Quantum relative entropy

Noncommutative generalization of Kullback-Leibler divergence for density operators.
Quantum relative entropy

Let ρ\rho and σ\sigma be s on a finite-dimensional Hilbert space HH. The quantum relative entropy (also called Umegaki relative entropy) is defined by

D(ρσ):={Tr ⁣(ρ(logρlogσ)),if supp(ρ)supp(σ),+,otherwise. D(\rho\|\sigma) := \begin{cases} \operatorname{Tr}\!\big(\rho(\log\rho-\log\sigma)\big), & \text{if }\operatorname{supp}(\rho)\subseteq \operatorname{supp}(\sigma),\\[4pt] +\infty, & \text{otherwise.} \end{cases}

The support condition is needed because logσ\log\sigma is not finite on the kernel of σ\sigma. The operators logρ\log\rho and logσ\log\sigma are defined via spectral calculus (see ).

Classical (commuting) case

If ρ\rho and σ\sigma commute, they are simultaneously diagonalizable and the formula reduces to the classical relative entropy (Kullback-Leibler divergence) of their eigenvalue distributions; compare .

Key properties

  • Nonnegativity (Klein inequality): D(ρσ)0D(\rho\|\sigma)\ge 0, with equality iff ρ=σ\rho=\sigma.
  • Not symmetric: typically D(ρσ)D(σρ)D(\rho\|\sigma)\ne D(\sigma\|\rho); it is not a metric.
  • Data processing inequality: for any completely positive trace-preserving map Φ\Phi, D(ρσ)  D(Φ(ρ)Φ(σ)). D(\rho\|\sigma)\ \ge\ D(\Phi(\rho)\|\Phi(\sigma)). In particular, taking Φ=TrB\Phi=\operatorname{Tr}_B (see ) gives D(ρABσAB)  D(ρAσA),ρA=TrB(ρAB), σA=TrB(σAB). D(\rho_{AB}\|\sigma_{AB})\ \ge\ D(\rho_A\|\sigma_A), \quad \rho_A=\operatorname{Tr}_B(\rho_{AB}),\ \sigma_A=\operatorname{Tr}_B(\sigma_{AB}).

Relation to von Neumann entropy

If d=dimHd=\dim H and τ=I/d\tau=I/d is the maximally mixed state, then

D(ρτ)=logdS(ρ), D(\rho\|\tau)=\log d - S(\rho),

where S(ρ)S(\rho) is the .