Generalized Measurement

Coupling a system to an ancilla and measuring the ancilla realizes measurements richer than any projective observable.

Beyond projective measurement

The textbook projective measurement is a special case. A generalized measurement is specified by measurement operators {M_k} with sum_k M_k-dagger M_k = I. Outcome k has probability p(k) = Tr(M_k-dagger M_k rho) and leaves the system in M_k rho M_k-dagger / p(k). The positive operators E_k = M_k-dagger M_k form the POVM that governs probabilities; the M_k additionally fix the post-measurement state.

Realized through an ancilla

Every generalized measurement can be built from standard parts: attach an ancilla in a fixed state, apply a joint unitary that correlates ancilla with system, then perform an ordinary projective measurement on the ancilla. This is the content of Naimark's dilation. It means no new physical postulate is needed; generalized measurements are projective measurements on a larger system, seen from the part we keep.

import numpy as np
# weak-ish measurement operators
eps=0.3
M0=np.array([[np.cos(eps),0],[0,np.sin(eps)]])
M1=np.array([[np.sin(eps),0],[0,np.cos(eps)]])
print(np.round(M0.conj().T@M0 + M1.conj().T@M1,3))  # I

Why it is more powerful

Generalized measurements can extract partial information with tunable disturbance, discriminate non-orthogonal states optimally, and implement measurements with more outcomes than the system's dimension. By varying the coupling strength one interpolates from a gentle probe that barely disturbs the state to a full projective collapse. This flexibility is the foundation of weak measurement and of feedback control.

Role in computing

In quantum error correction, syndrome extraction is a generalized measurement engineered to reveal error information while leaving the encoded data untouched. Ancilla-based readout, parity checks, and non-demolition measurements are all generalized measurements. The framework also underlies teleportation and measurement-based computation, where carefully chosen measurements drive the entire computation rather than unitary gates.