The Stability Landscape in Wave-Packet Scattering: Geometric Rigidity and Sharp Sobolev Thresholds

Source: arXiv:2607.21578 · Published 2026-07-23 · By Max Getter, S. Ivan Trapasso

TL;DR

This paper addresses the fundamental tension between the resolution power and deformation stability of nonlinear multiscale representations in harmonic analysis, motivated by the wavelet scattering transform originally introduced by Mallat. Specifically, it investigates how the geometric properties of frequency decompositions underlying general wave-packet scattering systems—such as wavelets, curvelets, shearlets, and wave atoms—affect stability under small smooth coordinate deformations (diffeomorphisms). The authors identify a sharp resolution-robustness trade-off governed by the transverse scale of frequency tiles: finer transverse resolution (as in curvelets/shearlets) enables better approximation but causes instability since small deformations can displace frequency content across adjacent channels, breaking stability even at the first scattering layer. They rigorously rule out uniform Lipschitz deformation-stability estimates below a critical Sobolev regularity threshold H^{1-eta}, where β measures transverse tile narrowness, while establishing matching positive stability results at and above this threshold via anisotropic paradifferential commutator bounds. These results provide the first intrinsic deformation-stability theory for scattering beyond wavelets, clarifying the precise geometric and functional-analytic conditions under which scattering transforms remain robust to nonlinear coordinate changes.

Key findings

For wave-packet systems with transverse scale exponent β < 1 (e.g., curvelets/shearlets), arbitrarily small smooth deformations can induce significant instability in first-layer scattering coefficients, with a uniform ℓ2(L2) difference bounded below by a positive constant independent of deformation size (Theorem 1.1).
Instability cannot be avoided by restricting to L2 inputs; no Lipschitz deformation-stability estimate holds below the critical Sobolev regularity s < 1 − β on inputs in H^s(R^2) (Corollary 1.2).
The critical Sobolev threshold H^{1−β} exactly characterizes stability: scattering transforms are Lipschitz-stable to deformations with gradient bounds on inputs measured in a cumulative energy E_{1−β}(f) that sums Sobolev norms of all propagated layers (Theorem 1.3).
Under additional energy localization assumptions on the filters and when the radial scale exponent α ≤ 1 (with 0 ≤ β ≤ α < 1), this multilayer Sobolev energy can be controlled by the initial input Sobolev norm ∥f∥_{H^{1−β}}, yielding sharp, uniform Lipschitz deformation bounds at the critical threshold (Corollary 1.5).
The geometric parameter β measuring transverse frequency tile width governs the scale of adversarial deformations: deformation size ∼2^{−(1−β)j} at scale 2^j can shift frequency content across channels, causing instability.
A constructive commutator lower bound shows the stability failure arises at the linear filter transform level, independent of the nonlinear modulus operation or input/deformation smoothness (Theorem 1.1).
The results unify and generalize prior wavelet scattering stability theorems, showing wavelets correspond exactly to the exceptional case β = 1 where geometric rigidity enforces stability without additional input regularity assumptions.
The positive deformation stability result relies on a novel anisotropic paradifferential commutator estimate that quantifies how smooth deformations interact with wave-packet multipliers at multiple scales (Section 6).

Threat model

The adversary is an entity applying a small, smooth, compactly supported spatial deformation (diffeomorphism) τ : R^2 → R^2 with bounded derivatives less than 1, aiming to disrupt the stability of multiscale scattering representations by shifting high-frequency content across narrow frequency channels. The adversary cannot perform large, non-smooth, or non-localized deformations and has no internal access to or ability to train the scattering filters; their power lies purely in nonlinear coordinate changes acting on the input signals.

Methodology — deep read

Threat Model & Assumptions: The adversary is modeled as a small smooth compactly supported diffeomorphism τ on the spatial domain R^2 with bounded derivatives (∥Dτ∥_{L∞} < 1). The analysis focuses on stability of scattering transforms under such spatial coordinate deformations, testing whether the representation approximately intertwines with the pullback operator. The adversary cannot perform non-smooth or large deformations nor access or train the filters.
Data and Function Classes: The input signals f are taken in the classical function spaces L2(R^2) and Sobolev spaces H^s(R^2), with s parameterizing input regularity. The main contrasting regimes consider s < 1 − β (instability) versus s ≥ 1 − β (stability), where β is a geometric parameter of frequency tile width. No empirical dataset is used; rather, explicit counterexamples (wave-packets f_j) and smooth diffeomorphisms τ_j are constructed in Schwartz space to witness instability gaps.
Architecture & Algorithms: The scattering transform S_χ is a nonlinear multiscale hierarchical feature map built on a Parseval frame filter bank of wave-packet multipliers W = {χ, ψ_λ}_{λ∈Λ}, where frequency tiles Q_{j,m,ℓ} have radial size ∼ 2^{αj} and transverse width ∼ 2^{βj}, controlled by parameters 0 ≤ β ≤ α ≤ 1. The transform applies iterative convolutions with these filters, followed by modulus nonlinearities and low-pass averaging convolution with χ at scale R. The layers output coefficients U[p]f = |...||f ∗ ψ_{λ_1}| ∗ ψ_{λ_2}|...|∗ ψ_{λ_n}| for paths p=(λ_1,...,λ_n). The overall transform maps into ℓ2(L2), the space of coefficients indexed by scattering paths.
Training Regime: No training is involved as the filters are fixed analytic constructions. The analysis is fully mathematical and theoretical, relying on harmonic analysis tools such as Littlewood-Paley decompositions, paradifferential calculus, and energy estimates. The proofs leverage intricate commutator calculations and geometric measure estimates.
Evaluation Protocols: Stability is assessed by computing norms of commutator defects [W, L_τ]f = W(L_τ f) − L_τ(Wf) and deformation differences ∥S_χ(L_τ f) − S_χ f∥_{ℓ2(L2)}. Lower bounds show uniform separation for certain sequences of inputs and deformations. Upper bounds use Lipschitz estimates involving the Sobolev regularity of inputs and deformation gradients. Stability is related to the Sobolev embedding scale s relative to the geometric β parameter. Results hold for infinite depth scattering and its finite truncations. Comparisons to previous wavelet-only stability results highlight generalization.
Reproducibility: The paper provides explicit filter bank constructions and detailed hypotheses for assumptions but no computational code or datasets are released as the work is fully analytic. The framework is general for Euclidean domains with two spatial dimensions. The proofs are constructive but technical, allowing reproducibility by mathematical derivation rather than experimental replication.
Concrete example walkthrough: The authors construct wave-packet sequences f_j localized at frequency scales 2^j with L2 norm 1 and explicit smooth compactly supported diffeomorphisms τ_j with ∥τ_j∥_{C^m} ≲ 2^{-(1−β)j}. These examples cause significant energy shuffling across transverse frequency tiles of width 2^{βj}, leading to a positive lower bound on the scattering transform difference ∥S_χ(L_{τ_j} f_j) − S_χ f_j∥ ≥ c. This demonstrates instability below the Sobolev threshold s < 1 − β, proving no uniform Lipschitz deformation bound can hold in that regime. Conversely, by analyzing commutators between the deformation flow and multipliers, the authors prove at and above the threshold, the scattering transform gains stability with Lipschitz bounds controlled by the cumulative Sobolev energy E_{1−β}(f).

Technical innovations

Identification of the geometric parameter β (transverse frequency tile width exponent) as the sharp determinant of deformation stability thresholds in wave-packet scattering systems.
Constructive demonstration that arbitrarily small smooth deformations can destabilize scattering transforms with β < 1 by shifting frequency content across adjacent narrow channels, even at first layer.
Introduction of anisotropic paradifferential commutator estimates tailored to wave-packet multipliers and transport vector fields, enabling sharp stability bounds at the critical Sobolev scale H^{1−β}.
Extension of Mallat’s wavelet scattering deformation stability theory to general anisotropic wave-packet systems including curvelets, shearlets, and wave atoms via precise geometric rigidity and energy propagation analysis.

Baselines vs proposed

Wavelet scattering transform: Lipschitz deformation stability proven at smoothness threshold H1(R2) corresponding to β = 1.
Curvelet/shearlet-based scattering (β < 1): instability below H^{1−β} shown by fixed positive ℓ2(L2) difference on first layer vs stability recovered at/above threshold with Lipschitz bound involving E_{1−β}(f).

Limitations

Results focus on two-dimensional Euclidean domains; extensions to higher dimensions involve additional technical challenges.
Stability proofs rely on smooth, bounded gradient diffeomorphisms; less regular or larger deformations are not covered.
Energy propagation control to reduce cumulative Sobolev energy to standard Sobolev norm requires restrictive filter bank localization assumptions.
The precise behavior exactly at the endpoint regularity s = 1 − β needs delicate geometric conditions (α < 1) to fully close the stability gap.
The instability constructions are somewhat pathological and constructed in Schwartz space; practical input signals may not realize worst-case behavior.
No empirical validation or numerical experiments accompany the theoretical results.

Open questions / follow-ons

Extension of the stability–instability characterization to three and higher spatial dimensions with potentially more complex geometric frequency tilings.
Relaxing the smoothness and bounded gradient assumptions on deformations, including possible stability results under more general or stochastic spatial perturbations.
Numerical validation of the theoretical instability and stability bounds in practical scattering implementations on real signals.
Design of adaptive or learned filter banks that balance fine resolution and deformation robustness optimally beyond fixed analytic wave-packet systems.

Why it matters for bot defense

For bot-defense and CAPTCHA practitioners, this work provides fundamental insights into how multiscale feature extraction pipelines behave under small geometric deformations—analogous to subtle image or audio transformations an adversarial bot might exploit. The identified geometric threshold explains why certain finely resolved directional decompositions (curvelets/shearlets) are more vulnerable to displacement-induced instability, potentially weakening robustness against adversarial transformations. This suggests caution when choosing or designing scattering-based features or similar CNN-like architectures for CAPTCHA verification, especially when expecting attackers to employ smooth but carefully crafted manipulations to evade detection. On the positive side, the results pinpoint the required input regularity assumptions and architectural properties (e.g., wavelet-like geometric parameters) necessary to guarantee deformation stability, guiding principled CAPTCHA feature design to enhance robustness. More generally, the commutator approach offers a rigorous toolset to analyze and certify deformation stability at multiple layers, informing defenses against adversarial geometric perturbations in bot and fraud detection systems.

Cite

bibtex

@article{arxiv2607_21578,
  title={ The Stability Landscape in Wave-Packet Scattering: Geometric Rigidity and Sharp Sobolev Thresholds },
  author={ Max Getter and S. Ivan Trapasso },
  journal={arXiv preprint arXiv:2607.21578},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.21578}
}

The Stability Landscape in Wave-Packet Scattering: Geometric Rigidity and Sharp Sobolev Thresholds ​

TL;DR ​

Key findings ​

Threat model ​

Methodology — deep read ​

Technical innovations ​

Baselines vs proposed ​

Limitations ​

Open questions / follow-ons ​

Why it matters for bot defense ​

Cite ​

Read the full paper ​