Reduction of Finite Sampling Error in Quantum Krylov Subspace Diagonalization

Within the realm of early fault-tolerant quantum computing (EFTQC), quantum Krylov subspace diagonalization (QKSD) has recently emerged as a promising quantum algorithm for the approximate Hamiltonian diagonalization through projection onto the quantum Krylov subspace. However, the application of this algorithm often entails solving an ill-conditioned generalized eigenvalue problem (GEVP) associated with an erroneous matrix pair, which can cause significant distortion to the solution. Because EFTQC assumes error correction albeit on a small scale, errors in the matrices are predominant due to finite sampling error. This work focuses on quantifying the sampling error within the measurement of matrix element of projected Hamiltonian by considering two measurement approaches based on the Hamiltonian decompositions: linear combination of unitaries and diagonalizable fragments. Furthermore, we propose two measurement strategies to minimize the sampling error with a given budget for quantum circuit repetitions: the shifting technique and coefficient splitting. The shifting technique removes redundant parts from the Hamiltonian that annihilate one of the bra or ket states. The coefficient splitting method optimizes the allocation of each common term that can be measured in different circuits. Numerical experiments with electronic structures of small molecules demonstrate the effectiveness of these strategies, showing a reduction in sampling costs by a factor of 20-500.