arXiv:2609.24915
Fisher-information training of optical sensing front ends from natural fluctuations
Abhinav Sinha, Kai Wang
Abstract
Fisher-information-based optimization of programmable optical front ends in situ generally requires a response model or parameter-labeled measurements. We show that natural parameter fluctuations can instead provide the training signal. We construct proxies for the Fisher information (FI) of one parameter and the prior-weighted FI spectrum of multiple parameters from measured covariances after subtracting conditional detection noise, assuming locally affine responses. We use simulated photon counts from two incoherent point sources to train a mode sorter that approaches the quantum limit for separation and the Nagaoka--Hayashi single-copy bound for joint centroid--separation estimation. We demonstrate count-only updates that require no instrument response model and achieve performance comparable to model-based gradients. Our results provide a route to adaptive optical measurements without controlled parameter scans.
AI Summary
The paper asks whether programmable optical measurements can be trained to maximize Fisher information (FI) without a response model, parameter-labeled data, or controlled parameter scans. It uses naturally fluctuating parameters as the training signal: subtracting conditional detection noise from measured outcome-fraction covariances isolates parameter-induced covariance. For locally affine responses, probability whitening connects this covariance to FI. For multiple parameters, its nonzero eigenvalues approximate those of $\Sigma_{\rm pr}^{1/2}F\Sigma_{\rm pr}^{1/2}$, enabling an A-optimal objective approximating $\operatorname{Tr}(\Sigma_{\rm pr}^{-1}F^{-1})$. This recovers a weighted spectrum, not the full FI matrix. The method is tested using simulated multinomial photon counts from two equally bright incoherent point sources with Gaussian point-spread functions. Programmable mode sorters are trained using either model-based gradients or count-only simultaneous-perturbation updates. For separation estimation, two detection outcomes attain at least $99.9\%$ of the quantum FI with model-based gradients and $94.1$–$99.9\%$ with count-only updates. For joint centroid–separation estimation, nine outcomes attain $97.6$–$99.8\%$ and $98.0$–$99.4\%$, respectively, of the Nagaoka–Hayashi single-copy benchmark, assessed using the exact FI of the trained measurement. All demonstrations are numerical. Operation requires approximately fixed parameters within each shot, stable fluctuation statistics across training evaluations, and modeled or calibrated conditional noise. Narrower fluctuations reduce nonlinear bias but weaken the covariance signal; larger batches reduce sampling error, not systematic bias. The proxy becomes inaccurate near stationary responses. Count-only training requires larger photon budgets, and information in parameter-dependent total detection rates is excluded.
Claude Opus 5 (Anthropic) was used to write, modify, and optimize the simulation code behind the numerical results and figures reported here;
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 11 pages main text, 20 pages supplemental material