Classical Limits of Spectral Filtering in Quantum Generative Models

Abstract: Spectral filtering has been proposed as a route to regularization in quantum generative models: the quantum Fourier transform exposes the amplitude spectrum of a quantum circuit Born machine, and a diagonal filter suppresses the high frequencies associated with finite-sample noise, an operation whose classical counterpart seemingly requires manipulating an exponentially long amplitude vector. We examine whether this coherent operation produces anything that classical post-processing of samples from the unfiltered model cannot match. Measuring the filter against convolution with a symmetric probability kernel at matched sampling cost, which accounts for the post-selection overhead of attenuation, we derive necessary and sufficient conditions for the gap between the two to vanish. Magnitude (attenuating) filters obey a dichotomy: at a fixed affordability threshold, the filtered output is either a constant-size Fourier object with an efficient classical sampler, or the passband must widen until no fixed frequency is attenuated and the filter no longer smooths. In neither case does the filter create a quantum-classical separation. Whatever separation survives is inherited from the spectral phase of the input state. Numerical experiments on trained circuit Born machines confirm the classification and show that the deciding phases are invisible to the Born-rule training loss and set by the initialization. Within the diagonal family, pure phase filters remain the only spectral operations exempt from these constraints.

• The paper establishes that affordable magnitude filters either yield constant-size Fourier descriptions that classical algorithms can sample efficiently or fail to attenuate any fixed frequency, preventing simultaneous smoothing and quantum advantage.
• The paper derives necessary and sufficient conditions for classical reproduction based on phase alignment and envelope realizability, showing that any residual separation comes from the input state's spectral phase rather than magnitude filtering.
• Experiments find exponentially widening affordable passbands in trained Born machines, persistent phase certificates, and filtered outputs roughly 0.31 total-variation distance from target distributions at ten times the sampling cost.

Spectral filtering has been proposed as a route to regularization in quantum generative models: applying a quantum Fourier transform (QFT) to a circuit Born machine exposes its amplitude spectrum, and a diagonal filter attenuates the high frequencies associated with finite-sample noise. Because the amplitude vector has length $N = 2^n$, this coherent operation has appeared classically prohibitive. The paper under review examines whether coherent spectral filtering produces anything that classical post-processing of samples from the unfiltered model cannot match, and answers negatively for magnitude filters. Its central result is a dichotomy: at a fixed affordability threshold, a magnitude (attenuating) filter either returns a constant-size Fourier object with an efficient classical sampler, or its passband must widen until no fixed frequency is attenuated and the filter no longer smooths. Whatever quantum-classical separation survives is inherited from the spectral phase of the input state, not produced by the filter.

Framework: coherent filtering versus classical smoothing

The object of study is a quantum circuit Born machine, $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$ prepared by a parameterized circuit, whose output distribution is $p(x) = |a_x|^2$. The regularization argument rests on a spectral fact: an empirical distribution from a finite training set is sparse and spreads its Fourier spectrum across all frequencies, so suppressing high frequencies acts as a smoothness prior. The coherent proposal applies a QFT to the state and multiplies amplitudes by a diagonal filter $g$, with $\|g\|_\infty = 1$, before an inverse QFT and measurement.

The paper's key methodological move is to fix a classical baseline with the same operational interface: post-processing samples of $p(x)$ by convolution with a symmetric probability kernel $K$, which rescales each Fourier mode of the distribution by a real, even factor $\hat K(m)$. By contrast, the coherent filter acts on the amplitude spectrum. Via the Wiener–Khinchin relation, the distribution's spectrum is the autocorrelation $\rho_m(k) = \hat a(k)\hat a^*(k-m)$, and the filter inserts the pair weight $W_m(k) = g(k)g^*(k-m)$ inside the autocorrelation sum rather than multiplying $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$0 pointwise. The filtered mode is

$\ket{\psi(\theta)} = \sum_x a_x \ket{x}$1

a complex reweighting, in contrast to the real scalar multiplication available classically. A second distinction is cost: attenuation ($\ket{\psi(\theta)} = \sum_x a_x \ket{x}$2) is non-unitary and requires ancilla post-selection with success probability $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$3, so filtering reduces the effective sample count. The analysis is conducted at a matched-cost affordability floor $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$4.

The classicality criterion and the coherent residual

The gap $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$5 is defined as the infimal total-variation distance between the coherently filtered distribution $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$6 and any classical smoothing $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$7. The per-lag diagnostic is the coherent residual $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$8, the component of $\ket{\psi(\theta)} = \sum_x a_x \ket{x}$9 orthogonal to the line $p(x) = |a_x|^2$0 reachable by classical smoothing. It admits the closed form

$p(x) = |a_x|^2$1

which locates the entire quantum content of a diagonal filter in a single scalar per lag. The simulability criterion (Proposition 1) states that $p(x) = |a_x|^2$2 if and only if two independent conditions hold: (a) phase alignment, $p(x) = |a_x|^2$3 at every regular lag; and (b) envelope realizability, the real envelope $p(x) = |a_x|^2$4 being the transfer function of a symmetric probability kernel (a Bochner-type positive-definiteness condition). Condition (a) constrains the imaginary part of $p(x) = |a_x|^2$5, which for a magnitude filter ($p(x) = |a_x|^2$6, so $p(x) = |a_x|^2$7 real) is sourced solely by the spectral phases of the input. Condition (b) is purely classical and phase-independent.

Affordability, compactness, and the dichotomy

A spectral compactness lemma bounds the mode count $p(x) = |a_x|^2$8 of the filtered output by a quantity depending only on the passband scale $p(x) = |a_x|^2$9, the filter's overlap profile, $g$0, and $g$1—independent of $g$2 and of the input phases. For a Gaussian low-pass, $g$3. Meanwhile, on broadband inputs whose spectral mass escapes every fixed window, the affordability constraint $g$4 at every fixed $g$5 forces the affordable passband scale $g$6 to diverge: an affordable magnitude filter cannot implement an $g$7-independent cutoff. Regularization and affordability compete.

Combining these yields the dichotomy theorem. Cell (I), classically reproducible: if $g$8, then $g$9 uniformly in $\|g\|_\infty = 1$0 and $\|g\|_\infty = 1$1 is reproduced to $\|g\|_\infty = 1$2 by a distribution samplable in $\|g\|_\infty = 1$3 time from the retained Fourier coefficients (via clipping and renormalization of the trigonometric polynomial). Cell (II), no affordable cutoff: the filter attenuates nothing at any fixed frequency. Within cell (II), a phase certificate $\|g\|_\infty = 1$4 lower-bounds the gap; in the certified branch ($\|g\|_\infty = 1$5) the separation is carried by the input's spectral phase, while in the uncertified branch no single-lag observable separates $\|g\|_\infty = 1$6 from classical smoothing, and any residual gap runs through phase-free channels such as envelope non-realizability. In neither cell is the separation attributable to the filter, which contributes only the real window $\|g\|_\infty = 1$7 and the post-selection cost.

A gap surviving in cell (I) is explicitly characterized as self-undermining: it places $\|g\|_\infty = 1$8 outside the smoothing class, but so does the constant-size classical sampler that reproduces $\|g\|_\infty = 1$9. The paper also concedes that the split at $p(x)$0 is chosen for simplicity, not because it is the boundary of classical simulability; for polynomially decaying overlap profiles, cell (I) extends to $p(x)$1.

Transfer to the hypercube

The framework transfers to any finite abelian group; on $p(x)$2 with the bit-flip low-pass $p(x)$3, the pair weight factorizes into the classical i.i.d. bit-flip transfer function times a $p(x)$4-dependent window. Crucially, because the Walsh characters are real, $p(x)$5 is real for every input and every diagonal filter, so the phase certificate is identically zero even for phase filters, and the certified branch of cell (II) is empty on the cube. Low-degree inputs admit a constant flip rate and confine the output to degree $p(x)$6 (a $p(x)$7-samplable object), while broadband inputs force $p(x)$8: no fixed degree is attenuated. The confinement for low-degree inputs is inherited from the input, not imposed by the filter.

Numerical evidence

Experiments with a Gaussian low-pass at $p(x)$9 confirm the classification. For a Gaussian input spectrum of width $K$0, the measured $K$1 matches the closed form $K$2 to a relative deviation of $K$3. Spectrally incoherent inputs exhibit a certificate decaying as $K$4 (fit $K$5 over $K$6, 128 seeds), consistent with the $K$7 washout bound, while their gap plateaus near $K$8 as phase speckle in the uncertified branch.

The most consequential experiments concern trained Born machines (hardware-efficient $K$9 ansatz, depth 4, trained on a bimodal target with KL loss). Both trained and random-parameter circuits have exponentially growing $\hat K(m)$0, placing them in cell (II), and the trained circuits' certificate is flat in $\hat K(m)$1 (slope $\hat K(m)$2), placing them in the certified branch generically. The mechanism is that the Born-rule loss sees only $\hat K(m)$3 and neither needs nor removes the spectral phases, which are inherited from initialization. Control experiments on an $\hat K(m)$4 state isolate this: erasing the phases (leaving the output distribution unchanged) collapses the model into cell (I) with certificate $\hat K(m)$5 and exact gap $\hat K(m)$6—roughly an order of magnitude below the trained value of $\hat K(m)$7—while keeping the trained phases but substituting exact target magnitudes preserves the gap to within $\hat K(m)$8.

The practical verdict is stark: the filtered output of a trained circuit lands at $\hat K(m)$9 from its own training target—fifteen times further than the unfiltered model, at ten times the sampling cost—with the gap scattering across seeds ($\rho_m(k) = \hat a(k)\hat a^*(k-m)$0–$\rho_m(k) = \hat a(k)\hat a^*(k-m)$1 at $\rho_m(k) = \hat a(k)\hat a^*(k-m)$2). If the goal is smoothing, the filter delivers it only in the regime $\rho_m(k) = \hat a(k)\hat a^*(k-m)$3, where a classical kernel does the same without post-selection cost; if the goal is $\rho_m(k) = \hat a(k)\hat a^*(k-m)$4 itself, any hardness claim attaches to the input state, which can be prepared and sampled without the filter. The quantum candidate is the input state, not the filter.

Limitations and open questions

The authors state several qualifications plainly. The cyclic group $\rho_m(k) = \hat a(k)\hat a^*(k-m)$5 is worked out in detail while the hypercube transfer is only sketched, and the analysis covers only filters diagonal in the Fourier basis—mode-coupling spectral operations are outside its scope. The classical competitor is convolution with a symmetric probability kernel, the class the filter's advertised semantics belongs to, not arbitrary classical post-processing; a positive gap against $\rho_m(k) = \hat a(k)\hat a^*(k-m)$6 is a separation from smoothing, not from classical simulation, although dropping reflection symmetry recovers a median $\rho_m(k) = \hat a(k)\hat a^*(k-m)$7 across all families with nonzero gap (with random-parameter circuits the loosest at a minimum ratio of $\rho_m(k) = \hat a(k)\hat a^*(k-m)$8). The gap is a large-sampling-limit statement; how the classification degrades at finite shot budgets, where $\rho_m(k) = \hat a(k)\hat a^*(k-m)$9 must be resolved against shot noise at a rate reduced by $W_m(k) = g(k)g^*(k-m)$0, is left open. Within the diagonal family, pure phase filters ($W_m(k) = g(k)g^*(k-m)$1) remain the only spectral operations exempt from the constraints—they incur no post-selection cost, admit no kernel surrogate, and escape the compactness bound—but whether they yield a useful inductive bias, rather than merely an unsimulable one, is the question the paper leaves unanswered. A second open direction follows from the observation that the decisive phases are invisible to Born-rule losses: training objectives that see the phase directly would convert that degree of freedom from an initialization accident into a design variable.

Conclusion

This paper provides necessary and sufficient conditions under which coherent spectral filtering of a quantum Born machine is equivalent to classical smoothing of its samples. The resulting dichotomy shows that magnitude filters cannot simultaneously regularize, remain affordable, and create quantum-classical separation: at matched cost they either produce a constant-size classically simulable object or attenuate nothing at any fixed frequency. The separation that survives is sourced entirely by the spectral phase of the input state—a quantity the Born-rule training loss does not control and that, on trained circuits, is set by initialization. The analysis thereby redirects attention from attenuating spectral design toward phase filters and phase-aware training objectives as the remaining candidates for a genuine quantum advantage in this setting.