<p>Reversing unknown quantum dynamics is vital for quantum control and learning, yet full unitary inversion requires a resource-intensive <i>O</i>(<i>d</i>²) queries. Because many applications only require the reversed evolution for a given observable, a critical gap remains in understanding the minimal resources for such targeted reversal. Here, we address this by introducing shadow unitary inversion. We establish a lower bound showing the query complexity must scale at least linearly with system dimension for spectrally biased observables, with the constant determined by the observable’s spectral properties. For qubit case, we construct an explicit, deterministic three-query sequential protocol achieving exact shadow inversion and completely characterize all admissible channels, with numerical evidence suggesting optimality. For higher dimensions, we develop a semidefinite-programming formulation and introduce a representation-theoretic symmetry reduction that decomposes the optimization into invariant blocks, substantially reducing the problem size. Shadow unitary inversion thus offers a resource-efficient path to inverse-dynamics estimation for future quantum control, diagnostics, and learning tasks.</p>