Measurement and Classical Data

[mpharrigan]

We need a measurement bloq. Measure-and-throw-out-the-result can be simulated by CNOTing to a measure qubit and then tracing it out (xref quantum channels roadmap item).

When we measure and keep the result, we want the result to be available in the compute graph for further processing or conditioning other gates on it. Measuring and keeping the result can be codified by CNOTing to a fresh register. This is a new type of register: a classical data register. From a simulation and tensor contraction standpoint it behaves identically to a new qubit. The restriction comes from the fact that we can only apply classical-logic bloqs to it. I.e.: all gates must be hadamard-free. Classical data can be copied and deleted but that’s just CNOT and tracing out, respectively.

The way measurements are used in Cirq is with a stochastic simulator and statistics are gathered by running the simulation many times. This must correspond to cutting or projecting your tensor network stochastically. Before we try to support this NISQy view of measurement, I would like to have a very clear idea of why we’re using it. Presumably we use these simulation techniques to write unit tests. The full everettian simulation of classical feedback could get too expensive and projecting the classical wires to specific values could let us “fuzz test” non-exhaustively. But if we’re projecting those things out for performance reasons; why wouldn’t we also project out qubits as well?

The "classical" data type would likely be implemented using data types: #446