arXiv:2610.38671
Homodyne versus heterodyne detection for classifying families of optical quantum states from finite records of detection events
Franciszek Grzywa, Sebastian Dolata, Karol Bartkiewicz
Abstract
Neural networks can identify an optical quantum state from an image of its phase-space function, but a real measurement gives a finite record of $N$ detection events. Heterodyne detection samples the Husimi $Q$ function directly, whereas homodyne detection gives the Wigner function only after a reconstruction from several phases. We ask which scheme better identifies one of seven families of states, including Gottesman-Kitaev-Preskill (GKP) states, under loss, dephasing and Gaussian noise. We simulate finite records on identical states for two sets of families, one with equal mean photon number and one built from the code of Ahmed et al. For convolutional networks, heterodyne detection outperforms homodyne detection with filtered back-projection by 14.1 percentage points at $N=100$ events and 20% dephasing in the first set (adjusted $p=0.003$ over $24$ cells across training runs), and it depends on the number of phases and on the reconstruction. At $N=10^3$ and $10^4$ no difference is significant, but the intervals are wide. At $N=100$, a homodyne network without reconstruction does not differ significantly from heterodyne detection in the first set and is nominally worse in the second. Near thermal states we show to leading order that heterodyne detection damps phase-space structure of order $n$ exponentially, whereas homodyne detection shares unit weight among $n+1$ angular components, a local result that does not explain the finite-record differences. Experimental records of a GKP state, photon-subtracted squeezed states and one mode of an Einstein-Podolsky-Rosen source are classified like simulated ones, but these low-power tests probe our counting-noise model, not the seven-class accuracy. A likelihood-ratio observer shows no such gap at $N=100$, so our evidence concerns a reconstruction-based pipeline, not homodyne detection in general.
AI Summary
The paper asks whether homodyne or heterodyne detection better identifies optical-state families from a fixed number $N$ of measurement events. It simulates seven-class classification under loss, dephasing and Gaussian noise, using an equal-energy state set and a second set based on an earlier benchmark. Convolutional networks receive heterodyne histograms, reconstructed homodyne Wigner images, or homodyne histograms directly. Controls vary reconstruction, phase count, architecture, training range and held-out state parameters. At $N=100$, heterodyne classification exceeds filtered-back-projection homodyne classification by 14.1 percentage points under 20% dephasing, 6.5 points under 50% loss and 9.8 points under 80% loss in the equal-energy set. These differences survive seed-level Holm correction across 24 comparisons; significance weakens when cross-validation folds are the replication unit. At $N=10^3$ and $10^4$, no baseline difference is significant, without establishing equivalence. Direct homodyne classification shows no significant heterodyne deficit in the equal-energy set, and a likelihood-ratio observer knowing the prototypes shows no heterodyne advantage in those four $N=100$ conditions. The observed advantage therefore concerns reconstruction and classification, not the detector generally. A local Fisher-information calculation near thermal states resolves measurement sensitivity into phase-space modes: heterodyne weights decay exponentially with mode order, while homodyne weights sum to unity across angular components. This does not explain the finite-record network differences. Public experimental homodyne records test counting-noise consistency, not seven-class accuracy. Stated limitations include overlapping training/test states, unoptimized phase counts, few training runs, closed-set classification and mostly simulated validation.
to write and run the analysis code and the plotting scripts, to check the numbers in the text against the result files, and to draft and edit parts of the text.
Error found: Unresolved numerical placeholders prevent independent verification of the abstract's adjusted p=0.003, and the paper reports that the dephasing comparison loses adjusted significance when cross-validation folds replace training seeds as the replication unit. The likelihood-ratio appendix incorrectly describes Bayes-optimal observer error as an upper bound, whereas it is a lower bound on achievable error under the stated known-prototype model. Equation EQ incorrectly equates realized integrated squared histogram error with its expected sampling contribution and omits binning bias if the target is the continuous Q function.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 21 pages, 7 figures, 4 tables. Code and results: https://github.com/barkol/qmi-finite-shot