Holomorphic Phasor Activations

• Holomorphic phasor activations are complex-valued functions leveraging analytic inversion (e.g., 1/z) to embed rational kernels and ensure closed-form invertibility.
• They are implemented in architectures like CauchyNet and compleX-PINN, achieving superior performance on near-singular, oscillatory, and data-scarce learning tasks through efficient computation.
• These activations offer practical benefits for scientific machine learning, time-series forecasting, and PDE-solving by maintaining numerical stability with regularized inversion operations.

Holomorphic phasor activations are a class of neural network activation functions defined over complex domains that leverage the analytic properties of holomorphic functions and the structural expressivity of phasors. Emerging from the direct application of Cauchy's integral formula, these activations—most notably those implementing inversion maps such as $z \mapsto 1/z$—embed a complex-analytic kernel into each neuron, providing a robust inductive bias for function approximation, time series forecasting, and scientific machine learning. Architectures such as CauchyNet and compleX-PINN systematically exploit holomorphic phasor activations to outperform conventional networks in resource-constrained, data-scarce, or near-singular regimes (Zhang et al., 11 Oct 2025, Si et al., 7 Feb 2025).

1. Holomorphic Inversion Activation and Mathematical Motivation

Holomorphic phasor activations are defined on the punctured complex domain

$$\mathbb{C}^N_{\ast} = \{ z = (z_1,\ldots,z_N)\in\mathbb{C}^N : z_i \ne 0\ \forall i \}$$

with the canonical form

$$\sigma(z) = \mathcal{X}(z) = \prod_{i=1}^N z_i^{-1}.$$

For $N=1$, this activation reduces to the scalar inversion $\sigma(z) = 1/z$, which is holomorphic and globally invertible on $\mathbb{C} \setminus \{0\}$ (Zhang et al., 11 Oct 2025). The theoretical foundation is the Cauchy integral formula:

$$f(z)=\frac{1}{2\pi i}\oint_C \frac{f(\xi)}{\xi-z}\,d\xi$$

where the kernel $(\xi-z)^{-1}$ supplies the core inversion operation. Embedding this kernel as an activation function not only guarantees holomorphicity (enabling the use of complex differentiation and the inverse function theorem) but also yields closed-form inverses critical for invertible learning architectures.

2. Invertibility and Practical Computation

The invertibility of the holomorphic inversion activation is rigorously justified:

• For $F(z) = 1/z$, the derivative $F'(z) = -1/z^2 \neq 0$ for $z \neq 0$, so $F$ is biholomorphic on its domain.
• In practice, a shift parameter $\epsilon > 0$ is added for numerical stability, yielding $\sigma(z) = 1/(z + \epsilon)$.
• The inverse map, for scalar input, is $z = 1/y - \epsilon$. For vectors, inversion is performed per coordinate: if $h_{k,i} = (H_{k,i}+\epsilon)^{-1}$, then $H_{k,i} = 1/h_{k,i} - \epsilon$ (Zhang et al., 11 Oct 2025).

Algorithmically, inversion is $O(1)$ per neuron with no need for iterative approximation:

def INV_ACTIVATION(y, epsilon): return (1.0 / y) - epsilon

Care must be taken to keep $|y|$ away from zero, regularizing as needed for numerical precision.

3. Parameterization and Complexity

Holomorphic phasor activations enable neural architectures that are both compact and theoretically expressive:

• Parameter footprint (CauchyNet): For $h$ hidden neurons and input dimension $m$, the total parameter count is $2h(m+1)$ real numbers (as each complex parameter counts as two real numbers).
• Operational complexity: Both forward and backward passes require $O(hm)$ complex arithmetic operations per layer; each complex multiply equates to four real multiplications (Zhang et al., 11 Oct 2025).
• Comparison with standard networks: While parameter counts may be similar to real-valued MLPs, CauchyNet often achieves comparable or superior expressivity with far fewer neurons, particularly on near-singular or highly oscillatory functions.

4. Learnable Holomorphic Rational Activations in Physical Modeling

Beyond fixed inversion maps, the rational form of the Cauchy kernel can be parameterized and learned:

$$\Phi(z;\mu_1,\,\mu_2,\,d) = \frac{\mu_1 z}{z^2 + d^2} + \frac{\mu_2}{z^2 + d^2}$$

where $\mu_1, \mu_2, d \in \mathbb{R}$ are trainable per neuron (Si et al., 7 Feb 2025). This discrete "Cauchy transform" allows the network to adapt the residue and pole locations for optimal expressivity. In compleX-PINN, a single Cauchy-activated hidden layer (with potentially many neurons) typically suffices to solve stiff or high-frequency PDEs, aided by standard PINN loss functionality and gradient-based optimization.

5. Empirical Performance and Inductive Bias

Holomorphic phasor activations confer significant empirical advantages:

• Function approximation: On sharp-peaked or rational-spike functions, CauchyNet tracks the spike precisely where ReLU MLPs substantially underfit. Training/validation loss decreases faster and to lower minima (Zhang et al., 11 Oct 2025).
• Missing-data imputation/gap-filling: CauchyNet generates faithful reconstructions with low MAE ($\approx 1.5$ vs. $>3.2$ for other baselines) (Zhang et al., 11 Oct 2025).
• Time-series and forecasting: On M4 trend subseries, CauchyNet achieves a median MAE $\approx 2.3$ using 256 parameters versus $>10$ MAE and $>130{,}000$ parameters for Transformer/N-BEATS (Zhang et al., 11 Oct 2025).
• Stiff/high-dimensional PDEs: compleX-PINN delivers one or two orders of magnitude improvement in relative $L^2$ error over PINN+Tanh and advanced architectures (RBA, gPINN) on wave, reaction, diffusion, and convection problems, requiring orders of magnitude fewer layers or neurons (Si et al., 7 Feb 2025).

The holomorphic, invertible structure imparts a bias towards modeling smooth oscillations and rational singularities, and the closed-form inverse enables use in invertible neural architectures and flow-based models.

6. Integration in Neural Architectures and Training

The practical integration of holomorphic phasor activations includes:

• Network embedding: Inputs are mapped by an affine transformation, then passed through the holomorphic inversion or learned rational activation, followed by a linear combination to produce the output.
• Layer composition: Single-layer implementations suffice for many challenging tasks, but multi-layer composition provides further capacity (Si et al., 7 Feb 2025).
• Optimization: All network weights, biases, and rational activation parameters are trained jointly via backpropagation and standard optimizers such as Adam or L-BFGS, with typical learning rates $10^{-4}$–$10^{-5}$ and $10^4$–$10^5$ iterations (Si et al., 7 Feb 2025).

7. Context, Limitations, and Outlook

Holomorphic phasor activations, as instantiated in CauchyNet, compleX-PINN, and related architectures, constitute a theoretically grounded advance in complex-valued neural computation. Their computational compactness, analytic invertibility, and empirically validated inductive biases on function approximation and scientific learning tasks underscore their applicability in resource-limited and data-scarce scenarios. A plausible implication is the further potential integration with invertible and flow-based deep models, as well as the extension to learnable parameterized rational activation forms for greater flexibility. Current limitations include the need for careful numerical regularization and domain restriction to avoid zeros in the complex input. The ongoing comparison with more conventional activations continues to clarify the specific domains of superiority for holomorphic phasor-based approaches (Zhang et al., 11 Oct 2025, Si et al., 7 Feb 2025).