pith. machine review for the scientific record. sign in

arxiv: 2605.08286 · v1 · submitted 2026-05-08 · 💻 cs.LG · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Diagnosing Spectral Ceilings in Equivariant Neural Force Fields

Hyunmog Kim
Authors on Pith no claims yet

Pith reviewed 2026-05-12 03:05 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords equivariant neural networksforce fieldsspectral analysisspherical harmonicsmolecular modelingaspirinangular frequenciesneural networks
0
0 comments X

The pith

Equivariant neural force fields preserve angular frequencies only up to a multiple of their layer degree.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a spectral-injection diagnostic to test which angular frequencies equivariant neural force field backbones preserve after training. It injects controlled angular-frequency perturbations into the force field and attaches a lightweight Spectral Prediction Network to the frozen backbone to read off recoverable frequencies. Experiments on aspirin demonstrate that an L=2 NequIP backbone recovers the signal at l=4 but collapses at l=5, showing an 11.7x cliff where p drops from 0.913 to 0.078. A finite-degree span theorem shows that degree-d polynomials of L spherical harmonics span frequencies exactly up to dL for single directions. Multiple controls rule out parameter count as the cause of the observed ceiling.

Core claim

On aspirin, a quadratic SPN attached to an L = 2 NequIP backbone recovers the boundary signal at l = 4 but collapses at l = 5: a 11.7x cliff at the predicted drL boundary, with p dropping from 0.913 to 0.078. The same boundary-vs-above contrast persists across n = 4 independently trained backbones and is corroborated by a denominator-free injected-residual metric. A finite-degree span theorem calibrates the diagnostic by showing that degree-d polynomials of degree-L spherical-harmonic features span exactly H ≤ dL with multiplicity-one saturation at the boundary, scoped to single-direction probes.

What carries the argument

The spectral-injection diagnostic that attaches a lightweight quadratic Spectral Prediction Network to the frozen equivariant backbone to measure recoverable angular frequencies, calibrated by the finite-degree span theorem.

If this is right

  • The equivariant backbone cannot represent or preserve angular force components with frequencies above the dL boundary.
  • The spectral ceiling scales directly with the product of polynomial degree d and layer degree L.
  • This limit is consistent across independently trained models and different architectures.
  • The ceiling is not an artifact of insufficient model capacity but a structural property.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Models aiming for higher accuracy in capturing fine molecular interactions may require increasing L or incorporating higher-degree polynomial expansions.
  • The diagnostic could be used to benchmark and improve future equivariant architectures for better frequency coverage.
  • In molecular simulations, this may lead to systematic errors in predicting high-frequency dynamics or rare events.
  • Generalizing the single-direction span theorem to full multi-atom interactions could provide tighter bounds on model expressivity.

Load-bearing premise

That the lightweight quadratic SPN attached to the frozen backbone isolates and accurately measures only the frequencies preserved by the backbone without the SPN's own training dynamics or capacity introducing bias, and that the single-direction finite-degree span theorem extends meaningfully to the full multi-atom molecular setting.

What would settle it

Applying the diagnostic to an L=3 backbone and finding recovery up to l=6 but not l=7 would support the scaling of the spectral ceiling with dL.

Figures

Figures reproduced from arXiv: 2605.08286 by Hyunmog Kim.

Figure 1
Figure 1. Figure 1: Three complementary measurements of the aspirin spectral cliff at the ideal [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ), while on the low-recovery backbone, all 16 seeds remain flat near the baseline (CV = 0.3%, ∆ = 0.005 kcal/mol/Å). The SPN is a stable instrument; the backbone is the stochastic element [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The spectral-stethoscope workflow. Given a trained equivariant backbone, practitioners [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Probe-degree invariance. At fixed L= 2 backbone, the SPN gain ∆ peaks inside each dr’s algebraic ceiling drL∈ {2, 4, 6} (shaded thresholds) and collapses above it. Curves for dr = 1, 2, 3 are individually consistent with Proposition 2; their convergence to a shared ∆<0.1 floor at ℓ≥5 reflects the fidelity bound (Appendix J). The same fidelity-bound attenuation observed on NequIP (above)—small lifts at ℓ ≤ … view at source ↗
Figure 5
Figure 5. Figure 5: Cubic (dr = 3) SPN self-improvement rself = (base−best)/base across architectures, split by regime. rself magnitudes are not directly comparable across architectures because each backbone’s baseline sits at a different absolute scale (e.g., NequIP L= 2 baseline is already close to its L = 4 reference, leaving little headroom for the cubic SPN, so rself ≈ 0.03; PaiNN baseline is ∼130 kcal/mol/Å, so rself ≈0… view at source ↗
Figure 6
Figure 6. Figure 6: Cliff-location shift with backbone angular cutoff [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: shows the per-atom force MAE gap (yL=2 − yL=4) at linj = 4 for aspirin. Atoms 0, 3 (body-frame carbons), and 10 (anchor atom) exhibit gaps of 5–6 kcal/mol/Å, dominating the aggregate spectral ceiling signal. Hydrogen atoms on the aromatic ring (16, 17) show moderate gaps (∼1.2 kcal/mol/Å), consistent with their distance from the frame center. This spatial pattern confirms that the spectral ceiling is not u… view at source ↗
Figure 8
Figure 8. Figure 8: Per-atom angular bandwidth ℓ ⋆ on a non-redundant RCSB subset (2,414 chains, 4.56 M heavy atoms; 5 Å neighbour ball, SOAP-style smoothed density). Median ℓ ⋆ = 4, P(ℓ ⋆ ≤4)= 88.9%. Small-molecule reference systems (MD17/MD22) overlay as narrower distributions, lining up with the small-molecule SPN cliff at d·L= 4 documented in §4.1. W Implications for Protein Structure Prediction The single-direction d·L a… view at source ↗
read the original abstract

We introduce a spectral-injection diagnostic for measuring which angular frequencies a trained equivariant force-field backbone preserves: inject a controlled angular-frequency perturbation into a molecular force field, attach a lightweight Spectral Prediction Network (SPN) to the frozen backbone, and read off which frequencies are recoverable. On aspirin, a quadratic SPN attached to an L = 2 NequIP backbone recovers the boundary signal at l = 4 but collapses at l = 5: a 11.7x cliff at the predicted drL boundary, with p dropping from 0.913 to 0.078. The same boundary-vs-above contrast persists across n = 4 independently trained backbones (raw-gain delta contrast, hierarchical cluster bootstrap) and is corroborated by a denominator-free injected-residual metric (R2_inj(4) = 0.374 versus R2_inj(5) = 0.006). A finite-degree span theorem calibrates the diagnostic: for a single marked direction, degree-d polynomials of degree-L spherical-harmonic features span exactly H less than or equal to dL with multiplicity-one saturation at the boundary (scoped to single-direction degree-bounded probes, not a function-class upper bound on multi-atom MPNNs). A synthetic C5 calibration plus capacity, activation, and cross-architecture controls rule out parameter count alone as the explanation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a spectral-injection diagnostic for equivariant neural force fields: a controlled angular-frequency perturbation is injected into a molecular system, a lightweight quadratic Spectral Prediction Network (SPN) is attached to the frozen backbone, and recoverability of specific frequencies is measured via signal strength and an injected-residual R² metric. On aspirin with an L=2 NequIP backbone, the method recovers the boundary signal at l=4 but shows collapse at l=5 (11.7x drop, p from 0.913 to 0.078), replicated across four backbones with bootstrap statistics; this is calibrated by a finite-degree span theorem (degree-d polynomials of degree-L spherical harmonics span H ≤ dL with saturation at the boundary) and supported by synthetic C5 controls plus capacity/activation/architecture ablations.

Significance. If the diagnostic holds, it supplies a practical, quantitative tool for identifying spectral ceilings in equivariant architectures used for force fields, which could guide improvements in expressivity without increasing L. Strengths include the denominator-free R²_inj metric, replication across independent backbones, and explicit scoping of the theorem; the synthetic calibration helps rule out trivial capacity explanations.

major comments (2)
  1. [Finite-degree span theorem and aspirin experiments] The finite-degree span theorem is explicitly scoped to single marked directions and degree-bounded probes (as stated in the abstract). The aspirin results involve multi-atom graphs, varying interatomic directions, and message-passing aggregation inside the frozen backbone; no derivation or control is shown demonstrating that the exact H ≤ dL span property (with multiplicity-one saturation) survives this aggregation, which is load-bearing for interpreting the l=4/5 cliff and R²_inj(4)=0.374 vs R²_inj(5)=0.006 contrast as direct evidence of the backbone's preserved frequencies.
  2. [Methods (SPN attachment and training)] The SPN is trained on backbone outputs, introducing potential dependence. While the paper employs a denominator-free injected-residual metric and cites a calibration theorem, the methods must explicitly demonstrate that the injection remains isolated and that SPN training dynamics do not introduce post-hoc selection bias affecting the reported p-value drop and cross-backbone consistency; this is central to the soundness of the quantitative contrasts.
minor comments (2)
  1. [Results (R²_inj reporting)] The exact definition and formula for the injected-residual R²_inj metric should appear in the main text (rather than solely in the appendix) to allow readers to verify the denominator-free property without cross-referencing.
  2. [Figures] Figure captions for the aspirin and synthetic C5 plots should state the precise number of independent backbone trainings and bootstrap iterations used for the p-values and cluster statistics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback on our manuscript. We appreciate the positive assessment of the diagnostic's potential and the noted strengths. We address each major comment below with clarifications and proposed revisions.

read point-by-point responses

  1. Referee: [Finite-degree span theorem and aspirin experiments] The finite-degree span theorem is explicitly scoped to single marked directions and degree-bounded probes (as stated in the abstract). The aspirin results involve multi-atom graphs, varying interatomic directions, and message-passing aggregation inside the frozen backbone; no derivation or control is shown demonstrating that the exact H ≤ dL span property (with multiplicity-one saturation) survives this aggregation, which is load-bearing for interpreting the l=4/5 cliff and R²_inj(4)=0.374 vs R²_inj(5)=0.006 contrast as direct evidence of the backbone's preserved frequencies.

    Authors: We agree that the finite-degree span theorem is derived and scoped strictly to single marked directions with degree-bounded probes, and we do not claim it provides a function-class bound for aggregated multi-atom MPNNs. The manuscript already states this scoping explicitly. The aspirin results are presented as empirical measurements of frequency recoverability from the frozen backbone outputs, with the observed l=4/5 cliff aligning with the dL boundary predicted by the theorem (d=2 for the quadratic SPN). The synthetic C5 calibration, capacity ablations, and cross-backbone bootstrap consistency are intended to support that the contrast is not explained by trivial factors. We acknowledge the absence of a full derivation showing propagation through message-passing aggregation. We will revise the discussion to emphasize the empirical nature of the aspirin contrast and to explicitly note this as a limitation of the current theoretical grounding, rather than interpreting it as direct evidence of the exact span property in the aggregated setting. revision: partial

  2. Referee: [Methods (SPN attachment and training)] The SPN is trained on backbone outputs, introducing potential dependence. While the paper employs a denominator-free injected-residual metric and cites a calibration theorem, the methods must explicitly demonstrate that the injection remains isolated and that SPN training dynamics do not introduce post-hoc selection bias affecting the reported p-value drop and cross-backbone consistency; this is central to the soundness of the quantitative contrasts.

    Authors: We will revise the Methods section to include explicit isolation controls: (i) verification that SPN predictions are near-zero in the absence of any injected signal, (ii) reporting of R²_inj stability across independent SPN training seeds, and (iii) confirmation that the bootstrap procedure for p-values and cross-backbone consistency uses the full distribution without selective thresholding. These additions will demonstrate that the injection remains isolated from the backbone and that training dynamics do not introduce bias into the reported contrasts or the 11.7x drop and R²_inj values. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical diagnostic and scoped theorem are independent

full rationale

The paper's derivation chain consists of an empirical spectral-injection procedure (inject perturbation, freeze backbone, train lightweight quadratic SPN, measure recovery via p and R2_inj) whose outputs are direct performance metrics on aspirin data, plus a finite-degree span theorem that is mathematically derived for single-direction probes and explicitly scoped as not applying to multi-atom MPNNs. No step reduces a reported result to its inputs by construction, renames a fit as a prediction, or relies on load-bearing self-citation. The observed 11.7x cliff and R2_inj contrast are measured quantities, not forced by the equations or theorem; controls (synthetic C5, capacity, cross-architecture) further separate the measurement from any definitional dependence. The central claim therefore retains independent empirical content.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on the newly introduced diagnostic procedure, the finite-degree span theorem as an ad-hoc calibration tool, and the SPN as a new measurement component. Experimental results on aspirin and controls supply the main support.

free parameters (2)
  • L = 2
    Maximum degree of spherical harmonics in the equivariant backbone (set to 2), a design choice that defines the predicted boundary.
  • SPN polynomial degree = quadratic
    Quadratic degree chosen for the Spectral Prediction Network to match the theorem's d parameter.
axioms (1)
  • ad hoc to paper Finite-degree span theorem: for a single marked direction, degree-d polynomials of degree-L spherical-harmonic features span exactly H ≤ dL with multiplicity-one saturation at the boundary
    Theorem introduced in the paper to calibrate the diagnostic; explicitly scoped to single-direction degree-bounded probes.
invented entities (1)
  • Spectral Prediction Network (SPN) no independent evidence
    purpose: Lightweight network attached to frozen backbone to read off recoverable angular frequencies from injected perturbations
    New component created as part of the diagnostic method; no independent evidence outside this work.

pith-pipeline@v0.9.0 · 5540 in / 1636 out tokens · 100043 ms · 2026-05-12T03:05:42.969749+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

  1. [1]

    Abramson et al

    J. Abramson et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630:493–500, 2024

    work page 2024

  2. [2]

    Anderson, T.-S

    B. Anderson, T.-S. Hy, and R. Kondor. Cormorant: Covariant molecular neural networks. In NeurIPS, 2019

    work page 2019

  3. [3]

    Batatia, D

    I. Batatia, D. P. Kovacs, G. N. C. Simm, C. Ortner, and G. Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. InNeurIPS, 2022. 9

    work page 2022

  4. [4]

    Batzner et al

    S. Batzner et al. E(3)-equivariant graph neural networks for data-efficient and accurate inter- atomic potentials.Nature Communications, 13:2453, 2022

    work page 2022

  5. [5]

    Chmiela, V

    S. Chmiela, V . Vassilev-Galindo, O. T. Unke, A. Kabylda, H. E. Sauceda, A. Tkatchenko, and K.-R. Müller. Accurate global machine learning force fields for molecules with hundreds of atoms.Science Advances, 9(2):eadf0873, 2023

    work page 2023

  6. [6]

    T. S. Cohen, M. Geiger, J. Köhler, and M. Welling. Spherical CNNs. InICLR, 2018

    work page 2018

  7. [7]

    Boucheron, G

    S. Boucheron, G. Lugosi, and P. Massart.Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, 2013

    work page 2013

  8. [8]

    G. Cybenko. Approximation by superpositions of a sigmoidal function.Mathematics of Control, Signals and Systems, 2(4):303–314, 1989

    work page 1989

  9. [9]

    K. Hornik. Approximation capabilities of multilayer feedforward networks.Neural Networks, 4(2):251–257, 1991

    work page 1991

  10. [10]

    R. Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B, 99(1):014104, 2019

    work page 2019

  11. [11]

    R. Drautz. Atomic cluster expansion of scalar, vectorial, and tensorial properties including magnetism and charge transfer.Physical Review B, 102:024104, 2020

    work page 2020

  12. [12]

    Bachmayr, R

    M. Bachmayr, R. Drautz, G. Dusson, S. Etter, C. van der Oord, and C. Ortner. Atomic cluster expansion: Completeness, efficiency and stability.Journal of Computational Physics, 454:110946, 2022

    work page 2022

  13. [13]

    Nigam, S

    J. Nigam, S. Pozdnyakov, and M. Ceriotti. Recursive evaluation and iterative contraction of N-body equivariant features.Journal of Chemical Physics, 153:121101, 2020

    work page 2020

  14. [14]

    Nigam, S

    J. Nigam, S. N. Pozdnyakov, K. K. Huguenin-Dumittan, and M. Ceriotti. Completeness of atomic structure representations.APL Machine Learning, 2(1):016110, 2024

    work page 2024

  15. [15]

    F. Bigi, S. N. Pozdnyakov, and M. Ceriotti. Wigner kernels: body-ordered equivariant machine learning without a basis.Journal of Chemical Physics, 161(4):044101, 2024

    work page 2024

  16. [16]

    S. N. Pozdnyakov, M. J. Willatt, A. P. Bartók, C. Ortner, G. Csányi, and M. Ceriotti. Incom- pleteness of atomic structure representations.Physical Review Letters, 125:166001, 2020

    work page 2020

  17. [17]

    Kondor, Z

    R. Kondor, Z. Lin, and S. Trivedi. Clebsch–Gordan nets: A fully Fourier space spherical convolutional neural network. InNeurIPS, 2018

    work page 2018

  18. [18]

    Musaelian et al

    A. Musaelian et al. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications, 14:579, 2023

    work page 2023

  19. [19]

    K. T. Schütt, O. T. Unke, and M. Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. InICML, 2021

    work page 2021

  20. [20]

    K. T. Schütt et al. SchNet: A continuous-filter convolutional neural network for modeling quantum interactions. InNeurIPS, 2017

    work page 2017

  21. [21]

    Duval et al

    A. Duval et al. FAENet: Frame averaging equivariant GNN for materials modeling. InICML, 2023

    work page 2023

  22. [22]

    Dym and H

    N. Dym and H. Maron. On the universality of rotation equivariant point cloud networks. In ICLR, 2021

    work page 2021

  23. [23]

    Geiger and T

    M. Geiger and T. Smidt. e3nn: Euclidean neural networks.arXiv:2207.09453, 2022

    work page arXiv 2022

  24. [24]

    C. K. Joshi et al. On the expressive power of geometric graph neural networks. InICML, 2023

    work page 2023

  25. [25]

    Jumper et al

    J. Jumper et al. Highly accurate protein structure prediction with AlphaFold.Nature, 596:583– 589, 2021. 10

    work page 2021

  26. [26]

    Liao and T

    Y .-L. Liao and T. Smidt. Equiformer: Equivariant graph attention transformer for 3D atomistic graphs. InICLR, 2023

    work page 2023

  27. [27]

    Y .-L. Liao, B. Wood, A. Das, and T. Smidt. EquiformerV2: Improved equivariant transformer for scaling to higher-degree representations. InICLR, 2024

    work page 2024

  28. [28]

    Passaro and C

    S. Passaro and C. L. Zitnick. Reducing SO(3) convolutions to SO(2) for efficient equivariant GNNs. InICML, 2023

    work page 2023

  29. [29]

    M. Rupp, A. Tkatchenko, K.-R. Müller, and O. A. von Lilienfeld. Fast and accurate modeling of molecular atomization energies with machine learning.Physical Review Letters, 108:058301, 2012

    work page 2012

  30. [30]

    O. Puny, M. Atzmon, H. Ben-Hamu, M. Galun, and Y . Lipman. Frame averaging for invariant and equivariant network design. InICLR, 2022

    work page 2022

  31. [31]

    Rahaman et al

    N. Rahaman et al. On the spectral bias of neural networks. InICML, 2019

    work page 2019

  32. [32]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

    N. Thomas et al. Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds.arXiv:1802.08219, 2018

    work page Pith review arXiv 2018

  33. [33]

    Batatia, S

    I. Batatia, S. Batzner, D. P. Kovács, A. Musaelian, G. N. C. Simm, R. Drautz, C. Ortner, B. Kozinsky, and G. Csányi. The design space of E(3)-equivariant atom-centred interatomic potentials.Nature Machine Intelligence, 2024

    work page 2024

  34. [34]

    D. A. Varshalovich, A. N. Moskalev, and V . K. Khersonskii.Quantum Theory of Angular Momentum. World Scientific, 1988

    work page 1988

  35. [35]

    Y . Xie, A. Daigavane, M. Kotak, and T. Smidt. The price of freedom: Exploring expressivity and runtime tradeoffs in equivariant tensor products. InICML, 2025

    work page 2025

  36. [36]

    Simeon and G

    G. Simeon and G. De Fabritiis. TensorNet: Cartesian tensor representations for efficient learning of molecular potentials. InNeurIPS, 2023

    work page 2023

  37. [37]

    Y .-L. Lee, M. Galkin, and S. Miret. Deconstructing equivariance in reinforcement-free molecular property prediction.arXiv:2410.08131, 2024

    work page arXiv 2024

  38. [38]

    Z. Xu, H. Yu, M. Bohde, and S. Ji. PACE: Poincaré-Asymptotic complete equivariance for interatomic potentials.Transactions on Machine Learning Research, 2024

    work page 2024

  39. [39]

    Pacini and R

    M. Pacini and R. Santin. On the universality of equivariant tensor-network architectures. arXiv:2506.02293, 2025

    work page arXiv 2025

  40. [40]

    Lin et al

    Z. Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023

    work page 2023

  41. [41]

    Wohlwend, G

    J. Wohlwend, G. Corso, S. Passaro, M. Reveiz, K. Leidal, W. Swiderski, T. Portnoi, I. Chinn, J. Silterra, T. Jaakkola, and R. Barzilay. Boltz-1: Democratizing biomolecular interaction modeling.bioRxiv, 2024. doi:10.1101/2024.11.19.624167

    work page doi:10.1101/2024.11.19.624167 2024

  42. [42]

    Cliff measurement

    J. D. Jackson.Classical Electrodynamics. John Wiley & Sons, 3rd edition, 1999. A Cross-Molecule Detailed Analysis We apply the diagnostic to four chemically diverse molecules from the sGDML CCSD/CCSD(T) dataset: aspirin (21 atoms, aromatic + multi-functional), ethanol (9 atoms, aliphatic flexibility), mal- onaldehyde (9 atoms, intramolecular proton transf...

    work page arXiv 1999

  43. [43]

    The spectral stethoscope measures whether the target property actuallyrequireshigher angular resolution, avoiding blind architectural scaling

    Diagnose before scaling.Raising L is expensive (O(L3) CG coupling cost). The spectral stethoscope measures whether the target property actuallyrequireshigher angular resolution, avoiding blind architectural scaling. The per-atom analysis (Appendix T) further localizes spectral deficits to specific atoms, guiding targeted interventions

    work page

  44. [44]

    Depth empirically improves the recovery fraction; the cliff location is depth-invariant. At fixed (LB, dr), deeper backbones yield higher within-ceiling ρ (Table 7); the cubic- readout clifflocationis at ℓ⋆ =d rLB independently of depth (Appendix N), as Lemma 2 predicts. We interpret the depth effect on ρ as a within-ceiling fidelity gain (and partial tru...

    work page

  45. [45]

    underfitting,

    Width cannot substitute for reach.The nf33 null result is unequivocal: adding parameters within a fixed polynomial structure provides zero spectral extension. Before attributing a performance gap to “underfitting,” practitioners should check that the gap is not a spectral ceiling—i.e., that the target content lies within the backbone’sd·Lrange. Angular co...

    work page arXiv 2000

  46. [46]

    Take a protein ensemble (CASP/mdCATH/ATLAS MD frames, or a single PDB with rotamer perturbations)

    work page

  47. [47]

    Compute a reference geometric field: atomic forces from a classical force field (Am- ber/CHARMM), or differential quantities from an AlphaFold-quality predictor

    work page

  48. [48]

    Project the field onto Yℓm(ˆr)in each atom’s local frame, yielding per-atom coefficients {cℓm(i)}

    work page

  49. [49]

    use equivariance?

    Compute the per- ℓ variance fraction w(ℓ) =∥c ℓ∥2/∥c∥2 (distinct from the main-text recovery fractionρ; this is an input-side basis-occupancy quantity). 40 The smallest ℓ⋆ withP ℓ≤ℓ⋆ w(ℓ)≥0.95 (theeffective protein bandwidth) is an architecture-free measurement of the one-hop input-density bandwidth. An ideal degree-bounded input readout would avoid the c...

    work page