Wavelet Scattering Transform

• Wavelet scattering transform is a multiscale, stable, and interpretable signal representation that cascades fixed wavelet convolutions, modulus nonlinearities, and low-pass averaging to achieve translation invariance and preserve local geometry.
• It bridges classical signal processing and deep learning by using predefined filters to capture localized multiscale structures without relying on backpropagation-based learning.
• WST has been successfully applied in areas such as image classification, gravitational-wave analysis, and RF sensing, providing robust features that highlight sparsity and morphological organization.

The wavelet scattering transform (WST) is a nonlinear, multiscale signal representation built by cascading wavelet convolutions, modulus nonlinearities, and low-pass averaging. It is commonly described as a convolutional network with predefined filters rather than learned ones, positioned between classical signal processing and deep networks: it is designed to produce coefficients that are translation invariant at a chosen scale, stable to small deformations, and still informative about local geometric structure. In the literature summarized here, WST appears in 1D, 2D, and 3D forms; with Morlet, monogenic, solid harmonic, diffusion, and group-theoretic wavelets; and on Euclidean as well as non-Euclidean domains such as graphs, manifolds, and finite groups (Andreux et al., 2018, Nicola et al., 2022, Chew et al., 2022).

1. Conceptual foundations and historical role

In the standard scattering perspective, the transform is a multilayer architecture that alternates three operations: convolution with a wavelet filter bank, a pointwise nonlinearity—typically the complex modulus—and low-pass averaging. This organization was introduced as a mathematically grounded model of the early layers of convolutional neural networks and as a way to encode stability and invariance analytically rather than by supervised filter learning. Several papers in the corpus explicitly describe scattering as a CNN-like system with fixed filters and no backpropagation-based learning of convolution kernels (Oyallon et al., 2013, Andreux et al., 2018, Chak et al., 2022).

A central motivation is that many tasks tolerate, or require, invariance to translation and robustness to small geometric deformations. In the scattering framework, these are not post hoc empirical properties but design goals of the architecture itself. The low-pass operator sets the invariance scale, while the wavelet cascade preserves local multiscale structure that would be lost under averaging alone. This makes scattering a representation rather than a classifier: the coefficients are typically fed to a separate downstream model such as a linear SVM, random forest, or task-specific classifier (Oyallon et al., 2013, Andreux et al., 2018, Licciardi et al., 2024).

The literature also emphasizes what scattering is not. It is not a learned deep network in the usual sense; it does not rely on max pooling in its canonical form; and it does not treat feature extraction as an opaque latent embedding problem. In Bruna and Mallat’s object-recognition study, a two-layer scattering network with no learning and no max pooling gave strong image-classification performance and was interpreted as approximating the first two layers of a pretrained ConvNet on Caltech benchmarks (Oyallon et al., 2013). Later work retained the same basic view while asking whether the fixed wavelet family itself should remain hand-designed or be learned in a low-parameter way (Gauthier et al., 2021).

2. Mathematical construction

A standard discrete formulation, used by Kymatio, defines scattering on a signal $x[n]$ with a wavelet family $\{\psi_\lambda\}$, a low-pass filter $\phi_J$, and a pointwise nonlinearity $\rho$. The zeroth-, first-, and second-order coefficients are

$S_0 x[n] = x \ast \phi_J[n],$

$$S_1 x[n,\lambda_1] = \rho(x \ast \psi_{\lambda_1}) \ast \phi_J[n],$$

$$S_2 x[n,\lambda_1,\lambda_2] = \rho\!\big(\rho(x \ast \psi_{\lambda_1}) \ast \psi_{\lambda_2}\big)\ast \phi_J[n],$$

with $\rho(t)=|t|$ in the common complex-valued case. The first wavelet transform resolves localized multiscale structure; the modulus demodulates oscillatory coefficients; the second wavelet transform recovers interactions lost after the first modulus; and the final averaging introduces translation invariance at scale $2^J$ (Andreux et al., 2018).

A path-based notation makes the recursive structure explicit. For a path $p=(\lambda_1,\dots,\lambda_m)$, one writes

$\{\psi_\lambda\}$0

with $\{\psi_\lambda\}$1 and $\{\psi_\lambda\}$2. This makes scattering a cascade of propagators followed by low-pass projection, and clarifies why the representation is naturally indexed by order, scale, and orientation rather than by learned feature channels (Chak et al., 2022).

In 2D image settings, the filters are often directional Morlet wavelets indexed by logarithmic scale $\{\psi_\lambda\}$3 and orientation $\{\psi_\lambda\}$4. For an image $\{\psi_\lambda\}$5, the Roman detector-analysis paper writes

$\{\psi_\lambda\}$6

$\{\psi_\lambda\}$7

$\{\psi_\lambda\}$8

In that interpretation, $\{\psi_\lambda\}$9 quantifies correlations between wavelet-detected features: it measures the clustering strength of features of size $\phi_J$0 on scales $\phi_J$1, with $\phi_J$2 used to avoid redundant symmetry (Velicheti et al., 2023).

Because raw coefficient tensors are often large, many applications compress them into summary statistics. Roman dark-image analysis uses the orientation-averaged quantities

$\phi_J$3

interpreted as a sparsity statistic, and

$\phi_J$4

interpreted as a shape or filamentarity statistic. In that usage, WST is not merely a black-box embedding: the coefficients are deliberately tied to morphology, clustering, anisotropy, and multi-scale organization (Velicheti et al., 2023).

3. Invariance, stability, and information beyond second-order spectra

The most widely cited structural properties of WST are translation invariance, deformation stability, and non-expansiveness. Low-pass averaging by $\phi_J$5 yields translation invariance at scale $\phi_J$6, while the wavelet cascade preserves localized multiscale structure. Kymatio’s summary states that scattering seeks the “sweet spot”: more invariant than raw signals, but more structured and interpretable than fully learned black-box representations (Andreux et al., 2018).

Theoretical work sharpens the stability claims. In gravitational-wave preprocessing, WST is described as stable to additive noise, locally translation invariant, and stable to small continuous deformations or time warping; under suitable assumptions it is stated to be Lipschitz-continuous under the action of $\phi_J$7-diffeomorphisms, and non-expansive in the sense that

$\phi_J$8

A more refined mathematical study identifies a regularity threshold in the Hölder scale: stability is still achievable for deformations of class $\phi_J$9, $\rho$0, whereas instability phenomena can occur for $\rho$1; the borderline Lipschitz or $\rho$2 case is handled only up to $\rho$3 losses (Licciardi et al., 2024, Nicola et al., 2022).

The information-theoretic motivation for WST is equally important. A power spectrum measures how variance is distributed across frequencies or wavenumbers, but it discards Fourier phase and thus cannot robustly distinguish many fields with identical or near-identical spectra. Multiple domain papers therefore present WST as a way to capture geometric information beyond second-order statistics: localization, sparsity, elongation, orientation, clustering, and cross-scale organization. The ocean-flow study states that ST distinguishes balanced dynamics, internal waves, and types of turbulence even when their power spectra are identical; the Roman detector study shows that very different patterns can occupy clearly separated regions in $\rho$4–$\rho$5 space even when power spectra are similar or only weakly distinctive (Skinner et al., 1 May 2025, Velicheti et al., 2023).

At the same time, the literature is explicit about limitations. WST is not a universal detector; summary statistics tailored to sparsity and filamentarity may miss other complex patterns unless the wavelet settings or summaries are expanded. Higher orders also decay rapidly in energy, which is why many implementations stop at second order or, less commonly, third order. In the gravitational-wave study, only 2 or 3 layers are said to be often sufficient to capture nearly all the energy, around 98% (Velicheti et al., 2023, Licciardi et al., 2024).

4. Filter families, variants, and generalized geometries

The canonical 1D and 2D implementations use complex Morlet wavelets. Kymatio adopts Morlet wavelets for 1D and 2D scattering, while 3D scattering uses solid harmonic wavelets $\rho$6 with a modulus defined over harmonic responses. In 3D line-intensity mapping, the first- and second-order solid harmonic coefficients are used to separate scale, angular structure, and multiscale morphology; the authors conclude that the full shape-preserving WST can outperform the combination of the power spectrum and the voxel intensity distribution, while reduced “shapeless” coefficients mainly probe scale interactions (Andreux et al., 2018, Chung, 2022).

Several papers modify the wavelet family itself. The Monogenic Wavelet Scattering Network replaces 2D Morlet wavelets with a monogenic construction based on a Gaussian high-pass filter and the Riesz transform, arguing that 2D monogenicity is a more faithful analogue of 1D analyticity. On CUReT, the monogenic variant reaches 97.34% classification accuracy, while the best tested Kymatio-STN baseline reaches 96.61% at $\rho$7 (Chak et al., 2022). Parametric scattering retains the scattering cascade but learns Morlet parameters $\rho$8, concluding that traditional tight-frame constructions may not always be necessary for effective representations and that learned versions can yield significant performance gains in small-sample classification settings (Gauthier et al., 2021).

Other extensions modify the symmetry group or the later layers. EqWS introduces “triglets” and a translationally invariant, rotationally equivariant scattering network whose coefficients are exactly equivariant to discrete rotations $\rho$9; it also notes that scale equivariance is substantially harder because of aliasing, boundary effects, and scale-dependent ringing (Saydjari et al., 2021). A hybrid scattering model for signals with isolated singularities uses wavelets in the first layer to promote sparsity and Gabor filters in the second layer to exploit the sparse singular structure created by the first layer (Perlmutter et al., 2021). Phase scattering, while no longer standard WST, replaces propagated magnitudes by phase derivatives derived from the STFT and is proposed as a complementary representation when precise localization of harmonic or transient structure matters (Haider et al., 2022).

The scattering paradigm has also been generalized far beyond Euclidean domains. Geometric scattering on measure spaces builds the transform from spectral filters of a self-adjoint operator on $S_0 x[n] = x \ast \phi_J[n],$0, recovering earlier graph- and manifold-scattering constructions and extending them to directed graphs, signed graphs, and manifolds with boundary (Chew et al., 2022). A finite-group construction defines wavelets directly from irreducible characters and proves non-expansiveness, energy preservation, stability, and equivariance with respect to left and right group translations on arbitrary finite groups, including nonabelian ones (Arias et al., 27 May 2025).

5. Empirical applications across scientific and engineering domains

The range of reported applications is unusually broad, but a common pattern recurs: WST is used when morphology, sparsity, or multiscale organization matter and a variance-only spectral summary is insufficient.

Domain	WST role	Reported outcome
Roman WFI dark images	Patch-level morphology screening	39,600 patches analyzed; about 0.2% contain rare filament-like or bubble-like structures (Velicheti et al., 2023)
Texture and object recognition	Fixed or monogenic feature extractor	97.34% on CUReT for MWSN; 68.8 ± 0.5 on Caltech-101 for 2-layer scattering (Chak et al., 2022, Oyallon et al., 2013)
Gravitational-wave glitches	Alternative or complement to Q-transform	91.03% test accuracy for WST + Q-transform ensemble; WST GPU runtime 0.063 ± 0.002 s/sample (Licciardi et al., 2024)
RF UAV detection	Scattergram and coefficient features	98.9% accuracy at 10 dB SNR with steady-state WST scattergrams and SqueezeNet (Medaiyese et al., 2021)
Low-resourced spoken LID	Alternative front-end to MFCC	EER reduced upto 14.05% and 6.40% in same-corpora and blind evaluations (Dey et al., 2023)
Ocean and cosmological fields	Morphology beyond power spectra	Spectrally identical flows separated; $S_0 x[n] = x \ast \phi_J[n],$1 with $S_0 x[n] = x \ast \phi_J[n],$2 gives $S_0 x[n] = x \ast \phi_J[n],$3 and $S_0 x[n] = x \ast \phi_J[n],$4 (Skinner et al., 1 May 2025, Jiang et al., 20 May 2025)

In detector characterization for the Nancy Grace Roman Space Telescope, WST is applied to dark-current-rate images split into $S_0 x[n] = x \ast \phi_J[n],$5 patches. The method separates common detector backgrounds from rare filament-like and donut-like artifacts in the $S_0 x[n] = x \ast \phi_J[n],$6–$S_0 x[n] = x \ast \phi_J[n],$7 plane, and the authors argue that it is particularly valuable for small-scale Roman detector characterization relevant to weak-lensing systematics (Velicheti et al., 2023).

In image classification, the same fixed-filter philosophy spans generic object recognition and domain-specific texture analysis. A two-layer scattering network using spatial and joint spatial-angular-scale wavelets achieves 68.8 ± 0.5 on Caltech-101 and 34.6 ± 0.2 on Caltech-256, and is explicitly compared with the first two layers of a pretrained ImageNet ConvNet (Oyallon et al., 2013). The monogenic variant then shows that changing the wavelet family can improve texture classification by better preserving 2D geometric structure (Chak et al., 2022).

In time-series physics, WST is used as a preprocessing stage rather than a classifier. For gravitational-wave glitches, first-order WST reaches 88.11% test accuracy, second-order WST 68.65%, combined WST 88.57%, Q-transform 87.41%, and a WST + Q-transform ensemble 91.03%; the paper argues that WST simplifies classification enough that a random forest suffices, while also being less sensitive to color-scale choices than spectrogram-based pipelines (Licciardi et al., 2024). In oceanography and cosmology, the transform is explicitly valued because it distinguishes regimes that are spectrally degenerate but geometrically distinct, such as 2D turbulence versus SQG turbulence, or different cosmological models under tracer-bias mitigation (Skinner et al., 1 May 2025, Jiang et al., 20 May 2025).

In RF sensing and speech, WST is used as an engineered front-end for difficult low-resource or low-SNR settings. The RF UAV study reports that WST scattergrams extracted from the steady state of RF control signals and fed to SqueezeNet achieve 98.9% accuracy at 10 dB SNR (Medaiyese et al., 2021). The spoken-language-identification study argues that WST preserves modulation detail discarded by mel/MFCC front-ends and reports EER reductions upto 14.05% in same-corpus evaluation and 6.40% in blind VoxLingua107 evaluation (Dey et al., 2023).

6. Software, computation, and methodological considerations

The most prominent software platform in this corpus is Kymatio, a Python implementation of scattering transforms in 1D, 2D, and 3D with frontends for NumPy, scikit-learn, PyTorch, and TensorFlow/Keras. Its implementation choices are algorithmically consequential: FFT-based periodic convolution, downsampling of low-frequency intermediate signals, and depth-first traversal of the scattering network to reduce memory footprint, especially on GPUs (Andreux et al., 2018).

Practical WST usage is highly sensitive to parameterization, reduction strategy, and domain geometry. In Roman detector analysis, $S_0 x[n] = x \ast \phi_J[n],$8 and $S_0 x[n] = x \ast \phi_J[n],$9 are used after testing $$S_1 x[n,\lambda_1] = \rho(x \ast \psi_{\lambda_1}) \ast \phi_J[n],$$0 and finding no substantial difference in the aggregate trends (Velicheti et al., 2023). In spoken language identification, low octave resolution is consistently best, specifically $$S_1 x[n,\lambda_1] = \rho(x \ast \psi_{\lambda_1}) \ast \phi_J[n],$$1, and frequency-scattering is found not to be useful overall (Dey et al., 2023). In 3D line-intensity mapping, raw WST covariances are often poorly conditioned; the paper therefore introduces reduced “shapeless” and rescaled coefficient sets, and finds that $$S_1 x[n,\lambda_1] = \rho(x \ast \psi_{\lambda_1}) \ast \phi_J[n],$$2 is the best overall compromise for that study (Chung, 2022).

Normalization and rendering choices can also materially affect performance. The gravitational-wave paper reports that WST classification works better with row-wise normalization and without applying a colormap, in contrast to Q-transform spectrograms whose performance depends strongly on capped color scaling (Licciardi et al., 2024). Such observations matter because they illustrate a recurring theme in the literature: WST often behaves more like a feature representation than like an image to be visually rendered.

Several recurrent cautions qualify the transform’s scope. WST is interpretable and stable, but it is not universally optimal. EqWS explicitly notes that performance on difficult natural-image benchmarks is not state of the art (Saydjari et al., 2021). Roman detector analysis states that WST is not a universal detector (Velicheti et al., 2023). Parametric scattering shows that fixed wavelet filterbanks are a strong prior, but not necessarily the best discriminative choice in every dataset (Gauthier et al., 2021). Taken together, these results situate WST not as a replacement for all learned representations, but as a mathematically structured family of multiscale representations whose main strengths are controlled invariance, deformation stability, compactness, and coefficient-level interpretability.