Traditional signal processing often forces a compromise: analyze a signal in the time domain, the frequency domain, or via fixed windows in a joint time-frequency domain (such as standard wavelets). However, real-world signals—ranging from non-stationary physical phenomena to multi-channel color images—frequently contain features that are poorly aligned with these rigid axes.
A powerful research direction addresses this limitation by merging two mathematical frameworks: fractional-order transforms and quaternion algebra. By generalizing traditional transforms to fractional orders, researchers can rotate the time-frequency plane to any arbitrary angle, capturing transient features along a continuous spectrum. Simultaneously, by lifting these transforms into the quaternion domain, they can process multi-dimensional, coupled data channels holistically rather than treating them as isolated streams.
The Fractional Dimension: Rotating Time and Frequency
To understand why fractional domains are useful, imagine a standard two-dimensional plot where the horizontal axis represents time and the vertical axis represents frequency. A standard Fourier transform projects a signal entirely onto the frequency axis, losing all localized time information.
A fractional transform introduces a continuous parameter, often interpreted as a rotation angle $\alpha$ of this time-frequency coordinate system. When $\alpha = 0$, the representation matches the pure time domain; when $\alpha = \pi/2$, it becomes the traditional frequency domain. Any fractional value in between creates a hybrid domain that interpolates between time and frequency. This allows the transform to perfectly align with "chirps" (signals whose frequency changes over time) and other non-stationary phenomena without smearing the signal's energy.
As explored by Xiang Qiu et al. in Fractional Wavelet-Based Generative Scattering Networks, embedding a fractional order parameter into wavelet scattering transforms allows deep learning architectures to shift along this time-frequency continuum. This flexibility enables the network to capture highly diverse feature sets that fixed-order transforms miss, which the authors demonstrate helps stabilize the training of generative models by providing a more expressive and continuous representation space.
The Quaternionic Dimension: Preserving Multi-Channel Geometry
While fractional transforms solve the problem of rigid time-frequency alignment, multi-dimensional signals (such as color images or vector fields) present a second challenge: cross-channel correlation. If you process an RGB image by applying a standard transform to the Red, Green, and Blue channels independently, you lose the structural geometry of how those colors interact.
Quaternion algebra solves this by extending real numbers into a four-dimensional hypercomplex space consisting of one real part and three imaginary parts ($i, j, k$). By mapping a multi-channel signal directly onto these imaginary components, the entire vector can be manipulated as a single quaternionic entity.
In Quaternionic shearlet transform, Firdous A. Shah demonstrated the power of this approach by extending shearlets—which are already highly effective at detecting directional geometric features like edges—into the quaternion domain. The author established that analyzing real-valued 2D signals within a quaternion-valued frequency framework preserves the localized geometric detection capabilities of traditional shearlets while providing a much richer, mathematically unified representation of multi-dimensional structures.
Unifying the Framework: The Quaternion Fractional Fourier Transform
The core of this research direction lies in the synthesis of these two concepts: the Quaternion Fractional Fourier Transform (QFRFT). By applying a fractional rotation to a quaternionic space, researchers gain a tool capable of analyzing multi-channel, non-stationary signals simultaneously.
Muhammad Adnan Samad et al. detailed this synthesis in Quaternion Fractional Fourier Transform: Bridging Signal Processing and Probability Theory. By extending the traditional Fourier transform to incorporate both quaternion algebra and fractional orders, the QFRFT allows for signal analysis in a rotated time-frequency plane tailored specifically for multi-dimensional stochastic processes. This mathematical bridge provides the exact flexibility needed to track complex statistical variations across multiple coupled dimensions over time.
Sampling and Reconstruction in Mixed Domains
For these theoretical advantages to be useful in practical applications like denoising or data compression, there must be a reliable way to sample and perfectly reconstruct signals within this rotated hypercomplex domain.
Because the QFRFT mixes time and frequency information across quaternionic channels, standard Nyquist-Shannon sampling theorems do not apply. To resolve this, Muhammad Adnan Samad et al. developed a framework outlined in Generalized Sampling Theory in the Quaternion Domain: A Fractional Fourier Approach.
The authors introduced a Generalized Sampling Expansion (GSE) specifically designed for signals that are bandlimited within the QFRFT domain. The key insight behind why this reconstruction works is the integration of signal derivatives alongside specialized quaternionic Fourier filters. Because a fractional transform spreads a signal's localized information across both time and frequency, measuring just the discrete values of the signal is insufficient. By capturing the signal's derivatives (its local rate of change) within the rotated plane, the reconstruction algorithm gains the necessary geometric context to accurately restore the original quaternion-valued signal, paving the way for robust multi-dimensional signal processing applications.
Key Foundational Work
Quaternionic shearlet transform (2018) — contributed 4 concept elements to this direction.
Fractional Wavelet-Based Generative Scattering Networks (2021) — contributed 1 concept elements to this direction.