When the same quantum state is prepared on several small, noisy quantum processors, or on one processor across several calibration windows, each run produces a slightly different estimate of that state. The standard way to combine them is to pool the measurement data, which amounts to taking the arithmetic average of the density matrices (the matrices that fully describe a quantum state). Averaging this way has a known flaw in other settings. Euclidean averaging of images or distributions blurs sharp features into ghosts, and averaging along the right geometry avoids this. Quantum states have the same problem. Suppose two devices over-rotate a qubit by $+\delta$ and $-\delta$. Their arithmetic average is a mixed state: the two coherent, device-specific errors become one incoherent error, and fidelity, purity and entanglement estimates all degrade. This proposal asks whether averaging in the Bures geometry, which is the quantum fidelity geometry and coincides with the Bures-Wasserstein metric from optimal transport, preserves what the devices agree on while cancelling how they disagree.
The core claim is concrete and testable. Under device-to-device coherent error spread, the Bures barycenter of per-device state estimates is closer to the target than the pooled mixture. For pure targets with symmetric spread, the barycenter recovers the target exactly, while the mixture loses fidelity at second order in $\delta$. Under purely incoherent noise, the two estimators should roughly tie. The open question is practical: whether this advantage survives finite measurement shots on hardware with few qubits. Shot noise turns low eigenvalues negative, which breaks the matrix square roots that barycenter algorithms rely on. We would build a shot-robust barycenter estimator for few-qubit states and reduced (local) marginals, and map exactly where it beats pooling.
Core Claim
For positive semidefinite matrices $\rho,\sigma$, the Bures-Wasserstein distance is
$$d_B(\rho,\sigma)^2=\mathrm{Tr}\,\rho+\mathrm{Tr}\,\sigma-2\,\mathrm{Tr}\big(\rho^{1/2}\sigma\rho^{1/2}\big)^{1/2}.$$
On trace-one matrices this is exactly the Bures distance between quantum states. Given per-device estimates $\rho_1,\dots,\rho_K$ with weights $w_k$, the barycenter is the minimizer $\bar\rho$ of $\sum_k w_k d_B(\bar\rho,\rho_k)^2$ over the positive semidefinite cone, followed by trace renormalization. It can be computed with the fixed-point iteration
$$\Sigma\leftarrow \Sigma^{-1/2}\Big(\textstyle\sum_k w_k(\Sigma^{1/2}\rho_k\Sigma^{1/2})^{1/2}\Big)^2\Sigma^{-1/2}.$$
On pure states the geodesic between rank-one matrices stays rank-one. The barycenter of $\pm\delta$-rotated copies of $|\psi\rangle$ is therefore proportional to $|\psi\rangle\langle\psi|$, while the mixture has fidelity $\cos^2(\delta/2)$. The trace shortfall of the unnormalized barycenter is itself a free diagnostic of how much coherent disagreement exists between devices.
Most quantum states that are measurable on small hardware are not pure. The targets we care about are reduced marginals of entangled states, and those are mixed. For mixed targets the claim becomes a hypothesis: the barycenter of conjugated copies $U_k\rho U_k^\dagger$ should distort the spectrum less than the mixture does, and so bias concave quantities like entanglement entropy less. Testing this hypothesis is part of the work.
Relation to Prior Work
The optimization side is supplied by the Bures-Wasserstein literature. Averaging on the Bures-Wasserstein manifold: dimension-free convergence of gradient descent and Gradient descent algorithms for Bures-Wasserstein barycenters prove that Riemannian gradient descent converges linearly at a rate set by the condition number $\kappa$ rather than the dimension. Barycenter Estimation of Positive Semi-Definite Matrices with Bures-Wasserstein Distance shows that a tangent-space "projection mean" converges in about 10 iterations and avoids the singularity failures of affine-invariant averaging. We reuse these algorithms. Both lines assume exact input matrices and a finite $\kappa$. Our inputs are shot-noisy and often nearly rank-deficient, so $\kappa$ is effectively infinite. That is precisely the regime their theory leaves open.
The motivation comes from the geometry-preserving averaging line and from quantum federated learning. Wasserstein Barycenters for Geometry Preserving Averaging and Wasserstein Barycenter Model Ensembling show, in classical settings, that transport-based averaging keeps structure that arithmetic averaging destroys. The latter reports 86.8% versus 80.7% accuracy on Animals with Attributes. On the quantum side, Toward Heterogeneous Quantum Federated Learning: Challenges and Solutions argues that identical inputs map to incompatible states on heterogeneous devices, so Euclidean aggregation is ill-posed. Its fix, like that of Sporadic Federated Learning Approach in Quantum Environment to Tackle Quantum Noise, is to scale down or drop noisy clients. Those methods aggregate circuit parameters and change only the weights. We aggregate the states themselves and change the geometry. Their noise scores could serve as our weights $w_k$.
The closest quantum-geometric work does something different. Quantum Wasserstein Compilation uses the order-1 quantum Wasserstein distance of The quantum Wasserstein distance of order 1 as a cost for compiling a single unitary, and does so only in noiseless simulation. It does not average anything. Quantum Multiple Rotation Averaging does average on a manifold with quantum hardware, but its manifold is $SO(3)$ rotations solved on an annealer; the quantum device is the solver, not the source of the objects being averaged. Here the quantum devices produce the objects and the averaging is classical.
One existing result constrains the design. The De Palma et al. paper notes that the exact quantum $W_1$ requires an exponentially large semidefinite program. We therefore choose the Bures distance, which has closed-form expressions on small matrices. We also restrict to states on at most about six qubits, or to 2–3-qubit marginals estimated by randomized measurements.
Research Program
The first strand is analysis. We would characterize the barycenter's bias relative to the target under three error models. The first is symmetric coherent spread, meaning random device-specific unitaries centred on the identity. The second is asymmetric coherent bias, where all devices err in the same direction and no averaging can help. The third is device-specific incoherent channels: amplitude damping, dephasing and depolarizing at differing rates. For each model we would compare the barycenter with the mixture on fidelity and on Rényi-2 entropy of marginals.
The second strand is a shot-robust estimator. Linear-inversion tomography and classical-shadow estimates, which reconstruct a state from many randomized single-shot measurements, are unbiased but not positive semidefinite. We would compare three ways of feeding them to the barycenter. The first projects each estimate onto the positive semidefinite cone before the fixed-point iteration. The second adds a small eigenvalue floor, standing in for the entropic regularization of Wasserstein Barycenters for Geometry Preserving Averaging. The third is a stochastic Riemannian gradient method that draws fresh shot batches at each step, in the spirit of the stochastic analysis in Averaging on the Bures-Wasserstein manifold. The matrix square root is nonlinear, so all three are biased at order roughly one over the number of shots. We would derive this bias and confirm it empirically.
The deliverable is a crossover curve. It shows, as a function of coherent spread $\delta$ and shots per device, where the geometric gain outweighs the extra estimation bias.
Baselines
The naive reference is the pooled mixture. All devices' measurement records are combined and reconstructed once, which equals the weighted arithmetic mean of the per-device estimates. This is what practitioners do by default, and it is the right null: any proposal that cannot beat it under coherent spread has no reason to exist. It receives the same total shot budget and the same reconstruction method, either maximum-likelihood or projected linear inversion.
The strongest existing competitor for pure targets is the pooled mixture followed by a rank or purity constraint, keeping the leading eigenvector or doing rank-$r$ maximum-likelihood reconstruction. For two symmetric rotated pure states, the leading eigenvector of the mixture coincides with the Bures barycenter, so this baseline may simply tie with us. It gets the same shots, and its rank is tuned on held-out devices with the same tuning budget we allow our eigenvalue floor.
The strongest non-geometric aggregation baseline is noise-aware weighting in the style of SpoQFL. Device estimates are down-weighted by $\exp(-\gamma|\xi|)$ of an estimated noise score and then mixed. We run it with the same scores as our barycenter weights, so the comparison isolates geometry from weighting.
An oracle per-device calibration serves as an upper bound. It inverts each device's true coherent error before mixing, and is available only in simulation.
Two ablations isolate the key ingredient. Affine-invariant geometric averaging tests whether any curved mean helps; we expect it to fail on near-singular states, as Barycenter Estimation of Positive Semi-Definite Matrices with Bures-Wasserstein Distance reports. The single-tangent-space projection mean, compared against the full Riemannian iteration, tests whether the cheap approximation suffices.
On task-level numbers, there is no reported result found on Lacuna for any method pooling quantum state estimates across devices. The relevant reported figures are algorithmic only. Riemannian gradient descent reaches $10^{-12}$ error in under 0.5 seconds versus over 15 seconds for an off-the-shelf semidefinite programming solver (Averaging on the Bures-Wasserstein manifold). The projection mean converges in about 10 iterations versus roughly $10^3$ for the affine-invariant inductive mean (Barycenter Estimation of Positive Semi-Definite Matrices with Bures-Wasserstein Distance).
"Beating" a baseline means lower trace distance to the target and lower absolute error in marginal Rényi-2 entropy. The difference must be significant across at least 50 random device draws, at equal total shots. Closing at least half the gap between the mixture and the oracle-calibration bound would count as a substantive win.
Why not just…?
Why not just pool all the data, since the mixture is unbiased for linear observables? The mixture is unbiased only with respect to the average device, not the target. Under symmetric coherent spread it carries second-order error in every observable. For nonlinear quantities such as purity and entanglement entropy its bias is first-class, because mixing inflates entropy. The incoherent-noise arm of our experiments is where pooling should tie with the barycenter. If pooling also ties under coherent spread at realistic shot counts, the proposal fails.
Why not just take the leading eigenvector of the pooled estimate? This is the real threat, and for pure targets it may match the barycenter. We concede that case in advance. The proposal's value then rests on mixed targets: reduced marginals of entangled states, and Gibbs-like or noisy targets. There, eigenvector truncation throws away genuine spectrum and the barycenter does not need to. The ablation on 2–3-qubit marginals of GHZ, cluster and small transverse-field Ising ground states settles this. If rank-constrained pooling with a tuned rank matches the barycenter there too, we would recommend it instead.
Why not just calibrate each device and undo its error? Calibration requires characterizing every device's coherent error, which drifts between calibration windows. The barycenter needs no error model. We include oracle calibration as the ceiling. The interesting result is how much of the gap a model-free estimator closes at zero characterization cost.
Why not just average circuit parameters, as quantum federated learning does? In the setting here, every device runs the same parameters, and the discrepancy lives in the hardware rather than in the parameter vector. Parameter averaging, as in Federated Quantum Machine Learning, therefore has nothing to correct.
Proposed Experiments
The main testbed is simulation with Qiskit Aer-style noise models, the simulator family used by the heterogeneous federated learning papers. We simulate 4 to 10 virtual devices. Each device draws its own coherent over-rotations, either symmetric or biased, together with its own T1/T2 relaxation times and readout error. Targets are 2–6-qubit GHZ, W and cluster states, small variational ground states, and Haar-random product states on 3–8 qubits as in Quantum Wasserstein Compilation. Per-device budgets range from $10^3$ to $10^5$ shots, using full tomography up to 4 qubits and randomized measurements for marginals beyond that. Metrics are trace distance and fidelity to the target, error in marginal Rényi-2 entropy, error in energy for the variational states, and wall-clock cost of aggregation.
Four checks are decisive. First, under symmetric coherent spread, the barycenter's advantage over the mixture should grow roughly as $\delta^2$, which is the order of the mixture's fidelity loss. Second, under incoherent-only noise the two should tie within shot error; a consistent loss there would mean the method harms the common case. Third, under shared biased errors both should fail equally, confirming that the gain comes from cancelling disagreement and not from denoising. Fourth, sweeping the number of shots should locate a crossover below which square-root bias makes the barycenter worse than pooling, and should show whether eigenvalue flooring or stochastic re-sampling pushes that crossover lower.
If the simulated crossover falls within reachable shot budgets, we would repeat the 2–3-qubit marginal experiment on a small number of publicly accessible devices, treating separate calibration days as separate devices.
References
- Averaging on the Bures-Wasserstein manifold: dimension-free convergence of gradient descent
- Gradient descent algorithms for Bures-Wasserstein barycenters
- Barycenter Estimation of Positive Semi-Definite Matrices with Bures-Wasserstein Distance
- Fast Computation of Wasserstein Barycenters
- Wasserstein Barycenter Model Ensembling
- Toward Heterogeneous Quantum Federated Learning: Challenges and Solutions
- Sporadic Federated Learning Approach in Quantum Environment to Tackle Quantum Noise
- Quantum Wasserstein Compilation: Unitary Compilation using the Quantum Earth Mover's Distance
- The Quantum Wasserstein Distance of Order 1
- Quantum Multiple Rotation Averaging
- Federated Quantum Machine Learning