Homodyne versus heterodyne detection for classifying families of optical quantum states from finite records of detection events

AI summary

Overview

Research area: Quantum physics — quantum state identification (quantum state tomography and machine-learning-based state classification) for continuous-variable optical states, with a focus on comparing measurement schemes.

Technical level: Intermediate to Advanced. The paper assumes familiarity with phase-space functions (Wigner, Husimi Q), homodyne/heterodyne detection, noise channels, and convolutional neural networks.

Scope (one sentence): The paper compares homodyne versus heterodyne detection, at equal numbers of recorded detection events N, for classifying seven families of optical quantum states with neural networks, using simulated finite records, a Fisher-information analysis near thermal states, and low-power tests on real experimental homodyne data.

What This Paper Is About

Neural networks can name an optical quantum state from a picture of its phase-space function with very high accuracy, but a real experiment never delivers a clean picture — it delivers a finite list of N detector clicks. Heterodyne detection yields the Husimi Q function directly (as a histogram of outcomes), whereas homodyne detection only yields the Wigner function after a reconstruction step from data taken at several local-oscillator phases, and that reconstruction amplifies counting noise.

The authors ask a practical question: at the same number of recorded events, and under the same loss, dephasing, and Gaussian noise, which detection scheme better lets a network decide which of seven state families — including Gottesman–Kitaev–Preskill (GKP) states — a source is actually producing? They build the study so that the mean photon number carries no information about the family, so that only the shape of the phase-space distribution can distinguish them.

Key Contributions

1. A like-for-like comparison of finite records. The authors model realistic finite records for both schemes on identical states and identical noise: multinomial-binned histograms of the Husimi Q function for heterodyne, and histograms at K phases plus filtered back-projection (FBP) reconstruction into a Wigner function for homodyne. In the base set of seven families, all families share the same mean photon number, so energy alone cannot reveal a failure.

2. A quantitative accuracy map over N and over noise, for two different sets of seven families (the equal-energy "base set", and a second set built with the released code of Ahmed et al.), benchmarked against networks trained on exact phase-space images. The authors derive a rule for choosing the measurement from this comparison.

3. Separation of the measurement from the reconstruction. A homodyne classifier that reads the marginal histograms directly (no reconstruction, called the S representation), plus maximum-likelihood reconstruction, plus a second network architecture, plus controls: training on exact images, restricted training noise range, held-out states, and a drifting local-oscillator phase.

4. A detector-level theory and experimental checks. Fisher information for both detections near (possibly displaced) thermal states, mode by mode, plus error exponents of the optimal decision on the events, to separate the property of the detector from the property of the network. Finally, tests of the finite-record model on published homodyne records of a GKP state, of photon-subtracted squeezed states, and of one mode of an Einstein–Podolsky–Rosen source.

Main Findings

• Heterodyne wins only for short records. 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 (base) set, with an adjusted p = 0.003 over 24 cells across training runs. After Holm step-down correction, significant base-set advantages for Q − W remain at N = 100 under dephasing (adjusted p = 0.003), at 80 % loss (0.004) and at 50 % loss (0.021).

• No significant difference at larger N. At N = 10³ and 10⁴ no difference is significant, though the authors note the intervals are wide. No difference in the second set (the Ahmed et al. code-based families) survives correction.

• The reconstruction, not homodyne detection itself, is the culprit. At N = 100, a homodyne network that reads the histograms without any reconstruction does not differ significantly from heterodyne detection in the first set, and is nominally worse in the second. A likelihood-ratio observer shows no such gap at N = 100, so the authors state their evidence concerns a reconstruction-based pipeline, not homodyne detection in general.

• Reconstruction inflates the error and colors the noise. In the authors' implementation, at equal N and pixel size, the integrated squared error of the reconstructed Wigner estimate is 48 to 63 times that of the direct Q estimate; referred to the signal size, the noise-to-signal ratio of W is 15 to 28 times that of Q. Unlike the white counting noise of Q, the reconstruction noise grows with wave number up to a cutoff and scales as the inverse cube of the pixel pitch, versus inverse square for Q.

• Heterodyne measurement is a low-pass filter. Because Q is W convolved with the vacuum Wigner function, the heterodyne data are a low-pass-filtered version of W; in Fourier space the filter is exp(−|k|²/4). For an even cat state with |α| = 2, whose fringes have wave number k ≈ 5.7, heterodyne data retain a fraction ≈ 3 × 10⁻⁴ of the fringe amplitude. This filter belongs to the measurement and cannot be undone by a better estimator, since inverting it multiplies counting noise by exp(k²/4).

• Exact images show near-perfect accuracy for both representations. With networks trained and tested on exact images, noiseless states are classified at 100.0 % for both Wigner and Q. At 60 % and 80 % loss both reach 100.0 %; at 90 % loss they reach 99.9 % (W) and 92.6 % (Q); at 100 % loss every state is the vacuum and accuracy falls to the chance level 1/7. Dephasing and Gaussian noise (σ² = 0.2 and σ² = 1) both give 100.0 % for both representations.

• A local theory that does not explain the data. Near thermal states, the authors 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. They describe this as a local result that does not explain the finite-record differences.

• Experimental records behave like simulated ones, but only as a counting-noise check. Real homodyne records of a GKP state, photon-subtracted squeezed states and one EPR mode are classified like simulated records, but the authors stress these low-power tests probe their counting-noise model, not the seven-class accuracy.

Methodology in Plain English

• Seven families, one energy. The base set contains seven single-mode families truncated at n_max = 63 — including squeezed vacuum (what a failed herald would leave instead of a GKP state), Fock, thermal, coherent, cat, GKP and binomial-family states. Every state is assigned a mean photon number from {2, 3, 4, 5}, with those four values occurring at equal frequency in every family, so the average energy is uninformative by construction.

• Three noise channels, swept. Photon loss (quoted as p = 100(1 − η) %, contracting everything toward vacuum), dephasing (γ = 3p/100, damping Fock-basis coherences), and Gaussian noise (variance σ² = p/100 per quadrature, convolving the Wigner function). Training data cover every strength of every channel, and every family sees the same sequence of noise files, so noise cannot act as a label. Four representative cells are used for the head-to-head comparison: noiseless, 50 % loss, 80 % loss and 20 % dephasing.

• Records are simulated from exact images at 128 × 128 pixels on [−12, 12]², so both schemes see identical discretization. Heterodyne: multinomial counts drawn from the exact Q image, rescaled to be unbiased. Homodyne: the N events are spread as evenly as possible over K = 36 phases (at N = 100, 28 phases get three events and 8 get two), the marginals are discrete Radon transforms, and the image is produced by filtered back-projection with a ramp filter.

• Networks and training. A convolutional network with 2.2 × 10⁶ weights (four blocks of two 3×3 convolutions with batch normalization and ReLU, max-pooling, a 256-unit dense layer with dropout 0.5, seven outputs), trained with AdamW on cross-entropy, 25 epochs, batches of 64, learning rate 3 × 10⁻⁴. Separate networks are trained per representation and per regime, with several independent runs; 135 networks in total.

• Size of the data. The noise-sweep test set has 100 samples per family at each of 21 strengths (0, 5, …, 100 %) for each of three channels — 44,100 samples. The mixed training set has 19,950 samples and the validation set 4,200. A held-out test trains on ⟨n̂⟩ ∈ {2, 3, 4} and tests on ⟨n̂⟩ = 5.

• Statistics. 95 % Wilson intervals, plus a paired interval for the difference Q − W because all networks see identical records, with Holm adjustment over the 24 cells of each comparison. Two decision-rule-free measures are also reported: the area under the ROC curve (macro-AUC) and the expected calibration error (ECE). Non-neural baselines are multinomial logistic regression and k-nearest neighbors on PCA-reduced, block-averaged images, with hyperparameters chosen by five-fold cross-validation. An energy control classifies k-nearest neighbors (k = 50) using only an estimated mean photon number per record.

• A quantum benchmark. The best single-copy success probability for the loss channel is computed as a semidefinite program over the loss-averaged family states. The authors explicitly note this is a bound only for N = 1 and not for their smallest record of N = 100.

• Experimental validation. Published homodyne records of a propagating GKP state (19,610 to 21,252 events at each of six phases from 0° to 150°) are checked for identical-and-independent-event assumptions and compared against records simulated from a maximum-likelihood-reconstructed density matrix. A detector offset of c = −0.25 ± 0.01 is subtracted.

Why This Matters

This paper turns a vague piece of experimental folklore — "heterodyne gives you the Husimi function for free, homodyne needs reconstruction" — into a measured, quantified trade-off at fixed measurement budget. It tells an experimenter whether the extra reconstruction step buys anything when only a finite record is available, and it isolates the reconstruction (rather than the detection principle) as the source of the short-record penalty.

Real-world applications:

• Source verification in photonic quantum computing. Propagating GKP states are heralded at about 10 Hz with tomography using of order 10⁴ events per phase, while photon-subtracted squeezed states are generated at up to 0.9 MHz. A fast classifier that says "this is still my target state, not a false herald" is a practical quality-control step.
• Distinguishing a target state from its typical failures. A false herald leaves squeezed vacuum; loss blurs fringes; phase noise dephases. The seven-family setup is built explicitly as a target-versus-failure decision problem.
• Detector and budget planning. Because the paper's records are counted in events (a record takes time N/f for repetition rate f), the results translate directly into "how long must I measure, and with which detector."
• Quantum communication and sensing pipelines that need cheap, noise-robust state identification without full tomography, which is inefficient in general.

Industry relevance: The trade-off bears on hardware choices for continuous-variable and photonic platforms — heterodyne is realized in microwave experiments with phase-preserving linear amplifiers, while optical homodyne is the standard tool. Knowing that the reconstruction, not the detection, causes the short-record penalty is directly actionable for groups building state-verification firmware around existing detectors.

Future Directions

• Explaining the finite-record gap theoretically. The paper's own leading-order analysis near thermal states is explicitly described as a local result that does not explain the observed differences, leaving the finite-N, finite-K behavior without a first-principles account.

• Tighter intervals at larger N. At N = 10³ and 10⁴ no difference is significant but the intervals are wide, so a larger, better-powered study could establish whether the schemes truly converge or whether a small residual advantage persists.

• Better reconstruction than filtered back-projection. The authors note that K = 36 phases are far from the Nyquist condition of roughly 200 for their disc, producing streak artifacts even at N → ∞. They already probe maximum-likelihood reconstruction and a no-reconstruction classifier; systematically optimizing the reconstruction, the number of phases, and the pixel pitch is the natural next step.

• Experimental tests at useful power. The current experimental validation covers a GKP state, photon-subtracted squeezed states and one EPR mode, but the authors state these low-power tests probe the counting-noise model rather than the seven-class accuracy — a genuine high-power, seven-class experimental test remains open.

Target Audience

Experimentalists in continuous-variable and photonic quantum optics who must choose a detection scheme under a fixed measurement budget; machine-learning researchers applying classifiers to quantum measurement records; and theorists interested in the information-theoretic comparison of homodyne and heterodyne detection beyond the single-copy setting. Readers should be comfortable with phase-space representations and basic neural-network terminology; the paper is not an introduction to either.

Authors’ 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.