Wavelet State Space Models

• Wavelet State Space Models are frameworks that integrate multiscale wavelet transforms with state-space dynamics to capture and model temporal or spatial signals.
• They enable robust inference by matching empirical wavelet variance to model-implied signatures, as seen in methodologies like the Generalized Method of Wavelet Moments.
• Recent developments extend these models to rational filter realizations, Bayesian VAR, and deep architectures, enhancing applications in signal processing, control theory, and computer vision.

Wavelet State Space Model denotes a family of constructions in which wavelet-domain representations and state-space formalisms are coupled, rather than a single standardized model class. In the literature, this coupling appears in several distinct senses: wavelet variance as a moment representation for latent linear state-space estimation; state-space realization of rational wavelet filters; time-varying Bayesian state-space models with wavelet-expanded predictor sets; latent stochastic processes defined directly in wavelet space; and modern deep sequence or vision architectures in which S4-, Mamba-, or VSSM-type state dynamics are built over wavelet frames or wavelet sub-bands (Balamuta et al., 2016, Alpay et al., 2011, Cekic et al., 2017, Solozabal et al., 25 Feb 2026).

1. Scope and taxonomic meaning

The term is best understood as an umbrella expression for models where wavelets provide multiscale localization, frequency partition, or frame coordinates, while the state-space component provides latent dynamics, recursive inference, transfer-matrix realization, or efficient long-range sequence modeling. This suggests that the phrase is polysemous across statistics, control, signal processing, neuroscience, and deep learning.

Family	Wavelet role	State-space role
Wavelet-variance estimation	Wavelet coefficients and wavelet variance	Latent linear state-space / structural model inference
Rational wavelet filters	Paraunitary filter-bank structure	Input-output and state-space realization
Multiscale Bayesian causality	$\textit{a trous}$ Haar predictors	Time-varying VAR in state-space form
Deep wavelet SSMs	Wavelet frames or DWT/IWT sub-bands	S4/Mamba/VSSM sequence modeling

A recurrent misconception is that a wavelet state-space model must always be a stochastic latent model whose hidden state is itself a wavelet decomposition. The literature does not support such a restriction. In some works, the wavelet object is a summary statistic or design matrix; in others, it is a transfer function factorization; in recent deep-learning papers, it can be the basis used to initialize the state matrix or the frequency domain in which sequence modeling is concentrated (Balamuta et al., 2016, Alpay et al., 2011, Babaei et al., 9 Jun 2025).

2. Wavelet variance and latent linear state-space inference

A classical statistical meaning of wavelet state-space modeling is exemplified by the gmwm framework, where wavelet decomposition yields scale-wise coefficients $W_{j,t}$ and the central summary is the wavelet variance

$\nu_j^2 = \mathrm{Var}(W_{j,t}).$

For a parametric latent model with parameter $\theta$, the model-implied multiscale signature is

$$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$

This representation is central because many observed processes are modeled as sums of latent components,

$Y_t = \sum_{k=1}^{K} X_t^{(k)},$

with additive wavelet variance under independence,

$$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$

The cited component library includes white noise, random walk, drift, quantization noise, and AR(1), together with structural combinations of these (Balamuta et al., 2016).

Estimation proceeds through the Generalized Method of Wavelet Moments (GMWM), a GMM-type criterion in which empirical and theoretical wavelet variances are matched: $$\hat{\theta} = \arg\min_{\theta \in \Theta} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right)^\top \boldsymbol{\Omega} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right).$$ The package paper describes this as a framework that “delivers a general framework for the robust estimation of many time series models as well as a quick and efficient estimation of many linear state-space models” (Balamuta et al., 2016). The associated workflow is explicit: compute the empirical wavelet variance, inspect the WV plot, specify a latent model as a sum of components, estimate parameters via GMWM, assess fit, and optionally use robust estimation.

The robust extension broadens this program. “Robust Inference for Time Series Models: a Wavelet-based Framework” states that the method covers models “going from ARMA to state-space models,” and provides estimators that are “consistent and asymptotically normally distributed,” “straightforward to implement,” and “computationally efficient” (Guerrier et al., 2015). It also introduces a robust estimator of wavelet variance and identifies conditions under which GMWM is identifiable for various classes of time series models. In this line of work, a wavelet state-space model is therefore a latent structural time-series model estimated through its multiscale wavelet-variance signature rather than through direct full likelihood or Kalman-filter likelihood evaluation (Balamuta et al., 2016).

3. State-space realization of wavelet filters

A different and older meaning arises in system theory, where the object is not a stochastic latent process but a rational wavelet filter bank. “Easy-to-compute parameterizations of all wavelet filters: input-output and state-space” studies $N\times N$ rational wavelet filters $W(z)$ satisfying paraunitarity and cyclic filter-bank symmetry, with class

$W_{j,t}$0

These filters are rational transfer matrices, and the paper’s main claim is that rational wavelet filters bounded at infinity admit state-space realization (Alpay et al., 2011).

The structural parameterization is built from the elementary wavelet filter

$W_{j,t}$1

and elementary paraunitary factors $W_{j,t}$2, yielding

$W_{j,t}$3

The integer $W_{j,t}$4 is the index of the filter, and the McMillan degree is

$W_{j,t}$5

The standard realization is

$W_{j,t}$6

with state equations

$W_{j,t}$7

The paper then gives a constructive recursion for cascade interconnection, so that increasing the index $W_{j,t}$8 by one augments the state dimension by $W_{j,t}$9 (Alpay et al., 2011).

This usage is important because it clarifies that “wavelet state-space model” can refer to exact control-theoretic realization of wavelet filter banks. It is therefore inaccurate to treat the phrase as intrinsically probabilistic. In this branch of the literature, the state-space model is a minimal or constructive dynamical-system realization of a wavelet transfer matrix, with stability, paraunitarity, and cascade structure as the primary concerns.

4. Multiscale Bayesian state-space models and wavelet-space latent fields

A more explicitly stochastic multiscale construction appears in the multiscale Bayesian state-space model (MSBSS) for time-varying, frequency-specific Granger causality. The starting point is a bivariate time-varying VAR written in state-space form,

$\nu_j^2 = \mathrm{Var}(W_{j,t}).$0

with $\nu_j^2 = \mathrm{Var}(W_{j,t}).$1 and $\nu_j^2 = \mathrm{Var}(W_{j,t}).$2 the stacked time-varying VAR coefficients. The multiscale extension replaces the ordinary lagged predictor matrix $\nu_j^2 = \mathrm{Var}(W_{j,t}).$3 by a wavelet-based predictor set $\nu_j^2 = \mathrm{Var}(W_{j,t}).$4 built from a redundant $\nu_j^2 = \mathrm{Var}(W_{j,t}).$5 Haar decomposition. The scale recursion is

$\nu_j^2 = \mathrm{Var}(W_{j,t}).$6

and the new design matrix collects lagged detail and smooth coefficients across scales (Cekic et al., 2017).

The result is not a latent wavelet-state evolution model in the strict sense. Rather, it is a Bayesian dynamic VAR whose observation design has been expanded into multiscale wavelet predictors. The paper’s stated purpose is to derive “dynamic and frequency-specific Granger-causality statistics,” and the estimation procedure combines variational Bayesian approximation with Kalman smoothing, including a multiple-trial extension. A significant Granger effect is assessed by testing whether the relevant coefficient subset excludes zero from its HPD region (Cekic et al., 2017).

A broader, non-Gaussian, non-finite-dimensional usage appears in continuous wavelet cascades for turbulence and multifractal processes. There the hidden multiscale field is

$\nu_j^2 = \mathrm{Var}(W_{j,t}).$7

and the observed process is synthesized as

$\nu_j^2 = \mathrm{Var}(W_{j,t}).$8

Its continuous wavelet transform obeys the stochastic scaling law

$\nu_j^2 = \mathrm{Var}(W_{j,t}).$9

with structure functions

$\theta$0

The paper explicitly notes that this is not a classical linear state-space model, but it does provide a latent multiscale field over scale-position $\theta$1 whose synthesis produces the observed signal (Muzy, 2018). A plausible implication is that wavelet state-space modeling can also denote continuous latent random fields in wavelet coordinates, not only finite-dimensional Markov recursions.

5. Wavelet frames and modern sequence state-space models

In recent long-range sequence modeling, the coupling between wavelets and state-space models has been recast in terms of S4, SaFARi, WaLRUS, and wavelet frames. W4S4 is presented as “WaLRUS for S4,” a class of SSMs constructed from redundant wavelet frames. Its role is not to replace S4 end-to-end but to supply a new SSM core or initialization scheme. The paper states the standard continuous-time and discrete-time equations

$\theta$2

$\theta$3

and then adopts the time-invariant translated form

$\theta$4

A central theorem states that only the first $\theta$5 eigenvalues of the diagonalized WaLRUS state matrix are greater than $\theta$6, the rest are exactly $\theta$7, and only the first $\theta$8 state components contribute to reconstruction. The practical recipe is: diagonalize $\theta$9, keep only the top $$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$0 modes, compute the reduced kernel, and absorb the relevant eigenvector block into $$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$1 (Babaei et al., 9 Jun 2025).

The empirical claim is that wavelet-based state dynamics improve long-horizon retention relative to HiPPO-based initialization. On a minimal delay reconstruction task with fixed delay $$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$2, WaLRUS reports MSE $$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$3 versus HiPPO MSE $$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$4, summarized as a “0.725 smaller log-MSE” and roughly “$$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$5” better. In reported downstream results, W4S4 attains 88.55 (0.23) on LRA Text with 215K parameters, and 94.37 (0.23) on Speech Commands autoregressive classification with 260K parameters (Babaei et al., 9 Jun 2025).

WaveSSM extends the same general direction but is framed explicitly as a collection of SSMs constructed over wavelet frames. Starting from the standard SSM

$$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$6

it uses the SaFARi construction for scaled and translated measures. For the scaled case,

$$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$7

and for the translated case,

$$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$8

The discretized wavelet frame $$\boldsymbol{\nu}(\theta) = \left(\nu_1^2(\theta), \ldots, \nu_J^2(\theta)\right).$$9 is “tightened” by

$Y_t = \sum_{k=1}^{K} X_t^{(k)},$0

and the continuous-time transition is estimated by

$Y_t = \sum_{k=1}^{K} X_t^{(k)},$1

Although the induced operator is dense, the paper states that the deployment uses S4’s DPLR parameterization for efficiency and stability (Solozabal et al., 25 Feb 2026).

Empirically, WaveSSM is reported to outperform orthogonal counterparts such as S4 on transient-rich data. On PTB-XL, the best overall AUROC in the provided table is WaveSSM$Y_t = \sum_{k=1}^{K} X_t^{(k)},$2 = 0.942; on Speech Commands, the best in-distribution result is WaveSSM$Y_t = \sum_{k=1}^{K} X_t^{(k)},$3=$Y_t = \sum_{k=1}^{K} X_t^{(k)},$4. At the same time, the paper reports that all WaveSSM variants fail on PathX (“55”), and under test-time frequency shift to 8 kHz the best result remains S4$Y_t = \sum_{k=1}^{K} X_t^{(k)},$5 = $Y_t = \sum_{k=1}^{K} X_t^{(k)},$6, above the reported WaveSSM variants (Solozabal et al., 25 Feb 2026). This is a useful corrective to any claim that waveletized SSMs dominate uniformly across tasks.

6. Wavelet-domain vision state-space models

In vision restoration and fusion, wavelet state-space models typically use DWT/IWT or wavelet feature modulation to partition low- and high-frequency information, then place Mamba/VSSM computation asymmetrically across those components.

Wave-Mamba for ultra-high-definition low-light image enhancement is an explicit wavelet-domain, frequency-decoupled vision SSM. The core wavelet identities are

$Y_t = \sum_{k=1}^{K} X_t^{(k)},$7

and the architectural claim is that most content information exists in the low-frequency component, while the high-frequency component exerts a minimal influence on low-light enhancement outcomes. Accordingly, the model applies a Low-Frequency State Space Block (LFSSBlock) to low-frequency sub-bands and a lighter High-Frequency Enhance Block (HFEBlock) to high-frequency sub-bands. The LFSSBlock is

$Y_t = \sum_{k=1}^{K} X_t^{(k)},$8

with VSSM

$Y_t = \sum_{k=1}^{K} X_t^{(k)},$9

The paper reports Average PSNR: 32.39, Average SSIM: 0.952, Average LPIPS: 0.102, with Params: 1.258M on UHD-LOL4K / UHD-LL average (Zou et al., 2024).

HSRMamba for single hyperspectral image super-resolution couples wavelet decomposition with a stripe-based scanning VSSM. DWT decomposes intermediate features into a low-frequency branch

$$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$0

and a high-frequency branch

$$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$1

These are processed by LFSE, HFSE, and a decoder-side HLFD, all built around an improved stripe-scanning VSSM intended to reduce artifacts from global unidirectional scanning. On PaviaU $$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$2, the paper reports PSNR $$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$3, SSIM $$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$4, SAM $$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$5, ERGAS $$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$6, with $$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$7 G FLOPs and $$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$8 M parameters (Li et al., 16 May 2025).

IRSRMamba adopts a looser coupling: the wavelet transform appears in the shallow feature extraction stage rather than inside the deep state dynamics. Its modulation path is

$$\nu_j^2(Y) = \sum_{k=1}^{K} \nu_j^2\!\left(X^{(k)}\right).$$9

$$\hat{\theta} = \arg\min_{\theta \in \Theta} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right)^\top \boldsymbol{\Omega} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right).$$0

after which the concatenated shallow features are processed by RSSGs, RSSBs, and VSSM blocks. In an ablation on $$\hat{\theta} = \arg\min_{\theta \in \Theta} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right)^\top \boldsymbol{\Omega} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right).$$1 super-resolution, the cumulative gain from “Pure MambaIR” to “+ wavelet modulation” is reported as +0.67 dB on result-A, +0.72 dB on result-C, and +0.52 dB on CVC10 (Huang et al., 2024).

For infrared and visible image fusion, the abstract of Wavelet-Mamba states that the proposed Wavelet-SSM module “incorporates wavelet-based frequency domain feature extraction and global information extraction through SSM, thereby effectively capturing both global and local features,” together with a “cross-modal feature attention modulation” module (Zhang et al., 24 Mar 2025). Because the supplied record contains only abstract-level content, this usage establishes the architectural intention but not the full mathematical specification.

7. Interpretation, limitations, and recurring misconceptions

Several boundary conditions recur across the literature. First, wavelet state-space modeling is not a single canonical equation. The state-space part may be a latent linear Gaussian model, a transfer-matrix realization, a dynamic VAR, a wavelet-frame SSM core, or a VSSM/Mamba block. The wavelet part may be wavelet variance, $$\hat{\theta} = \arg\min_{\theta \in \Theta} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right)^\top \boldsymbol{\Omega} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right).$$2 predictors, rational wavelet filters, redundant wavelet frames, or DWT/IWT sub-bands (Balamuta et al., 2016, Cekic et al., 2017, Alpay et al., 2011, Zou et al., 2024).

Second, “frequency-specific” does not always mean classical spectral factorization. In the MSBSS framework, the dynamic Granger statistic is extracted from scale-specific wavelet predictors, so the resulting quantity is a scale- or frequency-band-specific predictive influence rather than a Fourier transfer-function Granger decomposition (Cekic et al., 2017).

Third, wavelet downsampling and state-space efficiency are often linked through invertibility or reduced sequence length, but the exact mechanism varies. In Wave-Mamba, the formal claim of information-preserving downsampling is tied to the pair

$$\hat{\theta} = \arg\min_{\theta \in \Theta} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right)^\top \boldsymbol{\Omega} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right).$$3

A plausible implication is that preservation holds for the full collection of sub-bands, not for the approximation band $$\hat{\theta} = \arg\min_{\theta \in \Theta} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right)^\top \boldsymbol{\Omega} \left(\hat{\boldsymbol{\nu}} - \boldsymbol{\nu}(\theta)\right).$$4 in isolation (Zou et al., 2024).

Fourth, positive results are domain- and task-dependent. WaveSSM is strongest on transient-rich signals such as ECG and raw audio, but the same paper reports weaker behavior on PathX and under one particular frequency-shift test (Solozabal et al., 25 Feb 2026). Similarly, MSBSS can be weaker than the ordinary Bayesian state-space model for weaker causal effects with one trial (Cekic et al., 2017). The literature therefore supports a targeted, inductive-bias view: wavelets are most valuable when localization, multiscale separation, or frequency-domain disentanglement is central to the problem.

Taken together, the literature defines a wavelet state-space model not as a single architecture but as a research program. Its common thesis is that multiscale wavelet structure can regularize, localize, or compress the signal representation on which state-space dynamics operate. In classical time-series inference this yields interpretable latent-component estimation; in control theory it yields constructive realizations of wavelet filters; in neuroscience it yields time-varying, frequency-specific directed dependence; and in modern deep learning it yields alternative state dynamics and frequency-partitioned restoration pipelines for long-range sequence and image modeling (Guerrier et al., 2015, Babaei et al., 9 Jun 2025, Li et al., 16 May 2025).