1: \begin{abstract}

2: Fast estimation of the single-particle density matrix is key to many

3: applications in quantum chemistry and condensed matter physics. The

4: best numerical methods leverage the fact that the density matrix elements

5: $f(H)_{ij}$ decay rapidly with distance $r_{ij}$ between orbitals.

6: This decay is usually exponential. However, for the special case of

7: metals at zero temperature, algebraic decay of the density matrix

8: appears and poses a significant numerical challenge. We introduce

9: a gradient-based probing method to estimate all local density matrix

10: elements at a computational cost that scales linearly with system

11: size. For zero-temperature metals the stochastic error scales like

12: $S^{-(d+2)/2d}$, where $d$ is the dimension and $S$ is a prefactor

13: to the computational cost. The convergence becomes exponential if

14: the system is at finite temperature or is insulating.

15: \end{abstract}

16: