{"id":"88aab23f-f300-4394-af89-03a3d91083d9","arxiv_id":"2501.19370","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"G-SVGD modifies SVGD with a loss-based line search and a KDE-based local Gaussian approximation, but the convergence claims are not backed by independent validation.","lead":"This paper proposes G-SVGD, a variant of Stein Variational Gradient Descent that picks an adaptive step size by minimizing a new loss combining gradient norms and the Evidence Lower Bound. It tests the method on one-dimensional wave-prospection inverse problems and claims faster convergence than standard SVGD and MCMC.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The loss in Eq. (35) is affine in ω and is never shown to measure posterior fidelity; minimizing it cannot support the acceleration/accuracy claim until an external posterior diagnostic is reported.","rationale":"I agree with the reader's weakest assumption: the paper's central claim depends on L being a posterior-fidelity objective, and §2.7 does not establish that. The ω-derivative argument is invalid because L is linear in ω; the ELBO formula (32) is inapplicable to a KDE test distribution; and the reported comparisons use L as both objective and evaluation metric, making the acceleration partly circular. The concern is not that the loss disagrees with common practice—it is that no mechanism, theorem, or experiment ties lower L to better posterior approximation. The external-diagnostic test would settle whether G-SVGD actually accelerates posterior inference or only minimizes its own loss. I therefore maintain the reader's REJECT verdict; no new objection beyond the reader's is needed.","tokens_in":20884,"tokens_out":4997,"duration_ms":50536,"concrete_test":"Re-run the high-contrast experiment with an external posterior diagnostic independent of L: after budget-matched numbers of forward solves, compare G-SVGD, ADAM-SVGD, and t-walk by kernel Stein discrepancy (or 2-Wasserstein distance to a long t-walk reference) of the particle sets. If G-SVGD is not better on this external metric, the claim that minimizing L accelerates convergence to the posterior fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract; §4.2) is that G-SVGD accelerates convergence and improves posterior approximation because it minimizes L(Θ,q) = (1−ω)∥∇θ log π(Θ|y)∥∞ − ω ELBO(q). The load-bearing premise is that L is a valid objective for posterior fidelity. This is not established, and the derivation offered in §2.7 is unsound. For fixed Θ,q, L is affine in ω; ∂²ω L = 0, so minimizing L over ω∈[0,1] occurs at an endpoint, not at the interior stationarity condition (38), and ω is a user-selected weight rather than an optimization variable. More importantly, minimizing the gradient-norm term drives particles to zeros of ∇θ log π, i.e., local modes or flat regions; it imposes no constraint on the shape of q. No argument in §2.7–§2.9 connects minimizers of L to the posterior or to a controlled Gaussian local approximation. Independently, Eq. (32) inserts the Gaussian KL formula into the ELBO; since q is a KDE and p is not generally Gaussian, the ELBO estimate in Eq. (34) is not a valid lower bound. Figures 9–11 then evaluate all methods with the same L that G-SVGD is explicitly minimizing, so lower values are partly circular and do not show posterior fidelity. The 200-samples versus 10,000-samples comparison in §4.2 also omits runtime and number of forward solves. Thus the acceleration claim rests on an unvalidated objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Greedy Stein Variational Gradient Descent (G-SVGD), a modification of SVGD for Bayesian inference in wave prospection problems. The core novelty is a loss function L(Θ,q) = (1−ω)∥∇θ log π(Θ|y)∥∞ − ω ELBO(q), used to select a global step size at each iteration and to define a stopping rule. The method is applied to two 1D wave-equation models (high-contrast and low-contrast stratified media), with derivatives computed by a five-point finite-difference operator and gradients of the log posterior via the adjoint method. Numerical comparisons are made against SVGD with ADAM and the t-walk MCMC sampler.","tokens_in":21193,"tokens_out":2617,"duration_ms":28902,"significance":"If the central claims were established, G-SVGD would be a valuable contribution to variational inference for PDE-constrained inverse problems, particularly for expensive forward solvers. The paper contains some useful components: a careful derivation of five-point finite-difference error bounds, a clear treatment of the high-contrast layered wave model with reflection/transmission coefficients, and transparent acknowledgement of algorithmic limitations in Section 4.4. However, the central algorithmic innovation—the loss function in Eq. (35)—is not shown to measure posterior fidelity, and the convergence claims in the abstract are not supported by any theorem. The numerical comparisons are also circular because all methods are evaluated on the same loss that G-SVGD is explicitly designed to minimize. These issues concern the core contribution, not peripheral presentation.","major_comments":[{"comment":"The derivation of the loss-function minimum is flawed. For fixed Θ and q, L is affine in ω, so ∂²ω L = 0 and the minimum over ω ∈ [0,1] occurs at an endpoint; the interior stationarity condition ∥∇x log(Θ,y)∥∞ = −ELBO(q) in Eq. (38) does not identify a minimum. More fundamentally, minimizing ∥∇x log(Θ,y)∥∞ drives particles to zeros of the log-posterior gradient, i.e., modes or flat regions, and imposes no constraint on the shape of q. No argument in §2.7–§2.9 connects minimizers of L to the posterior or to a controlled Gaussian local approximation, so the claim that G-SVGD samples the posterior is unsupported.","section":"§2.7, Eqs. (35)–(38)"},{"comment":"The ELBO is computed using the closed-form KL divergence between two Gaussians in Eq. (32), but q is a kernel density estimate and p is the posterior, which is not generally Gaussian. Therefore Eq. (34) is not a valid lower bound on the log marginal likelihood; it is an ad hoc plug-in estimator. Consequently, the ω = 1 case does not actually maximize the ELBO, and the reported log(−ELBO) comparisons in Figures 12–14 are not measuring what they claim to measure.","section":"§2.6, Eqs. (32)–(34)"},{"comment":"The convergence comparison evaluates G-SVGD, A-SVGD, and t-walk using the same loss function L that G-SVGD is explicitly designed to minimize. Lower values of L for G-SVGD are therefore partly by construction and do not demonstrate improved posterior accuracy. The manuscript reports no independent posterior diagnostic (e.g., posterior predictive checks, a reference posterior from long MCMC, or a two-sample test). Additionally, the claim in §4.2 that t-walk needs 10,000 samples while G-SVGD needs only 200 for 'similar accuracy' omits both wall-clock runtime and the number of forward PDE solves, so the computational-cost advantage is not established.","section":"§3.2.3, Figures 9–11"},{"comment":"The abstract and introduction claim 'theoretical guarantees of convergence,' but no convergence theorem is stated or proved. The inequality chain in Eq. (39) of §2.8 assumes the KL reduction result for SVGD and then substitutes it into the proposed loss; it does not prove that the proposed step-size rule converges to a posterior sample. The theoretical claim should either be substantiated with a precise statement and proof or removed.","section":"Abstract and §1"}],"minor_comments":[{"comment":"The boundary condition u(L,t) = −c_L ∂x u(0,t) in system (5) appears dimensionally inconsistent and does not match the absorbing boundary condition used in the numerical system (26); please clarify the correct boundary condition and notation.","section":"§2.4.2, Eq. (5) vs. Eq. (26)"},{"comment":"The notation KDE_{Θ,ω}(θ) is undefined: the KDE bandwidth is not specified and does not depend on ω in the description. Please specify the bandwidth selection rule and whether ω enters the KDE.","section":"§2.9, Algorithm 2"},{"comment":"Several typos and formatting issues: 'sufficientelly' (§2.5.1), 'mínimum' (§2.7), 'Gaussain' (§2.9), 'completly' and 'becuase' (§4.2), and the broken cross-reference in Figure 1. These should be corrected.","section":"Throughout"},{"comment":"The histograms in Figure 8 do not show the true posterior or a reference posterior, so it is difficult to assess which method is more accurate; adding a contour or density overlay of a long-run MCMC reference would help.","section":"§3.2.3, Fig. 8"},{"comment":"The limitations section acknowledges sensitivity to ω and dependence on KDE quality, but these limitations are not connected to the numerical claims; please discuss how the reported speedups would change under different ω and bandwidth choices.","section":"§4.4, Limitations"}],"recommendation":"reject","confidential_remarks":"The paper is not in a publishable state. The core loss function is not a valid posterior-fidelity objective, the ELBO computation is mathematically incorrect for non-Gaussian targets, and the experimental section compares methods on a metric that favors the proposed algorithm by construction. These are load-bearing issues that cannot be repaired with minor edits; the manuscript would need a fundamentally different objective, external posterior diagnostics, and runtime/forward-solve comparisons to support its claims. I would encourage the authors to consider these points carefully for a substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper proposes a plausible-sounding SVGD variant with a custom loss, but the loss does not do what the authors claim, and the numerical comparisons are set up in a way that makes the acceleration claim partly circular. I would not trust the main result as it stands.\n\nWhat is actually new: the G-SVGD algorithm is clearly specified (Algorithm 2) and the idea of using a sup-norm gradient term plus an ELBO term to steer particles is a legitimate thing to try. The line search in Eq. (29) and the loss in Eq. (35) are not in the cited SVGD or boosting papers. The five-point operator analysis in Section 2.3 is a real, if minor, contribution: they derive error bounds and show how many grid points per wavelength you need for cosine and exponential-decay waves, and they demonstrate better absorbing-boundary behavior than simple backward/centered differences. The citation practice is fine; they engage with the relevant SVGD, boosting VI, and t-walk literature.\n\nThe soft spots are serious. First, the loss function is never shown to measure posterior fidelity. Minimizing ∥∇ log π∥∞ pushes the worst particle toward a critical point of the log-posterior, which could be a saddle or a flat region; it says nothing about the shape of the distribution. The ELBO term could in principle fix that, but the ELBO is computed with Eq. (32), which is the KL divergence between two Gaussians, while q is a KDE and p is not Gaussian. So Eq. (34) is not a valid ELBO. Second, the derivation in §2.7 is mathematically off: L is affine in ω, so minimizing over ω∈[0,1] happens at an endpoint, not at the interior stationarity condition (38). The condition ∥∇ log π∥∞ = −ELBO just says the two terms cancel; it does not establish that the minimum of L corresponds to the posterior. Third, the convergence comparisons in Figures 9–11 evaluate all methods with the same L that G-SVGD is explicitly minimizing; lower L values for G-SVGD are largely by construction. No independent posterior diagnostic (e.g., distance to a long MCMC run, posterior predictive checks) is reported. Fourth, the 200-samples vs 10,000-samples claim in §4.2 omits runtime and number of forward solves, so the computational advantage is not demonstrated.\n\nWho is this for? Someone working on SVGD for expensive inverse problems might find the algorithm idea worth experimenting with, but they should re-derive the objective and add proper diagnostics. As it stands, the central claim is not supported. I would not cite it, and I would not send it to a serious referee without major revisions. It is more of a working note than a finished paper.","headline":"A clearly-specified SVGD variant whose custom loss is never shown to track posterior fidelity; the acceleration claim rests on circular comparisons.","tokens_in":21753,"tokens_out":4940,"would_cite":false,"duration_ms":46985,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes G-SVGD, a variational inference algorithm that appends a weighted gradient-norm term to the ELBO and optimizes the step size against the combined loss, claiming faster posterior approximation in wave prospection models.","keywords":["gradient descent","learning coefficient","bayesian model","surrogate model","full wave inversion","Stein variational gradient descent","G-SVGD","variational inference"],"falsifier":"Run G-SVGD, A-SVGD, and t-walk on a Bayesian model with a known Gaussian-mixture posterior and known true parameters; if G-SVGD's 200-particle output has substantially larger KL divergence to the true posterior than the t-walk output, for comparable wall-clock time, the claim that the loss drives convergence to the posterior would be contradicted. A cheaper check: on a unimodal Gaussian posterior, verify whether the stopping condition $\\lVert \\nabla \\log p\\rVert_\\infty = -\\mathrm{ELBO}$ holds at convergence while minimizing $L$ actually decreases KL divergence.","tokens_in":20647,"feed_emoji":"🌊","tokens_out":6195,"duration_ms":58189,"temperature":0.7,"pith_summary":"This paper proposes G-SVGD, a particle-based variational inference algorithm for Bayesian inverse problems, and tests it on one-dimensional wave prospection models. The central claim is that adding a weighted gradient-norm term to the Evidence Lower Bound, and choosing the step size by minimizing the resulting loss function, makes particles converge to a Gaussian local approximation of the posterior faster than SVGD with the ADAM optimizer, and with many fewer samples than the MCMC t-walk. The practical payoff is cheaper uncertainty quantification in settings where each gradient evaluation requires solving a wave equation. The paper also contributes a five-point finite-difference operator for the forward solver and error bounds that specify how many grid points per wavelength are needed.","feed_headline":"Greedy sampler claims 50x fewer samples for wave inversion","feed_subtitle":"By tuning step size with a gradient-ELBO loss, particles approximate the posterior far faster than SVGD-ADAM or t-walk.","key_machinery":"The load-bearing object is the loss function $L(\\Theta,q) = (1-\\omega)\\lVert \\nabla_x \\log(\\Theta,y)\\rVert_\\infty - \\omega\\,\\mathrm{ELBO}(q)$, which blends the supremum norm of the log-posterior gradient across particles with the evidence lower bound of a kernel density estimate $q$. Minimizing this loss over the step size $\\alpha$ at each iteration steers particles along the SVGD direction $\\phi$ while controlling the trade-off between mode-seeking ($\\omega$ near 0) and distribution-shape alignment ($\\omega$ near 1) through KDE. The five-point finite-difference operator and adjoint method enter as supporting machinery that makes the gradient evaluations accurate enough for the loss to be meaningful. The paper's convergence argument rests on the condition $\\lVert \\nabla_x \\log(\\Theta,y)\\rVert_\\infty = -\\mathrm{ELBO}(q)$ derived from minimizing $L$ with respect to $\\omega$, which it takes as the stopping criterion.","core_discovery":"On the paper's own terms, the discovery is an algorithm rather than a theorem: replacing SVGD's hand-set or per-parameter learning rate with a global step size $\\alpha^* = \\arg\\min_{0\\le \\alpha\\le \\xi} L(\\Phi(\\alpha), q_\\alpha)$, where $L(\\Theta,q) = (1-\\omega)\\lVert \\nabla_x \\log(\\Theta,y)\\rVert_\\infty - \\omega\\,\\mathrm{ELBO}(q)$, yields a Greedy Stein Variational Gradient Descent that drives particles toward local modes while approximating the local density with kernel density estimation. In the high-contrast experiment, G-SVGD reached a posterior approximation comparable to t-walk with 200 samples where t-walk needed 10,000, and it converged faster than SVGD-ADAM across $\\omega = \\{0, 0.5, 1\\}$. The paper interprets this as evidence that the loss function provides an effective stopping condition and step-size selection rule, and that the method is well suited to wave-prospection models where gradient evaluation is computationally expensive.","pith_inferences":["The condition $\\lVert \\nabla \\log p\\rVert_\\infty = -\\mathrm{ELBO}$ suggests a cheap convergence diagnostic, but the paper does not prove that this condition is sufficient for convergence to the target distribution; a toy experiment with a known posterior would clarify whether the balance point tracks KL divergence.","If the loss is used as a stopping rule, then any bias introduced by KDE smoothing, which the paper lists as a limitation, will leak directly into the inferred posterior width; comparing G-SVGD output with a Gaussian mixture of known covariance would quantify that bias.","The 50x sample reduction over t-walk is demonstrated in a 2-parameter high-contrast model; in the 15-parameter low-contrast spline model the reference methods were not run, so the efficiency claim is not yet established at higher dimension.","The same loss could be tested as a step-size selector for other particle-based samplers, since the mechanism is agnostic to the forward model; that would be a quick extension the paper does not pursue."],"forward_implications":["Gradient-expensive inverse problems would become tractable with far fewer forward solves: 200 particles versus 10,000 MCMC samples in the demonstrated high-contrast case.","Uncertainty quantification in wave prospection can be done with a global, adaptive step size and an explicit stopping rule, both of which the original SVGD leaves unspecified.","The weight $\\omega$ acts as a dial between mode-hunting and shape-matching, giving practitioners a practical trade-off for multimodal posteriors.","The five-point operator with its error bound and node-per-wavelength criterion gives a concrete recipe for discretizing the wave forward model accurately enough for gradient-based inference."],"supporting_citations":[{"why":"Supplies the SVGD update direction $\\phi$ and the particle transformation framework that G-SVGD modifies.","marker":"(Liu and Wang, 2016)"},{"why":"Provides ADAM, the adaptive learning-rate method used as the A-SVGD comparison baseline.","marker":"(Kingma and Ba, 2014)"},{"why":"Establishes ELBO-based variational inference, the objective component G-SVGD combines with the gradient norm.","marker":"(Blei et al., 2017)"},{"why":"Provides the t-walk MCMC sampler used as the reference for posterior approximation and sample-count comparison.","marker":"(Christen and Fox, 2010)"},{"why":"Proposes greedy particle optimization with per-particle step sizes, the closest prior approach G-SVGD contrasts with a global step size.","marker":"(Futami et al., 2019)"},{"why":"Supplies the randomized Halton sequence used to initialize the particle set near local modes.","marker":"(Owen, 2017)"},{"why":"Source of the first-order absorbing boundary conditions discretized by the five-point operator.","marker":"(Buckman et al., 2012)"},{"why":"Documents the five-point operator in computational seismology and motivates its use for derivative approximation in wave models.","marker":"(Igel, 2017)"}],"fun_headline_variants":["Greedy SVGD samples 50x faster for wave prospection","Greedy step size speeds Bayesian wave inversion 50x","Gradient-ELBO loss tunes SVGD for costly wave gradients","Algorithm slashes samples to 1/50 for wave posterior","G-SVGD: smarter step size for wave prospection Bayes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the loss $L(\\Theta,q)$ is a valid measure of how close the particle set is to the posterior, so that minimizing it drives particles to the posterior; the paper derives a balance condition on $\\omega$ but never connects $L$ to any divergence from the posterior, and if $L$ does not measure posterior fidelity the acceleration and accuracy claims lack a foundation.","fun_headline_variants_meta":{"raw":{"variants":["Greedy SVGD samples 50x faster for wave prospection","Greedy step size speeds Bayesian wave inversion 50x","Gradient-ELBO loss tunes SVGD for costly wave gradients","Algorithm slashes samples to 1/50 for wave posterior","G-SVGD: smarter step size for wave prospection Bayes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1391,"prompt_tokens":1020,"completion_tokens":371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":295}},"tokens_in":636,"tokens_out":371,"duration_ms":3834,"temperature":1.0,"reasoning_tokens":295,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:20:13.851793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run G-SVGD, A-SVGD, and t-walk on a Bayesian model with a known Gaussian-mixture posterior and known true parameters; if G-SVGD's 200-particle output has substantially larger KL divergence to the true posterior than the t-walk output, for comparable wall-clock time, the claim that the loss drives convergence to the posterior would be contradicted. A cheaper check: on a unimodal Gaussian posterior, verify whether the stopping condition $\\lVert \\nabla \\log p\\rVert_\\infty = -\\mathrm{ELBO}$ holds at convergence while minimizing $L$ actually decreases KL divergence.","supporting_citations":[{"cited_title":", author Kucukelbir, A","cited_arxiv_id":null,"evidence_quote":"Establishes ELBO-based variational inference, the objective component G-SVGD combines with the gradient norm."},{"cited_title":", author Fox, C","cited_arxiv_id":null,"evidence_quote":"Provides the t-walk MCMC sampler used as the reference for posterior approximation and sample-count comparison."},{"cited_title":", author Cui, Z","cited_arxiv_id":null,"evidence_quote":"Proposes greedy particle optimization with per-particle step sizes, the closest prior approach G-SVGD contrasts with a global step size."},{"cited_title":", author Sanchez-Vizuet, T","cited_arxiv_id":null,"evidence_quote":"Source of the first-order absorbing boundary conditions discretized by the five-point operator."},{"cited_title":", year 2017","cited_arxiv_id":null,"evidence_quote":"Documents the five-point operator in computational seismology and motivates its use for derivative approximation in wave models."}],"review_version":1}