Quantum Hamiltonian Descent (QHD)¶
Implements Quantum Hamiltonian Descent (QHD), a quantum dynamical optimizer that quantizes the continuous-time limit of accelerated gradient descent.
QHD takes the classical Bregman-Lagrangian view of accelerated first-order methods, whose continuous-time limit is a damped second-order ODE, and replaces the classical trajectory \(X(t)\) with a quantum wave function \(\Psi(t)\) over the search space. The objective \(f(x)\) becomes a potential term and the momentum becomes the kinetic (Laplacian) term of a time-dependent Hamiltonian; the state then evolves under the Schrödinger equation. Because the evolution is quantum, the wave function tunnels through and delocalizes across barriers rather than following a single descent path, which lets QHD escape spurious local minima on nonconvex landscapes. A candidate minimizer is read out by measuring the position observable \(\hat{x}\) at the final time \(T\), the measured \(x\) being the returned parameters \(\theta\).
The classical second-order ODE that QHD quantizes, and the resulting quantum dynamics, are
where \(\Psi(t)\) is the wave function over the parameter space, \(\hat{H}(t)\) the time-dependent Hamiltonian, \(\Delta\) the Laplacian (the kinetic/momentum term), \(f(x)\) the objective acting as the potential, \(\theta = x\) the parameters, and \(e^{\phi_t}, e^{\chi_t}\) time-dependent damping coefficients inherited from the Bregman-Lagrangian schedule \((\alpha_t, \beta_t, \gamma_t)\). The ratio \(e^{\phi_t}/e^{\chi_t}\) is required to vanish for large \(t\) so that kinetic energy is gradually drained and the state concentrates near the global minimizer.
Reference: Jiaqi Leng, Ethan Hickman, Joseph Li, Xiaodi Wu, "Quantum Hamiltonian Descent", 2023. https://arxiv.org/abs/2303.01471