Classical Limits of Spectral Filtering in Quantum Generative Models

Technical Summary: Classical Limits of Spectral Filtering in Quantum Generative Models

Problem Statement
Quantum generative models, specifically Quantum Circuit Born Machines (QCBMs), utilize the exponential size of the Hilbert space to represent complex probability distributions. A proposed method to regularize these models and mitigate finite-sample noise involves spectral filtering: applying a Quantum Fourier Transform (QFT) to expose the amplitude spectrum of the quantum state, applying a diagonal filter to suppress high frequencies, and then performing an inverse QFT and measurement.

The central question addressed is whether this coherent quantum operation offers a genuine advantage over classical post-processing. Specifically, can a coherent filter produce a distribution that cannot be matched by a classical competitor performing convolution with a symmetric probability kernel on samples drawn from the unfiltered model, assuming both operate under matched sampling costs (accounting for the post-selection overhead of attenuation)?

Methodology
The author establishes a rigorous framework to compare the coherent quantum route against a classical baseline:

1. Quantum Route: A quantum state $|\psi\rangle = \sum a_x |x\rangle$ is transformed via QFT to $\hat{a}(k)$. A diagonal filter $g(k)$ is applied, yielding $\hat{a}'(k) = g(k)\hat{a}(k)$. The state is inverse-transformed and measured to produce a distribution $p_g(x)$. This process involves non-unitary attenuation, requiring post-selection with success probability $p_{succ}$.
2. Classical Route: Samples are drawn from the unfiltered distribution $p(x) = |a_x|^2$. These samples are post-processed via convolution with a symmetric probability kernel $K$, resulting in $p \star K$.
3. Metric: The performance gap is quantified by the Total Variation (TV) distance $\Phi_g$ between the coherent output $p_g$ and the optimal classical smoothing $p \star K$.
4. Analysis: The author decomposes the filter action into magnitude filters ($|g(k)| \leq 1$, attenuating) and phase filters ($|g(k)| = 1$, unitary). They analyze the "coherent residual" $R_g(m)$, which measures the deviation of the filtered spectral modes from what is achievable by real scalar rescaling (classical smoothing).

Key Contributions and Results

1. Dichotomy of Magnitude Filters:
   The paper proves a fundamental dichotomy for magnitude (attenuating) filters operating at a fixed affordability threshold ($p_{succ} \geq p^\star$):

   • Cell (I) - Classically Reproducible: If the affordable passband scale $\sigma^\star$ remains bounded (i.e., $O(1)$), the filtered output is a constant-size Fourier object. Such an output can be efficiently sampled classically, regardless of the input phases. The filter provides no separation.
   • Cell (II) - No Affordable Cutoff: If the input spectrum is "broadband" (spectral mass escapes fixed windows), maintaining a fixed success probability forces the passband scale $\sigma^\star$ to diverge. Consequently, the filter cannot implement a fixed-frequency cutoff and ceases to smooth. In this regime, any remaining separation from classical smoothing is not created by the filter but is inherited from the spectral phase of the input state.

2. Role of Spectral Phase:
   The author derives necessary and sufficient conditions for the gap $\Phi_g$ to vanish. They show that magnitude filters (where $g(k)$ is real) cannot generate a phase shift in the distribution's spectrum that classical smoothing cannot replicate. Any non-zero gap in magnitude filtering arises solely from the spectral phase of the input state ($\rho_m(k) = \hat{a}(k)\hat{a}^*(k-m)$). If the input is "phase-trivial" (spectrum is real up to global phase/translation) or spectrally incoherent, the coherent residual vanishes, and the filter is classically simulable.

3. Phase Filters as the Only Quantum Candidate:
   Pure phase filters ($|g(k)| = 1$) are identified as the only spectral operations exempt from these constraints. They incur no post-selection cost ($p_{succ}=1$) and do not suffer from the spectral compactness collapse that forces magnitude filters into Cell (I) or Cell (II). However, the paper notes that even for phase filters, the classification framework suggests that any advantage must be established by mechanisms other than the magnitude-based certification used for attenuation.

4. Numerical Validation:
   Numerical experiments on trained QCBMs confirm the theoretical classification. The results show that for trained models, the deciding phases are often invisible to the Born-rule training loss and are determined by the initialization. When these phases are present, they can create a gap, but this gap is an artifact of the initialization rather than a capability of the filtering operation itself.

Significance and Claims
The paper claims to establish the classical limits of spectral filtering in quantum generative models. Its primary conclusion is that magnitude filters do not create a quantum-classical separation; they either produce outputs that are efficiently classically simulable (Cell I) or fail to smooth at a fixed cost (Cell II).

The author argues that the "quantum advantage" often attributed to spectral filtering is an illusion when viewed through the lens of operational cost and classical post-processing capabilities. The only potential for separation lies in the spectral phase of the input state, which the filter merely preserves or windows, rather than generating. Consequently, the paper suggests that if a genuine quantum advantage exists in this domain, it must rely on pure phase filters, as magnitude-based regularization is fundamentally constrained by classical simulability.

The work reframes the discussion of quantum generative models by shifting focus from the "exponential access" to the amplitude vector to the operational equivalence of the resulting distributions, demonstrating that the "curse of dimensionality" in training often masks the fact that the filtered output may not possess properties inaccessible to classical methods.