acceptodds
Under review as a conference paper at ICLR 2027

When Does Learned Posterior Geometry Recover Weighted Manifold Forms?

Abstract

When does accurate local posterior geometry yield accurate weighted function energies? We study a fixed ambient trial space, separating learning error, finite-noise geometric bias, and numerical integration. A gradient-dependent bound connects local matrix error to the weighted form, while circle examples, including an exact Gaussian observation posterior, show how small normal components produce large form errors through ambient extensions. On controlled image-translation manifolds, we compare matched covariance models trained with sample residuals or conditional second moments. These objectives have the same ideal population minimizers but different conditional label randomness. With oracle-assisted centering and known-dimensional filtering, conditional supervision reduces mean linear-form error from to , with improvements numerically distinguished in all four image–initialization pairs. Same-form attribution identifies learning-induced normal error, rather than finite-noise posterior variance, as the dominant raw discrepancy in this cohort. Improving a learned mean instead improves tangent directions, worsens restricted metric scales, and produces large opposing fitting and centering contributions. Together, these findings connect extension-sensitive posterior forms, same-objective supervision effects, and signed geometric errors. They help researchers choose supervision and ambient probes by their effect on weighted energies, beyond aggregate local matrix accuracy.

Primary area:
learning theory
Keywords:
Posterior geometry, manifold learning, weighted Dirichlet forms, conditional moment regression, ambient extensions, geometric error analysis

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_EZMxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is answering the question “When does accurate local posterior geometry yield accurate weighted function energies?” Prior work shows that under small-noise regimes, normalized covariance approaches the tangent projector matrix. Some works do this via training a neural network to learn this operator instead of approximating it in traditional ways (e.g. graph Laplacian, diffusion maps). Thus, each point can be assigned a matrix via this learned matrix field. This paper is saying – if we assume that this local matrix field is accurate and accessible at each point, does that mean we can accurately capture global quantities like energy of a function across the whole manifold? The authors conclude that this is not necessarily true,

Critical strengths

The paper is continuing the line of work where people are investigating normalized covariance approaching tangent projector, and what are its implications in terms of performing learning tasks along the manifold as a whole.

The circle example is clean and intuitive as a counterexample.

Critical weaknesses

The introduction reads as a dense list of facts. It does not tell the story of why this is an important or interesting problem. It does not highlight the main contributions of the paper. Also, there are a surprisingly small number of references mentioned in the introduction (only 4). The introduction needs a rewrite that would help the reader understand the bigger picture. Seems like the authors tried to make things compact to fit in 9 pages – I completely understand that this is challenging. But introduction is the most important section so I would suggest not cutting corners here. Also, lines 042-043 contain the word “from” twice.

A picture/drawing of the circle example could be great for illustration purposes.

Questions for authors

Could you please justify why this work is important? What does it add to the community? Why is this a next step to the described line of work? What are its main limitations? The appendix line 1701 mentions that limitations are described in section C.2, but that section is called "Training, numeric posterior, and error definitions” and does not mention limitations.

5 comments
AuthorsOpenReview ↗

Reply to Reviewer EZMx (1/3)

Thank you for the feedback on motivation and presentation. We agree that the introduction should explain the intended use of learned geometry before listing the technical results. The additional analyses now provide a concrete example of both the benefit and the limits of transferring local geometry to outputs.

1. Why the problem matters and what the paper adds

A learned geometric matrix is an intermediate object when the intended output is a function energy or a spectral quantity. In a variational calculation, it enters through ∇f⊤C∇f\nabla f^\top C\nabla f. Matrix errors in directions with large ambient derivatives can therefore have a disproportionate effect. Researchers need to know which additional quantities to assess before using local geometry in such a calculation.

Our contribution has three connected parts:

  1. An obstruction and sufficient conditions. The exact-posterior circle example shows that the same intrinsic function can receive different finite-noise energies through different ambient extensions. The transfer bounds identify gradient and error quantities whose control suffices for form accuracy, while preserving fixed-extension small-noise convergence.
  2. A controlled intervention with measurable output consequences. Matched supervision preserves the ideal population objective while changing final form accuracy. Our additional analysis of the saved concentrated oracle-centered rank-two matrices shows mean worst relative ordered-eigenvalue error decreasing from 14.784% to 1.898% for linear probes and from 2.231% to 0.681% for cubic probes. Each mean covers two images and two initializations.
  3. An explanation of successful and unsuccessful comparisons. Signed attribution separates learning from posterior bias. The new per-case analysis explains the uniform-density reversals through reduced positive tangent/normal compensation against a persistent negative cross contribution. It also retains an eigenspace deterioration despite improved eigenvalues.

The benefit to the community is a way to evaluate learned geometry through the functions and outputs it will actually serve. Local matrix accuracy, additive form accuracy, eigenvalue accuracy, and eigenspace accuracy answer different questions; the paper supplies conditions and controlled evidence connecting them. The new eigenvalue calculations concern actual finite-matrix outputs, with numerical sensitivity checked separately, and required no new training.

AuthorsOpenReview ↗

Reply to Reviewer EZMx (2/3)

2. Main limitations

The experiments use controlled translation orbits with constant positive-definite intrinsic metrics, supplied densities, and fixed linear and cubic dictionaries. We have not established that these probes represent those in realistic applications. The strongest recovery uses oracle-assisted labels and a supplied intrinsic dimension.

The matched cohort contains two images, two covariance initializations, and one noise level. Additional matrix analyses do not enlarge that cohort. Additive-form and eigenspace outcomes remain mixed, and learned centers retain substantial errors. Numerical screens are descriptive sensitivity checks, not statistical tests or rigorous quadrature certificates. The results concern finite trial spaces; full-operator recovery would require additional approximation arguments.

AuthorsOpenReview ↗

Reply to Reviewer EZMx (3/3)

3. Presentation revisions

We will rewrite the introduction to follow the intended downstream quantity, the gap between local and form accuracy, the circle mechanism, and the three contributions above. We will expand the discussion of posterior geometry, ambient extensions, and variational spectral computation where it clarifies that progression.

We will add a geometric circle schematic showing identical intrinsic function values, different normal derivatives, and the resulting leaked energy, to complement the existing energy plot. We will also add an explicit limitations paragraph, repair the appendix limitations cross-reference, and check the reported duplicated “from” against the submitted text.

Reviewer_ScsZOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies when a learned posterior covariance field yields accurate weighted manifold energy forms on a fixed ambient trial space. It derives a gradient-dependent local-to-form error bound and presents circle examples, including an exact Gaussian observation posterior, showing that small local geometric errors can cause large energy errors through sensitivity to ambient extensions. It also separates learning error, finite-noise geometric bias, and numerical integration error. Experiments on controlled image-translation manifolds compare sample-residual and conditional-second-moment supervision, which share the same ideal population minimizers. Under oracle-assisted centering and known-dimensional filtering, the reported mean normalized linear-form error decreases from 24.80% to 3.73% in the two-image, two-initialization cohort. Signed attribution identifies learning-induced normal components as the dominant source of raw form error in this setting. The paper further distinguishes finite-space energy and eigenvalue accuracy from the additional conditions needed for stable eigenspace recovery.

Critical strengths
  1. Clear and informative formulation of the local-to-form transfer problem.

The paper identifies an important distinction between recovering local geometric matrices and accurately evaluating the energies of functions that use those matrices. The ambient-extension examples, including the exact-posterior construction, make the failure mechanism concrete: correct tangent directions and small matrix errors do not by themselves provide extension-independent energy guarantees. The gradient-dependent bound clarifies the additional quantities that must be controlled. This conceptual distinction is the main reason I find the paper informative and potentially valuable.

  1. Carefully structured controlled experiments and error attribution.

The experiments connect the theoretical mechanism to trained covariance models under known geometric references. Matched supervision comparisons preserve the ideal population objective while changing conditional label randomness, and the analysis separately examines learning error, posterior geometric bias, and integration sensitivity. Retaining signed tangent, cross, and normal contributions also helps expose cancellation that aggregate error norms would conceal. These controls strengthen the interpretation of the observed effects within the studied setting.

Critical weaknesses
  1. The practical significance of the identified failure mode remains insufficiently established.

The experimental manifolds are generated by translating individual CIFAR images under known distributions. Although the image content is real, the underlying geometric family is highly controlled, and the evaluated linear and cubic dictionaries are not linked to a concrete downstream application. The paper therefore provides limited evidence about how often comparable extension sensitivity occurs in realistic pipelines or whether it materially affects their final outputs. This limits my assessment of the practical importance of the phenomenon, while leaving its mathematical existence and the controlled observations intact.

  1. The strongest positive recovery evidence has a narrow empirical and oracle-assisted scope.

The main matched supervision result uses two base images and two covariance initializations, with oracle-assisted centering and a supplied intrinsic dimension. The conditional second-moment targets are obtained using the known generative setup. Moreover, the reported improvement is less consistent for cubic functions, and reversals occur in some uniform-density comparisons. These limitations are acknowledged, but they leave uncertainty about the robustness of the supervision effect and the extent to which the findings provide actionable guidance beyond the controlled setting.

Questions for authors
  1. Clarification of application relevance:

Which concrete downstream use cases motivate the chosen ambient trial functions? In particular, can the authors explain whether the normal-gradient magnitudes and extension sensitivity observed for the whitened linear and cubic dictionaries are representative of probes arising in an existing application of learned posterior geometry? A concrete connection would materially strengthen my assessment of the significance of the failure mode.

  1. Clarification of the conditions under which conditional supervision helps:

The concentrated linear comparisons show consistent improvement, whereas the cubic comparisons and some uniform-density cases are less consistent. Do the existing per-case diagnostics identify factors that explain these differences, such as gradient amplification or the balance of tangent, cross, and normal errors? Understanding whether the results support a useful condition for improvement would affect my assessment of how broadly informative the supervision findings are.

6 comments
AuthorsOpenReview ↗

Reply to Reviewer ScsZ (1/3)

Thank you for the specific questions about application relevance and mixed supervision outcomes. We have completed two analyses of the saved matrices: a comparison of actual finite-space spectral outputs and a per-case signed explanation of the supervision reversals.

1. Concrete spectral consequences and the remaining application limitation

A concrete use of a learned field is variational spectral computation:

Kij=Eν[∇ϕi⊤C∇ϕj],K_{ij}=\mathbb E_\nu[\nabla\phi_i^\top C\nabla\phi_j],

followed by the generalized eigenproblem with the mass matrix. Our linear dictionary isolates averaged-field errors; cubic gradients introduce spatially varying weights. These are controlled diagnostic choices. We agree that their normal-gradient magnitudes have not been shown representative of probes from an existing application.

We can now quantify their effect on a spectral output. Using the saved matrices, we computed all 12 ordered eigenvalues and

\frac{|\lambda_j(\widehat H_C)-\lambda_j(H_P)|}{\lambda_j(H_P)}.$$ Here $\widehat H_C$ is the computed mass-whitened stiffness matrix and $H_P$ its analytic geometric reference. For oracle-centered rank-two fields, the means of these per-model maxima are: | Density | Probes | Sample $e_\lambda$ | Conditional $e_\lambda$ | Improving pairs | Improvements passing numerical screen | | --- | --- | ---: | ---: | ---: | ---: | | Uniform | Linear | 1.955% | 0.897% | 4/4 | 4/4 | | Uniform | Cubic | 1.864% | 0.897% | 4/4 | 4/4 | | Concentrated | Linear | 14.784% | 1.898% | 4/4 | 3/4 | | Concentrated | Cubic | 2.231% | 0.681% | 4/4 | 4/4 | All 16 paired evaluations improve, and 15 pass the descriptive numerical screen. Each model's eigenvalue sensitivity is the maximum relative change in its ordered eigenvalues when replacing grid 256 by grid 128 or shifted grid 256, using the same analytic reference eigenvalues. The screen compares the paired effect with ten times the sum of those sensitivities. It is not a statistical test or a rigorous integration bound. The 16 evaluations reuse two images and two initializations across densities and dictionaries; they are not 16 independent replications. Eigenspace outcomes are less uniform. We evaluated all lowest-mode prefixes, with sizes 2, 4, and 8 designated for primary reporting in this post-review analysis. For concentrated cubic probes, image 0 / initialization 1 has a four-mode projector error increase from 0.04326 to 0.06373, passing the same type of numerical screen, despite improved eigenvalues. For concentrated linear two-mode projectors, no paired improvement passes the screen. Uniform-density primary-prefix errors are already approximately $10^{-8}$–$10^{-7}$, with no resolved paired changes. Prefixes cutting numerically repeated eigenvalues are marked undefined. These results establish a measured spectral consequence in the fixed trial spaces and demonstrate that eigenvalue improvement need not imply eigenspace improvement. They do not establish how prevalent the mechanism is in realistic pipelines. Evaluating application-derived probes and their normal gradients remains necessary to settle that part of your question.
AuthorsOpenReview ↗

Reply to Reviewer ScsZ (2/3)

2. What explains the supervision reversals?

We analyzed all 160 matched pairs represented by the 320 saved model/dictionary/field comparisons. The primary subset contains all 16 oracle-centered rank-two pairs. Both models in each pair are evaluated on the same sample-worst probe and, separately, on the same conditional-worst probe.

Uniform-density reversals: loss of compensation. For image 1 / initialization 0, the following signed form contributions use the fixed sample-worst probe, normalized by ∣HP∣op\\|H_P\\|_{\mathrm{op}}. The tangent column includes the analytic-reference quadrature correction.

ProbesLabelsTangentCrossNormalSigned total
LinearSample+0.002541−0.005871+0.000049−0.003281
LinearConditional+0.000645−0.005979+0.000010−0.005324
CubicSample+0.002341−0.006088+0.000148−0.003599
CubicConditional+0.000639−0.005765+0.000013−0.005113

The negative cross contribution persists, while the positive tangent and normal terms that compensated it shrink. The tangent reduction contributes most to both reversals. The tangent and normal changes individually pass the tenfold sensitivity screen; the smaller cross changes do not. Both total additive-form reversals pass the screen. Changing the maximizing direction contributes only approximately 10−810^{-8} to each norm change, so the reversals are already visible on a common probe.

Concentrated cubic case: a different worst probe, with an unresolved net increase. For the same image and initialization, additive error changes from 0.00415870 to 0.00419785. On the sample-worst probe, signed error improves from −0.00415870 to −0.00207229. On the conditional-worst probe, it worsens from −0.00278961 to −0.00419785. Along that latter direction, the cross term becomes more negative (−0.00356588 to −0.00456955), and positive normal compensation decreases (0.00075270 to 0.00005096). However, the global norm increase, 0.00003915, is below the summed matrix sensitivity, 0.00010753. We therefore retain this as numerically unresolved rather than evidence of a robust harmful effect.

What condition does the evidence support? The operator norms of the compressed normal form contributions decrease in all 16 primary comparisons, but additive-form errors improve in 13. Thus normal-contribution reduction alone does not predict improvement: the signed balance with tangent and cross terms must be checked on the intended functions. In the two largest concentrated-linear improvements, reduced normal contribution dominates the change on the sample-worst probe; in the uniform reversals, loss of compensation dominates.

The selected output also matters. The uniform additive-form reversals improve their worst relative eigenvalue errors: 0.7945% to 0.5732% for linear probes and 0.7895% to 0.5727% for cubic probes. These observations explain the saved outcomes retrospectively; they do not provide a universal predictor of training outcomes or new measurements of local gradient amplification.

AuthorsOpenReview ↗

Reply to Reviewer ScsZ (3/3)

3. Scope of the supervision conclusion

The population identity and fixed-parameter raw-gradient variance reduction do not guarantee better finite-budget form accuracy under Adam, clipping, and EMA. With learned centers, conditional labels also retain bmbm⊤/tb_mb_m^\top/t, where bm=μt−mb_m=\mu_t-m is the difference between the posterior mean and learned center; removing label randomness does not remove centering contamination.

The added analyses use no retraining, new queries, or selected checkpoints. They explain the existing mixed results and measure a downstream consequence, while preserving the limitations of two base images, two covariance initializations, one noise level, known density, oracle-assisted labels, and supplied dimension. The original 16-image study is separate from this matched intervention. Our conclusion is that supervision can materially change the forms and spectral outputs in this controlled setting, with benefits depending on the error balance and the output being assessed.

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

This paper asks when learned posterior covariance fields give accurate weighted manifold forms. Through theoretical bounds, counterexamples, and controlled translation-orbit experiments, it shows how small local errors can still lead to large form errors.

Critical strengths
  1. I think the exact-posterior counterexamples are generally reasonable. They make the key message quite concrete, namely getting local geometry roughly right does not necessarily give accurate forms.

  2. The experimental work is ok. The matched comparisons, error decomposition, and numerical checks make the attribution to learned normal leakage convincing within what the authors claimed.

Critical weaknesses
  1. My main concern is the limited theoretical advance. Most of your arguments use standard tools, and the bounds look a bit conservative and I can not clearly see what are their practical implications. Can you provide some quantitative guidance on when recovery should work? It should make your contribution more significant. Put it short, I am not yet convinced how much additional theoretical insight this provides over existing work.

  2. The positive recovery results are also quite narrow and limited. The strongest result, as far as I understand, relies heavily on oracle centering, conditional posterior labels, and known-dimensional filtering. The paired study covers only two images, two initializations, and one noise level, while errors remain large with learned centers as shown in Table 1. Can you explain why it is the case?

Questions for authors

Your related work partially explains the difference in the context of existing posterior-geometry and closest-point methods. Could you make the theorem-level distinction more explicit from them? Particularly, for example, which finite-noise conclusion of Corollary 3 goes beyond those results, and under what assumptions?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer wxFd (1/3)

Thank you for recognizing the exact-posterior counterexamples and the controlled error attribution. Your questions prompted us to clarify the theorem-level contribution and to compute the actual spectral outputs from the saved forms.

1. What Corollary 3 adds at finite noise

The additional conclusion is that an accurate posterior covariance need not give an extension-independent energy guarantee at a fixed positive noise level. Corollary 3 uses a uniform prior on the unit circle, Gaussian observations Yt=Z+t,ϵY_t=Z+\sqrt t\\,\epsilon, and clean queries on the circle. The genuine normalized posterior covariance satisfies

At(x)=mtP(x)+ntxx⊤,nt>0.A_t(x)=m_tP(x)+n_txx^\top,\qquad n_t>0.

The smooth ambient functions ψa\psi_a all restrict to the same unit-norm intrinsic function, but

EAt(ψa,ψa)=mt+nt(a+1)2.\mathcal E_{A_t}(\psi_a,\psi_a)=m_t+n_t(a+1)^2.

At every fixed t>0t>0, these energies are unbounded as the extension varies, although the intrinsic function, mass, and reference energy remain fixed. In the two-function compression, the lowest mode becomes orthogonal to the reference lowest mode when nt(a+1)2>3mtn_t(a+1)^2>3m_t.

The distinction from local small-noise convergence is the quantification over extensions: ∣At−P∣F=O(t)\\|A_t-P\\|_F=O(t), and every fixed extension retains small-noise form convergence, but this gives no uniform guarantee over extensions at finite noise. A closest-point extension removes the normal derivative by construction; our result concerns specified ambient extensions that need not satisfy that condition. We will make this distinction explicit in the theorem discussion.

We agree that the transfer inequalities use standard tools. Their contribution is the identification of the additional quantities needed for a sufficient local-to-form guarantee, together with an exact-posterior obstruction and controlled evidence about which errors learning introduces. result. We will make these scopes more prominent.

AuthorsOpenReview ↗

Reply to Reviewer wxFd (2/3)

2. Quantitative recovery conditions and measured spectral outputs

With analytic mass, Appendix B gives the sufficient condition

\le \tau\\|H_P\\|_{\mathrm{op}}$$ for normalized additive form error at most $\tau$. Here $\eta_Q$ is the stiffness integration error; $e_T,e_X,e_N$ measure tangent, cross, and normal field errors; and the $\kappa$ factors contain moments of the corresponding whitened ambient gradients. Large normal derivatives require tighter normal error control under this budget. Rank filtering can leave cross contributions. If the normal derivative vanishes, the cross and normal terms disappear. We agree that these sufficient bounds are conservative: in the original concentrated-linear study, the mean block bound is approximately 8,680, versus measured error 56.38. We therefore use the bounds to state sufficient conditions, and direct form comparisons to establish the observed accuracy. Failure to satisfy the bound does not imply failed recovery. A certified application would also require a valid bound on $\eta_Q$, which grid sensitivity alone does not provide. To quantify the output-level implication, our additional analysis computes $$e_\lambda=\max_{1\le j\le12} \frac{|\lambda_j(\widehat H_C)-\lambda_j(H_P)|}{\lambda_j(H_P)}$$ from all 12 ordered eigenvalues. Here $\widehat H_C$ is the computed mass-whitened stiffness matrix and $H_P$ its analytic geometric reference. For the concentrated oracle-centered rank-two cohort: | Probes | Mean $e_\lambda$, sample | Mean $e_\lambda$, conditional | Improving pairs | Improvements passing numerical screen | | --- | ---: | ---: | ---: | ---: | | Linear | 14.784% | 1.898% | 4/4 | 3/4 | | Cubic | 2.231% | 0.681% | 4/4 | 4/4 | These are means of per-model worst relative eigenvalue errors, distinct from both the normalized additive form error and the relative-energy upper bound. The numerical screen requires a paired improvement to exceed ten times the sum of the two models' output sensitivities to grid 128 and shifted grid 256, relative to grid 256. It is descriptive, not a significance test or a rigorous quadrature bound. The calculations use existing models and queries. The output distinction matters empirically. Across both densities and dictionaries, the operator norm of the compressed normal form contribution decreases in all 16 oracle rank-two comparisons, while additive-form error increases in three. The two resolved uniform-density reversals result from losing positive tangent/normal compensation against a persistent negative cross contribution. Their relative eigenvalue errors nevertheless improve. Conversely, one eigenspace comparison deteriorates despite improved eigenvalues. Thus direct evaluation of the intended form or spectral output remains necessary for an empirical assessment; a smaller individual error component is insufficient.
AuthorsOpenReview ↗

Reply to Reviewer wxFd (3/3)

3. Why learned-center errors remain large, and what the positive result establishes

For a frozen learned center mm, both supervision schemes have the ideal target

Bm=At+bmbm⊤t,bm=μt−m.B_m=A_t+\frac{b_mb_m^\top}{t},\qquad b_m=\mu_t-m.

Conditional supervision removes conditional label randomness but preserves the centering contamination. This explains why the intervention alone cannot correct learned-center bias.

The existing results also distinguish direction from scale. Replacing the old learned mean with the improved mean, across the eight image–density–initialization cells with sample supervision, mean bias energy decreases from 35.05 to 15.54 and tangent-projector RMS error improves from 0.3704 to 0.1732, while restricted cometric RMS error worsens from 0.2323 to 0.3334. Own-target fitting remains poor, and signed fitting and centering contributions can cancel. Better means and tangent directions therefore do not ensure accurate scales or forms under this training budget.

The strongest positive result remains an assisted control: two base images, two covariance initializations, one noise level, supplied density, oracle-assisted labels, and known-dimensional filtering. Oracle means construct labels rather than extra network features. The new analysis supplies quantitative output evidence and explains individual outcomes; it does not expand the training cohort or remove its oracle assumptions. The original 16-image study provides broader failure evidence, not a 16-image replication of the matched positive result. We will make these scopes more prominent.

Thanks for the clarification and additional spectral results. Corollary 3 is clear to me now, and I also find the eigenvalue measurement a useful addition.

Although all the unclarity has bee addressed, my main concerns is still the scientific contribution. The recovery evidence still comes from the same small, oracle-based setting. To me, it is insufficient and I would not recommend this paper for acceptance. Nevertheless, I would encourage the author to extend both the theory and experiment in a more general direction.

Overall, I appreciate author's effort in revising the paper, responding to my concerns. I would like to keep my rating for now.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.