arXiv:2610.41721
Leveraging Quantum Measurement Randomness for Shot-Efficient Diffusion Models
Tianhao Guo, Dianxin Zheng, Yuhua Qian, Yaohui Zheng, Yuzhi Yang, Merouane Debbah, Chau Yuen
Abstract
In quantum computing, the readout always carries finite-shot sampling noise. Suppressing this noise through repeated measurements raises the cost of a quantum neural network (QNN). Meanwhile, a stochastic diffusion sampler deliberately adds randomness at each reverse (generation) step. This paper turns the measurement noise from a liability into a resource. Specifically, we use the noise quantum measurement produces anyway to replace part of the noise the sampler would otherwise inject. We characterize the measurement error as multinomial sampling noise propagated through a trainable linear readout. To connect the circuit's Born probabilities to the denoising task, we further formulate the diffusion process on a bounded zero-sum hyperplane mapped to the probability simplex. Therefore, we can employ a compensation noise to match the mean and covariance of the shot-noise-free linear update under some minimum shot conditions. Solving this condition for the shot count gives a state-dependent adaptive budget. For the denoising diffusion implicit model (DDIM), this budget increases as the sampler becomes more deterministic. On an eight-dimensional Gaussian mixture, the schedule achieves comparable generation quality with $303$ shots per sample against a high-shot reference of $1.6\times10^{6}$, about $5300\times$ fewer. Latent diffusion experiments on MNIST further validate the proposed mechanism for generation in a learned latent space.
AI Summary
Stochastic diffusion samplers deliberately inject noise, while quantum denoisers require repeated measurements to suppress readout noise. This paper asks whether finite-shot fluctuations can replace part of the injected randomness, reducing measurement costs without changing the intended reverse-step statistics. The construction combines a quantum-circuit denoiser with a classical linear readout and represents data on a bounded zero-sum hyperplane mapped into the probability simplex. Multinomial measurement covariance is propagated through the readout. If $C_t$ denotes the readout covariance before division by shot count, covariance compensation requires $\sigma_t^2\Xi-k_t^2C_t/M_t\succeq0$, where $k_t$ propagates denoiser error and $\sigma_t$ sets injected noise. This yields $M_t\ge k_t^2\lambda_{\max}(C_t)/\sigma_t^2$ and approximately $M_t\propto\eta^{-2}$ near deterministic DDIM sampling. Conditional means and covariances match exactly under the idealized conditions; distribution matching additionally uses a Gaussian approximation. Simulation experiments use a three-component Gaussian mixture and MNIST latents supplied by a fixed VAE. For the mixture, adaptive sampling uses 303 shots per generated sample versus a $1.6\times10^6$-shot reference, with resolved fraction 0.941 and $\mathrm{MMD}^2=0.0049$. Uniform allocation needs roughly four times the adaptive budget for comparable resolved fraction. Three-qubit MNIST generation uses 430 shots per sample, with image-space $\mathrm{MMD}^2=0.0312$. Limitations include simulator-exact probabilities for shot allocation, exact-probability training, non-Gaussian errors at low shot counts, estimated residual covariances, feasibility projections, shot caps, and denoiser approximation. Their effects lack distributional guarantees. Validation excludes gate and decoherence errors; hardware operation and nonlinear readouts remain open.
Error found: Algorithm 1's plug-in residual covariance produces an excess update covariance of k_t^2 C/M^2 before clipping, so the implemented sampler does not inherit the ideal exact moment-matching guarantee. The reported 303-shot allocation relies on uncharged exact simulator probabilities and therefore does not demonstrate that measurement cost for a hardware-accessible adaptive rule. The per-dimension model's diagonal noise derivation assumes independent coordinate measurements, which is not generally compatible with the simultaneous entangled-circuit readout used in its execution accounting. The proposed affine encoding covers only a restricted part of the probability simplex, and its claimed clamped inverse can violate the zero-sum constraint.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 13 pages, 13 figures, 3 tables