Hellinger distance on a finite type #
A measure on a finite type gives a vector of Euclidean space, the square roots of the masses of its
atoms. The Hellinger distance between two measures is the Euclidean distance between their vectors,
divided by √2. The inner product of the vectors is the Bhattacharyya coefficient
∑ a, √(μ {a} ν {a}), so the squared distance of two finite measures is the mean of their total
masses less that coefficient, and probability measures lie within distance one of each other, where
the Kullback–Leibler divergence may be infinite.
Main definitions #
MeasureTheory.Measure.sqrtMass μ: the vectora ↦ √(μ {a})ofEuclideanSpace ℝ α.InformationTheory.hellingerDist μ ν: the distance betweenμ.sqrtMassandν.sqrtMassdivided by√2.
Main statements #
hellingerDist_sq: the squared distance is the mean total mass less the Bhattacharyya coefficient.hellingerDist_le_one: probability measures lie within distance one.hellingerDist_dirac_left: the distance from a Dirac mass.
Implementation notes #
As a distance on Euclidean space the Hellinger distance is symmetric and satisfies the triangle
inequality, the L² argument by which it is a metric. The factor 1/√2 bounds the distance
between probability measures by one; Polyanskiy and Wu omit it. Mathlib has no Hellinger
distance; in general it is an f-divergence of Radon–Nikodym derivatives, of which this is the
case of a finite type.
References #
The square roots of the masses of a measure's atoms, as a vector of Euclidean space.
Equations
- μ.sqrtMass = WithLp.toLp 2 fun (a : α) => √(μ.real {a})
Instances For
The Hellinger distance √(∑ a, (√(μ {a}) - √(ν {a})) ^ 2) / √2, the Euclidean distance between
square-root mass vectors divided by √2.
Equations
- InformationTheory.hellingerDist μ ν = dist μ.sqrtMass ν.sqrtMass / √2
Instances For
The squared Hellinger distance of two finite measures is the mean of their total masses less the Bhattacharyya coefficient.
Probability measures, and the zero measure, lie within Hellinger distance one of each other.
The Hellinger distance from a Dirac mass to a probability measure is fixed by the measure's mass at the atom, and is one where that mass vanishes.