We propose an adaptive, two step strategy, for the estimation of mixed qubit states. We show that the strategy is optimal in a local minimax sense for the trace norm distance as well as other locally quadratic figures of merit. Local minimax optimality means that given n identical qubits, there exists no estimator which can perform better than the proposed estimator on a neighborhood of size n ^−1/2 of an arbitrary state. In particular, it is asymptotically Bayesian optimal for a large class of prior distributions.

We present a physical implementation of the optimal estimation strategy based on continuous time measurements in a field that couples with the qubits. The crucial ingredient of the result is the concept of local asymptotic normality (or LAN) for qubits. This means that, for large n, the statistical model described by n identically prepared qubits is locally equivalent to a model with only a classical Gaussian distribution and a Gaussian state of a quantum harmonic oscillator. The term ‘local’ refers to a shrinking neighborhood around a fixed state ρ ₀. An essential result is that the neighborhood radius can be chosen arbitrarily close to n ^−1/4. This allows us to use a two step procedure by which we first localize the state within a smaller neighborhood of radius n ^−1/2+ϵ, and then use LAN to perform optimal estimation.