Mojtaba Asadollahi, Danial Khajehzadeh and Abolfazl Ramezanpour of Shiraz University report that a quantum annealing network with a width of just W = 2 can exponentially reduce the computational steps needed to solve quantum search problems. Their work extends quantum annealing to a quantum neural network, revealing that network “width”, the learnable evolution space, is critical for enhancing the algorithm’s success. The researchers found the search problem can be solved with exponentially small errors using a quantum perceptron of depth L = 1 and width W = 2, demonstrating a potentially highly efficient configuration.
Quantum Annealing Limitations in Discontinuous Phase Transitions
This improvement occurs despite limitations inherent in discontinuous quantum phase transitions, where the energy gap, a key factor in annealing success, shrinks exponentially with system size, traditionally demanding exponentially longer computation times. The research demonstrates a pathway to enhance performance not by increasing the complexity of the network, but by optimizing its width, the learnable evolution space within the annealing process.
The core of the limitation lies in the behavior of the energy gap during discontinuous quantum phase transitions; as system size increases, this gap diminishes rapidly, necessitating exponentially larger annealing times to reliably reach the ground state. Researchers observed this phenomenon in a system defined by the Hamiltonian = -N(1/Nsum_(i = 1)^Nσ_i^z)^3, where the energy gap becomes exponentially small around ~ 0. 44, hindering efficient annealing.
This observation prompted an exploration of quantum neural networks as a means to circumvent this scaling issue and expand the available evolution space for the algorithm. Specifically, the minimum depth, , needed to achieve a desired success probability in a quantum search problem scales from to N, representing a significant reduction in computational cost.
This suggests that a relatively small network width can unlock substantial performance gains by providing more dynamical paths for the annealing process, a departure from approaches focused solely on increasing network depth. “It seems that a linear annealing of width W provides a larger number of dynamical paths to the annealing process than the Grover annealing,” the researchers note, highlighting the benefit of expanding the evolution space.
This is particularly evident when comparing the performance to Grover annealing, where a large number of layers remain nearly identical near the phase transition and contribute little to expanding the evolution space. The researchers acknowledge the need for more powerful learning algorithms and classical simulations to extend these findings to larger systems, and are currently exploring a similar extension of classical simulated annealing to a neural network framework.
They anticipate that, as in the quantum case, the width of the network will prove beneficial in mitigating the destructive effects of classical phase transitions by enlarging the search space. Further investigation will focus on refining these techniques and applying them to more complex optimization problems, potentially unlocking new capabilities for quantum annealing in the face of challenging computational landscapes.
Quantum Neural Network Architecture for Enhanced Annealing
This extension of the quantum annealing process incorporates a defined “width” at each layer of the annealing dynamics, effectively creating multiple evolution paths and increasing the probability of bypassing energy gaps that typically hinder optimization. The architecture, illustrated in the paper, moves beyond standard quantum annealing, which is recovered when W = 1, by introducing learnable parameters at each layer.
The researchers evaluated the performance of this network using the ferromagnetic p-spin model and the quantum search problem, employing exact numerical simulations to quantify the impact of network width. This contrasts with the Grover algorithm, where the quantum evolution follows a linear progression. The network’s evolution is non-linear, offering a different approach to problem-solving.
Figures within the study compare the performance of standard quantum annealing with both linear and Grover annealing schedules against the quantum neural network for varying widths, W. The weights within the neural network, , are assigned as independent random variables uniformly distributed between zero and one, while the phases, , are drawn from a normal distribution with a mean of zero and a variance of one. This randomization is a key component of the network’s ability to explore a wider range of potential solutions.
Width Two Networks Enable Exponential Search Reduction
Simulations indicate that a quantum perceptron with a remarkably small architecture, a width of just two, can solve quantum search problems with exponentially small errors. The simulations, focused on the quantum search problem, reveal that tuning network parameters is key to this efficiency. Researchers observed that a randomly constructed quantum annealing network with linear annealing can exponentially reduce the number of evolution layers or depth of the annealing process in the quantum search problem.
The team used the “differential_evolution” and “minimize” methods of “scipy.optimize” to identify optimal solutions for these parameters. These parameters, detailed in a table within the published work, demonstrate the network’s ability to adapt to varying problem complexities. While the simulations show promise, the researchers acknowledge the need to move beyond the mean-field approximation to accurately model larger systems. This work suggests that focusing on network width, alongside depth, is a promising avenue for improving the efficiency and scalability of quantum and classical annealing algorithms.
Quantum Perceptron Depth and Success Probability
A quantum perceptron requiring only a single layer, depth L = 1, can solve a quantum search problem with exponentially small errors when its width, W, is set to 2, a configuration demonstrating unexpectedly efficient performance. This finding challenges conventional optimization strategies that prioritize increasing the depth of quantum annealing networks, suggesting that network width is a critical, and often overlooked, parameter. The work reveals that a wider network provides a significantly expanded and learnable evolution space for the annealing process, leading to substantial improvements in success probability.
The performance boost associated with increased width extends beyond the specific case of the quantum search algorithm; simulations indicate benefits across a range of optimization problems. Unlike the Grover algorithm, where numerous layers can remain largely unchanged near a phase transition, a wider network offers a greater number of dynamical paths for the annealing process to explore.
This expanded search space allows the algorithm to more effectively navigate complex landscapes and identify optimal solutions. The observed relationship between width and performance is not merely a matter of providing more computational resources; the architecture itself, the way the network is structured, enhances its ability to find solutions.
The relative error of an optimal quantum perceptron was also compared to that of the optimal quantum annealing, with the perceptron exhibiting slower error growth as the problem size increased. the success probability of achieving the correct superposition is inversely proportional to W raised to the power of L, or 1/W^L, in the worst-case scenario of orthogonal states. This mathematical relationship highlights the importance of keeping both depth and width growth slower than the system size, N, to maintain computational feasibility.
Superposition of States Improves Robustness and Speedup
A quantum annealing network, even one with a minimal width of two, can substantially reduce the computational layers needed for quantum search, shifting expectations for efficient network design. This suggests that carefully structuring the network’s evolution space, rather than solely increasing its depth, can yield significant speedups in complex optimization tasks. The architecture of this quantum neural network (QNN) avoids the complexities introduced by reinforcement learning, instead relying on the superposition of states from preceding layers to model the process.
While the complexity of the network grows exponentially with depth as , where W is the network width and L is the number of layers, this growth is managed by the network’s structure and optimized parameters. The researchers found that even a randomly generated network employing linear annealing can achieve this reduction in layer requirements, highlighting the potential for inherent efficiency in the network’s design.
This outcome contrasts with the slower annealing process observed around quantum phase transitions where the coefficient = = 1/2, indicating a sensitivity to annealing parameters. The initial state for the dynamics begins with a product state, a superposition of all qubits in the | + ⟩ state, and evolves through a series of unitary operations.
Each node at layer l generates a superposition of normalized product states, ultimately contributing to the computation of a loss function based on the original problem Hamiltonian, E =. Detailed calculations regarding this process are available in Appendix B. The number of unitaries required is of order L( + W), and each unitary can be implemented with approximately N/ε local gates using the Trotter approximation with an error of order , further detailing the computational resources involved.
The cumulative distribution of minimum energies found by the mean-field random neural network in the random spin-glass model provides further insight into the network’s performance. This exponential growth, while a potential limitation, is mitigated by the network’s structure and the ability to optimize its parameters, suggesting a pathway toward more efficient quantum annealing.
👉 More information
🗞 Quantum Annealing with a Quantum Neural Network: Width of Annealing Matters
✍️ Mojtaba Asadollahi, Danial Khajehzadeh and Abolfazl Ramezanpour (Department of Physics, College of Science, Shiraz University, Shiraz, Iran)
🧠 ArXiv: https://arxiv.org/abs/2611.24089v1




See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
