acceptodds
Under review as a conference paper at ICLR 2027

Intrinsic Geometry Matching: learning observation-equivariant process geometry

Abstract

Real data are often arbitrary, indirect observations of an unknown underlying process, and those observations can distort its geometry. Existing geometric methods equip data with the induced geometry of the ambient space, but this entangles the true geometry of the process with the distortion of the observation. To address this, we introduce **Intrinsic Geometry Matching**: a scalable approach to learning the intrinsic geometry of the underlying process. Specifically, we learn the *carré du champ* operator of the process, allowing us to recover its diffusion geometry. Our method uses the fact that data often come with natural perturbations, like steps in a time series, to recover the intrinsic carré du champ while transforming naturally with changes in observation: it is **observation-equivariant**. Furthermore, since this equivariance relates the geometry of different spaces, we can train a model using additional sources of data, even when that data lies in another space. We evaluate IGM on molecular dynamics, image, and video data and find better performance and scaling than existing neural and graph-based geometric tools.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Geometric deep learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_XniDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose Intrinsic Geometry Matching (IGM), a framework for learning the intrinsic geometry of stochastic process from observed data perturbations. By estimating CDC, IGM aims to recover process geometry without knowing the underlying state space or ground-truth geometry.

The authors introduce a sample-wise learning objective based on Bregman divergences and suggest using LogDet for keeping observation invariance. They further extend the framework to learn geometry from observations in different spaces. Experiments on synthetic manifolds, images, videos, and molecular dynamics demonstrate promising geometry recovery and practical applications.

Critical strengths
  1. The paper provides detailed mathematical principled framework. The authors develop a coherent mathematical formulation connecting stochastic processes, CDC operators, and observation equivariance. The theoretical analysis provides a clear foundation for proposed learning objective. Based on this construction, the choice of LogDet divergence is theoretically well motivated. The authors clearly establish a characterization of LogDet as the unique observation-invariant Bregman divergence under some assumptions.
  2. The cross-space geometry learning framework is interesting. The authors proposed framework that enables learning geometric information from observations in different spaces. In particular, the identifiability analysis for combining multiple partial observations is an interesting contribution.
Critical weaknesses
  1. Experiments are not well organized. The video experiments do not fully disentangle the effects of perturbation choice and geometry learning. In section 5.3, IGM is trained on temporal transitions, while the Gaussian-based baseline is trained on denoising pairs constructed from the same video dataset. Since the evaluation measures the likelihood of held-out temporal transitions, the comparison naturally favors a model trained on temporal transitions. While this comparison demonstrates the value of process-specific perturbations, it does not fully isolate the advantage of the proposed geometry learning method. In addition, It would be good to suggest additional ablation for the molecular trajectory experiment. In section 5.4, the proposed IGM bridge combines a learned conditional mean dynamics model with CDC-based endpoint conditioning. However, the baseline is latent linear interpolation (LERP), which does not use the learned mean dynamics, rather it only interpolate endpoints linearly. Therefore, it is unclear how much of the reported improvement is attributable to the learned CDC rather than the mean dynamics model.
  2. The theoretical scope of the framework could be clarified. The theoretical formulation relies on diffusion Markov process, while the proposed method is also applied to general data-driven perturbations, including image augmentations. It is unclear under what conditions arbitrary perturbation kernels correspond to diffusion processes with a nontrivial infinitesimal CDC. In particular, general finite-scale augmentations may not satisfy the diffusion assumptions or the chain rule used in the theoretical development.
Questions for authors

Since different Bregman divergences share the same population minimizer for conditional covariance estimation, could the authors further investigate the practical benefits of LogDet's observation-invariance under finite data and model capacity? For example, comparing equivariance error or observation consistency between LogDet and Frobenius might be interesting.

5 comments

We thank the reviewer for the thoughtful review and the concrete suggestions. We appreciate that the reviewer finds the mathematical framework principled and coherent, the characterisation of LogDet as the unique observation-invariant Bregman divergence well motivated, and the cross-space framework interesting. The review asks us to further separate the choice of perturbations from the geometry-learning method, to separate the learned mean from the CDC in the molecular bridges, and to compare the observation consistency of LogDet and Frobenius. We have added experiments for all three. They are described in our general response ("Response to All Reviewers", Parts A and B) and summarised below, together with our answer on the theoretical scope.

[W1.1] Separating the perturbations from the geometry-learning method. We agree that evaluating on held-out transitions favours models trained on transitions. This is the point of this evaluation, which tests the choice of perturbations, as LD-RMM differs from IGM only in its perturbations (Principle 2). The two models therefore learn genuinely different geometries, IGM the intrinsic geometry of the process and LD-RMM the ambient geometry of the observations, a real-world analogue of the synthetic experiments. As there, they can still be compared by how well they capture the low-rank tangent spaces of the data manifold, which explained variance approximates (Sec. 5.3 (ii)).

The choice of loss is tested by keeping the perturbations fixed (Principle 3). With known ground truth, the advantage of LogDet over Frobenius grew with the intrinsic dimension (Sec. 5.1), which is why we adopted LogDet for the real-world experiments . On real transitions, IGM attains better held-out NLL and explained variance than FGM on video and molecular dynamics (Response to All Reviewers: [A.3]) We also consider a classical nearest-neighbour estimator of the local transition covariance, with up to 5 million reference frames, which approximates the same intrinsic object. This has a significant performance-cost trade-off, and stays behind both neural models on video while becoming substantially more expensive in time and memory as the dataset increases in size (App. D.5).

[W1.2] Mean and CDC in the MD bridges. Thank you for this suggestion, which helped us sharpen our claim (general response, B). The drift sets the mean motion, and the CDC describes the variability around it, so replacing the CDC with an isotropic or any constant covariance lowers the improvement of the bridge over LERP by only about 2 percentage points (from 30.1% to 28.2%). This comparison scores a single mean path against a single held-out trajectory, which is one realisation of a stochastic process, while the CDC acts on how such paths spread around the mean. We therefore now evaluate the bridges as a distribution, and held-out MD paths are more likely under IGM bridges than under isotropic bridges with the same mean, on all 20 test trajectories at every rank and lag (App. D.4). Claim (ii) of Sec. 5.4 now reads "this geometry improves the sampling of entire transition paths".

[W2] Theoretical scope for general perturbations. This is a really interesting question about modelling assumptions. Our initial aim was to discuss the geometry of general Markov processes, which significantly generalises the manifold hypothesis. Diffusion geometry provides a theoretical model for geometry that uses only the carré du champ, and so simply requires that as input. In this sense, our finite-scale objects and methods are well defined for general perturbations, because an arbitrary perturbation kernel PtP_t induces a finite-scale generator (I−Pt)/t(I - P_t)/t and CDC GtG_t as its first two moments. When we evaluate geometric objects like the intrinsic derivative via Gt∇(f)G_t \nabla(f), we do have the chain rule by construction. A separate question, which you rightly identify, is whether the underlying process we used to construct GtG_t itself has the chain rule. This is really a modelling assumption, but applies in very general cases of Itô processes even when the coefficients are history-dependent (so the process is not Markov), which seems to encompass all the examples we consider here.

[Q1] Observation consistency of LogDet and Frobenius. We followed this suggestion directly (Response to All Reviewers: [A.2]) ). As the reviewer notes, all Bregman losses share the population minimiser, so any benefit of LogDet must appear with finite data and capacity. We observe the torus process of Sec. 5.1 through ill-conditioned linear distortions, with condition number up to 10001000, retrain IGM and FGM with the same data and architecture (Sec. 5.1 (ii)). We then measure the observation consistency of each method, the relative distance between the geometry learned from distorted data and the one learned from undistorted data, after pulling both back to latent coordinates. This is the equivariance error the reviewer suggests. IGM agrees with its undistorted estimate, and remains far more consistent than FGM, for every fixed loss regulariser we tested, λ∈{0.1,1,10}\lambda \in \{0.1, 1, 10\}. At the level of the loss (Response to All Reviewers: [A.1]), the regularised LogDet ranks candidate estimates almost identically in distorted coordinates while λ\lambda is small compared to the smallest eigenvalue of the target, and better than Frobenius at every λ\lambda (App. A.1). On video, LogDet also trains more stably, with gradient norms within 1414 times their median against up to 436436 for Frobenius (App. D.3, Response to All Reviewers: [A.4]).

We hope these additions address your concerns. If so, we would be grateful if you would consider raising your score, and we would be glad to answer any further questions.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer XniD, thank you again for your time and constructive feedback. We have addressed your concerns with new experiments and revisions. Since the discussion period is about to end, we would appreciate it if you could consider our improved manuscript and take these changes into account in your assessment. We would be very happy to respond to any further questions you may have.

With these revisions, we now train Frobenius and nearest-neighbour baselines on the same transitions as IGM, which disentangles the geometry-learning method from the choice of perturbations. In the molecular bridges, we separate the contribution of the learned mean from that of the CDC. Following your suggestion, we also compare the observation consistency of LogDet and Frobenius under ill-conditioned distortions. With the same finite data and capacity, LogDet stays far more consistent, and on video it also trains more stably. Finally, we clarify which parts of the theory hold for general perturbations (details in our reply above).

Reviewer_RLzvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the issue that standard geometric learning methods conflate the intrinsic geometry of an underlying dynamical process with distortions from the ambient observation space. Using diffusion geometry, the authors propose Intrinsic Geometry Matching, which models the underlying process via the carré du champ operator and learns its coordinate matrix from observed perturbations such as time-series steps or image augmentations. The paper shows that matching rank-one increment targets works properly with Bregman divergences, and identifies the LogDet divergence as uniquely invariant under general linear congruence, making the objective equivalent to heteroscedastic Gaussian negative log-likelihood. The approach is demonstrated on synthetic manifold dynamics, image deformations, video frames, and molecular dynamics trajectories.

Critical strengths
  1. The theoretical framing is clean and well-motivated. Formulating the learned geometry via the pushforward of an intrinsic operator and using linear congruence invariance to justify the LogDet loss over the Frobenius norm provides a coherent story that also unifies several existing metric learning and denoising methods.

  2. The empirical demonstrations span diverse domains, and the molecular dynamics application is particularly interesting, showing that pushing the learned latent geometry through a decoder recovers physically meaningful residue mobility patterns.

Critical weaknesses
  1. The experimental comparisons consistently conflate the data source with the estimation method. In the video and image tasks, the proposed method is trained on actual transitions or structured deformations, whereas the primary baseline only receives isotropic Gaussian noise on static frames. In the synthetic manifold experiment, the method even uses the ground-truth observation differential, despite the introduction noting that Jacobian knowledge is rarely available in practice. Furthermore, the video and molecular dynamics models are evaluated on held-out transition likelihood, which is the exact training objective of the proposed method, while comparing against baselines never trained on transitions. Similarly, in the molecular bridge experiment, the appendix reveals that trajectories are generated by rolling out a learned conditional mean network and then applying covariance conditioning, yet the only baseline is naive linear interpolation without the mean network. To isolate the value of the proposed LogDet covariance estimator, the paper needs fair baselines trained on the exact same transition data across the real-world tasks, such as Frobenius regression or local sample covariance estimators, as well as an isotropic bridge baseline that uses the same learned mean rollout.

  2. The methodological novelty over classical stochastic process identification is somewhat overstated, and the practical invariance guarantees are weakened by implementation approximations. Once the objective is written as heteroscedastic Gaussian likelihood on increments, it becomes very close to standard diffusion matrix estimation in stochastic differential equations and Kramers-Moyal expansions, which should be discussed more transparently. Moreover, while theoretical invariance holds for full-rank matrices, the practical implementation relies on a low-rank parameterization plus isotropic regularization, which breaks exact linear congruence invariance. Finally, the paper lacks an analysis of how additive measurement noise and finite step sizes bias the increment covariance in real data.

Questions for authors
  1. For the video and molecular dynamics tasks, can you report a Frobenius regression baseline and a local nearest-neighbor covariance estimator trained on the exact same transitions? In addition, how does your molecular bridge compare against a bridge that uses the exact same learned mean rollout but with an isotropic covariance?

  2. Given that the practical model uses a low-rank parameterization and additive identity regularization, how much of the theoretical linear invariance remains in practice? Can you show empirically under an ill-conditioned linear transformation that your regularized estimator recovers the transformed metric while Frobenius regression fails?

  3. How sensitive is the estimator to additive observation noise and the choice of time lag in real data, where measurement noise can dominate small increments?

  4. Could you clarify in the related work what is fundamentally new in your training objective compared to classical Gaussian likelihood estimation of state-dependent diffusion coefficients in stochastic system identification?

  5. There are minor naming inconsistencies regarding the Gaussian baseline acronym in the video section and slight numerical discrepancies in the mean Fréchet distances between the main text and the appendix that should be fixed.

4 comments
AuthorsOpenReview ↗

Part 1 [W1]

We thank the reviewer for the detailed review and the concrete requests. We appreciate that the reviewer finds the theoretical framing clean and well motivated, with congruence invariance justifying LogDet over Frobenius and unifying existing metric learning and denoising methods. The review asks for baselines trained on the same transitions, an isotropic bridge with the same learned mean, evidence of how much invariance remains in the regularised model, and an analysis of measurement noise and lag. We have added experiments for all four. They are described in our general response ("Response to All Reviewers", Parts A and B) and summarised below, together with our answers to the remaining questions.

[W1.1 & Q1] Baselines trained on the same transitions. The comparison with LD-RMM changes the data source by design, since it tests the choice of perturbations with the same LogDet loss (Principle 2). The two models learn genuinely different geometries, IGM the intrinsic geometry of the process and LD-RMM the ambient geometry of the observations, as in the synthetic experiments, and explained variance compares them by how well they capture the tangent spaces of the data (Sec. 5.3 (ii)). To isolate the estimator, as the reviewer requests, we now train a Frobenius regression (FGM) and a classical nearest-neighbour estimator of the local transition covariance on exactly the same transitions as IGM. With the same data and architecture, IGM attains better held-out NLL and explained variance than FGM on video and molecular dynamics, at every rank on molecular dynamics, although explained variance is the training objective of neither loss (Response to All Reviewers: [A.3]). The nearest-neighbour estimator targets the same intrinsic object as IGM. On video, even with 55 million reference frames, it stays behind IGM and LD-RMM, with −2.32-2.32 against −3.38-3.38 bits/dim for IGM and 0.360.36 against 0.850.85 of the held-out transition variance in 128128 directions, while its storage grows to 63.863.8 GiB against 0.080.08 GiB for the network (App. D.5).

[W1.2] The ground-truth differential in the synthetic experiment. In short, the differential is used only in the synthetic experiment, where it gives an exact target that isolates the loss, while all real-world experiments, on images, video and molecular dynamics, train IGM directly on observed transitions or augmentations, with no differential. The differential is also available in many cases of interest, whenever the observation map is a known differentiable function, such as a neural network. We have now also retrained the synthetic experiment on observed transitions instead of the differential, with all other settings fixed. A finite-lag transition also carries the curvature of the embedding, which sets a floor on the error of both methods. For d≤8d \leq 8 both reach this floor, while for d≥16d \geq 16 IGM stays on it and FGM does not, with latent cometric errors of 0.170.17 and 0.350.35 at d=32d = 32 (App. D.1).

[W1.3 & Q1] Isotropic bridge with the same learned mean. Thank you for this request, which helped us sharpen our claim (Response to All Reviewers: B). In the learned forward process, the drift sets the mean motion and the CDC describes the variability around it, so the location of a transition path is mainly set by the drift. Indeed, replacing the CDC by an isotropic or any constant covariance lowers the improvement of the bridge over LERP by only about 2 percentage points (from 30.1% to 28.2%). This comparison scores a single mean path against a single held-out trajectory, which is one realisation of a stochastic process, while the CDC acts on how such paths spread around the mean. We therefore now evaluate the bridges as a distribution, and held-out MD paths are more likely under IGM bridges than under isotropic bridges with the same mean, on all 2020 test trajectories at every rank and lag (App. D.4). Claim (ii) of Sec. 5.4 has been sharpened to "this geometry improves the sampling of entire transition paths".

AuthorsOpenReview ↗

Part 2 [W2, Q2, Q3, Q4, Q5]

[W2.1 & Q4] Relation to classical diffusion estimation. This is an important distinction, and we have added a further experiment benchmark to clarify it. Classical approximators to diffusion processes use covariance estimators over a time window, and extend this to test data via kNN-based graph estimators. As such, these techniques suffer from the same issues as general graph-based geometry: they are not very scalable, they use pairwise distances which are cursed by dimension, and they are transductive within the data rather than inductive.

We construct a thorough ablation against this technique in App. D.5, measuring the performance against our neural methods and the relative costs. We evaluate on the video dataset, which is large and high-dimensional, and use the same test data. This challenges all aspects of the kernel method in practice, and it does not attain NLL or explained variance results close to the neural methods. The computational time and memory also explode as the dataset size increases, leading to prohibitively expensive models. We feel that these results clearly demonstrate the value of amortised neural methods over classical techniques, and open several opportunities for extending classical work to new domains with these tools.

[W2.2 & Q2] Invariance of the regularised model. In short, only the isotropic term λI\lambda I can break the invariance, and under ill-conditioned distortions the regularised estimator recovers the transformed metric while Frobenius regression fails (Response to All Reviewers: [A.1] and [A.2]). The low-rank factor keeps its form under any linear change of coordinates. The regulariser λI\lambda I enters both arguments of the loss, so it leaves the minimiser unchanged, where invariance remains exact, and near the minimiser the loss stays approximately invariant while λ\lambda is small compared with the smallest eigenvalue of the target (App. A.1). Empirically, we observe the torus process of Sec. 5.1 through linear maps with condition number up to 10001000 and retrain IGM and FGM on the distorted data (Sec. 5.1 (ii)). Pulled back to latent coordinates, IGM recovers the true metric with error at most 0.0430.043 up to κ=100\kappa = 100 and 0.380.38 at κ=1000\kappa = 1000, while FGM fails with errors of 1.111.11 and 6.516.51. IGM remains far more consistent than FGM for every fixed loss regulariser we tested, λ∈{0.1,1,10}\lambda \in \{0.1, 1, 10\}, and at the level of the loss, the regularised LogDet ranks candidate estimates almost identically in distorted coordinates while λ\lambda is small compared to the smallest eigenvalue of the target, and better than Frobenius at every λ\lambda (App. A.1).

[W2.3 & Q3] Measurement noise and lag. In short, measurement noise biases the learned metric at short lags, so the best lag grows with the noise level, while the learned tangent stays robust. We added i.i.d. Gaussian measurement noise of standard deviation σ∈{0,0.03,0.1,0.3}\sigma \in \{0, 0.03, 0.1, 0.3\} to every observation in the torus experiment (d=8d=8, the tube has radius 11), at both ends of each transition, and trained IGM on these noisy transitions at three lags τ∈{0.0125,0.05,0.2}\tau \in \{0.0125, 0.05, 0.2\}. The finite-lag bias grows with τ\tau. Without noise, the latent cometric error is 0.0210.021, 0.0750.075 and 0.240.24 at the three lags. Measurement noise adds about σ2/τ\sigma^2/\tau to the increment covariance and therefore dominates short lags. The best lag thus grows with the noise level. The learned tangent space is much more robust because measurement noise mostly shifts the eigenvalues of the increment covariance rather than its eigenvectors. We have added this analysis to the App. D.1, with the full table.

[Q5] Minor fixes. Thank you, minor discrepancies have been fixed in the manuscript. The Gaussian baseline is now called LD-RMM throughout.

We hope these additions address your concerns. If so, we would be grateful if you would consider raising your score, and we would be glad to answer any further questions.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer RLzv, thank you again for your time and constructive feedback. We have addressed your concerns with new experiments and revisions. Since the discussion period is about to end, we would appreciate it if you could consider our improved manuscript and take these changes into account in your assessment. We would be very happy to respond to any further questions you may have.

With these revisions, we now train Frobenius and nearest-neighbour baselines on the same transitions as IGM. In the molecular bridges, we separate the contribution of the learned mean from that of the CDC. We also show that the regularised loss is exactly invariant at its minimiser and nearly invariant around it, and that under ill-conditioned distortions the regularised estimator recovers the transformed metric while Frobenius regression fails. Finally, we analyse how measurement noise and the choice of lag bias the estimate, discuss the relation to classical diffusion estimation, and retrain the synthetic experiment on observed transitions instead of the differential (details in our replies above).

Dear Reviewer RLzv,

We hope that we have addressed your concerns.

With less than 24 hours remaining before the discussion deadline (November 18), we would be grateful if you would incorporate these revisions into your final assessment, or ask any remaining questions for us to clarify.

Thank you again for your time and consideration.

Reviewer_a28cOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes IGM, which learns a process's local geometry by matching the carré du champ with Bregman losses, and gives theory on why LogDet is preferred and when the geometry is identifiable. Experiments span synthetic, image, video, and molecular data. However, coordinate sensitivity is not measured and the molecular gains do not isolate the covariance.

Critical strengths
  1. The goal is meaningful. The paper learns a locally varying second-order structure, the carré du champ (CDC), from real process transitions or meaningful perturbations, and uses it to describe the geometry of the process. The paper asks whether this geometry still reflects the underlying process when the way it is observed changes. This question is well motivated in both theory and practice.
  2. The method, loss, and theory fit together. Conditional metric matching for learning the CDC with neural networks was introduced in earlier work on Riemannian metric matching. This paper extends it in three ways. First, it generalizes the loss to Bregman divergences. Second, under the stated conditions on positive-definite matrices and invertible congruence transforms, it shows that LogDet is unique and relates it to the Gaussian NLL. Third, it gives identifiability conditions for cross-space training and partial observation, which widen the method's use. So the contribution is mainly in these links and how they are turned into an algorithm, not in introducing the CDC, LogDet, or neural geometry estimation.
  3. There is controlled evidence for the choice of loss. Comparing LogDet and Frobenius under the same process supervision is more convincing than comparing across different sources of supervision. In Figure 2(b), at intrinsic dimension 32, the latent-cometric error is 0.029 for LogDet and 0.390 for Frobenius. The multi-view experiments in Section 4.3 also confirm the identifiability threshold in the settings tested. One small note: standard CBMM can use observed transition samples directly and does not need a known Jacobian. Using differentials is just another way to set up cross-space learning.
  4. The applications are broad, and some evaluations are careful. The experiments cover synthetic processes, images, video, and molecular dynamics. In the molecular experiments, the data is split by trajectory, and uncertainty is reported over 20 test trajectories. The paper also varies the covariance rank while keeping the same mean model, and measures the explained variance after removing the mean (Eq. 28). This gives some evidence about the covariance that does not depend on how well the mean fits.
Critical weaknesses
  1. How much the regularized model depends on the coordinates is not measured. The paper says it uses a finite-scale approximation, and that invariance to nonlinear observations holds only asymptotically under suitable conditions. However, the low-rank-plus-scalar-floor model, and the version regularized with a fixed λI, generally do not keep their form under arbitrary congruence transforms. The smoothing ablation in Figure 10 does not show whether the learned geometry stays consistent when the same process is observed in different coordinates.
  2. The molecular path results do not separate the effect of the covariance. The rank sweep with a shared mean, the explained variance, and the NLL all support the covariance structure. But the bridge in Eq. 29 uses both the learned mean and the learned covariance, and it is mainly compared with linear interpolation. So it is unclear how much of the reported 30.1% path improvement comes from each.
  3. One geometric claim needs conditions in the non-injective case. Eq. 5 uses the posterior-averaged CDC, and Section 4.3 and Appendix C discuss identifiability limits under partial observation. However, right after Eq. 5, the paper says that geodesics in observation space are pushforwards of the true geodesics. It is not stated whether this still needs injectivity or some other condition.
Questions for authors
  1. Which properties under observation changes still hold after adding the scalar floor or fixed isotropic smoothing? Can you check whether the estimates agree across different observation coordinates of the same process?
  2. If the path in Eq. 29 uses the same mean but an isotropic or constant covariance, how much of the 30.1% gain remains?
  3. What extra conditions does the geodesic pushforward claim after Eq. 5 need when the observation map is not injective?
  4. Does the LogDet advantage over Frobenius in Figure 2(b) also hold on a real transition task with the same data and architecture?
5 comments

We thank the reviewer for the careful reading and the constructive questions. We appreciate that the reviewer finds the question of observation invariance well motivated, sees the method, loss and theory as fitting together, and values the controlled evidence for the choice of loss and the breadth of the applications. The review names two concerns, that coordinate sensitivity is not measured and that the molecular gains do not isolate the covariance. We have added controlled experiments for both, in distorted coordinates and with the bridge mean held fixed. They are described in our general response ("Response to All Reviewers", Parts A and B) and summarised below, together with our answers to the remaining questions.

[W1 & Q1] Coordinate dependence of the regularised model. In short, we now measure coordinate sensitivity both for the loss and the trained models, and IGM learns the same geometry in distorted coordinates (Response to All Reviewers: [A.1] and [A.2]). The model keeps its form under any linear change of coordinates, so only the isotropic terms break the exact invariance. The fixed ridge enters both arguments of the loss, so it leaves the minimiser unchanged, and near the minimiser the loss stays approximately invariant while λ\lambda is small compared with the smallest eigenvalue of the target (App. A.1). To check whether the estimates agree, we observe the torus process of Sec. 5.1 through ill-conditioned linear distortions, with condition number κ\kappa up to 10001000, and retrain on the distorted data (Sec. 5.1 (ii), Fig. 4). IGM agrees with its undistorted estimate, and remains far more consistent than FGM, for every fixed loss regulariser we tested, λ∈{0.1,1,10}\lambda \in \{ 0.1, 1, 10 \}.

[W2 & Q2] Mean and CDC in the MD bridges. Thank you for this question, which helped us sharpen our claim (Response to All Reviewers: B). In the learned forward process, the drift sets the mean motion and the CDC describes the variability around it, so the location of a transition path is mainly set by the drift. Indeed, replacing the CDC by an isotropic or any constant covariance lowers the improvement of the bridge over LERP by only about 2 percentage points (from 30.1% to 28.2%). This comparison scores a single mean path against a single held-out trajectory, which is one realisation of a stochastic process, while the CDC acts on how such paths spread around the mean. We therefore now evaluate the bridges as a distribution, and held-out MD paths are more likely under IGM bridges than under isotropic bridges with the same mean, on all 20 test trajectories at every rank and lag (App. D.4). Claim (ii) of Sec. 5.4 has been sharpened to "this geometry improves the sampling of entire transition paths".

[W3 & Q3] Injectivity of the geodesic pushforward. Thank you for this correction. The claim after Eq. 5 requires Φ\Phi to be injective, and we have added this condition in Sec. 3, in the sentence after Eq. 5. Without injectivity, Φ♯Γ\Phi_\sharp\Gamma is the posterior average of Eq. 5, and its geodesics are in general not images of latent geodesics. The claim may still hold in special non-injective cases, which we do not consider here.

[Q4] LogDet versus Frobenius on real transitions. The advantage also holds on real transition tasks (Response to All Reviewers: [A.3]). We retrained the video and MD models with Frobenius loss (FGM), keeping the data and architecture fixed. IGM attains better held-out NLL and explained variance than FGM on both (Sec. 5.3 and 5.4).

We hope these additions address your concerns. We would be grateful if you would then consider increasing your score, and would be very glad to answer any further questions.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer a28c, thank you again for your time and constructive feedback. We have addressed your concerns with new experiments and revisions. Since the discussion period is about to end, we would appreciate it if you could consider our improved manuscript and take these changes into account in your assessment. We would be very happy to respond to any further questions you may have.

With these revisions, we now show that the regularised loss is exactly invariant at its minimiser and nearly invariant around it, which we confirm in experiments under ill-conditioned distortions. We also separate the learned mean from the CDC in the molecular bridges, compare LogDet and Frobenius on real transitions, and add the injectivity condition after Eq. 5 (details in our reply above).

Dear Reviewer a28c,

We hope that we have addressed your concerns.

With less than 24 hours remaining before the discussion deadline (November 18), we would be grateful if you would incorporate these revisions into your final assessment, or ask any remaining questions for us to clarify.

Thank you again for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers - Part A

We are grateful to the reviewers for their careful and thoughtful evaluation of our work, and many useful suggestions that have greatly improved it. We appreciate that all reviewers noted the clear motivation for the problem, theoretical justification and framing of the method, and broad experimental evaluation.

There were two central questions raised by all the reviewers that we have substantially addressed in the new version of the paper, with multiple new experiments. We provide a general response to those questions here, in addition to the specific replies to each reviewer.

A. LogDet vs Frobenius losses in practice. Is LogDet more observation invariant? Does this hold when it has isotropic smoothing?

We thank the reviewers for raising these questions, and we have added much more thorough empirical evaluations across four new studies. The results add much stronger evidence for the advantages of LogDet: it performs better than Frobenius consistently, the regularised model remains nearly invariant, and training is more stable.

[A.1] The regularised loss stays nearly invariant (a28c W1/Q1, RLzv W2/Q2). We analyse this in a new App. A.1. While the true LogDet loss is conjugation-invariant, we sometimes add a small isotropic regulariser λI\lambda I to the model and target in the loss. The model keeps its form under any change of coordinates, but the isotropic term λI\lambda I breaks the exact invariance, because in the original coordinates the regulariser becomes λ(J⊤J)−1\lambda(J^\top J)^{-1}. The regulariser does not change the minimiser, where the invariance remains exact, but near it, the regularised loss weighs errors relative to the eigenvalues of the target where these exceed λ\lambda, as LogDet does, and in absolute terms where they fall below λ\lambda, as Frobenius does. It therefore stays approximately invariant while λ\lambda is small compared with the smallest eigenvalue. To evaluate this, we sample a ground truth matrix GG and 200200 perturbations GiG_i, and rank their distance to GG using LogDet with different regularisation levels λ\lambda as well as Frobenius. We then distort GG and the GiG_i with increasingly ill-conditioned linear maps and measure how the ranking changes. As expected, the ranking is fixed for λ=0\lambda = 0, only slightly degrades when λ\lambda is small, and even at λ=1\lambda = 1 is better preserved than with Frobenius.

[A.2] Trained IGM learns the same geometry in distorted coordinates, and FGM does not (a28c W1/Q1, RLzv W2/Q2, XniD Q1). We observe the torus process of Sec. 5.1 through increasingly ill-conditioned linear maps and retrain IGM and FGM on the distorted data (Sec. 5.1 (ii), Fig. 4). IGM learns the same geometry as from the undistorted data up to moderate distortions, and remains far more consistent than FGM as the distortion grows, for every regularisation λ\lambda we tested. At condition number 10001000, IGM still recovers the true geometry with error 0.380.38, while FGM fails with error 6.516.51 (App. D.1).

[A.3] LogDet beats Frobenius on real data (a28c Q4, RLzv W1/Q1, XniD W1/Q1). We retrained the molecular dynamics and video models with the Frobenius loss (FGM). On molecular dynamics, IGM has better held-out NLL and explained variance than FGM at every rank, even though FGM is trained for 50% more epochs (Sec. 5.4). On video, IGM again has better NLL and explained variance, and the gap is largest for the smaller-eigenvalue principal components (Sec. 5.3). Explained variance is not the training objective of either loss.

[A.4] LogDet greatly improves training stability (XniD Q1). On video, the gradient norms of FGM spike throughout training, up to 436436 times their median, and require gradient clipping, while those of IGM stay within 1414 times their median (App. D.3).

AuthorsOpenReview ↗

Response to All Reviewers - Part B

B. Separating the learned mean from the learned geometry in the MD bridges

We thank the reviewers for this question, which helped us sharpen our claim. In the learned forward process, the drift sets the mean motion and the CDC describes the variability around it, that is, the directions and magnitudes of stochastic change. The NLL and explained variance in Sec. 5.4 (i) already shows that this geometry is learned from single transitions and is far from isotropic, with the leading 512 directions capturing 74.5% of the residual variance. However, the location of a transition path is mainly set by the drift. Indeed, the submission compared a single mean bridge (Eq. 29) with a single held-out trajectory, and replacing the CDC by an isotropic or any constant covariance lowers its improvement over LERP by only about 2 percentage points (from 30.1% to 28.2%). A held-out trajectory is one realisation of a stochastic process, and the CDC acts on how such paths spread around the mean.

We therefore now evaluate the bridges as a distribution. We sample bridges from the learned transition model that follow the learned drift, draw each step from the CDC at the current state and are guided to the observed endpoint. IGM and isotropic bridges then share the same mean path by construction and differ only in their noise. Scored by their likelihood given their endpoints, held-out MD paths are more likely under IGM bridges on all 20 test trajectories, at every rank and every lag. We have revised Sec. 5.4 accordingly. Claim (ii) now reads "this geometry improves the sampling of entire transition paths", Fig. 9(a) shows the sampled bridges, and the table in Sec. 5.4 reports the Fréchet distance of the shared bridge mean together with the path NLL. App. D.4 describes the sampler, and Table 2 reports all ranks and lags.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.