Learning functions of Hamiltonians with Hamiltonian Fourier features

We propose a quantum machine learning task that is provably easy for quantum computers and arguably hard for classical ones. The task involves predicting quantities of the form $\mathrm{T}\mathrm{r}[f(H)\rho ]$, where $f$ is an unknown function, given descriptions of $H$ and $\rho$. Using a Fourier-based feature map of Hamiltonians and linear regression, we theoretically establish the learnability of the task and implement it on a superconducting device using up to 40 qubits. This work provides a machine learning task with practical relevance, provable quantum easiness, and near-term feasibility.