Quaternion Discrete Fourier Transform

• Quaternion Discrete Fourier Transform is a hypercomplex extension of the DFT that handles quaternion-valued signals and their non-commutative properties for multidimensional data analysis.
• It features left-sided, right-sided, and two-sided formulations alongside efficient computational strategies, such as fast algorithm implementations using complex FFT decompositions and matrix-exponential frameworks.
• QDFT is applied in color image enhancement, spectral filtering, and neural network optimization, leveraging its ability to fuse multichannel information and preserve cross-channel correlations.

The Quaternion Discrete Fourier Transform (QDFT) generalizes the classical Discrete Fourier Transform to quaternion-valued signals, accommodating the algebraic structure and non-commutativity of quaternions. The QDFT framework is employed for the spectral representation, convolution, and filtering of quaternion signals, with prominent applications in multidimensional signal processing, hypercomplex neural networks, and color image analysis. Its definition and properties are intrinsically linked to the choice of pure unit quaternion axes, leading to a family of related transforms and necessitating specialized treatment of eigenstructure and computational strategies.

1. Algebraic Foundations and Transform Definitions

Let $\mathbb{H}$ denote the division algebra of quaternions, with a generic quaternion $q = a + b i + c j + d k$, where $a,b,c,d \in \mathbb{R}$ and $i^2 = j^2 = k^2 = -1$ with $i j = k = - j i$ and similar multiplication rules. The non-commutativity of $\mathbb H$ fundamentally distinguishes any discrete Fourier transform in this setting.

1D QDFT

For a sequence $x[n] \in \mathbb{H}$ $(n = 0,\dots,N-1)$ and a pure unit quaternion $\mu$ $(\mu^2 = -1,\, \mathrm{Re}(\mu)=0)$, the main variants are:

• Left-sided QDFT: $q = a + b i + c j + d k$0
• Right-sided QDFT: $q = a + b i + c j + d k$1

Both are invertible, with explicit inverse formulas using $q = a + b i + c j + d k$2. The direction of multiplication reflects non-commutativity and influences eigenstructure and convolution results (Sfikas et al., 2023).

2D Two-Sided QDFT

For $q = a + b i + c j + d k$3 on an $q = a + b i + c j + d k$4 grid, the two-sided QDFT is

$q = a + b i + c j + d k$5

with $q = a + b i + c j + d k$6 orthogonal imaginary units. The inverse uses sign reversal in exponents (Dar, 2024, Grigoryan et al., 2018, Grigoryan et al., 2017).

2. Theoretical Properties and Matrix Structures

Orthogonality and Unitarity

The QDFT matrix $q = a + b i + c j + d k$7 defined by $q = a + b i + c j + d k$8 is unitary: $q = a + b i + c j + d k$9, ensuring energy preservation (Parseval/Plancherel theorem). Any choice of $a,b,c,d \in \mathbb{R}$0 yields a valid transform; all such axes form the 2-sphere $a,b,c,d \in \mathbb{R}$1, and different QDFTs are connected by similarity transformations: $a,b,c,d \in \mathbb{R}$2 with $a,b,c,d \in \mathbb{R}$3 for $a,b,c,d \in \mathbb{R}$4 a unit quaternion rotating $a,b,c,d \in \mathbb{R}$5 to $a,b,c,d \in \mathbb{R}$6 (Sfikas et al., 2023).

Circulant Matrices and Convolution

The quaternionic circulant $a,b,c,d \in \mathbb{R}$7 for kernel $a,b,c,d \in \mathbb{R}$8 is diagonalized by $a,b,c,d \in \mathbb{R}$9, and its left-eigenvalues are the right-sided QDFT of $i^2 = j^2 = k^2 = -1$0. This gives:

$i^2 = j^2 = k^2 = -1$1

where $i^2 = j^2 = k^2 = -1$2 (Sfikas et al., 2023). Convolution in the time/spatial domain corresponds to multiplication in the QDFT space, requiring consistent use of left/right-side QDFTs.

Matrix-Exponential Formalism

The QDFT is equivalently realized via a matrix-exponential framework, replacing the imaginary unit by a $i^2 = j^2 = k^2 = -1$3 real matrix $i^2 = j^2 = k^2 = -1$4 with $i^2 = j^2 = k^2 = -1$5, and Euler's formula by $i^2 = j^2 = k^2 = -1$6. This approach formalizes the unity of complex, quaternionic, and Clifford-algebra DFTs, and allows implementation via standard matrix algebra libraries without explicit quaternion code (Sangwine et al., 2010).

3. Algorithmic Implementations and Hardware Efficiency

Fast Algorithms

A quaternion-valued array can be decomposed into two complex arrays, enabling the QDFT or 2D QDFT to be computed via a pair of complex FFTs:

• Let $i^2 = j^2 = k^2 = -1$7
• Compute $i^2 = j^2 = k^2 = -1$8 and $i^2 = j^2 = k^2 = -1$9
• Combine as $i j = k = - j i$0

The resulting computational complexity is $i j = k = - j i$1 (Dar, 2024). In hardware, QDFT pipeline architectures utilize specialized multipliers exploiting i-quaternion and j-quaternion structures, reducing real multiplier count per product from 16 to 6 for left/right-sided multiplication, and to 9 for two-sided multiplication, with significant area, throughput, and power advantages in FPGA and ASIC deployments (Cariow et al., 2017).

Multiplication Kernel	Multipliers	2-Input Adders	4-Input Adders
Conventional q·t	16	12	0
Optimized sq, qt	6	6	0
Optimized sqt	9	6	4

Two-Side Transform Flow

A 2D QDFT on images is typically performed by sequentially applying row-wise left multiplications (with i-quaternion kernels), followed by column-wise right multiplications (with j-quaternion kernels). The modular arithmetic structure and pipeline tiling support high parallelism, and systolic architectures enable efficient butterfly computations (Cariow et al., 2017).

4. Applications in Multichannel and Color Image Processing

A principal application of QDFT is the joint spectral analysis of multichannel signals, particularly in color image processing. Treating an RGB pixel as a pure quaternion $i j = k = - j i$2, the QDFT processes all channels together, preserving cross-channel correlations that are lost in channel-by-channel DFTs (Grigoryan et al., 2018, Grigoryan et al., 2017). This property supports advanced tasks such as:

• Color image enhancement: QDFT-based α-rooting methods apply nonlinear magnitude scaling in the quaternion-frequency domain, followed by histogram equalization. These approaches consistently yield higher color enhancement measure estimation (CEME) values and more natural color balance compared to channel-wise DFT methods.
• Spectral filtering and watermarking: Custom quaternionic frequency masks can be designed and applied, leveraging the non-commutativity for orientation-sensitive operations.
• Deep neural networks: QDFT provides a direct spectral-norm bound for quaternionic convolutional layers, essential for controlling Lipschitz constants, robustness, and generalization in deep architectures (Sfikas et al., 2023).

Enhancement Method	Output Metric (CEME, dB)	Feature
Channel-wise 2D DFT	23.5 – 39.5	Looser color coherence
QDFT α-rooting + HE	29.5 – 49.0	Higher contrast, preserved color correlations

5. Structural Generalizations and Extensions

The QDFT is a principal case within a broader family of quaternionic linear transforms. For instance, the discrete quaternion quadratic-phase Fourier transform (DQQPFT) reduces to the 2D QDFT when quadratic parameters vanish. This provides QDFT with a foundation in advanced time-frequency analysis and signal representation (Dar, 2024).

Every QDFT is indexed by the choice of pure imaginary axes; axes transformations correspond to unit quaternion similarities, implying that many classical theorems (energy conservation, invertibility, convolution diagonalization) extend with nontrivial adjustments for non-commutativity.

Further, the matrix-exponential formalism provides a universal perspective on hypercomplex DFTs—including those based in Clifford algebras—streamlining the derivation and implementation across algebras (Sangwine et al., 2010).

6. Open Issues and Research Directions

The non-commutative structure of the quaternion algebra induces complexities in defining and analyzing eigenstructures, convolution theorems, and spectral regularizations, which remain subjects of active theoretical development. Matrix-based approaches facilitate software portability, but custom quaternion-FFT implementations remain superior for computational efficiency in large-scale systems.

The QDFT's capacity for channel fusion, orientation-aware filtering, and spectral regularization positions it as a core technology in modern hypercomplex signal and image processing, with ongoing research focusing on novel applications (e.g., spectral-constrained neural networks), hardware implementations, and theoretical unification with other hypercomplex transforms (Sfikas et al., 2023, Cariow et al., 2017, Dar, 2024, Sangwine et al., 2010).