{"id":"0fdaf0a2-1e62-4583-9741-bde72a18f0f5","arxiv_id":"2501.02978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A material's energy law can be recovered uniquely from boundary energy measurements whenever the loading set separates candidate laws, and maxout neural networks provide provably convergent approximations.","lead":"A new variational framework identifies a material's hidden energy law from only the total energy measured under prescribed boundary displacements, with no full-field imaging. It proves existence, gives a separating-data condition for exact recovery, and shows maxout neural networks converge to the true law.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Maxout density theorem in §4.2 is false as stated: the nonnegativity constraint on neurons prevents approximation of u(ξ)=ξ^2, so Cor 4.6's convergence claim collapses.","rationale":"The central claim has two pillars: variational identifiability and maxout approximation. The reader's concerns about Cor 3.12 compactness and exact data are real but largely repairable: sublevel sets of J are bounded because any unbounded sequence in U is uniformly large by Lipschitzness and nonnegativity, and then J→∞; the Γ-convergence constraint issue can also be fixed by adding vanishing constants to the recovery sequence. The density failure is different: it is a concrete counterexample, not a gap. The nonnegativity of f_i in (66) is not a technical convenience; it changes the approximating class. In 1D, the only nonnegative affine functions pointwise below ξ^2 are identically zero, and any nonnegative affine functions whose max is within ε of ξ^2 must have small intercepts but then a neuron reaching near 1 forces the max to be at least about δ at small δ, far above δ^2. Hence Uh is not dense, Prop 4.5 is false, and Cor 4.6 does not follow. This directly undermines the paper's advertised convergent approximation by maxout networks. The paper could be repaired by redefining Uh to allow arbitrary affine polyaffine neurons (possibly negative) plus a zero neuron, yielding max(hξ,0) in the construction, but this is a non-minor revision of the central approximation theorem. Therefore the current version should not be accepted as stated.","tokens_in":20090,"tokens_out":35018,"duration_ms":352831,"concrete_test":"Test Prop 4.5 in the scalar case K=[0,1], ℓ=2, u(ξ)=ξ^2. For any h∈U_N as in (66), let ε=∥h−u∥∞. From h(0)≤ε and h(1)≥1−ε, a neuron attaining h(1) satisfies h(1/2)≥(1−ε)/2, hence ε≥h(1/2)−u(1/2)≥(1−ε)/2−1/4, so ε≥1/6. Thus d(U_N,u)≥1/6 for every N, disproving density. For a direct check of the proof, apply the Prop 4.5 tangent construction to g(s)=s^2 at s=1/2: the tangent line 2s−1/4 is negative on [0,1/8], so hξ∘Minors∉U; replacing hξ by max(hξ,0) is not allowed by (66) because that function is not polyaffine.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing defect is in Prop 4.5 and the definition of Uh in (66), not in the noiseless-data idealization. U in A2 consists of nonnegative energy densities, and (66) requires each maxout neuron f_i to belong to U. In one dimension with K=[0,1], take u(ξ)=ξ^2 in U (any ℓ≥2). If h∈Uh and ∥h−u∥∞≤ε, then h(0)<ε forces every intercept b_i<ε, while h(1)>1−ε forces some neuron j to have a_j+b_j≥1−ε. But then h(δ)≥f_j(δ)=(a_j+b_j)δ+b_j(1−δ)≥(1−ε)δ for every δ∈(0,1). For δ=1/2 this gives h(1/2)−u(1/2)≥(1−ε)/2−1/4, so the sup error is bounded below by about 1/4−ε/2. Hence no sequence in Uh converges uniformly to u; the density property fails. The paper's proof uses tangent affine minorants hξ which can be negative, and asserts hξ∘Minors∈U, which is false because U requires nonnegative values. This is not a small proof gap: with f_i∈U the class Uh is simply too small to be dense. Consequently Cor 4.6 and the abstract's convergence claim are false as stated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a variational optimal-control formulation for identifying the strain-energy density of a hyperelastic material from measurements of total elastic energy under prescribed boundary displacements. The control space U consists of polyconvex, ℓ-Lipschitz, nonnegative energy densities, and the cost J(u) is the supremum over the data set M of the positive part of the difference between the trial energy E(u;g) and the measured energy E0(g). The authors prove lower semicontinuity and, in Corollary 3.12, existence of minimizers of J; they define a separating property of M and show in Proposition 3.15 and Corollary 3.16 that, when M is separating and E0 is generated by some w∈U, the unique minimizer equals w. They then propose approximation by 'maxout' neural networks (maxima of finitely many polyaffine functions) and claim in Proposition 4.5 that such networks are dense in U, leading in Corollary 4.6 to convergent approximations. The paper also contains one- and two-bar examples illustrating the minimax structure.","tokens_in":20425,"tokens_out":8623,"duration_ms":83128,"significance":"If the main results held, the paper would make a useful contribution to inverse problems in finite elasticity: it reduces the difficult question of identifying a full material law to a separation condition on scalar boundary-energy data, and it provides a natural lattice-theoretic connection to maxout networks. The identifiability theorem (Proposition 3.15) is honest and does not fit free parameters to the target conclusion. The Γ-convergence arguments for the state functional (Propositions 3.3 and 3.6) and the continuity of the energy map (Lemma 3.10) are standard and clearly presented. However, the paper's two central claims are compromised: the proof of existence in Corollary 3.12 relies on an unproved compactness statement, and the density theorem for maxout networks in Proposition 4.5 is false as stated because the admissible class U contains only nonnegative functions while the proof uses tangent affine minorants that may take negative values. Since the abstract explicitly advertises the 'requisite density property' and the convergent approximation by maxout networks, these are load-bearing defects.","major_comments":[{"comment":"The density claim is false as stated. The set U defined in A2 consists of nonnegative energy densities (g: R^{τ(n)} → [0,∞)), and each neuron f_i in (66) is required to belong to U. The proof constructs tangent affine minorants h_ξ and asserts that u_ξ = h_ξ∘Minors ∈ U, but h_ξ can be negative even when g≥0. A concrete counterexample in one dimension with K=[0,1]: take u(ξ)=ξ^2, which is in U for any ℓ≥2. If h∈U_h and ∥h−u∥∞<ε, then h(0)<ε forces every intercept b_i<ε, while h(1)>1−ε forces some neuron j to satisfy a_j+b_j≥1−ε. For δ=1/2 this gives h(1/2)≥(1−ε)/2, so the sup error is at least 1/4−ε/2, which cannot be made arbitrarily small. Hence no sequence in U_h converges uniformly to u, and Proposition 4.5, Corollary 4.6, and the abstract's convergence claim are false as stated.","section":"§4.2, Proposition 4.5 and equation (66)"},{"comment":"The existence proof is incomplete. Corollary 3.12 invokes 'compactness of U proven in Prop. 3.5', but Proposition 3.5(ii) only proves compactness modulo constants: it yields a subsequence u_{j_k}−u_{j_k}(0) converging uniformly, not a convergent subsequence of u_{j_k} itself. The cost J is not invariant under adding constants (for c≥0, J(u+c)=J(u)+c|Ω|; for sufficiently negative c the constraint in (10) fails and J=+∞), so a minimizing sequence could in principle escape to infinity in the constant direction. One must prove a priori bounds on u_j(0) for minimizing sequences, which is not done. This gap is likely repairable, but the argument as written does not establish existence.","section":"§3.3, Corollary 3.12 and Proposition 3.5"},{"comment":"The uniqueness and identifiability theorem is stated under the exact-data condition E0(g)=E(w;g) for some w∈U. The paper correctly notes in Remark 3.17 that for nonquasiconvex w only the relaxed energy can be identified, but it does not discuss the effect of measurement noise or the case w∉U. These are genuine limitations for applications; they do not invalidate the mathematical statement, but the paper should frame the identifiability result as explicitly idealizing noiseless data and an admissible ground truth.","section":"§3.4, Proposition 3.15 and Remark 3.17"}],"minor_comments":[{"comment":"The condition (30) uses the norm ∥·∥_{Lip(K)}, but K is the compact set of deformation gradients, not the domain of the displacement maps y; presumably a norm on Y such as ∥·∥_{W^{1,∞}} is intended. Please clarify.","section":"Proposition 3.8"},{"comment":"The displayed formula for J(u_N) contains an unbalanced parenthesis and appears to have the opposite sign from the cost defined in (9); as written it is the negative of the maximum discrepancy. Please check and correct.","section":"Equation (76)"},{"comment":"The separating property for the two-bar example is proved only under the additional local Hölder condition (59) on w. This extra assumption should be stated more prominently, since the general framework does not impose any such regularity beyond Lipschitz continuity.","section":"Example 3.5.2"},{"comment":"There are minor typographical issues, such as the spaced 'V ARIA TIONAL' in the title, and some references to [27] as 'in preparation' may be updated if available. These do not affect the mathematics.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper contains a clear and well-presented variational framework, and the identifiability result under exact data is a useful contribution. However, the false density theorem for maxout networks in §4.2 is a central advertised result; it cannot be repaired by minor edits within the manuscript's current scope, since the admissible class U_h defined in (66) is genuinely too small. The existence gap in Corollary 3.12 is also nontrivial, though probably fixable. Given that the abstract and conclusions rest on the convergent approximation claim, I cannot recommend acceptance in the present form. The authors may wish to consider reformulating the approximation class (e.g., using maxima of positive parts of affine functions) or restricting the setting so that the density statement becomes true."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead Conti and Ortiz on identifying material laws from boundary data. The core idea is good: set up material identification as an optimal control problem, with energy densities as controls and measured boundary energies as data, and reduce identifiability to a separating-data condition. The existence and identifiability results (Cor 3.12, Prop 3.15, Cor 3.16) are the honest part of the paper. Cor 3.12 as written is incomplete: they cite compactness of U from Prop 3.5, but Prop 3.5 only proves compactness modulo constants, and the cost J is not invariant under adding constants. That gap looks repairable by using the sublevel set of J to bound the constants, but it needs to be done.\n\nThe larger problem sits in Section 4.2. Prop 4.5, the density of the maxout spaces Uh, is false as stated. The definition (66) requires each neuron f_i to belong to U, and U consists of nonnegative energy densities. The proof constructs tangent minorants hξ∘Minors and claims they are in U, but those tangent functions can be negative on K, so they are not in U. The failure is not a proof slip; the theorem is wrong. For u(ξ)=ξ² on [0,1], any function in Uh is a max of nonnegative affine functions. If it is within ε of ξ² in sup norm, the intercepts must be < ε, yet some neuron must have slope near 1 to reach 1 at ξ=1; that neuron's value at ξ=1/2 is then about 1/2, giving sup error about 1/4. No sequence in Uh approximates ξ². So Prop 4.5, Cor 4.6, and the abstract's convergence claim collapse.\n\nOn the positive side, the paper is candid about its idealizations: it assumes exact noiseless boundary energy data, and it notes that for nonquasiconvex w only the relaxed energy is identifiable. The two-bar separability example needs the extra Hölder condition (59), which is a minor caveat. Citation practice looks fine.\n\nWho should read this? People working on variational inverse problems in mechanics will find the existence/identifiability framework worth engaging, and the density counterexample is a useful caution for anyone using maxout networks under nonnegativity constraints. The paper deserves a serious referee, but it needs major revision: repair the existence proof and either fix or substantially soften the maxout density claim. I would not pass it as is.","headline":"A worthwhile variational framework for material identification, but the maxout density theorem is false as stated and the existence proof needs a missing compactness argument.","tokens_in":20916,"tokens_out":6893,"would_cite":true,"duration_ms":62937,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49J45","74B20","49K20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that, when boundary-energy measurements are exact and the applied-displacement menu separates admissible densities, the unknown strain-energy density is the unique minimizer of a minimax cost, and maxout-network…","keywords":["material identification","finite elasticity","optimal control","minimax cost","polyconvexity","separating boundary data","maxout neural networks","Gamma-convergence"],"falsifier":"Test the injectivity of the energy-response map on $U$ for a proposed experimental menu. Concretely, in the two-bar example with unequal lengths and data interval $M$ containing $0$, look for a nonzero difference $w$ of two admissible densities solving $w(1+\\lambda\\beta) + \\lambda w(1+\\beta)=0$ for all $\\beta$ with $1+\\beta \\in K$; any such $w$ makes $E(w;\\delta)=0$ for all $\\delta \\in M$ and so $M$ is not separating. A concrete family is $w(1+\\beta)=\\beta\\,\\varphi(\\log \\beta)$ with $\\varphi(s+\\ln\\lambda)=-\\varphi(s)$, whose growth near $\\xi=1$ is only linear—exactly the case excluded by the paper's condition (59). If such a difference lies in $U-U$, uniqueness fails for that menu; if it cannot, condition (59) is shown to be necessary.","tokens_in":19845,"feed_emoji":"📐","tokens_out":14210,"duration_ms":125594,"temperature":0.7,"pith_summary":"The paper aims to make material identification a well-posed inverse problem: from only the total elastic energy of a specimen measured under prescribed boundary displacements, recover the unknown strain-energy density. It proposes an optimal-control formulation whose cost is the worst-case energy mismatch over the entire menu of applied displacements, proves that this cost has minimizers, and proves that if the menu is separating and the data come from an admissible density, the minimizer is exactly that density. It then shows that maxout neural networks—maxima of affine or polyaffine pieces—are dense in the admissible class, so the discrete minimizers converge weakly to the true law as the number of neurons grows. The point of caring is that a procedure normally treated as heuristic fitting becomes one with existence, uniqueness, and provable convergence of the numerical scheme.","feed_headline":"Boundary energy data alone identify the material law","feed_subtitle":"A minimax cost over prescribed displacements makes identification well-posed; maxout networks converge to the true energy density.","key_machinery":"The load-bearing object is the minimax cost functional $J(u) = \\sup_{g\\in M}(E(u;g)-E_0(g))$ with the majorization constraint $E(u;g) \\ge E_0(g)$, together with the separating property of the boundary-data set $M$: for any distinct admissible densities there must be some $g \\in M$ where their minimal energies differ. Existence rests on compactness of $U$ modulo constants and lower semicontinuity of $J$; uniqueness rests on the separating property; convergence rests on the lattice structure of $U$, namely that finite maxima of admissible densities remain admissible, which makes maxout networks (maxima of affine or polyaffine functions) a dense and convergent Galerkin family.","core_discovery":"Within finite elasticity, the energy-response map $u \\mapsto E(u;g) = \\inf_{y} \\int_\\Omega u(Dy)\\,dx$ under boundary displacement $g$ is continuous and concave in the trial density $u$. The cost $J(u) = \\sup_{g\\in M} (E(u;g) - E_0(g))$, with the constraint $E(u;g) \\ge E_0(g)$ enforced (otherwise $+\\infty$), is weak-star lower semicontinuous on the compact admissible class $U$ of polyconvex, $\\ell$-Lipschitz densities, and therefore attains its minimum. If the measured energies $E_0$ come from a true density $w \\in U$ and $M$ separates $U$ in the sense that any two distinct densities differ in energy for at least one applied displacement $g \\in M$, then the minimizer is unique and equals $w$. For approximation, the finite-dimensional subspaces built from maxima of polyaffine functions (maxout networks) are dense in $U$, the restricted costs $\\Gamma$-converge to the lower-semicontinuous envelope of $J$, and sequences of their minimizers converge weak-star up to subsequence to a minimizer of $J$.","pith_inferences":["If the measured energies carry noise or the true law lies outside the admissible class U, the hard constraint E(u;g) ≥ E0(g) can make the feasible set empty and the unique-recovery conclusion collapses; a penalized or statistical version of J would be needed for practical data.","The separating condition is the real content of identifiability: the paper gives a sufficient criterion (affine displacements forming a dense set are separating), but in richer specimens it remains the condition to verify for any proposed experimental design.","The two-bar example shows that separability can fail for non-quasiconvex or merely linearly growing differences near the reference state; a testable design rule is to choose specimen geometries and loadings that make the energy-response map injective on U, not merely well-conditioned.","The framework suggests an optimal experimental-design principle: choose the smallest finite menu Mh that is separating within the approximating class Uh, so that the number of tests is driven by identifiability rather than by ad hoc sampling."],"forward_implications":["If the boundary-displacement menu is separating and the data are exact, the energy density of the tested solid is variationally identified without full-field measurements such as DIC.","At every finite approximation level the identification problem becomes a finite-dimensional minimax (Chebyshev) program, so optimal testing programs require only finitely many applied displacements.","Maxout-network approximations do not require tuning the network topology, because iterated maxima collapse to a single layer.","Enlarging the experimental menu M increases the cost J and therefore makes the identification more constraining, giving a monotone design principle for experiments.","For non-quasiconvex ground-truth energies, only the relaxed (quasiconvexified) energy can be identified, so the method targets the effective material law rather than microscopic energy landscapes."],"supporting_citations":[{"why":"Supplies the existence theory in finite elasticity that motivates displacement control.","marker":"[30]"},{"why":"Provides the Γ-convergence tools used to prove convergence of the Galerkin/maxout approximations.","marker":"[36]"},{"why":"Tonelli's theorem is invoked for existence of minimizers of the state and control problems.","marker":"[37]"},{"why":"Names and motivates maxout neural networks as the approximation family.","marker":"[35]"},{"why":"Supports approximation of Lipschitz functions by maxima of affine functions, used in the density proof.","marker":"[34]"},{"why":"Supplies the definitions of polyconvexity and the minors representation of energy densities.","marker":"[39]"},{"why":"Establishes the equivalence of polyaffine and quasiaffine functions, grounding the maxout representation in the minors.","marker":"[40]"},{"why":"Provides the convex-programming/Kuhn-Tucker framework used in the optimal-test examples.","marker":"[42]"}],"fun_headline_variants":["Boundary energy alone pins down material law","Material law from boundary energy, no DIC","Maxout nets learn law from boundary energy","Energy measurements only give material law"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The uniqueness and identifiability claims assume the measured energies are exact noiseless values produced by a ground-truth density $w$ that lies inside the admissible class $U$, and that the boundary-displacement menu $M$ separates $U$; if any of those fails—noisy data, $w$ outside $U$, or a non-separating menu—the conclusion that the minimizer is $w$ no longer follows.","fun_headline_variants_meta":{"raw":{"variants":["Boundary energy alone pins down material law","Material law from boundary energy, no DIC","Maxout nets learn law from boundary energy","Energy measurements only give material law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001134,"raw_usage":{"total_tokens":4764,"prompt_tokens":1054,"completion_tokens":3710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":3656}},"tokens_in":670,"tokens_out":3710,"duration_ms":26763,"temperature":1.0,"reasoning_tokens":3656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:59:50.486700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Test the injectivity of the energy-response map on $U$ for a proposed experimental menu. Concretely, in the two-bar example with unequal lengths and data interval $M$ containing $0$, look for a nonzero difference $w$ of two admissible densities solving $w(1+\\lambda\\beta) + \\lambda w(1+\\beta)=0$ for all $\\beta$ with $1+\\beta \\in K$; any such $w$ makes $E(w;\\delta)=0$ for all $\\delta \\in M$ and so $M$ is not separating. A concrete family is $w(1+\\beta)=\\beta\\,\\varphi(\\log \\beta)$ with $\\varphi(s+\\ln\\lambda)=-\\varphi(s)$, whose growth near $\\xi=1$ is only linear—exactly the case excluded by the paper's condition (59). If such a difference lies in $U-U$, uniqueness fails for that menu; if it cannot, condition (59) is shown to be necessary.","supporting_citations":[{"cited_title":"Convexity conditions and existence theorems in nonlinear elasticity","cited_arxiv_id":null,"evidence_quote":"Supplies the existence theory in finite elasticity that motivates displacement control."},{"cited_title":"dal Maso","cited_arxiv_id":null,"evidence_quote":"Provides the Γ-convergence tools used to prove convergence of the Galerkin/maxout approximations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Tonelli's theorem is invoked for existence of minimizers of the state and control problems."},{"cited_title":"Goodfellow, D","cited_arxiv_id":null,"evidence_quote":"Names and motivates maxout neural networks as the approximation family."},{"cited_title":"Sorting out Lipschitz function approx- imation","cited_arxiv_id":null,"evidence_quote":"Supports approximation of Lipschitz functions by maxima of affine functions, used in the density proof."},{"cited_title":"Direct methods in the calculus of variations, volume 78 of Applied Mathematical Sciences","cited_arxiv_id":null,"evidence_quote":"Supplies the definitions of polyconvexity and the minors representation of energy densities."},{"cited_title":"Conti, G","cited_arxiv_id":null,"evidence_quote":"Establishes the equivalence of polyaffine and quasiaffine functions, grounding the maxout representation in the minors."},{"cited_title":"Tyrrell Rockafellar","cited_arxiv_id":null,"evidence_quote":"Provides the convex-programming/Kuhn-Tucker framework used in the optimal-test examples."}],"review_version":1}