Proceedings of the 15th European Conference on Artificial Intelligence
- I Lynce, J Marques-Silva
- Proceedings of the 15th European Conference on…
We present an analysis of the quantum adiabatic algorithm for solving hard instances of 3-SAT ͑an NP-complete problem͒ in terms of random matrix theory ͑RMT͒. We determine the global regularity of the spectral fluctuations of the instantaneous Hamiltonians encountered during the interpolation between the starting Hamiltonians and the ones whose ground states encode the solutions to the computational problems of interest. At each interpolation point, we quantify the degree of regularity of the average spectral distribution via its Brody parameter, a measure that distinguishes regular ͑i.e., Poissonian͒ from chaotic ͑i.e., Wigner-type͒ distributions of normalized nearest-neighbor spacings. We find that for hard problem instances—i.e., those having a critical ratio of clauses to variables—the spectral fluctuations typically become irregular across a contiguous region of the interpolation parameter, while the spectrum is regular for easy instances. Within the hard region, RMT may be applied to obtain a mathematical model of the probability of avoided level crossings and concomitant failure rate of the adiabatic algorithm due to nonadiabatic Landau-Zener-type transitions. Our model predicts that if the interpolation is performed at a uniform rate, the average failure rate of the quantum adiabatic algorithm, when averaged over hard problem instances, scales exponentially with increasing problem size.