acceptodds
Under review as a conference paper at ICLR 2027

Quantum Sliced Wasserstein Embedding: Geometry and Learning from Homodyne Data

Abstract

Continuous-variable quantum optics underlies photonic computing, bosonic error correction and precision sensing. Information is obtained by homodyne detection of the quadratures of an optical mode, and each shot produces one sample from a quadrature distribution and destroys the state. Extracting what is needed from few shots is the central practical problem. We bring sliced optimal transport, developed in machine learning for exactly this kind of data, to the homodyne record, and compare states through their measured slices with no reconstruction. Our instrument is the quantum sliced Wasserstein embedding, a map from a homodyne record to a Euclidean vector whose distances are the sliced Wasserstein distance between quadrature marginals. That relationship gives the embedding a metric structure in which displacement and squeezing of the source are locally orthogonal. We confirm the geometry on a grid of displaced and squeezed states, whose multidimensional scaling image reproduces the displacement lattice at physical spacing and stretches along squeezing as the metric predicts. We then apply the embedding to cat states, a resource widely used in quantum computing, and regress their displacement and squeezing from shot records, comparing the error to the Cramer-Rao bound. Moment, energy-distance and kernel mean-embedding features of the same records serve as baselines. Under a linear map fitted on a few dozen labeled states the embedding comes within a factor two of the bound on both parameters and is the best feature at every shot budget. The kernel features match it only with a nonlinear regressor and ten times the training states. The embedding is therefore the representation to use when a Gaussian operation on a non-Gaussian state must be read from few shots and few labels.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
continuous-variable quantum optics, homodyne detection, quantum state tomography, Wigner function, Cramér-Rao bound, cat states, sliced Wasserstein distance, optimal transport

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RtFKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper maps a homodyne record to a Euclidean vector by sorting the shots at each measured phase and averaging them into equal-mass quantile bins, which is the Radon cumulative distribution transform applied to measured quadrature marginals. Euclidean distances between these vectors equal the empirical sliced Wasserstein distance between the recorded distributions. The authors prove that the population object is a metric and an isometric embedding, derive an exact displacement law, a closed form for displaced squeezed Gaussians and a local line element in which displacement and squeezing are orthogonal, show that reducing the reference size is an orthogonal projection, and give a finite-shot rate. Experiments report MDS embeddings of displaced and squeezed grids and a regression of displacement and squeezing on cat states against a nuisance-adjusted Cramer-Rao benchmark, with moments, energy distance and an RBF mean embedding as baselines.

Critical strengths
  1. Proposition 1 is exact and useful. A displacement shifts every quantile by a constant, so the embedding is affine in the displacement for any state, and the displacement readout is a fixed linear map that does not need refitting per state family. The grid experiment shows this survives 64 shots per phase, which is the kind of check that separates a geometric identity from a practical tool.

  2. The geometry experiment reports the population limit next to the finite-shot result, so representation error and sampling error are separated. Table 2 does this explicitly, and the note that the squeezing grid error of 0.101 persists at infinite shots, and therefore belongs to the distance and not to the estimator, is exactly the right way to present this.

Critical weaknesses
  1. The evaluation never compares against the estimator a practitioner would actually use. The primary testbed is a family with a closed-form likelihood, and Appendix D.4 shows a maximum-likelihood fit reaching 1.01 times the benchmark with no training states at all, while the pipeline sits at 1.6 to 1.9 times the benchmark. On the paper's own main experiment, the method is therefore less accurate than a zero-label estimator. The case where the embedding could pay off, a family with no closed form, is the Kerr cat in Appendix E.2, and that is where the main table should be.

  2. Table 4 does not isolate the representation, and two of the baselines look handicapped. The embedding's coordinates contain the per-phase sample mean as an equal-weight average over the bins, and the moment baseline contains it directly, so both can express the sufficient statistic for the displacement. Both nonetheless sit well above the benchmark, at 1.58 and 3.05 times it, and the feature set that contains the statistic most directly is the worse of the two. That ordering points at the ridge fit rather than at the geometry, so "best feature at every budget" may mean "least badly regularized".

Questions for authors
  1. Table 4, ridge regression, displacement. What error does a fixed-weight projection of the twelve phase means onto the displacement direction produce at 3072 copies? If it lands near the benchmark, Table 4 needs that arm, and the penalty path for each feature family should be reported.

  2. Table 1. The record carries two budgets, labeled states and shots. Why is the comparison run along the label axis, where each extra label costs one simulation, rather than along the shot axis, where the physics actually charges you?

  3. Figure 2 and Table 2. The grid error rewards linearity in the displacement and squeezing coordinates after an affine alignment, and the embedding is built to be affine in the displacement. What does the comparison with Bures or with the kernel establish beyond the fact that those distances are not parameter-linear?

  4. Proposition 4 gives a shot term of order n to the minus one quarter, while Section 6.1 reports the error falling faster than that from 16 to 4096 shots. Is the bound meant to guide the choice of the phase count and the shot budget, or only to establish a rate?

  5. Section 5 notes that with finitely many phases the distance is only a pseudometric. With twelve fixed phases, what can two states that agree on those twelve slices still differ in, and does the regression experiment inherit that blind spot?

  6. A drift of the local oscillator phase appears in the record as a displacement. For the sensing use case in Section 7, is there any way to separate a real displacement from a phase drift without reference shots?

5 comments

We thank the reviewer for their time and consideration, and for the careful and candid evaluation of our paper. The review is remarkably thorough, and its questions led us to new analyses that have substantially improved the paper. We are pleased that the reviewer finds Proposition 1 exact and useful. We believe this is what makes QSWE a practical tool rather than only a geometric identity, in the reviewer's own words. Below we address the points raised by the reviewer.

Response W1. We agree that a family without a closed form, such as the Kerr cat of Appendix E.2, is where the embedding can pay off, and that it belongs in the main table. In the revision, the main regression table will be the Kerr cat. The closed-form cat will remain as the controlled test bed, where a likelihood fit attains the floor because it is given the correct model. A second reason to use QSWE even when a likelihood is available is the cost of inference. Its readout is a fixed linear map, fitted once from labeled records, so estimating a new state costs a sort and a dot product. A likelihood fit instead solves a separate nonconvex optimization for every state. In our implementation this takes 0.2 to 0.6 s per state, against under 0.2 ms for QSWE.

Response W2. We ran the penalty path the reviewer asks for in Q1, and the ordering in Table 4 does not come from the ridge fit. We refit every feature set of Table 4 at every penalty on a grid from 10−1210^{-12} to 10410^{4}, against the paper's 10−810^{-8} to 1010, and recorded its test error at each. Even when the penalty is chosen in hindsight on the test states, which is more favorable than any procedure available in practice, no baseline comes close. QSWE at its validated penalty stays below every baseline at that baseline's best penalty, by 1.05 [0.94, 1.17] times the floor on displacement against the closest one, the energy CDF, and by 0.30 [0.22, 0.39] on squeezing against the Gaussian moments.

Feature setq, validatedq, best fixed penaltyq, best penalty per repeatζ, validatedζ, best fixed penaltyζ, best penalty per repeat
QSWE r = 161.58 ± 0.061.54 ± 0.051.52 ± 0.051.87 ± 0.091.86 ± 0.091.83 ± 0.08
QSWE r = n1.59 ± 0.051.58 ± 0.051.56 ± 0.051.95 ± 0.091.95 ± 0.091.94 ± 0.09
Gaussian moments3.05 ± 0.103.01 ± 0.092.97 ± 0.092.25 ± 0.072.18 ± 0.072.17 ± 0.07
Moments 1–66.18 ± 0.546.09 ± 0.495.45 ± 0.334.93 ± 0.404.72 ± 0.194.34 ± 0.18
Energy CDF2.79 ± 0.122.70 ± 0.122.63 ± 0.102.96 ± 0.142.90 ± 0.142.88 ± 0.13
RBF mean embedding3.00 ± 0.193.01 ± 0.192.78 ± 0.132.81 ± 0.132.74 ± 0.112.68 ± 0.10

Ridge regression at 3,072 copies and 48 labeled states, on the records of Table 4 (30 repeats). Entries are test RMSE divided by the square root of the Cramér–Rao floor, mean ± 95% interval. Validated: the penalty chosen on the 24 validation states, as in Table 4. Best fixed penalty: the single penalty with the lowest mean test error. Best penalty per repeat: the penalty with the lowest test error in each repeat. For the RBF features the bandwidth is chosen together with the penalty: on validation in the first column, as the single best bandwidth in the second, and per repeat in the third.

We used the records, splits and seeds of Table 4 and changed nothing but the penalty, so the fits at the validated penalties reproduce Table 4 exactly. We will add the penalty paths of every feature set to the appendix.

Response Q1. The per-phase average is the efficient statistic for a displacement only when each slice is Gaussian. For a coherent state, the projection of the twelve averages onto the displacement direction attains the Cramer-Rao bound exactly. A cat's slice along its lobe axis instead consists of two narrow peaks far apart; the noise of the average is set by their separation, while the information about a shift is set by the width of each peak, so the average discards most of that information. A linear map on QSWE is a weighted sum of the sorted shots, so it can weight them by rank and in effect locate each peak separately, as efficient estimators of location do (Chernoff, Gastwirth & Johns, 1967). We ran the reviewer's projection as a simple check on the records of Table 4. Fitting the twelve phase averages to qcos⁡θ+psin⁡θq\cos\theta+p\sin\theta, with no training, gives 3.00±0.093.00\pm0.09 times the floor on qq, far from the benchmark and the same as ridge regression on the Gaussian moments (3.05), which contain these averages. Ridge regression on QSWE reaches 1.58 from 48 labeled states, and an untrained weighting of the same coordinates derived from the model reaches 1.56. The penalty path is in our response to W2, and we will add this check to the appendix.

Response Q2. The comparison is run along both axes. Figure 4 holds the label budget at 48 states and varies the shot budget from 192 to 12,288 copies per state (16 to 1,024 shots at each of 12 phases), and Tables 4 and 5 give the full results at 3,072 and 768 copies. Along the shot axis the advantage of QSWE grows. At the smallest budget shot noise dominates and all feature sets lie within a few percent of one another; as the budget grows, QSWE read by ridge regression runs parallel to the Cramer-Rao floor, at about 1.6 times it for displacement and 1.9 times for squeezing, while the other feature sets flatten, so that at the largest budget their errors are two to three times that of QSWE. Table 1 adds the label axis at a fixed 3,072 copies. We agree that labels are cheap to obtain in some settings, and with plentiful labels the gap narrows as shown in Table 1. The kernel features with a nonlinear regressor match QSWE on displacement once given about ten times as many labeled states.

Response Q3. We agree that the grid error rewards linearity in the physical parameters, and this is precisely the property our representation is designed to exploit. The significance of this property goes beyond geometric alignment: it determines the complexity of the model required to recover physical parameters from the representation. Following the principle of Occam's razor, we favor representations that enable accurate parameter estimation with the simplest possible readout. When the parameters depend linearly on the embedding, a linear model suffices, providing a strong inductive bias and enabling sample-efficient learning. In contrast, representations with nonlinear parameter dependence generally require more expressive readouts, which can increase sample complexity and the risk of overfitting in low-data regimes.

The grid experiment therefore serves as an indicator of downstream learning complexity, rather than merely a measure of geometric linearity. This connection is supported by the regression results in Section 6.2: with only 48 labeled states, ridge regression on QSWE accurately estimates displacement and squeezing, whereas a nonlinear regressor on kernel features requires approximately ten times as many labeled states to match its displacement accuracy. Moreover, using a nonlinear regressor on QSWE actually worsens squeezing estimation compared with ridge regression. These findings support our central argument: the value of a representation lies not only in the information it preserves, but also in how simply and efficiently that information can be extracted.

Response Q4. The estimate provided in the proposition is the conservative upper bound. It assumes only that the quadrature distributions of the two states have a bounded moment of order q>4q>4, and at that generality n−1/4n^{-1/4} cannot be improved. The measured error falls faster because the distances between the states of Figure 3 are large compared with the shot noise, so the noise changes them only at second order. A faster rate in general needs more than moments. The standard n−1/2n^{-1/2} rate requires a condition on the density, which fails where the density vanishes, as at the nodes of a cat's fringes (Bobkov & Ledoux, 2019). We therefore use the bound only to establish a rate, and the faster decay measured in Section 6.1 reflects the particular states there rather than this worst case.

S. Bobkov and M. Ledoux, One-dimensional empirical measures, order statistics, and Kantorovich transport distances, Memoirs of the AMS 261(1259), 2019.

Response Q5. This is a very good question, and answering it has made the result sharper. Each homodyne slice shows the state from one direction in phase space, and turning the measurement phase reveals its angular structure. A coherence between Fock states kk photons apart makes the slices oscillate kk times as the phase goes once around. Twelve phases, together with the symmetry between opposite directions, view the state from 24 equally spaced directions. That is enough to resolve angular patterns that repeat fewer than 12 times around phase space, while faster patterns can be confused with slower ones or missed. Two states that agree on the twelve slices therefore share every quadrature moment of order below 12, including the mean and variance of every quadrature and the mean photon number, and they can differ only if they differ in some coherence between Fock states 12 or more photons apart. Among states with at most 11 photons, twelve phases determine the state (Leonhardt & Munroe, 1996). For the states of the regression experiment, the blind spot costs essentially nothing. They carry little population at 12 photons or more (a median of 5×10−45\times10^{-4}), and at the same number of copies, the Cramer-Rao floor with twelve phases is within 0.3% of the floor with all phases for every target parameter. We will add this analysis to the appendix.

U. Leonhardt and M. Munroe. Number of phases required to determine a quantum state in optical homodyne tomography. Physical Review A, 54(4):3682–3684, 1996. doi:10.1103/PhysRevA.54.3682

Response Q6. We agree with the reviewer, and we will make the sentence in Section 7 more clear. A drift of the local-oscillator phase rotates the state, and the record sees the resulting motion of its center as a displacement. Whether the two can be separated without reference shots depends on the probe. For a coherent probe they cannot, since a rotated coherent state is exactly a coherent state with a rotated displacement. A probe with an orientation, such as a squeezed state or a cat, carries its own reference. A drift turns its shape as well as moving its center, while a real displacement moves only the center and leaves the shape unchanged (Proposition 1). In QSWE the two changes are also orthogonal - a displacement adds the same constant to every quantile at a phase, whereas a change of the centered shape averages to zero over quantile levels. The drift can therefore be read from the shape part of the same record and removed from the displacement part, provided the probe's orientation is tied to the frame in which the signal is defined. Appendix E.1 shows that a rotation of a squeezed state's axis is visible in the running embedding, although there the Gaussian likelihood monitor detects it sooner. We have not built such an estimator and see it as a direct follow-up.

Reviewer_B2P3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces quantum sliced Wasserstein embedding (QSWE), which represents quantum states directly from homodyne measurement records by sorting samples, averaging them into quantile bins, and concatenating the results into a Euclidean vector. Building on classical quantile-based transport representations, it establishes an equivalence with empirical sliced Wasserstein distance at full resolution, derives exact displacement geometry and local displacement–squeezing orthogonality, and analyzes compression and finite-sample error. Validation uses simulated state grids and parameter regression on non-Gaussian cat states, with moment, energy-distance, and kernel features as baselines. With 48 labeled training states and 3,072 measurements per state, QSWE with ridge regression estimates displacement and squeezing at approximately 1.6 and 1.9 times the square root of the nuisance-adjusted classical Cramér–Rao bound; substantial feature compression preserves performance in these experiments. The results support QSWE’s usefulness for learning physical transformations with few labels, while its weaker performance on shape parameters and exclusively simulated validation limit broader conclusions.

Critical strengths

The paper introduces the quantum sliced Wasserstein embedding (QSWE), a representation of continuous-variable quantum states constructed directly from homodyne measurement records. It applies optimal transport to the measured quadrature distributions, which are nonnegative probability distributions even when the underlying Wigner function contains negative values. This provides a natural way to compare quantum states without reconstructing their full phase-space representation. The connection between experimentally accessible measurements and transport-based representations is interesting, particularly because the resulting embedding has interpretable geometric properties. Displacement acts affinely on the embedding, while displacement and squeezing correspond to orthogonal directions in the derived local geometry. These properties provide a useful theoretical foundation for learning physically meaningful parameters from measurement data.

Critical weaknesses

The empirical advantages of QSWE are demonstrated within a relatively restricted evaluation setting. The main learning comparisons focus on alternative feature representations followed by linear or nonlinear regression, so they primarily establish the effectiveness of QSWE within this shared framework. Broader comparisons with methods that learn directly from measurement records would help clarify its advantages as a general estimation approach. The appendix’s model-based maximum-likelihood estimator also achieves lower errors when the correct parametric model is available, highlighting the importance of specifying the intended application setting. Furthermore, QSWE’s comparative advantage decreases as the number of labeled training states grows: nonlinear regression on kernel features becomes competitive and matches its displacement-estimation performance at the largest tested training size. Its performance is also weaker for cat-state shape parameters, such as coherence visibility and fringe phase. These results support a focused claim of effective displacement and squeezing estimation with limited labels, while leaving its broader usefulness less established.

Questions for authors

How does QSWE compare with models trained directly on homodyne records under matched measurement and label budgets?

3 comments

We thank the reviewer for their time and consideration, and for the candid evaluation of our paper. We are pleased that the reviewer finds the connection between experimentally accessible quantum measurements and transport-based representations interesting. We believe this connection has the potential to attract broad interest from both the quantum computing and optimal transport communities. We are also glad that the reviewer sees the geometry of the embedding as a useful foundation for learning physical parameters. Below we address the points raised by the reviewer.

Response W. We thank the reviewer for this candid point. QSWE creates an embedding that naturally fits the measurement geometry of quantum detection, making it a suitable representation for learning. We address the reviewer's first point, a comparison with models that learn directly from the measurement records, with new experiments in our response to Q1. On the second point, the likelihood fit reaches the Cramer-Rao floor because it is given the correct parametric model, and the paper uses it to show that the floor is attainable. QSWE needs neither a likelihood nor an optimization at inference. Its readout is a fixed linear map and estimating a new state costs a sort and a dot product. The likelihood fit instead solves a separate nonconvex optimization for every state, which takes 0.2 to 0.6 s per state in our implementation against under 0.2 ms for QSWE. When the family has no closed form, as for the Kerr cat of Appendix E.2, the likelihood has to be built numerically, whereas QSWE needs only labeled records. On coherence visibility and fringe phase, these parameters change the shape of a slice rather than moving or stretching it, so they are not transport parameters, and QSWE on its own is less suited to them. Combining QSWE with a feature that is sensitive to shape is a natural way to cover them, and Appendix E.4 takes a first step in that direction. QSWE is designed for the parameters that move or stretch the quadrature distributions, displacement and squeezing, which are the quantities of interest in homodyne-based sensing, phase tracking and the characterization of squeezed sources. We will state this setting explicitly in the introduction.

Response Q1. The two approaches suit different regimes of labeled data. QSWE builds the geometry of the measurement into the representation. Because it is a transport embedding, the parameters can be read by a linear map, which few labeled states suffice to fit. A network trained directly on the records has to learn this structure from the labels, and with 10410^4 to 10610^6 weights it overfits when labels are few, which is the regime QSWE targets. With many labeled states, a flexible network is no longer limited by overfitting, and we expect it to close the gap and eventually model dependences that a linear readout cannot. The experiments below compare the two in both regimes.

On the records of Table 4 (3,072 copies per state), with the same labeled, validation and test states, we trained neural networks that read the record directly at 48, 160 and 480 labeled states. At the paper's budget of 48 labeled states these networks, with thousands to millions of weights, overfit, as expected. With more labels the networks improve, but QSWE given the same labels remains more accurate at every label count up to 480. At 10,000 labeled states, the two best networks no longer overfit, with training and test errors that nearly coincide, and they reach an accuracy comparable to QSWE's. The paired differences lie within their 95% intervals on both parameters, over three repeats.

LabelsMethodqqζ\zetaqq, train / testζ\zeta, train / test
480QSWE, ridge1.17 ±\pm 0.051.24 ±\pm 0.060.034 / 0.0380.113 / 0.135
480MLP1.44 ±\pm 0.071.39 ±\pm 0.100.030 / 0.0470.076 / 0.151
480CNN1.63 ±\pm 0.101.43 ±\pm 0.070.047 / 0.0530.117 / 0.156
10,000QSWE, ridge1.23 ±\pm 0.271.15 ±\pm 0.310.037 / 0.0410.123 / 0.124
10,000MLP1.26 ±\pm 0.311.16 ±\pm 0.290.038 / 0.0410.113 / 0.125
10,000CNN1.43 ±\pm 0.391.10 ±\pm 0.220.042 / 0.0470.111 / 0.119

Cat-state regression at 3,072 copies per state; QSWE at r=16r = 16; MLP and CNN: a multilayer perceptron and a one-dimensional convolutional network on the sorted record, the two most accurate networks. Columns 3-4: test RMSE divided by the square root of the Cramer-Rao floor. Columns 5-6: RMSE as a fraction of the prior standard deviation on the training and test states. Mean ±\pm 95% interval over 10 repeats at 480 labels and 3 at 10,000; each label count has its own test states. Network weight decay, early stopping and, at 480 labels, architecture variant chosen on validation states; predictions averaged over initializations.

Once again, we thank the reviewer for their constructive feedback, which led to the new comparison with networks trained directly on the homodyne records. We would welcome the opportunity to address any additional comments or concerns the reviewer may have before the rebuttal deadline.

I would like to thank the authors to address my concerns. This improves my understanding on the estimation of the exponential setup and comparison with end-to-end neural networks. The linear representation does outperform the NN learning in some setting, so I will keep my current rating.

Reviewer_X25xOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper constructs a quantile-based representation of homodyne measurements, characterizes its displacement and squeezing geometry, and evaluates its ability to estimate physical parameters from limited measurements and labeled states.

Critical strengths

(1) The theoretical analysis connects directly to the learning task and the embedding’s geometry explains why simple predictors can recover displacement and squeezing efficiently.

(2) The empirical evaluation is carefully controlled, including shared measurement budgets, unknown nuisance parameters, repeated trials, nonlinear readouts, and a likelihood-based check of the Cramér–Rao benchmark.

Critical weaknesses

(1) The methodological novelty seems incremental. Quantile representations, their Wasserstein geometry, and their favorable behavior under translations build on established CDT/Radon-CDT constructions. The contribution is primarily their quantum-measurement interpretation, analysis, and validation.

(2) The learning evidence remains concentrated on simulated, low-dimensional state families. The loss/noise and Kerr-cat experiments are useful, but do not fully establish robustness to state-family mismatch or phase-calibration errors.

(3) The preservation of regression accuracy under reference compression is empirical. The contraction result alone does not guarantee preservation of parameter-estimation information.

Questions for authors

(1) What kind of results constitute the main advance beyond existing CDT/Radon-CDT theory and previous transport-based comparisons of optical states?

(2) Does the low-label advantage persist when the test-state family or measurement phases differ modestly from those used for training?

(3) Can the reference-size analysis relate compression error more directly to parameter-estimation accuracy?

4 comments

We thank the reviewer for their time and consideration, and for the candid evaluation of our paper. We are pleased that the reviewer finds that our theory connects directly to the learning task, and believe the link between the geometry of sliced optimal transport and quantum measurements would be of interest to the community. We are also glad that the reviewer finds the experiments carefully controlled. Below we address the points raised by the reviewer.

Response Q1 and W1. We agree with the reviewer that the novelty of this work lies in its connection to quantum measurement rather than in advances to Radon-CDT theory. The Radon-CDT is an established classical construction (Kolouri et al., 2016; Park et al., 2018a). It has been applied in classical optics, where the projections are computed from a nonnegative camera image (Park et al., 2018b), and transport distances have been applied to optical tomograms as diagnostics of nonclassical states (Paul et al., 2025). Our contribution is to use the sliced Wasserstein construction as an embedding for learning in a quantum setting, where quantum mechanics itself imposes the slicing geometry. Each homodyne shot measures a single quadrature. While the underlying Wigner function is not necessarily a probability density function, each shot provides a sample from a valid one-dimensional probability distribution, making Sliced Wasserstein a natural and mathematically well-founded framework for comparing quantum states directly from experimentally accessible measurements.

Response Q3 and W3. We agree that the contraction of Proposition 3 alone does not guarantee that estimation information is preserved. It is possible to analyze the relationship between reference size and estimation theoretically, and we will add the more detailed result to the appendix. A linear readout of QSWE is a weighted sum of the sorted shots, and compressing to rr bins forces the weights to be equal within each bin. We call the Fisher information that this constraint discards the information loss. It is zero when the weights of the best linear readout are already constant within each bin, and it grows with how much they vary there. The compression error instead measures how much the sorted shots themselves vary within a bin, so the two can separate. For a coherent state, for example, the best estimate of a displacement is the mean of the shots at each phase, which weights them equally, so a single bin per phase has no information loss although its compression error is large. For the cat states of Section 6.2 we computed the information loss from the exact tomograms of all test states, with the other parameters free as in the paper, and report it as the Cramer-Rao floor of the compressed embedding relative to that of the full record. At r≤2r\le2 the information loss is large, with a floor 2.8 times the full-record floor in RMSE for displacement and far higher for squeezing. From r=8r=8 on it is small, within 8% for displacement and 3% for squeezing, and within 0.5% for both from r=64r=64, while the compression error still falls by a factor of 20 between r=8r=8 and r=64r=64.

Whether a larger rr helps the readout depends on a second effect. The readout has 12r12r coordinates fitted from a fixed number of labeled states, so refining the bins reduces the approximation error, the information that compression discards, but increases the estimation error of fitting the readout. Once the information loss is small, further refinement adds coordinates to fit without adding usable information, so the estimation error can dominate and the test error may even increase. The reference size is therefore a tradeoff between these two errors, and the floor of the compressed embedding indicates where refining the bins stops paying.

Response W2 and Q2. While we acknowledge the inherent limitations of purely simulated experiments, we can assess the method's robustness by incorporating imperfections characteristic of real homodyne measurements, particularly by evaluating its performance under degradations not encountered during training.

When out-of-distribution test samples differ from training data in the degree of displacement or squeezing, a linear model operating on a transport-based representation can naturally extrapolate beyond the training range, provided the transformation acts within a fixed linear subspace of the representation. Shifat-E-Rabbi et al. (2021) exploited this property in a Radon-CDT classifier by incorporating the translation directions cos⁡θ\cos\theta and sin⁡θ\sin\theta into each class subspace, demonstrating strong out-of-distribution generalization. Proposition 1 establishes the same property for displacement in our setting. Although the embedding is not linear with respect to squeezing, it varies smoothly with the squeezing parameter, suggesting that a linear readout can extrapolate moderately beyond the training range, albeit with increasing approximation bias as the extrapolation distance grows.

For other distribution shifts, such as changes in the underlying state parameters or measurement phases, it is difficult to characterize theoretically whether the resulting representations remain within the subspaces learned during training. We therefore evaluate their effects empirically. As shown below, QSWE consistently outperforms the competing methods across all tested shifts. While its normalized estimation error increases under some shifts, most notably for squeezing estimation when the cat-state amplitude falls outside the training range, its relative advantage persists across all evaluated conditions.

Test conditionq, QSWEq, Momentsq, Energyq, RBFζ, QSWEζ, Momentsζ, Energyζ, RBF
No shift1.58 ± 0.073.05 ± 0.132.75 ± 0.162.89 ± 0.181.90 ± 0.132.24 ± 0.083.00 ± 0.192.89 ± 0.16
Amplitude α ∈ [1.9, 2.3]1.80 ± 0.123.84 ± 0.174.14 ± 0.575.08 ± 0.733.95 ± 0.484.92 ± 0.507.07 ± 0.576.54 ± 0.56
Detector noise ν = 0.051.54 ± 0.072.89 ± 0.132.73 ± 0.152.96 ± 0.161.80 ± 0.122.43 ± 0.102.86 ± 0.162.80 ± 0.16
Detector noise ν = 0.11.51 ± 0.072.77 ± 0.132.77 ± 0.143.13 ± 0.161.76 ± 0.122.96 ± 0.162.93 ± 0.172.95 ± 0.19
Phase offset 1°1.59 ± 0.073.05 ± 0.142.75 ± 0.152.91 ± 0.181.92 ± 0.142.29 ± 0.093.02 ± 0.192.92 ± 0.16
Phase offset 3°1.69 ± 0.083.21 ± 0.232.85 ± 0.153.08 ± 0.192.06 ± 0.172.62 ± 0.213.14 ± 0.203.06 ± 0.18
Phase offset 7.5°2.24 ± 0.194.05 ± 0.683.52 ± 0.254.02 ± 0.392.76 ± 0.424.51 ± 0.863.93 ± 0.504.00 ± 0.52
Phase jitter 1°1.58 ± 0.073.06 ± 0.142.76 ± 0.152.90 ± 0.171.95 ± 0.152.32 ± 0.123.01 ± 0.182.91 ± 0.16
Phase jitter 3°1.67 ± 0.103.17 ± 0.192.85 ± 0.163.08 ± 0.152.23 ± 0.352.87 ± 0.403.27 ± 0.243.14 ± 0.29

Cat-state regression at 3,072 copies per state and 48 labeled states, with QSWE at reference size r = 16 and the moment, energy-distance and kernel mean-embedding baselines of the paper, all read by ridge regression. Every readout was fitted, and its penalty chosen, on unshifted states only, then applied unchanged to the shifted test states. Entries are test RMSE divided by the square root of the Cramér–Rao floor of the record actually measured; mean ± 95% interval over 20 repeats. Under detector noise the floor itself rises, so a lower entry there does not mean a lower error.

We conducted this experiment using the cat-state regression task of Section 6.2, retaining the measurement records, data splits, and random seeds of Figure 4 at 3,072 copies per state. As in the paper, the readouts were trained on 48 labeled states, with regularization parameters selected using 24 validation states, and then evaluated without retraining on the same 96 test states under four types of distribution shift: (i) cat-state amplitudes outside the training range (α∈[1.9,2.3]\alpha\in[1.9,2.3] instead of [1.5,1.9][1.5,1.9]); (ii) Gaussian detector noise with variance 0.05 or 0.1, relative to a vacuum variance of 1/21/2; (iii) global phase offsets of 1∘1^\circ, 3∘3^\circ, or 7.5∘7.5^\circ, the latter corresponding to half the phase spacing; and (iv) independent phase jitter of 1∘1^\circ or 3∘3^\circ. Under phase perturbations, all methods use the nominal measurement phases, and evaluations are paired with their unshifted counterparts.

We defined persistence of QSWE's advantage a priori as requiring the paired 95% confidence interval for its error difference relative to each baseline to lie entirely below zero. This criterion is satisfied across all eight shifted conditions for both parameters at r=16r=16, and at r=nr=n in all but the squeezing estimation task under the amplitude shift.

The amplitude shift presents the greatest challenge because the effects of amplitude and squeezing are partially confounded. The cat-state lobes are located at ±2αe−ζ\pm\sqrt{2}\alpha e^{-\zeta} along the cat axis, so increasing the amplitude or decreasing the squeezing produces similar broadening of quadrature distributions near that axis. A linear readout trained on α∈[1.5,1.9]\alpha\in[1.5,1.9] can partially disentangle these effects using measurements near the conjugate axis, but extrapolation beyond this range introduces a systematic bias in the squeezing estimate. Nevertheless, under this shift, QSWE's displacement error increases by only 10%, compared with 22%–70% for the baselines. Squeezing estimation is more sensitive across all feature sets, with errors approximately doubling and estimates exhibiting systematic bias.

We also conducted a second experiment in which measurement and state-preparation imperfections were present in both training and test data, reflecting the practical setting where a readout is calibrated using measurements from an imperfect experimental source. Extending the loss and noise analysis of Appendix E.3, we considered impure state preparation, detector inefficiency, electronic noise, and per-shot phase jitter. Specifically, each state was assigned an impurity level of up to 20%, modeled by mixing the cat state with a displaced squeezed vacuum sharing the same (q,p,ζ)(q,p,\zeta) parameters, together with per-shot phase jitter of standard deviation 10 mrad, detector efficiency η∼U[0.85,1]\eta\sim U[0.85,1], and electronic noise variance ν∼U[0,0.05]\nu\sim U[0,0.05]. These nuisance parameters were unknown to all methods, and the readouts were trained and evaluated on states drawn from the same imperfect source.

Despite these imperfections, QSWE consistently achieved the lowest estimation errors for both parameters, with RMSEs corresponding to 6.9% and 23.8% of the prior standard deviation for displacement and squeezing, respectively, compared with 10.1% and 27.4% for the best-performing baselines. For reference, QSWE achieved 5.2% and 20.6% on clean measurements. These results demonstrate that QSWE retains its advantage even when trained and evaluated on realistically imperfect measurements. We will include these additional experiments in the revised appendix.

Once again, we thank the reviewer for their constructive feedback, which led to the new robustness experiments and the analysis relating reference compression to estimation accuracy. We would welcome the opportunity to address any additional comments or concerns the reviewer may have before the rebuttal deadline.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.