Quantum Hellinger Distance

• Quantum Hellinger distance is a square-root-based metric defined for density matrices, capturing differences via the trace of square roots (quantum affinity).
• It induces a Riemannian structure on quantum state space with key features like contractivity under quantum operations and joint convexity.
• Generalized forms link with operator means and f-divergences, underpinning applications in quantifying quantum coherence, discord, and nonlocal correlations.

Searching arXiv for recent and foundational papers on quantum Hellinger distance to ground the article in the cited literature. arXiv search query: "quantum Hellinger distance" Quantum Hellinger distance is a square-root-based distance on quantum states, typically defined for density matrices $\rho$ and $\sigma$ by

$$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$

with the overlap term $\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$ often called the quantum affinity (Spehner et al., 2016). In the commuting case it reduces to the classical Hellinger distance between probability distributions, while in the noncommutative case it defines a contractive, Riemannian metric that has become a basis for geometric quantifiers of coherence, discord, non-classical correlation, and measurement-induced nonlocality (Spehner et al., 2016, Jin et al., 2018). A parallel operator-theoretic line of work treats the usual geometric-mean-based expression as one member of a broader family of generalized quantum Hellinger divergences built from Kubo–Ando means, thereby linking the subject to maximal quantum $f$-divergences, Bregman geometry, and barycenter problems (Pitrik et al., 2019).

1. Definitions, affinity, and normalization conventions

For density matrices on a finite-dimensional Hilbert space, the survey literature defines the quantum Hellinger distance by

$$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$

and equivalently as the Hilbert–Schmidt norm of the difference of square roots (Spehner et al., 2016). Closely related papers instead work directly with the squared form

$$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$

or with a rescaled version such as

$$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$

These formulas differ by constants and square roots, so the literature is not uniform in normalization (Jin et al., 2018, Wissmann et al., 2013).

The common structural quantity is the affinity

$$A(\rho,\sigma)=\mathrm{Tr}\big(\sqrt{\rho}\sqrt{\sigma}\big),$$

so that $d_H^2$ is expressed as $\sigma$0 in one widespread convention (Kumar et al., 2024). One source states the bound

$\sigma$1

while another notes that some authors include an extra factor $\sigma$2, rescaling the range to $\sigma$3 (Dajka et al., 2011, Kumar et al., 2024). The same square-root structure underlies the classical reduction: for commuting states $\sigma$4 and $\sigma$5, the quantum formula collapses to

$\sigma$6

the ordinary Hellinger distance between probability vectors (Spehner et al., 2016).

This dependence on $\sigma$7 rather than $\sigma$8 is the defining technical feature of the subject. It gives a noncommutative overlap simpler than Uhlmann fidelity, yet still close enough to fidelity-based geometry to support contractivity and several resource-theoretic constructions (Marian et al., 2014, Spehner et al., 2016).

2. Metric structure, convexity, and information geometry

The quantum Hellinger distance is presented as a genuine metric: it is nonnegative, symmetric, vanishes exactly when $\sigma$9, and satisfies the triangle inequality (Spehner et al., 2016). It is also contractive under quantum operations,

$$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$0

and its square is jointly convex (Spehner et al., 2016). These properties place it among the standard contractive distances used in quantum information geometry.

The same survey identifies it as a Riemannian distance and gives the induced metric tensor

$$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$1

for $$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$2 and traceless Hermitian tangent vector $$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$3 (Spehner et al., 2016). In this sense the distance is not merely algebraic; it induces a differential geometry on state space. The same source contrasts it with the Bures metric and records the inequality

$$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$4

together with the fact that Bures coincides with the Fubini–Study distance on pure states whereas Hellinger does not (Spehner et al., 2016).

A central operational interpretation comes through quantum uncertainty. For unitary dynamics generated by $$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$5, the infinitesimal Hellinger speed is linked to the Wigner–Yanase skew information,

$$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$6

which underlies the connection between Hellinger geometry and Local Quantum Uncertainty (Spehner et al., 2016). This distinguishes it from the Bures metric, whose infinitesimal form is tied instead to quantum Fisher information (Spehner et al., 2016).

The square-root overlap also interacts naturally with relative-entropy-type quantities. The survey records

$$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$7

placing the metric within the $$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$8-$$d_{\mathrm{Hel}}(\rho,\sigma)=\|\sqrt{\rho}-\sqrt{\sigma}\|_2 =\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2},$$9 relative entropy framework (Spehner et al., 2016). This suggests that quantum Hellinger distance occupies an intermediate position between fidelity geometry and divergence geometry, a theme made explicit in the generalized operator-mean formulation.

3. Operator means, generalized divergences, and barycenters

A major extension replaces the geometric mean in the standard formula by an arbitrary Kubo–Ando mean. For positive definite operators $\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$0, the generalized quantum Hellinger divergence is defined by

$\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$1

or equivalently

$\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$2

where $\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$3 is the weight of the Kubo–Ando mean $\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$4 (Pitrik et al., 2019). The usual Hellinger-type quantity studied by Bhatia, Gaubert, and Jain is recovered by taking $\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$5, $\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$6, and the arcsine measure, giving

$\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$7

as the special geometric-mean case (Pitrik et al., 2019).

The same paper shows that these generalized divergences are maximal quantum $\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$8-divergences. Writing

$\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}$9

one has

$f$0

and because $f$1 is operator convex, the divergence is jointly convex and satisfies the data processing inequality for completely positive trace-preserving maps (Pitrik et al., 2019). It is also related to operator-valued Bregman divergences, which yields nonnegativity, definiteness, vanishing first derivative at the diagonal, and positive second derivative at the diagonal (Pitrik et al., 2019).

This framework leads naturally to barycenters. For positive definite operators $f$2 with weights $f$3, the barycenter is the unique minimizer of

$f$4

Its characterization is the matrix equation

$f$5

which has a unique positive definite solution (Pitrik et al., 2019).

An important correction follows. The claim that the barycenter for the geometric-mean-based Hellinger divergence is the weighted multivariate $f$6-power mean is true in the commuting case but false in general (Pitrik et al., 2019). A numerical $f$7 counterexample shows that the minimizer of the Hellinger divergence need not coincide with the weighted $f$8-power mean when the matrices do not commute (Pitrik et al., 2019). This is a precise noncommutative obstruction rather than a matter of notation.

4. Quantum coherence, discord, and measurement-induced nonlocality

Quantum Hellinger distance has been used extensively to define distance-to-free-set resource measures. For coherence in a fixed computational basis, one paper defines

$f$9

and derives the closed form

$$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$0

with the optimal incoherent state

$$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$1

(Jin et al., 2018). The same paper proves faithfulness, convexity, and strong monotonicity under incoherent selective operations, so $$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$2 satisfies the full set of standard coherence axioms (Jin et al., 2018).

For non-classical correlation, the same work defines

$$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$3

proves local-unitary invariance and contractivity under CPTP maps on the unmeasured subsystem, and shows

$$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$4

so the measure vanishes exactly on classical-quantum states (Jin et al., 2018). For qubit–qudit states an analytic formula is obtained,

$$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$5

with $$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$6 the largest eigenvalue of a matrix built from $$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$7 and Pauli operators (Jin et al., 2018).

A broader geometric program uses Hellinger distance to define geometric discord, measurement-induced geometric discord, and discord of response (Roga et al., 2015). In that framework,

$$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$8

and for pure states with Schmidt coefficients $$d_{\mathrm{Hel}}(\rho,\sigma)=\left(2-2\,\mathrm{tr}\sqrt{\rho}\sqrt{\sigma}\right)^{1/2}$$9,

$$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$0

(Spehner et al., 2016). For qubit–qudit systems, both the Hellinger geometric discord and the Hellinger discord of response are fully computable, and the latter is connected to Local Quantum Uncertainty (Roga et al., 2015, Spehner et al., 2016).

A related line quantifies measurement-induced nonlocality by

$$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$1

where the optimization is over locally invariant von Neumann measurements on subsystem $$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$2 (S et al., 2020). This Hellinger-distance MIN is presented as a bona fide measure of nonlocal correlation, is resistant to the local ancilla problem, and coincides with skew-MIN in the formulation used there (S et al., 2020). For a bipartite pure state with Schmidt coefficients $$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$3,

$$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$4

and closed formulas are also obtained for general $$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$5 mixed states (S et al., 2020).

The multipartite extension appears in a Hellinger-based generalized geometric discord defined by

$$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$6

with exact evaluation for bipartite pure states via Schmidt decomposition and direct extensions to symmetric multipartite settings (Cui et al., 2014). For permutation-invariant and translation-invariant states, the nearest classical state is proposed to inherit the same symmetry (Cui et al., 2014).

5. Gaussian states and open-system distinguishability

In continuous-variable Gaussian settings, Hellinger distance has been used to define a geometric Gaussian discord by minimizing over product Gaussian states (Marian et al., 2014). A central caveat is built into that construction: because the zero-discord Gaussian states are precisely the product Gaussian states, the resulting Hellinger quantity measures all intermode correlations, not only the genuinely quantum part, and is therefore an upper bound for the geometric discord in the unrestricted sense (Marian et al., 2014). The corresponding affinity for $$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$7-mode Gaussian states is written explicitly in terms of the covariance matrices of the square-root states, and in the two-mode case the closest Gaussian product state can be determined exactly (Marian et al., 2014).

For symmetric two-mode squeezed thermal states,

$$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$8

while for mode-mixed thermal states analogous closed forms are obtained (Marian et al., 2014). The same work argues that, in these Gaussian families, the Hellinger-based quantity behaves more like a geometric analogue of quantum mutual information than a strict measure of quantum discord (Marian et al., 2014). This is a recurrent source of confusion in the Gaussian literature, and the distinction is explicit in the source.

In open-system dynamics, the quantum Hellinger distance has also been compared with trace distance, Bures distance, and Jensen–Shannon divergence as a witness of initial system–environment correlations (Dajka et al., 2011, Wissmann et al., 2013). The main conclusion is metric dependence. In an infinite-environment dephasing model, the Hellinger distance can decrease at early times, reach a minimum, and then rise again, but it remains below its initial value; unlike the trace distance, it does not show the same correlation-induced growth above the initial distinguishability (Dajka et al., 2011). In the bosonic dephasing model of a later comparative study,

$$D_H(\rho,\sigma)=\mathrm{Tr}\,(\sqrt{\rho}-\sqrt{\sigma})^2,$$9

so the Hellinger distance does not witness initial correlations there (Wissmann et al., 2013). In a spin-bath model it can increase, but much less often than the trace distance (Wissmann et al., 2013).

These results rule out a common overgeneralization: contractivity and geometric regularity do not imply that a distance is an effective dynamical witness of initial correlations in reduced open-system dynamics (Dajka et al., 2011, Wissmann et al., 2013).

6. Random states and matrix-analytic extensions

The quantum Hellinger distance has also been studied statistically for random density matrices. For two states $$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$0, one paper defines

$$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$1

with $$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$2, and derives exact mean and variance of the squared distance $$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$3 when one or both states are random (Kumar et al., 2024). The random ensembles considered are the Hilbert–Schmidt and Bures–Hall ensembles, and the analysis reduces the first two cumulants of $$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$4 to the first two moments of the affinity (Kumar et al., 2024). Matching those cumulants yields a gamma-distribution approximation for the law of $$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$5, which is reported to agree well with Monte Carlo simulations (Kumar et al., 2024). The same work states that, for matched parameters, the average distance is smallest for two Hilbert–Schmidt states, largest for two Bures–Hall states, and intermediate for mixed Hilbert–Schmidt/Bures–Hall pairs (Kumar et al., 2024).

A related matrix-analysis literature studies “matrix versions of the Hellinger distance” on the cone of positive definite matrices (Bhatia et al., 2019). The literal square-root analogue is

$$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$6

while the Bures–Wasserstein metric is

$$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$7

(Bhatia et al., 2019). Two further constructions replace the overlap term by the Pusz–Woronowicz geometric mean $$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$8 and the log-Euclidean mean $$D_H(\rho_1,\rho_2)=\sqrt{1-\mathrm{Tr}\big(\sqrt{\rho_2}\sqrt{\rho_1}\big)}.$$9, giving quantities whose squares are divergences rather than metrics (Bhatia et al., 2019). The squared forms

$$A(\rho,\sigma)=\mathrm{Tr}\big(\sqrt{\rho}\sqrt{\sigma}\big),$$0

are shown to be jointly convex and strictly convex in each variable separately, and their barycenters are characterized by nonlinear matrix equations (Bhatia et al., 2019).

This broader matrix setting does not identify a single noncommutative Hellinger geometry. Instead, it exhibits a family of square-root- and mean-based noncommutative distances and divergences, with the standard quantum Hellinger distance corresponding to the Hilbert–Schmidt norm of square-root differences and the generalized Kubo–Ando framework giving a systematic extension beyond that basic case (Bhatia et al., 2019, Pitrik et al., 2019).