The standard Fourier Transform is a cornerstone of signal processing, but it operates under a rigid geometric constraint: it projects a signal onto axes that are strictly orthogonal to time. Geometrically, moving from the time domain to the frequency domain represents a fixed 90-degree rotation in the time-frequency plane.
For stationary signals, like a continuous sine wave, this 90-degree projection works perfectly, collapsing the signal's energy into a sharp, localized peak. However, many real-world signals are non-stationary. Consider a linear frequency modulated (LFM) chirp, whose frequency changes linearly over time. In the time domain, its energy is spread out; in the frequency domain, it is similarly smeared across a wide band. Because its energy is distributed across multiple frequencies simultaneously, it behaves as a wideband signal.
The core insight of fractional domain signal processing is that we do not have to limit ourselves to 90-degree rotations. By rotating the time-frequency plane by an arbitrary fractional angle, we can align our perspective with the internal dynamics of the signal, effectively transforming complex wideband signals into simple, stationary, narrowband-like tones.
The Geometry of Fractional Rotation
The Fractional Fourier Transform (FrFT) generalizes the standard Fourier Transform by introducing a continuous parameter—the transform order or rotation angle. As conceptualized by Wang et al. (2016), this creates a Unified Time-Frequency Domain (UTFD).
To understand why this works, imagine an LFM chirp plotted on a two-dimensional time-frequency graph. The signal appears as a diagonal line whose slope is determined by its chirp rate. If we project this line directly onto the time axis or the frequency axis, the projection is wide and diffuse.
Frequency
^
| / <-- LFM Chirp (Wideband in both Time and Frequency)
| /
| /
+---------> Time
If we instead rotate the coordinate axes by an angle $\alpha$ that perfectly matches the slope of the chirp, the signal becomes perpendicular to our new fractional axis. Projecting the chirp onto this optimal fractional domain collapses its entire energy profile into an intensely sharp, localized peak.
Fractional Axis
^
| | <-- Aligned Chirp (Becomes a Narrowband Impulse)
| |
+---------> Rotated Axis
This mathematical sharpening is highly effective for isolating signals from background noise. For example, Shi et al. (2023) utilized this rotation to denoise chirp spread spectrum signals in aerial acoustic communications, applying a tailored 2-D mask filter in the FrFT domain to isolate the compressed chirp pattern while discarding multi-path interference and noise. Similarly, Aziz et al. (2021) integrated fractional rotation with adaptive windowing to sharpen transient features in electrocardiogram (ECG) data, making subtle heartbeat anomalies highly distinguishable without sacrificing temporal localization.
Freezing Time-Varying Dynamics
Transforming a wideband signal into a narrowband state does more than just denoise; it fundamentally simplifies downstream algorithmic complexity.
In array signal processing—such as sonar or radar tracking—algorithms designed to estimate the Direction of Arrival (DOA) typically rely on a "narrowband assumption." This assumption dictates that the spatial signature (the steering vector) of a incoming wave remains constant across the sensor array during the observation window. When a signal is wideband and time-varying, its steering vector changes continuously, causing traditional high-resolution algorithms like MUSIC to break down.
By passing the received signals through the FrFT, we can "dechirp" them before performing spatial analysis. As demonstrated by Liu et al. (2017) and Yu et al. (2015), rotating the signal into an intermediate fractional domain converts time-variant steering vectors into time-invariant ones. This mechanism effectively "freezes" the non-stationary nature of the signal. Because the spatial signatures are stabilized, powerful narrowband localization techniques can be applied directly to wideband sources using only a handful of data snapshots (Liu et al., 2017; Liu et al., 2018).
This "Rotate + Decimate" philosophy also yields significant hardware efficiencies. In chirp radars, pulse compression traditionally requires memory-intensive, frequency-domain convolutions. To bypass this, Yan et al. (2009) implemented a time-domain coordinate rotation algorithm (CORDIC) on an FPGA. By rotating the phase of the incoming complex chirp to cancel its quadratic component, they converted the wideband chirp into a single-frequency tone, eliminating the need to store massive reference waveforms and drastically reducing the computational footprint.
Characterizing Phase and Structure
Because the fractional domain is highly sensitive to the exact rate of a signal's frequency variation, it serves as an exceptional tool for parameter estimation. When an LFM signal is transformed across varying fractional orders, an energy peak appears only at the exact order corresponding to the signal's chirp rate. Liu et al. (2019) leveraged this property to precisely estimate the chirp rates of signals intercepted by Nyquist Folding Receivers, transforming an intricate modulation tracking problem into a straightforward peak-detection task.
This capacity to view signals from multiple perspectives also provides a powerful framework for solving inverse problems, such as phase retrieval and waveform design. For instance, Pinilla et al. (2025) utilized the FrFT to rotate the radar Ambiguity Function in the time-frequency plane, allowing them to reconstruct missing signal phase information from multiple fractional viewpoints.
On a theoretical level, fractional rotations help map out the fundamental limits of signal recovery. Liehr (2022) employed the FrFT to analyze Short-Time Fourier Transform (STFT) measurements along rotated lines, proving foundational discretization barriers and showing how sampling lattices behave when they do not align with standard time or frequency axes.
Generalizing Phase Shifts and Wavelets
The philosophy of fractional transformation extends to other core signal processing operators, notably the Hilbert transform. While a standard Hilbert transform applies a fixed 90-degree phase shift to create a quadrature analytic signal, the Fractional Hilbert Transform (fHT) allows for arbitrary phase shifts on finite-energy signals (Chaudhury, 2009).
This generalized view of phase provides the mathematical foundation for understanding advanced wavelet architectures. Chaudhury et al. (2010) established that the enhanced shift-invariance of the Dual-Tree Complex Wavelet Transform (DT-CWT)—which overcomes the severe shift-sensitivity of traditional discrete wavelet transforms—is directly linked to the translation and dilation invariance properties of the underlying fHT.
Bridging Spatial and Frequency Domains in Deep Learning
Recently, the intuition of fractional domain rotation has been adapted into deep learning architectures to handle complex, non-stationary multi-dimensional data. Standard convolutional neural networks operate locally in the spatial (or time) domain, while standard Fourier neural operators operate globally in the frequency domain.
By treating data as a signal that can be rotated to any intermediate angle between these two extremes, Yu et al. (2023) introduced Multi-order FRactional Fourier Convolutions. This framework allows deep networks to capture features that are neither purely local nor purely global.
This unified spatial-frequency representation is uniquely suited for handling non-stationary distortions, such as spatially varying image blur or rapidly changing speech dynamics:
- Image Deblurring: Traditional deconvolution struggles when a camera shake creates a blur that varies across an image. The authors of F2former (2025) integrated the FrFT into a transformer architecture, exploiting its unified domain to effectively isolate and remove non-uniform, spatially varying blurs.
- Speech Enhancement: In audio processing, fixed-resolution transforms often fail to resolve rapid, non-stationary speech transitions in noisy environments. The MFSE (2025) framework addresses this by adaptively rotating speech representations into an optimal fractional time-frequency domain, providing a more focused spectral representation that separates clean speech from complex background noise.
- Time-Frequency Squeezing: To sharpen highly non-stationary components, Jiang et al. (2022) developed a synchrosqueezing transform featuring a phase transformation tailored to specific signal components, allowing precise separation of signals with rapidly changing instantaneous frequencies.
Ultimately, whether implemented in hardware-efficient radar pipelines, high-resolution sonar arrays, or deep neural networks, the power of fractional domain transformation rests on a simple geometric truth: when a signal's behavior is spread across time and frequency, changing your angle of perspective can collapse a complex wideband problem into a clear, manageable narrowband solution.
Key Foundational Work
Design and FPGA implementation of digital pulse compression for chirp radar based on CORDIC (2009) — contributed 2 concept elements to this direction.
Gabor wavelet analysis and the fractional Hilbert transform (2009) — contributed 2 concept elements to this direction.
On the shiftability of dual-tree complex wavelet transforms (2010) — contributed 2 concept elements to this direction.