Seite - 275 - in Differential Geometrical Theory of Statistics
Bild der Seite - 275 -
Text der Seite - 275 -
Entropy2016,18, 407
3.GeneralizationofRényiDivergence
In this section, weprovide a generalization of Rényi divergence, which is given in terms of
aϕ-function. Thisgeneralizationalsodependsonaparameterα∈ [−1,1]; forα=±1, it isdefined
as a limit. Supposing that theunderlying ϕ-function is continuouslydifferentiable,we show that
this limitexistsandresults in theϕ-divergence [12]. Inwhat follows,allprobabilitydistributionsare
assumedtohavepositivedensity. Inotherwords, theybelongto thecollection
Pμ= {
p∈L0 : ∫
T pdμ=1and p>0 }
,
where L0 is the space of all real-valued, measurable functions on T, with equality μ-a.e.
(μ-almosteverywhere).
TheRényidivergenceoforderα∈ (−1,1)betweentwoprobabilitydistributions pandq inPμ is
definedas
D(α)R (p‖ q)= 4
α2−1 log (∫
T p 1−α
2 q 1+α
2 dμ )
. (4)
Forα=±1, theRényidivergence isdefinedbytakinga limit:
D(−1)R (p‖ q)= limα↓−1D (α)
R (p‖ q), (5)
D(1)R (p‖ q)= limα↑1D (α)
R (p‖ q). (6)
Under some conditions, the limits in (5) and (6) are finite-valued, and converge to the
Kullback–Leiblerdivergence. Inotherwords,
D(−1)R (p‖ q)=D(1)R (q‖ p)=DKL(p‖ q)<∞,
whereDKL(p‖ q)denotes theKullback–Leiblerdivergencebetween pandq,which isgivenby
DKL(p‖ q)= ∫
T p log (p
q )
dμ.
TheseconditionsarestatedinProposition1,givenintheendof thissection, for thecase involving
thegeneralizedRényidivergence.
TheRényidivergence in its standardformisgivenby
D(α)(p‖ q)= 1
1−α log (∫
T pαq1−αdμ )
, forα∈ (0,1). (7)
Expression(4) is relatedto this formby
D(α)R (p‖ q)= 2
1−αD ((1−α)/2)(p‖ q).
Beyondthechangeofvariables,whichresults inα rangingin [−1,1], expressions (4)and(7)differ
bythe factor2/(1−α).Weoptedto insert the term2/(1−α) so that somekindofsymmetrycouldbe
maintainedwhenthe limitsα↓−1andα↑1areconsidered. Inaddition, thegeometry inducedbythe
version(4)conformswithAmari’snotation[5].
TheRényi divergenceD(α)R (· ‖ ·) canbedefined for every α ∈R. However, for α /∈ (−1,1),
theexpression(4)maynotbefinite-valuedforevery pandq inPμ. Toavoidsometechnicalities,we
justconsiderα∈ [−1,1].
275
Differential Geometrical Theory of Statistics
- Titel
- Differential Geometrical Theory of Statistics
- Autoren
- Frédéric Barbaresco
- Frank Nielsen
- Herausgeber
- MDPI
- Ort
- Basel
- Datum
- 2017
- Sprache
- englisch
- Lizenz
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03842-425-3
- Abmessungen
- 17.0 x 24.4 cm
- Seiten
- 476
- Schlagwörter
- Entropy, Coding Theory, Maximum entropy, Information geometry, Computational Information Geometry, Hessian Geometry, Divergence Geometry, Information topology, Cohomology, Shape Space, Statistical physics, Thermodynamics
- Kategorien
- Naturwissenschaften Physik