Distributed Quantum Inference

• Distributed quantum inference is a decentralized approach that partitions quantum states, data, and operations across spatially separated nodes under resource limitations.
• It leverages entanglement and optimized communication protocols to facilitate tasks like state certification, tomography, and machine-learning inference across distributed quantum processors.
• Practical implementations include secure multi-party computation and modular quantum computing, demonstrating improved scalability and efficiency in handling large quantum systems.

Distributed quantum inference denotes a family of inference tasks in which data, quantum states, or quantum subroutines are partitioned across spatially separated nodes, and the final estimate, decision, or prediction must be produced under explicit constraints on communication, coherence time, or hardware size. In the current literature, the term covers at least three distinct but related settings: inference from distributed copies of unknown quantum states under one-way communication constraints; modular execution of quantum circuits across several Quantum Processing Units (QPUs) connected by quantum links; and distributed quantum machine-learning architectures in which multiple small quantum models jointly produce a prediction that would otherwise require a larger monolithic device (Doosti et al., 7 Apr 2026, Promponas et al., 2024, Kawase, 2023).

1. Formal settings and problem classes

The literature does not employ a single universal model. Instead, distributed quantum inference appears in several formalisms, each tied to a different bottleneck. In “distributed computing with limited communication” (DCLC), Alice and Bob receive classical inputs $\mathbf{x}$ and $\mathbf{y}$, send bounded-capacity physical systems to Charlie, and Charlie must compute

$$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$

with each sender restricted to a system of operational dimension at most $2^{n-1}$ (Saha et al., 2020). In the communication-constrained state-inference model, $m$ distributed nodes each receive one copy of an unknown $d$-dimensional state $\rho$, send at most $n_c$ classical bits and $n_q$ qubits to a central node, and may or may not share public randomness or Bell pairs; this is formalized as the $(n_c,n_q,R,E)$-model (Doosti et al., 7 Apr 2026). In compositional quantum-network models for learning, Alice and Bob alternately apply local circuit blocks to a transmitted quantum state,

$\mathbf{y}$0

and inference consists of estimating

$\mathbf{y}$1

by repeated state preparation and measurement (Gilboa et al., 2023). In distributed quantum computing, the object being “inferred” is often the output of a circuit too large for one processor, so the central problem becomes online compilation across multiple QPUs with probabilistic entanglement generation and a decoherence deadline (Promponas et al., 2024).

Paradigm	Distributed resource	Canonical task
DCLC / DC$\mathbf{y}$2LC	Bounded operational-dimension messages	Compute $\mathbf{y}$3 with $\mathbf{y}$4
Communication-constrained state inference	One-way classical/quantum messages from many nodes to a center	Certify whether $\mathbf{y}$5 or $\mathbf{y}$6
Compositional network inference	Alternating local circuit layers across parties	Estimate $\mathbf{y}$7
Modular DQC	Multiple QPUs, local coupling graphs, probabilistic EPR links	Execute a circuit before decoherence

These formulations differ in what is distributed—input bits, copies of an unknown state, learned circuit layers, or hardware capacity—but they share the same structural feature: the target quantity is global, whereas the available operations are local and resource-limited.

2. Communication complexity, separations, and foundational tasks

A foundational result is that quantum resources can change the exact solvability of distributed inference tasks under fixed communication budgets. For DCLC(2), a task $\mathbf{y}$8 is nontrivial when no perfect classical strategy exists with at most $\mathbf{y}$9-cbits from each sender, and Theorem 1 states that a nontrivial $$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$0 is perfectly computable in quantum theory iff $$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$1 is balanced. The canonical protocol uses the shared maximally entangled state

$$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$2

local Pauli encodings, and the projective test $$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$3. The same paper further shows that symmetric extreme polygon-model GPTs cannot perfectly perform the nontrivial DCLC(2) tasks, including the delayed-choice variant DC$$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$4LC, thereby isolating the joint role of entangled states, entangled effects, and reversible transformations (Saha et al., 2020).

For distributed quantum state certification, the central question is whether a central node can decide

$$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$5

when each remote node can send only a short quantum message. In the $$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$6-setting with $$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$7, the sample complexity is

$$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$8

achieved by public-coin Haar-random compression $$\mathcal{F}(\mathbf{x},\mathbf{y})=\mathbb{F}(z_1,\ldots,z_n), \qquad z_i=f(x_i,y_i),$$9 followed by centralized Hilbert–Schmidt certification on the compressed states. Under the mixedness-preserving assumption $2^{n-1}$0, this upper bound is matched by a lower bound of

$2^{n-1}$1

while the private-coin setting requires

$2^{n-1}$2

A central structural consequence is that public randomness is necessary to achieve the stronger $2^{n-1}$3 scaling (Doosti et al., 7 Apr 2026).

A complementary benchmark is distributed quantum inner product estimation: Alice receives $2^{n-1}$4 copies of $2^{n-1}$5, Bob receives $2^{n-1}$6 copies of $2^{n-1}$7, and they must estimate $2^{n-1}$8 up to additive error $2^{n-1}$9 using LOCC. The sample complexity is exactly

$m$0

across all measurement and communication settings considered, from non-adaptive single-copy measurements with simultaneous message passing to adaptive multi-copy LOCC. This settles a common misconception that more elaborate interaction or coherent multi-copy processing must necessarily improve the distributed rate: here it does not (Anshu et al., 2021).

At the opposite end of the spectrum, compositional distributed quantum-network models can exhibit exponential communication savings. When the communicated intermediate representation is a $m$1-qubit quantum state, inference requires $m$2 qubits per copy over $m$3 rounds, whereas the classical lower bound for the corresponding general distributed inference problem is $m$4. The same line of work also delineates the boundary of this advantage by proving that no exponential quantum communication advantage can hold for distributed linear classification (Gilboa et al., 2023).

3. Distributed estimation of states and state properties

Distributed quantum inference is often reconstructive rather than decisional. In distributed quantum state tomography, the unknown density matrix is estimated from measurements partitioned across several classical workers. “Local Stochastic Factored Gradient Descent” formulates low-rank QST as

$m$5

then factorizes $m$6 and solves the nonconvex distributed problem by local stochastic gradient steps between synchronization rounds. The analysis uses the Procrustes metric

$m$7

and proves local linear convergence to a neighborhood with constant step size,

$m$8

as well as exact local $m$9 convergence with diminishing step size. The numerical validation is performed on GHZ-state tomography (Kim et al., 2022).

A distinct approach replaces optimization over a point estimate with parallel posterior sampling. In Bayesian QST, the posterior over an overparameterized vector $d$0 is sampled with a parallelized preconditioned Crank–Nicholson Metropolis–Hastings algorithm, and the Bayesian mean density matrix is estimated by pooling many independent chains: $d$1 The method is explicitly “unorthodox,” because for fixed chain length $d$2, taking $d$3 does not in general imply $d$4. Nonetheless, with sufficient thinning the reported integrated autocorrelation times at $d$5 are $d$6 for $d$7, and for $d$8 the simulations show roughly two orders of magnitude lower error than a single chain at the same average wall-clock time. On IBM Quantum Kyoto, the reconstructed $d$9 states reach Bayesian fidelities $\rho$0 and $\rho$1 for $\rho$2 (Nguyen et al., 28 Jan 2025).

The same inferential shift from full reconstruction to targeted property learning appears in distributed Quantum Extreme Learning Machines. The standard three-layer QELM (S3L) and the spatially multiplexed architecture (SM) remain linear in $\rho$3, with outputs of the form $\rho$4. By contrast, the multiple-injection architecture (MI) and the novel distributed architecture with inter-subsystem entanglement (D) access nonlinear targets through

$\rho$5

which enables reconstruction of polynomial quantities, Rényi entropies, concurrence, and negativity using only projective measurements in the computational basis. A key resource result is that for the distributed nonlinear architecture with equal reservoirs,

$\rho$6

so higher-order nonlinearities can be learned by increasing the number of interacting subsystems rather than the size of a single reservoir (Gili et al., 12 Feb 2026).

4. Distributed quantum machine-learning inference

In quantum machine learning, distributed inference often means decomposing the model itself. “Distributed Quantum Neural Networks via Partitioned Features Encoding” replaces one large QNN with $\rho$7 smaller circuits, each receiving a partition $\rho$8 of the input. The final output is the sum of the local expectation values,

$\rho$9

followed by $n_c$0 for classification. The circuits use $n_c$1, $n_c$2, and $n_c$3 gates; the loss is cross entropy; and optimization uses Adam with learning rate $n_c$4. On Semeion, the distributed models slightly outperform a single QNN on the reduced task, but excessive partitioning increases loss; on MNIST, a 14-QNN model reaches accuracy $n_c$5 with loss $n_c$6. The experiments therefore support both halves of the design claim: distribution can improve practicality and accuracy, but over-partitioning degrades performance (Kawase, 2023).

A different line of work frames distributed quantum inference as secure multi-party execution. In the distributed distance-based classifier, normalized training points $n_c$7 with labels $n_c$8 and a new point $n_c$9 are classified via

$n_q$0

The global classifier state is prepared distributively using amplitude encoding, shared GHZ states, and non-local CNOT gates, so that multiple parties can contribute private data without explicitly revealing them. The construction emphasizes phase encoding on entangled states and measurement-based recovery of the final label rather than classical disclosure of the underlying features (Neumann et al., 2022).

Distributed probabilistic inference also appears in multi-agent Quantum Gaussian Processes. In that setting, each agent $n_q$1 holds $n_q$2, trains a local QGP with quantum kernel $n_q$3, and participates in a distributed consensus optimization over circuit hyperparameters $n_q$4. The consensus constraint

$n_q$5

is enforced with a Distributed consensus Riemannian ADMM on the torus manifold $n_q$6, using a circular-mean consensus update and local primal/dual updates. The predictive equations retain the standard GP form,

$n_q$7

with the covariance now supplied by a projected quantum kernel. On SRTM and synthetic data, the reported average improvements are about $n_q$8 lower NRMSE than FACT-GP and about $n_q$9 lower NRMSE than apxGP, while the experiments remain entirely simulator-based (Gandhi et al., 16 Feb 2026).

5. Modular execution and runtime control on distributed hardware

For hardware-limited quantum inference, the decisive issue is often not the abstract communication complexity of the task but whether a wide circuit can be executed before decoherence. A reinforcement-learning compiler for distributed quantum computing addresses exactly this bottleneck. The architecture assumes multiple QPUs connected by quantum links that generate EPR pairs probabilistically; those EPR pairs support gate teleportation for nonlocal CNOTs and qubit teleportation for moving logical states between QPUs. The distributed compiler state is

$(n_c,n_q,R,E)$0

and the deadline-aware objective uses

$(n_c,n_q,R,E)$1

The optimization target

$(n_c,n_q,R,E)$2

formalizes “finish as early as possible, or fail catastrophically at the deadline.” In simulation on two QPUs, 18 logical qubits, and deadline $(n_c,n_q,R,E)$3, Double Deep Q-Networks outperform DQN and PPO, remain effective for $(n_c,n_q,R,E)$4, and improve the chance of completing 30- and 40-gate circuits before decoherence, whereas 50-gate circuits cannot always be finished within the deadline (Promponas et al., 2024).

Near-term distributed quantum computation also admits explicitly approximate schemes. One approach partitions a many-body system into fragments, augments each fragment with auxiliary qubits that mimic selected environment qubits, and corrects the local Hamiltonian by mean-field terms: $(n_c,n_q,R,E)$5 The short-time fragmentation error obeys

$(n_c,n_q,R,E)$6

which induces a variance-based rule for selecting auxiliaries. The framework extends to limited quantum transfer through selective qubit shuffling or teleportation and is then used for fragmented VQE pre-training. In the reported benchmarks, mean-field-corrected fragment pre-training reduces the geometric mean of the final energy error by roughly three orders of magnitude and reduces the number of iterations to convergence by nearly an order of magnitude (Gomez et al., 2023).

A systems-level counterpart is DQuLearn, a manager-worker architecture for distributed quantum learning. It divides a large workload into subtasks, generates logical circuits from the segments, and assigns them to quantum workers according to available resources: $(n_c,n_q,R,E)$7 Pending circuits are dispatched to workers satisfying $(n_c,n_q,R,E)$8, with selection based on the lowest classical resource usage. The paper is explicitly about training, not inference, but a plausible implication is that the same co-management and multi-tenant scheduling mechanisms can support batched inference services. In a 4-worker multi-tenant setting, the reported training results show runtime reduction up to $(n_c,n_q,R,E)$9 and circuits-per-second improvement up to $\mathbf{y}$00 times (Jr. et al., 2023).

6. Misconceptions, boundaries, and open directions

Several recurrent simplifications are inaccurate. Distributed quantum inference is not synonymous with circuit cutting, nor with federated training, nor with communication-constrained state testing alone. The current literature includes exact local-global Boolean computation with entanglement assistance, sample-complexity-optimal testing under one-way quantum communication, modular execution of large circuits over QPUs, distributed Bayesian tomography, property learning from copies of quantum states, and distributed quantum classifiers on classical data (Saha et al., 2020, Doosti et al., 7 Apr 2026, Promponas et al., 2024, Nguyen et al., 28 Jan 2025).

Nor is quantum advantage universal. One paper proves an exponential quantum communication advantage for a broad compositional class of distributed learning models, but also proves that such an exponential advantage cannot hold for linear classification (Gilboa et al., 2023). Another shows that for distributed inner product estimation, even the strongest adaptive multi-copy LOCC protocols do not improve the asymptotic sample complexity beyond what is already achievable by non-adaptive single-copy simultaneous message passing (Anshu et al., 2021). In state certification, shared randomness is not a cosmetic convenience but a necessary resource for the $\mathbf{y}$01 scaling under mixedness-preserving channels (Doosti et al., 7 Apr 2026). In distributed QNNs, more partitions do not monotonically help: the empirical pattern is that distributed models can outperform a single QNN, but excessive partitioning reduces performance (Kawase, 2023). In parallel Bayesian QST, pooling many independent chains is practically effective but does not restore asymptotic unbiasedness in the number of chains $\mathbf{y}$02 when the individual chain length $\mathbf{y}$03 is fixed (Nguyen et al., 28 Jan 2025).

A neighboring but distinct area is distributed quantum interactive proofs, where the task is distributed verification rather than statistical or predictive inference. There, any constant-$\mathbf{y}$04 classical distributed interactive protocol can be compressed to a $\mathbf{y}$05-turn distributed quantum interactive protocol of the same asymptotic certificate size, or to $\mathbf{y}$06 turns with shared randomness (Gall et al., 2022). This boundary is conceptually useful: it shows that “distributed quantum inference” sits at the intersection of communication complexity, estimation, learning, and verification, but it is not exhausted by any one of them.

The main open directions stated in the literature are correspondingly heterogeneous. For distributed compilation, proposed extensions include initial qubit mapping, circuit tokenization for larger models, repeater or switch architectures, and heterogeneous noise objectives such as maximizing final fidelity rather than minimizing time (Promponas et al., 2024). For distributed QNNs, proposed extensions include quantum communication between QNNs, ensembles with different feature partitions, and multichannel images (Kawase, 2023). For communication-constrained state inference, the framework explicitly opens learning, tomography, closeness testing, limited entanglement, and mixed classical/quantum communication as further problems (Doosti et al., 7 Apr 2026). This suggests that distributed quantum inference is best understood not as a single algorithmic technique, but as a research program for extracting global quantum or quantum-enhanced information from systems that are physically, computationally, or informationally decentralized.