Non-Stoquastic Adiabatic Quantum Optimization

• NS-AQO is a quantum optimization method that introduces non-stoquastic catalysts or counterdiabatic terms to modify spectral gaps and enable advanced tunneling paths.
• It leverages Hamiltonians with positive or complex off-diagonal elements that disrupt classical simulation and provide potential quantum speedups for combinatorial challenges.
• Effective implementation requires precise calibration of catalyst strengths and tailored annealing schedules to convert first-order transitions into avoided crossings with improved gap scaling.

Non-Stoquastic Adiabatic Quantum Optimization (NS-AQO) algorithms are quantum optimization protocols that employ non-stoquastic Hamiltonians during the adiabatic interpolation, in contrast to standard stoquastic adiabatic quantum optimization. While stoquastic Hamiltonians possess only non-positive off-diagonal matrix elements in the computational basis and admit efficient classical simulation via Quantum Monte Carlo, non-stoquastic Hamiltonians contain strictly positive (or complex) off-diagonal entries, introducing sign structure that inhibits classical simulability and is conjectured to enable quantum computational advantages. The essential aim of NS-AQO is to exploit these quantum interference effects, specifically through non-stoquastic catalyst terms and counterdiabatic driving, resulting in altered spectral gaps and, under certain circumstances, improved scaling of runtime and success probabilities for hard combinatorial optimization problems.

1. Stoquastic versus Non-Stoquastic Hamiltonians

Stoquastic Hamiltonians are real symmetric matrices in the computational basis ($|x\rangle$) such that all off-diagonal elements satisfy $\langle x|H|y\rangle \leq 0$ for every $x\neq y$. This property ensures there is no sign problem in path-integral quantum Monte Carlo and allows for efficient classical simulation (Crosson et al., 2020). In contrast, a Hamiltonian is non-stoquastic if, in any local basis or at least in the computational basis, it possesses strictly positive or complex off-diagonal entries. This non-stoquasticity impedes classical techniques due to destructive interference, theoretically enabling universal quantum computation and potentially circumventing classical bottlenecks (Choi, 2021).

Common sources of non-stoquasticity are the addition of XX-coupling terms ($\sum_{i<j} \sigma^x_i \sigma^x_j$ with positive coefficients) or Pauli-$y$ operator products as seen in counterdiabatic driving (Hegade et al., 2022, Vinci et al., 2017). In the context of adiabatic optimization, the anticipation was that non-stoquastic drivers would open novel tunneling paths, lifting bottleneck gaps and improving computational scaling.

2. Adiabatic Interpolation, Non-Stoquastic Catalysts, and Counterdiabatic Driving

The standard adiabatic algorithm interpolates between an easily prepared initial driver Hamiltonian $H_\mathrm{init}$ (e.g., transverse field) and a problem Hamiltonian $H_\mathrm{prob}$ diagonal in the computational basis:

$$H(s) = (1-s) H_\mathrm{init} + s H_\mathrm{prob}, \quad s \in [0,1].$$

Non-stoquasticity is introduced via "catalyst" terms or counterdiabatic (CD) corrections. A non-stoquastic catalyst is typically added via an interpolation:

$$H(s) = (1-s) H_\mathrm{init} + s H_\mathrm{prob} + \lambda(s) H_\mathrm{catalyst},$$

where $H_\mathrm{catalyst}$ contains terms (e.g., $\sigma^x_i \sigma^x_j$ with $\lambda(s)>0$) yielding positive off-diagonal elements (Albash, 2018, Choi, 2021). In counterdiabatic optimization (CDQO), additional terms derived from the adiabatic gauge potential (Berry connection) are added to suppress undesirable non-adiabatic transitions. These CD terms inherently contain Pauli-$y$ operators and are highly non-stoquastic (Hegade et al., 2022).

Gate-based, digitized variants of NS-AQO (e.g., DCQO) employ the Suzuki–Trotter decomposition and allow explicit programming of arbitrary non-stoquastic interactions (Hegade et al., 2022).

3. Spectral Gap Structure and Runtime Scaling

The computational utility of adiabatic optimization is dictated by the minimum spectral gap $\Delta_\mathrm{min}$ between the ground and first excited states during the interpolation. The runtime required to reliably project onto the ground state generally scales as $T \sim \Delta_\mathrm{min}^{-2}$:

• Stoquastic Hamiltonians without catalysts frequently exhibit sharp first-order phase transitions with exponentially small gaps, $T \sim \exp({O(N)})$ (Albash, 2018).
• Non-stoquastic catalysts can modify the mean-field landscape, softening first-order transitions into sequences of avoided crossings or merging potential wells, resulting in polynomial or constant gaps for designed models (e.g., infinite-range $p$-spin, 2-local large-spin tunneling) (Albash, 2018).
• Non-stoquastic XX-driver graph constructions can bridge single anti-crossings into double-AC structures, enabling diabatic transitions in polynomial time (Choi, 2021).
• Counterdiabatic driving generally increases the minimum gap and yields polynomial enhancement in ground-state success probability for generic spin glass instances (Hegade et al., 2022).

Analytical and extensive numerical evidence, however, demonstrates that for generic random and local-Max-Cut problems, non-stoquasticity alone generically produces smaller spectral gaps, making such adiabatic paths less effective than their stoquastic counterparts (Crosson et al., 2020).

4. Key Algorithmic Frameworks and Practical Applications

Catalyst-Enhanced Adiabatic Optimization

Catalyst protocols add non-stoquastic terms during the interpolation to flatten barriers or split crossings:

• $p$-spin model: Non-stoquastic catalyst $(\sum_i \sigma^x_i)^2$ enables exponential advantage over stoquastic interpolation—transition becomes second-order and the gap scales polynomially (Albash, 2018).
• 2-local large-spin model: XX-term between composite spins $S^x_1 S^x_2$ eliminates barriers, leading to constant gap and $O(1)$ runtime scaling.

Driver-Graph Design

Identifies local minima (forming independent-clique structures) and inserts XX-couplers to bridge anti-crossings, tuning the coupling strength to induce proper non-stoquasticity and maximize the gap (Choi, 2021). This approach achieves exponential speedup in AC-distance over classical and stoquastic algorithms for specific MWIS-type problems.

Counterdiabatic Quantum Optimization

DCQO explicitly programs k-local CD terms, with 2-local approximations yielding size-dependent polynomial enhancements in success probability. Trotterized gate-model decomposition facilitates implementation on contemporary NISQ platforms (Hegade et al., 2022).

Flux-Qubit Realization and Geometric Phases

Non-stoquastic interactions emerge naturally in flux-qubit hardware via non-adiabatic geometric (Aharonov–Anandan) phases, appearing as σ$^y$, σ$^x\sigma^y$, and σ$^z\sigma^y$ terms—even in standard adiabatic paths. These geometric contributions can soften bottleneck transitions and require no additional coupler hardware (Vinci et al., 2017).

5. Analytical and Numerical Evidence

Comprehensive surveys through random-matrix models, Max-Cut instances, and symmetry-constrained problems establish:

Algorithm/Model	Non-Stoquastic Gap Scaling	Stoquastic Gap Scaling	Observed Speedup
Infinite-range $p$-spin	Polynomial ($N^{-\gamma}$) or constant	Exponential ($\exp(-\alpha N)$)	Exponential (for $p \geq 4$) (Albash, 2018)
2-local large-spin	Constant	Exponential	Exponential
Local Ising ring-of-rings	Weaker exponential	Exponential	Sub-exponential
Random dense matrices, Max-Cut	Smaller gaps (generically)	Larger gaps (beneficial)	None (generic settings) (Crosson et al., 2020)
DCQO on Ising spin glass	Gap increased, $P_s \sim N^{0.5}$	Gap decreases, $P_s$ exponentially small	Polynomial (Hegade et al., 2022)
XX-driver graph (double-AC)	Large 2–1 gap, polynomial runtime	Single small gap, exponential runtime	Exponential in AC-distance (Choi, 2021)

Numerical experiments consistently show that sophisticated design (e.g., driver-graph XX terms, non-stoquastic catalysts, CD driving) is required; non-stoquasticity alone does not generically confer advantage over stoquastic adiabatic algorithms (Crosson et al., 2020).

6. Implementation Constraints and Algorithm Design Principles

Physical implementation of non-stoquastic interactions requires circuits supporting sign-tunable XX couplers and/or programmable Pauli-$y$ terms. Many commercial devices (e.g., D-Wave) currently allow only stoquastic couplings; gate-based and circuit-model architectures offer greater flexibility (Hegade et al., 2022). In flux-qubit annealers, the presence of geometric phases ensures that non-stoquastic effects appear automatically during finite-speed evolution, bypassing hardware limitations (Vinci et al., 2017).

Effective design principles include:

1. Identify dominant classical barriers in the optimization landscape.
2. Insert catalyst terms that flatten barriers or merge wells (e.g., $\sigma^x \sigma^x$ coupling).
3. Tune catalyst/coupler strengths to optimize the minimum gap.
4. Employ locally adapted annealing schedules to allocate computational time where gaps are smallest.
5. Validate robustness against calibration errors and noise.

7. Implications, Limitations, and Outlook

NS-AQO, through either non-stoquastic catalysts, counterdiabatic terms, or driver-graph constructions, can transform gap scaling and runtimes for specially constructed or hard instances by converting first-order to second-order transitions, bridging anti-crossings, or amplifying ground-state success probabilities. However, extensive analytical and empirical results reveal that in generic random or locally structured problems, non-stoquasticity does not yield systematic advantage—in such cases, de-signing transformations can convert non-stoquastic adiabatic paths to stoquastic ones with equal or larger spectral gaps and no loss in performance (Crosson et al., 2020).

A plausible implication is that genuine quantum speedups via NS-AQO hinge critically on careful catalyst and schedule design, precise control of hardware couplers, and tailored problem instance selection. While worst-case hardness persists, paradigms such as DCQO and double-AC driver-graphs deliver polynomial or exponential speedup in practice for specifically engineered classes, and the gate-based evolution makes near-term implementation feasible on NISQ devices for moderate system sizes (Hegade et al., 2022, Choi, 2021).

Overall, algorithm designers must rigorously justify the introduction of non-stoquastic structure rather than assume generic quantum advantage; leveraging quantum interference remains a nuanced and problem-dependent strategy in adiabatic optimization.