Deep Complex Networks for Phase Aware Learning

This direction is merged into another direction: complex-valued neural networks. Its last version stays readable and citable here.

Version 1 of 110 papers20 conceptsfrozen 2026-10-04

Most deep learning architectures operate exclusively in the domain of real numbers. While this is sufficient for many tasks, it ignores the mathematical elegance and representational power of complex numbers. In fields like audio processing, radar, and optics, information is naturally encoded as a combination of magnitude and phase.

Chiheb Trabelsi and colleagues in Deep Complex Networks argue that by using complex-valued arithmetic, neural networks can more effectively learn these phase-dependent relationships. This shift isn't just about changing the data type; it requires a fundamental redesign of convolutions, normalization, and initialization to maintain mathematical consistency.

The Geometry of Complex Convolutions

In a standard network, a convolution is a simple dot product. In a complex-valued network, we treat the weights $W = A + iB$ and the inputs $z = x + iy$ as complex entities. The resulting operation follows the rules of complex multiplication:

$$W * z = (A * x - B * y) + i(B * x + A * y)$$

This structure is the "engine" of the network. It forces the real and imaginary parts to interact in a specific, coupled way. Unlike a real-valued network with twice as many filters, a complex convolution constrains the relationship between these components, allowing the network to learn rotations and phase shifts directly. Siwen Jiang demonstrated in Utilizing Complex-valued Network for Learning to Compare Image Patches that this allows models to better capture structural information like edges and orientations in images.

Why Standard Normalization Fails

The most significant hurdle in training deep complex networks is normalization. In real-valued networks, Batch Normalization scales and shifts activations to have zero mean and unit variance. If you apply this "naively" to complex numbers—normalizing the real and imaginary parts independently—you encounter a geometric problem.

Independent normalization ignores the correlation between the real and imaginary components. This can cause the distribution of activations to "collapse" into an elongated ellipse in the complex plane, leading to numerical instability and vanishing gradients.

To solve this, Chiheb Trabelsi introduced Complex Batch Normalization. Instead of simple scaling, it performs a 2D whitening transformation. It calculates a $2 \times 2$ covariance matrix to account for the joint distribution of the real and imaginary parts:

  1. Center the complex data by subtracting the mean.
  2. Whiten the data using the inverse square root of the covariance matrix.

This ensures the distribution remains "circular" in the complex plane. This circularity is the key to training stability. As noted by Dong Li in CompEvent, this specialized whitening is essential for managing the statistical dependencies that standard normalization layers miss.

Preserving Phase in Audio and Beyond

The intuition behind this research direction is most visible in speech enhancement. Audio signals are typically transformed into the frequency domain using a Short-Time Fourier Transform (STFT), resulting in complex-valued spectrograms.

Traditional models often discard the phase and only process the magnitude. However, Hyeong-Seok Choi showed with the Deep Complex U-Net that keeping the phase allows for much cleaner speech reconstruction, especially in noisy environments. By using complex-valued operations, the network can perform "phase-aware" masking, essentially rotating and scaling the signal in the complex plane to cancel out noise.

Building on this, Yanxin Hu developed DCCRN, which combines complex convolutions with recurrent layers. The success of these models—DCCRN won the Interspeech 2020 DNS Challenge—proves that complex-valued architectures aren't just a mathematical curiosity; they are a superior way to handle signals where the relationship between "what" (magnitude) and "when" (phase) is inseparable.

Efficiency and Implementation

While full whitening is mathematically ideal, it can be computationally expensive. Siwen Jiang proposed a "decoupled" complex batch normalization as a faster alternative, showing that for some vision tasks, independent normalization is a sufficient approximation if the network is shallow enough. However, for deep architectures, the consensus across these papers remains: to unlock the full potential of complex-valued learning, one must respect the joint geometry of the real and imaginary components.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_deep_complex_networks_for_phase_aware_le_a00ac546,
  title = {Deep Complex Networks for Phase Aware Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-complex-networks-for-phase-aware-learning/rdr_a00ac54641f74a8da80a6d2d58c8a4be},
  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_deep_complex_networks_for_phase_aware_le_a00ac546_v1,
  title = {Deep Complex Networks for Phase Aware Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-complex-networks-for-phase-aware-learning/rdr_a00ac54641f74a8da80a6d2d58c8a4be/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}