acceptodds
Under review as a conference paper at ICLR 2027

Continuous Equivariant Learning with Linearly Many Frames

Abstract

Weighted frame averaging makes an ordinary network equivariant by evaluating it in several input-dependent coordinate systems. How many frames are needed to preserve continuity? For labeled vectors spanning , with fixed, the minimum worst-case support of a continuous - or -equivariant weighted frame grows linearly in . The lower bound follows from characteristic classes of Grassmannians. An explicit family of probes, adapted from Wronskian rank condensers, gives the upper bound. Determinant weights of order preserve regularity across probe failures, and this guarantee is sharp. For planar rotations the minimum support is exactly . In synthetic force-learning experiments, disagreement between transported predictions produces jumps under hard frame selection, while continuous averaging removes them. At particles and a common update budget, trained Wronskian hard selection reduces force error by relative to a Gram-matrix model; averaging over probes reduces the hard model's error by a further , using times its recorded worker time.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
equivariant learning, weighted frame averaging, canonicalization, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_fmakOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the minimum worst-case support required for continuous equivariant weighted frame averaging, establishing Θ(n)\Theta(n) bounds for labeled, full-rank inputs in a fixed spatial dimension and providing an explicit construction.

Critical strengths

The paper asks a very natural question, how many frames do we actually need for continuous equivariant frame averaging? This is an important problem in this area. I really appreciate the effort on the theory, the analysis looks solid and detailed. I also like the moving-boundary experiment. They keep the prediction branches fixed and only move the selection boundary, so we can see more clearly where the discontinuity comes from, rather than just comparing overall prediction errors.

Critical weaknesses

(i) The introduction gives very little background, I feel it assumes the reader already knows this subfield quite well. It jumps almost straight from frame averaging to probability measures and topological arguments. For someone without that background, it's quite hard to even get into the paper.

(ii) Correct me if I'm wrong, Appendix A.3 gets a linear lower bound from Dold’s theorem, and Appendix A.5 says the Wronskian coverage principle is already known. So I'm curious, how much of the Θ(n)\Theta(n) result can we get by putting these existing results together? The sharper bounds and regularity analysis may bring something new, but I'd like the authors to explain this a bit more in the rebuttal.

(iii) The writing is not very easy to follow. There are a lot of definitions, formulas and symbols, one after another, but not much explanation of what they mean or what they tell us in practice. Some geometric intuition and a simple illustration of the construction would help, also a bit more explanation of how to interpret the main results.

(iv) Finally, I'm still not quite sure what this result changes for geometric deep learning in practice. The construction needs labeled, full-rank inputs, and it doesn't guarantee permutation equivariance. Also, having linearly many frames doesn't mean the runtime is linear. I'd like to see a few concrete examples here, which applications would benefit from this, and when would using fewer frames actually make the method cheaper to run?

Questions for authors

See weakness

1 comment

Thank you for your helpful review. We have revised the manuscript accordingly.

(ii) What is new. Both the lower and the upper bounds are new. The known tools you identify are ingredients that had not been used in this context before. The contribution is connecting them to weighted frames.

Linear lower bounds. Dold's theorem gives a bound of ~n/(d(d − 1)/2 + 1). Building on Knudsen–Weinberger, Theorem 3 identifies the minimum number of frames with a sectional category, and characteristic classes of Grassmannians give constant 1/2, and 1 for even d. These bounds answer the question of Dym et al. (ICML 2024) and apply to their frames (Corollary 5).

A linear upper bound, explicitly. Wronskians can be adapted to give d(n − d) + 1 fixed linear probes (Theorem 6). Theorem 2 weights each probe's frame by det(AP_j)^{2k}, giving C^{2k−1} predictions, which is sharp. For fixed d ≥ 3 this improves order n^{d−1} of Dym et al. to linear: 61 frames at 24 particles against 1,012 for Dym et al.

Planar case: exactly n − 1 (Corollary 4).

(i) and (iii) Background and intuition. Thanks for these comments. We have entirely rewritten the abstract and introduction. The introduction now proceeds in steps with minimal notation, defines each term at first use, and includes a planar example of a jump and Figure 1. We believe this is much more accessible to a broader audience.

(iv) Practice.

  • Scope: full-rank inputs are the generic case and are fully covered. The lower bounds hold on perfectly conditioned inputs and for permutation-invariant frames. Appendix A.10 treats fixed lower rank, and Section 8 lists the open extensions.

  • Runtime: scoring and forming frames costs O(mnd^2 + md^3). At 24 particles, full averaging over our 61 frames had 4.1 and 10.3 times lower median latency than the 1,012 frames of Dym et al. at batch sizes 1 and 32 (Section 6, Table 13). Fewer frames matter most when network evaluations dominate, for example with large networks or limited memory.

  • Examples: force models for a fixed molecule (Appendix B.11), motion-capture markers, labeled robot keypoints.

  • Experiments: a network trained with one frame per input and then averaged over our frames without retraining has 48.2% lower force NRMSE than a frame-free Gram baseline at comparable training time (24 particles, Section 7.4). The moving-boundary experiment confirms that single-frame predictions jump at frame switches and averaging removes the jumps.

Thank you again for your feedback.

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

Any non-equivariant neural network (under several assumptions) can be made equivariant using weighted frame averaging in several input-dependent coordinate systems. The paper theoretically bounds the required number of frames for n labeled vectors spanning R^d to preserve continuity. The empirical results on the synthetic force-learning task show that hard frames selection leads to jumps, while continuous averaging removes them.

Critical strengths

I appreciate the problem tackled. The paper gives both a theoretical answer to the question of how many coordinate systems are needed to keep an equivariant model continuous and a way to construct them.

Critical weaknesses

In my view, the presentation of the content is too dense in the main part. There is a lot of notation, and some terms are used without clear definitions (I guess because there is no space for precise definitions).

The empirical evaluation seems too weak compared with the theoretical part. The experiments are synthetic, the performance advantage appears mainly for large numbers of particles, and continuous averaging is more expensive. So, the theoretical result is interesting, but the empirical section does not demonstrate that the significance ‘translates’ into practice.

Questions for authors

My doubts concern mainly the significance of the results and the presentation of the content. The latter, I believe, might be easily corrected. Regarding the former, I have the following questions. Which applications (practical or theoretical) are the main ones for the work? I mean, what becomes possible because the number of frames grows linearly rather than quadratically (or worse)? Which applications do satisfy full-rank assumptions?

2 comments

Thank you for a constructive review. The revision addresses both of your concerns.

Presentation. The introduction now explains the problem with minimal notation. It states the result in words before any formal definition. The setup and Theorem 1 form a new Section 2, and termnology in the main text is explained where it first appears.

Empirical significance. Section 7 compares our construction throughout with a Gram-matrix model, a frame-free baseline that predicts from pairwise inner products of positions. Error is NRMSE, the RMS force error divided by the RMS target force.

A clear gain over the baseline.

  • At 24 particles, with training time comparable to the baseline's, our continuous compact-blending predictor has 48.2% lower NRMSE (Section 7.4).
  • We obtain it by training with one frame per input, the most compute-efficient of the tested training rules, and averaging afterwards without retraining.
  • Full blending over all 61 frames is continuous by Theorem 2 and has 6.3% lower NRMSE than single-frame evaluation of the same networks. It is our most accurate 24-particle predictor.

An advantage that grows with the number of particles. At an equal number of training updates, averaging over our frames has 14.5%, 36.1% and 54.5% lower NRMSE than the Gram baseline at 12, 16 and 24 particles (Section 7.2).

Cost. Training with hard selection evaluates the network once per input. Compact blending takes about as long per input as single-frame evaluation at batch size 1, and 11.3 ms against 4.9 ms per batch of 32.

The theory's mechanism appears in trained models. Single-frame predictions jump where the frame switches, and averaging removes the jumps (Section 7.1). Jumps remain at the final checkpoints of the trained 24-particle networks. The median jump is about half their median error at the same inputs.

Why a synthetic task. It gives exact targets and controlled frame switches. Appendix B.11 works through a molecular-dynamics application. Together with the new lower and upper bounds, these results give practitioners both the continuity guarantee and a cost-effective way to obtain it.

What does linear rather than quadratic growth make possible?

An answer to an open question. The linear lower bounds answer a question posed in ICML 2024 by Dym et al., who asked for stronger lower bounds on the number of frames for rotations.

A known minimum. At 24 particles in 3D, at least 11 frames are needed at some input. Frames from the principal axes, averaged over the axes' signs, use at most 8. They therefore cannot form a continuous weighted frame on all such inputs.

A vastly smaller construction. On labeled inputs whose vectors span the space, and for fixed dimension d ≥ 3, our construction grows linearly in n, against order n^{d−1} for the construction of Dym et al. It needs 61 network evaluations at 24 particles and 289 at 100. Their O(3) construction needs up to 1,012 and 19,404, respectively.

Main applications. In practice: wrapping an existing network so that it becomes exactly and continuously rotation-equivariant, where predictions are used along trajectories or through derivatives. In theory: the growth rate of the minimum number of frames, with lower and upper bounds that are both new, and the exact answer n − 1 for planar rotations.

Which applications satisfy full rank? Full rank means that the input vectors span the space. It is the generic case. Data of a fixed lower rank, such as planar molecules, are treated for O(d) in new Appendix A.10.

Thanks again for your helpful review.

I thank the authors for their detailed response. I apologize for the delayed reply, as I was ill over the past few days; however, I kept in mind that I needed to respond. I thank the authors for the corrections made and would like to raise my score to 3.

Reviewer_5fAkOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses the question of how many frames are needed within the framework of weighted frames [Dym et al.] for the groups SO(d) and O(d) acting on n labeled vectors that span Rd\mathbb{R}^d. The answer is linear in n. The lower bound 1+⌊(n−d)/2⌋1+\lfloor (n-d)/2 \rfloor (n−d+1n-d+1 for even d) comes from nonvanishing powers of characteristic classes of the oriented Grassmannian, combined with a rational cup length theorem from a recent preprint (Daundkar and García-Calcines, 2026). For d = 2, 3 there is also a self contained proof (Thm 10). The upper bound d(n−d)+1d(n-d)+1 uses Wronskian probes. The Wronskian of d independent polynomials of degree less than n has degree at most d(n-d), so one of d(n-d)+1 nodes always gives a valid frame. With weights det⁡(APj)2k\det(AP_j)^{2k} the averaged model is C2k−1C^{2k-1} and in general not C2kC^{2k}. For SO(2) the answer is exactly n-1.

Then, the authors empirically demonstrate on synthetic force fields for labeled particles in R3\mathbb{R}^3 that hard frame selection jumps when the transported predictions disagree, and that averaging removes the jump. At N = 24, averaging over 61 probes has 21.4% lower error than Wronskian hard selection and 54.5% lower error than a Gram matrix model, at about 6 times the worker time.

Critical strengths
  1. The question is natural and the answer looks elegant. Dym etal. showed that continuous canonicalization is impossible and built weighted frames with n(n−1)⋯(n−d+2)n(n-1)\cdots(n-d+2) atoms (or non-trivial green functions) for SO(d). This paper shows that on full rank inputs we have that linearly many atoms are necessary and sufficient, with the exact count n−1n-1 for SO(2). The construction of Dym et al. also handles permutations and rank-deficient inputs (lines 157 to158), so the two counts are not directly comparable, but on full rank inputs the gap is real. The upper bound is elementary and explicit (Lemma 12, Thm 6), so it can actually be used.

  2. Thm 2 is correct and sharp. The proof in A.6.2 (lines 1132-1147) follows from Lemma 14, the bound ∣det⁡(AP)∣≤∥AP∥d−1σmin⁡(AP)|\det(AP)| \le \|AP\|^{d-1}\sigma_{\min}(AP) and the Leibniz rule. The example in lines 1149-1161 gives F(A(t))=t2ksign⁡(t)/(1+t2k)F(A(t)) = t^{2k}\operatorname{sign}(t)/(1+t^{2k}), which is C2k−1C^{2k-1} and not C2kC^{2k}. So the weight exponent buys exactly the stated regularity.

Critical weaknesses
  1. The paper is very hard to follow. It is written way too densely and mathematical notation is given prematurely in the intro (lines 41-48, "Let Md,nfr={A∈Rd×n:rank⁡A=d}\mathcal{M}^{\mathrm{fr}}_{d,n} = \{A \in \mathbb{R}^{d\times n} : \operatorname{rank} A = d\}, whose columns are n vectors spanning Rd\mathbb{R}^d ..."). Symbols are used before they are defined. The gjg_j in eq. (2) (lines 188-190) are only defined at lines 211-213, as polar factors of APjAP_j. The lower bound in eq. (8) rests on Thm 4.21 of a 2026 preprint that is never stated, and its hypotheses are checked in one sentence (line 289). The self contained Thm 10 recovers the same bound only for d = 2, 3, and gives a weaker linear bound for d ≥ 4. So the constants of Thm 1 for d ≥ 4 cannot be verified from the paper.

  2. The scope and the experiments limit the impact. Frames that are SO(3) and O(3) equivariant are important, yet the general case includes permutations.The construction only works for labeled points (lines 116-120), and all experiments use labeled particles. The continuity guarantee also holds only on full rank inputs. At a planar point cloud in R3\mathbb{R}^3 every probe determinant vanishes, and the averaged model has direction dependent limits there, although SO(3) still acts freely at rank 2 (lines 119-122). So the model cannot be extended continuously to planar configurations, which are common for molecules, and it varies arbitrarily fast near them. The Lipschitz bound of Prop. 13 also degrades with the rank margin, and the experiments exclude near planar inputs.

  3. In the experiments the Gram model is more accurate than averaging at N = 5 and N = 8 (Table 7, 0.0015 vs 0.0032 and 0.0047 vs 0.0058). At Gram's worker time it is also better at N = 12 and N = 24 (lines 1877-1880). In the N = 5 trajectory study Gram, the e3nn model and Dym-24 all beat soft averaging (Table 5, lines 1615-1616). The authors say this in lines 129-135, but the abstract reports only N = 24. So the experiments show that averaging beats hard selection, not that frames are a good choice over invariant or equivariant models.

Questions for authors

Clarifications that could change my assessment:

  1. Lines 227-229 ("For higher derivatives, set s=σmin⁡(APj)s = \sigma_{\min}(AP_j) ... They therefore extend by zero across a singular probe.") are not clear to me. The argument in A.6.2 works, but line 1136 ("Lemma 14 and the chain rule give ∥DrY(A)∥≤Crs(A)−r\|D^rY(A)\| \le C_r s(A)^{-r}") skips the Faà di Bruno expansion of ρ(g(A))f(g(A)−1A)\rho(g(A))f(g(A)^{-1}A) and the bound on the derivatives of ff on G⋅KG\cdot K. Please write this step out. Please also say in the main text that kk is a free integer and that ss only enters the estimate. I first read it as kk being chosen from the minimal singular value of the polar decomposition.

  2. Please state Thm 4.21 of Daundkar and García-Calcines (2026) with its exact hypotheses and check them for q:Std,nrow→Grd+(Rn)q: \mathrm{St}^{\mathrm{row}}_{d,n} \to \mathrm{Gr}^+_d(\mathbb{R}^n). For d ≥ 4 the constants of Thm 1 rest on it, since Thm 10 only gives slope 2/(d(d−1)/2+1)2/(d(d-1)/2+1) in n.

  3. Can you compare averaging and hard selection at matched compute, for example hard selection trained for about 6.5 times more updates, and averaging against Gram at equal worker time on the test set? Line 1612 says the soft vs hard comparison does not separate continuity, averaging and branch specialization. Without this, it is not clear the gain comes from continuity.

Suggestions for future work (not part of my assessment):

  1. Lines 116-120 say the lower bound also holds for frames that are permutation invariant and that linear sufficiency is open. Do you expect a linear count for Sn×O(d)S_n \times O(d), or is there a reason it must grow faster?

  2. For SO(3) the action is free on rank 2 inputs, but every probe fails there. Can the probes be extended to inputs of rank at least d-1, and how many frames are needed on that domain? For O(d), what changes with weak equivariance at the reflection stabilizer of a planar cloud?

2 comments

Thank you for your careful review. We have revised the exposition and added the following results and clarifications.

Relation to Dym et al. (ICML 2024). Our linear lower bounds answer their question about stronger lower bounds for continuous weighted rotation frames. They hold already on perfectly conditioned inputs and also constrain frames defined on all inputs (Corollary 5). On full-rank labeled inputs, our explicit construction uses d(n−d)+1 frames, improving their order n^{d−1} to linear growth for every fixed d≥3. Section 3 distinguishes this scope from their treatment of permutations and degenerate inputs.

Q3/W3: comparable compute and accuracy. We retrained from scratch at approximately equal GPU time (Section 7.3, Table 1, Appendix B.7). Hard selection evaluates the network once, using the frame with largest determinant score. Gram predicts from pairwise inner products.

At approximately 11.5 GPU-hours per run and N=24, hard selection achieves NRMSE 0.00296 versus Gram’s 0.00555. Averaging is better per update, reducing NRMSE by 13–23% at equal updates (Section 7.2). Gram leads at short budgets and N=5,8.

Hard training and continuous inference combine without retraining (Section 7.4, Table 11; 20,000 new test configurations):

  • Full blending over 61 frames is continuous by Theorem 2 and lowers NRMSE by 6.3% relative to hard routing of the same networks, giving our best N=24 predictor.
  • Compact blending discards low-weight frames through a cutoff and is C² for a C² backbone. Its NRMSE is 48.2% below the separately trained Gram baseline and 2.7% below hard routing.

We do not attribute the accuracy gains to continuity: uniform top-8 averaging recovers 92% of full blending’s improvement despite lacking a globally continuous weighted frame. Continuity is a separate guarantee. The switching experiments isolate its mechanism. With networks fixed, moving the switching location moves the jump. Its size approaches the disagreement between frame predictions, and averaging removes it.

Q2: the cited theorem. Appendix A.4 now states the hypotheses of Daundkar–García-Calcines Theorem 4.17 in v2 (4.21 in v1). They hold because q is a principal SO(d)-bundle over a paracompact base, hence a fibration, between compact connected manifolds. Knudsen–Weinberger Proposition 5.2 identifies its analog and distributional sectional categories.

Q1: derivative estimates. Appendix A.6, eq. (21), now expands ρ(g(A))f(g(A)⁻¹A) using Faà di Bruno and Leibniz, including bounds on derivatives of f over G·K. Theorem 2’s outline clarifies that k≥1 is fixed independently of the input; s=σ_min(AP_j) enters only the derivative estimates.

W1: exposition. The formal setup and Theorem 1 now occupy Section 2, leaving the introduction free of formal definitions. Frames, weights and other terminology are explained at first use, with eq. (2).

W2: rank assumptions. Full rank is generic. Our lower bounds hold even on perfectly conditioned inputs and constrain robust frames on larger domains (Corollary 5). New Appendix A.10 treats fixed rank r<d, including planar molecules, for O(d), using r(n−r)+1 probes (Proposition 19). Proposition 15 quantifies Lipschitz dependence on the rank margin.

Q4: permutation symmetry. A linear frame count is impossible already in d=4. Two regular m-gons in orthogonal planes give a full-rank input with n=2m columns. Their independent rotations form a group H of order m^2 acting by column permutations. Permutation invariance and equivariance force the weighted frame to be H-invariant. Since H acts freely on O(4), its support has at least m^2=n^2/4 elements. This requires no continuity and leaves the labeled-input theorem unchanged. Appending independent vectors gives the bound (n−d+4)^2/4 for d>4 along this family.

Q5: rank at least d−1. This is a natural question. The group SO(d) acts freely on this larger domain, where continuous equivariant weighted frames exist. Our determinant weights vanish at rank d−1, so an explicit extension requires different weights. Section 8 lists linear-size constructions there, and optimal support bounds with permutation symmetry, as open questions.

Thank you again for your helpful review.

Thanks for addressing most of my concerns. Overall, the paper solved the lower and upper bounds for the continuity of weighted frames, which is an important question, and showed empirically that it holds promise. The experimental verification aligns with what I would expect. The lack of permutation frames is less of a concern since downstream models are usually permutation invariant. I trust the authors will substantially rewrite the paper and improve its presentation.

Reviewer_Ste8OpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors present theoretical results on equivariance learning. They show that there exists a lower bound on the number of frames needed "to preserve continuity".

Critical strengths

I can't judge the strengths of the paper, because I can't parse or understand what the paper is about or trying to achieve.

Critical weaknesses

I can't speak directly to weaknesses in the theory, experimental design, or evaluation of the method.

However, I would like to point out that it's generally concerning for an ICLR submission if even the Introduction section of a paper is not easily accessible to a wider audience at ICLR.

I may not be an expert on the theory of equivariance learning in general, but I have done some work in the general field of equivariant representation learning and am familiar with the general goals and some of the methods used to achieve equivariance. And I have a very hard time getting past the first 3 paragraphs (1st page) of your paper. If I'm honest, I'd have a hard time explaining what the first two paragraphs are about to someone else not working on equivariance learning themselves. The introduction reads like it assumes I already have 6 other papers "loaded into context," and this one simply stacks on top.

I guess this is quite a bit a question of reading "taste" and a question of how broad one assumes the audience of a paper to be, but I'd argue that ICLR is one of the main ML venues that does cater to relatively broad audiences, and papers submitted here should reflect that. Of course, the authors don't need to start from scratch and explain all the basics. Some level of previous knowledge in this subfield can definitely be expected, but I think at least the Intro shouldn't assume that I have read multiple other directly related papers before, only so I can understand the problem this paper solves or why this problem is relevant to solve.

Btw: The fourth paragraph of the paper is already the main Theorem... Even for a pure theory paper I'd expect a lot more introduction and background on the problem, why the problem is important, how the problem arises and giving some intro to the important concepts that are required to then understand the details of the main result.

Questions for authors

Here are the questions I'd like to see answered as part of a push towards a more accessible paper (intro):

  1. What is frame averaging? Intuitively explain: How does it work? And how does it lead to equivariance
  2. Why is it an important tool for equivariant learning?
  3. What is the limitation of frame averaging? (-> discontinuities) -> At this point, explain what is intuitively meant by (dis)continuities.
  4. Following that: how does weighted frame averaging solve the existing problem of discontinuities? How does it improve equivariance learning?
  5. What is the limitation of weighted frame averaging?
  6. Given all of this, explain your research question / contribution / main results. Preferably, first in something close to "layman's terms" or some more intuitive explanation, and then back it up with a more rigorous definition and connect the intuition you provided to the formal definition, so your audience can then more easily dive into the actual theorems and derivations or skip ahead to see experimental verification.

Bonus question: Explain why anyone should care to know "how many frames are needed?" What exactly is the implication of your finding for any potential real-world application of weighted frame averaging methods? Even better if you can back this up with some experimental results.

1 comment

Thank you for your review, and in particular, for spelling out what a broad ICLR audience needs. Your six questions gave the new introduction its structure. The introduction uses minimal notation and explains terminology where it first appears. The formal setup and main theorem now form a new Section 2. In brief, the introduction answers as follows.

  1. Frame averaging. Choose input-dependent coordinates Q (a frame), let the network predict in those coordinates, and rotate the answer back. Averaging repeats this over several frames.
  2. Why it matters. It makes any network, including a pretrained one, exactly equivariant without changing its architecture.
  3. Limitation. • With enough input vectors spanning the space, no single frame can be chosen continuously for all inputs. • The frame must switch somewhere. The prediction can then jump under an arbitrarily small change of input, whenever the two frames give different predictions. 4. Weighted averaging. Predictions from several frames are blended, with weights that suppress failing frames. If the weighted collection varies continuously and rotates with the input, the averaged prediction is continuous and equivariant for every continuous network. 5. Its limitation. Cost. Each frame with positive weight needs its own network evaluation, so the number of frames that continuity requires is the key quantity. 6. Our result. • For labeled vectors spanning a space of fixed dimension, the minimum worst-case number of frames of a continuous equivariant weighted frame grows linearly with the number of vectors. • Both the lower and the upper bound are new. • The lower bounds answer a question posed in ICML 2024 by Dym et al. • The upper bound is a substantial improvement on their construction. For fixed dimension d ≥ 3 it grows linearly rather than as n^{d−1}. • For 24 particles in 3D, at least 11 frames are needed at some input. Our explicit construction uses 61, against 1,012 for theirs.

Bonus question. • Why the count matters: the bound determines how many frames continuity requires. For example, Puny et al.’s eight frames in O(3) cannot give continuous predictions for 24 particles in R^3, where our lower bound is eleven. • Experiment (Section 7): we trained a network on predicting forces with one frame per input, the cheapest option, and then averaged over the frames of our construction without retraining. This gives continuous predictors. • At 24 particles, blending with low-weight frames disregarded, has 48.2% lower force error (NRMSE) than a frame-free baseline built on pairwise inner products and trained for comparable time. • Full blending over all 61 frames has 6.3% lower error than single-frame evaluation of the same networks. • Mechanism: Single-frame predictions jump where the frame switches, and averaging removes the jumps. • Next: long simulations and energy-based force models, where smoothness is used directly, are the next settings to test.

Thank you again for your thoughtful comments.

Comments on the paper

1 comment

We thank all four reviewers for their time. We appreciate Reviewer 5fAk’s close reading of the proofs and their engagement with the revision. We expanded the two proof steps they flagged and ran a comparison at matched compute.

We are disappointed that the other three reviewers did not engage with our response and revision. fmak asked which parts of the main theorem were new. Our reply explained that our main result proves a new lower bound sought by Dym et al (ICML 2024) on the number of frames and shows this bound is asymptotically tight, substantially improving earlier constructions. ZvvA asked what fewer frames buy in practice. Table 13 now gives a direct measurement; for 24 particles at batch size 32, the construction of Dym et al. took 10.3 times as long as ours at inference. We would have valued hearing which concerns these reviewers believe still apply to the revised manuscript.

We used Ste8’s precise explanation of the difficulty with the exposition to structure the new introduction. We strongly disagree with their soundness score of 1. Our results and proofs are correct. We would be happy to address any specific technical concern.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.