⚛️ quantum physics

Hardware Demonstration of Neural-Surrogated Quantum-Enhanced MCMC for Large-Scale Constrained Boltzmann Sampling

This paper presents a hardware-efficient implementation of neural-surrogated quantum-enhanced MCMC on an IBM Quantum Heron r2 processor, demonstrating significantly accelerated autocorrelation decay for large-scale constrained Boltzmann sampling with up to 1,064 binary variables.

Original author: Yuya Kawamata, Keisuke Fujii

Published 2026-11-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original author: Yuya Kawamata, Keisuke Fujii

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast landscape of modern computing, a persistent challenge involves finding the best possible arrangement among countless options. Imagine trying to organize a massive library where every book must be placed on a specific shelf, but the rules for placement are so complex that checking every single possibility would take longer than the age of the universe. This is the realm of optimization, a field where scientists and engineers seek the most efficient solution to problems ranging from logistics to drug design. For decades, the standard approach has relied on classical computers to explore these possibilities step by step, often getting stuck in local valleys of inefficiency. However, a new frontier has emerged with the advent of quantum computers, machines that use the strange laws of physics to process information in ways classical machines cannot. While today's quantum devices are still noisy and limited in size, researchers are learning how to combine their unique strengths with powerful classical algorithms to solve problems that were previously out of reach.

A team of researchers has now demonstrated a practical way to bridge this gap, using a small quantum processor to help a classical computer solve a massive sampling problem. The goal was to generate random samples from a complex probability distribution, a task essential for understanding how systems behave under specific constraints. To do this, they employed a technique called Markov chain Monte Carlo, a method that builds a sequence of random steps to explore a landscape of possibilities. The innovation lies in how they guided these steps. Instead of letting the computer wander blindly, they used a quantum computer to generate high-quality "proposals" for the next step in the sequence. These proposals were then distilled into a neural network, a type of artificial intelligence that learned the patterns of the quantum samples. This hybrid system allowed the researchers to update large groups of variables simultaneously, moving through the solution space much faster than traditional methods.

The team tested this approach on an IBM Quantum Heron r2 processor, a device containing 156 superconducting qubits. They began with a system of 144 binary variables, which can be thought of as switches that are either on or off. The challenge was to find arrangements of these switches that met a specific constraint: the total number of "on" switches had to remain constant. The researchers broke this large problem into smaller, overlapping pieces. For one representative piece, they tuned the quantum computer to generate optimal samples. Remarkably, they found that the settings used for this single piece worked well for all the other pieces, even though the pieces were different. This "parameter transfer" meant they did not need to spend valuable time re-tuning the quantum machine for every single sub-problem, a crucial step that made the process efficient enough to run on current hardware.

The results showed that this method significantly accelerated the exploration of the solution space. When the researchers used the neural network, trained on the quantum-generated samples, to guide the classical computer, the system's ability to forget its starting point and reach a true random state improved dramatically. Specifically, the rate at which the system's memory of its past faded was about five times faster than using a standard global method and more than twelve times faster than a local method. This speedup held true even when the researchers scaled the problem up to a system with 1,064 variables, a size far beyond the capacity of the quantum processor itself. By keeping the quantum sub-problems small and reusing the learned settings, they successfully demonstrated constrained Boltzmann sampling with faster autocorrelation decay rates on a system more than seven times larger than the quantum device could handle directly.

However, the study also revealed important limitations that must be addressed for future growth. As the researchers increased the size of the sub-problems they tried to solve with the quantum computer, the performance did not continue to improve indefinitely. Instead, it began to degrade. The neural networks struggled to learn the complex patterns required for these larger blocks, leading to a higher rate of rejected proposals. This suggests that while the quantum computer can provide a powerful boost, the classical neural network must be able to accurately mimic the quantum behavior, and this becomes harder as the complexity of the sub-problem increases. The researchers found that simply making the quantum sub-problems larger did not automatically lead to better results; there is a sweet spot where the quantum advantage is maximized before the difficulty of learning the patterns overwhelms the system.

The significance of this work lies in its demonstration of a scalable path forward. It proves that we do not need a massive, error-corrected quantum computer to gain an advantage today. Instead, by using a small quantum device to generate high-quality samples for a few representative problems and then transferring that knowledge to a larger system, we can demonstrate efficient sampling for problems involving thousands of variables. The study confirms that the quantum advantage is not just a theoretical possibility but a measurable reality that can be extended to large-scale systems. While challenges remain in refining the learning process for larger blocks, the core finding is clear: a hybrid approach, where a small quantum processor acts as a specialized engine for a larger classical system, can deliver substantial speedups in constrained Boltzmann sampling for complex problems. This opens a practical door for using today's imperfect quantum machines to tackle real-world optimization challenges that are currently too difficult for classical computers alone.

Technical Summary: Hardware Demonstration of Neural-Surrogated Quantum-Enhanced MCMC for Large-Scale Constrained Boltzmann Sampling

Problem Statement
The paper addresses the challenge of performing Boltzmann sampling for large-scale constrained systems (specifically, binary variables with a fixed Hamming weight) using current Noisy Intermediate-Scale Quantum (NISQ) devices. While Quantum-Enhanced Markov Chain Monte Carlo (QeMCMC) offers a pathway to accelerate sampling by using quantum-generated proposals, scaling this approach to systems with thousands of variables is hindered by the high cost of optimizing Quantum Approximate Optimization Algorithm (QAOA) parameters for every subproblem in a divide-and-conquer strategy. Independent optimization for numerous subproblems requires excessive quantum circuit executions, making large-scale implementation impractical on current hardware.

Methodology
The authors propose a hardware-efficient implementation of neural-surrogated QeMCMC that reduces the quantum resource overhead through two primary mechanisms:

Linear-Ramp QAOA and Parameter Transfer: Instead of optimizing the full set of variational parameters for every subproblem, the authors employ a linear-ramp QAOA ansatz, which reduces the parameter space to just four independent variables ( $\gamma_0, \Delta\gamma, \beta_0, \Delta\beta$ ). They optimize these four parameters on a single representative subproblem and transfer them to all other subproblems. This strategy exploits the transferability of QAOA parameters across problem instances with similar graph topologies and optimization landscapes, eliminating the need for repeated parameter optimization on the quantum processor.
Divide-and-Conquer with Neural Surrogates: The large system is divided into overlapping blocks (subproblems). For each block, QAOA samples are generated using the transferred parameters. These samples are then distilled into conditional Masked Autoencoders for Distribution Estimation (MADEs). These neural networks learn the conditional probability distribution of the block variables given the block's Hamming weight.
MCMC Execution: During the Markov Chain Monte Carlo process, a block is selected, and the corresponding conditional MADE proposes a new configuration. The proposal is accepted or rejected based on the Metropolis–Hastings criterion using the full-system energy function, ensuring the chain converges to the target constrained Boltzmann distribution.

Key Contributions

Hardware-Efficient Scaling: The work demonstrates a method to scale QeMCMC to systems with $N > 1000$ variables by drastically reducing the number of quantum-circuit executions required for parameter optimization (from $M$ optimizations to just one).
Hardware Validation: The method is implemented and tested on an IBM Quantum Heron r2 processor (156 qubits), moving beyond purely classical simulations.
Empirical Analysis of Block Size: The paper provides experimental data on how the performance of neural-surrogated QeMCMC scales with block size on real hardware, identifying a non-monotonic relationship between block size and mixing speed that differs from previous classical simulation results.

Experimental Results
The study consists of two main experimental phases:

$N = 144$ System (Block-Size Scaling):
- Experiments were conducted on a system of 144 binary variables with varying subproblem sizes ( $b \in \{2, 12, 30, 49, 82, 144\}$ ).
- Parameter Transfer Efficacy: The transferred QAOA parameters successfully generated samples at lower effective temperatures than uniform sampling across different block sizes, validating the transferability assumption.
- Mixing Performance: For appropriate subproblem sizes (e.g., $b=49$ ), the neural-surrogate proposals achieved faster autocorrelation decay rates compared to both global and local Kawasaki dynamics.
- Limitations: Unlike previous classical simulations which suggested continuous improvement with larger block sizes, the hardware experiments showed a deterioration in mixing performance for larger blocks ( $b > 49$ ). This was attributed to an increase in constraint leakage, where the conditional MADEs failed to generate proposals satisfying the required block Hamming weight, leading to higher rejection rates.
$N = 1064$ System (Large-Scale Demonstration):
- The method was extended to a system of 1064 binary variables using a $3 \times 3$ array of the $N=144$ topology.
- Using the optimal block size ( $b=49$ ) identified in the first experiment and transferring the same optimized parameters, the authors demonstrated constrained Boltzmann sampling on the larger system.
- Performance Gain: The autocorrelation decay rate per MCMC step for the proposed method was approximately 5.2 times faster than global Kawasaki dynamics and 12.3 times faster than local Kawasaki dynamics.
- These improvements were comparable to those observed in the $N=144$ system, indicating that the acceleration is maintained even when the total system size exceeds one thousand variables, provided the quantum subproblem size remains fixed.

Significance and Claims
The paper claims that this work demonstrates the feasibility of using fixed-size quantum subproblems to accelerate sampling in large-scale systems ( $N > 1000$ ) via neural surrogates. The significance lies in proving that:

The "divide-and-conquer" approach can be made practical on real NISQ hardware by leveraging parameter transfer, avoiding the prohibitive cost of independent optimization for every subproblem.
Quantum-generated samples, even with hardware noise and transferred parameters, can provide nonlocal updates that significantly accelerate MCMC mixing compared to classical baselines.
While the method is effective, scaling the block size indefinitely is not a panacea; the study highlights the critical role of neural surrogate learning quality (specifically, the ability to satisfy hard constraints like Hamming weight) as a limiting factor for larger subproblems.

The authors conclude that while the current implementation shows substantial acceleration, future work is needed to disentangle the effects of training data coverage, model capacity, and training optimization on constraint leakage to further improve the method's scalability.

Technical Summary: Hardware Demonstration of Neural-Surrogated Quantum-Enhanced MCMC for Large-Scale Constrained Boltzmann Sampling

More like this