Mitigating Exponential Mixed Frequency Growth through Frequency Selection and Dimensional Separation in Quantum Machine Learning

Michael Poppel, David Bucher, Maximilian Zorn, Nico Kraus, Jonas Stein, Claudia Linnhoff-Popien

Introduction

Quantum Machine Learning (QML) models often utilize angle encoding to map classical data into quantum states. Mathematically, this transformation allows a quantum model to be represented as a truncated Fourier series. While this provides the theoretical framework for universal function approximation, it introduces a significant challenge known as the "Frequency Explosion."

Quantum models using angle encoding can be represented as truncated Fourier series, where increasing the number of frequencies (n) allows for better approximation of target functions.

The Problem: Frequency Explosion and Parameter Insufficiency

As QML models scale with more qubits or layers to accommodate high-dimensional data, the number of available frequencies grows exponentially. In multi-dimensional contexts, these frequencies interact to create "mixed frequencies," resulting in a double exponential scaling.

The authors identify a critical bottleneck: parameter insufficiency. To effectively control the frequency spectrum during training, a model requires at least as many trainable parameters as there are unique frequencies. When redundant or unnecessary frequencies overwhelm the model, the optimizer fails to converge, rendering high-dimensional QML practically intractable on current hardware.

Proposed Solutions

To address this, the paper proposes two primary strategies to restore model trainability:

1. Strategic Frequency Selection

Instead of dense encoding schemes (like ternary encoding) that generate a vast, redundant spectrum, the authors advocate for sparse frequency selection. By utilizing specific "prefactors" that align with the known frequency components of a target function, they eliminate irrelevant frequency "noise." This approach drastically reduces the parameter budget required to manage the model's expressivity.

2. Near-Zero Weight Initialization

Standard random initialization often prematurely "activates" a wide range of frequencies, creating a complex optimization landscape full of local minima. The authors propose scaling initial weights by factors of 0.1 or 0.01. This "near-zero" initialization keeps frequency coefficients low at the start, allowing the optimizer to selectively "turn on" only the frequencies necessary to fit the training data.

Experimental Results

The effectiveness of these methods was tested using Variational Quantum Circuits (VQCs) on synthetic "white-box" datasets where target frequencies were known.

Impact of weight initialization scaling on model performance. Near-zero scaling (0.01) significantly improves the R² score by preventing the activation of redundant frequencies across varying parameter counts.

Key findings include:

  • Enhanced Parameter Efficiency: The frequency selection method achieved near-optimal performance (median $R^2 \approx 0.95$) using only 78% of the parameters required by standard dense approaches.
  • Overcoming the 2D Wall: In two-dimensional experiments where standard models typically failed ($R^2 \approx 0$), the combination of near-zero initialization and frequency selection enabled architectures to reach $R^2$ scores above 0.9.
  • Initialization as Regularization: Near-zero scaling acts as an implicit filter, forcing the model to justify the use of each frequency and preventing the "parameter collapse" observed with standard initialization.

Conclusion and Future Work

While these techniques significantly improve QML performance, they currently rely on a priori knowledge of target frequencies. Future research will focus on data-driven selection—using classical Fourier analysis on data subsets to identify dominant frequencies—and automated prefactor tuning to allow quantum models to learn their optimal frequency spectra dynamically. These advancements are crucial for applying sparse QML techniques to real-world datasets where the underlying frequency structure is unknown.

Summary updated

Abstract

Quantum machine learning research has expanded rapidly due to potential computational advantages over classical methods. Angle encoding has emerged as a popular choice as feature map (FM) for embedding classical data into quantum models due to its simplicity and natural generation of truncated Fourier series, providing universal function approximation capabilities. Efficient FMs within quantum circuits can exploit exponential scaling of Fourier frequencies, with multi-dimensional inputs introducing additional exponential growth through mixed-frequency terms. Despite this promising expressive capability, practical implementation faces significant challenges. Through controlled experiments with white-box target functions, we demonstrate that training failures can occur even when all relevant frequencies are theoretically accessible. We illustrate how two primary known causes lead to unsuccessful optimization: insufficient trainable parameters relative to the model's frequency content, and limitations imposed by the ansatz's dynamic lie algebra dimension, but also uncover an additional parameter burden: the necessity of controlling non-unique frequencies within the model. To address this, we propose near-zero weight initialization to suppress unnecessary duplicate frequencies. For target functions with a priori frequency knowledge, we introduce frequency selection as a practical solution that reduces parameter requirements and mitigates the exponential growth that would otherwise render problems intractable due to parameter insufficiency. Our frequency selection approach achieved near-optimal performance (median $R^2 \approx 0.95$) with 78\% of the parameters needed by the best standard approach in 10 randomly chosen target functions.
BibTeX
@misc{poppel2025mitigating,
  author = {Michael Poppel and David Bucher and Maximilian Zorn and Nico Kraus and Jonas Stein and Claudia Linnhoff-Popien},
  title = {{Mitigating Exponential Mixed Frequency Growth through Frequency Selection and Dimensional Separation in Quantum Machine Learning}},
  year = {2025},
  eprint = {2508.10533},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.10533},
}
Figure