Undecimated Wavelet Transform

• Undecimated Wavelet Transform is a shift-invariant, redundant multiscale representation that retains full spatial resolution by omitting downsampling.
• It enhances denoising, scaling estimation, and feature extraction by leveraging all signal positions to reduce variance in spectral assessments.
• The transform can be efficiently implemented via recursive filtering or matrix formulations, ensuring perfect reconstruction and flexibility with arbitrary signal sizes.

The undecimated wavelet transform (UWT), also referred to as the non-decimated wavelet transform (NDWT), stationary wavelet transform, “à trous” transform, or maximal-overlap DWT, is a shift-invariant, highly redundant multiscale representation of discrete signals or images. Unlike the critically sampled discrete wavelet transform (DWT), the UWT omits the downsampling step after filtering, so each decomposition level retains full spatial resolution, resulting in redundancy and translation-invariance. These structural properties enable superior performance in denoising, scaling estimation, feature extraction, and signal analysis tasks where sensitivity to translation and local structure is crucial (Kang et al., 2016, Kang et al., 2022, Kong et al., 2019).

1. Core Principles and Mathematical Formalism

Let $y=(y_0,\ldots,y_{m-1})$ be a discrete 1D signal. The UWT decomposes $y$ by repeated convolutions with dyadically upsampled low-pass and high-pass filters $h$ and $g$, producing at each scale $j$ a set of approximation $c^{(j)}$ and detail $d^{(j)}$ coefficients, both of length $m$. The $j$-th level filters are constructed by inserting $2^{j-1}-1$ zeros between the filter coefficients: $y$0 At each scale, the convolutions proceed without decimation: $y$1 with $y$2. The resulting transform is highly redundant: for a signal of length $y$3 and $y$4 decomposition levels, the coefficient vector has length $y$5 (Kang et al., 2016, Kang et al., 2022).

The transform admits a matrix formulation: for moderate signal sizes, all convolution and shift operations are encapsulated in a matrix $y$6, such that the UWT is a single matrix-vector product $y$7. Inversion uses the transpose and a diagonal weight matrix $y$8, enabling perfect reconstruction independent of signal length or shape.

2. Structural Properties: Redundancy, Shift-Invariance, and Flexibility

A defining property of the UWT is its intrinsic redundancy: each scale outputs coefficients at every input position, yielding a representation robust to translations. This shift-invariance is critical in scenarios where translation sensitivity would introduce artifacts or impair statistical estimation (e.g., in denoising, local spectrum analysis, or feature localization) (Kang et al., 2016, Kang et al., 2022, Kong et al., 2019, Wang et al., 2019).

Unlike the orthogonal DWT, which is restricted to dyadic lengths and produces a hierarchy of subsampled subbands, the UWT has no power-of-two limitation and operates seamlessly on non-dyadic, arbitrary-length signals and, in 2D, rectangular images. Boundary handling is typically by periodic (circular) or symmetric extension, neither of which disrupts shift-invariance (Kang et al., 2022).

3. Algorithmic Realization, Computational Complexity, and Software

Traditional implementations of UWT rely on recursive filtering and explicit upsampling. However, in MATLAB and similar platforms, the explicit matrix approach is advantageous: once $y$9 is constructed, arbitrary signals of matching length can be transformed and reconstructed with a single matrix multiply, leveraging BLAS/LAPACK-optimized kernels (Kang et al., 2016, Kong et al., 2019).

The computational cost is a tradeoff: construction of $h$0 has $h$1 but, in practice, for $h$2, matrix multiplication is faster than repeated 1D or 2D filtering. For repeated transforms with fixed size and filter, $h$3 is reused, further improving efficiency. The main limitation is memory: the storage of $h$4 matrices can exhaust RAM for large $h$5 or deep decompositions. For comparison, classic convolutional NDWT has per-level cost $h$6 via FFT (Kang et al., 2016).

In higher dimensions, 2D UWT is defined via separable filtering or scale-mixing: an image $h$7 is transformed by row and column UWT matrices, e.g., $h$8, with inversion via transposes and weight matrices (Kang et al., 2016, Kong et al., 2019). This supports arbitrary, non-square, non-dyadic matrices.

4. Signal Analysis, Spectral Estimation, and Scaling Assessment

The redundancy of UWT enables improved spectral estimation through averaging over the full set of coefficients at each scale. For self-similar or fractal processes, the wavelet spectrum,

$h$9

is regressed against scale $g$0. The resulting slope estimates scaling exponents or the Hurst parameter with variance significantly lower than for orthogonal DWT-based estimators, due to the large number of non-aliased coefficients per scale. In 2D, these principles extend to images via scale-mixing and anisotropy/correlation ratios over main-diagonal and off-diagonal subbands (Kang et al., 2016, Kang et al., 2022).

In biomedical applications, such as breast cancer mammogram analysis, UWT-based scaling and anisotropy features yield mean diagnostic accuracy in excess of 80%. The variance reduction in scaling parameter estimation has been quantitatively shown: for 2D fractional Brownian fields, mean-squared error of UWT-based Hurst estimation is less than half that of DWT-based approaches (Kang et al., 2022, Kong et al., 2019).

5. Denoising, Feature Extraction, and Machine Learning Integration

UWT is widely used in denoising, exploiting its shift-invariance and redundancy to avoid pseudo-Gibbs artifacts. Universal thresholding is directly applicable: noise variance is estimated from fine-scale detail coefficients and coefficients below a theoretically-motivated threshold are zeroed, followed by perfect reconstruction. The WavmatND package demonstrates this for atomic force microscopy signals and Doppler test cases (Kang et al., 2016).

UWT integration into modern machine learning, especially denoising autoencoders, has been adopted for structured image priors in compressed sensing MRI (Wang et al., 2019). Multi-scale, multi-orientation UWT coefficients—arranged as channel stacks—serve as input to deep denoising networks. This redundancy cross-encodes spatial and scale information, enhancing the expressive power of network-based priors over standard DWT or unstructured image representations.

Similar principles apply in audio signal separation, where undecimated wavelet packet decompositions tuned to psychoacoustic critical bands as perceptual filterbanks, followed by kurtosis maximization, enable enhanced non-Gaussianity for ICA-based blind speech separation (Missaoui et al., 2012).

UWT has also been applied to higher-dimensional and graph-structured data, including spherical 3D cosmological fields where an “à trous” transform in the Spherical Fourier-Bessel domain provides a redundant, isotropic decomposition for denoising large-scale structure data (Lanusse et al., 2011).

6. Quantum Undecimated Wavelet Transform: Theory and Circuits

Quantum generalizations of UWT explicitly exploit its redundancy and translation invariance. Two core quantum circuits have been developed: the epsilon-decimated unitary, encoding all circularly shifted DWTs in superposition, and the Hadamard-test approach, which queries multiscale energy scalograms via diagonal phase operators. In these settings, redundancy, typically a computational burden classically, is harnessed for efficient quantum algorithms supporting denoising, feature extraction via cross-scale statistics, and scalogram computation with resource costs scaling polylogarithmically in signal length (Vidakovic, 25 Dec 2025).

Quantum circuits exploit ancilla-driven completely positive trace preserving maps for coefficientwise shrinkage, and statistical estimation of scaling exponents from measured energy summaries mimics the classical advantage of UWT in spectral estimation, but now extended to quantum signal processing.

7. Summary Table: Key UWT Properties and Differences

Property	UWT/NDWT	Critically Sampled DWT
Downsampling	None (no decimation)	By 2 at each level
Redundancy	High: $g$1 coefficients for length $g$2, $g$3 levels	Minimal: sum of $g$4 over all scales $g$5
Shift-invariance	Yes	No (sensitive to input translation)
Input size constraint	None (arbitrary length/shape)	Dyadic power of two preferred
Spectrum estimation	Lower-variance, all positions/subbands	Higher-variance, fewer coefficients
Denoising performance	Robust, artifact-free (pseudo-Gibbs suppressed)	Sensitive to translation, artifacts possible

The undecimated wavelet transform forms the backbone of a wide spectrum of modern signal processing, statistical, machine learning, and quantum algorithms, offering translation-invariant, multiscale, and redundant representations with significant advantages in denoising, spectral analysis, scaling estimation, and feature extraction (Kang et al., 2016, Kang et al., 2022, Kong et al., 2019, Wang et al., 2019, Vidakovic, 25 Dec 2025, Missaoui et al., 2012, Lanusse et al., 2011).