{"id":"9cca285b-2f7e-4907-bbc9-202b055bc231","arxiv_id":"2411.14409","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A combined inexact and generalized Golub-Kahan method with hybrid regularization provides an iterative solver for generalized Tikhonov problems with inexact matrix-vector products.","lead":"This paper introduces a numerical method for large-scale Bayesian inverse problems when the forward model is only approximately known. The method adapts inexact Golub-Kahan iterations to generalized norms and tests it on simulated CT reconstructions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on the noisy projected problem (4.5) being a faithful surrogate for the exact Tikhonov problem, but the paper gives no condition on E_k and F_k and no perturbation bound linking the two solutions.","rationale":"The reader's weakest assumption is exactly the missing perturbation analysis: no condition on the MVP errors E_k and F_k is given, and Section 4.2 does not show that the projected problem (4.5) is a faithful surrogate for the true Tikhonov problem (2.2). My reading confirms this as the most load-bearing gap. The algebraic construction of the decomposition is internally consistent, and the numerical results are encouraging, so the concern does not overturn the paper; it does mean the central claim is not yet rigorously supported. The reader already assigned CONDITIONAL, and my analysis reinforces that verdict rather than changing it. I also note the residual used in the discrepancy principle is biased by the model error, which is a concrete manifestation of the same missing surrogate analysis. No ad hominem or overstatement is intended: the issue is a missing proof and missing robustness study, not an observed failure of the method.","tokens_in":13296,"tokens_out":7038,"duration_ms":67500,"concrete_test":"Use the Section 5.1 CT setup and run igenHyBR for inexactness levels eta in {1e-6, 1e-4, 1e-2, 1e-1}, with at least 20 random seeds per level. Record at each iteration (a) the true relative residual ||AQV_k y_k - b||_{R^{-1}}/||b||_{R^{-1}}, (b) the projected residual ||M_k y_k - beta e_1||_2/beta, and (c) the relative reconstruction error of igenHyBR versus genHyBR. If the gap between (a) and (b) grows with eta, or if the error gap between the inexact and exact methods does not decrease linearly with eta, then the missing perturbation condition is material. An analytic complement: derive a first-order bound for y_k(lambda) - y_k^exact(lambda) in terms of max_k ||E_k|| / sigma_min(M_k^T M_k + lambda^2 I) and verify it numerically.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's central assertion (end of Section 1, Figure 3) is that the inexact generalized hybrid method reproduces the accuracy of its exact counterpart. The load-bearing step is Section 4.2: equation (4.5) replaces the original generalized Tikhonov problem (2.2) by a projected problem built from the noisy relation (4.2), (A+E_k)QV_k = U_{k+1}M_k. This means the reduced objective is minimized with respect to the perturbed operator A+E_k, not the true A. Because E_k is only defined implicitly by the random MVP errors, there is no stated bound on ||E_k|| or ||F_k||, no perturbation analysis relating y_k from (4.5) to x_lambda from (2.2), and no control on how the inexactness propagates through the R^{-1}- and Q-orthogonalizations. All evidence is a single synthetic CT run. The concern is concrete: for a given level eta = max(||E_k||/||A||, ||F_k||/||A||), the surrogate error could be O(eta/lambda) or worse, which would invalidate the claim outside the tested range. The Section 4.3 discrepancy principle is also affected: it uses ||M_k y_k(lambda) - beta e_1||_2 as the residual, but the true residual norm ||A s_k(lambda) - d||_{R^{-1}} differs by a term involving E_k QV_k y_k, so the stopping rule is biased by the model error. Table 1 verifies A^T U_k = V_k L_k^T rather than the defining relation (4.2), so the numerical verification does not cover the identity that (4.5) relies on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an inexact generalized Golub-Kahan (igenGK) decomposition for solving large-scale Bayesian inverse problems when the forward operator is available only through approximate matrix-vector products. The method combines the inexact Golub-Kahan approach of Gazzola and Landman with the generalized Golub-Kahan bidiagonalization of Chung and Saibaba, and it avoids explicit factorizations of the prior covariance matrix Q by using only MVPs. A hybrid scheme with discrepancy principle and WGCV is developed for automatic regularization parameter selection. Numerical experiments on a CT reconstruction problem compare the method against standard, generalized, and inexact counterparts, and a separate experiment considers inexactness arising from inaccurate projection angles.","tokens_in":13557,"tokens_out":5616,"duration_ms":51662,"significance":"The paper addresses a relevant and timely problem: Bayesian inverse problems with an uncertain forward model. The algebraic derivation of the inexact generalized decomposition (Section 4.1) is straightforward but sound, and the ability to work with Q only through matrix-vector products is a practical strength. The numerical results on a synthetic CT problem, including a realistic angle-inexactness scenario, suggest that the method can be effective. However, the central claim—that the inexact generalized hybrid method achieves accuracy comparable to its exact counterpart—is supported only by a single numerical example, and the paper provides no perturbation analysis to establish when the noisy projected problem is a faithful surrogate for the true Tikhonov problem. There is also a structural issue in the core algorithm and a mismatch in the numerical verification. The work is an incremental combination of existing techniques, but it could be a useful contribution if these gaps are addressed.","major_comments":[{"comment":"The k=1 branch of Algorithm 4.1 is incomplete: it computes v_1 (line 3) but never computes u_2, the scalar m_{2,1}, or the updated entry in M_2, yet the return statement outputs U_{k+1} = U_2 and M_{k+1}. Consequently, the algorithm as written does not produce a well-defined basis for the first iteration. Please provide the missing initialization or restructure the loop so that the update for u_{k+1} is performed for k=1 as well.","section":"Section 4.1, Algorithm 4.1"},{"comment":"The projected problem (4.5) is derived using the inexact relation (A+E_k)QV_k = U_{k+1}M_k, but no bound is given relating the solution y_k of (4.5) to the exact solution x_λ of (2.2) in terms of ||E_k||, ||F_k||, and λ. The paper's main claim (end of Section 1) that the inexact hybrid method achieves accuracy comparable to its exact counterpart is therefore unsupported beyond the single tested example, and the method's reliability for larger inexactness levels is unclear. Please add a perturbation analysis or at least a precise discussion of the conditions on E_k and F_k under which the surrogate problem is a faithful approximation.","section":"Section 4.2, Eq. (4.5)"},{"comment":"The verification in Table 1 checks the identity A^T U_k = V_k L_k^T, but the decomposition derived in (4.2) is (A+F_{k+1})^T R^{-1} U_{k+1} = V_{k+1} L_{k+1}^T. The table tests a different relation: it omits the R^{-1} weighting and uses the exact transpose A^T rather than the inexact operator. As a result, the numerical verification does not actually validate the relation that the projected problem (4.5) relies on. Please verify the correct identity (4.2), using the inexact transpose and the appropriate inner products.","section":"Section 5.1, Table 1"},{"comment":"The residual in the discrepancy principle is defined as r_k(λ) = A s_k(λ) - d = M_k y_k(λ) - β_1 e_1. This equality is only valid if E_k = 0, because the second expression is the residual of the perturbed projected problem. Consequently, the DP stopping rule uses a residual that is biased by the model error, and the paper does not analyze how this bias affects the chosen regularization parameter. Please derive the relation between the true residual norm and the projected residual norm under the inexactness model, and discuss the implications for the DP criterion.","section":"Section 4.3, Eq. (4.8)"}],"minor_comments":[{"comment":"There are several typos: 'cooresponds' in the Figure 2 caption, 'correponds' in Section 5.1, and 'Matérn' is spelled inconsistently. Please proofread carefully.","section":"Throughout"},{"comment":"The recurrence for v_{k+1} uses α_{k+1} twice, which is inconsistent with the standard genGK relation and with (3.10). Please check the coefficients and correct the equation.","section":"Section 3.3, Eq. (3.9)"},{"comment":"The notation for E_k and F_k is confusing: E_k is used both for the per-iteration error matrix in (3.5) and for the accumulated error defined after (4.2). Please introduce distinct notation for these two objects.","section":"Section 4.1"},{"comment":"It is not clear how iGK is applied to the generalized problem (2.2); if it requires a transformation or preconditioning, please specify the exact implementation to allow reproducibility.","section":"Section 5.1"},{"comment":"The WGCV results are reported to be worse than DP and optimal, but the paper does not provide the chosen weight parameter or any tuning strategy; a brief discussion would help the reader assess the practical usefulness of WGCV in this setting.","section":"Section 5.3, Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like an undergraduate research report; the novelty is the combination of two existing methods, and the main contribution is incremental. The structural bug in Algorithm 4.1 and the absence of any perturbation analysis are serious gaps that prevent acceptance in the current form. The authors should also reconsider the verification in Table 1 and the bias in the DP residual. If these issues are fixed, the paper could become a modest but useful contribution to the literature on inexact Krylov methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2411.14409. It is a natural combination of two known algorithms—inexact Golub-Kahan (Gazzola–Landman) and generalized Golub-Kahan (Arioli; Chung–Saibaba)—yielding a hybrid method for large-scale Bayesian inverse problems when matrix-vector products with A and A^T are noisy. The core algebra is standard and the empirical behavior on the CT example is consistent with the claim that the inexact method tracks its exact counterpart. But the paper is not close to publishable in its current form: Algorithm 4.1 drops a branch, Table 1 verifies the wrong relations, and there is no analysis of how the MVP errors E_k and F_k affect the projected problem.\n\nThe genuinely new piece is the specific combination. The paper defines an inexact generalized Golub-Kahan decomposition that maintains R^{-1}- and Q-orthogonality under inexact MVPs, and builds a hybrid Tikhonov solver on top. Section 4.1 derives the decomposition cleanly. The angle-error experiment in Section 5.4 is a nice touch, showing the method tolerates decreasing inexactness, a realistic scenario.\n\nThe soft spots are real. First, the pseudocode in Algorithm 4.1 is incomplete: for k=1 it never computes u_2 or updates M, yet returns U_{k+1} and M_{k+1}. A reader cannot reproduce the method. Second, Table 1 checks (5.1) A^T U_k = V_k L_k^T and (5.2) A Q V_k = U_{k+1} M_k, but the defining relations (4.2) are (A+F_{k+1})^T R^{-1} U_{k+1} = V_{k+1} L^T and (A+E_k) Q V_k = U_{k+1} M_k. The numerical verification does not test the relations the method actually relies on—a load-bearing gap. Third, the stress-test concern holds up: the projected problem (4.5) is solved with the noisy M_k, but there is no bound on ||E_k|| or ||F_k|| and no perturbation result relating the noisy solution to the exact Tikhonov solution. The discrepancy principle (4.8) also uses the projected residual, which is biased by model error. None of this is fatal if the method is offered as a heuristic, but the paper claims accuracy \"comparable to its exact counterpart\" on the strength of one synthetic experiment. That is overclaiming. Experimental details are also thin: no error bars over random inexactness, no information on how Q is applied, and only one phantom. Typos like \"this projects aims\" suggest hasty preparation.\n\nThis paper is for readers working on hybrid projection methods for Bayesian inverse problems who want an initial attempt at combining inexact and generalized GK. It deserves a serious referee because the idea is reasonable and the gaps are fixable, but it needs major revision: fix the algorithm, redo the verification against the correct relations, add a perturbation or at least a heuristic error bound, and broaden the experiments. I would not cite it in its current form. If it goes to review, I would recommend major revision rather than rejection.","headline":"A plausible but underdeveloped combination of inexact and generalized Golub-Kahan, with fixable pseudocode bugs and a missing perturbation analysis; worth a round of peer review if the authors can tighten it.","tokens_in":14234,"tokens_out":3131,"would_cite":false,"duration_ms":28670,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F22","65F10","65K10","15A29"],"pacs":[],"model":"deepseek-v4-flash","headline":"An inexact generalized Golub-Kahan solver matches exact-method accuracy for Bayesian inverse problems with uncertain forward models.","keywords":["Bayesian inverse problems","generalized Tikhonov regularization","inexact Golub-Kahan decomposition","hybrid projection methods","Krylov subspace methods","Matern covariance kernel","computed tomography reconstruction","regularization parameter selection"],"falsifier":"Take a small matrix $A$, a covariance $Q$, and a known true solution; run the proposed igenGK hybrid method with controlled deterministic error matrices $E_k$ and $F_k$ of increasing norm and compare its reconstruction to the exact generalized Golub-Kahan method on the same problem. If a moderate error level (say $\\|E_k\\|/\\|A\\|$ around $10^{-2}$) yields relative reconstruction error substantially worse than the exact counterpart, or if a non-random error pattern of the same size causes divergence where random-error experiments stayed stable, the paper's central claim is falsified.","tokens_in":12943,"feed_emoji":"🩻","tokens_out":8625,"duration_ms":74350,"temperature":0.7,"pith_summary":"Bayesian inverse problems become hard in two realistic ways: the forward operator may only be known approximately, and the prior covariance matrix may be too large or irregular to factor. This paper develops an inexact generalized Golub-Kahan (igenGK) decomposition that accepts matrix-vector products with $A$ and $A^T$ that carry errors, and it uses the covariance matrix $Q$ only through matrix-vector products, never through its square root or inverse. On top of this decomposition the paper builds a hybrid iterative solver that projects the generalized Tikhonov problem into a small least-squares problem and chooses the regularization parameter automatically at each iteration. The central claim, supported by computed-tomography experiments, is that this inexact hybrid method reaches reconstruction accuracy comparable to methods that assume the exact forward operator.","feed_headline":"Inexact forward models still yield high-accuracy CT reconstruction","feed_subtitle":"A Golub-Kahan hybrid solver handles uncertain forward operators and picks Tikhonov parameters automatically.","key_machinery":"The carrying object is the inexact generalized Golub-Kahan (igenGK) decomposition: an iterative two-sided orthogonalization that combines the inexact Golub-Kahan idea with the generalized Golub-Kahan idea. At each step it multiplies by $(A+E_k)Q$ and by $(A+F_{k+1})^T R^{-1}$, orthogonalizes against the existing basis with respect to the $R^{-1}$ and $Q$ inner products, and records the coefficients in an upper Hessenberg matrix $M_k$ and a lower triangular matrix $L_{k+1}$. These two matrices convert the original large problem into a small projected Tikhonov problem, so the algorithm can be driven by standard parameter-choice rules on the projected problem.","core_discovery":"After the change of variables $x = Q^{-1}(s - \\mu)$, the MAP estimate becomes a generalized Tikhonov problem in which the forward operator acts as $AQ$. The paper's central object is a decomposition that, at iteration $k$, produces vectors $U_{k+1}$, $V_k$ and matrices $M_k$, $L_{k+1}$ satisfying $(A+E_k)QV_k = U_{k+1}M_k$ and $(A+F_{k+1})^T R^{-1}U_{k+1} = V_{k+1}L_{k+1}^T$, with $U$ and $V$ orthonormal in the $R^{-1}$ and $Q$ inner products respectively. Because the projected solution is found from the small problem $\\min_y \\|M_k y - \\beta e_1\\|_2^2 + \\lambda^2 \\|y\\|_2^2$, the method never needs to form or factor $Q$, and it absorbs per-iteration errors $E_k$, $F_k$ in the forward-model products. The paper reports that the orthogonality conditions hold to machine precision and that the resulting reconstruction errors track the exact generalized method closely, including in a CT experiment where the inexactness comes from uncertain projection angles rather than synthetic random noise.","pith_inferences":["A natural next step, not taken in the paper, is to prove a perturbation bound showing that the noisy projected problem stays close to the exact one when $\\|E_k\\|$ and $\\|F_k\\|$ are small; the paper's evidence is entirely empirical.","The decreasing-inexactness experiment suggests a self-correcting loop: use the current reconstruction to improve estimates of uncertain parameters such as projection angles, then rerun the inexact solve with smaller errors.","The same decomposition should apply to other operator-uncertainty problems such as blind deblurring or super-resolution, where the error is structured model mismatch rather than Gaussian random noise."],"forward_implications":["The method solves the generalized Tikhonov MAP problem when the forward operator is available only through noisy matrix-vector products, so no exact $A$ or factorization of $Q$ is required.","In the computed-tomography experiments, the inexact generalized hybrid method's relative reconstruction error stays close to the exact genGK method at the same iterations.","Using the discrepancy principle for the regularization parameter gives results similar to using the optimal parameter, without needing the true solution.","The method remains stable when inexactness comes from realistic uncertain projection angles, provided the initial angle error is not too large.","Because $Q$ is accessed only via matrix-vector products, smooth stationary covariance kernels whose products can be accelerated (for example by Fourier methods) can be used directly."],"supporting_citations":[{"why":"supplies the inexact Golub-Kahan decomposition whose structure the proposed method generalizes.","marker":"[11]"},{"why":"provides the generalized Golub-Kahan bidiagonalization and hybrid projection framework for Bayesian inverse problems that this work extends.","marker":"[8]"},{"why":"motivates treating inexact matrix-vector products in Krylov subspace methods.","marker":"[25]"},{"why":"surveys hybrid projection methods, the basis for the automatic regularization parameter selection used here.","marker":"[3]"},{"why":"provides the parallel-beam computed-tomography test problem used in all numerical experiments.","marker":"[10]"},{"why":"contributes theory for generalized Golub-Kahan bidiagonalization and stopping criteria.","marker":"[1]"},{"why":"introduces the classical Golub-Kahan bidiagonalization underlying the exact and inexact variants.","marker":"[13]"},{"why":"defines the weighted generalized cross-validation parameter choice tested in the experiments.","marker":"[6]"}],"fun_headline_variants":["CT reconstruction robust to imperfect forward operators","Uncertain projection angles? Still get accurate CT scans","Golub-Kahan method tames inexact forward operators","Automatic Tikhonov selection for large-scale inverse problems","Inexact forward model? Hybrid solver picks parameters automatically"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if the errors in the approximate matrix-vector products are small enough that the reduced problem built from those approximate products still represents the original reconstruction problem; the paper gives no explicit bound on this, so this fidelity is the load-bearing assumption.","fun_headline_variants_meta":{"raw":{"variants":["CT reconstruction robust to imperfect forward operators","Uncertain projection angles? Still get accurate CT scans","Golub-Kahan method tames inexact forward operators","Automatic Tikhonov selection for large-scale inverse problems","Inexact forward model? Hybrid solver picks parameters automatically"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001214,"raw_usage":{"total_tokens":5007,"prompt_tokens":969,"completion_tokens":4038,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":3961}},"tokens_in":585,"tokens_out":4038,"duration_ms":31789,"temperature":1.0,"reasoning_tokens":3961,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:12:55.797020+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small matrix $A$, a covariance $Q$, and a known true solution; run the proposed igenGK hybrid method with controlled deterministic error matrices $E_k$ and $F_k$ of increasing norm and compare its reconstruction to the exact generalized Golub-Kahan method on the same problem. If a moderate error level (say $\\|E_k\\|/\\|A\\|$ around $10^{-2}$) yields relative reconstruction error substantially worse than the exact counterpart, or if a non-random error pattern of the same size causes divergence where random-error experiments stayed stable, the paper's central claim is falsified.","supporting_citations":[{"cited_title":"Gazzola and M","cited_arxiv_id":null,"evidence_quote":"supplies the inexact Golub-Kahan decomposition whose structure the proposed method generalizes."},{"cited_title":"Chung and A","cited_arxiv_id":null,"evidence_quote":"provides the generalized Golub-Kahan bidiagonalization and hybrid projection framework for Bayesian inverse problems that this work extends."},{"cited_title":"Simoncini and D","cited_arxiv_id":null,"evidence_quote":"motivates treating inexact matrix-vector products in Krylov subspace methods."},{"cited_title":"Chung and S","cited_arxiv_id":null,"evidence_quote":"surveys hybrid projection methods, the basis for the automatic regularization parameter selection used here."},{"cited_title":"Golub and W","cited_arxiv_id":null,"evidence_quote":"introduces the classical Golub-Kahan bidiagonalization underlying the exact and inexact variants."},{"cited_title":"Chung, J","cited_arxiv_id":null,"evidence_quote":"defines the weighted generalized cross-validation parameter choice tested in the experiments."}],"review_version":1}