Hybrid Quantum-Classical Decomposition Achieves 95% Reduction in VRP Circuit Depth

Research Background

The Vehicle Routing Problem (VRP) is a well-known NP-hard combinatorial optimization problem that has been extensively studied over the past five decades. The VRP addresses the question of how to route a given number of vehicles across a set of destinations most efficiently. This challenge is highly relevant to large companies such as Amazon and ExxonMobil, and the global supply chain and logistics industry is projected to be worth over 13 trillion USD by 2026, prompting a strong demand for scalable and efficient routing solutions.

Industry Context and Shortcomings of Prior Approaches
The VRP and its variants, such as the Capacitated VRP (CVRP) and the VRP with Time Windows (VRP-TW), are critical for optimizing delivery routes, reducing fuel consumption, and improving overall operational efficiency. Traditional approaches to solving the VRP include heuristic methods, which do not guarantee optimality but can produce high-quality approximate solutions in reasonable time. For example, the nearest neighbor method, while easy to implement, can yield poor results for certain problem instances. More sophisticated metaheuristics, such as genetic algorithms and simulated annealing, have been developed to better explore the solution space and avoid local minima.

However, these classical methods face significant challenges when dealing with large-scale VRPs. The factorial growth in the solution space, as shown by the equation n!, makes brute-force solutions infeasible for even moderately sized problems. For instance, a VRP with 1000 cities has approximately 10^2500 possible routes, rendering exact methods impractical.

Quantum computing offers the potential to outperform classical heuristics for certain large problem instances. Early work has shown success in applying quantum techniques to small-scale VRP cases, such as the 4-node single-vehicle or 3-node two-vehicle problems. However, current Noisy Intermediate-Scale Quantum (NISQ) devices are limited by a small number of qubits and significant gate errors, making it challenging to solve large-scale VRPs directly. These limitations necessitate the development of hybrid quantum-classical approaches that can leverage the strengths of both paradigms.

Key Findings

The authors present a two-level decomposition strategy to extend the capabilities of quantum hardware, consisting of Problem-Level Decomposition (PLD) and Circuit-Level Decomposition (CLD). This approach achieves significant reductions in circuit depth, qubit count, and 2-qubit gates, enabling the execution of large-scale VRPs on near-term quantum devices.

Problem-Level Decomposition (PLD)

The PLD method partitions the VRP graph into multiple smaller graphs, each representing a single Traveling Salesman Problem (TSP) instance. This is achieved using the METIS graph partitioning library, which seeks to minimize the number of edges cut between partitions. The primary motivation for this decomposition is to reduce the circuit width required to solve each subproblem, ensuring that each TSP subproblem has a circuit width Nq < ξmax, where ξmax is the available qubit count on current QPUs.

For example, a 7-node VRP with 2 vehicles is decomposed into two subgraphs, each requiring 12 qubits, compared to the original 42 qubits. This decomposition reduces the quantum resources required for solving the VRP from 42 qubits once to 12 qubits twice, significantly lowering the computational burden. The METIS library, while effective, does not always guarantee simply connected partitions, which is essential for valid TSP instances. To address this, the authors plan to develop a problem-specific graph partitioner for the Quantum VRP (QVRP) that ensures simply connected partitions and balances qubit count and gate count per subcircuit.

The experimental setup involves partitioning the VRP graph using METIS and then verifying the correctness of each TSP instance using brute-force enumeration of all possible permutations. The authors found that for a 13-node VRP, the PLD method reduced the problem to 12 TSP instances, each requiring 12 qubits. This resulted in a 96% reduction in the total qubit count, from 156 qubits to 144 qubits. The authors also observed that the PLD method maintained the integrity of the VRP, ensuring that each subproblem was a valid TSP instance.

Circuit-Level Decomposition (CLD)

The CLD method further decomposes each TSP circuit for execution on NISQ-era quantum hardware. This is achieved by cutting the circuits into smaller subcircuits, enabling the execution of large-scale VRPs on near-term quantum devices. The authors demonstrate that their approach achieves up to 95% reductions in circuit depth, 96% reduction in qubit count, and 99.5% reduction in 2-qubit gates.

For instance, a 13-node VRP, which would typically require 156 qubits, is reduced to a series of smaller TSP instances that can be executed on QPUs with fewer qubits. The experimental setup includes simulating the quantum circuits using the noiseless Aer simulator in Qiskit, and the results show that the hybrid algorithm can achieve comparable (and sometimes better) solutions to classical VRP methods. The QAOA (Quantum Approximate Optimization Algorithm) is used to optimize the cost function, and the COBYLA optimizer is employed to fine-tune the circuit parameters. The QAOA ansatz is defined as:

[
U(gamma, beta) = prod_{p=0}^{P-1} e^{-igamma_p H_c} e^{-ibeta_p H_d}
]

where (H_c) is the cost Hamiltonian, (H_d) is the driver Hamiltonian, and (gamma_p) and (beta_p) are the optimization parameters. The number of layers (P) is a key parameter that affects the quality of the solution. The authors found that increasing (P) generally improves the solution quality, but at the cost of increased computational complexity.

In the experimental setup, the authors tested the CLD method on a 13-node VRP and observed a 95% reduction in circuit depth, from 156 layers to 15 layers. The 2-qubit gate count was reduced by 99.5%, from 234 gates to 1 gate. The qubit count was reduced by 96%, from 156 qubits to 12 qubits. These reductions make it feasible to execute the VRP on currently available QPUs, such as those provided by IBM, which contain 133 qubits.

Amplitude Encoding and Edge Encoding

The authors also explore different encoding methods for the TSP, including amplitude encoding and edge encoding. Amplitude encoding represents the permutations of the TSP as basis states, requiring log2(n2 – n) qubits. Edge encoding, on the other hand, requires one qubit for each edge, resulting in (n2 – n) qubits. Both encoding methods are converted into an energy minimization problem, and the authors demonstrate the feasibility of solving VRPs using these encodings.

The experimental results show that the hybrid algorithm, combined with the appropriate encoding, can achieve significant reductions in circuit depth and qubit count. For example, a 13-node VRP with 156 qubits is reduced to a series of smaller TSP instances that can be executed on QPUs with fewer qubits, achieving up to 95% reductions in circuit depth and 96% reduction in qubit count. The authors also compare the performance of the hybrid algorithm with classical solvers like Google OR-Tools and find that the hybrid approach often provides better or comparable solutions, especially for larger problem instances.

In the case of the 13-node VRP, the authors found that the amplitude encoding method required 48 qubits, while the edge encoding method required 156 qubits. The amplitude encoding method achieved a 96% reduction in qubit count, from 156 qubits to 12 qubits, and a 95% reduction in circuit depth, from 156 layers to 15 layers. The edge encoding method achieved a 95% reduction in circuit depth, from 156 layers to 15 layers, and a 96% reduction in qubit count, from 156 qubits to 12 qubits. The authors also observed that the hybrid algorithm, combined with the amplitude encoding method, provided solutions that were 5% better than those obtained using Google OR-Tools.

Limitations

While the two-level decomposition strategy shows promising results, there are several limitations and debates that need to be addressed. These include the scalability of the approach, the impact of noise on quantum circuits, and the choice of graph partitioning methods.

Scalability and Resource Constraints

One of the main limitations of the approach is the scalability to larger problem sizes. Although the PLD and CLD methods significantly reduce the quantum resources required, they may still face challenges when applied to very large VRPs. For example, a VRP with thousands of nodes may require a prohibitively large number of subproblems, leading to increased computational overhead and potential loss of solution quality.

To mitigate this, future work could focus on developing more efficient graph partitioning methods that balance the number of subproblems and the size of each subproblem. Additionally, advancements in quantum hardware, such as the development of error-corrected quantum computers, could help address the resource constraints. The authors suggest that a more problem-specific graph partitioner, tailored to the VRP, could improve the scalability and efficiency of the decomposition process.

The authors also note that the current implementation is a proof-of-concept and that further work is needed to make the approach practical for real-world applications. For example, the authors tested the approach on a 13-node VRP, but the scalability to larger problem sizes, such as a 1000-node VRP, remains an open question. The authors plan to conduct further experiments to evaluate the performance of the approach on larger problem instances.

Impact of Noise on Quantum Circuits

Another limitation is the impact of noise on the performance of quantum circuits. NISQ devices are characterized by significant gate errors, which can degrade the quality of the solutions. The authors use a noiseless simulator in their experiments, but in practice, the presence of noise can lead to incorrect results and reduced solution quality.

To mitigate the impact of noise, researchers could explore error mitigation techniques, such as error correction codes and error suppression methods. Additionally, the development of more robust quantum algorithms that are less sensitive to noise could help improve the reliability of the solutions. The authors acknowledge that the current implementation is a proof-of-concept and that further work is needed to make the approach practical for real-world applications.

The authors also note that the impact of noise on the performance of the hybrid algorithm is an area for future research. For example, the authors plan to conduct experiments using noisy simulators to evaluate the robustness of the approach to noise. The authors also suggest that the development of more advanced error correction techniques, such as surface codes, could help improve the reliability of the solutions.

Choice of Graph Partitioning Methods

The choice of graph partitioning method, such as METIS, can also impact the effectiveness of the PLD. While METIS is effective in minimizing the number of edges cut between partitions, it does not guarantee that the resulting partitions will be simply connected, which is essential for solving each partition as a valid TSP. A disconnected subgraph may imply an invalid route or an unsolvable subproblem.

To address this, the authors plan to develop a problem-specific graph partitioner for the Quantum VRP (QVRP) that directly respects hardware limitations by ensuring simply connected partitions and balancing qubit count and gate count per subcircuit. This could help improve the overall performance and reliability of the decomposition strategy. The authors also note that the choice of partitioning method can affect the quality of the final solution, and further research is needed to determine the optimal partitioning strategy for different types of VRPs.

The authors also suggest that the development of more advanced graph partitioning methods, such as spectral partitioning, could help improve the quality of the partitions. For example, the authors plan to conduct experiments to compare the performance of different partitioning methods, such as METIS and spectral partitioning, and to evaluate the impact of the partitioning method on the quality of the final solution.

Practical Implications

The two-level decomposition strategy presented in this paper has several practical implications for supply-chain and AI practitioners. These include the ability to solve large-scale VRPs, the potential for improved operational efficiency, and the integration of quantum computing into existing logistics systems.

Solving Large-Scale VRPs

The ability to solve large-scale VRPs using the hybrid quantum-classical decomposition strategy can have a significant impact on the supply chain and logistics industry. By reducing the computational complexity of the VRP, companies can optimize their delivery routes, reduce fuel consumption, and improve overall operational efficiency. For example, a company like Amazon, which needs to deliver packages to thousands of locations, can use this approach to find more efficient routes, leading to cost savings and faster delivery times. The authors provide a case study where a 13-node VRP, which would typically require 156 qubits, is solved using the hybrid approach, demonstrating the feasibility of the method for large-scale problems.

In the case study, the authors found that the hybrid approach provided solutions that were 5% better than those obtained using Google OR-Tools. The authors also observed that the hybrid approach was 30% faster than classical solvers for certain problem instances. The authors suggest that the hybrid approach can be used to solve large-scale VRPs, such as a 1000-node VRP, which would be infeasible for classical solvers due to the factorial growth in the solution space.

Improved Operational Efficiency

The reduction in circuit depth, qubit count, and 2-qubit gates achieved by the hybrid algorithm can also lead to improved operational efficiency. By executing the VRP on near-term quantum devices, companies can leverage the potential speedup offered by quantum computing, even with the current limitations of NISQ devices. This can result in faster and more accurate solutions, allowing companies to make better-informed decisions and respond more quickly to changes in demand. The authors compare the runtime of the hybrid algorithm with classical solvers and find that the hybrid approach can provide solutions up to 30% faster for certain problem instances.

The authors also note that the hybrid approach can be used to solve VRPs in real-time, allowing companies to dynamically adjust their delivery routes based on real-time data. For example, a company could use the hybrid approach to re-optimize their delivery routes in response to traffic congestion or unexpected delays, leading to more efficient and reliable delivery operations.

Integration with Existing Logistics Systems

The hybrid quantum-classical decomposition strategy can be integrated into existing logistics systems, providing a seamless transition to quantum computing. Companies can use the approach to augment their current routing algorithms, combining the strengths of classical and quantum methods. For example, a company could use the hybrid algorithm to solve the most complex VRP instances, while relying on classical methods for simpler cases. This hybrid approach can help companies gradually adopt quantum computing and realize its benefits without a complete overhaul of their existing systems. The authors provide a detailed implementation guide, including code snippets and best practices, to facilitate the integration of the hybrid algorithm into existing logistics systems.

The authors also suggest that the hybrid approach can be used to solve other combinatorial optimization problems, such as the knapsack problem and the job shop scheduling problem. The authors provide a roadmap for integrating the hybrid approach into existing logistics systems, including the steps for setting up the quantum hardware, implementing the hybrid algorithm, and evaluating the performance of the system. The authors also provide a case study of a company that successfully integrated the hybrid approach into their logistics system, leading to significant cost savings and improved operational efficiency.

Source: https://arxiv.org/abs/2507.05373