But the proof is absent, and no reference is given. The same unstated dependence appears in Theorem 41 for noncompact subanalytic X, where finiteness is dismissed with 'finiteness follows from o-minimality'. Thus the central finiteness theorem is conditional on a nontrivial definability assertion that the paper does not justify.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Grassmann Distance Complexity (GDC) for subanalytic submanifolds of the real Grassmannian G(k,n), counting, for a generic external plane L, the number of critical points of the intrinsic distance function δ_L restricted to X, excluding points on the cut locus. The main theoretical results are: a description of the Clarke subdifferential of δ_L (Theorem 30); a proof that the number of critical points outside the cut locus is generically finite and uniformly bounded (Theorems 41 and 42); a Pfaffian bound on GDC for algebraic hypersurfaces (Theorem 49); and an analysis of the nearest-point problem for simple Schubert varieties, including a nonlinear Eckart-Young theorem identifying the unique global minimizer and an explicit family of global maximizers (Theorems 54, 61, 64).","tokens_in":38385,"tokens_out":17858,"duration_ms":176736,"significance":"If the gaps identified below are repaired, the paper makes a valuable contribution: it proposes a natural Riemannian analogue of Euclidean Distance Degree, correctly identifies the role of the cut locus and of Clarke critical points, and gives an explicit and apparently correct solution of the nearest-point problem for simple Schubert varieties. The subdifferential computation in Theorem 30 and the Eckart-Young critical points in Theorem 54 are concrete and useful, and the explicit global minimizer and maximizer in Theorems 61 and 64 are appealing. The Pfaffian bound in Theorem 49, though astronomically large, is a genuine effectivity statement. The work is not a repackaging of known results: the o-minimal framework, the treatment of the cut-locus strata, and the Schubert-variety analysis are new.","major_comments":[{"comment":"The proof of Theorem 42 asserts, immediately after the definition of A = {(S,L) in X×G(k,n) | S notin cut(L), S critical for δ_L|_X}, that 'By Proposition 23 and the definability of X, it follows that also the set A is definable.' This assertion is load-bearing: the o-minimal trivialization argument that produces the uniform bound C_X and the finiteness of GDC applies only if A is definable. Proposition 23 establishes that the principal angles and the distance function δ are globally subanalytic, but criticality of S for δ_L|_X involves the gradient of δ_L at S, which is the tangent vector obtained from the inverse exponential map exp_L^{-1}(S), and the orthogonality of that gradient to T_S X. The paper neither proves definability of this gradient map nor provides a reference for this step. The same missing definability is needed in Theorem 41 for noncompact subanalytic X, where the sentence 'finiteness follows from o-minimality' presupposes that the critical-point set is definable. Please supply a proof or a precise citation for the definability of the critical-pair set.","section":"Section 3.5 (Theorem 42)"},{"comment":"The semi-Pfaffian set S is defined by the Lagrange multiplier equations (36) together with the maximal-rank condition rank([∂(U,V) p̃; D_(U,V)F]) = k(k+1). These equations are necessary only at regular points of the constraint Y = ~exp_L^{-1}(X\\cut(L)). The parametrization Φ(U,V,μ) = U diag(μ) V^T is not a submersion at points with repeated singular values, so Y can be singular there, and a critical point of δ_L|_X whose every SVD lift lies in such a singular locus would not be represented in S. The proof does not show that every critical point admits a lift satisfying the maximal-rank condition (37), nor does it treat the repeated-singular-value case separately. Since the estimate GDC(X) ≤ b0(S) in (40) requires every critical point to have at least one preimage in S, this gap affects the validity of Theorem 49 as stated.","section":"Section 3.7 (Theorem 49)"},{"comment":"The Grassmann Distance Complexity is defined as a maximum over 'generic' L, but the proof of Theorem 42 only produces a definable set ~M, the union of strata with zero-dimensional fibers, on which the count is bounded by C_X; the complement of ~M has codimension at least one. It is not stated that the count is constant on each connected component of ~M, nor is the meaning of 'generic' made precise, for example as belonging to a nonempty Zariski-open or definable dense subset. Without these clarifications, GDC(X) is not unambiguously a single number. The trivialization theorem does imply constancy of the fiber cardinality on each stratum, so this should be stated explicitly in the definition or in the proof of Theorem 42.","section":"Definition 43 / Theorem 42"}],"minor_comments":[{"comment":"There are numerous typos, including 'Grassmann Dist ance' in the title, 'isomoprhism' in Definitions 13 and 14, 'Riemmanian sphere' in Section 3.6, and 'π 4 2' in the statement of Theorem 64, which should be π^2/4.","section":"Throughout"},{"comment":"In equation (22), the symbol W is used both for the orthogonal matrix variable in the convex hull and later for the fixed subspace W defining Schubert varieties; this notational clash should be resolved.","section":"Theorem 30"},{"comment":"The sentence 'every connected component of S corresponds to exactly one critical point' is imprecise; the intended statement is that the continuous map from S to X sends each connected component to a single critical point and that every critical point has at least one preimage, so that GDC(X) ≤ b0(S).","section":"Section 3.7 (Theorem 49)"},{"comment":"In the uniqueness proof, the assertion that a minimizer E is an orthogonal direct sum E = E_L ⊕ E_W is not justified in the text. It follows from the equality conditions in the interlacing argument together with the uniqueness of principal vectors for distinct angles, but this step should be spelled out.","section":"Theorem 61"},{"comment":"The paper assumes k ≤ n−k from Section 2.2 onward, but Theorem 5 and some introductory statements claim results 'for every 0≤k≤n'; the two statements should be harmonized.","section":"Introduction and Theorem 5"}],"recommendation":"major_revision","confidential_remarks":"The reader's stress-test concern is real: the definability assertion in Section 3.5 is exactly the load-bearing step for the central finiteness theorem, and the current text does not justify it. The SVD-regularity gap in Theorem 49 is also genuine and needs a separate argument. These appear to be technical gaps rather than fatal flaws, since the claims are plausible and the Schubert variety analysis is largely independent and convincing. I recommend major revision rather than rejection, and I would expect the authors to fill the definability proof and to clarify the generic-maximum definition before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper introduces the Grassmann Distance Complexity (GDC), a sensible analogue of the Euclidean Distance Degree for subanalytic subsets of the Grassmannian, and it solves the nearest point problem for simple Schubert varieties with an explicit Eckart–Young type minimizer. The main theorems are genuinely new and mostly well proved. The soft spot is a specific unproved definability assertion in Theorem 42, which is load-bearing for the finiteness of GDC. It is likely fixable, but as written it is a gap.\n\nWhat is actually new: the definition of GDC, the subdifferential analysis at the cut locus (the convex hull of O(j) description is nice), the proof that minima generically avoid the cut locus, the uniform finiteness theorem, and the nonlinear Eckart–Young theorem for Schubert varieties. The explicit unique minimizer, with distance (θ_1² + ... + θ_s²)^{1/2}, is a clean and surprising result. The family of global maximizers on a Grassmannian of dimension (k−s)(n−k−s) is also very concrete. The paper is honest: Remark 60 admits the authors do not yet understand why the Eckart–Young critical points are exactly the Grassmannian critical points, and the Pfaffian bounds are acknowledged to be astronomically large but existentially meaningful.\n\nWhere the soft spots are: the stress-test note is on target. In Theorem 42 the set A of critical pairs is asserted definable with no proof and no reference. The claim is plausible—away from the cut locus the inverse exponential map is definable, and derivatives of definable functions in o-minimal structures are definable—but this is not automatic and the paper does not supply the argument. Since the entire definition of GDC as a finite number depends on Theorem 42, this needs to be fixed. Similarly, the noncompact subanalytic case in Theorem 41 is dismissed with a one-line 'finiteness follows from o-minimality' that deserves a sentence or two more.\n\nEverything else looks solid. The Schubert variety analysis is careful, the genericity arguments are standard, and the non-semialgebraicity example is convincing.\n\nWho should read this: anyone working in metric algebraic geometry or on optimization over Grassmannians. It deserves a serious referee. The right request is: repair the definability gap (or cite a theorem that covers it), and the paper will be in good shape.","headline":"A new complexity measure for the Grassmannian with a clean nonlinear Eckart–Young theorem, but the finiteness theorem rests on a plausible, unproved definability assertion.","tokens_in":38861,"tokens_out":2358,"would_cite":true,"duration_ms":24895,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["53C22","14M15","32B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Nearest-point search on Grassmannians has finite generic complexity, and Schubert varieties are solved by Eckart–Young.","keywords":["Grassmann distance complexity","Euclidean distance degree","subanalytic sets","o-minimal geometry","Lipschitz critical point theory","Schubert varieties","Eckart-Young theorem","Pfaffian functions"],"falsifier":"For a concrete smooth subanalytic $X$, write the critical-point relation explicitly and check whether it is definable in an o-minimal structure; a single $X$ for which this relation is not definable, or a sequence of generic $L_m$ with an unbounded number of critical points outside $\\mathrm{cut}(L_m)$, would refute Theorem 42. A cheaper numerical check for a Schubert variety like $\\Omega_s\\subset G(3,6)$ is to count, for random $L$, all critical points of $\\delta_L|_{\\Omega_s}$ outside the cut locus and see whether the count exceeds $\\binom{k}{s}$.","tokens_in":37851,"feed_emoji":"🎯","tokens_out":8708,"duration_ms":74046,"temperature":0.7,"pith_summary":"This paper introduces the Grassmann Distance Complexity (GDC) of a subanalytic submanifold of a Grassmannian: the maximum, over generic external $k$-planes $L$, of the number of critical points of the intrinsic Riemannian distance from $L$ that lie outside $L$'s cut locus. The central claim is that this number is finite and bounded by a constant $C_X$ depending only on the submanifold $X$ (Theorem 42). Because the Grassmann distance is neither smooth nor semialgebraic, the proof works with Clarke's subdifferential for Lipschitz functions and with o-minimal geometry, a tame logical framework in which principal angles are definable. For real algebraic hypersurfaces the bound is made explicit in the degree via Pfaffian function theory. The paper also proves a nonlinear Eckart–Young theorem: for simple Schubert varieties $\\Omega_s$, the unique closest point to a generic $L$ is obtained by zeroing the $s$ smallest principal angles, with distance $(\\theta_1(L)^2+\\cdots+\\theta_s(L)^2)^{1/2}$.","feed_headline":"Nearest-point complexity on Grassmannians is finite","feed_subtitle":"New GDC bounds the points to check and solves Schubert varieties via Eckart–Young.","key_machinery":"The load-bearing object is the Clarke subdifferential $\\partial_S\\delta_L$ of the Lipschitz distance function, together with the description of the cut locus as the Schubert variety $\\mathrm{cut}(L)=\\{E:\\theta_k(L,E)=\\pi/2\\}$, stratified by the number of principal angles equal to $\\pi/2$. At a point in the $j$-th stratum the subdifferential is a convex set linearly isomorphic to the convex hull of the orthogonal group $O(j)$, of dimension $j^2$. Finiteness of critical points away from the cut locus comes from the normal exponential map restricted to the normal bundle $NX$, whose generic fibers are discrete by Sard's theorem; o-minimal definability of the principal-angle functions then makes the fiber count uniformly bounded. The Pfaffian bound for hypersurfaces encodes the singular-value/exponential description of the distance in a Pfaffian chain and controls its Betti numbers by known estimates.","core_discovery":"The central discovery is that the nearest-point problem on a Grassmannian, despite the non-smoothness and non-semialgebraicity of the metric, has a well-defined finite complexity in the generic case. Away from the cut locus, critical points of the restricted distance function are exactly points where the unique length-minimizing geodesic to $L$ is normal to $X$; Sard's theorem applied to the normal exponential map makes this set discrete, and o-minimal definability upgrades discreteness to a uniform bound $C_X$. For simple Schubert varieties the optimization is exactly solvable: the Eckart–Young construction produces critical points, the unique global minimizer spans the first $s$ principal vectors of the fixed plane $W$ together with the remaining principal vectors of $L$, and the global maximizers form a Grassmannian of dimension $(k-s)(n-k-s)$ that lies entirely in the cut locus.","pith_inferences":["Our inference: the same subdifferential and o-minimal machinery should define an analogous complexity for partial flag manifolds, replacing principal angles by iterated flag angles.","Our inference: the existential bound in Theorem 42 and the huge Pfaffian constants suggest that the practical value of GDC lies in qualitative finiteness, not in the size of the bound; computing exact GDC for a concrete variety will likely need a geometric rather than Pfaffian route.","Our inference: for small examples such as $\\Omega_s\\subset G(3,6)$, numerical sampling of generic $L$ and counting critical points outside the cut locus could test whether the Eckart–Young points are the only critical points there, which the paper does not fully settle for global counts."],"forward_implications":["For any smooth subanalytic submanifold $X\\subset G(k,n)$, the number of candidate nearest points that must be checked is bounded by a constant independent of the query plane $L$.","For smooth hypersurfaces of degree $d$, $\\mathrm{GDC}(X)\\le c_1(k,n)d^{c_2(k,n)}$, so the complexity grows at most polynomially in the degree with constants depending only on $k,n$.","For simple Schubert varieties $\\Omega_s$, the Eckart–Young construction gives $\\binom{k}{s}$ distinct critical points away from the cut locus, so $\\mathrm{GDC}(\\Omega_s)\\ge\\binom{k}{s}$.","Generic minimizers never lie on the cut locus, so the non-differentiable part of the distance function can be ignored when solving the nearest-point problem.","The global maximizers of distance to $\\Omega_s$ form a Grassmannian of dimension $(k-s)(n-k-s)$, all lying on the cut locus, which explains why the cut locus must be excluded from the definition of GDC."],"supporting_citations":[{"why":"defines the Euclidean Distance Degree that GDC generalizes from the algebraic to the o-minimal setting","marker":"[5]"},{"why":"Eckart–Young theorem supplies the critical points for low-rank matrix approximation that become critical points for Schubert varieties","marker":"[6]"},{"why":"Clarke's subdifferential and critical point theory for Lipschitz functions is the framework for defining critical points of the distance","marker":"[3]"},{"why":"identifies the cut locus of the Grassmannian distance as planes with a principal angle equal to pi/2","marker":"[20]"},{"why":"o-minimal cell decomposition and uniform finiteness results underpin the uniform bound in Theorem 42","marker":"[19]"},{"why":"Pfaffian Betti-number estimates give the explicit degree bound for algebraic hypersurfaces","marker":"[9]"},{"why":"principal angles between subspaces and their Schubert-variety inequalities are used to prove the global minimum and maximum results","marker":"[21]"},{"why":"singular value decomposition computes principal angles, connecting the Riemannian distance to Frobenius norm data","marker":"[2]"}],"fun_headline_variants":["Finite check: Grassmann nearest point complexity","Non-smooth distance, finite critical points on Grassmannian","Eckart–Young extends to Schubert varieties on Grassmannians","Grassmannian distance: finite complexity despite non-smoothness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The uniform bound depends on the unproved assertion that the set of pairs $(S,L)$ with $S$ critical for $\\delta_L|_X$ is definable in the paper's o-minimal structure; if that tameness fails, the generic finiteness of GDC could collapse.","fun_headline_variants_meta":{"raw":{"variants":["Finite check: Grassmann nearest point complexity","Non-smooth distance, finite critical points on Grassmannian","Eckart–Young extends to Schubert varieties on Grassmannians","Grassmannian distance: finite complexity despite non-smoothness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1401,"prompt_tokens":855,"completion_tokens":546,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":475}},"tokens_in":471,"tokens_out":546,"duration_ms":5623,"temperature":1.0,"reasoning_tokens":475,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:56:34.257721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a concrete smooth subanalytic $X$, write the critical-point relation explicitly and check whether it is definable in an o-minimal structure; a single $X$ for which this relation is not definable, or a sequence of generic $L_m$ with an unbounded number of critical points outside $\\mathrm{cut}(L_m)$, would refute Theorem 42. A cheaper numerical check for a Schubert variety like $\\Omega_s\\subset G(3,6)$ is to count, for random $L$, all critical points of $\\delta_L|_{\\Omega_s}$ outside the cut locus and see whether the count exceeds $\\binom{k}{s}$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the Euclidean Distance Degree that GDC generalizes from the algebraic to the o-minimal setting"},{"cited_title":"The approximation of one matr ix by another of lower rank","cited_arxiv_id":null,"evidence_quote":"Eckart–Young theorem supplies the critical points for low-rank matrix approximation that become critical points for Schubert varieties"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Clarke's subdifferential and critical point theory for Lipschitz functions is the framework for defining critical points of the distance"},{"cited_title":"Differential geometry of Grassmann man ifolds","cited_arxiv_id":null,"evidence_quote":"identifies the cut locus of the Grassmannian distance as planes with a principal angle equal to pi/2"},{"cited_title":"Complexity of co mputations with Pfaffian and Noetherian functions","cited_arxiv_id":null,"evidence_quote":"Pfaffian Betti-number estimates give the explicit degree bound for algebraic hypersurfaces"},{"cited_title":"Schubert varieties and distance s between subspaces of different dimensions","cited_arxiv_id":null,"evidence_quote":"principal angles between subspaces and their Schubert-variety inequalities are used to prove the global minimum and maximum results"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"singular value decomposition computes principal angles, connecting the Riemannian distance to Frobenius norm data"}],"review_version":1}