SBO-QAOA: Enhanced QAOA via Bayesian & Subspace Methods

• SBO-QAOA is a hybrid quantum-classical framework that integrates Bayesian methods, subspace reduction, and Boltzmann encoding for efficient QAOA parameter tuning and fair sampling.
• It employs Gaussian process surrogates with adaptive region techniques to dramatically reduce circuit calls and mitigate hardware noise in NISQ devices.
• By leveraging backbone decomposition and qubit embedding, SBO-QAOA addresses scalability and constraint handling challenges in large combinatorial optimization problems.

The acronym SBO-QAOA encompasses several distinct but related quantum optimization frameworks in the literature, leveraging subspace reduction, Bayesian optimization, and Boltzmann sampling strategies for enhanced performance, scalability, and sampling fidelity in the Quantum Approximate Optimization Algorithm (QAOA). SBO-QAOA approaches address the core challenges of variational quantum optimization—optimizer inefficiency, hardware noise, limited qubit count, constraint handling, and biased solution sampling—by integrating classical statistical learning, structural decomposition, temperature-dependent cost functions, and subspace embeddings. The following sections provide a rigorous exposition of the primary SBO-QAOA methodologies and their theoretical underpinnings, implementation details, practical trade-offs, and benchmarked results.

1. Sequential Bayesian Optimization for QAOA Parameter Tuning

Sequential Bayesian Optimization QAOA (SBO-QAOA) leverages Gaussian process (GP) surrogates and acquisition functions to efficiently optimize the $2p$ variational angles $$\theta=(\gamma_1,\ldots,\gamma_p,\beta_1,\ldots,\beta_p)$$ of QAOA circuits, minimizing expensive quantum cost function evaluations. The GP prior (typically zero-mean, with Matérn $\nu=3/2$ kernel) models the unknown energy landscape $E(\theta)=\langle\theta|H_C|\theta\rangle$; the kernel $k(\theta,\theta')$ encapsulates both smoothness and local roughness. After collecting $n$ evaluations $\{(\theta_i, y_i)\}_{i=1}^n$, the GP posterior provides mean $\mu_n(\theta)$ and variance $s_n^2(\theta)$ updates via conditioning, with noise terms $\sigma_N^2$ reflecting measurement shot statistics ($\sigma_N^2 \propto N_S^{-1.1}$).

The acquisition function, typically Expected Improvement (EI), balances exploration ($s_n$) and exploitation ($\mu_n$). EI is maximized by population-based metaheuristics (Differential Evolution, population size $\sim15 \cdot 2p$), with convergence governed by the drop in EI standard deviation and population diameter. Posterior prediction and hyperparameter refitting ($\{\sigma^2, \ell, \sigma_N^2\}$ by L-BFGS maximum marginal likelihood) drive the iterative loop, making SBO-QAOA highly sample-efficient and robust against gate-level and measurement noise. With $N_{\text{BAYES}}\sim300$–$600$ circuit calls, SBO-QAOA achieves approximation ratios $R\approx95\%$ for 10-qubit Max-Cut at $p=7$—cutting quantum circuit calls by a factor $3$–$20$ over gradient-free baselines and maintaining resilience up to gate errors $\sigma_{QN}\sim0.01$ and shallow depths $p\le7$ (Tibaldi et al., 2022).

2. Bayesian Adaptive Region Optimization: The DARBO Framework

DARBO (Double Adaptive-region Bayesian Optimization) refines QAOA parameter search via two nested, adaptively sized hyperrectangles: the trust region (TR), centered at the incumbent best parameter vector $x^*$, and the search region (SR), which toggles between the full domain and a narrowed neighborhood. The GP surrogate (Matérn–5/2 kernel) is locally trained on TR samples for acute landscape estimation; the acquisition (EI or UCB) is maximized only within $TR_i \cap SR_i$. TR side length $L_i$ is doubled upon consecutive sampling "successes" ($y<y^*$) and halved after failures, dynamically scaling search tightness, while SR alternates based on acquisition trap frequency. This double-region scheme prevents over-contraction and premature local minima trapping.

DARBO's core loop compiles shot-noisy QAOA circuits, updates GP surrogates, and adapts regions in response to optimization trajectory. Experimental deployments on superconducting hardware employed layout benchmarking, readout error mitigation, and zero-noise extrapolation (ZNE), confirming the method's robustness—yielding a $1$–$2\times$ improvement in final gap and using $\sim50\%$ fewer circuit calls than Adam, COBYLA, and SPSA (Cheng et al., 2023). When combined with full quantum error mitigation (QEM), Bayesian strategies in QAOA deliver sample success probabilities $5$–$10\times$ above random and outperform conventional optimizers under noise.

3. Boltzmann-encoded Hamiltonians for Fair Sampling

One SBO-QAOA variant targets the problem of biased sampling among degenerate ground states in combinatorial optimization, employing a temperature-dependent Hamiltonian $H_S(T)$ whose ground state encodes the Gibbs distribution: $$|\psi_{\text{Gibbs}}(T)\rangle = \frac{1}{\sqrt{Z(T)}}\sum_\sigma e^{-H_0(\sigma)/(2T)}|\sigma\rangle$$ Here, the cost Hamiltonian $H_C$ is replaced by

$$H_S(T) = -e^{-\alpha/T} \sum_{i=1}^N [\sigma^x_i - e^{H_i/T}]$$

where $H_i$ is the local energy, and $\alpha$ normalizes against $T\to0$ divergence. The QAOA circuit retains the standard transverse-field mixer, and variational parameters $\{\gamma_k,\beta_k\}$ (in full or linearized four-parameter schedule) are optimized to minimize total variation distance to the Gibbs target.

Numerical evidence shows SBO-QAOA achieves uniform sampling among degenerate ground states and target finite temperature distributions, with total ground-state probability saturating at the correct Gibbs value and within-manifold probabilities equilibrating ($P_1 = P_2 = P_3$ in $N=5$ spin toy models). Full convergence is attained in both the full $2p$ and linearized four-parameter domains, with $D_{\text{TVD}}(p)\to0$ as depth increases. Practical circuit decompositions of multi-body $e^{H_i/T}$ remain a scalability challenge, and further advances in decompositions are required for large $N$ implementations (Abe et al., 22 Jan 2026).

4. Subspace and Backbone-driven Problem Decomposition

Backbone-driven SBO-QAOA employs adaptive tabu search to identify high-stability "backbone" variables in large QUBO cost functions, decomposing the global problem into NISQ-compatible windows for quantum subproblem optimization. The algorithm:

• Preprocesses via tabu search, extracting top-$k$ backbone bits by local flip-cost magnitude;
• Constructs quantum-tractable subspaces (windows $W$ of size $n_w$), fixing non-window bits and forming reduced cost Hamiltonians;
• Applies shallow QAOA ($p\sim1$) to each window's induced Ising Hamiltonian for variational optimization;
• Iteratively refines solutions by sliding the window, updating couplings, and accepting global updates.

Experimental benchmarks on G-set and Karloff graphs ($N=800$–$3500$) report approximation ratios $[0.987,1.0]$, matching or surpassing classical baselines (Goemans–Williamson $\sim0.92$), with high success probabilities and explicit scalability via $k\sim0.2N$, $n_w=15$ (Gou et al., 13 Apr 2025). The framework efficiently orchestrates classical preprocessing with quantum resources, allowing hierarchical search regimes exceeding raw device qubit count.

5. Quantum-constrained and Stochastic Optimization

SBO-QAOA is adapted for stochastic constrained binary optimization by employing dual decomposition: the original expectation-constrained QCQP is sampled over scenarios, forming empirical Hamiltonians, and then solved as a sequence of penalized QUBOs. Lagrange multipliers $\lambda$ are updated via subgradient ascent on constraint violation, with the QAOA (or VQE) ansatz addressing the unconstrained penalized binary optimization at each step. Optimization alternates between primal (sampling via QAOA, angle optimization, shot-based measurement) and dual (multiplier update) steps.

Numerical results on small ($n=2,5$) QCQP instances demonstrate near-optimal solution recovery and constraint satisfaction—with dual variables stabilizing in $10$–$20$ iterations and sampler convergence within $100$–$200$ optimizer steps. QAOA-based samplers matched LP-optimal distributions and maintained feasibility within $5\%$ tolerance in constrained cases. The complexity is bounded by QAOA circuit depth ($p\sim2$), total gate count ($\sim200$ for $n=5$), and scenario-shot sampling rates (Gupta et al., 2023).

6. Subspace Embedding and Qubit Efficiency

Recent SBO-QAOA implementations address hardware limitations by embedding the $n$-bit optimization problem into $m\ll n$ qubits. The approach partitions variables into blocks (size $d$, count $G=n/d$), with label and data qubits (total $m=d+\log_2 G$), and uses an embedding operator to map bitstrings into entangled wavefunctions. The variational ansatz alternates mixer and wavefunction-dependent cost Hamiltonians, estimating cost via block-wise postselection and empirical mean.

Parameter concentration—instance independence of optimal angles—has been empirically and theoretically confirmed for Sherrington–Kirkpatrick spin glasses. At depth $p=3$, $n=64$, SBO-QAOA ($d=1$, $m=7$) achieves $r\approx0.69$ (matching standard QAOA at $p=1$), with parameters transferable across instances and extensible via rescaling. Large-$n$ asymptotic guarantees follow: $r_p^\infty$ at fixed depth matches QAOA performance, converging to the Parisi constant at $p\to\infty$ (Sundar et al., 2024).

Variant	Key Principle	Typical NISQ Features
Bayesian optimizer	GP surrogate + acquisition function	Low shot/iteration, robust to noise
Backbone/subspace	Tabu backbone + windowed QAOA	NISQ-scale optimization via subproblem
Boltzmann encoding	Gibbs distribution ground state	Fair sampling among degenerate solutions
Dual decomposition	Penalty-based constraint handling	Stochastic QCQPs, expectation feasibility
Subspace embedding	Blockwise labeling, few-qubit ansatz	Qubit-efficient, parameter concentration

7. Implementation Guidelines and Limitations

Implementation recommendations include warm-up sampling ($N_W\approx10$–$15$ for $2p\le20$), Matérn kernel selection, hyperparameter optimization via multiple L-BFGS restarts, shot-count tuning ($N_S\ge128$ for GP accuracy), and circuit depth constraints ($p\le7$ for NISQ devices with error mitigation). For backbone methods, parameter selection trades classical effort against quantum depth/window size. Fair sampling SBO-QAOA shifts complexity to cost Hamiltonian engineering, with many-body terms requiring scalable decomposition. Algorithmic success in constraint satisfaction and hardware realization is presently confined to small-to-medium problem sizes ($n\le20$), with larger $N$ awaiting advances in quantum device capacity and circuit compilation strategies.

In sum, SBO-QAOA encompasses a suite of quantum–classical hybrid methodologies that substantially enhance QAOA’s sample efficiency, scalability, constraint handling, and sampling fidelity, leveraging Bayesian statistics, subspace reduction, backbone decomposition, and temperature-targeted Hamiltonian engineering. These approaches exhibit strong empirical and theoretical performance, precise resource allocation, and adaptability to contemporary quantum hardware regimes (Tibaldi et al., 2022, Cheng et al., 2023, Abe et al., 22 Jan 2026, Gou et al., 13 Apr 2025, Gupta et al., 2023, Sundar et al., 2024).