{"id":"4b635209-3a29-49e0-a993-c1d96fe13151","arxiv_id":"1908.08742","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In a finite-dimensional space with a smooth strictly convex norm, the distance to any convex body is differentiable outside the body, with norm gradient equal to the normalized outer normal of the metric projection.","lead":"The paper builds convex analysis in finite-dimensional normed spaces using a Legendre transform adapted to the norm instead of the standard inner product, and shows that distance functions to convex bodies are differentiable outside the body whenever the norm is smooth and strictly convex. This gives a unified picture linking norm gradients, subgradients, normal cones, Birkhoff orthogonality, and metric projections in Minkowski spaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Theorem 5.1 needs a corrected Proposition 3.5 and a proof of parallel-body smoothness; the key orthogonality step is not justified as written.","rationale":"I read the paper's central argument in good faith. The abstract and Section 5 promise that, for a smooth and strictly convex norm, the distance function to any convex body is differentiable outside the body, with norm gradient equal to the normalized metric-projection direction. The overall strategy is coherent: build a norm-adapted subdifferential via the Legendre transform, show that differentiability of a convex function is equivalent to a singleton norm subdifferential (Corollary 4.2), and then prove that the norm subdifferential of d_K at x is exactly {η_K(x)}. The algebra in that last step is sound once the geometric input is granted. The reader's conditional verdict identifies the smooth-and-strictly-convex assumption as the weakest point; I agree that it is essential, but the more immediate, concrete weakness lies inside the proof of Theorem 5.1. The theorem needs the unique supporting hyperplane of the parallel body K+cB at x, and it needs Proposition 3.5 to connect that hyperplane to η_K(x). The uniqueness is merely cited, and Proposition 3.5's proof contains a translation error: it concludes that h+δu supports K+δB at x even though that hyperplane does not pass through x. The statement of Proposition 3.5 is plausibly correct and fixable, and I do not see evidence that Theorem 5.1 is false. However, a reader cannot currently verify the central step from the written proof. This does not change the verdict from the reader's CONDITIONAL, so I recommend UNCHANGED, but with the request that Proposition 3.5 be rewritten with the correct supporting hyperplane and that the smoothness of K+cB be either proved or cited more precisely. My concern is about correctness risk in the proof, not about novelty or external consensus.","tokens_in":23428,"tokens_out":36775,"duration_ms":386425,"concrete_test":"Re-derive Proposition 3.5 with the correct affine support plane and check it in a concrete case. Take K=[-1,1]^2, B the Euclidean unit disk, δ=1, z=(1,1), u=(1,1)/√2, and x=z+u. Compute the supporting line of K+B at x; it is the line through x perpendicular to u, i.e. h0+x, not h0+δu. Then verify whether the manuscript's containment K+δB⊆(h+δu)^- implies that h+δu supports K+B at x; it does not, since that line does not pass through x. If Proposition 3.5 cannot be reproved with the corrected hyperplane, then Theorem 5.1's orthogonality step is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central proof of Theorem 5.1 reduces differentiability of d_K outside K to showing that the norm subdifferential at x is contained in span{η_K(x)}. This relies on two geometric facts that are not adequately established in the manuscript. First, the paper assumes without proof that the parallel body K+cB has a unique supporting hyperplane at x, citing [18]. This uniqueness is load-bearing: if K+cB had several supporting hyperplanes at x, Theorem 4.4 would only show that a subgradient is orthogonal to one of them, and the conclusion ∂d_K(x)⊆span{η_K(x)} would not follow. Second, Proposition 3.5, which asserts that η_K(x) is a Birkhoff normal vector of K+cB at x, is proved incorrectly as written. The proof translates the hyperplane h to pass through x, then claims K+δB⊆(h+δu)^- and concludes that 'h+δu supports K+δB at x'. But h+δu does not pass through x (since x=z+δu and z∉h in general), so it cannot be the supporting hyperplane at x. The correct support plane is h translated to x (equivalently h0+x), and the containment should be K+δB⊆h^- with h through x. As written, the key step 'v is a multiple of η_K(x)' is not justified. The proposition's statement is likely true, but the proof gap is real and sits directly on the path to Theorem 5.1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a convex-analytic framework for finite-dimensional normed spaces in which the usual identification of the space with its dual is replaced by the Legendre transform associated with the norm. On this basis it defines norm gradients and norm sub-gradients, relates them to Birkhoff orthogonality, and proves a main theorem (Theorem 5.1): if the ambient norm is smooth and strictly convex and K is a convex body, then the distance function d_K is differentiable on R^n\\K, and its norm gradient is the normalized outer normal direction from the metric projection, ∇d_K(x)=η_K(x). The paper also proves auxiliary results on uniqueness and continuity of metric projections, continuity and self-duality of the Legendre transform, a norm-subgradient characterization of differentiability, and a description of the boundary subdifferential of d_K as the Birkhoff normal cone.","tokens_in":23634,"tokens_out":23580,"duration_ms":219751,"significance":"If the proof is completed, Theorem 5.1 settles a natural question: under only smoothness and strict convexity of the norm, the distance to an arbitrary convex body is differentiable outside the body, with an explicit formula for the gradient. The Legendre-transform framework is elegant and gives a genuinely norm-dependent replacement for classical Euclidean gradient and subgradient theory, with connections to Birkhoff orthogonality that appear to be new. The paper is largely self-contained, uses no fitted parameters or data, and makes a concrete, falsifiable prediction about the norm gradient. Several auxiliary results, such as the self-duality of the Legendre transform and the norm-cyclic-monotonicity version of Rockafellar's theorem, are of independent interest. The main caveat is that the proof of the central theorem currently depends on two geometric facts that are not adequately established in the manuscript.","major_comments":[{"comment":"The proof of Proposition 3.5 is not valid as written. After the sentence 'we assume that h is translated to pass through x', the subsequent containment K+δB ⊆ (h+δu_K(z))^- treats h as the supporting hyperplane at z rather than at x; with h translated through x, the hyperplane h+δu_K(z) does not pass through x=z+δu_K(z) and therefore cannot be the supporting hyperplane asserted. This is not a cosmetic issue, because Theorem 5.1 uses Proposition 3.5 to conclude that an arbitrary norm subgradient v is a multiple of η_K(x). The statement is repairable: take h to be the supporting hyperplane of K at z with u_K(z) ⊣_B h, and let φ=L(u_K(z)); then for y∈K and b∈B one has φ(y)+δφ(b) ≤ δ, showing that the translate h+δu_K(z) supports K+δB at x. The manuscript should be corrected accordingly.","section":"§3, Proposition 3.5, and its use in §5, Theorem 5.1"},{"comment":"Theorem 5.1 relies on the uniqueness of the supporting hyperplane of K+cB at x: if K+cB had several supporting hyperplanes at x, Theorem 4.4 would only give that a subgradient is orthogonal to one of them, and the conclusion ∂d_K(x)⊆span{η_K(x)} would not follow. The manuscript asserts smoothness of K+δB with only a citation to [18] and gives no proof or precise statement. Since this is load-bearing, please add a self-contained proof, for instance using normal cones: for a boundary point x=a+b of K+B with b∈∂B, the normal cone N_{K+B}(x) is contained in N_B(b), and smoothness of B makes N_B(b) a ray, so uniqueness of the supporting hyperplane follows. Alternatively, quote the exact theorem from [18].","section":"§3, parallel sets, and §5, Theorem 5.1"}],"minor_comments":[{"comment":"The proposition states an 'if and only if', but the proof establishes only the direction 'non-unique metric projection implies parallel boundary segments'; the converse is asserted without proof. The converse is not used in the proof of Theorem 5.1, but the statement should either be proved in full or explicitly weakened.","section":"§3, Proposition 3.2"},{"comment":"In the displayed formula for f'_+(x,-u), the symbol L(w) should be L(v); the variable w is undefined in that display.","section":"§4, Corollary 4.2"},{"comment":"The proof uses the assertion that Birkhoff orthogonality is a continuous relation without proof. This follows in one line from the definition by passing to the limit in ||x_n+t z_n|| ≥ ||x_n||, and the manuscript should include that observation.","section":"§4, Proposition 4.2"},{"comment":"The notation h is used sometimes for a hyperplane through the origin and sometimes for a translate (through z, through x, or shifted by δu). Distinguishing the linear subspace from its affine translates would prevent the kind of confusion that occurs in the proof of Proposition 3.5.","section":"Throughout Sections 3–5"}],"recommendation":"major_revision","confidential_remarks":"The main theorem is likely correct after the indicated repairs to Proposition 3.5 and the smoothness assertion for parallel bodies; these are local fixes rather than fundamental flaws. The introduction should also sharpen the statement of novelty relative to the existing literature on distance functions in Banach spaces (Fitzpatrick, Giles, Zajíček), and the authors should verify that [18] indeed contains the exact parallel-body smoothness statement they need."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nQuick take: Theorem 5.1 is true, and the norm-gradient/subgradient framework is a useful way to organize convex analysis in Minkowski spaces. The printed proof has a fixable translation slip in Proposition 3.5; don't let that drive the verdict.\n\nWhat is actually here: the paper builds convex analysis on a smooth strictly convex norm using the Legendre transform of the norm. That gives norm gradients, norm subgradients, and a clean link to Birkhoff orthogonality. The payoff is the differentiability theorem: for a convex body K in a smooth strictly convex normed space, the distance function d_K is differentiable on R^n\\K and its norm gradient is the normalized metric projection direction. The proof is coherent and the C^1 corollary follows. I also liked Theorem 4.6 and Corollary 4.4, which recover differentiability of the norm and identify the Legendre transform with the norm derivative; those are clean results.\n\nWhere it needs work: Proposition 3.5, as written, is wrong. The proof translates the supporting hyperplane to pass through x and then says h+δu supports K+δB at x. That is a translation error: if h has been translated to x, h+δu doesn't pass through x; if h is the original support at z, then h+δu is the right plane and the containment can be proved in one line with the Legendre functional. So the proposition is true, but the proof needs a corrected write-up. The smoothness of K+δB is cited from [18] rather than proved; if the citation is solid that is acceptable, but I'd ask for a precise statement. Proposition 3.2 states an iff but only proves the forward direction. It is unused in the strict-convexity regime, so this is a minor completeness issue. Proposition 4.2's \"Birkhoff orthogonality is continuous\" is actually immediate from continuity of the norm, so I don't count that as a gap.\n\nThe bigger referee request is novelty support. The authors claim the differentiability question under smooth plus strictly convex norm was not explicitly answered, but they don't compare with the known subdifferential characterizations of distance functions (Fitzpatrick, Giles, Zajicek). It may well be that the norm-gradient apparatus is the new part and Theorem 5.1 is a corollary of known subdifferential formulas; either way, a few paragraphs of comparison would settle it.\n\nWho it is for: convex analysts and Minkowski geometers who want a norm-adapted calculus; they will get a working toolkit and a clean theorem. It deserves serious refereeing, not a desk reject. I'd accept and ask for the Prop 3.5 fix and the comparisons.","headline":"The main theorem is sound and the norm-gradient framework is genuinely useful; the proof has a fixable slip in Proposition 3.5 and needs a comparison with known distance-function subdifferential results.","tokens_in":24231,"tokens_out":14910,"would_cite":true,"duration_ms":163657,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["41A50","41A65","46B20","46G05","52A20","52A21","53C23","58C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"With a smooth, strictly convex norm, distance to a convex body is differentiable outside the body.","keywords":["Birkhoff orthogonality","convex body","distance functions","differentiability","Legendre transform","sub-gradient","metric projection","normed spaces"],"falsifier":"Take the plane with an $\\ell^p$ norm for $1<p<\\infty$, which is smooth and strictly convex, and let $K$ be a square. For a point $x$ whose metric projection is a corner $v$, the theorem predicts that $d_K$ is differentiable at $x$ and that every one-sided directional derivative equals $L((x-v)/\\|x-v\\|)\\cdot u$. Computing these derivatives directly for $u$ along the two edge directions from the corner, checking that the derivatives in opposite directions are negatives, would settle the claim; any mismatch is a counterexample.","tokens_in":23142,"feed_emoji":"📐","tokens_out":13299,"duration_ms":124979,"temperature":0.7,"pith_summary":"This paper establishes that in a finite-dimensional normed space, if the norm's unit ball is smooth and strictly convex, then the distance function to any convex body is differentiable at every point outside the body, and its gradient is the unit vector pointing from the nearest point in the body toward the point. This answers a question the authors say was open: regularity of the distance function follows from the geometry of the ambient norm alone, with no regularity assumed for the body. The route is a new convex-analysis toolkit for normed spaces: instead of identifying vectors with linear functionals via an inner product, the paper uses the Legendre transform built from the norm, and develops norm versions of gradients and sub-gradients. These tools also give an explicit characterization of the distance function's sub-differential on the boundary and a clean link between metric projections and the norm's own orthogonality relation. A reader should care because distance functions and metric projections are basic objects in approximation, optimization, and geometric analysis, and the paper pins down exactly when their classical Euclidean regularity survives.","feed_headline":"Distances to convex bodies stay differentiable in smooth normed spaces","feed_subtitle":"The gradient is the unit vector from the body's nearest point, so no inner product is needed.","key_machinery":"The engine is the Legendre transform of the norm: for each nonzero vector $x$ it is the unique linear functional whose kernel is the hyperplane Birkhoff right-orthogonal to $x$ (the supporting hyperplane of the unit ball in direction $x$), normalized by $L(x)\\cdot x=\\|x\\|^2$. This map is a norm-preserving bijection between the space and its dual, and it replaces the standard inner product in every convex-analysis step. On top of it the paper defines the norm gradient and norm sub-gradient of a convex function via the inequality $f(y)-f(x) \\ge L(\\nabla f(x))\\cdot(y-x)$, and proves the norm analogues of the classical facts: the sub-differential is nonempty, the one-sided derivative is the maximum of $L(w)\\cdot u$ over sub-gradients $w$, and a convex function is differentiable exactly when its norm sub-differential is a singleton. That last criterion, applied to the distance function with its parallel-body sub-level sets, is what produces the main theorem.","core_discovery":"The central discovery is Theorem 5.1: for a convex body $K$ in $(\\mathbb{R}^n, \\|\\cdot\\|)$ with a smooth and strictly convex norm, the distance function $d_K(x)=\\operatorname{dist}(x,K)$ is differentiable on $\\mathbb{R}^n\\setminus K$ and its norm gradient is $\\nabla d_K(x)=\\eta_K(x)=(x-p_K(x))/\\|x-p_K(x)\\|$, the normalized vector from the metric projection. Consequently $d_K$ is of class $C^1$ outside $K$, because both the distance function and the metric projection are continuous. On the boundary the picture is different: $d_K$ is not differentiable, and its norm sub-differential equals the Birkhoff normal cone of $K$ at the point, the cone of outward vectors that are left-orthogonal (in the norm's sense) to a supporting hyperplane. The proof identifies the sub-level sets $\\{d_K \\le c\\}$ with the parallel bodies $K+cB$, uses smoothness of the unit ball to get a unique supporting hyperplane there, and shows the norm sub-differential is a singleton; a general characterization from the paper then converts that singleton into differentiability.","pith_inferences":["The paper does not pursue infinite-dimensional versions; a natural next step would be to ask whether reflexivity plus smooth strict convexity suffices in Banach spaces, though the compactness arguments used here would need replacement.","Since $\\nabla d_K$ is the unit vector pointing from the nearest point to $x$, the identity suggests that metric projections can be recovered by integrating the gradient flow of the distance function, which could give practical algorithms for nearest-point queries under arbitrary norms.","The geometric definition of the Legendre transform, which avoids assuming differentiability of the norm, connects directly to semi-inner products; this may give a coordinate-free gradient calculus in Finsler geometry.","A testable boundary of the result: allow the body to be nonconvex while keeping the norm smooth and strictly convex; the proof leans on convexity at nearly every step, so a counterexample there would show the phenomenon is really about convexity, not just the norm."],"forward_implications":["In any finite-dimensional normed space whose unit ball is smooth and strictly convex, the distance function to every convex body is $C^1$ outside the body; no smoothness or strict convexity of the body itself is needed.","The metric projection, when it is unique (guaranteed by strict convexity), is continuous, and the formula $\\nabla d_K=\\eta_K$ makes the gradient an explicit, computable function of the nearest point.","At boundary points, the distance function is never differentiable, and its norm sub-differential is exactly the Birkhoff normal cone of the body.","The same Legendre-transform machinery shows that a smooth norm is itself $C^1$ away from the origin, with norm gradient $x/\\|x\\|$, recovering a known regularity fact from the new tools.","The norm subgradient calculus reproduces the standard Euclidean subdifferential results (max formula, cyclic monotonicity, convexity detection) with the norm's Legendre transform in place of the inner product."],"supporting_citations":[{"why":"supplies the convex-analysis foundation: one-sided derivatives, epigraph support, the separating-hyperplane theorem, and the Euclidean subdifferential max formula used as a template.","marker":"[28]"},{"why":"supplies the Birkhoff orthogonality facts used throughout: smoothness is equivalent to unique right-orthogonal hyperplanes and strict convexity to unique left-orthogonal directions.","marker":"[2]"},{"why":"supplies the Minkowski-space geometry facts, including the equivalence between strict convexity of the norm and strict triangle inequality, used for uniqueness of metric projections.","marker":"[20]"},{"why":"gives the smoothness of parallel bodies $K+\\delta B$ when the unit ball is smooth, which identifies the unique supporting hyperplane of the sub-level sets of the distance function.","marker":"[18]"},{"why":"provides the Euclidean subdifferential theorem the paper generalizes by replacing the inner product with the Legendre transform (the max formula for one-sided derivatives).","marker":"[6]"},{"why":"supplies the Birkhoff normal cone notion used to identify the norm sub-differential of the distance function on the boundary.","marker":"[25]"}],"fun_headline_variants":["Smooth norms make convex-body distances differentiable","Smooth unit ball gives C^1 distance to convex bodies","Distance to convex bodies is C^1 in smooth normed spaces","No inner product needed: smooth norms give metric projection gradient"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ambient norm's unit ball is smooth and strictly convex: every boundary point has exactly one supporting hyperplane and the boundary contains no line segment, which together make the metric projection unique and the Legendre transform a bijection.","fun_headline_variants_meta":{"raw":{"variants":["Smooth norms make convex-body distances differentiable","Smooth unit ball gives C^1 distance to convex bodies","Distance to convex bodies is C^1 in smooth normed spaces","No inner product needed: smooth norms give metric projection gradient"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001021,"raw_usage":{"total_tokens":4259,"prompt_tokens":851,"completion_tokens":3408,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":3341}},"tokens_in":467,"tokens_out":3408,"duration_ms":22370,"temperature":1.0,"reasoning_tokens":3341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:31:49.359378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the plane with an $\\ell^p$ norm for $1<p<\\infty$, which is smooth and strictly convex, and let $K$ be a square. For a point $x$ whose metric projection is a corner $v$, the theorem predicts that $d_K$ is differentiable at $x$ and that every one-sided directional derivative equals $L((x-v)/\\|x-v\\|)\\cdot u$. Computing these derivatives directly for $u$ along the two edge directions from the corner, checking that the derivatives in opposite directions are negatives, would settle the claim; any mismatch is a counterexample.","supporting_citations":[{"cited_title":"Schneider, Convex Bodies: The Brunn-Minkowski Theory","cited_arxiv_id":null,"evidence_quote":"supplies the convex-analysis foundation: one-sided derivatives, epigraph support, the separating-hyperplane theorem, and the Euclidean subdifferential max formula used as a template."},{"cited_title":"Alonso, H","cited_arxiv_id":null,"evidence_quote":"supplies the Birkhoff orthogonality facts used throughout: smoothness is equivalent to unique right-orthogonal hyperplanes and strict convexity to unique left-orthogonal directions."},{"cited_title":"Martini, K","cited_arxiv_id":null,"evidence_quote":"supplies the Minkowski-space geometry facts, including the equivalence between strict convexity of the norm and strict triangle inequality, used for uniqueness of metric projections."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the smoothness of parallel bodies $K+\\delta B$ when the unit ball is smooth, which identifies the unique supporting hyperplane of the sub-level sets of the distance function."},{"cited_title":"Borwein and A","cited_arxiv_id":null,"evidence_quote":"provides the Euclidean subdifferential theorem the paper generalizes by replacing the inner product with the Legendre transform (the max formula for one-sided derivatives)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Birkhoff normal cone notion used to identify the norm sub-differential of the distance function on the boundary."}],"review_version":1}