{"id":"20f5ffee-9205-469f-8a09-2f55fc3850d1","arxiv_id":"2507.11130","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An adaptive reduced-basis trust-region Gauss-Newton method is extended to parabolic parameter identification, with POD-based enrichment, achieving 5 to 18x speedups in four reaction-diffusion tests.","lead":"This paper extends an accelerated inversion method for elliptic equations to time-dependent parabolic equations, reconstructing reaction and diffusion fields from noisy data. The approach uses adaptively built low-dimensional surrogate models inside a trust-region Gauss-Newton iteration, reporting 5 to 18 times speed-ups over the full-order method in four test problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No-stagnation is not merely unproven: after a (4.34) rejection the algorithm shrinks η while keeping the spaces and q_r fixed, so the new trust region is a subset of the one just searched; shrinking cannot create a trial satisfying (4.34), so the certified-termination claim is structurally…","rationale":"The reader correctly identified Remark 4.1 as the weakest link. My stress test sharpens that concern: the rejection handling in Algorithm 2 makes the situation worse than a mere absence of proof. Since T(η) shrinks monotonically with η and neither q_r nor the reduced spaces change upon rejection, there is no mechanism by which repeating the same subproblem with a smaller trust region can create a point satisfying (4.34) if none existed before. This is a structural flaw in the certification claim, not just a missing convergence theorem. The numerical experiments are encouraging and the code is a real asset, but they only show that stagnation did not occur in four hand-picked scenarios; they do not discharge the assumption. The verdict should remain CONDITIONAL: the authors need to either prove termination (or a suitable substitute) for a relevant problem class or explicitly soften the certification wording, and provide repeated-run statistics.","tokens_in":28262,"tokens_out":14194,"duration_ms":194958,"concrete_test":"On a coarse version of Run 1 (e.g., 20×20 grid, K=10), instrument the released implementation to log, for every rejected trial, the triple (η, ||q_trial−q_r||, Jh(q_trial)−Jr(q_AGC)). Run with η(0)=10^-1,10^-2,10^-3,10^-4. If the signed gap stays positive and the number of rejections does not decrease as η shrinks, the shrink-on-reject loop cannot terminate; that would confirm the no-stagnation assumption is not an artifact of the tested parameters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Remark 4.1 concedes that the inner IRGNM may never find a trial q_trial satisfying (4.34), and Algorithm 2's only response to rejection is to keep q(i+1)=q(i)_r, keep the reduced spaces, shrink η to β3η, and continue (lines 25-27). Because T(η) is nested, if no point in T(η) satisfies (4.34), then no point in the smaller T(β3η) satisfies it either. The shrink rule cannot manufacture an acceptable trial; it can only exclude candidates that were already unavailable. The situation is made worse by condition (4.29), which fixes the comparison value Jr(q_AGC)<Jh(q_r): as the allowed step shrinks, q_trial tends toward q_r, and Jh(q_trial) tends to Jh(q_r), so the required inequality Jh(q_trial) ≤ Jr(q_AGC) becomes harder, not easier, to meet. Thus finite termination is not a harmless missing proof: the prescribed repair is the wrong direction. The reported runs never hit this case, but the abstract's claim that the framework 'certifies' reliability depends on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TR-IRGNM, an adaptive reduced-basis trust-region variant of the iteratively regularized Gauss-Newton method for parameter identification in parabolic PDEs. It extends the authors' earlier elliptic method [44] by combining simultaneous state and parameter space reduction, POD-based adaptive enrichment, and a residual-based a posteriori error estimator that defines an error-aware trust region. Numerical experiments on four reaction-diffusion scenarios (stationary and time-dependent reaction/diffusion fields) report speedups between 4.66 and 17.89 over a full-order IRGNM benchmark, with reconstructed parameters matching the full-order reconstructions to within 0.1% to 16% relative error. The code is released for reproducibility.","tokens_in":28620,"tokens_out":7035,"duration_ms":81669,"significance":"If the identified gaps are resolved, the method is a practically useful contribution to large-scale dynamic inverse problems: it demonstrates that combined parameter/state reduction can be made adaptive and local inside a Gauss-Newton iteration for parabolic problems, and it provides systematic numerical evidence across four scenarios. Strengths include the independent validation against FOM-IRGNM, the sweep over POD tolerances, the release of reproducible code [55], and the clear discussion of computational bottlenecks such as O(NV) residual evaluations. However, the paper's central claim of a 'certified' framework is currently stronger than its theoretical support: the error estimator is imported with an unproven omission, and termination of the trust-region loop is assumed rather than guaranteed. The numerical evidence is persuasive but does not by itself establish the reliability assertions in the abstract.","major_comments":[{"comment":"The algorithm's termination is not guaranteed. Remark 4.1 concedes that a trial point satisfying (4.34) may never be found, and the only response to rejection in Algorithm 2 (lines 25–27) is to keep q_r and the reduced spaces fixed and shrink η to β3η. Since the trust regions T(η) are nested, any point in T(β3η) was already in T(η), and condition (4.34) compares Jh(q_trial) with Jr(q_AGC); as η shrinks, both q_trial and q_AGC tend to q_r, so (4.34) tends to the false inequality Jh(q_r) ≤ Jr(q_r), which is false by (4.29). Thus the shrink rule cannot, by itself, manufacture an acceptable trial. The empirical observation that stagnation never occurred does not support the abstract's claim that the framework 'certifies' reliability. Please either prove finite termination under explicit assumptions, modify the rejection step (e.g., enrich the spaces and recompute the AGC), or restrict the certification claim accordingly.","section":"§4.1, Remark 4.1 and Algorithm 2, lines 17–27"},{"comment":"The proof of the a posteriori error estimator is delegated to [53, Lemma 8, Theorem 9] with the statement that 'the last term ... can be omitted, because the reduced primal ur and adjoint states pr are elements of the same space V_r^K'. This justification is not substantiated: the omitted term in the cited theorem couples the primal and adjoint residuals, and co-membership in V_r^K does not by itself eliminate that coupling. Because this estimator defines the trust-region radius and underpins the sufficient/necessary tests (4.35)–(4.36), please provide a complete, self-contained proof or quote the precise theorem with a rigorous derivation of the omission.","section":"§3.2.1, Proposition 3.2"},{"comment":"When q_trial_r = q_AGC_r, the algorithm accepts the AGC as the next iterate without checking (4.34). Condition (4.29) only ensures Jr(q_AGC) < Jh(q_r), which does not imply Jh(q_AGC) ≤ Jr(q_AGC). Therefore the monotone decrease Jh(q(i+1)) < Jh(q(i)) asserted after (4.34) is not established for this branch. Please verify (4.34) before accepting the AGC, or prove that the construction of the AGC makes this condition automatic.","section":"§4.1, Algorithm 2, lines 10–12"}],"minor_comments":[{"comment":"The assignment q(i+1)_r = q(i)_AGC,r appears to be a typo; it should presumably be q(i)_trial,r to match the surrounding acceptance logic and the description in the text.","section":"Algorithm 2, line 21"},{"comment":"The while condition uses q(i,l)_r, but i is the outer iteration index and l is the inner IRGNM index; the condition should refer to the current outer iterate, e.g., q(i)_r, for consistency.","section":"Algorithm 2, line 2"},{"comment":"There are several typos: 'Additonally, it is has to be verfied' and 'startegies' should be corrected.","section":"§4.1 and §4.2"},{"comment":"The trust region T(i) is defined as a subset of QK,(i)_ad,r ⊂ Qad, but Qad is the infinite-dimensional admissible set; for consistency with the discrete setting it should be QK_ad,h or the appropriate discrete admissible set.","section":"§4, definition of T(i)"}],"recommendation":"major_revision","confidential_remarks":"The no-stagnation issue in Remark 4.1 is the most serious concern; if the authors cannot prove termination or modify the algorithm, the paper's certification claims would need to be substantially weakened. The Proposition 3.2 gap appears fixable with a detailed proof. I recommend major revision rather than rejection because the numerical evidence is strong and the central algorithmic idea is sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does what it says: it extends the adaptive reduced-basis trust-region IRGNM from elliptic to parabolic parameter identification, adds truncated POD to the enrichment step, and ships open-source code. The numerical validation is believable: speedups of 4.7–17.9x over FOM-IRGNM across four reaction–diffusion scenarios, with reconstructed parameters reasonably close to the full-order results. That is a real contribution, and the implementation makes it easy to reproduce.\n\nWhat is genuinely new is the parabolic setting itself, the combined state/parameter reduction with POD-based safeguards, and the explicit treatment of the admissible set. The paper is honest about its limitations, which is to its credit.\n\nThe soft spots are real but not fatal. First, the proof of Proposition 3.2 is delegated to Qian et al. and the omission of a term is asserted without derivation. That may be fixable, but as written it is a gap in a central error-estimation claim. Second, and more seriously, Remark 4.1 concedes that a trial satisfying the sufficient decrease condition (4.34) may never be found. The stress-test is right: after a rejection, the algorithm keeps the spaces and q_r fixed, shrinks eta, and searches a subset of the trust region it just searched. Shrinking cannot manufacture an acceptable point; if anything, as the allowed step tends toward q_r, condition (4.34) becomes harder, not easier. So the algorithm's own repair rule is structurally the wrong direction. The authors say stagnation was never observed, and that is honest, but it means the word \"certifies\" in the abstract is stronger than the theory supports. Third, the experiments are single runs per tolerance with no repeated-noise statistics; minor, but easy to add.\n\nWho is this for? Numerical analysts and practitioners working on reduced-order methods for time-dependent inverse problems. The method is likely to be useful, and the code is a solid starting point.\n\nRecommendation: send it to peer review. A serious referee should push on the termination question and the error-estimator proof. The authors should either prove a no-stagnation result for a relevant subclass, or redesign the rejection step—for example, by enriching the bases before shrinking the trust region—and soften the certification language. After those revisions I would be comfortable relying on the method.","headline":"Useful, reproducible extension of elliptic TR-IRGNM to parabolic inverse problems, but the trust-region rejection step cannot guarantee termination — shrinking the region cannot create an acceptable trial.","tokens_in":29079,"tokens_out":2795,"would_cite":true,"duration_ms":36260,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35R30","35K90","65M32","35K57"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that adaptively constructed reduced parameter and state spaces, embedded in an error-aware trust region, can accelerate IRGNM for parabolic inverse problems by factors of 4.66 to 17.89 in benchmark reaction-diffusion…","keywords":["parameter identification","model reduction","inverse problems","parabolic PDEs","Gauss-Newton methods","trust-region methods","proper orthogonal decomposition","reduced basis methods"],"falsifier":"Run Algorithm 2 on a parabolic parameter identification problem and record at each outer iteration whether a trial parameter satisfying condition (4.34) is found; an instance in which the trust-region radius keeps shrinking and no accepted iterate is produced, or in which the outer loop fails to terminate, would refute the method's reliability claim. A milder quantitative check is to vary random initializations, noise draws, and $\\epsilon_{\\mathrm{POD}}$ values and look for unbounded growth in rejected trial steps or stagnation of the full-order objective.","tokens_in":28067,"feed_emoji":"⚡","tokens_out":8214,"duration_ms":88543,"temperature":0.7,"pith_summary":"The paper proposes a way to make iteratively regularized Gauss–Newton (IRGNM) much cheaper for recovering unknown coefficients in parabolic partial differential equations from noisy data. Instead of solving the full-order state and adjoint equations at every trial, the algorithm builds low-dimensional reduced parameter and state spaces adaptively from snapshots, compresses them with proper orthogonal decomposition, and solves each IRGNM subproblem inside an error-aware trust region that checks the reduced model's accuracy before accepting a step. On four reaction-diffusion benchmark problems covering stationary and time-dependent reaction and diffusion fields, the resulting TR-IRGNM runs 4.66 to 17.89 times faster than the full-order IRGNM while matching the full-order reconstructions to within roughly 0.1% to 16% relative error. A sympathetic reader would take the paper's main contribution to be a practical, tunable speedup for parabolic parameter identification that keeps the regularization structure of IRGNM intact.","feed_headline":"Reduced-basis trust region fits PDE parameters up to 18x faster","feed_subtitle":"A certified adaptive reduced-order loop cuts full-order solves while keeping reconstructions within 0.1-16 percent.","key_machinery":"The load-bearing object is the error-aware trust-region reduced-basis loop. At each outer iteration a reduced parameter space $Q_r^{(i)}$ and a reduced state space $V_r^{(i)}$ are constructed by POD-truncated enrichment from full-order gradients, primal states, and adjoint states, and the IRGNM subproblem is solved only within a trust region defined by the ratio of a residual-based a posteriori error estimator to the reduced objective value. This loop does two jobs: it certifies that the reduced model is accurate enough at accepted steps through the sufficient-decrease and error-estimator checks, and it keeps the reduced bases minimal by adding only the POD modes needed to meet a chosen tolerance $\\epsilon_{\\mathrm{POD}}$.","core_discovery":"The central claim is that the computational bottleneck of IRGNM for parabolic inverse problems—the repeated high-dimensional solution of linearized primal and adjoint problems—can be overcome by a combined reduction of the parameter and state spaces, and that the resulting reduced surrogate can be made reliable through an error-aware trust-region strategy. The algorithm TR-IRGNM enriches the reduced parameter space with POD-compressed full-order gradients and the reduced state space with POD-compressed primal and adjoint trajectories, solves each local subproblem by projected gradient descent in the reduced admissible set, and only accepts a trial step after a sufficient-decrease check in the full-order objective supported by residual-based a posteriori error estimates. In the numerical experiments the reconstructed reaction and diffusion fields differ from the full-order reconstructions by about 0.1% to 16% relative error, while the total computation time drops by factors between 4.66 and 17.89 and the number of full-order solves decreases sharply.","pith_inferences":["Beyond the paper's experiments, the same combined reduction strategy should transfer to other affine-parameter parabolic forward operators, such as convection-diffusion or linear wave equations, provided adjoint solutions and residual estimators are available.","A natural next step the paper leaves open is to couple the POD tolerance, the trust-region radius, and the reduced-space dimension to the noise level $\\delta$, which would turn the empirical reliability into a convergence result as $\\delta \\to 0$.","The observed dependence of speedup and accuracy on $\\epsilon_{\\mathrm{POD}}$ suggests an automatic tolerance-selection rule: choose the smallest tolerance for which the estimated time to completion still beats the full-order method.","Because the diffusion reconstructions smooth sharp material interfaces, an extension with spatially adaptive basis enrichment would be a worthwhile stress test for the method."],"forward_implications":["If the reported behavior holds beyond the four benchmarks, parabolic parameter identification can be run at a fraction of full-order cost while preserving the discrepancy-principle stopping and the qualitative reconstruction quality.","The adaptive enrichment strategy removes the need for a globally accurate reduced model; only locally accurate surrogates along the IRGNM path are required.","For time-dependent parameter fields the time-dependent gradients provide 50 snapshots per enrichment, so the reduced parameter space becomes expressive quickly and the outer iteration count drops (three instead of twelve in the reaction test).","Tightening the POD tolerance $\\epsilon_{\\mathrm{POD}}$ trades slightly larger reduced spaces for smaller reconstruction error, giving a single tuning knob for the accuracy-speed balance.","The largest reported speedup, 17.89, occurs for time-dependent diffusion reconstruction, indicating that the method helps most where the full-order IRGNM is most expensive."],"supporting_citations":[{"why":"The elliptic predecessor whose combined parameter-and-state reduction and adaptive trust-region framework this paper extends to parabolic problems.","marker":"[44]"},{"why":"Supplies the convergence and regularization theory for the full-order IRGNM that TR-IRGNM accelerates.","marker":"[24]"},{"why":"Introduces the error-aware trust-region model-order-reduction framework whose acceptance criteria the algorithm adopts.","marker":"[26]"},{"why":"Provides the adaptive trust-region reduced-basis machinery and error estimators for PDE-constrained optimization that this work adapts.","marker":"[28]"},{"why":"Gives the certified trust-region reduced-basis method whose residual-based a posteriori bounds underlie Proposition 3.2.","marker":"[53]"},{"why":"Defines hierarchical approximate POD, the truncation algorithm used to keep the enriched reduced bases minimal.","marker":"[43]"},{"why":"The open-source model reduction library used to implement the numerical experiments and reproducibility artifacts.","marker":"[56]"}],"fun_headline_variants":["Adaptive reduced-basis trust region speeds parabolic inverse problems up to 18x","Certified reduced-order trust region reconstructs PDE fields with up to 18x less compute","Error-aware trust region + reduced basis cuts full-order solves in parabolic PDEs","Certified adaptive reduced basis trust region accelerates parabolic inverse problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the optimization never stagnates: the algorithm assumes that at every outer iteration a trial reduced parameter satisfying the sufficient-decrease condition (4.34) can be found, while Remark 4.1 concedes that this is not guaranteed and reports only that it never failed in the numerical experiments.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive reduced-basis trust region speeds parabolic inverse problems up to 18x","Certified reduced-order trust region reconstructs PDE fields with up to 18x less compute","Error-aware trust region + reduced basis cuts full-order solves in parabolic PDEs","Certified adaptive reduced basis trust region accelerates parabolic inverse problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001189,"raw_usage":{"total_tokens":4907,"prompt_tokens":943,"completion_tokens":3964,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":3881}},"tokens_in":559,"tokens_out":3964,"duration_ms":30035,"temperature":1.0,"reasoning_tokens":3881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:15:47.022575+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 on a parabolic parameter identification problem and record at each outer iteration whether a trial parameter satisfying condition (4.34) is found; an instance in which the trust-region radius keeps shrinking and no accepted iterate is produced, or in which the outer loop fails to terminate, would refute the method's reliability claim. A milder quantitative check is to vary random initializations, noise draws, and $\\epsilon_{\\mathrm{POD}}$ values and look for unbounded growth in rejected trial steps or stagnation of the full-order objective.","supporting_citations":[{"cited_title":"Computational Science and Engineering 1(3) (2024) https://doi.org/10.1007/ s44207-024-00002-z","cited_arxiv_id":null,"evidence_quote":"The elliptic predecessor whose combined parameter-and-state reduction and adaptive trust-region framework this paper extends to parabolic problems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the certified trust-region reduced-basis method whose residual-based a posteriori bounds underlie Proposition 3.2."}],"review_version":1}