We propose a quantum inverse algorithm (QInverse) to
directly determine
general eigenstates by repeatedly applying the inverse power of a
shifted Hamiltonian to an arbitrary initial state. To properly deal
with the strongly entangled inverse power states and the resultant
excited states, we solved the underlying linear equation, both variationally
and adaptively, to obtain a faithful inverse power state with a shallow
quantum circuit. QInverse is singularity-free and successfully obtains
the target excited states with an energy closest to the shift ω,
which is difficult to reach using variational methods. We also propose
a subspace expansion approach to accelerate convergence and show that
it is helpful to determine the two nearest eigenvalues when they are
equally close to ω. These approaches were compared with the
folded-spectrum method, which aims to generate excited states through
variational optimization. It is shown that, whereas the folded-spectrum
approach often fails to predict the target state by falling into a
local minimum owing to its variational features, the success rate
and accuracy of our algorithms are systematically improvable.