AI 中文总结
该研究针对可编程光子处理器中相位波动限制性能的问题，分析相位漂移代价，提出自适应求积读出规则与感知增量估计器，经模拟验证其性能优于传统方法。
AI 中文摘要
光学输入间的相位波动会限制可编程光子处理器的性能，因为其输出功率依赖于相干干涉。我们研究了依次而非同时测量正弦和余弦求积时产生的相位漂移代价。该分析的动机来自对一款八模可编程光子处理器的测量，包括每个通道以约125个样本/秒的速率采集的35次300秒自由运行记录。这些记录为估计选定重构间隔下的相位增量方差提供了经验途径，该估计定义在第二次测量时刻。对于固定求积阶，atan2重构的扰动给出$e_{C\to S}=-\tau_\tau\boldsymbol{\text{sin}}^2\boldsymbol{\text{φ}}_0+O(\tau_\tau^2)$和$e_{S\to C}=-\tau_\tau\boldsymbol{\text{cos}}^2\boldsymbol{\text{φ}}_0+O(\tau_\tau^2)$。令$Q_\tau=\text{Var}(\tau_\tau)$，均匀相位平均给出一阶漂移均方误差$3Q_\tau/8$。一种相位预测排序规则先测量局部信息较少的求积，再测量信息较多的求积，其一阶均匀代价为$(3/8-1/\boldsymbol{\text{π}})Q_\tau$，比固定阶值低84.9%。我们还从局部状态空间模型推导出感知增量的估计器，边缘化未知相位增量会使过时相位观测的方差增加$Q_\tau$，将其费舍尔信息从$I$降至$I/(1+IQ_\tau)$。对于理想平衡泊松探测，每个求积的费舍尔信息等于其检测到的信号光子数，这在空间信息和相位增量方差中产生了无量纲架构边界。非线性蒙特卡洛模拟验证了微扰定律，量化了对预测误差的鲁棒性，并在共同噪声模型下比较了同时、固定阶、感知增量和自适应接收机的性能。
英文摘要
Phase fluctuations between optical inputs limit programmable photonic processors because their output powers depend on coherent interference. We study the phase-drift penalty that arises when sine and cosine quadratures are measured sequentially rather than simultaneously. The analysis is motivated by measurements from an eight-mode programmable photonic processor, including 35 free-running recordings of 300 s acquired at approximately 125 samples per second per channel. These recordings provide an empirical route for estimating the phase-increment variance at a selected reconfiguration interval. The estimate is defined at the time of the second measurement. For fixed quadrature order, perturbation of the atan2 reconstruction gives $e_{C\to S}=-δ_τ\sin^2ϕ_0+O(δ_τ^2)$ and $e_{S\to C}=-δ_τ\cos^2ϕ_0+O(δ_τ^2)$. Writing $Q_τ=\operatorname{Var}(δ_τ)$, uniform phase averaging gives the first-order drift mean-square error $3Q_τ/8$. A phase-predicted ordering rule measures the locally less informative quadrature first and the more informative quadrature second. Its uniform first-order penalty is $(3/8-1/π)Q_τ$, which is 84.9 percent below the fixed-order value. We also derive an increment-aware estimator from a local state-space model. Marginalizing the unknown phase increment increases the variance of a stale phase observation by $Q_τ$, reducing its Fisher information from $I$ to $I/(1+IQ_τ)$. For ideal balanced Poisson detection, the Fisher information of each quadrature equals its detected signal-photon number. This yields dimensionless architecture boundaries in spatial information and phase-increment variance. Nonlinear Monte Carlo simulations validate the perturbative laws, quantify robustness to prediction error, and compare simultaneous, fixed-order, increment-aware, and adaptive receivers under a common noise model.
