We introduce an approach to improve the accuracy and reduce the sample complexity of near term quantum-classical algorithms. We construct a simpler initial parameterized quantum state, or ansatz, based on the past causal cone of each observable, generally yielding fewer qubits and gates. We implement this protocol on a trapped ion quantum computer and demonstrate improvement in accuracy and time-to-solution at an arbitrary point in the variational search space. We report a $\sim 27\%$ improvement in the accuracy of the calculation of the deuteron binding energy and $\sim 40\%$ improvement in the accuracy of the quantum approximate optimization of the MAXCUT problem applied to the dragon graph $T_{3,2}$. When the time-to-solution is prioritized over accuracy, the former requires $\sim 71\%$ fewer measurements and the latter requires $\sim 78\%$ fewer measurements.