Quantum state ensemble ($\{p_i,|{\psi }_i⟩\},\rho ={\sum }_i{p}_i|{\psi }_i⟩⟨{\psi }_i|$)

A quantum state ensemble specifies a random preparation: normalized pure states $|{\psi }_i⟩$ are selected with probabilities $p_i\ge 0$, ${\sum }_i{p}_i=1$. Without an accessible preparation label, its measurement statistics are those of the density operator $\rho ={\sum }_i{p}_i|{\psi }_i⟩⟨{\psi }_i|$. Indeed every POVM effect $E$ has probability ${\sum }_i{p}_i⟨{\psi }_i|E|{\psi }_i⟩=\mathrm{Tr}(E\rho )$. Different ensembles can therefore have identical operational statistics: equal mixtures of computational-basis qubits or Hadamard-basis qubits both have density operator $I/2$. The ensemble decomposition contains information about the preparation, not additional information accessible from the system alone.

Three equally weighted pure qubit states with coplanar unit Bloch vectors separated by $120$ degrees. One example is $|t_j⟩=(|0⟩+{\omega }^j|1⟩)/\sqrt{2}$, with $\omega =e^{2\pi i/3}$ and $j=0,1,2$. The roots-of-unity identity $1+\omega +{\omega }^2=0$ cancels the average off-diagonal matrix entries, so $\frac{1}{3}{\sum }_j|t_j⟩⟨t_j|=I/2$.

This is an ensemble of preparation states, whereas a trine qubit POVM consists of measurement effects $E_j=\frac{2}{3}|t_j⟩⟨t_j|$. The POVM–ensemble duality for the maximally mixed state relates them but does not identify their operational roles.

For normalized pure states $|u⟩,|v⟩$, a two-member quantum state ensemble gives the maximally mixed state of a qubit exactly when its members are orthogonal and equally probable. To prove necessity, take expectation values of its density matrix in the two members: $1/2=p+(1-p)|⟨u|v⟩{|}^2\ge p$ and $1/2=1-p+p|⟨u|v⟩{|}^2\ge 1-p$. Hence $p=1/2$ and the overlap vanishes. Sufficiency is the resolution of the identity in an orthonormal basis.

A convex roof extension extends a function $\mu$ on pure states to arbitrary density operators by minimizing its ensemble average over all pure state decompositions. It agrees with $\mu$ on pure states: every positive-weight state in a decomposition of a rank-one density operator has the same one-dimensional support. Concatenating nearly minimizing ensembles proves that $F$ is a convex function. Moreover, every convex function that is no larger than $\mu$ on pure states is no larger than $F$, by applying convexity to each decomposition. Infima need not be attained; an arbitrarily small additive error suffices when $F$ is finite.

If a convex roof extension is non-increasing on average for every pure input to a quantum instrument, then it is non-increasing on average for mixed inputs too. For an input ensemble $\{p_j,{\pi }_j\}$, let $q_{k|j}$ and ${\sigma }_{jk}$ be its conditional outcome probabilities and states. Linearity of each unnormalized outcome map gives $P_k{\rho }_k={\sum }_j{p}_j{q}_{k|j}{\sigma }_{jk}$. Convexity gives ${\sum }_k{P}_{k}F({\rho }_k)\le {\sum }_{j,k}{p}_j{q}_{k|j}F({\sigma }_{jk})\le {\sum }_j{p}_j\mu ({\pi }_j)$. Taking the infimum over input ensembles proves the result, even when conditional outputs of pure inputs are mixed, provided the averages are well defined.