Pith. sign in

REVIEW 1 major objections 4 minor 21 references

Tensor trains plus backward SDEs solve the high-dimensional HJB equation that drives reverse diffusion sampling, yielding faster and more accurate samples from multimodal unnormalized densities than neural methods.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 20:11 UTC pith:FQXJUOO6

load-bearing objection Solid methods paper: FTT+BSDE for the diffusion HJB gives a real, faster sampler on multimodal targets when the low-rank assumption holds. the 1 major comments →

arxiv 2607.06841 v1 pith:FQXJUOO6 submitted 2026-07-07 stat.ML cs.LG

Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling

Robert Gruhlke , Julius Berner , David Sommer , Lorenz Richter This is my paper
classification stat.ML cs.LG
keywords tensor trainfunctional tensor trainscore-based samplingHamilton-Jacobi-Bellmanbackward SDEdiffusion modelshigh-dimensional samplinglow-rank approximation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Sampling from an unnormalized high-dimensional density is hard because the normalizing constant is unknown and modes are often separated. Diffusion-style methods reverse a noising process whose score is the gradient of the log-density; that log-density obeys a Hamilton-Jacobi-Bellman PDE that is expensive to solve with neural networks. This paper shows that the same PDE can be solved by representing the value function in functional tensor-train format and marching backward in time with a sequence of cheap least-squares regressions that come from the associated backward SDE. The resulting Tensor Train Diffusion sampler needs far fewer density evaluations, converges in a few outer iterations, and produces higher-fidelity samples on multi-well and Ginzburg-Landau targets than existing neural diffusion samplers while running substantially faster. A sympathetic reader cares because the method turns a previously brittle, hyper-parameter-sensitive training problem into a stable, rank-adaptive linear-algebra procedure that scales with the intrinsic low-rank structure of the target.

Core claim

The authors establish that the value function of the HJB equation governing the log-density of a noising process admits an accurate, low-rank functional tensor-train approximation along the trajectories of an iteratively refined reverse SDE. By discretizing the corresponding BSDE into a sequence of regularized least-squares problems and solving them with alternating least squares on the tensor-train cores, one obtains both the value function and its gradient (the score) at every time step. The resulting control produces high-quality samples from challenging multimodal targets after only a few outer iterations and with far lower wall-clock cost than neural competitors.

What carries the argument

Functional tensor-train (FTT / extended TT) representation of the HJB value function, combined with a BSDE-derived empirical loss that is linear in the value function and its gradient; the loss is minimized by alternating least squares on the TT cores, with adaptive rank, basis degree, regularization, and domain-extension strategies that keep the approximation stable outside the training trajectories.

Load-bearing premise

The value function of the HJB equation must stay low-rank enough in the functional tensor-train format on the moving domains visited by the sampling trajectories; if the ranks grow exponentially the storage and solve cost become prohibitive.

What would settle it

On a multi-well target whose modes are known exactly, replace the learned TT control by the true score (available from a high-resolution 1-d finite-difference solve) and check whether the effective sample size and log-variance of the TT sampler match the true-score baseline within sampling error; a systematic gap that cannot be closed by increasing TT rank or number of time steps would falsify the claim that the FTT approximation is sufficiently accurate for high-fidelity sampling.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 4 minor

Summary. The paper proposes Tensor Train Diffusion (TTD), a method for sampling from unnormalized high-dimensional densities by solving the HJB PDE that governs the log-density of a noising process. The value function is approximated in the functional/extended tensor-train format and learned via a sequence of BSDE-derived empirical regression losses (Eqs. 9–10) that are minimized by alternating least squares with adaptive rank, basis degree, and regularization. An outer iteration starting from annealed Langevin refines the control so that training trajectories concentrate on the relevant support. Numerical experiments on Multiwell (d=10/50), Ginzburg–Landau, and a Kitagawa state-space model report high ESS, low log-variance, and accurate log-Z estimates, with runtimes competitive with or better than neural diffusion samplers (DIS, PIS and several other baselines).

Significance. If the low-rank hypothesis holds for the target class, TTD supplies a concrete, non-SGD alternative to PINN- and neural-SDE-based diffusion samplers that is both faster and more accurate on the reported multimodal benchmarks. The combination of HJB–BSDE theory with carefully stabilized FTT regression (domain extension, adaptive τ, rank and degree selection) is technically solid and the appendix supplies explicit Gaussian rank bounds (Thms. A.5–A.6) together with reproducible algorithmic detail. The public code link further strengthens the contribution. The work therefore advances the practical toolkit for high-dimensional sampling from unnormalized densities while remaining honest about the modelling assumption that limits its scope.

major comments (1)
  1. The central modelling hypothesis—that V(·,t) admits a sufficiently low-rank FTT representation on the trajectory-induced domains—is correctly identified as a limitation (§B.1 and conclusion) and is supported by the Gaussian analysis and the rank-adaptivity diagnostics (Figs. 5, 11). No load-bearing inconsistency is present; the method is simply inapplicable when the hypothesis fails. The experimental evidence on Multiwell and Ginzburg–Landau is therefore sufficient for the claimed target class, and no further major technical revision is required on this point.
minor comments (4)
  1. Notation for time inversion (⃗f) and the discrete loss (10) is dense; a short glossary or expanded first appearance would help readers less familiar with the BSDE literature.
  2. Figures 3–4 and 14–15 would benefit from explicit wall-clock units and a brief statement of hardware used, so that the runtime comparisons are fully reproducible.
  3. The hyper-parameter study (Figs. 12–13) is useful; a short table summarizing the default values of γ, q, p, δ and m_i would make the experimental protocol easier to replicate.
  4. A few typographical inconsistencies appear (e.g., “HJB type” vs. “HJB-type”, occasional missing spaces around citations); a light copy-edit would polish the manuscript.

Circularity Check

0 steps flagged

No significant circularity: HJB–BSDE derivation and TT regression are self-contained; performance metrics are external importance weights against the known target.

full rationale

The paper’s central chain (time-reversal of an OU noising SDE o score = abla log p o Hopf–Cole transform to HJB PDE (Lemma 2.1) o Itô-derived BSDE loss (Eqs. 5–10) o discrete backward regression on trajectory samples o ALS on extended tensor-train cores) is classical optimal-control / BSDE theory plus a standard low-rank ansatz; none of the equations is definitionally equivalent to its own inputs. The outer-loop refinement and adaptive rank/regularization/basis procedures are algorithmic heuristics, not fitted free parameters that force the reported ESS / log-variance / log-Z numbers. Those metrics are computed from path-space importance weights that explicitly evaluate the unnormalized target ho_target (Section B.2), so they constitute independent external validation. Self-citations (Richter et al. 2021/2024, Gruhlke et al. 2026, Berner et al.) supply prior TT–BSDE machinery and the sampling setup, but the uniqueness or correctness of the present claims does not rest on an unverified self-citation; the Gaussian FTT-rank bounds (Theorems A.5–A.6) are proved in the appendix. The low-rank hypothesis is openly listed as a limitation (§B.1), not smuggled in as a prediction. Consequently the derivation is self-contained against external benchmarks and exhibits no circular reduction.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central claim rests on the classical HJB–score equivalence, the BSDE representation of that HJB, and the modelling hypothesis that the value function is low-rank in the FTT sense. Free parameters control regularization, domain padding, rank truncation and basis size; they are chosen by hand or by simple adaptive rules rather than fitted to the final performance metric. No new physical entities are postulated.

free parameters (5)
  • relative regularization weight γ = 0.1
    Fixed to 0.1 so that the H-norm penalty is a fixed fraction of the empirical BSDE residual; chosen by hand and shown necessary for stability (Fig. 9).
  • domain extension / shrinkage factors q, p = 0.1
    Both set to 0.1 to pad the sample-supported domain and to damp boundary oscillations; empirical sensitivity plots (Figs. 12–13) show performance degrades outside a narrow range.
  • singular-value rank threshold δ = 1e-4
    Relative threshold 10^{-4} used to truncate TT ranks; controls the accuracy–cost trade-off of the adaptive rank procedure.
  • number of univariate basis functions m_i = 3–13 (problem-dependent)
    Treated as a hyper-parameter (3–13 Fourier modes) tuned to observed approximation quality; larger m improves ESS but increases sample complexity.
  • number of time steps N = 2^8 … 2^13
    Discretization parameter; performance improves with N (Figs. 3, 6) at linear cost; chosen by the user for each experiment.
axioms (4)
  • domain assumption The log-density of the time-reversed diffusion satisfies the HJB PDE (4) (Lemma 2.1).
    Standard consequence of the Fokker–Planck equation under the Hopf–Cole transform; taken from Berner et al. (2024) and classical references.
  • standard math The BSDE residual (7)/(9) vanishes if and only if the approximate value function coincides with the true HJB solution (up to discretization).
    Uniqueness of viscosity solutions of the HJB and Itô’s formula; classical BSDE theory (Pardoux, Bouchard–Touzi, Gobet et al.).
  • ad hoc to paper Target densities of interest possess latent low-rank structure that is inherited by the HJB value function along sampling trajectories.
    Stated as a modelling hypothesis in the introduction and limitations (§B.1); no general proof is given that HJB solutions stay low-rank.
  • standard math The set of fixed-rank tensor trains is a closed Riemannian manifold, permitting stable ALS optimization.
    Standard TT geometry (Holtz et al. 2012b, Oseledets 2011).
invented entities (1)
  • Tensor Train Diffusion (TTD) algorithm independent evidence
    purpose: Concrete solver that couples FTT approximation, BSDE regression, outer trajectory refinement and adaptive rank/degree/regularization for score-based sampling.
    The algorithmic object itself; not a physical entity. Independent evidence is the released code and the reported numerical benchmarks.

pith-pipeline@v1.1.0-grok45 · 43798 in / 3259 out tokens · 36342 ms · 2026-07-10T20:11:56.541151+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling." pith.science (2026). https://pith.science/paper/FQXJUOO6

@misc{pith2026260706841,
  author       = {Pith},
  title        = {Pith review of: Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FQXJUOO6}},
  note         = {Machine review of arXiv:2607.06841}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion models offer a powerful framework for sampling from complex probability densities by learning to reverse a noising process. A common approach involves solving for the time-reversed stochastic differential equation (SDE), which requires the score function of the evolving sample distribution. The logarithm of this distribution's density is governed by a Hamilton-Jacobi-Bellman (HJB) type partial differential equation (PDE). However, current methods for solving this PDE, such as PINNs or trajectory-based techniques, often suffer from long training times and significant sensitivity to hyperparameter tuning. In this work, we introduce a novel and efficient solver for the underlying HJB equation based on the functional tensor train (FTT) format. The FTT representation leverages latent low-rank structures to efficiently approximate high-dimensional functions, enabling both model compression and rapid computation. By integrating this efficient representation with a backward-in-time iterative scheme derived from backward stochastic differential equations (BSDEs), we develop a fast, robust and accurate sampling method. Our approach overcomes primary bottlenecks of existing techniques, enabling high-fidelity sampling from challenging target distributions with improved efficiency.

Figures

Figures reproduced from arXiv: 2607.06841 by David Sommer, Julius Berner, Lorenz Richter, Robert Gruhlke.

Figure 1
Figure 1. Figure 1: Overview of the proposed TTD method. From left to right: Annealed Langevin dynamics serve as an initialization for our approach. By learning the score function along relevant trajectories, TTD iteratively refines the sampling process. As a result, new modes are discovered and the quality of samples improves. The histograms illustrate the terminal samples of the trajectories in comparison to the target dens… view at source ↗
Figure 2
Figure 2. Figure 2: We plot one- and two-dimensional marginals of our Mul￾tiwell problem in d = 50, showing very high sampling accuracy that aligns with the metrics in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: We display the singu￾lar values of TT component 3 over time and reduce the rank whenever a corresponding value is below a prespecified threshold. In this section, we study highly anisotropic Gaus￾sian targets to analyze the rank structure of the induced transport prob￾lems and to assess the adaptivity of our algo￾rithm. We consider tar￾get densities with po￾tential log ρtarget(x) = −x ⊤Mx, where M is a ran… view at source ↗
Figure 3
Figure 3. Figure 3: We consider two instances of the Multiwell problem defined in (18) and evaluate performance using the effective sample size (ESS), log-variance divergence, as well as the log-normalizing constant and its relative error; see Section B.2. As expected, increasing the number of steps N leads to improved performance and the outer loop can provide additional gains. Remarkably, both the ESS and relative error rem… view at source ↗
Figure 4
Figure 4. Figure 4: We compare the performance versus runtime of our TTD sampler with DIS (Berner et al., 2024). By design, our algorithm produces one result per chosen number of steps N (shown as blue and orange dots), whereas DIS can improve over training time. Accordingly, we evaluate DIS at equally spaced runtime intervals. In both experiments, our algorithm is not only significantly faster but also achieves better result… view at source ↗
Figure 6
Figure 6. Figure 6: We consider two instances of the Ginzburg-Landau model defined in (19) and assess performance using the effective sample size (ESS) and the log-variance divergence. In both settings, we obtain strong results, with performance improving as the number of steps N and the number of outer iterations increase. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Consider a spline basis with 3 knots, local polynomial degree 4, and smoothness parameter s = 1, used to approximate a target function from a limited number of samples concentrated on the left side of the origin. Despite the weak coupling induced by the smoothness constraint s, the resulting regression based on these samples performs very poorly. • Localization: The localization properties are comparable t… view at source ↗
Figure 8
Figure 8. Figure 8: Illustration of the linear extension procedure. Left: the log target density (gray), the approximated value function at time step t = tN−1 (blue), and the training samples used for the approximation (orange). Right: the resulting approximate policy, where the shaded region indicates the domain in which the linear extension is activated. A.4.5 REPRESENTATION CHANGE ON DIFFERENT DOMAINS In view of the iterat… view at source ↗
Figure 9
Figure 9. Figure 9: Example of adaptively determining the regularization parameter τn for the Multiwell problem in dimension d = 10. Left: the empirical loss value LbK n (VbC). Middle: the value ∥VbC∥ 2 H. Right: the adaptive evolution of τn. for a matrix Au i,n ∈ R K,ri−1miri . In particular, the original empirical BSDE loss can be efficiently evaluated at each micro step of the proposed ALS algorithm as LbK n (VbC) = [PITH… view at source ↗
Figure 10
Figure 10. Figure 10: Illustration of basis-degree adaptivity for a one-dimensional example. Left: Trajectory transformation from the log-target distribution (log-multimodal) to the log-prior distribution (log-Gaussian). Middle: The selected polynomial degree decreases when moving from right (t = T = 2) to left (t = 0). In the limit t → 0, the solution approaches a log-Gaussian, which requires only a degree of 2. Legendre poly… view at source ↗
Figure 11
Figure 11. Figure 11: We display the singular values of all tensor train components for an example in dimension d = 6 over time and reduce the rank whenever a corresponding value is below a prespecified threshold. A.7.2 LOW-RANK AND SMOOTHNESS This section is devoted to building intuition for the relationship between rank behavior and smoothness (or regularity). To this end, we distinguish between two cases for the FTT represe… view at source ↗
Figure 12
Figure 12. Figure 12: We assess the sensitivity of our algorithm by varying key hyperparameters and recording the effective sample size (ESS), the log-variance divergence, the log-normalizing constant, and the relative error in estimating the normalizing constant. Whenever a value is absent from a plot, the algorithm diverged for that configuration. The results indicate that performance improves with increasing batch size, but… view at source ↗
Figure 13
Figure 13. Figure 13: As in [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: We compare the performance of our TTD sampler against several state-of-the-art baselines, including PIS (Zhang & Chen, 2022), DDS (Vargas et al., 2023), DIS (Richter & Berner, 2024), GBS (Blessing et al., 2024), CMCD (Vargas et al., 2024), SCLD (Chen et al., 2024), CRAFT (Matthews et al., 2022), AFT (Arbel et al., 2021), MFVI (Bishop, 2006), and MCD (Doucet et al., 2022). We see that TTD converges signifi… view at source ↗
Figure 15
Figure 15. Figure 15: We compare the performance versus runtime of our TTD sampler with DIS (Berner et al., 2024) and PIS (Zhang & Chen, 2022). By design, our algorithm produces one result per chosen number of steps N (shown as blue and orange dots), whereas DIS and PIS can improve over training time. Accordingly, we evaluate DIS and PIS at equally spaced runtime intervals. In both experiments, our algorithm is not only faster… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 9 internal anchors

  1. [1]

    Albergo, M. S. and Vanden-Eijnden, E. NETS: A non-equilibrium transport sampler.arXiv preprint arXiv:2410.02711,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Approximation Theory of Tree Tensor Networks: Tensorized Univariate Functions -- Part I

    Ali, M. and Nouy, A. Approximation with tensor net- works. Part I: Approximation spaces.arXiv e-print, pp. arXiv:2007.00118, 2020a. Ali, M. and Nouy, A. Approximation with tensor networks. Part II: Approximation rates for smoothness classes. arXiv e-print, pp. arXiv:2007.00128, 2020b. Ali, M. and Nouy, A. Approximation with tensor networks. Part III: Mult...

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  3. [3]

    Approximation by tree tensor networks in high dimensions: Sobolev and compositional functions

    Bachmayr, M., Nouy, A., and Schneider, R. Approx- imation by tree tensor networks in high dimensions: Sobolev and compositional functions.arXiv preprint arXiv:2112.01474,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Blessing, D., Berner, J., Richter, L., Domingo i Enrich, C., Du, Y ., Vahdat, A., and Neumann, G

    URL https:// openreview.net/forum?id=Q1QTxFm0Is. Blessing, D., Berner, J., Richter, L., Domingo i Enrich, C., Du, Y ., Vahdat, A., and Neumann, G. Trust region con- strained measure transport in path space for stochastic optimal control and inference.Advances in Neural Infor- mation Processing Systems, 38:165462–165510, 2026a. Blessing, D., Richter, L., B...

    work page arXiv

  5. [5]

    Sequential Controlled Langevin Diffusions

    Chen, J., Richter, L., Berner, J., Blessing, D., Neumann, G., and Anandkumar, A. Sequential controlled Langevin diffusions.arXiv preprint arXiv:2412.07081,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Hierarchical singular value decomposition of tensors.SIAM journal on matrix analysis and applica- tions, 31(4):2029–2054,

    Grasedyck, L. Hierarchical singular value decomposition of tensors.SIAM journal on matrix analysis and applica- tions, 31(4):2029–2054,

    work page 2029

  7. [7]

    MCMC for multi-modal distributions

    Latuszy´nski, K., Moores, M. T., and Stumpf-F ´etizon, T. MCMC for multi-modal distributions.arXiv preprint arXiv:2501.05908,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Flow Annealed Importance Sampling Bootstrap

    Midgley, L. I., Stimper, V ., Simm, G. N., Sch ¨olkopf, B., and Hern´andez-Lobato, J. M. Flow annealed importance sampling bootstrap.arXiv preprint arXiv:2208.01893,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Backward stochastic differential equations and viscosity solutions of systems of semilinear parabolic and elliptic PDEs of second order

    Pardoux, ´E. Backward stochastic differential equations and viscosity solutions of systems of semilinear parabolic and elliptic PDEs of second order. InStochastic Analysis and Related Topics VI: Proceedings of the Sixth Oslo—Silivri Workshop Geilo 1996, pp. 79–127. Springer,

    work page 1996

  10. [10]

    Diffusion-PINN Sampler

    Shi, Z., Hu, Z., Lin, M., and Kawaguchi, K. Stochastic taylor derivative estimator: Efficient amortization for ar- bitrary differential operators.Advances in Neural Infor- mation Processing Systems, 37:122316–122353, 2024a. Shi, Z., Yu, L., Xie, T., and Zhang, C. Diffusion-PINN sampler.arXiv preprint arXiv:2410.15336, 2024b. Song, Y ., Sohl-Dickstein, J.,...

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Dynamical Measure Transport and Neural PDE Solvers for Sampling

    Sun, J., Berner, J., Richter, L., Zeinhofer, M., M ¨uller, J., Azizzadenesheli, K., and Anandkumar, A. Dynamical measure transport and neural PDE solvers for sampling. arXiv preprint arXiv:2407.07873,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    FP64 is all you need: Rethinking failure modes in physics-informed neural networks.arXiv preprint arXiv:2505.10949,

    Xu, C., Liu, D., Nassereldine, A., and Xiong, J. FP64 is all you need: Rethinking failure modes in physics-informed neural networks.arXiv preprint arXiv:2505.10949,

    work page arXiv

  13. [13]

    Artificial Intelligence for Science in Quantum, Atomistic, and Continuum Systems

    Zhang, X., Wang, L., Helwig, J., Luo, Y ., Fu, C., Xie, Y ., Liu, M., Lin, Y ., Xu, Z., Yan, K., et al. Artificial intelli- gence for science in quantum, atomistic, and continuum systems.arXiv preprint arXiv:2307.08423,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Using the notation of contractions of tensors from Section A.1, we can compactly write C=C 1C2 · · ·C d.(21) 14 Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling First, we note that for any i= 1, . . . , d−1 we can choose an arbitrary invertible matrix Gi ∈GL(r i)⊂R ri,ri and insert it and its inverse between...

    work page 2011

  15. [15]

    The classical representation of a TT from(20) is prone to rounding errors, when trying to accessC[α] with α= (α 1,

    Then, it holds that C=U 1 · · ·U k−1CkUk+1 · · ·U d.(27) The resulting updated non-orthonormal componentC=C k is called thecoreat positionkofC. The classical representation of a TT from(20) is prone to rounding errors, when trying to accessC[α] with α= (α 1, . . . , αd). Instead, one first defines a core representation with core C, e.g. with a core positi...

    work page 2017

  16. [16]

    This is expected to provide better control over the magnitude of ∥ · ∥2 H, thereby reducing the sensitivity of the algorithm to the choice of τn

    RemarkA.3.An alternative approach would be to construct the tensor basis orthonormally with respect to the (empirical) weighted Hilbert spaces, which corresponds to orthonormalization using empirical Gram matrices. This is expected to provide better control over the magnitude of ∥ · ∥2 H, thereby reducing the sensitivity of the algorithm to the choice of ...

    work page 2026

  17. [17]

    the multi-modal setup from Section 4.1 with d= 1

    To illustrate the principle underlying our degree-adaptivity strategy, we consider a one-dimensional example, i.e. the multi-modal setup from Section 4.1 with d= 1 . In this setting, there are no rank interactions, which allows us to isolate the effect of basis-degree adaptivity. Figure 10 illustrates this behavior in practice: forT= 2 , the algorithm sel...

    work page 2026

  18. [18]

    Proof.Direct consequence of the sub-diagonal rank boundsℓ i ofMby application of Theorem A.5. 0 1 2 10 2 10 1 100 Singular value ratio TT component 1 0 1 2 TT component 2 0 1 2 TT component 3 0 1 2 TT component 4 0 1 2 TT component 5 0 1 2 time 2 3 4 5Rank 0 1 2 time 0 1 2 time 0 1 2 time 0 1 2 time Figure 11.We display the singular values of all tensor t...

    work page 2013

  19. [19]

    Unless stated otherwise, we fix the number of basis functions to10, the batch size toK= 2 15, and setp= 0.1. 210 211 212 213 214 215 batch size 0.85 0.90 0.95 ESS iteration 0 iteration 1 210 211 212 213 214 215 batch size 10 1 3 × 10 2 4 × 10 2 6 × 10 2 2 × 10 1 Log-variance 210 211 212 213 214 215 batch size 0.000 0.001 +7.312 estimated log Z 210 211 212...

    work page 2024

  20. [20]

    35 Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling 100 101 102 0.0 0.2 0.4 0.6 0.8 1.0 Multiwell, d = 10, m = 3 ESS 100 101 102 10 4 10 3 10 2 10 1 error estimated log Z 100 101 102 0.0 0.2 0.4 0.6 0.8 1.0 Multiwell, d = 50, m = 5 ESS 100 101 102 10 3 10 2 10 1 100 error estimated log Z TTD, iter 1 TTD, ite...

    work page 2022

  21. [21]

    By design, our algorithm produces one result per chosen number of steps N (shown as blue and orange dots), whereas DIS and PIS can improve over training time

    and PIS (Zhang & Chen, 2022). By design, our algorithm produces one result per chosen number of steps N (shown as blue and orange dots), whereas DIS and PIS can improve over training time. Accordingly, we evaluate DIS and PIS at equally spaced runtime intervals. In both experiments, our algorithm is not only faster but also achieves better results, partic...

    work page 2022