Phase-Drift Limits and Adaptive Quadrature Readout in Programmable Photonic Processors

Technical Summary: Phase-Drift Limits and Adaptive Quadrature Readout in Programmable Photonic Processors

Problem Statement
Programmable photonic processors rely on coherent interference within meshes of tunable Mach–Zehnder interferometers (MZIs) to perform linear transformations. The output powers of these systems depend critically on the relative optical phases between inputs. However, environmental factors such as mechanical vibration, temperature fluctuations, and airflow induce phase drifts that alter interference patterns even when the programmed transfer matrix remains fixed.

A fundamental challenge arises in phase monitoring and estimation: inferring an unknown relative phase typically requires measuring two orthogonal quadratures (proportional to $\cos \phi$ and $\sin \phi$). While spatial receivers can measure these simultaneously, temporal receivers must reconfigure the system to measure them sequentially. This sequential acquisition introduces a "phase-drift penalty" because the two measurements refer to different instantaneous phases. The paper addresses the specific question of how much estimation error is caused by this sequential acquisition and under what conditions a temporal receiver might outperform a simultaneous one, given the trade-off between resource efficiency and phase instability.

Methodology
The study is grounded in experimental data from an eight-mode programmable photonic processor. The authors utilized 35 free-running recordings (300 seconds each) where two equal-power inputs were mapped to four output powers, allowing for the extraction of sine and cosine quadratures via balanced detection.

1. Empirical Modeling: The authors derived an empirical route to estimate the phase-increment variance ($Q_\tau$) associated with a specific reconfiguration interval $\tau$ without assuming a specific drift model (e.g., Brownian motion). They calculated the variance of phase increments from the synchronized four-port data.
2. Perturbative Analysis: The paper derives the estimation error for sequential measurements using a first-order perturbation of the exact atan2 reconstruction. The error is analyzed for two fixed orders ($C \to S$ and $S \to C$) and compared against a simultaneous spatial benchmark.
3. Adaptive and Increment-Aware Strategies:
   • Adaptive Ordering: A rule is proposed to measure the "less informative" quadrature first and the "more informative" one second, based on a phase predictor from the previous cycle.
   • Increment-Aware Estimation: A local state-space model is developed to marginalize the unknown phase increment. This leads to an estimator that down-weights "stale" measurements based on the variance of the intervening phase change.
4. Fisher Information and Resource Boundaries: The abstract Fisher information is connected to ideal balanced Poisson photon counting. This allows the authors to define dimensionless architecture boundaries in the plane of spatial information ($I_{sp}$) and phase-increment variance ($Q_\tau$).
5. Validation: The analytical laws are validated using exact nonlinear Monte Carlo simulations, which account for circular phase wrapping and higher-order drift effects.

Key Contributions and Results

• Fixed-Order Drift Penalty: For a fixed measurement order, the mean-square error (MSE) due to phase drift is derived as $3Q_\tau/8$ (where $Q_\tau$ is the variance of the phase increment). The error terms are explicitly $e_{C \to S} = -\delta_\tau \sin^2 \phi_0$ and $e_{S \to C} = -\delta_\tau \cos^2 \phi_0$.
• Adaptive Ordering Gain: By measuring the quadrature with the lower local Fisher information first, the uniform first-order MSE is reduced to $(3/8 - 1/\pi)Q_\tau$. This represents an 84.9% reduction in drift-induced MSE compared to the fixed-order approach.
• Information Loss in Stale Measurements: The paper derives that the Fisher information of a "stale" phase observation (measured at $t$ but used at $t+\tau$) is reduced from $I$ to $I/(1 + IQ_\tau)$. This provides a physical interpretation where the uncertainty of the phase increment acts as a noise source that degrades the utility of the earlier measurement.
• Architecture Crossover Conditions: The authors establish explicit dimensionless boundaries for when temporal readout is preferable to spatial readout. These boundaries depend on the information-gain factor $g$ (the ratio of total information in temporal vs. spatial modes). For an idealized symmetric case with $g=1.6$, the crossover point for fixed-order readout is $I_{sp}Q_\tau \approx 1.00$, while the increment-aware and adaptive strategies tolerate significantly higher phase variance before spatial readout becomes superior.
• Simulation Validation: Monte Carlo simulations confirm the perturbative laws in the small-drift regime. They also quantify the robustness of adaptive ordering to prediction errors; even with a prediction uncertainty of 0.1 rad, the adaptive strategy retains an ~83% reduction in error compared to fixed ordering.

Significance and Claims
The paper claims to provide a rigorous theoretical framework for designing receivers in programmable photonic processors that must operate under phase instability. Its significance lies in:

1. Quantifying the Cost of Sequentiality: It precisely defines the penalty of measuring quadratures sequentially, moving beyond qualitative statements about phase drift.
2. Resource-Aware Design: By linking Fisher information to detected photon numbers, it offers a physical resource interpretation that allows engineers to define "architecture boundaries" without assuming a specific time-scaling law for phase drift.
3. Practical Mitigation: It demonstrates that simple adaptive ordering rules can drastically reduce drift errors (by ~85%) without requiring complex feedback stabilization or simultaneous hardware for all quadratures.
4. Empirical Grounding: The theory is directly motivated by and validated against real-world measurements from an eight-mode processor, providing an empirical method to estimate the critical parameter $Q_\tau$ for specific hardware implementations.

The authors conclude that while temporal readout introduces phase inconsistency, it can be advantageous if the system gains sufficient information (e.g., through detector reuse or higher integration times) to offset the drift penalty, provided that adaptive or increment-aware estimation strategies are employed.