Quantum Simulated Annealing

• Quantum simulated annealing is a framework that leverages quantum walk dynamics and quantum tunneling to achieve quadratic speedup over traditional annealing methods.
• It encompasses fully quantum, hybrid, and stochastic variants that integrate quantum-enhanced proposals and Monte Carlo techniques to overcome local minima.
• Practically, QSA is applied in combinatorial optimization, sampling, and inference, demonstrating efficiency improvements in problems like maximum independent set and non-convex quadratic programming.

Quantum simulated annealing (QSA) encompasses a family of quantum algorithms and quantum-inspired classical algorithms designed to accelerate optimization and sampling tasks by exploiting quantum walk dynamics, quantum fluctuations, or quantum-enhanced proposals. QSA can refer both to fully quantum implementations that yield quadratic speedups relative to classical simulated annealing (SA) and to hybrid or classical algorithms that mimic aspects of quantum annealing, such as quantum tunneling. QSA techniques span theoretical frameworks for discrete and continuous optimization, rigorous complexity analyses, experimental quantum-classical hybrids, and classical path-integral Monte Carlo emulations of quantum dynamics.

1. Foundational Principles and Algorithmic Construction

QSA fundamentally extends the classical SA protocol, which sequentially applies Markov transition kernels (stochastic matrices) that satisfy detailed balance with respect to a Boltzmann-Gibbs distribution at an annealing temperature. The classical process relies on thermal fluctuations to escape local minima and is bottlenecked by the minimal spectral gap $\Delta$ of the stochastic transition matrices along the annealing path, resulting in overall runtime $O(1/\Delta)$ for hard instances (Boixo et al., 2015).

QSA algorithms, exemplified by the Somma-Ortiz-Batista protocol, “quantize” each stochastic matrix $S(\beta)$ (at inverse temperature $\beta$) to a unitary quantum walk $W(\beta)$ whose phase gap scales as $\sqrt{\Delta}$. The adiabatic-like evolution is orchestrated through a sequence of quantum walk operators combined in randomized phase estimation steps (“Zeno projection”), gradually deforming an initial uniform superposition to the low-temperature Gibbs quantum state:

• On an $n$-qubit system encoding $d=2^n$ configurations, QSA prepares the state

$$|\psi_{\beta_m}\rangle = \sum_{i=0}^{d-1} \sqrt{\pi_{\beta_m}(\sigma_i)}\, |\sigma_i\rangle |0\rangle$$

from an initial uniform superposition at infinite temperature.

• The runtime for preparing this state and sampling an optimal solution scales as $O(\log(d)\,E_{\max}/\sqrt{\Delta})$, representing a quadratic improvement over the $O(1/\Delta)$0 runtime of classical SA (Boixo et al., 2015, Harrow et al., 2019).

The QSA circuit comprises:

• An isometry $O(1/\Delta)$1 encoding transition probabilities into amplitudes,
• The SWAP operator, reflection about $O(1/\Delta)$2, and
• The quantum walk composition $O(1/\Delta)$3, which inherits spectral properties directly from the underlying Markov process.

This scheme is theoretically rigorous for discrete, sparse, reversible Markov chains; the quadratic speedup is guaranteed under a known spectral gap and efficient oracle access to nonzero transitions (Boixo et al., 2015, Harrow et al., 2019).

2. Variants: Hybrid and Stochastic Quantum Simulated Annealing

Beyond purely quantum implementations, QSA encompasses a broad spectrum of quantum-classical hybrid and classical stochastic algorithms that leverage quantum or quantum-inspired mechanisms to enhance annealing:

• Quantum-Enhanced Simulated Annealing (QESA): Combines quantum subroutines (e.g., Rydberg atom arrays or quantum annealers) to produce “warm starts” for a classical SA loop, reducing the number of epochs required to reach target approximation ratios, as empirically demonstrated on the maximum independent set (MIS) problem (Jeong et al., 16 Jun 2025). In continuous optimization domains, QESA utilizes quantum annealing hardware for low-energy search directions inside a classical SA loop, showing superior solution quality and scalability compared to classical or pure quantum approaches in non-convex quadratic programming (Djidjev, 2 Apr 2025).
• Stochastic Simulated Quantum Annealing (SSQA): Implements quantum Monte Carlo principles with stochastic computing, representing quantum fluctuations via large ensembles (Trotter replicas) of probabilistic bits (p-bits) on classical architectures. Parallel p-bit updates and Trotter-coupling emulate quantum tunneling, yielding order-of-magnitude reductions in time-to-solution for hard combinatorial tasks. SSQA has demonstrated capability for fully connected graphs two orders of magnitude larger than current hardware QA can simulate (Onizawa et al., 2023).
• Hybrid Annealing: Alternates quantum evolution (short time evolution under quantum fluctuations, followed by measurement) with classical Metropolis acceptance. Especially effective for random-energy models with quasi-degenerate minima where adiabatic QA is hampered by small gaps; hybrid cycles maintain short coherence and utilize repeated quantum mixing (Graß et al., 2016).

3. Path-Integral SQA and Mapping to Quantum Dynamics

The principal classical simulation method for QSA is path-integral Monte Carlo (PIMC) based on the Suzuki–Trotter decomposition. For a transverse-field Ising model

$O(1/\Delta)$4

the quantum partition function at inverse temperature $O(1/\Delta)$5 is mapped to a classical Ising model in $O(1/\Delta)$6 dimensions (the extra dimension corresponding to imaginary time/Trotter slices), with effective Hamiltonian

$O(1/\Delta)$7

where $O(1/\Delta)$8. Local and cluster updates sample the quantum Gibbs distribution, enabling simulation of quantum tunneling and barrier crossing (Tanaka et al., 2012, Crosson et al., 2014, Inack et al., 2015).

SQA in this sense has been proven, both empirically and by rigorous mixing time bounds, to inherit polynomial-time barrier tunneling for certain thin, high-spike cost landscapes that are classically exponential for SA but polynomial for QA and SQA (Crosson et al., 2016, Crosson et al., 2014).

4. Complexity, Spectral Gap, and Phase Transitions

Quantum speedup in QSA is intimately linked to the spectral properties of the underlying Markov (or quantum) process. QSA achieves quadratic scaling in the inverse of the minimum spectral gap $O(1/\Delta)$9 compared to SA:

• Classical SA: $S(\beta)$0 scaling,
• Quantum SA: $S(\beta)$1 scaling (Boixo et al., 2015, Harrow et al., 2019).

However, both QSA and QA can experience exponential slowdown when the gap closes exponentially, typically at first-order phase transitions (Nishimori et al., 2014). Advanced iterative techniques, such as the half-hot constraint for Potts models, can transform first-order transitions (exponentially small gap, exponential runtime) into second-order transitions (polynomial gap scaling, polynomial runtime) under both SA and QA (Okada et al., 2019).

QSA further supports adaptive schedule construction, utilizing quantum amplitude estimation and overlap testing to minimize the number of annealing steps required for convergence (improving from $S(\beta)$2 to $S(\beta)$3 schedule lengths) (Harrow et al., 2019).

5. Applications, Implementation, and Benchmarking

QSA methods are applicable to a variety of fundamental computational tasks:

• Optimization: Discrete combinatorial optimization (Ising, QUBO, Potts, MIS), box-constrained quadratic programming, constraint satisfaction (e.g., XORSAT) (Jeong et al., 16 Jun 2025, Djidjev, 2 Apr 2025, Ramezanpour, 2018).
• Sampling and Inference: Partition-function estimation, Bayesian posterior sampling, mixing of complex Gibbs states (Harrow et al., 2019).

Empirical studies show that QSA and QESA implementations on quantum hardware (e.g., Rydberg atom arrays) or effective classical surrogates (e.g., stochastic p-bit processors or projective QMC) can outperform standalone SA, with greater advantages as the dimensionality and ill-conditioning increase or when initialization with high-quality quantum states is possible (Jeong et al., 16 Jun 2025, Onizawa et al., 2023, Inack et al., 2015).

Key performance metrics used in benchmarking include:

• Approximation ratio $S(\beta)$4,
• Hamming distance to optimal or initial configurations,
• Time-to-solution (epochs, wall-clock),
• Residual energy and success probability.

High-quality warm starts (quantum-prepared or quantum-inspired) directly translate into reduced epochs-to-solution and permit substantially larger problem instances solvable within fixed time budgets (Jeong et al., 16 Jun 2025, Djidjev, 2 Apr 2025).

6. Rigorous Correspondences and Limitations

A profound feature of QSA is the established spectral mapping between classical SA (Markov generator $S(\beta)$5) and corresponding quantum Hamiltonians ($S(\beta)$6), ensuring identical eigenvalue gaps and thus identical complexity bottlenecks for quasi-static (strictly adiabatic/slow-annealing) regimes (Nishimori et al., 2014). This guarantees that, except for differences in locality (classical $S(\beta)$7 quantum mapping preserves local structure; quantum $S(\beta)$8 classical mapping does not), quantum annealers can simulate classical SA efficiently, but the converse is only tractable for specifically structured quantum problems.

Nonetheless, SQA emulations (PIMC) only reproduce certain aspects of quantum dynamics. Numerical studies show that SQA captures average observable scaling (e.g., defect density in Kibble–Zurek mechanisms) but can fail to reproduce full quantum distribution functions, especially for closed systems, and only partially agrees with quantum dynamics in open/dissipative cases (Bando et al., 2021). Thus, caution is necessary when equating SQA results with those of genuine quantum annealers for detailed dynamical phenomena.

7. Future Directions and Open Problems

QSA research continues to address several open challenges and directions:

• Relaxing gap and sparsity assumptions: Current guarantees rely on known lower bounds of spectral gaps and sparsity, with active work on gap amplification and robust fidelity to Gibbs states (without explicit gap or sparsity knowledge) (Boixo et al., 2015).
• Hybrid algorithms and co-located hardware: Reducing classical-quantum communication overhead, exploiting in-loop quantum resources, and developing adaptive, dynamically scheduled hybrid protocols are key for scaling QSA to larger and broader classes of optimization problems (Djidjev, 2 Apr 2025).
• Benchmarking and scaling: Extensive empirical and theoretical work is ongoing for understanding performance in presence of frustration, glassiness, highly non-stoquastic terms, and system-environment coupling, as well as practical embedding and decoding on near-term noisy quantum hardware (Albash et al., 2016, Graß et al., 2016).
• Algorithmic extensions: Integration of reinforcement feedback, higher-order quantum-inspired moves, and more general non-equilibrium mappings (beyond detailed balance and time-local generators) to improve convergence, bypass bottlenecks, or further bridge the gap to quantum supremacy in optimization (Ramezanpour, 2018, Okada et al., 2019).

Quantum simulated annealing thus provides a rigorous, scalable, and physically motivated framework for quantum acceleration of classical and hybrid optimization algorithms, with established quadratic speedups in spectral gap and annealing step counts, direct mappings to quantum walks, and a rich interface to both hardware and classical computing paradigms. The field continues to develop towards practical large-scale deployment and deeper understanding of quantum-classical correspondences in complex optimization and sampling.