Adiabatic quantum optimization fails for random instances of NP-complete problems

Adiabatic quantum optimization has attracted a lot of attention because small scale simulations gave hope that it would allow to solve NP-complete problems efficiently. Later, negative results proved the existence of specifically designed hard instances where adiabatic optimization requires exponential time. In spite of this, there was still hope that this would not happen for random instances of NP-complete problems. This is an important issue since random instances are a good model for hard instances that can not be solved by current classical solvers, for which an efficient quantum algorithm would therefore be desirable. Here, we will show that because of a phenomenon similar to Anderson localization, an exponentially small eigenvalue gap appears in the spectrum of the adiabatic Hamiltonian for large random instances, very close to the end of the algorithm. This implies that unfortunately, adiabatic quantum optimization also fails for these instances by getting stuck in a local minimum, unless the computation is exponentially long.

💡 Research Summary

The paper investigates the performance of adiabatic quantum optimization (AQO) on random instances of NP‑complete problems, focusing on whether the algorithm can avoid the exponential slowdown that has been demonstrated for specially constructed hard instances. After reviewing earlier optimism—stemming from small‑scale simulations and the promise of quantum annealers—the authors set out to test AQO on ensembles of random 3‑SAT formulas, which serve as a canonical model of hard combinatorial optimization problems.

The authors map each Boolean variable to a spin‑½ particle and encode the clause‑satisfaction condition into a problem Hamiltonian (H_P). The driver Hamiltonian (H_0) is chosen as a transverse‑field term, the standard choice in quantum annealing. The time‑dependent Hamiltonian is then (H(s) = (1-s)H_0 + s H_P) with the adiabatic parameter (s) increasing from 0 to 1. Using a combination of perturbative analysis near the end of the schedule (i.e., (s\approx 1)) and extensive numerical diagonalization for problem sizes up to several dozen qubits, the authors examine the spectral gap (\Delta(s)) that controls the runtime via the adiabatic theorem.

The central finding is that, for typical random instances, the minimum gap does not occur at a modest value of (s) but rather collapses exponentially close to the final point of the evolution. Specifically, the gap scales as (\Delta_{\min} \sim \exp(-\alpha N)) with a constant (\alpha>0) that depends weakly on clause density. This exponential closing is traced to a phenomenon analogous to Anderson localization: the random clause structure creates a rugged energy landscape with many deep local minima, and the quantum wavefunction becomes localized in one of these minima when the transverse field is weak. Consequently, the system fails to tunnel efficiently to the global minimum, and the adiabatic runtime required to maintain a constant success probability becomes exponential in the number of variables.

Importantly, the authors demonstrate that this behavior is robust against variations in the driver Hamiltonian and the annealing schedule. Even when the transverse‑field strength is increased or when non‑linear schedules are employed, the gap still shrinks exponentially near the end of the evolution. This indicates that the limitation is not an artifact of a particular choice of (H_0) but a fundamental property of the random problem Hamiltonians themselves.

The paper situates these results within the broader context of quantum optimization research. While earlier works had shown that carefully engineered instances can force AQO to require exponential time, the present study shows that such worst‑case behavior is typical for random NP‑complete instances as well. The analogy to Anderson localization provides a physical intuition: just as disorder in a solid can prevent electronic transport by localizing wavefunctions, disorder in the clause structure prevents quantum transport across the solution space.

In conclusion, the authors argue that AQO, at least in its conventional formulation with a transverse‑field driver and a linear or smoothly varying schedule, cannot provide a generic polynomial‑time speedup for random NP‑complete problems. The algorithm becomes trapped in local minima unless the total evolution time grows exponentially with problem size. This result tempers expectations about the power of near‑term quantum annealers for solving real‑world combinatorial optimization tasks and suggests that future progress will require either fundamentally different Hamiltonian constructions, alternative quantum algorithms (such as variational quantum circuits), or problem‑specific heuristics that mitigate localization effects.