Heterodyne Data as a Physical Berezin Transform: Calibrated Optimal-Transport Tests for Continuous-Variable Quantum States

Suppose a lab wants to know whether two sources of light produce the same quantum state. The standard measurement is heterodyne detection. It records one point in the two-dimensional "phase plane" per copy of the state, and each point is drawn from a smooth density called the Husimi Q-function. This density is exactly the Berezin symbol of the state: the diagonal of the state operator in the basis of coherent states, which are the minimum-uncertainty wave packets of the oscillator. Put another way, the measurement already applies a Berezin transform, an averaging operator built from a reproducing kernel. That operator blurs the underlying phase-space picture by a Gaussian whose width is set by vacuum noise. It is also covariant under the phase-space symmetries: it commutes with displacements and rotations of the plane. Statisticians have spent a decade showing that smoothing is exactly what makes optimal transport (OT, the cost of moving one probability distribution onto another) statistically tractable. With smoothing, error rates stop depending on dimension, limit laws exist, and tests can be calibrated. Here nature supplies the smoothing, with a known kernel and a known scale.

The proposal is to build and analyze a two-sample test and a confidence interval for a transport-based distance between continuous-variable quantum states, computed directly from raw heterodyne samples. The regularization scale is fixed by the physics of the Berezin transform rather than tuned. The null distribution comes from second-order limit theory for entropic OT. There is also a curved variant on the disc, where the natural state space carries Möbius symmetry and Jaesung Lee's spectral results on invariant Berezin transforms describe which differences between states survive the smoothing. The payoff is an inference tool whose effect size has a physical meaning, which trace distance and kernel statistics lack. It needs no tomography and no hand-picked bandwidth.

Core Claim

Write $Q_\rho(\alpha) = \langle \alpha | \rho | \alpha \rangle / \pi$ for the Q-function of state $\rho$, where $|\alpha\rangle$ is the coherent state centered at the phase-space point $\alpha \in \mathbb{C}$. Heterodyne detection on $n$ copies returns $n$ independent draws $\alpha_1,\dots,\alpha_n \sim Q_\rho$. Our claim has three parts.

First, the entropic OT cost between two Q-functions, with the entropic parameter $\epsilon$ set to the vacuum variance and a displacement-invariant quadratic cost, is a well-defined, dimension-free estimand. The Q-functions of finite-energy states have subgaussian tails, so the parametric-rate and central-limit results of Statistical bounds for entropic optimal transport cover it. That paper proves $O(n^{-1/2})$ error and asymptotic normality for entropic OT under subgaussian marginals.

Second, under the null $\rho = \sigma$, the debiased (Sinkhorn) divergence $\bar S_\epsilon$ between two heterodyne samples obeys

$$ n\,\bar S_{\epsilon}(\hat Q^{(1)}_n, \hat Q^{(2)}_n) \xrightarrow{d} \chi_{Q_\rho}, $$

where $\chi_{Q_\rho}$ is a non-Gaussian limit that depends on the state. This is the second-order law that Limit theorems for entropic optimal transport maps and Sinkhorn divergence proved for compactly supported measures. We will extend it to the non-compact, subgaussian Q-function setting. We will also exploit the fact that the Berezin smoothing makes the relevant potentials real-analytic, which should simplify the Hölder-space differentiability argument.

Third, for Gaussian states the population quantity has a closed form. A Gaussian state with covariance $V$ produces heterodyne data with covariance $V + I/2$, so transport between two such data distributions reduces to transport between two Gaussians. We can therefore compare our estimand with the quantum transport distance of Quantum Optimal Transport with Quantum Channels, which computes optimal plans between thermal Gaussian states as attenuator and amplifier channels and shows the distance recovers classical OT in the semiclassical limit. This gives the effect size a physical reading: a heterodyne-computable surrogate that tracks the quantum transport distance.

What is new is not "OT on quantum data" in the abstract. It is the identification of the measurement itself as the smoothing operator that entropic and smooth OT theory requires, together with a calibrated null law and a local-power analysis built on that identification.

Research Program

Flat phase space. The first strand treats a single optical mode or a few modes, where phase space is $\mathbb{C}^m$ and the Berezin transform is the Gaussian heat semigroup at fixed time. We will prove the non-compact null limit and a bootstrap-consistency result. The $m$-out-of-$n$ bootstrap validated in Limit theorems for entropic optimal transport maps and Sinkhorn divergence is the starting point. The route follows the earlier Kantorovich-duality linearization of Central limit theorems for general transportation costs, which supplies the first-order method when $\rho \neq \sigma$.

We will then derive local power against alternatives $\sigma = \rho + t\Delta/\sqrt n$. The Berezin map from states to Q-functions is injective, so every alternative is detectable in principle. Its attenuation, however, is explicit in the coherent-state kernel: sub-vacuum-scale phase-space features, such as the interference fringes of a cat state, are strongly damped. Making this attenuation quantitative tells experimenters in advance how many copies a given alternative needs. Smooth p-Wasserstein limit theory shows that Gaussian smoothing before transport removes the curse of dimensionality. We supply the same kind of analysis where the smoothing is physical and cannot be chosen.

Curved phase space (disc). The second strand treats states whose coherent-state manifold is curved, namely SU(1,1) (Perelomov) coherent states on the unit disc. These arise for squeezing and two-mode systems. Here the Q-function is a weighted Berezin transform $T_\alpha$ on $\mathbb{D}$, with the weight fixed by the representation, and the symmetry group is the Möbius group $\mathrm{Aut}(\mathbb{D})$. We will use the hyperbolic squared distance as the transport cost, so the test is invariant under the symmetry that squeezing implements.

Lee's results then play a constraining role. A CHARACTERIZATION OF M-HARMONICITY shows that the spectrum of the Berezin transform on radial functions meets the unit circle only at 1. So under repeated smoothing, which physically corresponds to cascaded noisy symmetric channels, the only directions preserved are the invariant-harmonic ones. Iterates of weighted Berezin transform under invariant measure in the unit ball shows that $L^p$ differences wash out for $p>1$ while radial $L^1$ mass persists. We will turn these facts into a prediction: power against a perturbation decays with channel depth at a rate set by the Berezin spectrum. Radial "mass-shift" alternatives keep non-vanishing power, and localized alternatives do not.

The counterexamples in Weighted Berezin transform in the polydisc warn that invariance alone does not force harmonicity in $L^p$. We therefore state the power results under the radial-boundedness conditions Lee uses, and do not claim more.

Physical anchoring. The geometric-quantization literature motivates treating coherent-state symbols as the canonical classical shadow of a state. Generalized Berezin quantization, Bergman metrics and fuzzy laplacians builds Berezin and Toeplitz lifts on Kähler manifolds and shows that the choice of kernel metric matters away from homogeneous spaces. Supersymmetric ground states and Heisenberg algebras finds that Segal–Bargmann spaces, the natural home of the flat-case Q-function, arise as protected ground-state spaces. These works motivate the setting. None of them involves sampling, which is where our contribution sits. We restrict to flat and disc phase spaces because those are the settings where Lee's spectral theorems are actually proved.

Baselines

The naive reference point is "reconstruct, then compare." We would run maximum-likelihood tomography in a truncated photon-number basis on each sample, compute the trace distance between the reconstructions, and calibrate with a parametric bootstrap from the pooled reconstruction. This is what many labs do. It uses the same heterodyne data, the same total copy count, and a truncation chosen by held-out likelihood. Sample-optimal tomography of quantum states shows that independent product measurements such as heterodyne pay a dimension-factor penalty for full reconstruction. Testing that goes through tomography should therefore be wasteful, and beating it means higher power at equal size across the tested alternatives.

The strongest existing competitor is a kernel two-sample test, the maximum mean discrepancy with a Gaussian kernel, applied to raw heterodyne points. It shares our degenerate $1/n$ null scaling, needs no reconstruction, and Sinkhorn divergences interpolate toward it as $\epsilon$ grows. No paper in the file applies it to heterodyne data. We give it the same samples and calibration budget: permutation null, with the bandwidth chosen by the median heuristic and additionally tuned on held-out data with the same number of tuning trials we allow ourselves. There is no reported result found on Lacuna for kernel or OT two-sample tests on heterodyne data, so both methods are run fresh.

The closest OT baseline is the Euclidean Sinkhorn divergence with its $\epsilon$ freely tuned and the $m$-out-of-$n$ bootstrap null of Limit theorems for entropic optimal transport maps and Sinkhorn divergence. That paper verifies $n^{-1}$ null scaling on the uniform law on $(0,1/2)^2$ with $n=2500$ and $5000$, but reports no numerical error value. We also include the unregularized empirical Wasserstein distance with a permutation null.

The key ablations isolate our two ingredients. The first replaces the physics-fixed $\epsilon$ with a tuned one. The second, on the disc, replaces the hyperbolic cost with the Euclidean cost. Beating the baselines means a combination of three things: valid size, meaning type I error within Monte Carlo error of 5%; power at least matching the tuned kernel test on non-Gaussian alternatives with no tuning; and a confidence interval whose coverage holds for the transport effect size, which the kernel test does not provide.

Why not just…?

Why not just use trace distance or fidelity after tomography? Trace distance is unitarily invariant and treats a tiny displacement and a gross error alike once the states become nearly orthogonal. The quantum Wasserstein distance of order 1 documents this loss of locality in the qubit setting. Tomography also inherits the product-measurement penalty of Sample-optimal tomography of quantum states. The tomography baseline measures directly how much power this costs.

Why not just run a kernel (MMD) two-sample test on the heterodyne points? This is the real threat. For a pure yes/no test it may match us in power, and we will report that honestly if it does. Our case rests on two differences that the experiments check. The first is the calibrated effect size: for Gaussian states our estimand can be compared with a closed-form quantum-transport quantity, and its interval should cover it, while MMD has no such physical reading. The second is freedom from bandwidth tuning, with covariance under the disc symmetry. If MMD also matches our power on the disc benchmarks and on the effect-size coverage, the contribution shrinks to the limit theory. We will say so.

Why not just test means and covariances? For Gaussian states this is optimal and should win. It is blind to non-Gaussian differences, such as a Fock state versus a thermal state with the same second moments. The moment-matched non-Gaussian benchmarks exist to show exactly this.

Why not just use sliced Wasserstein distances? One-dimensional projections are cheap and have clean central limit theorems (Distributional Convergence of the Sliced Wasserstein Process). On the disc, however, linear slices do not commute with Möbius maps, so the test would not be symmetry-invariant. We include a sliced test in the disc ablation to measure whether this matters in practice.

Proposed Experiments

The file contains no continuous-variable quantum benchmark, so the experiments use simulated heterodyne sampling. This is exact, because $Q_\rho$ can be sampled by rejection from its Fock-basis expansion.

Calibration first reproduces the uniform-on-$(0,1/2)^2$ null check at $n=2500$ and $5000$ from Limit theorems for entropic optimal transport maps and Sinkhorn divergence. It then repeats the check for heterodyne data from vacuum, thermal, and squeezed states, where the non-compact extension is untested. We report P-P plots and empirical size.

The power study compares four pairs:

  • coherent states at small displacement,
  • squeezed states against thermal states with matched photon number,
  • the Fock state $|1\rangle$ against a thermal state with matched second moments,
  • even against odd cat states, whose difference lives almost entirely in sub-vacuum fringes and should be heavily attenuated.

We plot power against $n$ from $10^2$ to $10^5$ for all baselines. We check the predicted attenuation ordering: the cat-state pair should be hardest, by an amount our local-power formula predicts before the simulations run.

The effect-size study checks the coverage of 95% intervals for Gaussian pairs against the closed-form value.

The disc study uses SU(1,1) coherent-state symbols with the hyperbolic cost. It tests two predictions from Lee's spectral picture: power against localized perturbations decays with the depth of simulated cascaded channels, while power against radial mass-shift alternatives persists.

The decisive check is simple. If the physics-fixed $\epsilon$ yields valid size and matches the tuned kernel test's power on the non-Gaussian pairs, and the disc power-decay pattern follows the Berezin spectrum, the core claim holds. If either fails, the identification of heterodyne data with a usable smoothing operator is not enough on its own.