Measurement disturbance tradeoffs in three-qubit unsupervised quantum classification

We consider measurement disturbance tradeoffs in quantum machine learning protocols which seek to learn about quantum data. We study the simplest example of a binary classification task in the unsupervised regime. Specifically, we investigate how a classification of two qubits, that can each be in one of two unknown states, affects our ability to perform a subsequent classification on three qubits when a third is added. Surprisingly, we find a range of strategies in which a nontrivial first classification does not affect the success rate of the second classification. There is, however, a nontrivial measurement disturbance tradeoff between the success rate of the first and second classifications, and we fully characterize this tradeoff analytically.