acceptodds
Under review as a conference paper at ICLR 2027

Optimization Geometry of Equivalent Brownian RKHS Representations

Abstract

Equivalent finite parameterizations can represent the same functions and intrinsic norm yet induce different optimization algorithms. We study this effect in a controlled finite Brownian RKHS with nodal, increment, and spectral coordinates. Classical finite-element, RKHS-interpolation, Brownian-covariance, and mixed-boundary DCT identities make the shared hypothesis class, Brownian energy, approximation operator, and coordinate maps explicit. Our main results concern the optimization geometry of this fixed model. With mapped initialization, identical scalar steps, and identical minibatches, nodal and spectral GD/SGD have exactly the same mapped trajectories. Increment GD is an explicit Euler step for the constant Brownian/Sobolev metric, with factor . For Brownian-regularized least squares, , independently of grid resolution for fixed , , and the stated normalization. Under the stated standard-Adam convention, the universal orthogonal equivariance group is exactly the signed permutations; the block DCT-VIII transform is not one. Float64 tests over five grids numerically verify the finite identities, mapped one-layer and recursive trajectories, conditioning predictions, and theorem-matched Adam separation. Thus, coordinate effects are isolated without changing the represented functions, intrinsic regularizer, or approximation space.

Primary area:
optimization
Keywords:
Brownian, RKHS, Optimization geometry, Coordinate representations, Stiffness matrix, Equivariance

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DbX6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies finite Brownian RKHSs relevant to the VBKL and BKL models, studying its properties, interpolation in such space, optimization paths, and equivalent coordinate systems. The article then checks some of the theoretical results empirically.

Critical strengths
  • The problem of optimization in RKHSs and equivalent description of optimization problem is an important task.
  • The article contains experimental results in order to back-up some of the theoretical claims.
Critical weaknesses

The quality of the writing is very low, making it difficult to properly assess the contribution of the article. The following list highlights some of the most critical weaknesses of the present article.

  • The paper motivation and challenges are not properly stated or mentioned at any point in the article. More importantly, the results introduced in this article are not properly motivated nor discussed in relation with previous work. This makes it extremely difficult to navigate the results and assess the quality of the contributions in the manuscript.
  • With the current wording, the work seems to be incremental; in particular, the paper studies a discretization of the RKHS of a Brownian kernel that is known (Corollary 2) to be interpolation between nodes. Interpolation in RKHSs have been widely studied in the literature (Bitzer and Steinwart, 2026).
  • Some of the notations and properties introduced in this article are not properly defined. More importantly, this issue is present even when listing the contributions of the paper: a) the nodal, increment, and block DCT-VIII coordinates are not properly defined as to what they are, nor motivated as to why one may be interested on them, b) κ2\kappa_2 is not introduced nor defined as to what it is, and is only defined through the very complex formula in (3), and c) the Adam optimizer plays an important role for the contribution and it is not introduced in the text nor the challenges overcame with Proposition 5.
  • The paper centers around the use of the Brownian RKHS for BKLs and VBKLs, yet, these models are not introduced, defined nor motivated.

References

Bitzer, M., Steinwart, I. Spectral Representations of Interpolation Spaces of Reproducing Kernel Hilbert Spaces. Constr Approx (2026). https://doi.org/10.1007/s00365-026-09781-4

Questions for authors

Q1. What is the motivation for the present article? Which problems does it tackle and why are those problems important to the ML community?

Q2. What are BKLs and VBKLs? Why are they important?

Q3. What is the relation to other results in interpolation in RKHSs? Why are these results different?

Q4. What are the challenges this article had to overcome in order to derive the theoretical results?

Q5. Why are equivalent description of the coordinate systems important?

3 comments
AuthorsOpenReview ↗

Response to DbX6 (1/2): motivation and BKL/VBKL

Thank you for recognizing the importance of optimization in equivalent RKHS descriptions and the role of the experiments. We agree that the motivating question and contribution boundaries should be apparent before the detailed formalism.

Q1: motivation. Learnable nonlinear profiles can be regularized through their derivative energy rather than through the Euclidean length of an arbitrary coefficient vector. Once a profile is implemented on an anchored grid, its nodal values, increments and spectral coefficients encode the same piecewise-linear function. A change in training behavior need not represent a more expressive model, a better approximation space or a different intrinsic regularizer.

Our question is which changes preserve the algorithm and which alter it despite preserving that represented problem. The metric controls mapped GD/SGD; coordinatewise Adam may additionally distinguish bases after the metric and Hessian spectrum are matched. Scaled spectral whitening is therefore an essential control, not an inferior version of spectral coordinates.

The practical consequences are exact regularization and step-size conversions, a controlled comparison of finite implementations, and the new grouped-replay and grid-node minimum-energy results. The latter were developed during the rebuttal. Compatible initialization is required for useful fixed-input compression; the joint-training extension has an evaluation-plus-radial gradient condition and can require all original coordinates. No predictive superiority or measured runtime speedup is claimed.

Q2: definitions and relevance. Brownian Kernel Ladders build successive integral RKHSs using $ k_{\ell+1}(x,x')=\int k_B(u(x),u(x'))\,d\mu_\ell(u),\qquad k_B(a,b)=\frac{|a|+|b|-|a-b|}{2}, $ with support functions drawn from the preceding representation. Variation Brownian Kernel Ladders first compose normalized Brownian profiles into paths, then form signed-measure mixtures of completed paths with total-variation complexity. They do not take a variation hull at every intermediate layer.

The shared building block is a learnable profile satisfying f(0)=0f(0)=0 and ∣f∣HB2=∫∣f′(t)∣2,dt\\|f\\|_{\mathcal H_B}^2=\int |f'(t)|^2\\,dt. Its piecewise-linear realization has exact energy $ \sum_j\frac{(f(t_{j+1})-f(t_j))^2}{t_{j+1}-t_j}. $ This provides a directly interpretable functional regularizer with a finite computational implementation. Our paper asks how these finite components should be represented during optimization; the coordinate results do not require adopting either entire ladder framework. We will introduce the models and this connection earlier, while establishing this paper's contribution through its own optimizer statements rather than through companion-paper depth or statistical guarantees.

AuthorsOpenReview ↗

Response to DbX6 (2/2): interpolation, technical scope and notation

Q3: interpolation. We agree that nodal RKHS interpolation and minimum-energy interpolation are established tools, not new discoveries here. Bitzer and Steinwart study real interpolation spaces [L2(ν),H]θ,r[L^2(\nu),H]_{\theta,r} and their spectral representations. Our nodal interpolation matches function values on a finite grid. These are related RKHS subjects but different uses of "interpolation"; we will distinguish them and credit the literature explicitly.

In the anchored Brownian setting, nodal interpolation is the orthogonal projection onto the span of kernel sections at non-anchor grid nodes. This establishes the common represented space and norm. It does not determine an optimizer trajectory. A general linear parameterization changes the Euclidean metric, and standard Adam can distinguish orthogonal bases within that metric. Our additional grid-node result concerns an entire compatible Adam trajectory, including minibatching, not merely the existence of a minimum-energy minimizer. It is not asserted for arbitrary off-grid data.

Q4: technical steps. The original analysis consistently aligns the anchor, grid ordering, inverse maps and energy scaling. In particular, D0⊤D0=hK0D_0^\top D_0=hK_0 fixes both the increment preconditioner and the penalty conversion. We do not portray this as a new general reparameterization principle, nor the conditioning argument as a mechanism unique to Brownian increments.

The additional Brownian-specific steps are the 2rx2r_x bound on identical evaluation-column groups, exact transport of both Adam moments with the finite-epsilon normalization, and historical refinement when jointly trained internal inputs move. Preserving current fitted values alone does not preserve upstream derivatives; the joint proof therefore preserves the whole profile. Its useful count requires evaluation-plus-radial gradients. The first-step criterion separately optimizes each basis's scalar step, removing that particular tuning confound, but gives no universal or persistent ranking. Complete proofs and the boundary cases are retained in Response Supplement S1-S2.

Q5 and notation. Equivalent descriptions are the control that separates a change in the learning algorithm from a change in the model. Nodal parameters are grid values; increment parameters are neighboring differences; spectral parameters are orthogonal-basis coefficients. Diagonalizing the energy is distinct from whitening it.

We will define κ2(H)=λmax⁡(H)/λmin⁡(H)\kappa_2(H)=\lambda_{\max}(H)/\lambda_{\min}(H) at first use for an SPD Hessian and introduce Adam through its coordinatewise first and second gradient moments. Nodal and unwhitened-spectral GD/SGD agree after mapping; increments and scaled-whitened spectral GD/SGD also agree. Adam need not agree within these orthogonal pairs. This is why equal function spaces and equal conditioning are insufficient to identify the adaptive algorithm.

All statements retain their original initialization, objective, penalty and update conditions. Singular empirical covariance is allowed where stated; no empirical inverse is introduced. Extra normalization, projection and nonsmooth derivative conventions must be matched. The actual Salinas forward normalization, the original E2 penalized updates and the new Adam diagnostics are kept separate. We will use this distinction to organize the revised exposition more clearly.

I acknowledge the authors' response. I believe the paper is still very unclear in all aspects and highly question the relevance of all of its results.

Hence, I will maintain the score.

Reviewer_7qVgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

It is almost impossible to summarize the paper... it is absolutely nonsense : a kind of abstraction without any purpose and any explanations

Critical strengths

No strength here

Critical weaknesses

It is impossible to understand what is going on in this paper.

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Response to 7qVg: purpose and clarification

We are sorry that the purpose and organization were not clear. The question is concrete: can implementations of the same learnable function, with the same intrinsic regularizer, behave differently solely because they store the function in different coordinates?

Nodal coordinates store grid values, increments store neighboring differences, and spectral coordinates store coefficients in an orthogonal eigenbasis. The finite Brownian model makes all three descriptions and their inverse maps explicit, so the represented function space, energy and approximation space stay fixed.

The submitted optimizer results identify which mapped GD/SGD trajectories agree, the precise Brownian preconditioner induced by increments, the regularized least-squares conditioning bound, and the universal orthogonal symmetries of the specified Adam update. We agree that the conditioning benefit is shared by spectral whitening; it is not unique to increments, and the underlying interpolation and reparameterization principles are classical.

The additional rebuttal analysis makes the remaining adaptive-basis question more concrete. For fixed observations and compatible initialization, increment Adam admits exact grouped replay. On grid-node observations, each iterate is the minimum-energy function with its current fitted values. These results have explicit assumptions and do not establish universal optimizer superiority, a general nonlinear convergence rate, or measured runtime improvements. They are labeled as new additions, not as part of the submitted manuscript.

The accompanying Response Supplement provides the full arguments and actual trajectory plots, separating original experiments from new controlled studies. We will put the concrete optimization question and the meanings of the representations before the detailed formalism in the manuscript revision. A specific mathematical statement or experiment of concern would help us address the technical assessment directly.

Reviewer_xbMwOpenReview ↗
Rating
2(−1)
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how coordinate choices affect optimization within a fixed finite-dimensional Brownian reproducing kernel Hilbert space (RKHS). It compares nodal, increment, and spectral parameterizations that preserve the same function space, intrinsic Brownian norm, and approximation properties, thereby isolating the coordinate dependence of optimization.

The authors establish four main results. First, the orthogonal relationship between nodal and spectral coordinates implies identical GD/SGD trajectories when mapped to a common representation, provided that initialization, learning rates, and minibatches are matched. Second, Euclidean gradient descent in increment coordinates is equivalent to Brownian/Sobolev preconditioning by h−1K0−1h^{-1}K_0^{-1} in nodal coordinates. Third, for Brownian-regularized least squares, the Hessian in increment coordinates admits a grid-independent condition-number bound when the interval radius AA, regularization parameter ρ>0\rho>0, and objective normalization are fixed. Finally, for the specified formulation of standard Adam, the authors characterize signed permutations as its maximal universal orthogonal equivariance group. General DCT-VIII transformations fall outside this group, implying the existence of objectives for which nodal and spectral parameterizations produce different first-step updates.

Critical strengths
  1. The paper holds the function class and regularization fixed while matching initial functions, learning-rate schedules, and minibatches across parameterizations. This design isolates the effects of coordinate choice on optimization and provides a useful methodological framework.

  2. Appendices C.6–C.8 clearly present the chain-rule derivations and blockwise extensions, explaining why cross-layer coupling in recursive architectures preserves the identities induced by fixed linear reparameterizations. The distinction between discrete Euler updates and continuous gradient flow is also clearly articulated.

  3. The paper carefully delineates the scope of its results. It does not extrapolate the least-squares condition-number bound to general nonlinear convergence guarantees or interpret Adam’s trajectory differences as evidence that one coordinate system is universally superior. Proposition 5 explicitly specifies the relevant conditions, including zero-initialized moment estimates, bias correction, ε>0\varepsilon>0, and the absence of clipping.

Critical weaknesses

W1. Novelty relative to classical results needs clarification. The paper appropriately acknowledges its foundations in finite elements, kernel interpolation, Brownian covariance, and DCT theory. However, it remains unclear which insights from the unified framework go beyond direct consequences of established results. Proposition 2, the GD/SGD equivalence between nodal and spectral coordinates, and the blockwise extensions to recursive architectures follow from standard linear reparameterization identities. The proof of Proposition 4 is clear, but its main ingredients—whitening the regularizer and bounding the kernel diagonal—apply more generally.

In particular, defining z=Asv~z=A_s\widetilde v, where As=h,Λ1/2Q⊤A_s=\sqrt h,\Lambda^{1/2}Q^\top, gives $ A_s^\top A_s=hK_0=D_0^\top D_0. $ Thus, whitened spectral coordinates and increment coordinates yield identical mapped GD/SGD trajectories under corresponding initialization, identical learning rates, and shared minibatches. The improved conditioning can therefore be understood as a consequence of whitening the Brownian regularizer, rather than a property unique to the increment representation. The authors should also clarify what the necessity direction in Proposition 5’s characterization of the maximal universal orthogonal equivariance group adds to the cited work on Adam’s rotation dependence, including VectorAdam and Zhang et al. This concern does not establish that Proposition 5 duplicates prior results, but its precise theoretical contribution needs clearer positioning. A contribution table distinguishing classical results, Brownian-specific specializations, and new conclusions, together with a comparison of the original and whitened spectral coordinates, would help.

W2. The experiments provide limited insight into the optimization consequences.
Figure 1 and Tables 1–2 carefully verify the theoretical predictions, but most results primarily reproduce established identities numerically. Adam’s first-step discrepancy demonstrates the absence of universal orthogonal equivariance, yet does not establish whether the discrepancy persists, grows, or vanishes over subsequent iterations; whether it corresponds to differences in objective decrease; or under which objectives and initializations it becomes small. Since Appendix G.1 states that E1C already runs for 200 steps, the existing outputs could be used to report full objective trajectories, function-space discrepancies, and trajectory discrepancies normalized by update magnitude. The interpretation of absolute ℓ2\ell_2 discrepancies should also account for possible effects of increasing dimension.

Furthermore, each eigenvalue has multiplicity two, and Remark 1 acknowledges that the eigenbasis is not unique. A focused ablation varying the orthonormal basis within the corresponding two-dimensional eigenspaces would help assess how strongly Adam’s finite-step behavior depends on the selected blockwise DCT basis. These additions could provide more concrete optimization insights within the existing controlled setting, without requiring large benchmarks or state-of-the-art comparisons.

Questions for authors

Q1. Which results go beyond direct consequences of standard reparameterization identities, finite-element preconditioning, and whitening analyses for bounded kernels? In particular, what does the necessity direction in Proposition 5 add to the cited work on Adam’s basis dependence?

Q2. Could the authors report objective values and function-space discrepancies over the full 200-step trajectories, together with appropriately normalized discrepancies between the mapped trajectories? Are the findings robust to changes of basis within the two-dimensional eigenspaces associated with repeated eigenvalues, while preserving the function space and Brownian energy?

Q3. When varying GG, is the intrinsic regularization coefficient ρ\rho or the increment-penalty coefficient used in the implementation held fixed? Please specify the conversion between these coefficients, the learning rates used in Table 2, and how those learning rates scale across grids.

Q4. Is Brownian energy normalization implemented as a forward reparameterization or a post-update projection? Does E2 include the same operation? If not, how should the scope of the theoretical results be qualified when interpreting the actual training procedure?

7 comments
AuthorsOpenReview ↗

Response to xbMw (1/6): whitening and contribution

Thank you for the careful review and for identifying the importance of holding the represented model fixed. We agree with the whitening point: the conditioning gain is not unique to increments.

With nodal parameters vv, let w=D0vw=D_0v, K0=QΛQ⊤K_0=Q\Lambda Q^\top, and $ s=A_sv,\qquad A_s=\sqrt h\,\Lambda^{1/2}Q^\top,\qquad U=A_sD_0^{-1}. $ Because As⊤As=D0⊤D0=hK0A_s^\top A_s=D_0^\top D_0=hK_0, UU is orthogonal and s=Uws=Uw. Both maps recover the identical original profile, and $ v^\top K_0v=h^{-1}\|w\|^2=h^{-1}\|s\|^2. $ Increments and scaled spectral whitening therefore have identical mapped GD/SGD trajectories under corresponding initialization, shared scalar steps and minibatches, and orthogonally similar Hessians. This also holds blockwise with jointly trained coupled parameters. Any extra normalization, projection or nonsmooth derivative convention must be transported consistently. Unscaled whitening z=Λ1/2Q⊤vz=\Lambda^{1/2}Q^\top v instead requires ηz=ηw/h\eta_z=\eta_w/h to match increment descent.

We distinguish the contributions as follows.

ResultStatus
Finite Brownian space, interpolation, stiffness energy and DCT basisClassical foundations in a common anchored normalization
Mapped GD/SGD and the conditioning boundReparameterization identities and a Brownian specialization of whitening
Proposition 5A necessary-and-sufficient universal orthogonal symmetry statement for the specified Adam convention
Exact grouped replay and grid-node minimum energyAdditional Brownian-specific results developed during the rebuttal
Historical replay and the first-step criterionRebuttal-time extensions with explicit scope conditions

We have checked Bernacchia, Global curvature for second-order optimization of neural networks (ICML 2025), Appendix F, pp. 19-20. It already establishes signed-permutation equivariance of Adam in its invariant-loss setting. VectorAdam (Ling et al., 2022), Zhang et al. (2025, Appendix A.2), and Lau and Su (2026, Appendix C) also discuss rotation dependence or signed permutations. We do not claim discovery of that symmetry, nor make it the main novelty claim.

Our precise statement includes the universal converse for the specified standard Adam: zero initial moments, bias correction, scalar beta parameters, positive epsilon, and no clipping or weight decay. At the first step, Fϵ(g)i=gi/(∣gi∣+ϵ)F_\epsilon(g)_i=g_i/(|g_i|+\epsilon). Universal orthogonal equivariance, applied to u=Veju=Ve_j and g=τug=\tau u, forces $ \frac{\tau u_i}{\tau+\epsilon}=\frac{\tau u_i}{\tau|u_i|+\epsilon}. $ Every nonzero uiu_i consequently has magnitude one; orthogonality makes VV a signed permutation. The moment recurrence proves sufficiency. This specifies the theorem's content; it is not an exhaustive historical-priority claim.

The stronger additional contribution is what Adam can still change after the metric has been matched: exact grouped replay, the grid-node minimum-energy property, and quantitative basis effects. Part 2 states these results; the accompanying Response Supplement contains their complete proofs and scope qualifications.

AuthorsOpenReview ↗

Response to xbMw (2/6): additional optimizer results

These results were developed during the rebuttal, not claimed as already present in the submitted manuscript.

For the original fixed-input regression objective, write $ L(w)=\frac{\|Xw-y\|^2}{2n}+\frac{\rho}{2h}\|w\|^2,\qquad X_{ij}=b_j(x_i),\quad \rho>0, $ where bj(x)=h−1∫0x1(tj,tj+1)(t),dtb_j(x)=h^{-1}\int_0^x\mathbf1_{(t_j,t_{j+1})}(t)\\,dt. Group identical nonzero columns of XX. If rxr_x counts distinct non-anchor sites, the number of groups satisfies q≤min⁡G,2rxq\le\min\\{G,2r_x\\}.

Exact replay. Zero-start standard increment Adam keeps its parameters and both moments equal within each group, for arbitrary minibatches from this fixed sample. For a group of size dad_a with common parameter θa\theta_a, exact replay divides the reduced-objective derivative by dad_a before the original Adam recurrence. Naively applying scalar-step Adam to the summed reduced gradient would change the algorithm. With orthonormal aggregate coordinates instead, the exact conversions are ηt,a=ηtda\eta_{t,a}=\eta_t\sqrt{d_a} and ϵa=ϵda\epsilon_a=\epsilon\sqrt{d_a}.

Original-objective recovery. Let TT have normalized group-indicator columns and P=TT⊤P=TT^\top. Every original ww uniquely decomposes as w=Tz+uw=Tz+u, u⊥range⁡(T)u\perp\operatorname{range}(T), and $ L(Tz+u)=\frac{\|XTz-y\|^2}{2n}+\frac{\rho}{2h}\left(\|z\|^2+\|u\|^2\right). $ Both directions recover the original nodal profile through D0−1wD_0^{-1}w. In particular, $ XPw=Xw,\qquad L(w)-L(Pw)=\frac{\rho}{2h}\|(I-P)w\|^2. $ For grid-node observations, q=rxq=r_x and range⁡(T)=range⁡(X⊤)\operatorname{range}(T)=\operatorname{range}(X^\top). Every compatible increment-Adam iterate is the minimum-energy function with its current fitted values. Off-grid inputs only give the stated group-space invariance, not full representer-space invariance. Duplicated sites, singular designs and anchor-only data are included; anchor-only zero-start training has q=0q=0. Compatible nonzero initialization in the group space with zero moments is also allowed. This does not imply lower total loss than another trajectory.

Basis choice. For repeated observations at one non-anchor site, let b=b(x0)b=b(x_0), R=∣b∣2R=\\|b\\|^2, λ=ρ/h\lambda=\rho/h, and let μ≠0\mu\ne0 be the mean label. In whitening s=Vws=Vw, put c=Vbc=Vb, dj=cj/(∣cj∣+ϵ/∣μ∣)d_j=c_j/(|c_j|+\epsilon/|\mu|) and χ=(c⊤d)2/(R∣d∣2)\chi=(c^\top d)^2/(R\\|d\\|^2). The best scalar first step satisfies $ \frac{\min_{\eta\ge0}L(w_1^V(\eta))-L_{\star}}{L(0)-L_{\star}}=\frac{\lambda(1-\chi)}{\lambda+R\chi}. $ This is a first-step statement on the declared family, not a universal basis ranking. The supplement retains the zero-gap cases, the opposite ordering and a two-step reversal.

Joint training. Historical refinement exactly transports full profiles and both moments while all layers train. The useful site-count bound requires the actual block gradient to be an evaluation term plus a radial term, gb,t=Bb,t⊤ab,t+γb,twb,tg_{b,t}=B_{b,t}^\top a_{b,t}+\gamma_{b,t}w_{b,t}. Arbitrary nonradial coupling is not silently excluded: actual-gradient refinement still replays it, but may require all coordinates. Group counts may grow to GbG_b; they are not measured runtime or total-memory savings.

General tied-state invariance and convex grouping have precedents in Zhu and Chen (2026), Chen et al. (2024 preprint version), and Joudaki et al. (2026). Our added claims concern Brownian column counts, finite-epsilon state transport, historical refinement and the minimum-energy consequence. Response Supplement S1-S2 supplies the full statements and proofs.

AuthorsOpenReview ↗

Response to xbMw (3/6): full trajectories and basis ablation

We agree that a first-step discrepancy and a final-value table do not answer the full-trajectory question. The accompanying Response Supplement now contains the actual curves at all five grids, with medians and minimum-to-maximum bands over five seeds. The bands describe observed seed variation, not confidence intervals. Original and new experiments are explicitly separated.

Original E1C (Supplement S3). We re-executed the supplied E1C source and original numerical configuration: n=256n=256, ρ=0.1\rho=0.1, 200 full-batch steps, learning rate 0.0030.003, and matched initial functions of Brownian energy 0.250.25. These are nodal, unwhitened-spectral and increment arms. All 5,025 saved trajectory rows and 30 original numeric columns were reproduced exactly as parsed from the archived CSV in this execution. We added read-only logging of intrinsic function and relative update discrepancies; no objective, initialization, gradient or update was changed.

At G=128G=128, median final objectives are 0.06082520.0608252 (nodal), 0.13433310.1343331 (unwhitened spectral), and 0.04908140.0490814 (increments). The median nodal--spectral Brownian-norm discrepancy is 0.45900.4590 at step 200. This documents persistent finite-budget separation in the original protocol; it is not a tuned superiority comparison.

New matched-whitening ablation (Supplement S4). This is a separate n=128n=128 experiment with A=2A=2, ρ=0.1\rho=0.1, zero initial function and moments, and 200 full-batch Adam steps. The common schedule is η=0.01h\eta=0.01\sqrt h, with (β1,β2,ϵ)=(0.9,0.999,10−8)(\beta_1,\beta_2,\epsilon)=(0.9,0.999,10^{-8}). We compare increments, scaled spectral whitening and rotations by π/8\pi/8, π/4\pi/4 or seeded random angles within the repeated two-dimensional eigenspaces. The represented objective and whitening metric are held fixed within each mesh.

GIncrementSpectralpi/8pi/4Random
85.06e-51.75e-81.79e-62.62e-54.83e-6
169.55e-76.89e-91.20e-62.13e-53.45e-6
321.43e-96.06e-91.31e-62.22e-53.41e-6
645.42e-105.65e-91.46e-62.35e-53.53e-6
1283.71e-105.38e-91.58e-62.44e-53.64e-6

Entries are median final relative gaps (L200−L⋆)/(L0−L⋆)(L_{200}-L_{\star})/(L_0-L_{\star}); denominators are positive. At G=128G=128, median maximum increment--spectral function separation is 0.8730.873, declining to 8.32×10−58.32\times10^{-5} at step 200. At absolute objective tolerance 10−810^{-8}, spectral whitening is better in 10 settings and the remaining 15 are indistinguishable. We do not call sub-tolerance differences increment wins. Matched GD agrees to approximately 1.16×10−131.16\times10^{-13} in Brownian norm.

For both studies the plots show the original objective, dk=∣fk(a)−fk(b)∣HBd_k=\\|f_k^{(a)}-f_k^{(b)}\\|_{\mathcal H_B}, and $ r_k=\frac{\|\Delta f_k^{(a)}-\Delta f_k^{(b)}\|_{\mathcal H_B}}{\max\{\|\Delta f_k^{(a)}\|_{\mathcal H_B},\|\Delta f_k^{(b)}\|_{\mathcal H_B}\}},\qquad 0/0:=0. $ Absolute and relative discrepancies are displayed together so that nearly vanishing updates are not overinterpreted. The new whitening ablation does not replace or get relabeled as original E1C. Its fixed epsilon and sqrt(h) step schedule are not claimed to define exactly the same energy-normalized optimizer across different meshes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.