Parity Floors in Quantum Denoisers: A Closed-Form Benchmark for Fixed-Map Denoising Networks

AI summary

Overview

Research area: Quantum machine learning — specifically, the expressivity limits of fixed (non-trainable) quantum feature maps when used as conditioning components inside diffusion-model denoisers.

Technical level: Intermediate. The mathematics is stated cleanly (an orthogonal-decomposition argument in L²), but readers will benefit from familiarity with angle-encoding circuits (RY + CNOT), Pauli-Z readouts, and the denoising/score formulation of diffusion models. The conceptual message is accessible without the proofs.

Scope (one sentence): The paper introduces a synthetic torus-diffusion benchmark with analytically computable noising and proves, then measures, that depth-1 RY+CNOT+Pauli-Z feature maps can only represent even functions of the encoded angles, so their excess denoising risk is dominated by an irreducible "parity floor" equal to the odd-sector mass of the Bayes denoiser.

What This Paper Is About

Quantum circuits are increasingly inserted into diffusion models as fixed feature maps, and small gains are often reported — but standard image benchmarks like FID cannot say why a map helps or fails, because they conflate "the map cannot represent the target," "the map overfits," and "the classical baseline was weak." This paper builds a controlled testbed where the noise process, the target's symmetry decomposition, and a lower bound on achievable error are all known in closed form, and uses it to isolate one specific structural constraint — parity (even/odd symmetry under flipping the sign of all input angles) — as the dominant limitation of a standard single-layer encoding.

Key Contributions

1. CoupledPhaseTexture benchmark. A synthetic diffusion process on an 8-dimensional torus (θ ∈ 𝕋⁸) with a 3-body coupled-phase Gibbs prior and analytic heat-kernel noising (wrapped Gaussian, attenuating Fourier mode k by e^{−σ²k²/2}), a degree-3 both-parity trigonometric numerical reference fit at N = 8000, and a baseline suite with feature counts reported so that parity is not confounded with capacity.

2. A containment-free parity error-floor theorem. For any even-sector feature class Q, the excess risk decomposes exactly as R_σ(Q) − R*σ = ‖Π_odd m_σ‖² + inf{q∈Q}‖Π_even m_σ − q‖², with no assumption that Q is contained in a particular bank, linear, or closed — only that Q ⊆ {even}. The first term is an irreducible, σ-resolved lower bound.

3. A reference-relative diagnostic instrument. A measured excess is split into an empirical parity proxy P̂_σ (even-reference risk minus reference risk) and a signed empirical residual δ̂_{σ,N}, with finite-sample effects assessed independently via an N-sweep — turning a null result into a named diagnosis.

4. An empirical characterization of the depth-1 RY+CNOT+Z family as parity-dominated on two structurally distinct priors, with negative controls (a purely classical cosine-only bank is floored identically; adding the sine sector recovers the reference) confirming the deficit is parity, not quantumness or capacity.

Main Findings

• The parity decomposition holds exactly. Lemma 1 shows that for U = U_ent ⊗ᵢ RY(aᵢ) with any CNOT entangler, each readout is ⟨Z_S⟩(a) = ∏_{i∈S′} cos aᵢ for some set S′, hence even. CNOT conjugation maps Z-strings to Z-strings and never produces X or Y, so linear (ridge) readout preserves evenness.

• The measured excess is dominated by the parity proxy at every noise level. Across σ = 0.20, 0.35, 0.50, 0.70, 1.00, 1.40, the parity proxy is 0.250, 0.234, 0.212, 0.171, 0.101, 0.030 versus a quantum excess of 0.263, 0.250, 0.230, 0.190, 0.109, 0.024.

• The signed residual is small but statistically resolved. |δ̂_{σ,N}| ≤ 0.018 on 8 graphs, with paired 5000-resample bootstrap 95% CIs excluding zero at every σ. It is slightly negative at σ = 1.40 (−0.006, CI [−0.007, −0.005]) because the finite even reference need not contain the quantum even feature span. The map is therefore parity-dominated, not exactly parity-bound.

• The floor's σ-profile matches heat-kernel attenuation. Fitting the measured proxy to c·e^{−aσ²} gives R² = 0.99 (a = 1.10) on CoupledPhaseTexture and R² = 0.997 (a = 0.65) on the second prior, of the same order as the mode-1 value a = 1. The authors describe this as a descriptive fit, not an exact identity for the conditional denoiser.

• The deficit is representational, not a sample-complexity artifact. At σ = 0.5 over N = 100, 300, 1000, 3000, the depth-1 quantum map (0.304, 0.264, 0.243, 0.235) and the even-classical equivalent (0.323, 0.260, 0.229, 0.218) track closely by N ≥ 300, while the sin-carrying classical bank (0.225, 0.085, 0.031, 0.020) stays near the reference.

• Higher-body readouts help inside the even sector; entanglement does not close anything. At σ = 0.5, N = 3000: order-1 (0.397 ± 0.005) → order-2 (0.234 ± 0.005) → order-3 (0.219 ± 0.003). Entangler variants give none 0.215, line 0.214, ring 0.234, all-to-all 0.342.

• Re-uploading breaks evenness but fails to close the floor, non-monotonically. Depth 1/2/3 gives 0.234 ± 0.005 / 0.227 ± 0.012 / 0.400 ± 0.004, with max_a‖q(a) − q(−a)‖ jumping from 0 to O(1) at depth ≥ 2. The authors attribute the non-monotone trend to representation misalignment (the fixed 36-dimensional readout pushed to high frequency), not capacity.

• The result reproduces on a structurally different prior. On a 2-body frustrated signed-XY prior (β = 2.0, random 3-regular graph, J_ij ∈ {±1}), the quantum excess tracks the proxy with mean ratio 1.00 and residuals ≈ 0; the classical even bank is floored at the same level (+0.175 averaged over σ) and the classical full bank matches the reference (−0.004).

• Classical controls are decisive. Averaged over σ, a classical cosine-only bank (64 features) is floored at +0.165 versus +0.178 for the quantum even map (36 features), while the classical full bank (128 features, + sin) reaches +0.004. Among quantum fixes, odd ⟨Xᵢ⟩ readouts (100 features) leave +0.144 mean excess and a σ-coupled RZ(φ(σ)=σ) + ⟨Y⟩ variant (72 features) gives +0.176, a ≈ 0.003 reduction — neither matches the sine-carrying classical bank.

• A conditional-image check does not contradict the prediction. On 3-class conditional generation of structural civil images (classifier accuracy 1.00, 500 per class), controllability saturates near 1.0 for the quantum fixed map (1.000), a parameter-matched classical fixed map (0.999), and the no-map baseline (0.999); FID is 13.51 for the quantum map versus 12.65 for the parameter-matched classical map and 12.91 for the trainable classical SE core. The authors are explicit that this is an external-validity check, not an image-generation benchmark, and that the metric of record is controllability, never unconditional FID.

Methodology in Plain English

The authors pick a task where the answer is knowable. Data are 8 angles drawn from a prior that couples triples of angles and is symmetric under flipping every angle's sign. Noise is added by wrapping a Gaussian around the circle — the heat kernel on a torus — which is analytically tractable and damps fine angular detail as σ grows. The network is asked to predict the clean coordinates (cos θ₀, sin θ₀) from noisy angles, i.e. to act as the Bayes denoiser.

Because the prior and the noise are both sign-symmetric, functions split cleanly into an even part (cosines) and an odd part (sines) that are orthogonal. The authors prove that a single angle-encoding layer followed by CNOTs and Pauli-Z measurements can only produce products of cosines — strictly even. So the sine half of the Bayes target is unreachable in principle, and the excess risk splits into that unreachable part plus whatever remains within the even sector.

To measure this without an analytic oracle, they fit two references independently of any map under test, both at N = 8000 with cross-validated ridge regression: a degree-3 trigonometric bank using both parities (the reference R_ref) and a cosine-only bank (the even reference). The gap between the even reference and R_ref is the empirical "parity proxy"; the gap between a candidate map and the even reference is a signed residual. They then vary readout order, entangler topology, re-upload depth, sample size, and prior, and add classical banks with matched feature counts to separate "parity" from "quantum" from "capacity."

Why This Matters

Research impact. The paper supplies a template for converting disappointing or ambiguous quantum-ML results into mechanistic diagnoses. Its lower bound is task-resolved — it depends on the diffused Bayes denoiser's odd-sector mass, not just on the static Fourier spectrum of an encoding — which is something standard dequantization/expressivity results do not provide. It also draws a careful boundary: the results motivate but do not establish nonclassical data access or unmatched feature classes as routes to advantage.

Real-world applications (implications, not claims made by the paper):

• Auditing reported gains from quantum layers in generative models, using closed-form testbeds before committing to expensive image-scale studies.
• Designing hybrid generative models for data with periodic or toroidal structure, where symmetry constraints on the encoder can be identified in advance.
• Building diagnostic tooling for quantum-ML practitioners that says which constraint binds, rather than only reporting a single aggregate score.
• Informing benchmark design for scientific generative modeling on structured domains where heat-kernel noising is natural.

Industry relevance. For teams deciding whether a fixed quantum feature map belongs in a production conditioning path, the work argues that capacity-matched, feature-count-reported classical controls are mandatory, and that a null result should be reported with an attribution rather than a bare FID comparison. The image check — no controllability or FID edge over a parameter-matched classical map — is directly relevant to deployment decisions.

Future Directions

• Nonclassical data access. The authors repeatedly flag learning from quantum data as the regime their boundary argument leaves open, while warning that in any simulable instance, privileged access to the state-preparation circuit can eliminate the intended measurement-access separation.
• Feature classes without efficient classical surrogates. Since odd-augmented readouts (100 features, +0.144 mean excess) and a σ-coupled encoder (72 features, +0.176) did not match the sine-carrying classical bank, the search is for maps whose relevant basis is aligned with the denoiser's low-frequency statistics, not merely parity-complete.
• Extending the attribution beyond parity-invariant families. The instrument currently applies to parity-invariant fixed encoder/readout/entangler families exposing comparable feature vectors; generalizing the decomposition to other symmetry groups is the natural next step.
• Ruling out fine-tuning explanations for the re-upload result. Depth 3 rising to 0.400 ± 0.004 while conditioning improves is attributed to representation misalignment; whether a different readout or frequency budget could make re-uploading a genuine capacity gain is left open. The paper does not report whether the truncated appendices resolve this.

Target Audience

Quantum machine learning researchers and graduate students studying expressivity, dequantization, and hybrid generative models; benchmark designers who need capacity-matched comparisons; and practitioners evaluating whether to insert fixed quantum feature maps into diffusion pipelines. Readers wanting only the conceptual story can follow the Introduction, the theorem statement in Section 3, and the control results in Section 8, while the appendices serve those checking the proofs and Monte Carlo generation details (200 Metropolis sweeps per chain, acceptance ≈ 0.53, stability confirmed from 200 to 800 sweeps with changes < 0.004).

Authors’ abstract

Fixed quantum feature maps are increasingly inserted into diffusion denoisers, but standard image benchmarks do not reveal which structural constraint limits them. We introduce CoupledPhaseTexture, a torus-diffusion benchmark with analytic heat-kernel noising that separates parity, within-sector approximation, and sample-complexity limitations. For the depth-1 RY+CNOT+Pauli-Z family we prove a containment-free parity floor: all reachable features are even functions of the encoded angles while the sine components of the Bayes denoiser are odd, so the excess risk splits exactly into an inaccessible odd part and a within-sector residual. The first term is an irreducible, noise-scale-resolved lower bound holding for every even feature class, with no containment, linearity, or closedness assumption on the feature class. The obstruction is a property of the noise-conditioned denoising target rather than static representability: the floor is re-derived at each noise scale because the target's parity content changes with noise. The measured excess is dominated by the parity proxy on two distinct priors. Higher-order Z readouts improve the even sector, but entanglement does not lower the floor and re-uploading does not reliably close it. Classical controls confirm the deficit is parity rather than quantumness: a cosine-only bank is floored similarly, while adding the sine sector matches the reference. Among tested constructions, odd readouts and a noise-coupled encoder do not match the sine-carrying classical bank. These results motivate nonclassical data access or feature classes without efficient classical surrogates; they do not establish either as sufficient for quantum advantage.