Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Efficient Diffusion Models for Symmetric Manifolds

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A projection-based diffusion framework trains on symmetric manifolds with one gradient evaluation per step and provably polynomial sampling error, closing most of the runtime gap with Euclidean diffusion.

desk verdict Promising manifold-diffusion method, but the printed training objective contradicts the reverse SDE used in the sampling guarantee; the central theorem is unsupported as written. read the letter →

arxiv 2505.21640 v1 pith:HNCQOF6J submitted 2025-05-27 cs.LG cs.AIcs.DSmath.PRstat.ML

Oren Mangoubi , Neil He , Nisheeth K. Vishnoi This is my paper
classification cs.LGcs.AIcs.DSmath.PRstat.ML MSC 58J6560J6060H1068T07
keywords diffusionmodelssymmetricmanifoldsRiemannianmanifoldheat-kernel-freetrainingBrownianmotionprojectionaverage-caseLipschitznessrandommatrixtheoryoptimaltransportcoupling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to close the runtime gap between diffusion generative models in Euclidean space and those constrained to symmetric Riemannian manifolds: the torus, the sphere, the special orthogonal group $SO(n)$, and the unitary group $U(n)$. Existing manifold diffusion models lean on the heat kernel, which lacks a closed form and forces either exponential-in-dimension arithmetic or $d$ gradient evaluations per training step. The paper's construction chooses a forward diffusion with a spatially varying covariance so that the process is exactly the projection of Euclidean Ornstein-Uhlenbeck motion onto the manifold; training then minimizes a new objective, derived from Itô's Lemma, that costs one gradient evaluation and $O(d^{1.19})$ arithmetic per iteration on $SO(n)$ and $U(n)$. For trained models that $\varepsilon$-minimize this objective, the proposed sampler is claimed to output a distribution within total-variation distance $O(\varepsilon d^9 \log(d/\varepsilon))$ on $SO(n)/U(n)$, in $\mathrm{poly}(d)\cdot\log(d/\varepsilon)$ iterations. If these guarantees hold, manifold-constrained generation for robotics, drug discovery, and quantum-physics applications approaches the efficiency of Euclidean diffusion, with accuracy bounds that are polynomial rather than exponential in dimension.

What carries the argument

The load-bearing object is the projection map $\varphi:\mathbb{R}^d\to M$ together with the Itô-Lemma projection of the Euclidean time-reversal: the reverse SDE $dY_t = f^*(Y_t,t)\,dt + g^*(Y_t,t)\,dB_t$ on $M$ has drift given by the conditional expectation of $\nabla\varphi(H_t)^\top dH_t + \tfrac12 dH_t^\top \nabla^2\varphi(H_t) dH_t$ given $\varphi(H_t)$, and covariance $g^*$ satisfying $(g^*)^2 = \mathbb{E}[J\varphi(H_t)^\top J\varphi(H_t) \mid \varphi(H_t)]$. Because the underlying Euclidean process has a closed-form Gaussian kernel, both terms are computable in closed form; for $SO(n)/U(n)$, $\varphi$ is a singular value decomposition, and the covariance's symmetric structure reduces it to $n^2$ scalar functions $\alpha_{ij} = \mathbb{E}[1/(\lambda_i-\lambda_j)]$, which is what makes the per-iteration cost $O(d^{\omega/2})$ instead of $d^2$. The proof machinery is an optimal-transport analysis: a Gronwall-type inequality for SDEs on non-negative curvature manifolds (Lemma 6.3), the random-matrix-theory verification of 'average-case Lipschitzness' (Assumption 2.1, Lemma 6.4), and a Wasserstein-to-total-variation conversion that resets the coupling every $\tau = 1/c$ steps so error grows linearly rather than exponentially in time (Lemma 6.7).

What would settle it

Take a simple non-invertible case, the sphere projection $\varphi(z)=z/\|z\|$ in $d=3$: simulate the Ornstein-Uhlenbeck process $Z_t$ honestly, record the projected path $Y_t=\varphi(Z_t)$, and estimate its short-time transition kernels; separately solve SDE (2) with the paper's conditional-expectation drift and covariance, and compare the transition kernels. A measurable discrepancy at small times would show that the SDE representation every accuracy bound flows from is not the true projected process.

Watch

Extended reading notes

Core claim

The paper's central claim is that the heat-kernel bottleneck of manifold diffusion can be replaced by a projection trick. The forward process is the projection $X_t = \varphi(Z_t)$ of a Euclidean Ornstein-Uhlenbeck process $Z_t$ onto $M$, with a spatially varying covariance chosen so that this identity is exact even on curved manifolds, on which standard Brownian motion is not a projection. The reverse drift and covariance are then conditional expectations of the projected Euclidean reverse SDE, whose Gaussian transition kernel is closed-form; an indicator that excludes a high-probability 'average-case' set $\Omega_t$ keeps the projection's derivative singularities (eigenvalue collisions, the sphere's origin) out of the picture. Manifold symmetries make $\Omega_t$ depend only on the spectrum, so the conditional expectations smooth out the singularities and the true drift $f^*$ and covariance $g^*$ are $\mathrm{poly}(d)$-Lipschitz on the whole manifold. On that basis, the paper claims Theorem 2.2: for $\varepsilon$-minimizers $\hat{f}$, $\hat{g}$ of the training objective, Algorithm 2's output $\nu$ satisfies $\|\nu-\pi\|_{TV} = O(\varepsilon d^9 \log(d/\varepsilon))$ on $SO(n)/U(n)$, and $O(\varepsilon d^6 \log(d/\varepsilon))$ on the torus and sphere, in $\mathrm{poly}(d)\cdot\log(d/\varepsilon)$ iterations, each needing one evaluation of $\hat{f}$ and $\hat{g}$, one exponential map, and $O(d)$ arithmetic operations.

Load-bearing premise

The whole guarantee passes through the step where the projected time-reversed process $Y_t = \varphi(H_t)$ is asserted to follow the SDE whose drift and covariance are those conditional expectations; when the projection is many-to-one, a deterministic function of a Markov process is not automatically Markov, and the paper does not cite or prove a theorem licensing that SDE representation.

Editorial extensions

If this is right

  • Per training step on $SO(n)$ and $U(n)$ drops to one gradient evaluation and $O(d^{1.19})$ arithmetic, versus $d$ gradient evaluations and exponential-in-$d$ heat-kernel sums in earlier manifold diffusion models.
  • Sampling accuracy becomes polynomial: with an $\varepsilon$-optimal model the output is within $O(\varepsilon d^9 \log(d/\varepsilon))$ total variation on $SO(n)/U(n)$ in $O(d^{5.5}\log(d/\varepsilon))$ iterations, each iteration requiring one drift/covariance evaluation plus one exponential map.
  • The forward diffusion is simulation-free: a single Gaussian draw plus one projection (an SVD for matrix groups) replaces numerical SDE/ODE solvers inside the training loop.
  • The framework composes under direct products, so it covers spaces like $SE(n) \cong \mathbb{R}^n \times SO(n)$ used in rigid-body and protein-backbone generation.
  • Empirically, per-iteration training stays within a factor of 3 of a Euclidean diffusion model on $U(n)$ up to manifold dimension 1225, and C2ST and likelihood scores beat the prior heat-kernel and momentum-based baselines for $n \geq 9$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The conditional-expectation (projected) SDE is the gate: a numerical check of whether $\varphi(H_t)$'s transition kernels match SDE (2) on the sphere or a low-dimensional unitary group would directly stress-test the entire chain of guarantees.
  • The 'project plus average-case Lipschitz' template looks portable to other spaces admitting a projection oracle whose singular set is symmetry-invariant and rarely hit, such as the positive-semidefinite cone or shape spaces; the paper sketches polytopes but leaves this open.
  • A clean empirical test of the dimensional dependence is available: train with controlled objective values on wrapped Gaussians on tori of growing $d$, measure the TV error of the samples, and check whether it scales as $O(\varepsilon d^6 \log(d/\varepsilon))$ as Theorem 2.2 predicts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a diffusion generative model for symmetric Riemannian manifolds (torus, sphere, SO(n), U(n)) by projecting an Ornstein-Uhlenbeck process Z_t in R^d onto the manifold through a map φ, and simulating an associated reverse SDE. The training objective in Eq. (4) and Algorithm 1 is claimed to be derivable from Itô's lemma, and the sampling Algorithm 2 is claimed to inherit a polynomial-in-d total variation guarantee from an ε-minimizer of that objective (Theorem 2.2, Corollary 2.3). The main advertised contributions are O(1) gradient evaluations and O(d^{ω/2}) arithmetic per training iteration, with empirical demonstrations on tori, SO(n) and U(n). The theoretical core, however, contains an internal inconsistency between the printed training objective and the objective derived in the proof, as well as an unproven Markovian-projection step, so the main accuracy/runtime claims are not established as written.

Significance. If the central derivation were correct, the paper would make a meaningful contribution: it would reduce the per-iteration training cost of manifold diffusion models to nearly Euclidean levels and provide the first polynomial-in-d sampling guarantee for SO(n) and U(n). The paper also ships code and presents empirical comparisons with RSGM, TDM, and Euclidean baselines, which is a strength. However, the main theorem is gated by the correctness of the training objective and by the claim that the projected time-reversed process satisfies the SDE in Eq. (2); both gates fail in the current manuscript. The eigenvalue-gap verification of Assumption 2.1 also has a direction error that undermines the stated Lipschitz constants. These are load-bearing issues, not presentation defects.

major comments (4)
  1. [§3, Eq. (4); Algorithm 1, Line 1; §6.1, Eqs. (19), (31), (34)] The objective actually minimized by Algorithm 1 does not match the objective derived in Lemma 6.2. Eq. (4) and Algorithm 1, Line 1 train f against (∇φ)^T (Z_{T-t} − ψ(b)e^{−(T−t)/2})/(e^{−(T−t)}−1) + (1/2)tr(∇²φ), which equals (∇φ)^T ∇log q_{T−t|0}(Z_{T−t}|b) + (1/2)tr(∇²φ). Lemma 6.2's Eq. (34), however, targets (∇φ)^T( Z_{T−t} + 2∇log q_{T−t|0}(Z_{T−t}|b) ) + (1/2)tr(∇²φ), and Eq. (19) for the reverse OU process implies the drift should be (1/2)Z_{T−t} + 2∇log q. Thus the printed objective is missing the (1/2)(∇φ)^T Z_{T−t} drift term and has the wrong coefficient on the score term. An ε-minimizer of the objective in Eq. (4) is therefore not an ε-minimizer of the objective used in Theorem 2.2's hypothesis, so the theorem's accuracy guarantee does not follow for the algorithm as implemented.
  2. [§3, Eq. (2); §6.1, Eqs. (22)–(25)] The derivation assumes that Y_t = φ(H_t) satisfies the closed SDE dY_t = E[ dY_t | Y_t ] with drift f⋆ and diffusion g⋆. For a non-injective projection φ, a deterministic function of a Markov process is not generally Markov, and the conditional expectation of the infinitesimal increment does not in general produce a Markovian SDE for the projected process. The paper neither cites nor proves a Markovian projection theorem. This is not a cosmetic issue: for U(n) and SO(n), the spectral projection depends on the evolving eigenvalue gaps, so the projected process φ(H_t) is not obviously Markovian by itself. Without this step, Eq. (2), Lemma 6.2, and hence Theorem 2.2 are unsubstantiated.
  3. [§6.3, Lemma 6.4] The proof of Lemma 6.4 uses the wrong direction of the eigenvalue-gap bound. It states ∥∇φ∥ ≤ Σ_i 1/(λ_{i+1}−λ_i) ≤ d^{1.5}√t α^{−1/3}, and justifies this by the assertion that λ_{i+1}−λ_i ≤ α^{1/3}(1/√d)√t. An upper bound on the gaps yields a lower bound on the reciprocals 1/(λ_{i+1}−λ_i), not an upper bound, so the displayed inequality for Σ 1/(gap) does not follow. The high-probability statement (40) controls the probability that a gap is small; extracting a lower bound on gaps from it would give reciprocals of the form poly(d)√t/α^c, not d^{1.5}√t α^{−1/3}. The claimed constants L1, L2, and therefore the exponents in Corollary 2.3, are not justified.
  4. [§6.4, Lemma 6.6] The proof of Lemma 6.6 expresses f⋆(U,t) as an integral of (∇φ)^T ∇log q_{T−t|0}(z) + (1/2)tr(∇²φ) over Λ, but the drift derived in Eq. (31) and Eq. (34) also contains the term (1/2)(∇φ)^T H_t (or (∇φ)^T H_t in Eq. (34)). This term is absent from the integrated formula, so the Lipschitz bound in Eq. (47) is proved for a different function than the actual reverse-drift f⋆. Consequently, the passage from Assumption 2.1 to the c-Lipschitz condition used in Lemma 6.3 and Lemma 6.7 is not established.
minor comments (4)
  1. [Table 3 and §5] Table 3 is labeled U(n) but reports d=1225 for n=50, which is n(n−1)/2, the dimension of SO(50), whereas dim(U(50))=2500. The dimension convention should be stated consistently for U(n) and SO(n).
  2. [§4.1, Eq. (12) and Lemma 6.3] The displayed bound W2(Ŷ_t,Y_t) ≤ (ρ²(Ŷ_0,Y_0)+ε)e^{ct} omits the infimum term and constants present in Lemma 6.3; this is presumably an informal sketch, but the mismatch should be flagged so readers can reconcile the overview with the full proof.
  3. [§6.7] The thresholds for 'small' eigenvalue gaps are inconsistent: the text uses O(1/√d), O(1/n^{10}), and O(1/n^8) at different points without explaining the relations among them or how they combine in the two-body analysis.
  4. [Assumption 2.1] The displayed assumption lists the condition ∥d/dU ∇φ(x)∥_{2→2} ≤ L1 twice and never states the analogous condition for d/dU ∇²φ, although the text later uses such a bound; this appears to be a typographical error.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is conditional on stated epsilon-minimizer hypotheses and externally cited mathematical results, not on its own conclusion.

full rationale

The central chain is: Lemma 6.2 shows the ideal reverse-drift f* and covariance g* are the conditional-expectation minimizers of the squared training objectives; this is a standard regression characterization, not a circular definition. Theorem 2.2 then takes epsilon-minimizers f_hat and g_hat as a hypothesis and derives a coupling/TV bound; the predicted TV error is not the same quantity as the fitted objective value. The average-case Lipschitz verification for U(n) and SO(n) cites the authors' own prior random-matrix eigenvalue-gap bounds [24,25], and Lemma 6.6 cites Proposition 20 of [8]; both are published, parameter-free external results whose assumptions do not include the paper's TV guarantee, so they are independent support rather than a self-citation loop. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported. Two non-circular correctness concerns should be separated from circularity: (i) the printed objective (4) and Algorithm 1 Line 1 omit the (1/2)(nabla phi)^T H_t reverse-drift term and alter the score coefficient relative to Lemma 6.2's Eq (34), so the paper's hypothesis 'minimize our training objective' is internally inconsistent as written; (ii) the SDE form of the projected non-invertible process Y_t = phi(H_t) in Eq (2)/(25) is asserted, not proven, and the paper itself notes in Eq (22) that phi is not an invertible map. These are gaps in correctness, not cases where a result reduces to its inputs by construction. The paper's own Appendix D also admits that its third key property is harder outside symmetric spaces, and Section 7 notes that tightening the polynomial dimension dependence remains open; these are limitations rather than circular reasoning. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The paper introduces no new physical entities. Its theoretical load is carried by a projection-with-conditional-expectation construction, an average-case Lipschitz assumption, and cited random matrix theory estimates. No target result is assumed as an input, but the unproved Markovian projection step is a genuine additional premise.

free parameters (3)
  • Time horizon T = O(log(dC/epsilon))
    Chosen to make the Ornstein-Uhlenbeck forward process mix to within epsilon of N(0,I). Not fitted to data, but both the sampling guarantee and runtime scale with it.
  • Failure probability alpha in Assumption 2.1 = set <= epsilon
    Trades the probability that phi fails the average-case Lipschitz bounds against the Lipschitz constants L1 and L2. Not fitted to data, but it enters the later TV error.
  • Early stopping time t0 = min(alpha/C, alpha^2/d)
    Borrowed from Proposition 6.5 of Chen et al. to ensure the Euclidean score is Lipschitz after time t0.
assumptions (7)
  • standard math The Ornstein-Uhlenbeck process has a closed-form Gaussian transition kernel.
    Used throughout the training derivation to replace heat kernels with Gaussian sampling.
  • standard math Rauch comparison theorem bounds derivatives of squared geodesic distance on manifolds of non-negative curvature.
    Used in Lemma 6.3 to obtain the Gronwall-like Wasserstein bound for coupled SDEs.
  • domain assumption Eigenvalue-gap lower bounds for Gaussian perturbations hold uniformly over time intervals (Eqs. (13) and (40), citing [24,25,2]).
    This random matrix theory statement is the main external input for verifying Assumption 2.1 on U(n) and SO(n).
  • ad hoc to paper Assumption 2.1: average-case Lipschitzness of phi on an omega_t set that contains the forward diffusion with high probability.
    Introduced by the paper and verified heuristically for the target manifolds; it is the key smoothness premise for the sampling proof.
  • domain assumption The pushforward psi(pi) has bounded support in a ball of radius C (Assumption 6.1).
    Needed for OU mixing bounds and for Proposition 6.5; satisfied by the constructed psi maps.
  • ad hoc to paper Y_t = phi(H_t) follows the SDE in Eq. (2) with conditional-expectation drift and diffusion coefficients.
    Asserted in Section 3 and Lemma 6.2 without proving a Markovian projection theorem; the projected process is not obviously Markov.
  • standard math Proposition 6.5 (Proposition 20 of [8]): the Euclidean score is Lipschitz for t > t0.
    Cited external result used to bound the derivative of the reverse OU drift in Lemma 6.6.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Diffusion Models for Symmetric Manifolds." pith.science (2026). https://pith.science/paper/HNCQOF6J

@misc{pith2026250521640,
  author       = {Pith},
  title        = {Pith review of: Efficient Diffusion Models for Symmetric Manifolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HNCQOF6J}},
  note         = {Machine review of arXiv:2505.21640}
}
abstract

We introduce a framework for designing efficient diffusion models for $d$-dimensional symmetric-space Riemannian manifolds, including the torus, sphere, special orthogonal group and unitary group. Existing manifold diffusion models often depend on heat kernels, which lack closed-form expressions and require either $d$ gradient evaluations or exponential-in-$d$ arithmetic operations per training step. We introduce a new diffusion model for symmetric manifolds with a spatially-varying covariance, allowing us to leverage a projection of Euclidean Brownian motion to bypass heat kernel computations. Our training algorithm minimizes a novel efficient objective derived via Ito's Lemma, allowing each step to run in $O(1)$ gradient evaluations and nearly-linear-in-$d$ ($O(d^{1.19})$) arithmetic operations, reducing the gap between diffusions on symmetric manifolds and Euclidean space. Manifold symmetries ensure the diffusion satisfies an "average-case" Lipschitz condition, enabling accurate and efficient sample generation. Empirically, our model outperforms prior methods in training speed and improves sample quality on synthetic datasets on the torus, special orthogonal group, and unitary group.

Figures

Figures reproduced from arXiv: 2505.21640 by the authors.

Figure 1
Figure 1. C2ST scores when training on datasets of quantum evolution operators on [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗
Figure 2
Figure 2. Points generated by different models when training on a dataset sampled from a wrapped [PITH_FULL_IMAGE:figures/full_fig_p040_2.png] view at source ↗
Figure 3
Figure 3. Points generated by different models trained on a Gaussian mixture dataset on [PITH_FULL_IMAGE:figures/full_fig_p041_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Points generated on U(n) for different values of n, when training on datasets comprising time-evolution operators of quantum harmonic oscillators with random potentials. For n = 9 and n = 15, we observe that our model generates samples resembling the data distribution,…
Figure 5
Figure 5. Figure 5: A probability density µ with one mode (blue) on the torus. The map ψ, which maps points in the d-dimensional torus Td to Euclidean space R d , may break up the single mode on the torus into up to 2 d separated modes in R d . This can make the task of learning the pushf…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages

  1. [1]

    Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    Brian DO Anderson. Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    1982

  2. [2]

    Number 118

    Greg W Anderson, Alice Guionnet, and Ofer Zeitouni.An introduction to random matrices. Number 118. Cambridge university press, 2010

    work page 2010

  3. [3]

    Springer, 2014

    DominiqueBakry, IvanGentil, MichelLedoux, etal.Analysis and geometry of Markov diffusion operators, volume 103. Springer, 2014

    work page 2014

  4. [4]

    Matching normalizing flows and probability paths on manifolds

    Heli Ben-Hamu, Samuel Cohen, Joey Bose, Brandon Amos, Maximillian Nickel, Aditya Grover, Ricky TQ Chen, and Yaron Lipman. Matching normalizing flows and probability paths on manifolds. InInternational Conference on Machine Learning, pages 1749–1763. PMLR, 2022

    work page 2022

  5. [5]

    Nearly d-linear convergence bounds for diffusion models via stochastic localization

    Joe Benton, Valentin De Bortoli, Arnaud Doucet, and George Deligiannidis. Nearly d-linear convergence bounds for diffusion models via stochastic localization. InThe Twelfth Interna- tional Conference on Learning Representations, 2024

    work page 2024

  6. [6]

    Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions

    Hongrui Chen, Holden Lee, and Jianfeng Lu. Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions. InInternational Conference on Machine Learning, pages 4735–4763. PMLR, 2023

    work page 2023

  7. [7]

    Flow matching on general geometries

    Ricky TQ Chen, Meta FAIR, and Yaron Lipman. Flow matching on general geometries. In ICLR, 2024

    work page 2024

  8. [8]

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions

    Sitan Chen, Sinho Chewi, Jerry Li, Yuanzhi Li, Adil Salim, and Anru Zhang. Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions. InThe Eleventh International Conference on Learning Representations, 2023

    2023

Show all 39 references
  1. [9]

    Theory and algorithms for diffusion processes on Riemannian manifolds.arXiv preprint arXiv:2204.13665, 2022

    Xiang Cheng, Jingzhao Zhang, and Suvrit Sra. Theory and algorithms for diffusion processes on Riemannian manifolds.arXiv preprint arXiv:2204.13665, 2022

    2022 arXiv

  2. [10]

    Molecular design in drug discovery: a comprehensive review of deep generative models.Briefings in bioinformatics, 22(6):bbab344, 2021

    Yu Cheng, Yongshun Gong, Yuansheng Liu, Bosheng Song, and Quan Zou. Molecular design in drug discovery: a comprehensive review of deep generative models.Briefings in bioinformatics, 22(6):bbab344, 2021

    2021

  3. [11]

    Advances in machine-learning-based sampling motivated by lattice quantum chromody- namics.Nature Reviews Physics, 5(9):526–535, 2023

    Kyle Cranmer, Gurtej Kanwar, Sébastien Racanière, Danilo J Rezende, and Phiala E Shana- han. Advances in machine-learning-based sampling motivated by lattice quantum chromody- namics.Nature Reviews Physics, 5(9):526–535, 2023

    2023

  4. [12]

    Riemannian score-based generative modelling.Advances in Neural In- formation Processing Systems, 35:2406–2422, 2022

    Valentin De Bortoli, Emile Mathieu, Michael Hutchinson, James Thornton, Yee Whye Teh, and Arnaud Doucet. Riemannian score-based generative modelling.Advances in Neural In- formation Processing Systems, 35:2406–2422, 2022

    2022

  5. [13]

    A Brownian-motion model for the eigenvalues of a random matrix.Journal of Mathematical Physics, 3(6):1191–1198, 1962

    Freeman J Dyson. A Brownian-motion model for the eigenvalues of a random matrix.Journal of Mathematical Physics, 3(6):1191–1198, 1962

    1962

  6. [14]

    Rigid motion estimation using mixtures of projected Gaussians

    Wendelin Feiten, Muriel Lang, and Sandra Hirche. Rigid motion estimation using mixtures of projected Gaussians. InProceedings of the 16th International Conference on Information Fusion, pages 1465–1472. IEEE, 2013. 34

    2013

  7. [15]

    Note on the derivatives with respect to a parameter of the solutions of a system of differential equations.Annals of Mathematics, pages 292–296, 1919

    Thomas Hakon Gronwall. Note on the derivatives with respect to a parameter of the solutions of a system of differential equations.Annals of Mathematics, pages 292–296, 1919

    1919

  8. [16]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  9. [17]

    Number 38

    Elton P Hsu.Stochastic analysis on manifolds. Number 38. American Mathematical Soc., 2002

    2002

  10. [18]

    Riemannian diffusion models.Advances in Neural Information Processing Systems, 35:2750– 2761, 2022

    Chin-Wei Huang, Milad Aghajohari, Joey Bose, Prakash Panangaden, and Aaron C Courville. Riemannian diffusion models.Advances in Neural Information Processing Systems, 35:2750– 2761, 2022

    2022

  11. [19]

    Torsional diffusion for molecular conformer generation.Advances in neural information processing sys- tems, 35:24240–24253, 2022

    Bowen Jing, Gabriele Corso, Jeffrey Chang, Regina Barzilay, and Tommi Jaakkola. Torsional diffusion for molecular conformer generation.Advances in neural information processing sys- tems, 35:24240–24253, 2022

    2022

  12. [20]

    Generative modeling on manifolds through mixture of Riemannian diffusion processes

    Jaehyeong Jo and Sung Ju Hwang. Generative modeling on manifolds through mixture of Riemannian diffusion processes. InInternational Conference on Machine Learning, 2024

    2024

  13. [21]

    Denoising diffusion probabilistic models on SO(3) for rotational alignment

    Adam Leach, Sebastian M Schmon, Matteo T Degiacomi, and Chris G Willcocks. Denoising diffusion probabilistic models on SO(3) for rotational alignment. InICLR 2022 Workshop on Geometrical and Topological Representation Learning, 2022

    2022

  14. [22]

    Revisiting classifier two-sample tests

    David Lopez-Paz and Maxime Oquab. Revisiting classifier two-sample tests. InICLR, 2017

    2017

  15. [23]

    Scaling Riemannian diffusion models.Advances in Neural Information Processing Systems, 36, 2024

    Aaron Lou, Minkai Xu, Adam Farris, and Stefano Ermon. Scaling Riemannian diffusion models.Advances in Neural Information Processing Systems, 36, 2024

    2024

  16. [24]

    Private covariance approximation and eigenvalue- gap bounds for complex Gaussian perturbations

    Oren Mangoubi and Nisheeth K Vishnoi. Private covariance approximation and eigenvalue- gap bounds for complex Gaussian perturbations. InThe Thirty Sixth Annual Conference on Learning Theory, pages 1522–1587. PMLR, 2023

    2023

  17. [25]

    Private low-rank approximation for covariance matri- ces, Dyson Brownian motion, and eigenvalue-gap bounds for Gaussian perturbations.Journal of the ACM, 72(2):1–88, 2025

    Oren Mangoubi and Nisheeth K Vishnoi. Private low-rank approximation for covariance matri- ces, Dyson Brownian motion, and eigenvalue-gap bounds for Gaussian perturbations.Journal of the ACM, 72(2):1–88, 2025

    2025

  18. [26]

    Mathieu and M

    E. Mathieu and M. Nickel. Riemannian continuous normalizing flows. InAdvances in Neural Information Processing Systems, 2020

    2020

  19. [27]

    The imbedding problem for Riemannian manifolds.Annals of mathematics, 63(1):20–63, 1956

    John Nash. The imbedding problem for Riemannian manifolds.Annals of mathematics, 63(1):20–63, 1956

    1956

  20. [28]

    Video generation models as world simulators, 2023

    OpenAI. Video generation models as world simulators, 2023

    2023

  21. [29]

    Springer, 2006

    Peter Petersen.Riemannian geometry, volume 171. Springer, 2006

    2006

  22. [30]

    A contribution to differential geometry in the large.Annals of Mathe- matics, 54(1):38–55, 1951

    Harry Ernest Rauch. A contribution to differential geometry in the large.Annals of Mathe- matics, 54(1):38–55, 1951

    1951

  23. [31]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 35

    2022

  24. [32]

    Moser flow: Divergence- based generative modeling on manifolds.Advances in Neural Information Processing Systems, 34:17669–17680, 2021

    Noam Rozen, Aditya Grover, Maximilian Nickel, and Yaron Lipman. Moser flow: Divergence- based generative modeling on manifolds.Advances in Neural Information Processing Systems, 34:17669–17680, 2021

    2021

  25. [33]

    Hanson-Wright inequality and sub-Gaussian concen- tration.Electronic Communications in Probability, 8(82):1–9, 2013

    Mark Rudelson and Roman Vershynin. Hanson-Wright inequality and sub-Gaussian concen- tration.Electronic Communications in Probability, 8(82):1–9, 2013

    2013

  26. [34]

    Nisheeth K. Vishnoi. Geodesic convex optimization: Differentiation on manifolds, geodesics, and convexity.CoRR, abs/1806.06373, 2018

    2018 arXiv

  27. [35]

    De novo design of protein structure and function with rfdiffusion.Nature, 620(7976):1089–1100, 2023

    Joseph L Watson, David Juergens, Nathaniel R Bennett, Brian L Trippe, Jason Yim, Helen E Eisenach, Woody Ahern, Andrew J Borst, Robert J Ragotte, Lukas F Milles, et al. De novo design of protein structure and function with rfdiffusion.Nature, 620(7976):1089–1100, 2023

    2023

  28. [36]

    Se (3) diffusion model with application to protein backbone generation

    Jason Yim, Brian L Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, and Tommi Jaakkola. Se (3) diffusion model with application to protein backbone generation. InInternational Conference on Machine Learning, pages 40001–40039. PMLR, 2023

    2023

  29. [37]

    Triv- ialized momentum facilitates diffusion generative modeling on Lie groups

    Yuchen Zhu, Tianrong Chen, Lingkai Kong, Evangelos A Theodorou, and Molei Tao. Triv- ialized momentum facilitates diffusion generative modeling on Lie groups. InInternational Conference on Learning Representations, 2025. 36 A Additional simulation details A.1 Datasets Given ad...

    2025

  30. [38]

    Sample an indexiat random from{1,...,k}

  31. [39]

    smooth out

    SetX= exp mi(Z), whereexp x(·)denotes the exponential map at any pointx∈M. Datasets on the TorusTd.The synthetic dataset is sampled from a single-wrapped Gaussian distribution, with mean at the origin,(0,...,0) T and covariance matrix0.2Id. A total of 30,000 points were sample...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.