{"id":"c7f705b0-6db4-4c07-92a8-6fbcf0bed8df","arxiv_id":"2606.23691","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"An approximate second-order adjoint for the shallow-stream ice sheet equations produces Hessians whose leading ~4 eigenvector modes match the exact AD-computed Hessian, but whose subspaces diverge beyond that rank.","lead":"The paper derives an approximate second-order adjoint method for ice sheet models that use the shallow-stream approximation, comparing it against exact Hessians from algorithmic differentiation. It finds the approximation works for the first few modes but diverges spectrally after about four, guiding when cheap approximations suffice versus when full AD is needed.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The self-adjoint approximation is the central object of study, not a hidden assumption; the paper is appropriately cautious. The real load-bearing risk is generalizability of the spectral findings to more realistic flow regimes.","rationale":"The paper does what it sets out to do: derive the SOSA method, compare it against AD, and characterize the discrepancy. The derivation is correct and self-contained. The spectral diagnostics are appropriate and validated. The conclusions are appropriately cautious.\n\nThe reader's weakest_assumption correctly identifies the self-adjoint approximation as the central assumption, but frames it as if it were a hidden weakness. In fact, it is the explicit subject of the paper — the entire contribution is to characterize what this approximation costs. The paper does not claim the approximation is correct; it quantifies how it fails.\n\nThe real risk is that the quantitative findings (4-mode threshold, factor-of-3 eigenvalue ratio) are specific to the two test domains. Both are idealized: uniform thickness, simple geometry, no grounding line. The factor-of-3 argument via Glen's n is a post-hoc heuristic that works for simple shear but may not hold for mixed flow regimes. However, this is a generalizability concern, not an internal inconsistency, and the paper's hedged language ('case-dependent') partially addresses it.\n\nNo code repository is provided, which limits reproducibility, but the methods are described in sufficient detail for re-implementation. The two-domain test suite is limited but adequate for a methods paper whose conclusions are appropriately qualified.\n\nThe verdict of ACCEPT is appropriate. The paper makes a useful contribution: a derived method, a rigorous comparison, and honest assessment of limitations. The generalizability concern does not undermine the paper's claims as stated, but a grounding-line test case would substantially strengthen (or revise) the quantitative conclusions.","tokens_in":18594,"tokens_out":3987,"duration_ms":144594,"concrete_test":"Re-run the spectral comparison (principal angles + eigenvalue ratios) on a domain that includes a grounding line — e.g., a marine ice sheet with variable thickness transitioning from grounded to floating ice. If the '4 modes safe' threshold shifts by more than a factor of 2, or if the eigenvalue ratio departs from the range [3,4] by more than 50%, the generalizability of the paper's quantitative conclusions is compromised. This is the single most informative additional experiment because grounding-line dynamics are where the SSA nonlinearity is strongest and where most real inverse problems are posed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies the self-adjoint approximation (Eq. A.12) as the key assumption, but this is the approximation the paper explicitly sets out to characterize — it is the subject of the study, not a concealed premise. The derivation (Appendices A–B) is internally consistent: linearity of G in u follows from freezing μ̄, self-adjointness follows from linearity plus boundary conditions (Eq. A.11), and the resulting SOSA system (Eqs. 15–17, B.29–B.37) is correctly derived. The spectral comparison methodology is sound: eigenvalue residuals (Eq. 20) and orthonormality checks (Eq. 21) confirm numerical quality, and principal angles are an appropriate diagnostic for subspace alignment.\n\nThe genuine load-bearing concern is generalizability. The 'first 4 modes are safe' threshold and the 'factor of ~3' eigenvalue ratio are established on exactly two idealized domains: a uniform-thickness ice shelf and a synthetic snaking stream. Both have simple geometry, uniform thickness, and no grounding line. Real ice sheet inverse problems involve grounding zones (where the SSA changes character from floating to grounded), strongly variable thickness, and thermomechanical coupling. In such regimes, the relative contribution of the neglected viscosity nonlinearity to the Hessian structure could differ substantially, shifting the divergence threshold or the eigenvalue ratio. The paper's heuristic explanation for the factor-of-3 ratio — that '∂uG is roughly a factor of n=3 smaller when derivatives of the viscosity are included' (§5.2) — holds for simple shear but is not obviously correct for mixed compression/shear regimes where the effective strain rate invariant enters nonlinearly. If the ratio is not a stable ~n but varies with flow regime, the suggestion that 'a scaling factor could be applied to the eigenvalues' (§6.1) becomes unreliable.\n\nThat said, the paper's conclusions are appropriately hedged: 'case-dependent,' 'full Hessian preferable where feasible.' The claim do","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This manuscript derives a second-order self-adjoint (SOSA) formulation for the shallow-stream approximation (SSA) of ice sheet flow, in which the nonlinear dependence of viscosity on velocity is neglected (the standard linear-viscosity approximation). The resulting approximate Hessian-vector products are compared against exact Hessians computed via algorithmic differentiation (AD) in a JAX-based finite-volume code. The comparison uses two synthetic test domains (an ice shelf and a snaking ice stream) and employs spectral diagnostics—eigenvalue residuals, orthogonality residuals, and principal angles between invariant subspaces—to quantify the fidelity of the approximation. The main finding is that the SOSA and AD Hessians share similar leading eigenvector structure for approximately the first 4 modes, after which the subspaces diverge, with SOSA eigenvalues systematically smaller by a factor of roughly 3. The authors conclude that the SOSA Hessian is case-dependent in utility and recommend the full AD Hessian where high-fidelity second-order information is needed above very low rank.","tokens_in":19458,"tokens_out":1127,"duration_ms":405050,"significance":"The paper addresses a practically important question for the ice sheet modeling community: whether the widely used self-adjoint (linear-viscosity) approximation, well-established for first-order adjoints, can be extended to second-order derivatives with acceptable fidelity. The derivation (Appendices A–B) is self-contained and follows standard adjoint methodology cleanly. The spectral comparison is rigorous: eigenvalue residuals (Eq. 20) and orthogonality residuals (Eq. 21) confirm numerical quality of both Hessians, and principal angles between subspaces provide a principled similarity metric. The implementation in JAX enables a clean AD ground truth. The finding that subspace divergence begins after approximately 4 modes and the eigenvalue ratio of ~3 are concrete, falsifiable results that provide actionable guidance to practitioners. The honest assessment—that the approximation is case-dependent and inferior to the full Hessian above low rank—is appropriately cautious.","major_comments":[{"comment":"§5.2, paragraph on the factor-of-3 eigenvalue ratio: the heuristic explanation that '∂uG is roughly a factor of n=3 smaller when derivatives of the viscosity are included' is stated without derivation or reference. This is a load-bearing claim because it is the primary physical explanation for the systematic eigenvalue discrepancy that motivates the recommendation against using SOSA for uncertainty quantification (§6.2). A more precise justification—showing how the viscosity nonlinearity enters ∂uG and why it produces a factor of n in the eigenvalues—should be provided, or the claim should be softened to a conjecture.","section":null},{"comment":"§5.2 and §7: the conclusion that subspaces are safe 'up to the first 4 modes' is established on exactly two idealized domains, both with uniform thickness and no grounding line. The abstract and conclusion state this threshold as a general finding. Given that real ice sheet inverse problems involve grounding zones, variable thickness, and thermomechanical coupling, the generalizability of the '4 modes' threshold is a correctness-risk concern. The authors should add a brief statement in the conclusion acknowledging that the threshold is domain-specific and may shift under different flow regimes, or alternatively test a third domain with non-uniform thickness to strengthen the claim.","section":null}],"minor_comments":[{"comment":"Abstract: 'second order derivatives' and 'second-order derivatives' are both used; please unify hyphenation throughout.","section":null},{"comment":"§4.1, Ice Shelf description: 'righ-hand-side' should be 'right-hand-side' (missing 't').","section":null},{"comment":"§4.1, Twisty Stream: the formula for C contains nested parentheses that are hard to parse; please verify the closing delimiters match.","section":null},{"comment":"Figure 5 caption: panels (a) and (b) labels in the caption text appear swapped relative to the figure layout described; please check consistency.","section":null},{"comment":"§5.2: the introduction of the functional J in Eq. (19) comes after results from a different functional (J = ∫√(u·u)) are shown in §5.1; a brief note that the functional changes between subsections would help the reader.","section":null},{"comment":"Eq. (9): the definition φ(q) = φ₀e^q is used without specifying the meaning of φ₀ or the units/dimensions of q; a brief clarification would help.","section":null},{"comment":"Appendix B.2: the list of 9 terms (Eqs. B.11–B.19) is introduced as '12 terms' earlier in the text; please reconcile the count.","section":null},{"comment":"§6.1: 'indetically zero' should be 'identically zero' (also appears in Appendix B.1, Eq. B.6 description).","section":null}],"recommendation":"minor_revision","confidential_remarks":"The reader's report and stress-test note both flag the self-adjoint approximation (Eq. A.12) as the key assumption. On reading the paper, I agree with the skeptic's assessment that this is the central object of study, not a concealed premise—the paper is appropriately explicit about it. The more substantive concern is generalizability of the spectral findings, which I have elevated to a major comment. The factor-of-3 explanation is the other item I consider load-bearing; it deserves either a derivation or a softer statement. Apart from these, the manuscript is well-structured and the derivation is sound."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. Both major comments are well-taken and will be addressed in revision. On the factor-of-3 eigenvalue ratio, we agree the heuristic explanation is insufficiently justified and will either provide a more precise derivation or soften the claim to a conjecture. On the generalizability of the '4 modes' threshold, we agree it is domain-specific and will add an explicit caveat in the conclusion.","responses":[{"response":"The referee is correct that the factor-of-3 explanation as currently stated is a heuristic assertion without sufficient justification. We will revise this in one of two ways. Our preferred approach is to provide a more precise derivation: the SSA effective viscosity scales as |u|^{(1-n)/n} under Glen's law, so the derivative of viscosity with respect to velocity introduces a factor of (1-n)/n, which for n=3 yields a factor of -2/3. The way this factor enters the operator ∂uG, combined with the structure of the Hessian-vector product (Eq. 17), can plausibly account for the observed ratio. However, we acknowledge that a fully rigorous derivation showing this factor propagates unchanged into the eigenvalues is not straightforward, as it depends on which terms dominate in each eigenmode. If we cannot complete this derivation to our satisfaction, we will instead soften the claim to a conjecture, explicitly stating that the factor-of-3 ratio is consistent with the viscosity nonlinearity scaling but that a rigorous derivation is not provided. Either way, the recommendation in §6.2 against using SOSA for uncertainty quantification does not depend solely on this explanation—it rests on the observed eigenvalue discrepancy itself, which is an empirical result from the spectral comparison.","revision_made":"partial","referee_comment":"§5.2, paragraph on the factor-of-3 eigenvalue ratio: the heuristic explanation that '∂uG is roughly a factor of n=3 smaller when derivatives of the viscosity are included' is stated without derivation or reference. This is a load-bearing claim because it is the primary physical explanation for the systematic eigenvalue discrepancy that motivates the recommendation against using SOSA for uncertainty quantification (§6.2). A more precise justification—showing how the viscosity nonlinearity enters ∂uG and why it produces a factor of n in the eigenvalues—should be provided, or the claim should be softened to a conjecture."},{"response":"We agree that the '4 modes' threshold is established on only two idealized domains and should not be stated as a general finding without qualification. We will add an explicit caveat in both the abstract and the conclusion acknowledging that this threshold is domain-specific and may shift under different flow regimes, particularly in the presence of grounding zones, variable thickness, or thermomechanical coupling. We considered adding a third test domain with non-uniform thickness, but we feel this is beyond the scope of a minor revision and would not by itself resolve the generalizability question—two or three idealized domains cannot establish a universal threshold. The honest and appropriate response is to qualify the claim rather than to overstate it. We will also adjust the language in §5.2 to make clear that the threshold of approximately 4 modes is specific to the Twisty Stream domain and that the broader, more robust finding is that subspace divergence begins at low rank and does not accelerate thereafter.","revision_made":"yes","referee_comment":"§5.2 and §7: the conclusion that subspaces are safe 'up to the first 4 modes' is established on exactly two idealized domains, both with uniform thickness and no grounding line. The abstract and conclusion state this threshold as a general finding. Given that real ice sheet inverse problems involve grounding zones, variable thickness, and thermomechanical coupling, the generalizability of the '4 modes' threshold is a correctness-risk concern. The authors should add a brief statement in the conclusion acknowledging that the threshold is domain-specific and may shift under different flow regimes, or alternatively test a third domain with non-uniform thickness to strengthen the claim."}],"tokens_in":18390,"tokens_out":848,"duration_ms":94728,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper derives a PDE-level second-order adjoint for the shallow-stream approximation (SSA) that freezes the viscosity (the standard self-adjoint trick), then systematically compares the resulting approximate Hessian against an AD-computed exact one using spectral diagnostics. The finding is that the two Hessians agree well for the first ~4 eigenmodes and then diverge, with SOSA eigenvalues systematically smaller by a factor of ~3. The paper is honest about the limitations and does not oversell the method.","headline":"Paper derives a second-order self-adjoint (SOSA) approximation for SSA ice sheet Hessians and benchmarks it spectrally against AD. The derivation is clean, the comparison is rigorous, and the conclusions are appropriately cautious. The main limitation is generalizability from two idealized domains.","tokens_in":19433,"tokens_out":610,"would_cite":true,"duration_ms":33947,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Approximate ice-sheet Hessian breaks after 4 modes","keywords":["ice sheet modelling","adjoint methods","algorithmic differentiation","Hessian","shallow-stream approximation","uncertainty quantification","inverse problems","self-adjoint approximation"],"falsifier":"If, on a domain with different flow regime (e.g., strong shear or grounding-line dynamics), the eigenvector subspaces were to diverge immediately at mode 1 rather than mode 4, the 'first 4 modes are safe' conclusion would not generalize.","tokens_in":18703,"feed_emoji":"","tokens_out":1223,"duration_ms":85597,"temperature":0.7,"pith_summary":"The paper derives a PDE-level second-order adjoint method for the shallow-stream approximation (SSA) of ice sheet flow, called SOSA (second-order self-adjoint), which neglects the nonlinear dependence of ice viscosity on velocity. The motivation is practical: many ice sheet models lack algorithmic differentiation (AD) infrastructure, and SOSA can be implemented as a drop-in using existing forward-model stencils. The authors compare SOSA Hessian-vector products against exact AD-computed Hessians on two synthetic domains. The key finding is that while the two Hessians look superficially similar in direct comparisons, spectral analysis reveals their eigenvector subspaces align only for the first 4 modes (principal angles below π/8), then diverge and become persistently orthogonal after mode 33. Additionally, SOSA eigenvalues are systematically smaller than AD eigenvalues by a factor of approximately 3, attributable to the neglected viscosity nonlinearity (Glen's law exponent n=3). The authors conclude that SOSA is case-dependent in utility, adequate only at very low rank, and that a full Hessian computed via AD should be preferred when high fidelity is required.","feed_headline":"Approximate ice-sheet Hessian breaks after 4 modes","feed_subtitle":"A cheaper second-derivative shortcut for ice sheet models aligns with exact computation only in its first four eigenvectors, then diverges —","key_machinery":"The central objects are the SOSA (second-order self-adjoint) Hessian and the AD (algorithmic differentiation) Hessian for the SSA momentum balance. The comparison tool is principal angle analysis between eigenvector subspaces: the cosine of each principal angle is obtained from the SVD of the matrix formed by the inner products of the two sets of eigenvectors. Small angles mean the subspaces are nearly parallel; π/2 means orthogonal. The eigenvalue ratio of ~3 is traced to the Glen's flow law exponent n=3 entering the viscosity derivative.","core_discovery":"The central discovery is a quantitative characterization of how the self-adjoint approximation degrades second-order derivative information. The divergence is not catastrophic—the eigenvector subspaces start close and the divergence does not accelerate after approximately 50 modes—but it is real and structurally significant. The factor-of-3 eigenvalue compression traces directly to the Glen's law exponent, providing a mechanistic explanation for the systematic underestimation of curvature. The subspace alignment pattern (good for 4 modes, orthogonal by mode 33, stable thereafter) also retrospectively explains why first-order self-adjoint gradient methods have worked well in practice: there's","pith_inferences":["The 'first 4 modes are safe' threshold is established on only two synthetic domains. Domains with stronger shear margins, grounding-line proximity, or different flow regimes could shift this threshold in either direction—particularly if nonlinear viscosity terms contribute more to the Hessian structure in those settings.","The factor-of-3 eigenvalue ratio is tied to n=3 (Glen's law). If a different rheological exponent were used, the ratio would scale accordingly, suggesting the eigenvalue compression is a predictable structural feature rather than a domain-specific artifact.","The persistent orthogonal subspace after mode 33 suggests there exists a class of curvature directions that are entirely invisible to the self-adjoint approximation—directions driven by velocity-dependent viscosity feedbacks. Identifying what physical structures these eigenvectors correspond to could guide when the approximation is safe versus dangerous.","The methodology (principal angle analysis between approximate and exact Hessians) is generalizable beyond ice sheets to any PDE-constrained optimisation where a self-adjoint approximation is used, and could serve as a diagnostic tool for assessing approximation fidelity in other geophysical inverse problems."],"forward_implications":["Ice sheet models using SOSA for Newton-type optimisation should expect reliable curvature information only in the first 3-4 search directions, with progressively degraded directions beyond that.","For Bayesian uncertainty quantification requiring the inverse Hessian for posterior covariance, SOSA will systematically underestimate uncertainty in every direction by roughly a factor of 3, and the eigenvector misalignment beyond mode 4 makes faithful covariance reconstruction unlikely without thousands of modes.","The factor-of-n eigenvalue compression suggests a simple scaling correction could partially rescue SOSA for leading-mode applications, though the subspace divergence limits its effectiveness at higher rank.","The subspace alignment pattern provides a retrospective explanation for the empirical success of self-adjoint approximations in first-order ice sheet inverse problems: gradient methods explore curvature directions sequentially, and the first few are the ones that matter most."],"fun_headline_variants":["Approximate ice-sheet Hessian diverges after 4 modes","Cheap Hessian shortcut for ice sheets fails beyond 4 eigenvectors","Viscosity approximation compresses ice-sheet Hessian eigenvalues 3x","Ice-sheet Hessian subspaces go orthogonal by mode 33","Self-adjoint Hessian matches exact only at very low rank"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The self-adjoint approximation assumes that the tangent linear operator of the SSA residual is self-adjoint, which holds only when the vertically-averaged effective viscosity is treated as independent of ice velocity. This is false for Glen's flow law with n=3, where viscosity depends nonlinearly on the strain rate and hence on velocity. The entire SOSA derivation and the resulting spectral comparison depend on this linearization.","fun_headline_variants_meta":{"raw":{"variants":["Approximate ice-sheet Hessian diverges after 4 modes","Cheap Hessian shortcut for ice sheets fails beyond 4 eigenvectors","Viscosity approximation compresses ice-sheet Hessian eigenvalues 3x","Ice-sheet Hessian subspaces go orthogonal by mode 33","Self-adjoint Hessian matches exact only at very low rank"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":688,"prompt_tokens":598,"completion_tokens":90,"prompt_tokens_details":null},"tokens_in":598,"tokens_out":90,"duration_ms":19045,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-05T10:31:54.827759+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If, on a domain with different flow regime (e.g., strong shear or grounding-line dynamics), the eigenvector subspaces were to diverge immediately at mode 1 rather than mode 4, the 'first 4 modes are safe' conclusion would not generalize.","supporting_citations":[],"review_version":1}