{"id":"b856a55b-b410-4a6f-8fb6-e0f93bd0a49e","arxiv_id":"2608.09759","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper derives forward and adjoint sensitivity equations for the information geometrically regularized Euler equations and numerically verifies them against finite differences and automatic differentiation.","lead":"Computing how shock flows respond to input changes is notoriously hard because shock sensors and limiters corrupt derivatives. This paper derives sensitivity equations for a model that smooths shocks without viscosity, and shows they agree with finite differences and automatic differentiation on several test flows.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own scaling study contradicts the claimed grid-refinement convergence under the natural √α ∝ Δx scaling, and the proposed Δx^{2/3} rescaling is unproven with its target limit unidentified.","rationale":"The reader's weakest assumption is precisely the commutativity of differentiation with the vanishing-regularization limit, and the paper's own scaling study is the strongest evidence that this assumption fails under the natural grid-refinement path. My stress-test pass confirms this: the abstract promises convergence under grid refinement without qualifying the scaling, while Section 5.2 demonstrates non-convergence for √α ∝ Δx and only apparent convergence for an empirically chosen √α ∝ Δx^{2/3}. The proposed sublinear scaling is not derived from any error analysis, and the text explicitly defers rigorous justification, so the central practical claim—that IGR provides reliable sensitivities of shock flows—remains a conjecture supported by favorable experiments. I do not see a reason to move the verdict further: the derivations in Sections 2 and 3 are formally sound given smooth IGR solutions, the fixed-α numerical checks are consistent, and the code is said to be available. The correct disposition is CONDITIONAL: accept only if the scaling question is resolved or the claims are restricted to fixed-α convergence. Since the reader already reached CONDITIONAL, no verdict change is needed.","tokens_in":20465,"tokens_out":3495,"duration_ms":33808,"concrete_test":"Run a one-dimensional shock-tube or sine-wave case with the sublinear scaling √α = s Δx^{2/3} over Ne = 128, 256, ..., 8192 and compare the IGR adjoint's directional derivative dJ/ds against the exact sensitivity of the entropy solution computed from the Rankine–Hugoniot shock-position derivative (or against a high-resolution reference using the Giles–Ulbrich viscous-regularization theory with regularization width √α). If the IGR value does not converge to the Euler sensitivity as Ne increases, the central claim fails; if it does converge, repeat the same test under the natural scaling √α = s Δx to confirm that the failure in Figure 5 is real and not an artifact of the specific test functional.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract; Section 5) is that the IGR forward/adjoint sensitivities converge, under grid refinement, to the sensitivities obtained by differentiating the numerical forward solve. The paper's own Section 5.2, Figures 4 and 5, show that under the natural scaling √α = s Δx—the scaling that keeps the regularized shock width resolved by a fixed number of cells—the adjoint solutions do not converge: the L1 differences between successive resolutions break upward at the finest resolutions for both s = 2 and s = 4 (Figure 5, right). Only the heuristic sublinear scaling √α = s Δx0^{1/3} Δx^{2/3} appears to converge, and the text explicitly defers a rigorous analysis of this phenomenon to future work. Even if the sublinear path converges, no argument identifies the limit with the sensitivity of the underlying discontinuous Euler solution: differentiating the smooth IGR system (2.1) and sending α to zero does not obviously commute with the vanishing-regularization limit, and no error estimates connect the IGR adjoint to the Euler adjoint in the sense of the Giles–Ulbrich theory cited in [20, 21]. The fixed-α experiments (Figures 6 and 7) only validate discretization of the continuous IGR adjoint; they say nothing about α → 0. Thus the abstract's unconditional 'convergence under grid refinement' overstates the evidence, and the load-bearing assumption that differentiation and α → 0 commute is both unproven and, under the standard scaling, contradicted by the paper's own numerics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives forward and adjoint sensitivity equations for the information geometric regularization (IGR) of the compressible Euler equations with periodic boundary conditions. The forward equations are obtained by direct differentiation of the IGR system, and the adjoint equations are derived through an abstract integration-by-parts argument extended to the nonlocal entropic-pressure flux. The resulting PDEs are discretized with a discontinuous Galerkin method, and numerical tests in one and two dimensions compare the PDE-based sensitivities against automatic differentiation and finite differences. The abstract and Section 5 claim convergence of these sensitivities under grid refinement to the discrete sensitivities of the forward solve.","tokens_in":20773,"tokens_out":2914,"duration_ms":27624,"significance":"If the central claim were fully established, this would be a valuable contribution to adjoint-based design and uncertainty quantification for compressible flows with shocks, since IGR offers an inviscid smooth regularization whose sensitivities could be computed by standard continuous adjoint methods. The formal derivation is self-contained and the fixed-regularization numerical validations (Figures 1-3, 6-7) are credible, with a publicly available code. However, the paper's most important claim—that the sensitivities converge in the joint zero-regularization and zero-mesh-size limit—is not supported by the presented evidence; the paper's own scaling study shows a breakdown under the natural scaling and defers rigorous analysis of the substitute scaling. The contribution is therefore best viewed as a derivation plus a fixed-α consistency study, with the vanishing-regularization question left open.","major_comments":[{"comment":"The abstract states that the paper 'demonstrates their convergence, under grid refinement' to the sensitivities obtained by finite differences or automatic differentiation. This is not supported by the paper's own scaling study. Under the natural scaling √α = s Δx, which keeps the number of cells per regularized shock width fixed, the L1 differences between adjoint solutions at consecutive resolutions break upward at the finest resolutions for both s = 2 and s = 4 (Figure 5, right). Only the ad hoc sublinear scaling √α = s Δx0^{1/3} Δx^{2/3} appears to converge, and the text explicitly states that a rigorous analysis is future work. The claim in the abstract should be qualified to fixed regularization or to the sublinear scaling as a heuristic.","section":"Abstract and Section 5.2, Figures 4-5"},{"comment":"The sublinear scaling √α = s Δx0^{1/3} Δx^{2/3} is introduced as a numerical observation motivated by Giles and Ulbrich [20, 21], but no argument is given that the limit of the IGR adjoint under this scaling is the sensitivity of the underlying discontinuous Euler solution. Differentiating the smooth IGR system (2.1) and then letting α → 0 does not obviously commute with the vanishing-regularization limit, and the paper provides no error estimates connecting the IGR adjoint to an appropriately defined Euler adjoint in the sense of [20, 21]. Without such an identification, the title's promise of 'sensitivities of flows with shocks' is not realized; the paper rigorously addresses only sensitivities of the fixed-α regularized system.","section":"Section 5.2, Figure 5"},{"comment":"The fixed-α two-dimensional experiments with mesh refinement at constant α validate only the discretization of the continuous IGR adjoint; they do not provide evidence about the α → 0 limit. These figures should be presented as consistency checks for the numerical scheme at fixed regularization, not as convergence of sensitivities to those of the unregularized Euler equations. The O(h^2) labels in these figures are also not defined in terms of the norm or the range of h used, so the claimed convergence order is not verifiable from the text.","section":"Section 5, Figures 6-7"}],"minor_comments":[{"comment":"The title as rendered contains spacing artifacts ('INFORMA TION', 'REGULARIZA TION'); these should be corrected in the final version.","section":"Title page"},{"comment":"The text says 'T = 1.7 ≫ ts' without defining ts or stating its value; please define the shock-formation time or rephrase.","section":"Section 5.1"},{"comment":"The sentence 'The above expression can be deduced by checking dimensional consistency and remembering that ˆq is contracted over the derivatives with respect to q' is too informal for a derivation of the adjoint operator; the tensor contraction should be shown explicitly or a reference provided.","section":"Section 3.2"},{"comment":"The 'small discretize-then-differentiate gap' mentioned for Figures 1 and 2 is not quantified; a table of the maximum or L1 difference would make the claim precise.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope as a numerical analysis contribution. The main issue is the mismatch between the abstract's convergence claim and the evidence in Section 5.2; the authors should either prove or clearly circumscribe the vanishing-regularization statement. I would not reject the paper, since the fixed-α derivations and validations are valuable, but the framing needs substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read Schäfer's IGR sensitivity paper. The useful core is real: he derives the forward and adjoint PDEs for information geometric regularization of Euler, and the treatment of the nonlocal entropic-pressure term via an L2 gradient and one elliptic solve is a genuine technical step. The adjoint derivation is a standard integration-by-parts argument, but the conservative form (3.5) and the nonlocal source (3.7) are consistent, and the fixed-alpha tests (Figures 3, 6, 7) show the discrete and continuous adjoints agreeing at the expected order.\n\nThe soft spot is the scaling claim, and the paper is more honest than its abstract. The abstract says sensitivities converge under grid refinement to discretize-then-differentiate results, but Section 5.2 shows the natural √α ∝ Δx scaling—the one that fixes cells per shock width—does not converge: Figure 5's consecutive-resolution differences break upward at fine resolutions for both s=2 and s=4. Only the special √α ∝ Δx^{2/3} scaling contracts, with no proof or identification of the limit. The fixed-alpha experiments validate discretization of the continuous IGR adjoint; they say nothing about α→0. So the load-bearing premise, that differentiating IGR and letting α tend to zero gives the desired sensitivity of the underlying Euler solution, is unproven and, under the standard scaling, contradicted by the paper's own numerics. To its credit, the text says a rigorous analysis is future work; the abstract just omits the caveat.\n\nFor whom: people doing adjoint-based design or UQ in compressible flow, and developers of inviscid regularizations. The derivations and the nonlocal trick are worth keeping.\n\nRecommendation: send it to peer review. The equations are new, the code is linked, and the scaling failure is a real research problem worth refereeing. The referee should ask for a revised abstract and a proof or much stronger evidence for the Δx^{2/3} scaling, with the limit identified. Accept after a major revision, not as is.","headline":"First derivation of IGR forward/adjoint sensitivities with a clever one-elliptic-solve nonlocal term; honest numerics, but the abstract overstates convergence under natural scaling.","tokens_in":21328,"tokens_out":3107,"would_cite":true,"duration_ms":26050,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35L65","76L05","65M60","76N25","49M41"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper derives forward and adjoint sensitivity equations for the IGR regularization of the Euler equations and shows they converge, under grid refinement, to finite-difference and automatic-differentiation sensitivities.","keywords":["adjoint sensitivities","information geometric regularization","Euler equations","shock waves","inviscid PDE regularizations","forward sensitivities","discontinuous Galerkin","vanishing regularization limit"],"falsifier":"Extend the mesh sweep of Section 5.2 to $N_e = 8192$ and beyond under the sublinear scaling $\\sqrt{\\alpha} = s\\,\\Delta x_0^{1/3}\\Delta x^{2/3}$: if the $L^1$ consecutive-resolution error of the adjoint solution stops decreasing or reverses, the paper's claim that the IGR adjoint converges under this scaling is refuted.","tokens_in":20231,"feed_emoji":"💥","tokens_out":9636,"duration_ms":71261,"temperature":0.7,"pith_summary":"Computing adjoint sensitivities of flows with shocks has long been unreliable because shock-capturing limiters and sensors produce spurious spikes when differentiated. This paper shows that a recently introduced regularization, information geometric regularization (IGR), which replaces shocks with smooth profiles through an auxiliary entropic pressure, admits clean forward and adjoint sensitivity equations. The paper derives those equations for periodic domains and verifies in one and two dimensions that solving them reproduces the sensitivities obtained by finite differences or automatic differentiation through the forward IGR solve, with agreement improving under mesh refinement at fixed regularization. If correct, this gives practitioners a route to gradient-based design optimization, uncertainty quantification, and scientific machine learning for flows with shocks without freezing limiters or accepting their systematic errors.","feed_headline":"Adjoint equations for shock flows now match finite differences","feed_subtitle":"Smooth entropic-pressure shock profiles give gradient-based design and UQ a way around spiky limiter sensitivities.","key_machinery":"The central object is the information geometric regularization of the compressible Euler equations, which replaces shocks with smooth profiles of width proportional to $\\sqrt{\\alpha}$ by augmenting the physical pressure $P$ with an entropic pressure $\\Sigma$ defined through an elliptic equation. Two mechanisms carry the argument: formal differentiation of the IGR system, which yields an evolution PDE for the forward sensitivity together with an additional elliptic PDE for the entropic-pressure sensitivity $\\hat\\Sigma$; and an abstract adjoint calculus for hyperbolic systems with nonlocal fluxes that splits the flux into a local part and a part mediated by $\\Sigma$. The nonlocal adjoint term is evaluated through the $L^2$ gradient $\\Sigma_q$, which the paper expresses using an auxiliary adjoint elliptic field $\\Pi$ (equation (3.6)) and the self-adjointness of the elliptic operator, yielding the nonlocal source term (3.7) at the cost of one elliptic solve per time step. The numerical implementation uses a nodal discontinuous Galerkin method with symmetric interior penalty discretization of the elliptic problems and local Lax--Friedrichs fluxes for both the forward and adjoint hyperbolic solves.","core_discovery":"The paper's central claim is that the IGR system, the compressible Euler equations augmented by an entropic pressure $\\Sigma$ solving the elliptic equation $\\Sigma/\\rho - \\alpha\\,\\mathrm{div}(\\rho^{-1}\\nabla\\Sigma) = \\alpha(\\mathrm{tr}^2(Du)+\\mathrm{tr}((Du)^2))$, supports a well-posed sensitivity calculus: the forward sensitivity equations (2.1) and the adjoint equations (3.5), closed by the nonlocal term (3.7) that expresses the $L^2$ gradient $\\Sigma_q$ through one additional elliptic solve, correctly describe how regularized shock solutions respond to parameter changes. Numerically, the PDE-based sensitivities agree with finite differences and with forward and reverse-mode automatic differentiation through the discretized IGR solve, and the agreement tightens under grid refinement at fixed $\\alpha$ for one- and two-dimensional problems including interacting blast waves and a blast--vortex interaction. The paper further reports that the adjoint solution appears to converge under the sublinear scaling $\\sqrt{\\alpha}\\propto\\Delta x^{2/3}$, whereas the more common scaling $\\sqrt{\\alpha}\\propto\\Delta x$ fails to converge at fine resolutions, and it defers a rigorous analysis of this phenomenon to future work.","pith_inferences":["If the sublinear-scaling convergence is later proven, the IGR adjoint would settle the long-standing tension between 'differentiate-then-discretize' and 'discretize-then-differentiate' for Euler flows with shocks, because the equation-based sensitivity would track the discretized regularized solution at every resolution.","The same splitting argument, a local hyperbolic flux plus an elliptic auxiliary field whose $L^2$ gradient is recovered from an adjoint elliptic solve, should transfer to other PDE-based regularizations with self-adjoint elliptic operators, such as artificial bulk viscosity or hyperviscosity.","A practical recommendation that follows from the scaling study, though the paper does not make it explicitly, is to choose $\\alpha$ according to the resolution demanded by the output functional rather than purely by $\\Delta x$, since the adjoint but not the primal solution distinguishes the two scaling paths."],"forward_implications":["For a fixed regularization strength $\\alpha$, the continuous IGR forward and adjoint sensitivities coincide with finite-difference and automatic-differentiation sensitivities under mesh refinement, so the adjoint can replace per-parameter forward solves.","The adjoint of a scalar output functional costs one backward-in-time PDE solve plus one elliptic solve per time step for the auxiliary field $\\Pi$, independent of the number of parameters.","Under the sublinear scaling $\\sqrt{\\alpha}\\propto\\Delta x^{2/3}$ the discrete IGR adjoint appears to converge across resolutions, whereas the standard linear scaling $\\sqrt{\\alpha}\\propto\\Delta x$ does not, making the choice of $\\alpha$ relative to the grid decisive for the reliability of the sensitivity.","In two dimensions, the PDE-based adjoint agrees with central finite differences for directional derivatives of kinetic-energy objectives on triple Sedov blasts and blast--vortex interactions, with the mismatch decreasing at second order in the mesh spacing."],"supporting_citations":[{"why":"Defines the IGR system whose forward and adjoint sensitivities are the subject of this paper.","marker":"[7]"},{"why":"Provides global strong solutions converging to entropy solutions in the pressureless case, supporting the smooth-profile premise.","marker":"[8]"},{"why":"Supplies the discontinuous Galerkin semidiscretization of IGR used in all numerical experiments.","marker":"[16]"},{"why":"Source of the sublinear $\\Delta x^{2/3}$ scaling idea that the paper's adjoint mesh sweep is modeled on.","marker":"[20]"},{"why":"Companion analysis of adjoint convergence for discontinuous solutions that this paper extends empirically to the IGR adjoint.","marker":"[21]"},{"why":"Documents the spurious sensitivities of shock-laden flows that motivate replacing shock-capturing limiters with smooth regularizations.","marker":"[5]"}],"fun_headline_variants":["Shock flow sensitivities via IGR match finite differences","IGR adjoints converge under grid refinement","Smooth entropic shock profiles fix adjoint gradients","Adjoint shock sensitivities now grid-converged","IGR regularization yields FD-matching adjoints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that differentiating the smooth IGR solution and then letting the regularization strength $\\alpha$ and the grid spacing $\\Delta x$ tend to zero together yields the true sensitivity of the underlying discontinuous Euler solution, so that the regularized sensitivity is not an artifact of the smoothing.","fun_headline_variants_meta":{"raw":{"variants":["Shock flow sensitivities via IGR match finite differences","IGR adjoints converge under grid refinement","Smooth entropic shock profiles fix adjoint gradients","Adjoint shock sensitivities now grid-converged","IGR regularization yields FD-matching adjoints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000816,"raw_usage":{"total_tokens":3553,"prompt_tokens":903,"completion_tokens":2650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2578}},"tokens_in":519,"tokens_out":2650,"duration_ms":17527,"temperature":1.0,"reasoning_tokens":2578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:17:32.941264+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Extend the mesh sweep of Section 5.2 to $N_e = 8192$ and beyond under the sublinear scaling $\\sqrt{\\alpha} = s\\,\\Delta x_0^{1/3}\\Delta x^{2/3}$: if the $L^1$ consecutive-resolution error of the adjoint solution stops decreasing or reverses, the paper's claim that the IGR adjoint converges under this scaling is refuted.","supporting_citations":[{"cited_title":"Discontinuous Galerkin Semidiscretization of the Information Geometric Regularized Compressible Euler Equations","cited_arxiv_id":"2608.02223","evidence_quote":"Supplies the discontinuous Galerkin semidiscretization of IGR used in all numerical experiments."},{"cited_title":"Giles and Stefan Ulbrich","cited_arxiv_id":null,"evidence_quote":"Source of the sublinear $\\Delta x^{2/3}$ scaling idea that the paper's adjoint mesh sweep is modeled on."},{"cited_title":"Giles and Stefan Ulbrich","cited_arxiv_id":null,"evidence_quote":"Companion analysis of adjoint convergence for discontinuous solutions that this paper extends empirically to the IGR adjoint."},{"cited_title":"Adjoint-based sensitivity of shock-laden flows","cited_arxiv_id":null,"evidence_quote":"Documents the spurious sensitivities of shock-laden flows that motivate replacing shock-capturing limiters with smooth regularizations."}],"review_version":1}