Page - 275 - in Differential Geometrical Theory of Statistics
Image of the Page - 275 -
Text of the Page - 275 -
Entropy2016,18, 407
3.GeneralizationofRƩnyiDivergence
In this section, weprovide a generalization of RƩnyi divergence, which is given in terms of
aĻ-function. Thisgeneralizationalsodependsonaparameterαā [ā1,1]; forα=±1, it isdeļ¬ned
as a limit. Supposing that theunderlying Ļ-function is continuouslydifferentiable,we show that
this limitexistsandresults in theĻ-divergence [12]. Inwhat follows,allprobabilitydistributionsare
assumedtohavepositivedensity. Inotherwords, theybelongto thecollection
Pμ= {
pāL0 : ā«
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).
TheRĆ©nyidivergenceoforderαā (ā1,1)betweentwoprobabilitydistributions pandq inPμ is
deļ¬nedas
D(α)R (pā q)= 4
α2ā1 log (ā«
T p 1āα
2 q 1+α
2 dμ )
. (4)
Forα=±1, theRĆ©nyidivergence isdeļ¬nedbytakinga limit:
D(ā1)R (pā q)= limαāā1D (α)
R (pā q), (5)
D(1)R (pā q)= limαā1D (α)
R (pā q). (6)
Under some conditions, the limits in (5) and (6) are ļ¬nite-valued, and converge to the
KullbackāLeiblerdivergence. Inotherwords,
D(ā1)R (pā q)=D(1)R (qā p)=DKL(pā q)<ā,
whereDKL(pā q)denotes theKullbackāLeiblerdivergencebetween pandq,which isgivenby
DKL(pā q)= ā«
T p log (p
q )
dμ.
TheseconditionsarestatedinProposition1,givenintheendof thissection, for thecase involving
thegeneralizedRƩnyidivergence.
TheRƩnyidivergence in its standardformisgivenby
D(α)(pā q)= 1
1āα log (ā«
T pαq1āαdμ )
, forαā (0,1). (7)
Expression(4) is relatedto this formby
D(α)R (pā q)= 2
1āαD ((1āα)/2)(pā q).
Beyondthechangeofvariables,whichresults inα rangingin [ā1,1], expressions (4)and(7)differ
bythe factor2/(1āα).Weoptedto insert the term2/(1āα) so that somekindofsymmetrycouldbe
maintainedwhenthe limitsαāā1andαā1areconsidered. Inaddition, thegeometry inducedbythe
version(4)conformswithAmariāsnotation[5].
TheRĆ©nyi divergenceD(α)R (Ā· ā Ā·) canbedeļ¬ned for every α āR. However, for α /ā (ā1,1),
theexpression(4)maynotbeļ¬nite-valuedforevery pandq inPμ. Toavoidsometechnicalities,we
justconsiderαā [ā1,1].
275
Differential Geometrical Theory of Statistics
- Title
- Differential Geometrical Theory of Statistics
- Authors
- FrƩdƩric Barbaresco
- Frank Nielsen
- Editor
- MDPI
- Location
- Basel
- Date
- 2017
- Language
- English
- License
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03842-425-3
- Size
- 17.0 x 24.4 cm
- Pages
- 476
- Keywords
- Entropy, Coding Theory, Maximum entropy, Information geometry, Computational Information Geometry, Hessian Geometry, Divergence Geometry, Information topology, Cohomology, Shape Space, Statistical physics, Thermodynamics
- Categories
- Naturwissenschaften Physik