{"id":"89164af8-6a5b-46eb-83e3-248ef79e612c","arxiv_id":"2501.11944","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Discontinuous Galerkin discrete minimisers are shown to Gamma-converge to minimisers of the quasiconvex envelope energy for a wide class of nonconvex variational problems.","lead":"This paper proves that discontinuous Galerkin methods converge to correct minimisers for a broad class of nonlinear elastic energies, including quasiconvex and relaxed nonconvex problems. The result matters because it places a flexible numerical tool on firm theoretical footing for materials simulations involving microstructure.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Penalty term (3.6) can be undefined for W satisfying (3.2); Remark 3.1's claimed equivalence with (3.8) fails, so Theorem 3.1 does not cover the stated scheme.","rationale":"The reader identified Lemma 2.2 as the weakest assumption, and it is true that the reconstruction estimate is load-bearing and only sketched. However, on inspection the proof of Lemma 2.2 appears to be a plausible extension of [28] via Lemma 2.1, and the apparent dimensional mismatch in the final inverse-inequality step actually yields a bound at least as strong as stated for all d≥2. The more concrete and decisive problem is the well-posedness of the penalty term. The paper itself flags the switch from (3.6) to (3.8) in Remark 3.1, but the asserted equivalence is not valid under (3.2)–(3.3) alone. A one-line counterexample W(ξ) = −M+(M+1)|ξ|^p shows the base of (3.6) can be negative, so the fractional power is undefined. This means the theorem statement and the proof analyze different functionals. The issue is correctable: one can add the assumption that W is bounded below by a nonnegative constant offset, or simply adopt (3.8) as the definition of the method throughout. Because the main Gamma-convergence strategy is otherwise coherent and the numerical experiments are consistent with the theory, the appropriate judgment remains CONDITIONAL rather than REJECT. The recommended change is to rerun the verification step described in concrete_test and, if it reproduces the failure, revise the penalty definition or the assumptions accordingly.","tokens_in":19673,"tokens_out":16004,"duration_ms":157113,"concrete_test":"Compute Pen(0) for W(ξ) = −M+(M+1)|ξ|^p, p=2, Ω=(0,1)^2, M=100, with u_h ≡ 0 on a single-element mesh. The base of (3.6) is 1+∫_Ω W(0) = 1−100 = −99, so (3.6) involves (−99)^{1/2} and is not real-valued. This directly contradicts the claim in Remark 3.1 that 1+|u_h|^p_{W^{1,p}(Ω,T_h)} ≲ 1+Σ_K∫_K W(∇u_h)+Σ_e h_e^{1-p}||Ju_hK||^p, since at u_h=0 the left side is 1 and the middle expression is −99. If the authors confirm this counterexample, Theorem 3.1 must be revised to assume 1+W ≥ 0 or to use penalty (3.8) as the actual scheme.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The discrete energy (3.5) is defined with the penalty (3.6), whose first factor is (1 + Σ_K ∫_K W(∇u_h) + Σ_e h_e^{1-p}||Ju_hK||^p_{L^p(e)})^{(p-1)/p}. Hypothesis (3.2) only implies W(ξ) ≥ c(−1+|ξ|^p), so 1+W(ξ) ≥ (1−c)+c|ξ|^p. For c>1, this quantity is negative at ξ=0. The paper's Remark 3.1 asserts that for the purposes of the proofs one may replace (3.6) by (3.8) because 1+|u_h|^p_{W^{1,p}(Ω,T_h)} is comparable to 1+Σ_K∫_K W(∇u_h)+Σ_e h_e^{1-p}||Ju_hK||^p. This comparison is false without an additional lower bound on W. A concrete admissible W is W(ξ) = −M+(M+1)|ξ|^p with M > 1; it satisfies (3.2)–(3.3), but at u_h ≡ 0 the base in (3.6) equals 1−M|Ω|, which is negative for |Ω|>1/M, making Pen(0) undefined in real arithmetic. Thus the functional minimized in Theorem 3.1 is not well-defined on the stated space for such W, while the proof genuinely uses (3.8), a different objective. This is a load-bearing gap: the central convergence result, as stated, applies to a scheme that may not be defined, and the proof covers a scheme that is not the one presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops a Gamma-convergence framework for discontinuous Galerkin discretizations of vector variational problems with quasiconvex and nonconvex energies. The two main theorems, Theorem 3.1 and Theorem 3.2, claim that sequences of discrete minimizers of two DG energy functionals, the method of [27] and the discrete-gradient method of [38], converge in L^p to minimizers of the relaxed energy E^{qc}(u)=∫ W^{qc}(∇u) dx, for W∈C^1 satisfying the growth, coercivity, and Lipschitz conditions (3.2)-(3.3). The proof strategy is: compactness from a coercivity estimate in Lemma 3.1; a liminf inequality obtained by reconstructing a continuous finite element function w_h from each discontinuous u_h (Lemma 2.2) and then applying quasiconvex lower semicontinuity; and a limsup (recovery sequence) argument via Dacorogna's relaxation theorem and interpolation estimates. The paper also reports numerical experiments for a polyconvex energy and for a two-well frame-indifferent energy, illustrating the behavior of the new penalty term and the emergence of microstructures.","tokens_in":19995,"tokens_out":8355,"duration_ms":94856,"significance":"If the stated results are correct, they are a significant advance: they remove the convexity requirement that had been essential in previous Gamma-convergence analyses of DG methods for variational problems, and they cover the two known convergent DG formulations. The continuous-reconstruction argument, used to control the non-vanishing curl of the discrete gradient, is a new and plausible mechanism for restoring lower semicontinuity in the quasiconvex setting. The numerical experiments are relevant and seem to support the qualitative claims. However, the paper is not yet acceptable in its current form: the penalty term defining the minimized functional is not well-defined for all W admitted by the stated assumptions, and the central reconstruction lemma is only sketched. Both are fixable, so I recommend a major revision rather than rejection.","major_comments":[{"comment":"The functional minimized in Theorem 3.1 is not well-defined on V_h^q for every W satisfying (3.2). Condition (3.2) only gives W(ξ) ≥ c(-1+|ξ|^p) with c>0, so W(0) may be arbitrarily negative. For example, take W(ξ)=-M+(M+1)|ξ|^p with M>1; this W satisfies (3.2)-(3.3), but for |Ω|>1/M the base 1+Σ_K∫_K W(∇u_h)+Σ_e h_e^{1-p}||Ju_hK||^p_{L^p(e)} is negative when u_h≡0 and u_0=0, so Pen(u_h) is undefined in real arithmetic because the exponent (p-1)/p is a non-integer. Thus Theorem 3.1 states convergence of minimizers of a functional that is not even a function on the discrete space for admissible data. Remark 3.1's asserted equivalence with the penalty (3.8) is also false in this case: the claimed lower bound 1+|u_h|^p_{W^{1,p}(Ω,T_h)} ≲ 1+Σ_K∫_K W(∇u_h)+Σ_e h_e^{1-p}||Ju_hK||^p fails because the right-hand side can be negative. This is a load-bearing gap, since the proof genuinely uses the penalty in the form (3.8). A concrete fix is to add an explicit lower bound such as W≥0 (or W(0)≥0) to the hypotheses, or to replace the first factor in (3.6) by its positive part, and then to recheck the estimates of Lemmas 3.1 and 3.2 under that assumption.","section":"§2.0.2, Lemma 2.2"}],"minor_comments":[{"comment":"The Lipschitz condition (3.3) contains a typo: the last term in the factor multiplying |ξ1-ξ2| should be |ξ2|^{p-1}, not a second copy of |ξ1|^{p-1}.","section":"§2, Eq. (3.3)"},{"comment":"The proof of the limsup inequality for Theorem 3.2 is delegated to 'adopting in similar fashion arguments from the previous section'. Since this is a second main theorem, the recovery sequence for E_{G,h} should be given in more detail, especially the control of the lifting term R_h(u_h) along the recovery sequence.","section":"§3.2, after Eq. (3.32)"},{"comment":"The smoothing bound (3.21), asserted for a sequence (u_δ)⊂C^∞(Ω̄) with |u_δ|_{W^{2,p}(Ω)} ≲ δ^{-1}|v|_{W^{1,p}(Ω)}, is stated without proof or citation; please provide a reference or a short justification.","section":"§3.3, Lemma 3.3"},{"comment":"The numerical section says that the computations use the penalty (3.6), but the implemented formula (4.8) rewrites the exponent and the second factor. Please clarify that the actual implemented penalty is (4.8), and mention whether this rewrite changes any of the theoretical requirements.","section":"§4, Eqs. (3.6) and (4.8)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core idea is promising, and the Gamma-convergence structure is likely repairable. However, the penalty definition issue in Theorem 3.1 is not a mere presentation detail: it makes the main theorem, as stated, invalid for a family of integrands admitted by the hypotheses. I would be comfortable with a revised version in which W is assumed nonnegative or the penalty is replaced by a manifestly nonnegative expression, and in which Lemma 2.2 is proved in complete form. I do not see grounds for rejection, as the corrections appear to be within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe headline: this is a genuinely new result—Gamma-convergence of DG methods for quasiconvex and relaxed vectorial variational problems—and the central idea, passing through continuous reconstructions to regain lower semicontinuity, is the right trick. But the manuscript as posted has a load-bearing gap in the definition of the penalty term, and a few proofs are only sketched or delegated. I would send it to a serious referee, but not without requiring major revision.\n\nWhat is new: all prior convergence results for these DG schemes were for convex W. Theorems 3.1 and 3.2 extend this to quasiconvex W and to the relaxed problem via the quasiconvex envelope. That answers an open question in the numerical analysis of microstructures. The liminf argument is structurally convincing: compactness in the broken Sobolev space, continuous reconstruction, quasiconvex lower semicontinuity, and control of the interface terms by the penalty. The numerical experiments with a two-well energy and microstructure formation are consistent with the theory and illustrate the main points.\n\nThe soft spots, in proportion. First, the penalty term (3.6) is not well-defined for all W satisfying (3.2). The lower bound only gives W(ξ) ≥ -C + c|ξ|^p, so at u_h=0 the base 1+∫W can be negative when the domain is large or W(0) is sufficiently negative. Taking the fractional power (p-1)/p of a negative number is not real. Remark 3.1 claims (3.6) is comparable to (3.8), and the proofs use (3.8). That comparison fails without an additional lower bound on W. This is not a minor typo: as stated, Theorem 3.1 applies to a functional that may not exist on the discrete space, and the proof analyzes a different objective. It is fixable—add a positive constant or assume W ≥ -a pointwise with a small enough, or redefine the penalty with max{·,0}—but it must be fixed.\n\nSecond, Lemma 2.2 is the load-bearing approximation result, and the proof is only sketched. For a general p and mesh geometry, this needs a full proof or a precise citation. Third, the limsup for the discrete-gradient method (Theorem 3.2) is delegated in one sentence; given that method's lifting operators, that should be written out. Also, no code or data accompanies the numerics, which limits reproducibility. And there is a typo in (3.3): the second |ξ1|^{p-1} should be |ξ2|^{p-1}.\n\nNone of this undermines the core idea. The reconstruction strategy is sound and should be publishable. The paper deserves refereeing, but the referees should push for a complete, well-defined statement and full proofs.\n\nBest,\n[You]","headline":"New Gamma-convergence results for DG methods in quasiconvex/relaxed problems, but the stated penalty is not always well-defined and the proofs are not fully complete.","tokens_in":20541,"tokens_out":4762,"would_cite":true,"duration_ms":44212,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","49J45","74B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that discontinuous Galerkin minimisers converge in L^p to minimisers of the quasiconvex-envelope (relaxed) energy, for quasiconvex and non-convex variational problems.","keywords":["discontinuous Galerkin","quasiconvexity","Gamma-convergence","relaxation","calculus of variations","nonlinear elasticity","finite element approximation","microstructure"],"falsifier":"Choose a family of meshes where the reconstruction constant in Lemma 2.2 can be computed exactly (for example, a single row of elements with increasing aspect ratio) and check whether the constant stays bounded for $p>2$. If it unbounded while jump terms vanish, the liminf inequality has no uniform control. Alternatively, run the DG scheme on a two-well energy with boundary data that do not align with the laminate directions: if the computed $L^p$ limit is not the known relaxed minimiser, or the discrete energy does not approach $\\min E^{qc}$, the convergence claim is false.","tokens_in":19453,"feed_emoji":"🧮","tokens_out":8051,"duration_ms":79553,"temperature":0.7,"pith_summary":"This paper proves that discontinuous Galerkin (DG) finite element methods converge for variational problems whose energy density $W$ is quasiconvex — the natural but weakest convexity notion in nonlinear elasticity — and, when $W$ is not convex, that discrete minimisers converge to minimisers of the relaxed problem $\\int_\\Omega W^{qc}$. This addresses a long-standing difficulty: DG approximations lack a true gradient structure, so the usual lower-semicontinuity argument for quasiconvex integrands does not apply directly. The proof restores the missing structure by reconstructing from each discontinuous $u_h$ a continuous piecewise polynomial $w_h$ whose error is controlled by the inter-element jumps. The result covers both the interior-penalty DG formulation of the authors' earlier work and the discrete-gradient formulation, and it is accompanied by numerical experiments on polyconvex and two-well (microstructure-forming) energies.","feed_headline":"Discontinuous Galerkin minimisers converge for non-convex energies","feed_subtitle":"Proof: despite the discrete gradient not being a true gradient, DG schemes reach the quasiconvex-envelope minimiser.","key_machinery":"The load-bearing object is a continuous reconstruction operator, generalised here from $p=2$ to any $p \\geq 2$ (Lemma 2.2): for every discontinuous piecewise polynomial $u_h$ in the DG space $V_h^q$, there is a continuous piecewise polynomial $w_h$ with $\\|\\nabla^\\alpha u_h - \\nabla^\\alpha w_h\\|_{L^p}$ bounded by inter-element jump terms, uniformly in $h$. This reconstruction lets the proof pass from $u_h$ to $w_h$, where $\\nabla w_h$ is a genuine gradient and $W^{qc}$ is lower semicontinuous under weak $L^p$ convergence. A second ingredient is the modified penalty $\\mathrm{Pen}(u_h)$, built from $(1+|u_h|^p_{W^{1,p}})^{(p-1)/p}$ times the jump term, which provides coercivity for non-convex $W$ and makes the consistency terms vanish as $h \\to 0$.","core_discovery":"On the paper's own terms, the central discovery is Theorem 3.1 and Theorem 3.2: for $W \\in C^1$ satisfying $p$-growth and coercivity with $p>1$, any sequence of discrete minimisers $u_h$ of the DG energy $E_h$ converges in $L^p(\\Omega)$, up to a subsequence, to a function $u$ that minimises the relaxed energy $E^{qc}(u)=\\int_\\Omega W^{qc}(\\nabla u)$, and $E^{qc}(u)$ equals the infimum over the admissible set. When $W$ is quasiconvex, $W^{qc}=W$, so the discrete minimisers converge to genuine minimisers of the original problem. The theorem holds even though the discrete gradient is not a true gradient — its curl need not vanish — because the liminf inequality is obtained through the continuous reconstruction $w_h$ rather than through the discontinuous gradient $\\nabla_h u_h$.","pith_inferences":["A natural testable extension is to check whether the reconstruction estimates remain uniform on anisotropic or adaptively refined meshes; the paper's proof assumes a family where the constants stay bounded as $h\\to 0$.","The same reconstruction-plus-liminf strategy could be applied to other relaxed objects such as polyconvex or rank-one convex envelopes, and to problems with Lavrentiev gaps, but the paper does not claim these extensions.","The paper's two-well numerics report roundoff degradation for large $p$ (the $p=8$ example), suggesting that a rescaled jump term is needed in practice; this is a practical adjustment the theory does not address."],"forward_implications":["For quasiconvex $W$, the discrete minimisers of the DG energy converge in $L^p(\\Omega)$ to a minimiser of the continuous energy $\\int_\\Omega W(\\nabla u)$, so the method can be trusted for elasticity-type problems without full convexity.","For non-quasiconvex $W$, the discrete minimisers converge to a minimiser of the relaxed energy $\\int_\\Omega W^{qc}(\\nabla u)$, meaning that finitely many mesh oscillations are correctly averaged into the macroscopic relaxed state.","Both the interior-penalty DG formulation and the discrete-gradient DG formulation satisfy the same convergence theorem, unifying the DG methods previously known to converge only for convex energies.","The numerical two-well experiments show discrete minimising sequences forming ever-finer laminates whose $L^2$ limit is the relaxed minimiser, consistent with the $\\Gamma$-convergence result.","The paper's pointwise-energy experiments suggest that quasiconvex envelopes $W^{qc}(F)$ can be approximated by solving a sequence of discrete minimisation problems."],"supporting_citations":[{"why":"Supplies the continuous reconstruction operator and its error estimates, which the paper generalises from $p=2$ to $L^p$ and uses as the main tool in the liminf inequality.","marker":"[28]"},{"why":"Provides the DG energy functional, Nitsche boundary treatment, and discrete-stability penalty that Theorem 3.1 analyses.","marker":"[27]"},{"why":"Defines the discrete-gradient DG method whose minimisers are treated in Theorem 3.2, with discrete gradients obtained through lifting operators.","marker":"[38]"},{"why":"Supplies the broken-Sobolev compactness and convex-case $\\Gamma$-convergence framework used for the $L^p$ compactness of bounded-energy sequences.","marker":"[11]"},{"why":"Provides the relaxation theorem used to produce recovery sequences that attain the quasiconvex-envelope energy $E^{qc}(u)$.","marker":"[21]"}],"fun_headline_variants":["DG schemes converge to relaxed minimisers for non-convex energies","No true gradient? DG still finds the quasiconvex envelope minimiser","Quasiconvex convergence: DG methods tackle non-convex variational problems","Relaxed variational minimisers via discontinuous Galerkin, even when gradients fail","Discontinuous Galerkin converges for quasiconvex and relaxed energies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the estimate that every discontinuous piecewise polynomial $u_h$ can be replaced by a continuous piecewise polynomial $w_h$ whose $L^p$ error, and the error of its gradient, is bounded by the size of the inter-element jumps uniformly in $h$; the paper gives the statement and a sketch but omits the full proof.","fun_headline_variants_meta":{"raw":{"variants":["DG schemes converge to relaxed minimisers for non-convex energies","No true gradient? DG still finds the quasiconvex envelope minimiser","Quasiconvex convergence: DG methods tackle non-convex variational problems","Relaxed variational minimisers via discontinuous Galerkin, even when gradients fail","Discontinuous Galerkin converges for quasiconvex and relaxed energies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000372,"raw_usage":{"total_tokens":1968,"prompt_tokens":902,"completion_tokens":1066,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":971}},"tokens_in":518,"tokens_out":1066,"duration_ms":8873,"temperature":1.0,"reasoning_tokens":971,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:43:08.451859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a family of meshes where the reconstruction constant in Lemma 2.2 can be computed exactly (for example, a single row of elements with increasing aspect ratio) and check whether the constant stays bounded for $p>2$. If it unbounded while jump terms vanish, the liminf inequality has no uniform control. Alternatively, run the DG scheme on a two-well energy with boundary data that do not align with the laminate directions: if the computed $L^p$ limit is not the known relaxed minimiser, or the discrete energy does not approach $\\min E^{qc}$, the convergence claim is false.","supporting_citations":[{"cited_title":"A posteriori error estimates for a discontinuous Galerkin approximation of second-order elliptic prob- lems","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous reconstruction operator and its error estimates, which the paper generalises from $p=2$ to $L^p$ and uses as the main tool in the liminf inequality."},{"cited_title":"A class of Discontinuous Galerkin methods for nonlinear variational problems","cited_arxiv_id":"2308.12891","evidence_quote":"Provides the DG energy functional, Nitsche boundary treatment, and discrete-stability penalty that Theorem 3.1 analyses."},{"cited_title":"Discontinuous Galerkin methods for non- linear elasticity","cited_arxiv_id":null,"evidence_quote":"Defines the discrete-gradient DG method whose minimisers are treated in Theorem 3.2, with discrete gradients obtained through lifting operators."},{"cited_title":"Dacorogna","cited_arxiv_id":null,"evidence_quote":"Provides the relaxation theorem used to produce recovery sequences that attain the quasiconvex-envelope energy $E^{qc}(u)$."}],"review_version":1}