Quantum circuits with tunable rotation angles cannot be trained by ordinary backpropagation. Gradients have to be measured on the hardware, typically with the parameter-shift rule. This rule evaluates the circuit at two shifted angles per parameter and repeats each evaluation thousands of times ("shots") to average away measurement noise. When such circuits are trained in a federated way, with many clients each holding private, differently distributed data and a central server combining their work, two costs pile up. The first is the shot bill. The second is client drift: local steps pull each client toward its own optimum, so averaging their angles no longer solves the shared problem. We propose a different arrangement in which clients never take local steps on their own losses. Along any few rotation angles, a circuit's expectation value is an exact low-degree trigonometric polynomial. Under squared-error or energy objectives, a client's whole loss along those angles is therefore described exactly by a small vector of coefficients. Clients send Bayesian estimates of these coefficients, the server adds them, and the server then knows the global loss along those angles exactly up to shot noise and can minimize it directly. Because no client ever optimizes its own objective, there is no drift. The posterior uncertainty tells the server how many shots each client should spend.
The opportunity is that "sufficient statistics add exactly" is the idea that made federated Gaussian processes scalable in Federated Bayesian Neural Regression. The physics-informed kernels of Physics-Informed Bayesian Optimization of Variational Quantum Circuits supply an exact, non-random finite feature space where that idea holds without approximation. No existing work joins the two. The result is a federated quantum optimizer whose behavior under data heterogeneity can be predicted and whose shots can be allocated across heterogeneous clients.
Core Claim
Take a circuit whose gates are rotations $\exp(-i\theta_l H_l/2)$ with Pauli generators. Write $\theta_B$ for a block of $B$ angles, with all other angles held fixed. Then each data point's expectation value is a trigonometric polynomial with $3^B$ terms, products of $\{1,\cos\theta_l,\sin\theta_l\}$. If client $m$'s loss is a mean squared error on such expectations, or a weighted sum of them as in a distributed energy objective, it follows that
$$f_m(\theta_B) = \beta_m^\top \phi(\theta_B),$$
where $\phi(\theta_B)$ is a fixed vector of trigonometric features. Its dimension is $5^B$ for squared error, since squaring raises the degree to two, and $3^B$ for linear objectives. $\beta_m$ is client $m$'s coefficient vector. The global objective $f=\frac1M\sum_m f_m$ then has coefficients $\beta=\frac1M\sum_m\beta_m$.
Each client places a Gaussian prior on $\beta_m$ and observes its loss at a few shifted angles with shot noise of known variance. Its posterior $\mathcal N(\mu_m,\Sigma_m)$ is exact Bayesian linear regression, and the server's posterior on the global coefficients is $\mathcal N(\frac1M\sum\mu_m,\frac1{M^2}\sum\Sigma_m)$. The server minimizes the posterior-mean global landscape over $\theta_B$ and broadcasts the new angles.
The claim has three parts. First, this procedure has no client drift term at all, because heterogeneity enters only through the coefficients being summed and never through local trajectories. Second, it removes the angle-averaging-on-a-circle problem, because the server alone holds the angles. Third, given a shot budget, it can choose shots per client to minimize the variance of the global landscape near its minimizer. Clients whose local landscapes are steep or noisy along the block then get more shots, which is something no parameter-averaging method can express.
Relation to Prior Work
Two papers supply the landscape model, and they are the closest single-machine work. Physics-Informed Bayesian Optimization of Variational Quantum Circuits builds a Gaussian process kernel from the trigonometric structure. It proves that three noisy points reconstruct a one-dimensional slice, and it identifies the parameter-shift rule and Nakanishi–Fuji–Todo coordinate updates as two faces of the same structure. Bayesian Parameter Shift Rules in Variational Quantum Eigensolvers recasts the parameter-shift rule as Gaussian process derivative inference. Its GradCoRe method reuses past measurements and chooses shot counts from posterior variance. Both papers are single-agent, use one Hamiltonian, and run full Gaussian process inference that scales cubically with the number of observations. We reuse their kernel and their shot-allocation logic. We change three things: we replace the full kernel with its exact finite feature expansion on small blocks, which removes the cubic cost; we move from one landscape to a sum of heterogeneous client landscapes; and we allocate shots across clients rather than over time.
The federated Gaussian process line supplies the aggregation idea. Federated Bayesian Neural Regression projects inputs through learned random features and sums client scatter matrices to recover a global posterior. It also shows that personalized federated Gaussian processes collapse outside their local data range. Federated Gaussian Process instead averages kernel hyperparameters and then conditions locally. Heterogeneous Federated Zeroth-Order Optimization using Gradient Surrogates, called FZooS, uses random-feature Gaussian process surrogates of client gradients with a drift correction. All three use generic stationary kernels on random features, which carry approximation error that FZooS notes grows along long trajectories. Our feature map is exact and has no random-feature error, and the quantity aggregated is the objective's coefficients rather than a regression dataset.
The federated quantum work motivates the problem. Federated Quantum Machine Learning runs FedAvg over parameter-shift gradients, but only on IID splits. Quorus handles devices of different circuit depths and non-IID data with layerwise losses, but uses plain parameter-shift gradients and no surrogate. A hybrid quantum neural network for split learning keeps the quantum circuit at the server and explicitly assumes IID clients.
Classical drift-correction work constrains the design. SCAFFOLD and FedProx correct drift with control variates or proximal anchors. Both assume low-noise gradients and Euclidean parameters, and neither assumption holds for shot-noisy periodic angles.
Research Program
The first step is to build the client and server protocol. In each round the server picks a block of $B\in\{1,2,3\}$ angles, sweeping over the circuit. Each client evaluates its data-averaged loss at a design of shifted points; the minimal design for $B=1$ is five points at multiples of $2\pi/5$. Clients condition a Gaussian prior on these observations and send $(\mu_m,\Sigma_m)$, which is at most 125 coefficients plus a covariance. The server sums the posteriors, minimizes the global landscape over the block, and broadcasts the result.
The client's prior comes from a sliding-window Gaussian process with the physics-informed kernel over the full angle vector, fit to that client's recent observations. This is the information-recycling step of Bayesian PSR, here used only to set the prior over the next block.
Shot allocation works as follows. Given a total per-round budget, the server picks per-client shot counts to minimize the posterior variance of the global landscape value at the predicted minimizer. That variance is a sum of per-client terms, each scaling like inverse shots, so the problem is a small convex allocation with a closed-form square-root rule. The server stops adding shots once a confident-region test in the style of GradCoRe passes on the global posterior.
The analysis covers three things. The first is that, with full client participation and correct noise variances, each round's update is the exact Bayesian block-coordinate step on the global objective. Convergence therefore follows coordinate-descent arguments, with no heterogeneity term and with error controlled by the summed posterior variance. The second is what breaks under partial participation: the server then sees a random subset of clients. Heterogeneity re-enters as sampling variance of the summed coefficients rather than as drift, and we will characterize it that way. The third is the role of losses outside the trigonometric family. Cross-entropy is not a finite trigonometric polynomial, so we restrict to squared-error and energy objectives and measure how much accuracy that restriction costs.
Baselines
The do-nothing reference is FedAvg with standard two-point parameter-shift gradients, as in Federated Quantum Machine Learning. Clients take $K$ local steps with fixed shots per evaluation, and the server averages angles. It is the method people actually deploy. We will tune $K$, the learning rate, and shots per evaluation on a validation split, under the same total shot budget and the same number of communication rounds as our method. Beating it means reaching the same test loss or energy with fewer total shots and fewer rounds, under Dirichlet label skew.
The classical drift-correction baselines are FedProx and SCAFFOLD with parameter-shift gradients, following FedProx and SCAFFOLD. They are the strongest generic answers to non-IID drift. We will give them the same tuning grid size as our method's single shot-budget parameter, plus the proximal weight or control-variate step. On classical non-IID problems, FedProx is reported to improve absolute test accuracy over FedAvg by 22% on average with 90% stragglers, and SCAFFOLD reached target accuracy on heterogeneous EMNIST where FedAvg failed within 1,000 rounds. For quantum circuits under shot noise, no reported result was found on Lacuna for either method. The comparison is meaningful only if our method beats them at matched shots, and especially if their advantage over FedAvg shrinks as shots per evaluation fall, which is our hypothesis about noisy control variates.
The strongest and most direct competitor is federated exact coordinate descent without any Bayesian machinery, which we call federated NFT. Each client measures its loss at the same five or $3^B$ shifted points with a fixed, uniform shot count. The server solves the resulting linear system for the coefficients, sums them, and minimizes. It inherits the absence of drift. The comparison isolates exactly what the Gaussian process prior and the variance-driven shot allocation buy. It will use identical block schedules and total shots.
As an upper reference we run the centralized single-agent optimizer on pooled data, using GradCoRe from Bayesian Parameter Shift Rules and EMICoRe from Physics-Informed Bayesian Optimization. The numbers to approach are those reported in the latter for the 5-qubit Ising model: EMICoRe reaches 98% ground-state fidelity after 4,000 observations, while baselines reach 92% after 6,000. GradCoRe is reported to need an order of magnitude fewer cumulative shots than fixed-shot SGD, but the file gives no absolute figure.
Two ablations isolate the key ingredients: removing the cross-round prior, which keeps allocation but uses a flat prior, and removing heterogeneity-aware allocation, which keeps the prior but uses uniform shots across clients.
Why not just…?
Why not just run FedAvg or SCAFFOLD with parameter-shift gradients and more shots? More shots shrink the noise but not the drift. FedAvg's failure under heterogeneity is a bias, as the SCAFFOLD paper argues. SCAFFOLD's correction depends on control states that are themselves shot-noisy estimates of full gradients, each costing two evaluations per parameter. Our experiments sweep shots per evaluation and the Dirichlet concentration jointly. If SCAFFOLD matches us at all shot levels, the drift-free claim has no practical value and we will report that.
Why not just do federated NFT with uniform shots, skipping Gaussian processes entirely? This is the real threat, because it already gets the drift-free property from the same trigonometric fact. The Gaussian process earns its place only if two effects hold. Recycling past observations should shrink the shots needed per block. Variance-aware allocation should matter more as client landscapes diverge, since under IID data uniform allocation is near-optimal. The decisive test is the second ablation plotted against Dirichlet concentration. If the gap is negligible even under strong skew, the contribution reduces to "federated exact coordinate descent is drift-free," and we will present it that way rather than defend the Bayesian layer.
Why not just use an FZooS-style federated surrogate with an RBF kernel and random Fourier features? That approach is built for black-box objectives. It pays random-feature approximation error, which its authors note degrades along long trajectories. It also has to learn periodic structure from data, whereas we know the structure exactly. We will run FZooS with the same shot budget on the Ising and classification tasks. If its random features match our exact features at equal message size, the physics-informed prior is not needed for communication efficiency.
Why not just keep the circuit at the server, as in quantum split learning? That approach sends client activations per sample every step and processes clients round-robin, so latency grows linearly with cohort size. Its reported gains of over 3% on Fashion-MNIST and over 1.5% on Speech Commands come from IID splits. It answers a different question, namely clients without quantum hardware, so we treat it as context rather than a baseline.
Proposed Experiments
The physics tasks are the 5-qubit Ising and Heisenberg chains from the Bayesian PSR and EMICoRe papers. We make them federated by giving each client a disjoint, unequally weighted subset of Hamiltonian terms, with term weights skewed by a Dirichlet draw. The global objective is then the full energy, a linear objective with $3^B$ features. Metrics are ground-state fidelity and energy error against cumulative shots and communication rounds, with 100 random seeds and Wilcoxon signed-rank tests as in the source papers.
The learning tasks are CIFAR-10 planes-versus-cars and Fashion-MNIST, reduced to 10 features by principal component analysis, following Quorus. They are trained with squared-error loss on a 10-qubit circuit across 10 to 50 clients under Dirichlet label skew. On planes-versus-cars, the IID FedAvg figure of 93.4% accuracy, against 93.65% centralized, from Federated Quantum Machine Learning anchors the easy end. We also report the accuracy cost of squared-error versus cross-entropy training in the centralized setting.
Four checks decide the outcome. The first is whether our final loss stays flat as Dirichlet concentration drops while FedAvg, FedProx, and SCAFFOLD degrade; this is the drift-free claim. The second is whether the heterogeneity-aware ablation gap grows with skew; this is the Gaussian process claim. The third is whether partial participation at 20% of clients per round degrades our method gracefully, as sampling variance rather than divergence. The fourth is whether gains persist when a noise model of thermal relaxation, dephasing, and gate infidelity distorts the trigonometric landscape. Under that noise the landscape keeps its trigonometric form only approximately, so this is where the exactness argument is most likely to fail.
References
- Federated Bayesian Neural Regression: A Scalable Global Federated Gaussian Process
- Physics-Informed Bayesian Optimization of Variational Quantum Circuits
- Bayesian Parameter Shift Rules in Variational Quantum Eigensolvers
- Federated Gaussian Process: Convergence, Automatic Personalization and Multi-fidelity Modeling
- Heterogeneous Federated Zeroth-Order Optimization using Gradient Surrogates
- Federated Quantum Machine Learning
- Layerwise Federated Learning for Heterogeneous Quantum Clients using Quorus
- A hybrid quantum neural network for split learning
- SCAFFOLD: Stochastic Controlled Averaging for On-Device Federated Learning.
- Federated Optimization in Heterogeneous Networks