WorksPhysics and Astronomy
Hardware Demonstration of Neural-Surrogated Quantum-Enhanced MCMC for Large-Scale Constrained Boltzmann Sampling
Abstract
Neural-surrogated quantum-enhanced Markov chain Monte Carlo (QeMCMC) uses neural networks trained on samples from the quantum approximate optimization algorithm (QAOA) to accelerate Boltzmann sampling. A divide-and-conquer strategy extends this approach to larger systems, but independently optimizing QAOA parameters for each subproblem is costly. In this paper, we develop a hardware-efficient implementation of neural-surrogated QeMCMC and demonstrate it using an IBM Quantum Heron r2 processor. Experiments with binary variables reveal faster autocorrelation decay for appropriate subproblem sizes while exposing challenges in neural-surrogate learning for larger blocks. We then extend the demonstration to a system with , where the autocorrelation decay rate per MCMC step is approximately and times those of global and local Kawasaki dynamics, respectively. These results demonstrate faster autocorrelation decay in a system exceeding one thousand variables using samples generated on quantum hardware.
Introduction
Quantum computers have the potential to outperform classical computers for certain computational problems and are expected to find applications in optimization, quantum chemistry, machine learning, statistical physics, and other fields [1, 2, 3, 4]. However, currently available quantum devices are noisy intermediate-scale quantum (NISQ) devices [5], and it remains difficult to run large-scale quantum algorithms with full error correction. Therefore, variational quantum algorithms (VQAs) have been widely studied as a framework for using NISQ devices [6]. A representative example is the quantum approximate optimization algorithm (QAOA) [7], which uses relatively shallow parameterized quantum circuits to minimize a cost function. However, scaling VQAs to larger problems remains challenging because of optimization difficulties such as barren plateaus [8], the large number of measurements required to estimate expectation values, and the increasing effect of hardware noise with circuit size [9].
One promising way to exploit current quantum devices is to use samples generated by quantum circuits in combination with classical computation [10, 11, 12]. In particular, quantum-enhanced Markov chain Monte Carlo (QeMCMC) uses quantum-generated samples as proposal distributions and can provide nonlocal updates that accelerate MCMC mixing [13, 14]. Previous work proposed distilling QAOA sample distributions into masked autoencoders for distribution estimation (MADEs) [15, 16, 17] and using the learned distributions as Metropolis–Hastings proposals for Boltzmann sampling [18]. Furthermore, a divide-and-conquer strategy extends this approach to systems larger than the available quantum processor by applying QAOA and surrogate learning to individual subproblems [19]. However, implementing this approach at large scale on real quantum hardware requires substantial quantum resources. In particular, optimizing QAOA parameters separately for many subproblems requires repeated expectation-value evaluations on the quantum processor, which leads to a large number of quantum-circuit executions as the number of subproblems increases [9].
In this work, we develop a hardware-efficient implementation of neural-surrogated QeMCMC and demonstrate constrained Boltzmann sampling for systems with up to binary variables using samples generated on an IBM Quantum Heron r2 processor. Specifically, we construct subproblems with similar graph topologies and employ linear-ramp QAOA [20], which uses only four variational parameters while retaining high transferability across problem instances and sizes. We optimize these four parameters only for one representative subproblem and reuse them for the remaining subproblems, exploiting the transferability of QAOA parameters between problems with similar graph structures and optimization landscapes [20, 21, 22]. The remaining subproblems therefore require only sampling with the transferred parameters, eliminating separate parameter optimization and substantially reducing the number of quantum-circuit executions on hardware. Although one might expect parameter transfer to degrade performance, we found that it still accelerates MCMC convergence, as demonstrated by the numerical experiments mentioned later.
To evaluate the proposed method, we investigate its performance for binary variables by varying the subproblem size using samples generated on an IBM Quantum Heron r2 processor with 156 qubits, and then demonstrate that it can be extended to a larger system with binary variables. More precisely, in the first experiment, we study constrained Boltzmann sampling with binary variables, using subproblem sizes , and . To reduce routing overhead in QAOA circuits, we construct QUBO problems whose interaction graphs match the connectivity of the quantum processor [23, 24, 25]. We found that the autocorrelation decay rate increases with subproblem size up to a certain point but then decreases, which contrasts with previous classical simulations that showed continued improvement for sizes up to [19]. Based on this finding, we construct a larger QUBO problem with binary variables using the same connectivity pattern and demonstrate constrained Boltzmann sampling with , a subproblem size that yielded fast autocorrelation decay in the first experiment. The resulting autocorrelation decay rate per MCMC step is approximately 5.2 and 12.3 times those of nonlocal and nearest-neighbor pair-flip dynamics, respectively. These results show that the improvement over the classical baselines is maintained when the full system size exceeds one thousand variables using samples from quantum hardware.
The remainder of this paper is organized as follows. Sections II and III introduce the sampling framework and its hardware-efficient implementation. Sections IV and V present the experiments and the demonstration, respectively. Section VI summarizes the results and outlines future directions.
Preliminaries
Markov Chain Monte Carlo (MCMC)
Markov chain Monte Carlo (MCMC) is a sampling method that constructs a Markov chain whose stationary distribution is a target probability distribution from which direct sampling is difficult. In this study, we consider an energy function defined for binary variables . Let denote the interaction graph with the vertices . The energy function is defined as
Here, and denote the quadratic and linear coefficients, respectively. In particular, we focus on Boltzmann sampling under the fixed Hamming-weight constraint
The feasible set is defined as
The corresponding constrained Boltzmann distribution is
where is the inverse temperature and is the constrained partition function,
We use the Metropolis–Hastings (MH) algorithm to construct the MCMC updates. Given the current state , a candidate state is generated from a proposal distribution and accepted with probability
If the candidate is rejected, the current state is retained. By repeating this procedure, the resulting Markov chain converges to the target distribution under appropriate conditions.
As classical baselines, we use global and local Kawasaki dynamics, both of which preserve the Hamming weight by exchanging a pair of bits with different values. Global Kawasaki dynamics selects a 1 bit and a 0 bit from the entire system, whereas local Kawasaki dynamics restricts the exchange to adjacent vertices in the interaction graph. Both proposal kernels are symmetric, so the MH acceptance probability reduces to
Because every Kawasaki update preserves the Hamming weight, the Markov chain remains within throughout the sampling process.
Quantum Approximate Optimization Algorithm (QAOA)
The quantum approximate optimization algorithm (QAOA) is a variational quantum algorithm for generating low-cost solutions to combinatorial optimization problems. For binary variables , a quadratic cost function can be mapped to an Ising Hamiltonian of the form
such that
For the Boltzmann sampling problems considered in this work, the cost function corresponds to the energy function of the target system. Here, , , and are determined from the QUBO coefficients through the standard binary-to-Ising transformation.
At circuit depth , the QAOA state is
where is the mixer Hamiltonian and and are variational parameters. The parameters are optimized to minimize the expectation value
which is estimated from quantum-circuit measurements. For the fixed Hamming-weight constraint considered in this work, we use the XY mixer [26, 27],
The XY mixer preserves the total Hamming weight. Therefore, an initial state in the Hamming-weight- subspace remains within the feasible subspace throughout the QAOA circuit.
Hardware-Efficient Neural-Surrogated QeMCMC
We develop a hardware-efficient implementation of neural-surrogated QeMCMC for large-scale systems. Our objective is to sample from the constrained Boltzmann distribution defined in Eq. (4), with the Markov chain restricted to the feasible set . In the divide-and-conquer approach, a large system is divided into smaller subproblems, and the QAOA sample distributions for each subproblem are learned by separate MADE models and used as MCMC proposal distributions [19]. However, independently optimizing QAOA parameters for every subproblem requires repeated expectation-value evaluations and therefore a large number of quantum-circuit executions. To reduce this cost, we employ linear-ramp QAOA and construct subproblems so that QAOA parameters can be efficiently transferred between them. As a result, separate QAOA optimization for each subproblem can be avoided, substantially reducing the number of quantum-circuit executions required on quantum hardware.
Figure 1 shows an overview of the proposed implementation. For the interaction graph introduced in Sec. II, let denote a collection of vertex blocks satisfying and . Each block defines a subproblem. We allow neighboring blocks to overlap so that successive block updates can propagate changes across block boundaries and throughout the full system. At the same time, to facilitate QAOA parameter transfer, the subproblems are constructed to have similar local graph structures [21, 22]. For general graph structures, constructing such subproblems may require problem-dependent heuristics that balance these two requirements. In this work, however, we focus on hardware-native subproblems constructed according to the topology of the IBM quantum processor. This allows us to explicitly construct subproblems with similar topologies while maintaining the required overlaps, without introducing a dedicated partitioning heuristic.
For each block , we define the local cost function using the interactions contained within the block,
where is a configuration of the variables in , and and are the QUBO coefficients defined in Sec. II. Interactions crossing the block boundary are omitted from this local cost function. They are nevertheless included in the full-system energy used for the MCMC acceptance test below.
The corresponding cost Hamiltonian is obtained through the same binary-to-Ising transformation introduced in Sec. II and satisfies
The block mixer has the XY form introduced in Sec. II, with its mixer graph defined on the vertices of .
We first optimize the QAOA parameters only for the first subproblem . To reduce the dimension of the optimization problem, we employ linear-ramp QAOA [20]. For a fixed circuit depth , the angles are parameterized as
The four independent parameters are collected into , which determines the angle vectors and .
The QAOA state for is
where the angles are determined by through the linear schedules above. The product is ordered so that layer acts first. The parameters are optimized by minimizing the expected local energy,
Thus, the optimized angle vectors and are obtained by optimizing only four parameters on the first subproblem.
Motivated by the transferability of QAOA parameters between problem instances with similar graph structures and objective landscapes [21, 22], we reuse and for all subproblems , . For each block, the corresponding QAOA state is
We measure each state in the computational basis and use the resulting bit strings as quantum samples for the corresponding block.
The initial state contains components in all block Hamming-weight sectors. The XY mixer preserves each Hamming-weight sector during QAOA evolution, but the resulting samples can still have different block Hamming weights. This allows the samples to be used for learning distributions conditioned on different Hamming weights, as required when the configuration outside the block changes during MCMC.
These samples are then distilled into a conditional MADE. Each sample is labeled by its measured Hamming weight . For the th autoregressive ordering of the block variables, the learned conditional likelihood is represented as
where denotes the model parameters and denotes the values of the variables preceding in that ordering. We use the uniform mixture over autoregressive orderings, , defined in (31), as the proposal distribution.
During MCMC, one block is selected at each step. We assume that the block-selection probabilities are independent of the current state. Let denote the complement of the selected block. For the current state , the Hamming weight inside is
A proposed block configuration is generated from . The conditional MADE does not enforce the Hamming-weight constraint by construction. If the generated configuration satisfies
the candidate state is defined by
Otherwise, the proposal is rejected and the Markov chain remains at the current state.
For a valid candidate state , the Metropolis–Hastings acceptance probability conditional on selecting block is
Here, is the full-system QUBO energy defined in Sec. II, rather than the local cost used to generate the quantum samples. The energy difference therefore includes interactions between the updated block and the unchanged variables outside it.
Because the variables outside remain unchanged, the conditioning value is identical for the forward and reverse proposals. The block-wise MH update therefore preserves , and selecting blocks with state-independent probabilities preserves the same target distribution. Consequently, the MH correction preserves the target distribution even when the learned MADE does not exactly reproduce the QAOA sample distribution or the local QAOA distribution differs from the target conditional distribution. Convergence to additionally requires the usual irreducibility and aperiodicity conditions on the resulting Markov chain within .
This implementation substantially reduces the cost of QAOA parameter optimization required to construct neural-surrogated QeMCMC. Linear-ramp QAOA uses only four independent variational parameters instead of optimizing the angles of standard depth- QAOA separately. Furthermore, instead of optimizing QAOA parameters independently for all subproblems, we optimize them only for one representative subproblem and transfer the optimized parameters to the remaining subproblems, reducing the number of independent optimizations from to one. Together, these two reductions significantly decrease the number of repeated expectation-value evaluations and quantum-circuit executions required on quantum hardware, making large-scale divide-and-conquer QeMCMC more practical. This efficiency, however, relies on the transferred QAOA parameters remaining effective for the other subproblems, which we examine experimentally on real quantum hardware in the following section.
Block-Size Scaling Experiment on Quantum Hardware
Experimental Setup
In this experiment, we implement the hardware-efficient neural-surrogated QeMCMC introduced in
§ 1.2 to sample from the constrained Boltzmann distribution defined in Sec. II, using hardware-generated QAOA samples. For the quantum hardware experiments, we use the IBM Quantum System Two ibm_kobe installed at the RIKEN Center for Computational Science (R-CCS) [28]. The system contains an IBM Quantum Heron r2 processor with 156 superconducting qubits connected through a sparse heavy-hex graph [29, 30, 31].
We use the QUBO energy function introduced in Sec. II for a system of binary variables. The interaction graph is constructed to match the hardware connectivity of the IBM Quantum Heron r2 processor, as shown in Figure 2. Matching the QUBO interaction graph to the hardware connectivity suppresses routing overhead during circuit execution [23, 24, 25]. For each edge , the quadratic coefficient is independently drawn from a standard normal distribution, , using a fixed random seed for reproducibility. The linear coefficients are set to . We focus on the half-filled case and set the total Hamming weight to .

Figure 2. Problem topology used in the hardware experiment. The interaction graph is constructed to be compatible with the hardware connectivity of the IBM Quantum Heron r2 processor.
To investigate how the MCMC performance scales with the number of variables updated at once, we vary block sizes while keeping the full system size fixed at . Let denote the block size, i.e., the number of binary variables in each block . We consider
The subproblems are constructed to share some vertices across blocks and to have similar graph topologies to facilitate QAOA parameter transfer. Figure 1 Figure 1 shows an example of the partitioning for . For each block size, we generate QAOA samples using the parameter-transfer procedure, distill them into conditional MADE surrogates, and evaluate the MCMC performance using the resulting block-wise proposal distributions. Additional details on MADE training, quantum-sampling budgets, and block definitions are provided in Appendix A.

Figure 1. Overview of the proposed implementation of neural-surrogated QeMCMC. (a) The full system is divided into overlapping subproblems. (b) The four parameters of linear-ramp QAOA are optimized for a representative subproblem using a quantum computer and a classical optimizer. (c) The optimized parameters are reused to generate QAOA samples for all subproblems without further optimization. (d) The samples from each subproblem are used to train a conditional MADE. (e) At each MCMC step, a block is selected, and a new configuration is proposed by the corresponding MADE conditioned on the block Hamming weight . Proposals that violate the Hamming weight constraint are rejected, while valid proposals are accepted or rejected using the Metropolis–Hastings algorithm.
QAOA Results
We first examine whether QAOA can still generate samples at low effective temperatures when the optimized parameters are transferred to other subproblems. To assess this, we compare the effective temperatures of the samples from the representative subproblem and the other subproblems, using uniform sampling as a reference. Previous studies have shown that QAOA can generate distributions that approximately resemble Boltzmann distributions [32, 33]. We therefore estimate an effective inverse temperature from the mean energy of the samples.
Figure 3 shows a representative procedure for estimating . For each target subproblem, we generate Boltzmann samples using Kawasaki dynamics at several inverse temperatures and evaluate their mean energies. We then define as the value of whose mean energy is closest to that of the QAOA samples in the same fixed-Hamming-weight sector. For comparison, the same fitting procedure is also applied to uniform samples. As shown in the figure, the fitted value for uniform sampling is close to , consistent with the fact that a uniform random distribution corresponds to the infinite-temperature limit, . Thus, a positive indicates a lower effective temperature than uniform sampling, based on the mean energy within the same subproblem and Hamming weight sector.

Figure 3. Example of the procedure for estimating the effective inverse temperature of QAOA samples. For the selected subproblem, Boltzmann samples are generated using Kawasaki dynamics at several inverse temperatures , and their mean energies are compared with the mean energy of the QAOA samples. The fitted inverse temperature is defined as the value of whose mean energy is closest to that of the QAOA samples. The same procedure is applied to uniformly random samples for comparison, for which the fitted value is expected to be close to the infinite-temperature limit .
Figure 4 shows for different block sizes and QAOA depths . For each QAOA depth , is relatively large for small blocks and decreases as the block size increases. For larger block sizes, however, the dependence on becomes weaker and approaches an approximately constant value. This behavior is qualitatively consistent with previous hardware-native QAOA studies, which reported that the system-size dependence of shallow-QAOA performance becomes weak for sufficiently large sparse graphs [23, 34]. The effective inverse temperatures show similar trends for the representative subproblem and the other subproblems using transferred parameters, across different subproblem sizes and QAOA depths. Although the values vary among subproblems, they generally remain positive. These results suggest that parameter transfer preserves the ability of QAOA to generate samples at lower effective temperatures than uniform sampling, without separate parameter optimization for each subproblem. In most of the tested settings, the estimated effective inverse temperatures are below . Based on these results, we evaluate MCMC at higher target inverse temperatures , , , and in the next subsection.
MCMC Results
We evaluate the MCMC mixing performance by varying the block size , QAOA depth , and target inverse temperature . For each setting, we compare the proposed method with global and local Kawasaki dynamics. Mixing is quantified using the autocorrelation of the normalized Hamming distance between independent replicas. For each method and target inverse temperature, we run independent replicas for MCMC steps. For replicas and , this distance at MCMC step is
The corresponding replica overlap is . The two observables therefore have identical normalized autocorrelation functions, making this evaluation equivalent to the replica-overlap analysis used in Ref. [19]. For each replica pair, we calculate
where is the lag in MCMC steps, is the number of recorded steps, and , with denoting the time-averaged distance. We average over all replica pairs and fit the resulting mean autocorrelation to . Figure 5 shows a representative example of the measured autocorrelation as a function of lag and the corresponding exponential fits. We use the decay rate as the mixing metric: a larger indicates faster autocorrelation decay per MCMC step.

Figure 5. Representative extraction of the autocorrelation decay rate for , , subproblem size , QAOA depth , and target inverse temperature . The horizontal axis is the lag in MCMC steps, and the vertical axis is the normalized autocorrelation averaged over replica pairs. The Hamming-distance autocorrelation is equivalent to the replica-overlap autocorrelation. Solid lines show the measured autocorrelations for the proposed method, global Kawasaki dynamics, and local Kawasaki dynamics. Dashed lines show fits to , from which is estimated.
Figure 6 presents the main MCMC result. For a range of block sizes, QAOA depths, and target inverse temperatures, the neural-surrogate proposals learned from hardware-generated QAOA samples yield larger than both global and local Kawasaki dynamics. These results demonstrate that the proposed framework can achieve larger autocorrelation decay rates per MCMC step than those of these classical baselines using samples generated on current quantum hardware. The improvement is obtained despite hardware noise and the use of transferred QAOA parameters, supporting the practical potential of the hardware-efficient implementation.

Figure 6. Autocorrelation decay rate as a function of the block size for different QAOA depths and target inverse temperatures . The rates are extracted from exponential fits to the replica-pair autocorrelations, as illustrated in Figure 5. Neural-surrogate proposals trained on hardware-generated QAOA samples are compared with global and local Kawasaki dynamics. A larger indicates faster autocorrelation decay per MCMC step.
The temperature dependence is broadly consistent with the effective inverse temperatures of the QAOA samples shown in Figure 4. The target inverse temperatures considered here are generally larger than , corresponding to targets colder than the effective temperatures inferred for the QAOA samples. As the target temperature is lowered further, this effective temperature mismatch increases, and the mixing performance generally deteriorates. This trend is consistent with the proposal distributions becoming less well matched to targets that place more probability on states with lower energies.

Figure 4. Effective inverse temperature of the QAOA sample distributions as a function of the block size for different QAOA depths . For each block size, the star denotes the result for the representative subproblem on which the QAOA parameters are optimized. The diamonds show the mean over the remaining subproblems obtained using the transferred QAOA parameters, and the error bars indicate the corresponding standard deviation across these subproblems. The effective inverse temperature is estimated by comparing the mean energy of the QAOA samples with that of a Boltzmann distribution for the same subproblem and Hamming weight sector. Uniform sampling corresponds to , whereas a positive indicates a lower effective temperature.
An important difference from the previous classical simulations [19] appears in the dependence on subproblem size. Increasing does not always increase the autocorrelation decay rate, even though a larger subproblem allows more variables to be updated in a single proposal. In contrast, the previous simulations always showed improved mixing as the subproblem size increased over the smaller sizes examined there. The hardware experiments therefore reveal a block-size-dependent limitation that was not apparent in those small-scale simulations.
Figure 7 shows the fraction of MADE-generated proposals that violate the requested block Hamming weight. To investigate possible reasons for the slower autocorrelation decay at larger subproblem sizes, we examine whether the MADE proposals satisfy the required Hamming weight. Specifically, a block configuration generated from does not necessarily satisfy . The fraction of such invalid proposals increases with , and each invalid proposal is rejected without updating the Markov chain. Because each quantum sample used for training is labeled by its measured Hamming weight, the conditional training data satisfy the specified constraint. The increasing leakage therefore suggests that reproducing the Hamming-weight-conditioned distributions becomes more difficult for the MADE models at larger block sizes under the present training settings.

Figure 7. Constraint-leakage rates of MADE-generated proposals as a function of the block size , pooled over target inverse temperatures . Leakage is the fraction of proposals that do not have the required Hamming weight , i.e., . Such proposals are rejected before the Metropolis–Hastings acceptance test. Larger blocks exhibit higher leakage rates under the model and training settings used here.
This increase in leakage occurs despite a substantial increase in the number of quantum samples. As summarized in Table I, we collect 100,000, 300,000, 500,000, and 1,000,000 hardware samples per block for , 49, 82, and 144, respectively. Thus, the difficulty in reproducing the requested Hamming-weight constraint persists even with large quantum-sampling budgets.
| 2 | 5,000 | 2 |
| 12 | 30,000 | 3 |
| 30 | 100,000 | 3 |
| 49 | 300,000 | 16 |
| 82 | 500,000 | 16 |
| 144 | 1,000,000 | 32 |
Table I. QAOA sampling budgets and numbers of autoregressive orderings for each block size . is the total number of hardware samples collected per block across Hamming-weight sectors, and is the number of orderings in the proposal ensemble.
These sample counts, however, are totals across different Hamming-weight sectors and do not by themselves guarantee sufficient training data for every conditioning value. Although leakage directly causes rejected MCMC steps, it does not by itself establish how much of the observed deterioration in mixing is attributable to surrogate-learning errors. In particular, the present results do not distinguish insufficient sample coverage from limitations in model capacity or training optimization. A more detailed investigation is needed to separate these effects and assess their contributions relative to other sources of mismatch between the proposal and target distributions.
Overall, the hardware experiments demonstrate larger autocorrelation decay rates per MCMC step than those of the classical Kawasaki baselines for appropriate block sizes, while showing that larger quantum subproblems do not automatically translate into faster MCMC mixing. The increase in constraint leakage suggests that neural-surrogate learning is one factor that warrants further investigation when scaling the method. Supplementary acceptance-rate results are provided in Figure 9 in Appendix A.

Figure 9. MCMC acceptance rates as a function of the block size for different QAOA depths . The dotted green and red lines indicate the local and global Kawasaki baselines, respectively. These results supplement the autocorrelation and constraint-leakage results presented in Sec. IV C.
Large-Scale Demonstration on Quantum Hardware
Experimental Setup
We consider constrained Boltzmann sampling from in Eq. (4), with , , and . The interaction graph is constructed by arranging the brick-pattern graph in Figure 2 in a array and identifying vertices and edges shared by neighboring patches. This produces a connected graph with 1064 vertices and 1252 edges. The construction extends the same hardware-compatible connectivity pattern used in the preceding experiment. The quadratic coefficients are drawn from a standard normal distribution, , using a fixed random seed, and the linear coefficients are set to .
We choose the block size and QAOA depth . Repeating the four-block decomposition of the graph within each patch yields overlapping blocks. Each block contains 49 vertices and 54 internal edges and retains a topology compatible with the quantum processor connectivity. Thus, increasing the full problem size does not require increasing the number of qubits used in an individual QAOA circuit.
Quantum samples are generated using the IBM Quantum Heron r2 processor in ibm kobe, following the same hardware workflow as in Sec. IV. The linear-ramp parameters optimized for the representative , subproblem of the experiment are transferred to all 36 blocks. We collect 300,000 hardware samples per block, corresponding to 10.8 million samples in total. The per-block sampling budget, conditional MADE architecture, training settings, and ensemble of 16 autoregressive orderings are the same as those used for in the preceding experiment.
We evaluate mixing using the normalized Hamming-distance autocorrelation defined in Sec. IV C. For each method, we run independent replicas for MCMC steps and record the state after every step. For the analysis below, these replicas are grouped into eight disjoint replica pairs. We average their autocorrelations and fit the resulting mean to . As before, a larger indicates faster autocorrelation decay per MCMC step.
Results
We first characterize the hardware-generated QAOA samples using the mean-energy-matching procedure described in Sec. IV B. For each block, the effective inverse temperature is estimated in a fixed-Hamming-weight sector of . Averaging over all 36 blocks yields , where the uncertainty denotes the sample standard deviation across blocks rather than the standard error of the mean.
Figure 8 shows the autocorrelation functions obtained using the learned proposals and the two Kawasaki baselines. The fitted autocorrelation decay rate per MCMC step is approximately 5.2 times that of global Kawasaki dynamics and 12.3 times that of local Kawasaki dynamics.

Figure 8. Hamming-distance autocorrelation for the system at , with block size and QAOA depth . Solid lines show the mean autocorrelations over eight disjoint replica pairs, and shaded regions indicate one standard deviation across these pairs. Dashed lines show exponential fits . The proposed method achieves larger decay rates than both global and local Kawasaki dynamics.
These improvements are comparable to those obtained for at the same block size, circuit depth, and target inverse temperature, where the corresponding ratios are approximately 4.71 and 9.95. A substantial acceleration is therefore retained when the system size is increased to while keeping the quantum block size fixed.
This observation is qualitatively consistent with the divide-and-conquer simulations of Ref. [19], which found sustained improvements over Kawasaki dynamics as the system size increased at a fixed block size. For noiseless simulations on 3-regular graphs with , that study reported decay-rate improvements averaged over system sizes of approximately 7.6 and 20.3 relative to global and local Kawasaki dynamics, respectively. Although the graph structure, block size, and hardware conditions differ here, our results show that the same qualitative advantage persists for a system exceeding one thousand variables. They demonstrate that proposals learned from hardware-generated QAOA samples can accelerate constrained Boltzmann sampling on a large system using fixed-size quantum subproblems.
Conclusion
In this work, we developed a hardware-efficient implementation of neural-surrogated QeMCMC for constrained Boltzmann sampling. By combining linear-ramp QAOA with parameter transfer between subproblems with similar graph topologies, the implementation requires optimizing only four parameters on a representative subproblem and reuses them for the remaining subproblems. This avoids independent parameter optimization for every subproblem and substantially reduces the number of quantum-circuit executions required for parameter optimization.
Using an IBM Quantum Heron r2 processor, we first investigated the block-size dependence of MCMC performance for a system of binary variables. Parameter transfer generally preserved the ability of QAOA to generate samples at lower effective temperatures than uniform sampling. For suitable subproblem sizes, the learned proposals yielded larger autocorrelation decay rates per MCMC step than those of both global and local Kawasaki dynamics. By varying the block size up to , we observed a deterioration in mixing at larger block sizes that was not apparent in the previous small-scale simulations. The accompanying increase in constraint leakage from MADE-generated proposals suggests difficulties in learning high-dimensional conditional distributions under the present model and training settings. However, leakage alone does not establish surrogate learning as the dominant cause of the deterioration in mixing.
Building on these observations, we extended the full system to while fixing the block size at , for which the learned proposals had yielded larger autocorrelation decay rates than those of both Kawasaki baselines. The QAOA parameters optimized for the representative , subproblem in the experiment were transferred to all 36 blocks without further optimization. At , the fitted autocorrelation decay rate per MCMC step was approximately 5.2 times that of global Kawasaki dynamics and 12.3 times that of local Kawasaki dynamics. These improvements were comparable to those obtained for at the same block size, circuit depth, and target inverse temperature. Thus, substantial acceleration was retained when the full system was enlarged without increasing the number of variables handled by each quantum circuit or neural surrogate.
Future work should disentangle the effects of training-sample coverage, model capacity, and training optimization on constraint leakage and MCMC performance. These studies should also distinguish surrogate-learning errors from mismatches between the quantum-generated sample distributions and the target conditional distributions, and guide the development of improved neural-network models and training procedures. Overall, our results demonstrate that proposals learned from hardware-generated samples of fixed-size quantum subproblems can increase the autocorrelation decay rate per MCMC step for constrained Boltzmann sampling in a system exceeding one thousand variables.
Acknowledgments
This work was supported by MEXT Quantum Leap Flagship Program (MEXT Q-LEAP) Grant No. JP-MXS0120319794, JST COI-NEXT Grant No. JP-MJPF2014, and JST CREST Grant No. JPMJCR24I3. This paper is based on results obtained from a project, JPNP20017, commissioned by the New Energy and Industrial Technology Development Organization (NEDO).
Experimental Details
For each block , we train a conditional MADE based on the architecture of [15], using hardware-generated QAOA samples obtained through the procedure described in § 1.2. The samples are split into 90% for training and 10% for validation, and the model is trained until the loss converges. To condition the model on the block Hamming weight , we supply a one-hot representation through auxiliary context units connected to the hidden layers: the unit corresponding to is set to 1, and all others are set to 0. We use 145 context units for and three for . Table I lists the number of quantum samples collected per block, , and the number of autoregressive orderings, . The sample counts are totals across block Hamming-weight sectors, and the sampling budget is increased for larger blocks. We combine multiple autoregressive orderings to reduce dependence on a particular variable ordering [15].
For each block , the proposal distribution is the uniform mixture
where denotes the conditional distribution associated with the th ordering. A proposal is generated by selecting an ordering uniformly and sampling autoregressively from the corresponding distribution. The forward and reverse proposal probabilities in the Metropolis–Hastings acceptance test are evaluated using the full mixture likelihood in Eq. (31).
Table II lists the vertex sets defining the blocks for each block size . The vertex indices refer to the binary variables of the full system, and the family labels A and B distinguish the block collections. These blocks are used in the parameter-transfer and block-wise MCMC procedures described in § 1.2.
| Family | Vertex sets | |
| 2 | A | |
| 2 | B | |
| 12 | A | |
| 30 | A | |
| 49 | A | |
| 82 | A | |
| 144 | A |
Table II. TABLE II. Vertex sets used for the subproblem partitions of the system. Each set specifies a block of size . The notation denotes , and semicolons separate blocks. A and B denote the block families.
Figure 9 presents supplementary MCMC acceptance-rate results for the experiments in Sec. IV C. The neural-surrogate proposals are compared with global and local Kawasaki dynamics for different block sizes and QAOA depths . The reduction in acceptance at larger block sizes accompanies the deterioration in mixing discussed in the main text. These acceptance rates alone, however, do not identify the causes of this deterioration.
References
- [1]P. W. Shor, in Proceedings of the 35th Annual Symposium on Foundations of Computer Science (IEEE, 1994) pp. 124–134.
- [2]L. K. Grover, Physical Review Letters 79, 325 (1997).
- [3]D. S. Abrams and S. Lloyd, Physical Review Letters 83, 5162 (1999).
- [4]A. W. Harrow, A. Hassidim, and S. Lloyd, Physical Review Letters 103, 150502 (2009).
- [5]J. Preskill, Quantum 2, 79 (2018).
- [6]M. Cerezo et al., Nature Reviews Physics 3, 625 (2021).
- [7]E. Farhi, J. Goldstone, and S. Gutmann, arXiv preprint arXiv:1411.4028 (2014), arXiv:1411.4028 [quant-ph].arxiv.org/abs/1411.4028
- [8]J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature Communications 9, 4812 (2018).
- [9]J. F. Gonthier et al., Physical Review Research 4, 033154 (2022).
- [10]F. Arute et al., Nature 574, 505 (2019).
- [11]H.-S. Zhong et al., Science 370, 1460 (2020).
- [12]L. S. Madsen et al., Nature 606, 75 (2022).
- [13]D. Layden, G. Mazzola, R. V. Mishmash, M. Motta, P. Wocjan, J.-S. Kim, and S. Sheldon, Nature 619, 282 (2023).
- [14]Y. Nakano, H. Hakoshima, K. Mitarai, and K. Fujii, Physical Review Research 6, 033105 (2024).
- [15]M. Germain, K. Gregor, I. Murray, and H. Larochelle, in Proceedings of the 32nd International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 37 (PMLR, 2015) pp. 881–889.
- [16]B. Uria, M.-A. Côté, K. Gregor, I. Murray, and H. Larochelle, Journal of Machine Learning Research 17, 1 (2016).
- [17]H. Larochelle and I. Murray, in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research, Vol. 15 (PMLR, 2011) pp. 29–37.
- [18]Y. Nakano, K. N. Okada, and K. Fujii, PRX Quantum 7, 010338 (2026).
- [19]Y. Kawamata, Y. Nakano, and K. Fujii, arXiv preprint arXiv:2604.20701 (2026), arXiv:2604.20701 [quant-ph].
- [20]J. A. Montañez-Barrera and K. Michielsen, npj Quantum Information 11, 31 (2025).
- [21]A. Galda, X. Liu, D. Lykov, Y. Alexeev, and I. Safro, in 2021 IEEE International Conference on Quantum Computing and Engineering (QCE 2021) pp. 171–180.
- [22]A. Galda, E. Gupta, J. Falla, X. Liu, D. Lykov, Y. Alexeev, and I. Safro, Frontiers in Quantum Science and Technology 2, 1200975 (2023).
- [23]E. Pelofske, A. Bärtschi, and S. Eidenbenz, in High Performance Computing, Lecture Notes in Computer Science (Springer, 2023) pp. 240–258, arXiv:2301.00520.
- [24]J. A. Montanez-Barrera, K. Michielsen, and D. E. Bernal Neira, arXiv preprint arXiv:2502.06471 (2025), arXiv:2502.06471 [quant-ph].
- [25]D. Bucher, M. Janetschek, M. Poppel, J. Stein, C. Linnhoff-Popien, and S. Feld, arXiv preprint arXiv:2604.02083 (2026), arXiv:2604.02083 [quant-ph].DOI
- [26]S. Hadfield, Z. Wang, B. O’Gorman, E. G. Rieffel, D. Venturelli, and R. Biswas, Algorithms 12, 34 (2019).
- [27]Z. Wang, N. C. Rubin, J. M. Dominy, and E. G. Rieffel, Physical Review A 101, 012320 (2020).
- [28]RIKEN, IBM and RIKEN unveil first IBM Quantum System Two outside of the U.S., RIKEN Center for Computational Science (2025), accessed: 2026-08-17.
- [29]Z. Wang, J. D. Teske, A. Bhardwaj, M. O. Takahashi, and S. Yunoki, arXiv preprint arXiv:2606.27856 (2026), arXiv:2606.27856 [quant-ph].arxiv.org/abs/2606.27856
- [30]H. Xu, T. Shirakawa, and S. Yunoki, arXiv preprint arXiv:2605.28040 (2026), arXiv:2605.28040 [quant-ph].arxiv.org/abs/2605.28040
- [31]K. Shinjo, K. Seki, and S. Yunoki, arXiv preprint arXiv:2603.11910 (2026), arXiv:2603.11910 [quant-ph].
- [32]P. Díez-Valle, D. Porras, and J. J. García-Ripoll, Physical Review Letters 130, 050601 (2023).
- [33]P. C. Lotshaw, G. Siopsis, J. Ostrowski, R. Herrman, R. Alam, S. Powers, and T. S. Humble, Physical Review A 108, 042411 (2023).
- [34]M. P. Harrigan, K. J. Sung, M. Neeley, K. J. Satzinger, F. Arute, K. Arya, J. Atalaya, J. C. Bardin, R. Barends, S. Boixo, M. Broughton, B. B. Buckley, D. A. Buell, B. Burkett, N. Bushnell, Y. Chen, Z. Chen, B. Chiaro, R. Collins, W. Courtney, et al., Nature Physics 17, 332 (2021).