Classical and Quantum Annealing in the Median of Three Satisfiability

We determine the classical and quantum complexities of a specific ensemble of three-satisfiability problems with a unique satisfying assignment for up to N=100 and N=80 variables, respectively. In the classical limit we employ generalized ensemble techniques and measure the time that a Markovian Monte Carlo process spends in searching classical ground states. In the quantum limit we determine the maximum finite correlation length along a quantum adiabatic trajectory determined by the linear sweep of the adiabatic control parameter in the Hamiltonian composed of the problem Hamiltonian and the constant transverse field Hamiltonian. In the median of our ensemble both complexities diverge exponentially with the number of variables. Hence, standard, conventional adiabatic quantum computation fails to reduce the computational complexity to polynomial. Moreover, the growth-rate constant in the quantum limit is 3.8 times as large as the one in the classical limit, making classical fluctuations more beneficial than quantum fluctuations in ground-state searches.

💡 Research Summary

The paper investigates the computational complexities of a specially constructed ensemble of 3‑SAT instances that possess a unique satisfying assignment (USA). The authors study both classical simulated annealing and quantum adiabatic quantum computation (AQC) on these instances, focusing on the median behavior rather than worst‑case scenarios.

In the classical regime, they employ multicanonical (MUCA) and Wang‑Landau Monte Carlo techniques to obtain an accurate estimate of the density of states n(E) over the full energy range. From this they compute the canonical partition function, internal energy, and specific heat, identifying a sharp peak in the specific heat that signals a thermodynamic phase transition. Using Binder’s method they define a nucleation barrier B₀ = ln(P_max,left / P_min), where P_max,left and P_min are the heights of the two peaks and the intervening minimum in the distribution of the ground‑state overlap observable. They also measure the average Monte‑Carlo search time τ₀, i.e., the number of MC steps required to reach the ground state from high energy, normalized by N² to remove the trivial random‑walk contribution. Both τ₀ and B₀ grow exponentially with the number of variables N, following A·exp(g N) with growth‑rate constants g_τ≈0.016 and g_B≈0.21 for clause‑to‑variable ratio r=8, and larger values for r=5. The exponential scaling is far slower than the trivial unstructured search (g=ln 2), yet it remains exponential, confirming that classical simulated annealing does not achieve polynomial time on these instances.

For the quantum analysis, the authors adopt the standard AQC Hamiltonian H(λ) = (1‑λ)H_D + λH₀, where H_D = Σ_i σ_i^x is a transverse‑field driver and H₀ encodes the 3‑SAT cost function. They discretize imaginary time using a Trotter‑Suzuki decomposition (N_τ=128 slices, Δτ=1) and work at effectively zero temperature (β→∞). By scanning the adiabatic control parameter λ they locate a quantum phase transition at λ* where the ground‑state overlap observable exhibits a discontinuous jump. The key quantum metric is the maximal spin‑spin correlation length ξ_max = 1/ΔE_min, obtained from the exponential decay of a two‑point correlation function Γ(τ) = ⟨O(0)O(τ)⟩_c. ξ_max also grows exponentially with N, fitting ln ξ_max = A + g_q N with g_q≈0.061 for r=8. According to Landau‑Zener theory, the runtime of an adiabatic algorithm scales as O(ξ_max²) ≈ exp(2g_q N). The authors find that 2g_q≈0.122, which is smaller than the classical growth constant g_B but also below the Grover bound g=ln 2/2≈0.347. Consequently, standard linear‑schedule AQC does not outperform Grover’s optimal quantum search for this problem class; in fact, the quantum growth constant exceeds the classical one by a factor of about 3.8, indicating that pure quantum fluctuations are less effective than classical thermal fluctuations for finding the ground state.

A notable observation is the linear correlation between the classical nucleation barrier B₀ and the quantum correlation length ξ_max across many instances, suggesting that the static free‑energy landscape provides predictive information about the quantum gap. The study emphasizes that while both classical and quantum annealing exhibit exponential scaling, the constants differ, and the quantum advantage is limited to a modest constant factor rather than a change of complexity class.

The authors conclude that standard AQC with a simple transverse‑field driver cannot reduce the exponential complexity of USA 3‑SAT to polynomial time. They suggest that alternative driver Hamiltonians, non‑linear annealing schedules, or entirely different quantum algorithms may be required to achieve a genuine quantum speedup. Moreover, the methodology of focusing on median‑complexity instances offers a realistic benchmark for evaluating both classical and quantum heuristics on NP‑hard problems.