Fourier Analysis of Variational Quantum Circuits for Supervised Learning

Technical Summary: Fourier Analysis of Variational Quantum Circuits for Supervised Learning

Problem Statement
Variational Quantum Circuits (VQCs) are increasingly utilized as function approximators in supervised learning. While it is established that the output of a VQC can be represented as a truncated Fourier sum of the input data, the set of available frequencies (the spectrum) is often assumed to be determined solely by the encoding gates. The authors identify a critical limitation in this assumption: the variational part of the circuit can constrain specific Fourier coefficients to zero, effectively removing frequencies from the spectrum that would otherwise be theoretically available based on the encoding alone. Furthermore, the functional dependence of these Fourier coefficients on the variational parameters has remained largely unexplored. This lack of precise spectral characterization hinders the ability to predict which VQC architecture is best suited for a specific dataset without resorting to full training.

Methodology
The paper proposes a two-pronged methodological approach:

1. Exact Spectrum and Coefficient Calculation:
   Building on the algorithm by Nemkov et al. [18], the authors develop a procedure to compute the exact Fourier spectrum and coefficients for any given VQC (specifically those transpilable into Clifford + Pauli circuits).

   • Algorithm: The method utilizes a binary tree construction where nodes represent circuits in a canonical normal form. By traversing paths through this tree and multiplying trigonometric factors associated with Pauli rotations, the algorithm decomposes the expectation value into a sum of trigonometric polynomials.
   • Coefficient Analysis: The authors derive that Fourier coefficients are trigonometric polynomials of the variational parameters $\theta$. They establish necessary and sufficient conditions for combinatorial parts of these coefficients to vanish, allowing for the identification of frequencies that are strictly zero regardless of parameter values.
   • Data Analysis: To compare the circuit's spectrum with a dataset, the authors employ the Non-uniform Fast Fourier Transform (NFFT) via iterative numerical inversion. This allows for the extraction of frequency representations from non-uniformly spaced real-world data.

2. Architecture Ranking Heuristic:
   The authors propose a dataset-aware method to rank VQC architectures without training. The core hypothesis is that an optimal architecture must contain the dominant frequencies of the dataset while maintaining a spectrum size small enough to ensure trainability (addressing the expressivity-trainability trade-off).

   • Scoring Metric: A score is calculated based on three components:
     • $R_\Omega$: The energy of the dataset's Fourier coefficients that lie outside the VQC's available spectrum.
     • $R_{corr}$: A Mahalanobis distance heuristic estimating the likelihood that the VQC's variational parameters can reproduce the specific correlations and magnitudes of the dataset's Fourier coefficients.
     • $R_{punish}$: A penalty term proportional to the size of the spectrum, discouraging overly expressive but potentially untrainable architectures.
   • The architecture with the lowest total score is predicted to fit the data best.

Key Contributions

• Refined Spectral Characterization: The paper demonstrates that the Fourier spectrum of a VQC is not solely determined by encoding gates but is significantly constrained by the variational layers. This corrects the overestimation of available frequencies found in previous literature [16].
• Functional Dependence: The authors provide the first description of Fourier coefficients as explicit trigonometric polynomials of the variational parameters, including an algorithm to compute the exact spectrum and coefficients for any given circuit.
• Predictive Architecture Selection: A novel algorithm is introduced that predicts the best-fitting VQC architecture for a given dataset by comparing the dataset's Fourier transform against the analytically derived spectrum of candidate circuits.

Results
The authors tested their methods on eleven different VQC architectures using two datasets: the synthetic Friedman regression function and the MNIST handwritten digit dataset (reduced to 2D via t-SNE).

• Spectral Variability: Despite using identical encoding gates across all architectures, the resulting Fourier spectra varied significantly, ranging from 7 to over 6,000 frequencies. This confirmed that variational layers actively prune the spectrum.
• Ranking Performance: The proposed scoring method successfully identified architectures (specifically circuits 2, 7, and 9) that achieved the lowest testing losses on the Friedman dataset and performed competitively on the MNIST dataset.
• Limitations: The study noted that while the heuristic generally correlated with performance, the $R_{corr}$ term (based on a subset of coefficients) sometimes failed to perfectly distinguish between architectures with similar spectral properties but different internal coefficient correlations. Additionally, the computational cost of calculating the full spectrum scales exponentially with circuit depth, limiting the current proof-of-concept to small circuits (4 qubits, 3 layers).

Significance and Claims
The paper claims to provide the first algorithm capable of determining the exact functional relationship between variational parameters and Fourier coefficients. By revealing that the variational part of a circuit constrains the spectrum, the authors argue that VQCs should be engineered specifically for a task rather than simply maximizing expressivity. The proposed method offers a classical simulation tool to predict model performance and select architectures, potentially saving the computational resources required for training suboptimal models. However, the authors modestly note that their method is currently limited to small circuits due to the exponential scaling of classical simulation and that the ranking heuristic, while promising, requires further refinement regarding the weighting of its components.