Simulated Quantum Annealing

• Simulated Quantum Annealing is a classical algorithm that emulates quantum annealing by mapping quantum Hamiltonians to extended classical spin configurations using Suzuki–Trotter decomposition.
• It leverages path-integral Monte Carlo methods to simulate quantum fluctuations and tunneling in high-dimensional and rugged optimization problems.
• SQA serves as a robust benchmarking tool for quantum annealers and a practical approach for tackling NP-hard optimization challenges across diverse hardware architectures.

Simulated Quantum Annealing (SQA) is a classical stochastic algorithm that emulates aspects of quantum annealing (QA) by leveraging quantum Monte Carlo (QMC) methods—especially path-integral Monte Carlo—on classical hardware. SQA is designed to explore the computational power of quantum tunneling and quantum fluctuation in complex optimization and sampling problems, particularly those involving rugged or high-dimensional energy landscapes. SQA is widely used both as a quantum-inspired classical algorithm and as a simulation framework to benchmark or probe physical quantum annealers.

1. Theoretical Foundations and Path-Integral Mapping

At the core of SQA is the path-integral representation of a quantum transverse-field Ising model, which maps the quantum Hamiltonian to a classical statistical model with an extended configuration space. For a general transverse-field Ising Hamiltonian with $N$ spins,

$$H(\Gamma) = -\sum_{\langle i,j \rangle} J_{ij} \sigma^z_i \sigma^z_j - \sum_i h_i \sigma^z_i - \Gamma \sum_i \sigma^x_i$$

the quantum partition function at inverse temperature $\beta$,

$Z = \operatorname{Tr} e^{-\beta H(\Gamma)}$

is approximated via Suzuki–Trotter decomposition into $P$ imaginary-time slices as

$$Z \approx \sum_{\{s_{i,k} = \pm 1\}} \exp\left(-K_{\mathrm{cl}}[\{s\}]\right)$$

where $\{s_{i,k}\}$ denotes Ising spins at space index $i$ and (imaginary) time index $k$, and

$$K_{\mathrm{cl}}[\{s\}] = -\sum_{k=1}^P \left[ \sum_{\langle i,j \rangle} \frac{\beta J_{ij}}{P} s_{i,k} s_{j,k} + \sum_i \frac{\beta h_i}{P} s_{i,k} \right] - \sum_{i=1}^N \sum_{k=1}^P J^\perp(\Gamma, \beta_P) s_{i,k} s_{i,k+1}$$

with $\beta_P = \beta / P$ and the inter-slice quantum coupling

$$J^\perp(\Gamma, \beta_P) = -\frac{1}{2} \ln\left[\tanh(\beta_P \Gamma)\right]$$

and periodic boundary conditions $s_{i,P+1} \equiv s_{i,1}$ (Mbeng et al., 2018, Bando et al., 2021).

In the limit $P \rightarrow \infty$, this mapping becomes exact. All quantum observables diagonal in the computational basis can be computed as averages over this classical extended ensemble.

The driving “quantum” parameter $\Gamma(t)$ is slowly decreased during the simulation (“annealing schedule”) to traverse from a quantum-fluctuating regime to a classical regime.

2. Algorithmic Implementation and Update Schemes

The SQA workflow proceeds as follows:

1. Initial Mapping: Map the quantum optimization or sampling problem via Suzuki–Trotter decomposition to a classical $(N \times P)$ spin system with both intra-slice (classical) and inter-slice (quantum) couplings.
2. Monte Carlo Updates: Perform Markov chain Monte Carlo (MCMC) sweeps on the extended system at each anneal step. Physical update choices include:
   • Single-spin Metropolis updates.
   • Time-only cluster updates (e.g. Swendsen-Wang clusters in the Trotter direction).
   • Space–time cluster updates (less “physical”—can lead to unphysical dynamics and sampling pathologies).
3. Annealing Schedule: The transverse field $\Gamma(t)$ is ramped, typically in a linear or geometric schedule. At each step, configurations are updated to sample the instantaneous equilibrium of the classical effective Hamiltonian (Mbeng et al., 2018).
4. Measurement: At the end of the anneal, quantities of interest (such as the residual energy or spin configuration) are computed, typically by averaging over Trotter slices.

The detailed implementation can vary:

• Path-integral QMC (“conventional” SQA) is standard for discrete Ising spins (Bando et al., 2021).
• For continuous variables (e.g. in quantum optimization with continuous coordinates), projective QMC methods (e.g. diffusion Monte Carlo) can be used for SQA, offering different scaling for residual energy and better handling of high barriers in some settings (Inack et al., 2015).
• Stochastic Simulated Quantum Annealing (SSQA) leverages parallel stochastic bit (“p-bit”) hardware for parallel updates and enhanced scalability (Onizawa et al., 2023).
• Variants incorporating multi-spin quantum fluctuations employ an extended driver Hamiltonian with two-spin or higher transverse couplings, which can accelerate tunneling and cluster update efficiency (Mazzola et al., 2017).

3. Performance Regimes, Scaling Laws, and Benchmark Results

Ordered One-Dimensional Systems

In ordered one-dimensional transverse-field Ising chains, SQA with physical update choices (e.g., time-only clusters) recovers the Kibble–Zurek scaling for residual energy,

$\varepsilon_{\mathrm{res}}(\tau) \sim \tau^{-1/2}$

with anneal time $\tau$, matching the result for coherent quantum annealing in the thermodynamic limit (Mbeng et al., 2018, Bando et al., 2021).

Disordered and Glassy Systems

Disorder introduces substantial changes:

• For random coupling instances, standard SQA with large $P$ often suffers a sampling crisis, where the local dynamics in the Trotter direction become non-ergodic, leading to non-convergent or pathologically biased energy estimates as $P \to \infty$. Moderate $P$ values (e.g., $P \approx 32 - 64$) give better convergence but incur unavoidable Trotter discretization errors (Mbeng et al., 2018).
• In two-dimensional Ising spin glasses, the apparent speedup of SQA over simulated annealing (SA) occurs only for discrete-time SQA at small $M$ (time discretization); this advantage disappears in the continuous-time ($M \to \infty$) limit and does not reflect true quantum annealing performance (Heim et al., 2014).

Hard Optimization Instances

For problems featuring thin or high energy barriers (e.g., the spike Hamiltonian),

• Classical SA is exponentially slow due to rare barrier crossing: $\tau_{\mathrm{SA}} \sim \exp[\Theta(\beta n)]$.
• SQA, via imaginary-time quantum fluctuations, can mix in polynomial time, $\tau_{\mathrm{SQA}} = O(n^4)$ for the basic spike instance when $P = O(n)$.
• Rigorous analysis shows that SQA—provided the quantum gap is polynomially bounded—matches QA’s scaling, precluding exponential speedups by QA over SQA on these instances (Crosson et al., 2014, Crosson et al., 2016, Bergamaschi, 2020).

Continuous-space and Frustrated Systems

• In continuous-variable models (e.g., double-well, multi-well, quasi-disordered), projective QMC-based SQA displays robust $$\varepsilon_{\mathrm{res}}(\tau_f) \sim \tau_f^{-1/3}$$ scaling for the residual energy, independent of potential symmetry or disorder, and outperforms both classical annealing (which can saturate at logarithmic scaling in complex landscapes) and finite-T path-integral SQA (which saturates at a finite temperature plateau) (Inack et al., 2015).

Table: Representative Residual Energy Scaling Exponents

System Type	Method	Scaling Exponent $b$	Source
1D TFIM (ordered)	SQA (closed)	$0.5$	(Bando et al., 2021)
1D TFIM (open)	SQA with coupling $\alpha$	$0.26$–$0.5$	(Bando et al., 2021)
Spike Ising	SQA	poly$(n)$	(Bergamaschi, 2020)
Double-well cont.	Projective SQA (DMC)	$1/3$	(Inack et al., 2015)

4. Applications, Hardware Agnosticism, and Architectural Realizations

SQA is extensively applied in the exploration and benchmarking of combinatorial optimization problems:

• NP-hard optimization (e.g., MAX-SAT, planted Ising, TSP) (Raimondo et al., 17 Mar 2025).
• Sampling in quantum models (e.g., lattice gauge theories with topological constraints) (Yan et al., 2021).
• Large-scale fully-connected problems (graph isomorphism) using parallel stochastic computing (SSQA) (Onizawa et al., 2023).

From a hardware viewpoint, SQA is highly adaptable:

• SQA mapping is entirely classical, and its principal elements—random number generation, nonlinearity, and networked p-bits—can be implemented in pure CMOS, hybrid CMOS–spintronic, photonic, or other analog–digital platforms (Raimondo et al., 17 Mar 2025).
• Probabilistic Ising machines (PIMs) using SQA demonstrate superior reliability, robustness to device nonuniformity, and higher success probabilities compared to SA and parallel tempering under practical hardware constraints. In CMOS, per-spin update times of 8 ns and energy costs $\lesssim 0.22$ mW have been demonstrated for 500-pbit systems (Raimondo et al., 17 Mar 2025).
• SSQA leverages bit-parallel operations to update thousands of p-bits simultaneously, achieving dramatic wall-clock speedups over serial SA and outperforming hardware-limited quantum annealers (e.g., D-Wave) by large margins in fully-connected benchmarks (Onizawa et al., 2023).

5. Limitations, Convergence Criteria, and Open System Dynamics

Despite its usefulness, SQA faces several key limitations:

Non-equivalence to Physical Quantum Dynamics

• SQA samples from an equilibrium (or near-adiabatic) quantum Boltzmann distribution in imaginary time, not from the real-time Schrödinger evolution as in coherent QA. No formal theorem connects SQA’s stochastic dynamics with physical QA, especially in nonequilibrium regimes (Mbeng et al., 2018, Heim et al., 2014, Bando et al., 2021).
• The choice of Monte Carlo update moves crucially affects dynamical behavior and scaling. Only “time-only” cluster updates can preserve Kibble–Zurek scaling, while space–time clusters can artificially enhance equilibration, rendering the simulation physically irrelevant (Mbeng et al., 2018).

Convergence Conditions and Scaling

• Systematic, size-dependent convergence conditions for SQA schedules have been derived by mapping the SQA master equation to an imaginary-time Schrödinger equation. A generic “adiabatic” schedule for the transverse field is

$\Gamma(t) \sim t^{-1/(2N)}$

which becomes prohibitively slow for large $N$ (Kimura et al., 2022). This schedule coincides, up to constants, with rigorously derived bounds for both closed and open SQA, and parallels those for real-time unitary QA (Kimura et al., 2022).

Sampling Crisis and Pathologies

• In disordered or frustrated systems, SQA with large $P$ can exhibit a “sampling crisis,” where local-in-Trotter updates become trapped in atypical worldline configurations, causing non-convergence or strong dependence on initialization, even in extremely long simulations (Mbeng et al., 2018).
• SQA’s residual energy scaling diverges from that of coherent QA in the presence of disorder or open-system couplings. Quantitative predictions of QA performance based on SQA are unreliable in these regimes (Bando et al., 2021).

Robustness and Validity by Observable

• For averaged quantities such as the residual energy density in simple ordered systems, SQA replicates analytical scaling laws. However, for higher-order statistics (e.g., defect number cumulants), observable distributions, and more subtle dynamical properties, SQA can exhibit discrepancies—even in well-controlled “classical” limits (Bando et al., 2021).
• No universal a priori criterion exists to determine for which observables or in which systems SQA reliably mimics quantum dynamics. Careful case-by-case validation is required (Bando et al., 2021).

6. Algorithmic Innovations and Hybrid Variants

Several advances have extended the SQA framework:

• Multi-Spin Quantum Drivers: Incorporation of two-spin (or higher) quantum fluctuation operators in the driver Hamiltonian enhances tunneling and improves scaling of residual energy, especially in disordered or glassy landscapes. Cluster loop updates and bond-decomposition methods support these schemes without introducing sign problems (Mazzola et al., 2017).
• Fast-SQA via Best-Slice Dynamics: Heuristics that dynamically select the lowest-energy Trotter slice as a reference for all imaginary-time couplings enable highly parallel updates and significant empirical speedup on large graphs, though they are not rigorously Markovian with respect to the full equilibrium ensemble (Murashima, 2023).
• Hardware-Agnostic and Stochastic SQA: SSQA architectures update massive spin arrays in parallel using stochastic computing, bypassing serial bottlenecks and enabling efficient scaling to problems far beyond current quantum annealer hardware connectivity (Onizawa et al., 2023).

7. Summary and Outlook

Simulated Quantum Annealing provides a powerful and flexible framework for probing the advantages of quantum-inspired optimization and sampling on classical resources. Its rigorous mapping to stoquastic Hamiltonians, broad hardware compatibility, and robust convergence analysis underpin its status as both a benchmark for physical quantum annealers and a leading classical metaheuristic. Nevertheless, SQA’s predictive power for true quantum dynamics is system- and observable-dependent, requiring careful control of discretization errors, update schemes, and schedule parameters. Ongoing developments in multi-spin drivers, hybrid fast-update strategies, and hardware/software co-design continue to expand the applicability and performance of SQA in challenging optimization and quantum simulation domains (Mbeng et al., 2018, Bando et al., 2021, Raimondo et al., 17 Mar 2025, Inack et al., 2015, Mazzola et al., 2017, Kimura et al., 2022, Onizawa et al., 2023).