One-sentence Summary

Researchers at the University of Innsbruck prove that augmenting shallow, bounded-connectivity unitary Born models with shared classical randomness—implemented via a single random bit controlling spatially separated local Pauli operations—establishes a strict scalable representational separation by generating long-range correlations that purely unitary shallow-depth models cannot reproduce and may require depth $\Omega(N)$Ω(N) in one-dimensional architectures.

Key Contributions

• Augmenting shallow, bounded-connectivity unitary circuits with shared classical randomness, specifically spatially separated local Pauli operations controlled by a single random bit, creates a strict and scalable representational separation by generating long-range correlations that no purely unitary shallow model of the same depth and connectivity can reproduce.
• For one-dimensional nearest-neighbor architectures, reproducing such distributions with a purely unitary model requires depth Ω(N) in the worst case.
• Measurement-based quantum computation provides a natural implementation of the required shared classical randomness, and numerical experiments on MBQC-based generative models support the analytical results.

Introduction

Quantum generative models such as quantum circuit Born machines produce classically hard-to-simulate distributions, but their expressivity is severely constrained in near-term devices by limited circuit depth and local qubit connectivity. Shallow local circuits cannot establish correlations between distant qubits because information propagation is restricted to a finite spatial range, and prior attempts to overcome this with mid-circuit measurements and feedforward introduce experimental overhead. The authors show that shared classical randomness alone, without additional quantum depth, long-range gates, or adaptive operations, strictly enlarges the family of output distributions accessible to shallow local quantum generative models. They construct a stochastic channel model where a tunable shared Pauli string coordinates local Pauli gates on spatially separated qubits, creating long-range correlations that a purely unitary Born model of the same depth and connectivity cannot reproduce.

Dataset

The paper does not introduce a new dataset in the traditional sense. Instead, the authors provide the code and numerical data that support the theoretical and numerical findings of the study. This material is available at a repository (the link is given in the paper). The work is primarily theoretical, proving shallow-depth separations between unitary and correlated channel models for quantum generative modeling, and the numerical experiments rely on synthetic constructions rather than real-world data. Therefore, there is no dataset composition, filtering, or training split to describe.

Method

The authors propose a quantum generative model that augments shallow local unitary circuits with shared classical randomness to enhance expressivity. The method is built upon a generic brickwall unitary circuit architecture, which is then extended into a channel model and implemented natively within a measurement-based framework.

Generic Brickwall Unitary and Channel Circuits

The foundation of the model is a shallow one-dimensional local quantum circuit on $N$N qubits, composed of $D$D layers. The unitary operator is defined as $U(\boldsymbol{\theta}) = U_D(\boldsymbol{\theta}_D) \dots U_1(\boldsymbol{\theta}_1)$U(θ)=UD​(θD​)…U1​(θ1​), where each layer consists of parametrized two-qubit gates acting on nearest-neighbor bonds. These layers alternate between sub-layers acting on disjoint even and odd bonds, forming a brickwall structure.

To extend this unitary model, the authors introduce correlated stochastic Pauli operations. A Pauli string $P_{\mathcal{M}}$PM​ is controlled by a single classical Bernoulli variable $s \in \{0, 1\}$s∈{0,1}. This variable determines whether the entire Pauli string is applied to a subset of qubits $\mathcal{M}$M at a specific insertion slot between circuit layers. This creates a spatially correlated stochastic operation. The resulting channel model $\mathcal{E}_{\boldsymbol{\theta}, p}$Eθ,p​ is a convex combination of unitary branches, where the Pauli string is applied with probability $p$p and omitted with probability $1-p$1−p.

As shown in the figure below:

The figure illustrates the advantage of this approach. While a standard shallow circuit (panel b) has disjoint backward light cones for distant observables $\hat{Z}_1$Z^1​ and $\hat{Z}_6$Z^6​ resulting in zero covariance, a shallow circuit augmented with shared randomness (panel c) can generate non-zero covariance between these distant qubits. This capability mimics the long-range correlation generation of a deep circuit (panel a) without increasing the circuit depth.

Learning Model Architecture

For the explicit learning model, the authors specify a depth-$D$D circuit where each layer $k$k consists of nearest-neighbor entangling rotations $R_{zz}$Rzz​ followed by local single-qubit rotations $R_z$Rz​ and $R_x$Rx​. The layer unitary is given by:

$$U_k(\boldsymbol{\theta}_k) = \bigotimes_{i=1}^N R_x(\theta_i^{k,x}) \bigotimes_{i=1}^N R_z(\theta_i^{k,z}) \prod_{i=1}^{N-1} R_{zz}(\theta_{i,i+1}^{k,zz})$$Uk​(θk​)=i=1⨂N​Rx​(θik,x​)i=1⨂N​Rz​(θik,z​)i=1∏N−1​Rzz​(θi,i+1k,zz​)

A shared classical bit $s^l \sim \text{Bernoulli}(p^l)$sl∼Bernoulli(pl) controls the insertion of a stochastic Pauli-Z string $Z_{\mathcal{M}}^{s^l}$ZMsl​ between layers $U_l$Ul​ and $U_{l+1}$Ul+1​. This string acts simultaneously on a subset of qubits $\mathcal{M}$M.

Refer to the framework diagram:

This diagram details the learning model structure. The input state $|\alpha\rangle^{\otimes N}$∣α⟩⊗N passes through parametrized layers $U_l$Ul​. At a designated slot, the stochastic Pauli string $Z_{\mathcal{M}}^{s^l}$ZMsl​ is inserted. The shared random variable $s^l$sl ensures that the local Z operations on the subset $\mathcal{M}$M are either applied jointly or all omitted, creating correlated branches in the circuit evolution.

Endpoint Pauli Correction

To ensure the stochastic Pauli string modifies the circuit dynamics non-trivially without merely relabeling the final measurement outcomes, an endpoint correction procedure is employed. The inserted Pauli string is propagated through the subsequent layers $U_{>l}(\boldsymbol{\theta})$U>l​(θ). During propagation, commuting gates remain unchanged, while anticommuting Pauli rotations undergo sign flips in their rotation angles. Specifically, $P_{\mathcal{M}}^{s^l} \exp(-i\theta Q) = \exp(-i(-1)^{s^l}\theta Q) P_{\mathcal{M}}^{s^l}$PMsl​exp(−iθQ)=exp(−i(−1)slθQ)PMsl​ if the operators anticommute.

A final Pauli correction layer is appended to cancel the accumulated Pauli string at the output. Since the final string is diagonal in the computational basis (for Z-strings) or corrected explicitly, it does not alter the output probabilities. The effective action of the stochastic Pauli string is thus reduced to branch-dependent sign flips of the rotation angles.

As shown in the figure below:

Panel (a) depicts the model with the stochastic Pauli string inserted after layer $l$l and applied again at the output for correction. Panel (b) shows the equivalent effective model where the propagation results in stochastic flipped angles $\tilde{\boldsymbol{\theta}}$θ~ in the subsequent unitary $U_{>l}$U>l​, while the final Pauli string is removed.

Native Implementation in VMBQC

The authors demonstrate that this correlated channel model can be realized natively within Variational Measurement-Based Quantum Computation (VMBQC). In MBQC, computation is performed via adaptive single-qubit measurements on an entangled resource state, such as a cluster state. Measurement outcomes induce Pauli byproducts. Standard MBQC corrects these byproducts via classical feedforward.

The authors introduce an effective classical control mechanism where byproducts can be deliberately retained or introduced (anti-correction) with a tunable probability $p^l$pl. By processing the raw measurement outcomes, a single shared binary variable $s^l$sl can control the collective Pauli byproduct $Z_{\mathcal{M}}^{s^l}$ZMsl​ acting on a subset of spatially separated qubits.

Refer to the framework diagram:

This figure illustrates a $4 \times 3$4×3 cluster state and its equivalent circuit representation. Panel (a) shows independent randomness where distinct classical variables control byproducts on different qubits ($Z_1^{s_1^1}$Z1s11​​ and $Z_4^{s_4^1}$Z4s41​​). Panel (b) shows shared randomness where a single variable $s^1$s1 controls correlated byproducts on spatially separated boundary qubits ($Z_1^{s^1}$Z1s1​ and $Z_4^{s^1}$Z4s1​), realizing the correlated channel model natively.

To prepare the specific product input state $|\alpha\rangle^{\otimes N}$∣α⟩⊗N required for the learning model, an auxiliary column of qubits is attached to the cluster state.

As shown in the figure below:

The auxiliary column (A) is initialized in the $|+\rangle$∣+⟩ state and measured in the $\{|\pm_\alpha\rangle\}${∣±α​⟩} basis. After applying the corresponding byproduct correction, the state $R_x(\alpha)|0\rangle$Rx​(α)∣0⟩ is teleported to every qubit of the first computational column (B), effectively preparing the desired input state $|\psi_{\text{in}}\rangle = (R_x(\alpha)|0\rangle)^{\otimes N}$∣ψin​⟩=(Rx​(α)∣0⟩)⊗N for the subsequent circuit layers.

Training Process

The model is trained using the squared Maximum Mean Discrepancy (MMD) as the loss function $\mathcal{L}(\boldsymbol{\theta}, \boldsymbol{p})$L(θ,p). This implicit loss quantifies the distance between the model's output distribution $P_{\mathcal{E}_{(\boldsymbol{\theta}, \boldsymbol{p})}}$PE(θ,p)​​ and a target distribution $Y$Y using a kernel function $K(x,y)$K(x,y):

$$\mathcal{L}(\boldsymbol{\theta}, \boldsymbol{p}) = \mathbb{E}_{x,y \sim P}[K(x,y)] - 2\mathbb{E}_{x \sim P, y \sim Y}[K(x,y)] + \mathbb{E}_{x,y \sim Y}[K(x,y)]$$L(θ,p)=Ex,y∼P​[K(x,y)]−2Ex∼P,y∼Y​[K(x,y)]+Ex,y∼Y​[K(x,y)]

The training involves updating both the circuit parameters $\boldsymbol{\theta}$θ and the application probabilities $\boldsymbol{p}$p. The gradients with respect to the probabilities $p^l$pl are computed by evaluating the loss twice: once with $p^l=1$pl=1 and once with $p^l=0$pl=0, while keeping other parameters fixed. The gradients with respect to the variational angles $\theta$θ are evaluated using standard parameter-shift rules. This allows the model to jointly learn the unitary evolution and the optimal stochastic branching probabilities to approximate the target distribution.

Experiment

The analytical results prove that for one-dimensional nearest-neighbor circuits, a correlated channel model with shared randomness can generate output distributions that are unreachable by any shallow-depth unitary circuit, requiring depth linear in the qubit separation to match. This separation extends to all finite-range local architectures and can be realized natively in measurement-based quantum computing. A numerical experiment on a 6-qubit, depth-2 circuit targeting a two-branch mixture distribution validates the representational advantage, with the channel model achieving substantially lower training loss and less variability than the purely unitary baseline.