acceptodds
Under review as a conference paper at ICLR 2027

Hard-Constrained Sampling on Embedded Riemannian Manifolds via Adjoint Schrödinger Bridges

Abstract

Probability measures confined to nonlinear feasible sets arise whenever physical invariants or structural relations must hold exactly. Such laws are often known only up to an intractable normalizing constant, making sample generation challenging. Diffusion samplers offer an efficient alternative to MCMC methods, yet these formulations are developed primarily for Euclidean domains. Building upon the foundations of adjoint Schr\"odinger bridge sampling, this paper provides a theoretically justified method, through the lens of stochastic optimal control, to address this problem on smooth, compact, path-connected embedded Riemannian manifolds. As an element of novelty compared to existing literature, feasibility is enforced at the level of the state space, meaning the controlled diffusion is defined intrinsically on the curved space. Empirical validations are provided for several physics applications, including multimodal spherical distributions, Gibbs laws on Stiefel manifolds, closed-loop inverse kinematics, and robust Wahba optimization.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Riemannian manifold sampling, Stochastic optimal control, Schrödinger bridge samplers

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HHSrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of sampling from a distribution over a smooth embedded manifold M. The authors establish the ASBS SOC-SB for embedded manifolds, and propose a R-ASBS that uses practical gemetric approximations for intractable terms, and then further propose Extended R-ASBS which uses even more practical projections/retraction operations. The authors evaluate their methods on the sphere, Stiefel manifold, robot kinematics, and SO(3) optimization.

Critical strengths
  1. The exact theory appears to be correct (I did not check the proofs carefully), and I appreciate the sequence of simplifications, from the ideal object to R-ASBS and then to Extended R-ASBS.
  2. Extended R-ASBS uses practical oracles that are computationally tractable on many manifolds.
  3. The experiments cover a broad range of manifolds.
Critical weaknesses
  1. It is unclear how bad the approximation error in (20) affects the error in the final sampled distribution. This does not seem to be a simple discretization error, as Appendix F.1 seems to suggest that these errors do not vanish under mesh refinement. It would help if the authors can bound these quantities by bounds on terms involving the intrinsic or extrinsic curvature, or barring that, empirically estimate the error in some of the experiment setups.

  2. The experiment evaluation seem unsatisfactory, many of them do not report reasonable quantifiable metrics: the earthquake experiment simply provides a visualization, the Stiefel experiment plots the expected energy, which is a good sanity check but is too weak to suggest distributional convergence, the robot experiment again only provides a visualization, the Wahba problem states the quality of the best sample, but this does not necessarily imply the quality of the distribution.

  3. The comparison against MCMC appears unfair. In figure 1(a), MCMC appears to starts at (0,0,1), but then \mu_M(x), which is the initial distribution of R-asbs, is defined to be the uniform distribution. This seems potentially unfair for MCMC, which could also easily have started from the uniform distribution. In particular, the uniform distribution already assigns equal probability to both hemispheres, which is the principal property used in the reported comparison. Can the authors provied a comparison against MCMC using matching starting distributions?

Questions for authors

see weaknesses. My main concerns are with the experiment validity. I will be happy to raise my score if the authors can convince me that their empirical advantage holds up under more careful evaluation.

3 comments
AuthorsOpenReview ↗

Response to Reviewer HHSr

All the changes in the manuscript resubmission are marked in magenta to facilitate the review. We also updated the GitHub repository.

Reviewer Q. 1

Author's response: We agree that this error is geometric, not a discretization error. Its intrinsic part can be bounded in terms of curvature. Writing damped transport as Wt,1=PTt,1 Qt,1\mathcal{W}_{t,1}=\mathrm{PT}_{t,1}\thinspace Q_{t,1} (App. D), Grönwall's inequality applied to the linear ODE for QQ gives

∥(Wt,1)∗−(PTt,1)∗∥op≤eK2r−1,K:=sup⁡x∈M∥Ric⁡x♯∥op,r:=∫t1σs2 ds,\big\Vert(\mathcal{W}_{t,1})^{\ast}-(\mathrm{PT}_{t,1})^{\ast}\big\Vert_{\mathrm{op}} \le e^{\frac{K}{2}r}-1, \qquad K:=\sup_{x\in\mathcal{M}}\Vert\operatorname{Ric}^{\sharp}_x\Vert_{\mathrm{op}}, \quad r:=\int_t^1\sigma_s^2\thinspace ds,

which vanishes as t→1t\to1. On the unit sphere Sn\mathbb S^n the omitted factor is explicit, e−n−12re^{-\frac{n-1}{2}r} (App. F.1). So the intrinsic error is governed by the Ricci curvature and by the remaining bridge variance. Extended R–ASBS adds an extrinsic error through projection-as-transport: on S1\mathbb S^1, where Ric⁡≡0\operatorname{Ric}\equiv0, it contracts by e−12re^{-\frac12 r} (App. F.1). We do not propagate these bounds to the sampled law, but we estimate the final error empirically:

(i) on the earthquake target on S2\mathbb S^2, samples are compared with exact draws from νM\nu_{\mathcal M} (Table 2): cell-mass TV 0.173, energy KS 0.040, and a mean-energy bias within one Monte Carlo standard error (z-score 0.77);

(ii) on S3\mathbb S^3 (Wahba), where both algorithms apply, R–ASBS and Extended R–ASBS give consistent results (Table 7), so the additional approximation layer does not cost measurable accuracy;

(iii) on the closed-loop chain, the energy KS distance to long constrained MCMC runs is 0.186 (Table 4).


Reviewer Q. 2

Author's response: We thank the reviewer, and we have made a significant effort to strengthen the numerical section. The revision adds:

  • Earthquake (Sec. 4.1, Table 2): cell-mass TV, energy KS, geodesic W1W_1, MMD² and energy bias against exact samples of νM\nu_{\mathcal M}, compared with Riemannian ULA, MALA and Riemannian flow matching.
  • Robot (App. G.3, Table 4): energy KS (overall and within mode) against long constrained MCMC runs, mode balance (0.490 vs. the ideal 0.5), constraint violation (≤7.5×10−6\le 7.5\times10^{-6}) and collision-free rate (100%), compared with HardFlow.
  • Wahba (App. G.4, Tables 6–7): beyond the best sample, the concentration of the whole sampled law (fraction within 1∘1^\circ, 2∘2^\circ, 5∘5^\circ of q⋆q^\star and median error), compared with MALA, multistart Riemannian gradient descent (RGD) and uniform sampling.
  • Stiefel (Sec. 4.2, Fig. 3): expected energy across β\beta, now compared with Riemannian ULA and geodesic Monte Carlo from identical Haar initialization.

Reviewer Q. 3

Author's response: We thank the reviewer for this comment, which was also raised by Reviewer d3sk. In the revision, all samplers start from the same uniform distribution μM=1/4π\mu_{\mathcal M}=1/4\pi (Sec. 4.1, Fig. 1(a), Table 2); we refer to our reply to Reviewer d3sk for details.

Dear Reviewer HHSr,

Thank you again for your detailed and constructive review. As the author–reviewer discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns.

I appreciate the added experiment evaluations. I have improved my score to 3.

Reviewer_d3skOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper's central idea is to extend the Adjoint Schrödinger Bridge Sampler (ASBS) from Euclidean spaces to embedded Riemannian manifolds, to sample from a distribution on the manifold given gradient access to its potential.

Critical strengths
  1. The paper is well motivated and presented cleanly. Sampling from a manifold has been studied in the recent years using diffusion models, and it is natural to consider Schrodinger-bridge samplers which are close relatives of diffusion models.
  2. The idea involved is natural and convincing. SOC–SB equivalence (maybe better known as Doob h-transform characterization of the Schrödinger bridge) is a core theory for Euclidean ASBS, and finding its right generalization is naturally the key step towards a manifold version of the algorithm.
Critical weaknesses
  1. The theoretical guarantee for continuous-time formulation mostly follows from established Schrödinger-bridge theory, standard stochastic analysis on manifolds, and existing geometric approximations. This part is more or less a straightforward combination of existing techniques.
  2. The more non-trivial part of the math is dedicated to the implementation of the algorithm, but it could not establish the correctness of the algorithm. Several approximations introduce errors that do not disappear as the discretization is refined (Appendix F.1). I am also uncertain about the locality argument implied by Proposition 2, as the tilting function required seems to depend on rr while Proposition 2 only established locality for fixed hh.
  3. The experiments described in Figure 1 contain incorrect designs. It claims that "all cases utilize μM=1/4π\mu_{\mathcal M} = 1/4\pi, but then initialize geometric Langevin MCMC at the north pole instead of using this uniform distribution. Therefore, the comparison "86% of the MCMC particles become trapped in a single energy basin, failing to cross the high-energy equatorial region. In contrast, R–ASBS allocates 43.8% of its particles to the northern hemisphere, close to the theoretical ratio of 0.5" makes no sense: that's merely different initialization.
Questions for authors

I have no particular questions.

2 comments
AuthorsOpenReview ↗

Response to Reviewer d3sk

All the changes in the manuscript resubmission are marked in magenta to facilitate the review. We also updated the GitHub repository.

Reviewer Q. 1

Author's response: We agree that Theorem 1 relies on established tools; its role is to state precisely how the SB–SOC equivalence carries over to embedded manifolds and which geometric objects it requires. The step that does not carry over verbatim is adjoint matching: on a curved manifold ∇Ptf≠Pt∇f\nabla P_t f\neq P_t\nabla f, and the exact controller identity requires the adjoint of damped stochastic parallel transport (Lemma 2), which has no Euclidean counterpart. Our main contributions build on this: the practical surrogates and their locality justification (Sec. 3.4–3.5, Prop. 2), and the strong empirical validation across different physical tasks.


Reviewer Q. 2

Author's response: The mesh-independent errors are quantified in App. F.1 in terms of curvature and of A1=∫01σs2 dsA_1=\int_0^1\sigma_s^2\thinspace ds (see also our reply to Reviewer mW8Y on curvature and bridge length), and the resulting error is measured against exact target samples in Table 2.

Regarding Proposition 2: it holds for a fixed tilt, which is exactly the setting where it is used. For a given noise schedule, the proof of Lemma 2 shows that the conditional law of X1X_1 given Xt=xX_t=x is φ1M(y) prM(x,y)/φtM(x)\varphi_1^{\mathcal M}(y)\thinspace p_r^{\mathcal M}(x,y)/\varphi_t^{\mathcal M}(x), with r=∫t1σs2 dsr=\int_t^1\sigma_s^2\thinspace ds. This is the tilted law of Proposition 2 with h=φ1Mh=\varphi_1^{\mathcal M}. The tilt is the terminal potential and does not depend on rr; only the normalizer φtM=Prφ1M\varphi_t^{\mathcal M}=P_r\varphi_1^{\mathcal M} does, and the proposition handles it explicitly. Since C=eosc⁡gC1C=e^{\operatorname{osc} g}C_1 with C1,c,r0C_1,c,r_0 depending only on M\mathcal M, we obtain E⋆[dM(Xt,X1)2∣Xt]≤C′r\mathbb E^\star[d_{\mathcal M}(X_t,X_1)^2\mid X_t]\le C'r uniformly as t→1t\to1. This is the locality used for parallel transport on short terminal segments.

The bound is not uniform as A1→0A_1\to0, because the potentials then depend on A1A_1. Indeed, no such locality can hold there, since E⋆[dM(X0,X1)2]≥W2(μM,νM)2>0\mathbb E^\star[d_{\mathcal M}(X_0,X_1)^2]\ge W_2(\mu_{\mathcal M},\nu_{\mathcal M})^2 > 0. The full-horizon corrector therefore rests on the short-time expansion of Ludewig (2019), not on Proposition 2, and its residual error is among the persistent approximation errors acknowledged in App. F.1.


Reviewer Q. 3

Author's response: We thank the reviewer and agree. That comparison was affected by initialization and has been removed. In the revision (Sec. 4.1, Fig. 1(a), Table 2), every sampler targets the same earthquake density on S2\mathbb S^2 and starts from the same uniform reference μM=1/4π\mu_{\mathcal M}=1/4\pi. The MCMC baselines are Riemannian ULA and MALA (3500 parallel chains, MALA step adapted to a 0.574 acceptance rate), and all methods are scored against exact samples of νM\nu_{\mathcal M} drawn with Wood's algorithm. Under matched initialization, R–ASBS still performs better than all energy-based samplers (see Table 2 for metrics and Fig. 2 for a visual representation) and outperforms Riemannina Flow Matching (RFM) on 4/5 metrics. RFM is included for additional comparison but belongs to a different class of samplers, which assume data rather than a differentiable energy function.

In general, the numerical section has been significantly extended to provide empirical evidence of the performance of the proposed approach.

Dear Reviewer d3sk,

Thank you again for your detailed and constructive review. As the author–reviewer discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns.

Reviewer_mW8YOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends a recent diffusion-based sampler, the adjoint Schrödinger bridge sampler, from flat space to distributions that live on smooth curved surfaces, e.g., the set of matrices with orthonormal columns. The sampler is built so that every generated point stays on the surface by construction, instead of being penalized or projected. The authors derive the continuous time control formulation on the surface and replace the exact geometric quantities with computable approximations.

Critical strengths
  • The problem is well chosen. Hard constraints are common, and satisfying constraints by construction is a real advantage over penalty methods.
  • The theoretical development is careful. The matching identities are derived rather than assumed.
  • The paper is honest that the practical algorithm uses approximations. It does not claim exactness for the implemented method.
  • The applications include a real data example and a robotics task.
Critical weaknesses
  • The gap between theory and algorithm is large. The implemented method relies on short time heat kernel and transport approximations. The paper does not offer a global convergence guarantee or an error bound.
  • The matrix manifold experiment checks known thermodynamic limits but does not compare against alternative samplers.
  • The rotation experiment compares only against the true objective value, not against fast certifiable solvers. A sampler that needs 10,000 samples to match a solver does not seem to be competitive.
  • The effect of curvature and bridge length on the approximation error is not studied in detail.
Questions for authors

The method is limited to small surfaces where the exponential map and transport are cheap. The extention needs a projection or Newton solve, correct?

2 comments
AuthorsOpenReview ↗

Response to Reviewer mW8Y

All the changes in the manuscript resubmission are marked in magenta to facilitate the review. We also updated the GitHub repository.

Reviewer Q. 1

Author's response: We agree, and the manuscript states this explicitly (Sec. 3.4, App. F.3, Conclusion): Theorem 1 and Lemmas 1–2 characterize the exact continuous-time targets, while the algorithm replaces the heat-kernel score, the reference bridge and damped stochastic transport with local surrogates, for which we do not claim a global convergence bound. What we provide is: (i) short-time support for these surrogates, since the tilted endpoint law stays within geodesic distance O(r)O(\sqrt r) with Gaussian tails (Prop. 2), and damped transport differs from parallel transport only by a bounded finite-variation factor (App. D); (ii) an explicit account of the errors that do not vanish under mesh refinement (App. F.1); (iii) an empirical measurement of the final error against exact target samples (Table 2).

This reflects a deliberate trade-off between generality and exactness. Methods specialized to structured geometries, e.g. unimodular Lie groups (Ye and Chirikjian; Ye, Jayaraman and Chirikjian), reduce approximations by exploiting closed-form operations. We instead target general embedded manifolds, as illustrated by the four applications (spheres, Stiefel manifold, closed-loop kinematics, SO(3)\mathrm{SO}(3)).


Reviewer Q. 2

Author's response: The revised Sec. 4.2 compares Extended R–ASBS with two Riemannian MCMC baselines on V(4,2)V(4,2), Riemannian ULA (Cheng et al., 2022) and geodesic Monte Carlo (Byrne and Girolami, 2013), all initialized from the same Haar-distributed particles. Figure 3 reports the expected energy of each method across β\beta, together with the two exact thermodynamic limits, showing consistency across the methods.


Reviewer Q. 3

Author's response: We agree that, as a pure optimizer, a sampler is not meant to compete with dedicated solvers; the Wahba problem tests whether the learned law concentrates at the global TLS minimizer under heavy outliers. Following the comment, App. G.4 (Table 7) now compares against multistart Riemannian gradient descent (RGD) and Riemannian MALA. RGD reaches the TLS minimizer, and Extended R–ASBS attains the same rotation accuracy (e.g. 0.83∘0.83^\circ vs. 0.81∘0.81^\circ at 7575% outliers, with a relative TLS gap below 10−510^{-5}), whereas MALA with 16×16\times the gradient budget stays at 2.1∘2.1^\circ.

Moreover, 10,000 samples are not required: with Ns=500N_s=500 the best error is already 0.85∘0.85^\circ (Table 5), and the sampled law itself is concentrated, with ≈94\approx 94% of the samples within 2∘2^\circ of q⋆q^\star for every NsN_s (Table 6). After training, sampling is amortized, with cost linear in the number of samples (App. F.4).


Reviewer Q. 4

Author's response: For the transport surrogate, writing damped transport as Wt,1=PTt,1 Qt,1\mathcal{W}_{t,1}=\mathrm{PT}_{t,1}\thinspace Q_{t,1} (App. D), Grönwall's inequality applied to the linear ODE for QQ gives

∥(Wt,1)∗−(PTt,1)∗∥op≤eK2r−1,K:=sup⁡x∈M∥Ric⁡x♯∥op,r:=∫t1σs2 ds.\big\Vert(\mathcal{W}_{t,1})^{\ast}-(\mathrm{PT}_{t,1})^{\ast}\big\Vert_{\mathrm{op}} \le e^{\frac{K}{2}r}-1, \qquad K:=\sup_{x\in\mathcal{M}}\Vert\operatorname{Ric}^{\sharp}_x\Vert_{\mathrm{op}}, \quad r:=\int_t^1\sigma_s^2\thinspace ds.

The error therefore grows with the Ricci curvature and with the remaining bridge variance rr, and vanishes as t→1t\to1. On the unit sphere Sn\mathbb S^n it is explicit: the omitted factor is e−n−12re^{-\frac{n-1}{2}r} (App. F.1). Projection-as-transport adds an extrinsic error: on S1\mathbb S^1, where Ric⁡≡0\operatorname{Ric}\equiv0, it still contracts by e−12re^{-\frac12 r} (App. F.1). The heat-kernel score uses the first-order expansion of Ludewig (2019), whose remainder is of order A1=∫01σs2dsA_1 = \int_0^1 \sigma_s^2 ds, and Prop. 2 keeps the endpoints within O(r)O(\sqrt{r}).

These dependencies motivate the choice A1∈[0.4,0.7]A_1\in[0.4,0.7] (App. G.5): a small A1A_1 makes the corrector target, of order 1/A11/A_1, dominate the terminal adjoint, whereas a large A1A_1 degrades the short-time approximations.


Reviewer Q. 5

Author's response: This interpretation is correct. We distinguish two settings. When Exp/Log maps and parallel transport are available in closed form (e.g. Sn\mathbb S^n, Tn\mathbb T^n, SO(3)\mathrm{SO}(3)), R–ASBS uses them (Sec. 3.4). For implicitly defined manifolds c(x)=0\\{c(x)=0\\}, Extended R–ASBS replaces them with a minimum-norm Newton retraction, projection-as-transport and a projected-chord corrector (Sec. 3.5, App. F.1), at a cost of O(CJc+m2d+m3)O(C_{J_c}+m^2d+m^3) per particle and projection (App. F.4). This covers settings without geometric primitives, such as the Stiefel manifold and the 10-joint closed-loop chain (Secs. 4.2–4.3).

Dear Reviewer mW8Y,

Thank you again for your detailed and constructive review. As the author–reviewer discussion period ends on November 18 (AoE), we would be grateful if you could let us know whether our responses have addressed your concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.