Why quantum neural networks need Fourier analysis

Quantum machine learning (QML) promises to one day solve problems that classical computers cannot handle — from simulating chemical reactions to optimizing financial portfolios. But there's a catch: training quantum models is extremely difficult. They often run into so-called barren plateaus — a phenomenon where gradients vanish toward zero and the model stops learning entirely.

Fourier analysis has become a key tool in recent years for understanding why quantum models work (or don't work). Just as sound can be decomposed into individual sine waves, the behavior of a quantum circuit can be broken down into frequency components. The richer the frequency spectrum, the greater the model's ability to capture complex patterns in data. Conversely — a poor spectrum means the model will be "dumb."

However, previous research on Fourier analysis of quantum circuits focused exclusively on so-called angle embedding — a method where classical data is encoded as rotation angles of individual qubits. The problem is that this method requires one qubit per input feature. For modern datasets with thousands or millions of features (images, text, video), this is completely unsustainable.

Amplitude encoding: more efficient, but more mysterious

The alternative is amplitude encoding (amplitude embedding), which stores data directly in the amplitudes of quantum states. Its main advantage? Encoding N features requires only log₂(N) qubits. A thousand features thus take 10 qubits instead of a thousand — an exponential saving.

"Amplitude encoding is practically indispensable for VQC algorithms," the authors state in the paper published on arXiv on June 12, 2026. "However, its systematic Fourier analysis has been lacking until now."

Researchers Haiyue Kang, Martin Sevior, and Muhammad Usman from the University of Melbourne (and CSIRO Data61) therefore decided to fill this gap. Across 26 pages and 12 figures, they built a mathematical framework for Fourier analysis of VQCs with amplitude encoding — and also incorporated the influence of noise, which is unavoidable on real quantum computers.

What did the researchers discover?

The first surprising finding concerns the zero frequency. With symmetric encoding of input values (around zero), the zero-frequency coefficient completely vanishes — making the model unable to learn any non-zero constant value. In practice, this means a fatal loss of expressive power.

"The difference in expressivity of a single — and yet the most important — Fourier coefficient can lead to a qualitatively different model performance," the authors explain. Their simulations on a 2-qubit circuit confirmed that the model with symmetric encoding barely improved at all, while the non-negative encoding trained normally.

The second key result: under the assumption that the parameterized unitary operations form at least a 2-design (a mathematical property ensuring sufficient "randomness" of the circuit), the authors used Weingarten calculus to derive that:

• The mean value of Fourier coefficients concentrates around zero
• The variance of coefficients decreases exponentially with increasing frequency — similar to angle embedding, but with different detailed behavior

The third crucial insight concerns noise. When a noise channel is introduced into the model (which corresponds to real operation on today's quantum processors), the variance of coefficients is further suppressed by a factor of (∑ pₖ²)^Q, where Q is the number of noise channel applications. In other words — noise systematically kills the model's expressive power, and in a measurable, predictable way.

What does this mean for developers and researchers?

For practitioners in quantum machine learning, this work provides concrete guidance on how to design more effective VQC models:

• Always use non-negative encoding (non-negative domain) for amplitude models — symmetric encoding causes blindness to the lowest frequencies
• Expect exponential decay of expressivity with increasing frequency dimension — complex target functions will require more sophisticated architectures
• The quantifiable influence of noise allows one to predict how much error the model can still tolerate before becoming untrainable

The work also opens the path for further research: what happens when the 2-design assumption does not hold? How will hybrid amplitude-angle encodings perform? And perhaps most intriguingly — can the theoretical bounds from the frequency domain be translated into predictions of training dynamics on specific benchmark tasks?

European and Czech context

Quantum machine learning research is not just a matter for laboratories in Melbourne. Several significant quantum initiatives are running in Europe, including the EuroQCS project, which is building a network of quantum computers across the continent. The Czech Republic has joined through IT4Innovations in Ostrava, where the national supercomputing center operates and the Czech AI Factory recently launched.

For Czech companies and researchers considering investments in quantum technologies, it is crucial to understand that quantum ML is still in the basic research phase. Works like this one from Melbourne are not about "quantum AI arriving tomorrow" — they are about building the mathematical foundations that will one day enable effective deployment on real hardware.

For those interested in deeper study, the full preprint is freely available at arXiv:2606.14206 under the CC BY 4.0 license. The work includes complete mathematical proofs, simulation code, and supplementary materials.