{"id":"a2d75ec4-78a4-4dbd-8863-3e9edfef0c4b","arxiv_id":"2412.07637","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A low-rank tensor-train solver for the continuity equation along an annealing path, combined with resampling and Langevin steps, samples Boltzmann densities with low energy distance on benchmarks.","lead":"This paper builds a sampler for unnormalized Boltzmann densities by learning the velocity field of an annealing path in a low-rank tensor format, then adding Langevin and resampling steps. It matters because training reduces to linear solves instead of deep-network optimization, offering a possible stable alternative to neural samplers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (6) is not implementable as written: C_t = ∂t log Z_t for an unnormalized target is never defined or estimated, so the ALS is minimizing an unspecified objective and the Table 1 numbers may not test the claimed PDE.","rationale":"The paper proposes to learn a velocity field vθ_t by minimizing the squared residual of the logarithmic continuity equation in a functional tensor train format. For that objective to be well-defined, the scalar C_t = ∂t log Z_t must be available at each time step. The manuscript never states how this quantity is obtained for an unnormalized Boltzmann target. This is upstream of the two concerns the reader emphasized: the choice of training samples and the approximate resampling density q. If C_t is missing or misestimated, the ALS is minimizing a different functional, so the learned vector field is not a solution of Eq. (3), and the subsequent ODE transport is not transporting p0 along the intended annealing path. The numerical experiments are detailed and the ablations are informative, but they cannot validate the central mechanism unless the objective actually optimized is the one written down. I do not think this warrants rejection: the gap may be fixable by specifying an estimator, and the stochastic steps may still produce reasonable samples. It does, however, strengthen the case for a conditional verdict, which is the reader's current position. The reader's rationale mentions C_t in passing, so my agreement is partial rather than full; the reader's stated weakest assumption is the training-distribution/mode-discovery issue, whereas I view the unspecified C_t as the more load-bearing condition for the central claim.","tokens_in":8567,"tokens_out":6234,"duration_ms":63476,"concrete_test":"Ask the authors to add an explicit C_t estimator to Algorithm 1, for example \\hat C_t = -(1/N) Σ_ℓ (f1 - f0)(x_ℓ) on the current transported samples, and rerun the many-well f^{4,8}_1 row of Table 1 under three settings: (a) this sample-based estimator, (b) C_t ≡ 0, and (c) exact C_t computed by high-resolution numerical quadrature of the known target. If (a) and (c) agree within the reported 2e-4 standard deviation and both differ from (b), then C_t is load-bearing and the paper must document how it was obtained. If (a) and (b) are indistinguishable, the paper should explain why the C_t term can be dropped despite Eq. (3). Independently, inspect the released code used for Table 1 to verify whether any C_t term appears in the residual assembled for the ALS solves.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central least-squares loss (6) contains the term C_t = ∂t log Z_t, the time-derivative of the log-normalizer of the annealing path ft = tf1 + (1-t)f0. For the unnormalized targets f1 considered in this paper, C_t is not known in closed form; it depends on the very intermediate densities pt = e^{-ft}/Zt that the method is supposed to approximate. The manuscript never states how C_t is computed or estimated. Algorithm 1 lists ALS, ODEsolve, resampling, and Langevin steps, but contains no C_t computation or update, and Section 4.1 discusses sample selection without addressing this term. This is not a cosmetic omission: a wrong constant C_t changes the residual in Eq. (3) by a spatially constant function, and since div v can produce constants, ALS can partially compensate by introducing a spurious divergent component into the learned velocity field. The reported energy distances in Table 1 may therefore be driven by the stochastic resampling and Langevin postprocessing rather than by a genuine solution of the continuity equation. The central claim that the method 'solves the underlying continuity equation in the low-rank tensor train format' is unsupported until the C_t term is explicitly defined, estimated, and shown to enter the implemented objective.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes sampling from an unnormalized Boltzmann density e^{-f_1}/Z_1 by approximating the velocity field of the continuity equation for the annealing path f_t = t f_1 + (1-t) f_0. The velocity field is represented as a functional tensor train (FTT) with an H^2-orthonormalized Fourier basis, and the coefficients are fitted by an alternating linear scheme (ALS) applied to the empirical squared residual of the logarithmic continuity equation. Training is performed on samples transported by the current approximation (ODEsolve), with optional Langevin and resampling steps inspired by sequential Monte Carlo, and TT ranks are adjusted adaptively by an exponential moving average with SVD truncation. The paper reports energy distances for Gaussian mixture and many-well targets, comparing a pure flow, a flow with post-processing, a flow with stochastic steps during training, and a stochastic-only baseline. The best reported result is an energy distance of 1.9e-3 +/- 2e-4 for the many-well problem f^{4,8}_1 with the flow+stochastic method (Table 1).","tokens_in":8848,"tokens_out":7536,"duration_ms":70729,"significance":"If the method works as claimed, it is a useful alternative to neural-network PDE samplers: the FTT parameterization avoids backpropagation through networks, the H^2 Fourier basis is a principled choice for PDE residuals, and the rank-adaptive ALS update is elegant. The ablation study in Table 1 separates the contributions of the flow, resampling, and Langevin steps and includes standard deviations. However, the main algorithmic claim is not yet supported because the normalizing-constant term C_t entering the objective is never specified or estimated, and because the stochastic corrections are explicitly left unanalyzed. The numerical experiments are therefore not reproducible in their current form, and the reported accuracy cannot yet be attributed to a solved continuity equation with certainty.","major_comments":[{"comment":"The empirical loss J_{t,N} in Eq. (6) contains the term C_t = d/dt log Z_t, but the manuscript never states how C_t is computed or estimated for the unnormalized targets. Algorithm 1 lists no C_t computation or update, and the statement in Section 5 that the ground truth is known refers to validation samples, not to the training objective. This is a load-bearing omission: if C_t is replaced by a wrong constant, the residual in Eq. (3) changes by a spatially constant function, which can be absorbed by a divergent component of div v, so the learned velocity field need not solve the intended continuity equation. Until C_t is explicitly defined, estimated, and shown to enter the implemented objective, the central claim that the method solves the continuity equation in the TT format is unsupported.","section":"Section 2, Eq. (6); Algorithm 1"},{"comment":"Sampling the empirical loss from the current trained flow means the PDE residual is only penalized on regions that already carry mass under the current approximation. Section 4.1 adopts this mode-seeking strategy from [19] without discussing its coverage failure mode: a target mode that is not discovered by the transported samples is never seen by the ALS residual, and no mechanism in Algorithm 1 provably discovers it. The Langevin and resampling steps are proposed as a mitigation, but their effect is heuristic. A concrete test would be to initialize from a prior that misses a well-separated mode and report whether flow+stochastic recovers it; without such a test or an analysis, the method's exploration behavior for multimodal Boltzmann densities is not established.","section":"Section 4.1"},{"comment":"The resampling step uses importance weights w^{(l)} = p(x^{(l)})/q(x^{(l)}), where q is obtained from the flow transport T_k only. Algorithm 1 nevertheless applies Langevin steps before and after resampling, and the authors explicitly state that the effect of the Langevin layers on the density q is left for future work and that only the flow layers are used. Hence the proposal density used in the weights is not the density of the samples that are actually resampled, so the stochastic correction can introduce an uncontrolled bias. Since the stochastic steps are presented as the mechanism that adjusts the relative mode weights and anneals to the correct path distribution, this limitation is load-bearing and should be addressed either by including the Langevin transition in q or by reporting a sensitivity analysis.","section":"Section 4.1, Algorithm 1"}],"minor_comments":[{"comment":"The inner product notation in Eq. (3) is missing a comma: it should read <nabla f_t, v_t> rather than <nabla f_t v_t>.","section":"Section 2, Eq. (3)"},{"comment":"The UpdateTT procedure depends on an SVD truncation tolerance and an EMA weighting factor, but neither quantity is specified in Section 5, so the rank-adaptivity claim is not reproducible.","section":"Section 4.2"},{"comment":"The number of Langevin steps, the Langevin step size h, and the exact Fourier basis setup (n, orthonormalization domain) are not reported for each experiment, and no runtime or TT-rank trajectories are given.","section":"Section 5"},{"comment":"The description of the stochastic baseline as taking 'the ratio for the resampling step to be p_{k+1}/p_k' is ambiguous, since for Boltzmann densities this ratio contains the unknown normalizing constants Z_k/Z_{k+1}; the baseline should state explicitly whether an unnormalized density ratio is used.","section":"Section 5, stochastic baseline"},{"comment":"The claim that polynomial bases perform poorly is not supported by a dedicated experiment or a comparison in Table 1, so it reads as speculation rather than a reported finding.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The main risk is an underspecified training objective rather than methodological dishonesty. If the authors can provide a concrete estimation procedure for C_t (for example, a Monte Carlo estimate of Z_t along the annealing path) and rerun the experiments, the core idea remains publishable. If they instead used exact normalizers from the known targets, the contribution to unnormalized sampling is substantially smaller. I would not reject on the mode-coverage issue alone, since it may be addressable with additional experiments, but the C_t issue must be resolved before the claims can be evaluated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper if you work on tensor-train sampling or physics-informed generative models. It introduces a functional tensor train (FTT) approach to solve the continuity equation along the annealing path, replacing neural-network training with ALS linear solves. That is genuinely new: the formulas for how the continuity equation acts on FTT vector fields with arbitrary basis functions, the adaptive-rank EMA updating, and the specific H2-orthonormal Fourier basis. The benchmark results, especially on the many-well problems, look strong—flow+stochastic reaches energy distances around 1e-3 on f^{4,8}_1, which is impressive if the pipeline is really doing what the paper claims. The ablation across flow, flow+, flow+stochastic, and stochastic is a nice touch and suggests the stochastic steps matter, but the pure flow still does a reasonable job.\n\nThe soft spot is real and load-bearing. The least-squares loss in Eq. (6) contains C_t = ∂t log Z_t, the time derivative of the log-normalizer along the annealing path. For an unnormalized target, this quantity is not known in closed form and depends on the very intermediate densities the method is supposed to approximate. The paper never states how C_t is computed, estimated, or avoided. Algorithm 1 lists no C_t step, and Section 4.1 does not address it. This is not a cosmetic omission: a wrong constant changes the residual by a spatially constant term, which the divergence term in the learned velocity field can partially absorb by adding a spurious divergent component. So, as written, the algorithm is not fully implementable, and the central claim that the method solves the continuity equation is not supported until C_t is defined and shown to enter the implemented objective. A second, related gap: the resampling density q is only approximately computed using the flow layers, with the effect of Langevin steps on q explicitly left unassessed. The paper admits this, but it means the stochastic mitigation is heuristic. Finally, training samples are drawn from the current trained flow, so the PDE residual is only enforced where the approximation already has mass—mode discovery is delegated to the Langevin/resampling steps, which is fine in practice but not analyzed.\n\nNone of this kills the work: the method may work well, and the numerical evidence suggests it does. But the gaps are substantial enough that the paper needs a serious revision, not just polishing. The authors should specify the C_t estimator, ideally with a convergence argument, and quantify the effect of Langevin steps on the resampling weights.\n\nWho benefits: people building low-rank samplers or working on tensor-train PDE solvers will find real value here. It deserves a serious referee and a revision request, not a desk reject.","headline":"Novel TT-based Boltzmann sampler with promising numbers, but the PDE loss is not fully specified because C_t is never defined.","tokens_in":9375,"tokens_out":1169,"would_cite":true,"duration_ms":12104,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","65C05","65N35"],"pacs":[],"model":"deepseek-v4-flash","headline":"A velocity field in functional tensor train format, trained on the continuity-equation residual and interleaved with Langevin and resampling steps, samples unnormalized Boltzmann densities efficiently.","keywords":["Boltzmann sampling","tensor train","functional tensor train","continuity equation","annealing path","Langevin dynamics","sequential Monte Carlo resampling","alternating linear scheme"],"falsifier":"Run Algorithm 1 on a two-dimensional Gaussian mixture with two modes of equal weight whose centers are separated by a large energy barrier, starting $p_0$ concentrated on only one mode; compute the fraction of final samples in the neglected mode over many trials. The central claim predicts both modes are populated in roughly equal proportion and a small energy distance, so a reproducible run where the second mode remains empty would disprove the method's mode-coverage claim.","tokens_in":8350,"feed_emoji":"🎲","tokens_out":6090,"duration_ms":49836,"temperature":0.7,"pith_summary":"This paper proposes a way to draw samples from a target density proportional to $\\exp(-f_1)$, where the normalization constant is unknown. It treats sampling as a transport problem: along the annealing path $f_t = t f_1 + (1-t) f_0$, one learns a velocity field that pushes an easy latent density $p_0$ onto the target. The velocity field is stored in a functional tensor train format, a low-rank tensor-product representation with a Fourier basis, and the continuity equation becomes a residual minimized by an alternating linear scheme. To keep all modes populated, the authors interleave deterministic transport steps with stochastic Langevin and resampling steps borrowed from sequential Monte Carlo. If the method works as reported, it gives a parameter-light, PDE-based sampler that avoids the mode collapse of reverse-KL normalizing flows; their best many-well experiment reaches energy distance $1.9\\times 10^{-3}\\pm 2\\times 10^{-4}$.","feed_headline":"PDE solver in tensor train format samples unnormalized densities","feed_subtitle":"Functional tensor train flow plus Langevin and resampling hits 1.9e-3 energy distance on many-well targets.","key_machinery":"The central object is the functional tensor train (FTT), a tensor-train coefficient tensor contracted with univariate basis functions, here an $H^2$-orthonormalized Fourier basis. The FTT represents the velocity field $v_t$ with cores that can be optimized one at a time; for each core, the PDE residual (6) is linear in the core entries, so the update reduces to solving a small quadratic program. Two auxiliary mechanisms carry the argument: an exponential moving average of successive TT approximations with TT-SVD truncation makes the ranks adaptive, and the alternating sequence of ODE transport, resampling, and Langevin steps keeps all modes of the target populated during training and evaluation.","core_discovery":"The central claim is that the continuity equation governing the annealing path can be solved directly in a functional tensor train (FTT) format, and that the resulting velocity field, when evaluated by numerical ODE integration, yields accurate samples from the Boltzmann density without ever evaluating the normalization constant. Specifically, for $f_t = t f_1 + (1-t) f_0$, the authors show the continuity equation reduces to the scalar condition $f_1 - f_0 + \\langle \\nabla f_t, v_t\\rangle - \\nabla\\cdot v_t + C_t = 0$, with $C_t = \\partial_t \\log Z_t$. The paper approximates $v_t$ as a low-rank FTT with $H^2$-orthonormalized Fourier basis functions, and trains it by minimizing the empirical squared residual (6) with ALS, where each core update is a quadratic program. Stochastic steps are added after transport: resampling reweights particle populations to handle mode-weight teleportation near $t=1$, and Langevin steps prevent collapse to point masses. The reported energy distances, for example $1.9\\times 10^{-3}\\pm 2\\times 10^{-4}$ for the many-well $f^{4,8}_1$ with flow+stochastic, support the claim that the combination works.","pith_inferences":["Implicit in the paper: because the residual is evaluated on samples from the currently trained flow, the velocity field is unconstrained in regions the flow has not yet discovered; a target mode outside the initial latent support could be invisible to the PDE loss.","A testable extension would be to add a small number of uniform or adaptive collocation points to residual (6); if this improves performance on separated modes, it would confirm that the residual's sample support, not the FTT expressivity, is the limiting factor.","The reported experiments are two-dimensional; whether the FTT ranks remain manageable in higher dimensions is open, since TT ranks for the annealed flow can in principle grow with the complexity of the target.","The effect of Langevin steps on the resampling proposal density $q$ is left unassessed in the paper; a correction that accounts for this bias could make the method asymptotically exact rather than approximate."],"forward_implications":["A user only needs the unnormalized target energy $f_1$ and an easy latent density $p_0$; no target samples and no normalizing constant are required.","The TT ranks are not fixed in advance: the EMA update with TT-SVD truncation adapts them during training.","The $H^2$-orthonormalized Fourier basis is what makes the PDE residual tractable; polynomial bases perform poorly, as stated in the conclusions.","For high-dimensional many-well targets, the full pipeline flow+stochastic reaches energy distances an order of magnitude smaller than the pure TT flow, showing the stochastic steps are necessary for good mode coverage.","Because every core update is a linear least-squares problem, training avoids the backpropagation and stability issues of physics-informed neural network solvers."],"supporting_citations":[{"why":"Supplies the core training strategy of drawing residual samples from the current trained flow.","marker":"[19]"},{"why":"Introduces the annealing-path setup and the teleportation issue that motivates the stochastic steps.","marker":"[6]"},{"why":"Defines the many-well target distributions used to benchmark the method.","marker":"[21]"},{"why":"Provides the annealed flow transport framework whose resampling and Langevin steps the method adapts to the TT setting.","marker":"[3]"},{"why":"Gives the alternating linear scheme used to solve each core update as a quadratic program.","marker":"[17]"},{"why":"Defines the TT decomposition and the TT-SVD truncation used for rank adaptation.","marker":"[26]"},{"why":"Supplies the sequential Monte Carlo resampling methodology used for mode-weight adjustment.","marker":"[10]"}],"fun_headline_variants":["Tensor train flow samples unnormalized Boltzmann densities","Low-rank transport for Boltzmann sampling without normalization","Annealing path solved in tensor train for Boltzmann sampling","Physics-informed low-rank flow samples Boltzmann densities","Tensor train continuity equation samples Boltzmann targets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that samples drawn from the current trained flow adequately cover the regions where the target density has mass, because the PDE residual is only evaluated on those samples; if the flow has not yet discovered a mode, the training signal never penalizes the velocity field there.","fun_headline_variants_meta":{"raw":{"variants":["Tensor train flow samples unnormalized Boltzmann densities","Low-rank transport for Boltzmann sampling without normalization","Annealing path solved in tensor train for Boltzmann sampling","Physics-informed low-rank flow samples Boltzmann densities","Tensor train continuity equation samples Boltzmann targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3246,"prompt_tokens":911,"completion_tokens":2335,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2267}},"tokens_in":527,"tokens_out":2335,"duration_ms":18710,"temperature":1.0,"reasoning_tokens":2267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:38:33.713109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a two-dimensional Gaussian mixture with two modes of equal weight whose centers are separated by a large energy barrier, starting $p_0$ concentrated on only one mode; compute the fraction of final samples in the neglected mode over many trials. The central claim predicts both modes are populated in roughly equal proportion and a small energy distance, so a reproducible run where the second mode remains empty would disprove the method's mode-coverage claim.","supporting_citations":[{"cited_title":"Máté and F","cited_arxiv_id":null,"evidence_quote":"Supplies the core training strategy of drawing residual samples from the current trained flow."},{"cited_title":"Arbel, A","cited_arxiv_id":null,"evidence_quote":"Provides the annealed flow transport framework whose resampling and Langevin steps the method adapts to the TT setting."},{"cited_title":"Thealternatinglinearschemefortensoroptimization in the tensor train format.SIAM Journal on Scientific Computing, 34(2):A683–A713, 2012","cited_arxiv_id":null,"evidence_quote":"Gives the alternating linear scheme used to solve each core update as a quadratic program."},{"cited_title":"Del Moral, A","cited_arxiv_id":null,"evidence_quote":"Supplies the sequential Monte Carlo resampling methodology used for mode-weight adjustment."}],"review_version":1}