Real-Imaginary Decomposition in Neural Networks

44 concepts47 papers1 proposal

Many physical and mathematical domains naturally operate in the complex plane. Whether dealing with the phase and amplitude of wireless communications, the frequency components of a Fourier transform, or the Hermitian matrices of directed graphs, data is frequently complex-valued ($z = x + iy$).

However, modern deep learning hardware and software stacks are fundamentally built for real-valued linear algebra. To bridge this gap without redesigning optimization algorithms and GPU kernels from scratch, researchers rely on real-imaginary decomposition. This core mechanism maps complex-valued operations into equivalent real-valued representations, enabling standard neural networks to natively process complex data while preserving its underlying mathematical structure.

The Geometry of Complex Mapping

To ingest a complex vector into a standard neural network, the data must first be translated into a real-valued format. There are two primary geometric strategies for this mapping: concatenation and multi-channel separation.

Complex Vector: [x1 + iy1, x2 + iy2]

Concatenation:  [x1, x2, y1, y2]  (Double the dimension)
Multi-channel:  Channel 1: [x1, x2]
                Channel 2: [y1, y2]

As demonstrated by Jonas Wacker et al. in Complex-to-Real Sketches for Tensor Products with Applications to the Polynomial Kernel and Stefano Fiorini et al. in SigMaNet: One Laplacian to Rule Them All, concatenating the real and imaginary components into a single vector ensures immediate compatibility with downstream real-valued models and layers.

Alternatively, for spatial or structural data, the components can be treated as independent channels. In the context of physical fields, Yu Sun et al. show in Efficient and accurate inversion of multiple scattering with deep learning that treating the real and imaginary parts as multi-channel inputs allows standard convolutional architectures to simultaneously extract features from both phase and amplitude.

Decomposing Linear Operations

The core elegance of real-imaginary decomposition lies in how it handles matrix multiplication. Consider multiplying a complex weight matrix $W = W_R + iW_I$ by a complex vector $z = x + iy$:

$$Wz = (W_R x - W_I y) + i(W_R y + W_I x)$$

This algebraic expansion can be perfectly mirrored inside a standard real-valued neural network using a structured block matrix:

$$\begin{bmatrix} \text{Re}(Wz) \\ \text{Im}(Wz) \end{bmatrix} = \begin{bmatrix} W_R & -W_I \\ W_I & W_R \end{bmatrix} \begin{bmatrix} x \\ y \end{bmatrix}$$

By structuring the weights this way, the network performs the exact equivalent of complex matrix multiplication using four standard real-valued operations.

The authors of Joint Design of Measurement Matrix and Sparse Support Recovery Method via Deep Auto-encoder and Semi-supervised Complex-valued GAN for Polarimetric SAR Image Classification utilize this exact formulation to embed complex-domain equations directly into real-valued deep learning frameworks. To make this computationally efficient, Yinchuan Li et al. design a dual-path architecture in Unfolded Deep Neural Network (UDNN) for High Mobility Channel Estimation that processes the real and imaginary components in parallel while interleaving information between the paths to maintain mathematical consistency with complex arithmetic.

The Activation Function Dilemma

While linear operations decompose cleanly, non-linear activation functions introduce a fundamental mathematical hurdle known as Liouville's theorem. In complex analysis, Liouville's theorem states that any bounded, fully analytic (complex-differentiable) function must be constant. This means a direct complex equivalent of a bounded activation function like Sigmoid or Tanh cannot exist without encountering singularities or losing differentiability.

To circumvent this, researchers employ split-domain activation processing. Instead of treating the input as a single complex number, the activation function is applied independently to the real and imaginary parts:

$$f(x + iy) = \sigma(x) + i\sigma(y)$$

In On the Appropriateness of Complex-Valued Neural Networks for Speech Enhancement, Lukas Drude et al. demonstrate that this split activation strategy introduces necessary non-linearities while remaining computationally efficient and avoiding complex differentiability constraints.

For more expressive modeling, Simone Scardapane introduces learnable, non-parametric kernel functions applied separately to each component in Complex-valued Neural Networks with Non-parametric Activation Functions. This decoupling allows the network to learn distinct features for each component, which is particularly useful when the cross-domain dependencies between the real and imaginary parts are minimal or asymmetric, as further detailed by Simone Scardapane et al. in Widely Linear Kernels for Complex-Valued Kernel Activation Functions.

Parallel Streams and Latent Fusion

Beyond element-wise splitting, a prominent architectural paradigm is the use of completely isolated parallel streams that fuse late in the network.

Input (Real) ------> Subnetwork A -------\
                                          >----> Joint Latent Space
Input (Imaginary) --> Subnetwork B -------/

In Steinmetz Neural Networks for Complex-Valued Data, Shyam Venkatasubramanian et al. show that routing the real and imaginary components through independent subnetworks before merging them into a joint latent space allows the model to isolate and filter noise specific to each component.

This separation is highly effective in medical imaging. For parallel MRI reconstruction, Qing Wu et al. implement a dual-network architecture in A scan-specific unsupervised method for parallel MRI reconstruction via implicit neural representation to preserve phase and magnitude streams. As noted by Dinggang Shen et al. in Guest Editorial Special Issue on Deep Learning in Medical Imaging, processing magnitude and phase information through separate streams isolates aliasing artifacts because the two components capture fundamentally distinct physical properties of the underlying tissue.

Why Decomposition Works

Decomposition is more than a software engineering trick; it provides distinct theoretical and practical advantages over treating complex numbers as arbitrary two-dimensional real vectors.

  1. Feature Preservation: In signal detection tasks, Hasan Albinsaid et al. demonstrate in Block Deep Neural Network-Based Signal Detector for Generalized Spatial Modulation that explicitly separating real and imaginary components into distinct vectors preserves individual scalar values and phase boundaries more effectively than joint processing methods.
  2. Theoretical Tractability: Proving the stability or convergence of a purely complex network is notoriously difficult due to the lack of bounded analytic functions. By decomposing the system, researchers can map complex networks directly to real-valued counterparts. Xiwei Liu leverages this in Finite time anti-synchronization of complex-valued neural networks with bounded asynchronous time-varying delays to apply classical real-valued stability theories to complex networks. Similarly, Jianrong Chen et al. show in Complex, Temporally Variant SVD via Real ZN Method and 11-Point ZeaD Formula from Theoretics to Experiments that transforming complex time-varying problems into real-valued ones makes them immediately solvable via established dynamical system techniques like Zeroing Neurodynamics.

By formalizing the algebraic and structural relationships between the real and imaginary components, decomposition allows standard neural network architectures to inherit the mathematical rigor of the complex domain. This framework underpins deep learning applications across signal processing, physics-informed modeling, and spectral graph analysis, as seen in works ranging from 5G channel estimation by An Le Ha et al. to directed hypergraph processing by Stefano Fiorini et al. in Let There be Direction in Hypergraph Neural Networks.

Key Foundational Work

Complex domain flexible non-linear function for blind signal separation (2001) — contributed 1 concept elements to this direction.
Low-power Butterfly Structure for DIT Radix-4 FFT Implementation (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_complex_to_real_decomposition_8433f91a,
  title = {complex-to-real decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complex-to-real-decomposition/rdr_8433f91a9a304646a9fe8d4f60558a61},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_complex_to_real_decomposition_8433f91a_v1,
  title = {complex-to-real decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complex-to-real-decomposition/rdr_8433f91a9a304646a9fe8d4f60558a61/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake