Vadym Kliuchnikov, Eddie Schoute (Mar 29 2024).
Abstract: Non-Clifford gates are frequently exclusively implemented on fault-tolerant architectures by first distilling magic states in specialised magic-state factories. In the rest of the architecture, the computational space, magic states can then be consumed by a stabilizer circuit to implement non-Clifford operations. We show that the connectivity between the computational space and magic state factories forms a fundamental bottleneck on the rate at which non-Clifford operations can be implemented. We show that the nullity of the magic state, $\nu(|D\rangle)$ for diagonal gate $D$, characterizes the non-local resources required to implement $D$ in the computational space. As part of our proof, we construct local stabilizer circuits that use only $\nu(|D\rangle)$ ebits to implement $D$ in the computational space that may be useful to reduce the non-local resources required to inject non-Clifford gates. Another consequence is that the edge-disjoint path compilation algorithm [arXiv:2110.11493] produces minimum-depth circuits for implementing single-qubit diagonal gates.