Search papers, labs, and topics across Lattice.
This paper introduces a temporal-access resource theory for continuous process tensors, enabling the quantification of how well quantum processes can protect logical tasks under time-resolved control by incorporating physical access interfaces and budgeted free conversions. Complete-record total-variation error serves as the performance measure, demonstrating monotonicity under conversions that maintain a common free baseline, which is $1-d^{-2}$ for unknown-state transmission. The theory reveals that two-qubit processes with identical endpoint channels can exhibit distinct optimal errors at the same finite pulse budget, highlighting the limitations of endpoint-only characterization and the importance of temporal structure and control investment.
Endpoint channels mask crucial temporal control information, leading to mischaracterization of quantum process capabilities that this new resource theory resolves.
The uncontrolled endpoint channel can conceal how well a quantum process can protect a logical task under time-resolved control. We formulate a fixed-task resource theory of continuous process tensors with physical access interfaces, entrance-erasing free objects, and budgeted free conversions. Complete-record total-variation error measures performance; the calibrated capability is monotone under conversions that preserve a common free baseline. For unknown-state transmission the baseline is exactly $1-d^{-2}$. Two qubit processes with the same endpoint channel have optimal errors $0$ and $p=\sin^2(2gh)$ at the same finite pulse budget in an ideal-transparent system-only window architecture, forcing any endpoint-only predictor to err by at least $p/2$ on one of them. An exactly solvable family shows how the allocation of a fixed interaction phase around a control window sets the attainable error, and a nontrivial perfect echo requires and attains total control action $π$ in that architecture. The separation extends directly to every qubit unitary target at a budget that includes its boundary gate, while unitary tasks in any dimension have baseline $1-d^{-2}$ whenever a budget-feasible free realization gives the depolarizing output. Complete-record implementation bounds quantify coherent and dissipative control imperfections. Continuous-process randomized compilation supplies physical instrument pullbacks; compatible completion conditions extend their action, and certified dynamical bounds yield achievable task-error and capability guarantees. The resulting theory identifies temporal structure and control investment that endpoint characterization alone cannot resolve.