{"id":"75510cd3-64ef-4c70-85a3-546ea52dad0f","arxiv_id":"2411.12430","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A tensor-train solver for the entropy-regularized JKO proximal step approximates Bayesian posteriors using only pointwise density evaluations, matching or beating Metropolis-Hastings in several tests.","lead":"This paper develops a numerical method that approximates high-dimensional probability distributions by solving a regularized optimal-transport flow on a fixed grid, using tensor-train compression to keep the problem tractable. It targets Bayesian inverse problems where the posterior density can be evaluated but its gradient is unavailable, and reports sample quality comparable to or better than Metropolis-Hastings MCMC with the same number of expensive model evaluations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The low-rank TT assumption is the load-bearing condition for the headline claim, and the reported experiments only exercise targets that are low-rank by construction, leaving the central tractability claim untested where the assumption breaks.","rationale":"The reader's weakest_assumption identifies the same load-bearing condition: the TT ranks of the posterior and intermediate entropic potentials must be small for the Eulerian discretization to be tractable. I agree with that assessment. The mathematical derivation of the regularized JKO step and the Hopf-Cole fixed-point formulation is coherent, and the reported experiments support the claim on the specific test problems. However, the experiments are all on targets whose low-rank structure is either exact or nearly exact, so they do not stress the assumption that makes the method work. This is precisely the condition under which the central claim could fail: if ranks are not small, the O(dNr^2) storage and evaluation costs make the method no longer competitive with MCMC at equal posterior-call budgets. The paper explicitly acknowledges the assumption in Section 1.3, which is honest but does not reduce its load-bearing role. The reader's CONDITIONAL verdict remains appropriate; my read does not move the verdict. I would only sharpen the requested revision: include at least one experiment where the target is deliberately chosen to have large TT ranks and report the observed ranks and reconstruction errors, so the scope of the claim is made explicit.","tokens_in":27571,"tokens_out":10607,"duration_ms":117155,"concrete_test":"Apply the method to a d=10 target whose TT ranks are not small by construction, e.g. ρ∞ ∝ exp(−½ x^T A x) with A = I + 11^T (all-to-all coupling) or A = R diag(1,...,1/d) R^T for a fixed random orthogonal R, using the same grid size, budget of ~15k–20k unique posterior calls, and one JKO step as in Table 3. During the fixed-point iteration record the TT ranks of η_m and \\hatη_m, and measure the TT-cross reconstruction error on a held-out set of grid points. Then compare S2ε(ρ_TT, ρ_ref) with the ref-to-ref distribution. If ranks stay low, held-out error is small, and S2ε matches ref-to-ref, the concern is resolved. If ranks grow to rmax and S2ε visibly exceeds ref-to-ref, the low-rank assumption is the binding limitation and the headline claim must be scoped to low-rank posteriors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the method is tractable in high dimension rests on TT-cross (Sec. 4.3) being accurate for the intermediate potentials η_m, \\hatη_m, and the posterior. Theorem 4.1 only bounds cross error conditional on existence of a rank-r approximation and on conditioning constants κ; it gives no a-priori guarantee. Section 1.3 explicitly labels low-rankness a hypothesis. In every reported experiment the target is low-rank by construction: the Gaussian mixture is a sum of 5 product Gaussians (rank ≤5), the nonconvex potential is rank 2, and the parabolic posterior has independent coefficients so its exact TT rank is 1. No experiment probes a target whose TT ranks grow with dimension or grid resolution, and the paper does not report the TT ranks of the intermediate η_m, \\hatη_m during fixed-point iteration. If those ranks are not small, storage and evaluation costs O(dNr^2) grow, the cache becomes useless, and the 'same number of posterior calls' comparison against MH collapses. This is not an internal inconsistency; the paper acknowledges the hypothesis, but it is the weakest point at which the headline claim could fail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Eulerian, entropy-regularized JKO scheme for approximating and sampling from Bayesian posteriors that are available only through unnormalized pointwise evaluations. The proximal step is reformulated via a Hopf-Cole transform as a pair of coupled heat equations, which is then solved as a fixed-point problem in the Tensor Train format using TT-cross approximation, matrix-exponential heat solves, and Anderson acceleration. The authors demonstrate the method on a Gaussian verification problem, several synthetic targets (Gaussian mixture, double-moon, nonconvex potential), and two PDE-based Bayesian inverse problems, comparing sample quality against Metropolis-Hastings MCMC with equal numbers of posterior evaluations. They also show that the fitted TT model can be reused to construct an importance distribution for a quantity of interest. The central claim is that low-rank TT compression makes the Eulerian approach tractable in dimensions up to d=30 and that the resulting samples are comparable or better than MCMC with the same resource budget.","tokens_in":27658,"tokens_out":9785,"duration_ms":100312,"significance":"If the central claims are fully supported, this would be a valuable contribution: it offers a gradient-free, deterministic alternative to MCMC with reusable posterior evaluations, access to marginals and normalized densities, and a natural way to reuse the fitted model for related tasks such as importance sampling. The derivation in Section 3 is clear and the algorithmic pipeline is well motivated. The authors are also transparent about the main hypothesis, and the release of code at the stated repository is a strength. However, the significance is currently limited by the fact that all numerical demonstrations use targets that are low-rank by construction, and the convergence properties of the fixed-point iteration are only empirical. The headline claim about overcoming the curse of dimensionality therefore remains conditional on an unverified structural assumption.","major_comments":[{"comment":"The central tractability claim—that the TT format overcomes the curse of dimensionality—is load-bearing but is only tested on targets that are low-rank by construction. In Section 1.3 the low-rank assumption is explicitly a hypothesis, and Theorem 4.1 (Eq. (4.3.1)) provides only a conditional error bound depending on the existence of a rank-r approximation and on conditioning constants κ. The Gaussian mixture is a sum of five product Gaussians (rank ≤ 5), the double-moon and nonconvex potentials have very low algebraic separability, and both PDE posteriors are reported with rmax = 1 (Table 3). No experiment studies a target whose TT rank grows with dimension or grid resolution, and the ranks (or truncation errors) of the intermediate potentials η_m and \\hatη_m during the fixed-point iteration are never reported. To support the headline claim, the paper should add at least one experiment with a genuinely non-product target (e.g., a correlated Gaussian with a dense covariance, or a rotated/elliptical distribution) and should report the evolution of TT ranks and approximation errors of η_m and \\hatη_m, together with memory and runtime scaling in d and N.","section":"1.3, 4.3, Tables 2–3"},{"comment":"The fixed-point iteration is the engine of each proximal step, but its convergence is only empirical. The operator G in (3.2.5)–(3.2.8) is derived from the optimality conditions; no theorem states that G is a contraction or that the Picard/Anderson iterates converge. The only theoretical result, inequality (3.3.1) in Section 3.3, concerns the outer proximal sequence under a subgradient bound Ξ_k, and the paper explicitly hypothesizes that Ξ_k can be controlled by (T_k, β_k). The convergence evidence is a single Gaussian verification (Figure 3a). Because divergence of the fixed-point iteration would invalidate every reported experiment, the paper should either prove convergence under explicit conditions on T, β, and the target, or provide convergence diagnostics for all test cases, including the effect of the truncation rank rmax and the Anderson relaxation parameter q.","section":"3.2, 3.3, 4.4"},{"comment":"The headline comparison against Metropolis-Hastings is made in terms of the number of unique posterior evaluations, but the TT method's total posterior calls (including cache hits) are substantially larger and its internal cost (TT-cross, SVD, Anderson updates, ODE/SDE integration) is not counted. If the abstract's 'same amount of resources' means only posterior evaluations, this should be stated explicitly; otherwise wall-clock time or a comparable cost model should be reported. As written, the claim is stronger than the measured quantity.","section":"5.1, Tables 2–3"}],"minor_comments":[{"comment":"The displayed initial condition for \\hatη is garbled; it should read \\hatη(0,x)=ρ_k(x)/η(0,x), as used in (3.2.6). Please fix the typo.","section":"3.1, Eq. (3.1.8)"},{"comment":"The notation W_2^2 and W_2 is used interchangeably for the squared and unsquared 2-Wasserstein distance; please standardize the notation.","section":"2.1, Corollary 2.2.1"},{"comment":"The theorem states a bound 'for sufficiently small ε>0', but ε is the best-approximation error; the quantifier should be clarified, and the dependence on κ should be discussed since κ is never estimated.","section":"4.3, Theorem 4.1"},{"comment":"The convergence criterion and the definition of 'not converged' (the '-' entries) are not given; please specify the maximum number of iterations and the stopping tolerance.","section":"5.2, Table 1"},{"comment":"The choice of ϵ_SDE is heuristic (0.01–0.001); a brief sensitivity study or a rule for selecting it would strengthen the sampling section.","section":"4.5"},{"comment":"The importance-sampling experiment reports only the spread of estimates; please provide a quantitative comparison of variance or mean squared error and state how the bias is controlled.","section":"6.1"},{"comment":"The related methods in [32] and [34] are discussed but never compared numerically; adding such a comparison would clarify the practical advantage of the proposed approach.","section":"1.4, 5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a numerical analysis journal, but the theoretical contribution is mainly a formal derivation plus a numerical study, and the central low-rank hypothesis remains unverified outside the tested regime. The authors are transparent about this limitation, and the experimental design is otherwise careful. I would not reject on novelty grounds; the revision should focus on strengthening the evidence for the tractability claim and on clarifying the resource metric used in the MCMC comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth engaging with. What's genuinely new is that it solves the full nonlinear entropic JKO proximal step as a fixed point in tensor-train format, rather than the linearized functional used in Han-Osher-Li [34] or the different Fokker-Planck setting of Chertkov-Oseledets [32]. The Hopf-Cole transform turns the coupled system into two heat equations, the heat kernels are applied in TT format without explicit time stepping, and the terminal condition is handled by TT-cross. That combination, plus the caching strategy on a fixed grid, is a real algorithmic contribution. The derivation in Section 3 is clean, and the experiments are careful: Sinkhorn distances to reference samples, comparisons to Metropolis-Hastings with matched posterior-call budgets, and code is released. I believe the main performance claim—comparable or better than MH with the same resources—is supported for the tested problems.\n\nThe soft spots are real but not fatal. Most importantly, the tractability claim in high dimension rests on the low-rank assumption for the intermediate entropic potentials and the posterior. Every test target is low-rank by construction: the Gaussian mixture has rank ≤5, the nonconvex potential rank 2, and the parabolic posterior rank 1. The paper flags this as a hypothesis (Section 1.3), and Theorem 4.1 only gives conditional error bounds with unknown constants. The ranks of η_m and η̂_m during iteration are not reported. If those ranks grow with dimension or grid resolution, the O(dNr²) storage eats the advantage and the cache stops helping. I'd like to see one experiment where the posterior is not low-rank by construction—e.g., a rotated covariance or a sum of non-factorizing terms—to see where the method degrades.\n\nThe fixed-point convergence is empirical; there's no contraction proof. That's a minor concern for a numerical paper, but worth stating. More concrete: key hyperparameters for the main experiments (grid size, β, T, q, rmax) aren't fully reported, which makes reproduction harder. And since [32] and [34] are cited as closest prior work, not benchmarking against them is a missed opportunity.\n\nWho gets value: researchers in Bayesian inversion with expensive forward models and posteriors that plausibly have low-rank structure. The importance sampling extension is a nice byproduct. The paper deserves a serious referee. I'd recommend major revision: add a non-low-rank test, report hyperparameters, benchmark against [32] and [34], and discuss the rank-growth issue honestly. It's not a desk reject.","headline":"A genuine algorithmic advance for structured Bayesian posteriors, but the high-dimensional tractability claim is only tested on targets that are low-rank by construction.","tokens_in":28343,"tokens_out":2659,"would_cite":true,"duration_ms":25241,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","65K10","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that entropy-regularized JKO proximal steps, discretized on a fixed grid and compressed as tensor trains, produce a tractable Eulerian surrogate for high-dimensional Bayesian posteriors whose sampling quality, at equal…","keywords":["Wasserstein gradient flow","JKO scheme","entropic regularization","tensor train","low-rank approximation","Bayesian inverse problems","sampling","fixed-point iteration"],"falsifier":"Construct a 30-dimensional target with provably high TT-rank, e.g. a random mixture of many well-separated Gaussian components, and measure the maximal rank and the number of unique posterior evaluations TT-cross needs to reach a fixed Sinkhorn distance to reference samples; if the rank grows with the number of components and the unique-evaluation count approaches or exceeds the Metropolis-Hastings chain length needed for the same accuracy, the central tractability claim fails.","tokens_in":27203,"feed_emoji":"🧮","tokens_out":8295,"duration_ms":69870,"temperature":0.7,"pith_summary":"Bayesian inverse problems are usually sampled with Markov-chain methods that burn in, correlate, and can reuse almost none of the expensive posterior evaluations. This paper proposes instead to treat the posterior as the stationary point of a Wasserstein gradient flow and to march toward it with entropy-regularized Jordan-Kinderlehrer-Otto (JKO) proximal steps, discretized once on a fixed grid. Because the grid is fixed, posterior evaluations can be cached; because the densities and potentials are stored in the tensor-train (TT) format, the exponential grid growth is compressed away when the target has low-rank structure. The paper shows that, with the same number of unique posterior evaluations as a Metropolis-Hastings chain, the TT model produces samples whose Sinkhorn distance to a reference is comparable or smaller across synthetic and PDE-constrained inverse problems (Tables 2 and 3). If this holds, expensive forward-model calls stop being wasted on burn-in and correlated draws, and a reusable model of the posterior becomes available for marginals and importance sampling.","feed_headline":"A heat-equation route to Bayesian posteriors matches or beats MCMC","feed_subtitle":"A tensor-compressed Eulerian scheme needs pointwise posterior calls and, at equal budget, matches Metropolis-Hastings.","key_machinery":"The mechanism is the entropy-regularized JKO proximal step in its dynamic form, converted by the Hopf-Cole change of variables $\\Phi = 2\\beta_k \\log \\eta$, $\\rho = \\eta\\hat\\eta$ into two linear heat equations $\\partial_t \\hat\\eta = \\beta_k \\Delta \\hat\\eta$ and $\\partial_t \\eta = -\\beta_k \\Delta \\eta$ coupled at the terminal time by $\\tilde\\eta = (\\rho_\\infty/\\hat\\eta)^{1/(1+2\\beta_k)}$. The paper treats this system as a fixed-point equation $\\eta = G(\\eta)$, solves each heat application exactly in TT format via Kronecker matrix exponentials, and reconstructs the nonlinear terminal coupling with TT-cross, so the only calls to the unnormalized posterior happen during cross approximation on the fixed grid. This object makes the method gradient-free, cacheable, and, under low-rank assumptions, subexponential in the dimension.","core_discovery":"The central claim is that a fully Eulerian, gradient-free approximation of a high-dimensional posterior is tractable when each regularized JKO proximal step is reformulated as a fixed point of two heat equations. With the Hopf-Cole transform, the optimality conditions for the proximal step become a forward heat equation for one entropic potential and a backward heat equation for another, coupled through the terminal condition by the explicit formula $\\tilde\\eta_m = (\\rho_\\infty / \\hat\\eta_m)^{1/(1+2\\beta_k)}$. The paper solves these equations on a tensor grid with finite differences in tensor-train format, applies TT-cross to build the coupled terms from few pointwise posterior evaluations, and accelerates the fixed-point iteration with Anderson acceleration. The resulting TT density yields a deterministic transport ODE, optionally with a final stochastic correction, for sampling, and the same model gives normalized marginals and an importance distribution without extra posterior calls.","pith_inferences":["The same fixed-point machinery could be applied to any energy whose first variation is computable in TT format, turning the method into a general-purpose constrained sampling tool for fairness, safety, or interpretability constraints rather than a KL-specific one.","Because the Eulerian model is a discretized function, one could build a joint density $\\rho(x,y)$ over parameters and measurements and condition on new data by tensor contractions, letting a single fitting amortize many measurement campaigns.","The observed superlinear convergence of Anderson acceleration suggests that the relaxation parameter, currently chosen heuristically, might be tuned automatically, removing one user-dependent robustness knob.","If the low-rank hypothesis is confirmed across a wider class of posteriors, the resource metric for comparing samplers may shift from chain length to unique posterior evaluations, which better reflects expensive forward-model calls."],"forward_implications":["A single JKO step with large $T$ and small $\\beta$ can already give a good posterior model (KL $\\sim 10^{-6}$ in the Gaussian verification), so multi-step flow may often be unnecessary.","Fixed-grid caching makes posterior evaluations reusable across fixed-point iterations and outer steps, multiplying the effective budget by a factor between 1.5 and 7.3 in the test problems.","From the fitted TT model one obtains the normalized density, marginals, credible intervals, and samples, all without additional posterior calls.","The model can be re-purposed as an importance distribution for a quantity of interest, reducing estimator variance without new posterior calls, as demonstrated for the parabolic inverse problem.","The method is gradient-free, so it applies when the posterior is only available through a black-box unnormalized density, such as measurements tied to a PDE or ODE forward solver."],"supporting_citations":[{"why":"Establishes the Wasserstein gradient-flow structure of the Fokker-Planck equation and thereby the JKO scheme the method is built on.","marker":"[9]"},{"why":"Supplies the minimizing-movements and EVI theory used to state convergence rates for the proximal sequence.","marker":"[10]"},{"why":"Proves convergence of entropic schemes for optimal transport and gradient flows, justifying the regularization limit.","marker":"[17]"},{"why":"Introduces TT-cross approximation, the tool used to reconstruct tensors from few pointwise posterior evaluations.","marker":"[25]"},{"why":"Shows how to solve Fokker-Planck equations by cross approximation in the tensor-train format, the template for the heat-equation solver here.","marker":"[32]"},{"why":"Applies regularized Wasserstein proximal operators with tensor trains, the closest prior method this work extends and compares against.","marker":"[34]"},{"why":"Provides the Fisher-information regularization scheme for Wasserstein gradient flows that motivates the entropic regularization used here.","marker":"[47]"},{"why":"Gives the Kronecker-product matrix-exponential identity that lets the heat equation be solved in TT format without time stepping.","marker":"[52]"}],"fun_headline_variants":["Heat equations match MCMC for Bayesian inversion","Tensorized heat flow beats MCMC on Bayesian posteriors","Gradient-free Eulerian scheme rivals MCMC for inversion","Low-rank heat equations sample high-dimensional posteriors","Regularized JKO with tensors beats MCMC in Bayesian tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the posterior and the intermediate entropic potentials are close to low-rank tensor trains, so TT-cross can reconstruct them from a few pointwise evaluations; if that fails, storage and evaluation cost blow up and the Eulerian grid becomes exponential again.","fun_headline_variants_meta":{"raw":{"variants":["Heat equations match MCMC for Bayesian inversion","Tensorized heat flow beats MCMC on Bayesian posteriors","Gradient-free Eulerian scheme rivals MCMC for inversion","Low-rank heat equations sample high-dimensional posteriors","Regularized JKO with tensors beats MCMC in Bayesian tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1366,"prompt_tokens":971,"completion_tokens":395,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":317}},"tokens_in":587,"tokens_out":395,"duration_ms":4595,"temperature":1.0,"reasoning_tokens":317,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:33:26.019858+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a 30-dimensional target with provably high TT-rank, e.g. a random mixture of many well-separated Gaussian components, and measure the maximal rank and the number of unique posterior evaluations TT-cross needs to reach a fixed Sinkhorn distance to reference samples; if the rank grows with the number of components and the unique-evaluation count approaches or exceeds the Metropolis-Hastings chain length needed for the same accuracy, the central tractability claim fails.","supporting_citations":[{"cited_title":"The variational formulation of the Fokker–Planck equation","cited_arxiv_id":null,"evidence_quote":"Establishes the Wasserstein gradient-flow structure of the Fokker-Planck equation and thereby the JKO scheme the method is built on."},{"cited_title":"Gradient flows: in metric spaces and in the space of probability measures","cited_arxiv_id":null,"evidence_quote":"Supplies the minimizing-movements and EVI theory used to state convergence rates for the proximal sequence."},{"cited_title":"Convergence of entropic schemes for optimal transport and gradient flows","cited_arxiv_id":null,"evidence_quote":"Proves convergence of entropic schemes for optimal transport and gradient flows, justifying the regularization limit."},{"cited_title":"TT-cross approximation for multidimensional arrays","cited_arxiv_id":null,"evidence_quote":"Introduces TT-cross approximation, the tool used to reconstruct tensors from few pointwise posterior evaluations."},{"cited_title":"Solution of the Fokker–Planck equation by cross approximation method in the tensor train format","cited_arxiv_id":null,"evidence_quote":"Shows how to solve Fokker-Planck equations by cross approximation in the tensor-train format, the template for the heat-equation solver here."},{"cited_title":"Tensor train based sampling algorithms for approximating regularized Wasserstein proximal operators","cited_arxiv_id":"2401.13125","evidence_quote":"Applies regularized Wasserstein proximal operators with tensor trains, the closest prior method this work extends and compares against."},{"cited_title":"Fisher information regularization schemes for Wasserstein gradient flows","cited_arxiv_id":null,"evidence_quote":"Provides the Fisher-information regularization scheme for Wasserstein gradient flows that motivates the entropic regularization used here."}],"review_version":1}