Seite - 371 - in Differential Geometrical Theory of Statistics
Bild der Seite - 371 -
Text der Seite - 371 -
Entropy2016,18, 98
Proposition3states theconvergenceof themaximumlikelihoodestimate YˆN to the truevalueof
theparameter Y¯. It isbasedonLemma1,whichstates that theparameter Y¯ is theRiemannianmedian
of thedistributionL(Y¯,σ) in thesenseofdefinitionEquation(12).
Proposition2 (MLEandmedian). Themaximumlikelihoodestimateof theparameter Y¯ is theRiemannian
medianYˆN ofY1, . . . ,YN.Moreover, themaximumlikelihoodestimateof theparameterσ is the solution σˆN of:
σ2× d
dσ logζm(σ)= 1
N N
∑
n=1 d(Y¯,Yn) (20)
Both YˆN and σˆN exist andareunique foranyrealizationof the samplesY1, . . . ,YN.
ProofofProposition2. The log-likelihoodfunction,of theparameters Y¯andσ, canbewrittenas:
N
∑
n=1 log p(Yn|Y¯,σ) = N
∑
n=1 log (
1
ζm(σ) e− d(Y¯,Yn)
σ )
= −N logζm(σ)− 1
σ N
∑
n=1 d(Y¯,Yn)
Asthefirst terminthe lastexpressiondoesnotcontain Y¯,
argmaxY¯ N
∑
n=1 log p(Yn|Y¯,σ)=argminY¯ N
∑
n=1 d(Y¯,Yn)
Thequantityontheright isexactly YˆN byEquation(11). Thisprovesthefirstclaim.Now,consider
the function:
F(η)=−N log(ζm(−1
η ))+η N
∑
n=1 d(YˆN,Yn), η< −1
σm
This function isstrictlyconcave, since it is the logarithmof themomentgeneratingfunctionofa
positivemeasure.Note that limη→−1σm F(η)=−∞, andadmit foramoment that limη→−∞F(η)=−∞.
By the strict concavityofF, there exists aunique ηˆN< −1σm (which is themaximumofF), such that
F′(ηˆN)= 0. It follows that σˆN = −1ηˆN lies in (0,σm)andsatisfiesEquation (20). Theuniquenessof σˆN
isaconsequenceof theuniquenessof ηˆN. Thus, theproof iscomplete.Now, it remains tocheckthat
limη→−∞F(η)=−∞or just limσ→+∞ 1σ log(ζm(1σ))=0.Clearly:
∏
i<j sinh (|ri−rj|
2 )
≤AmeBm|r|
whereAm andBm are twoconstantsonlydependingonm. Usingthis, it followsthat:
1
σ log(ζm( 1
σ ))≤ 1
σ log(cmAm)+ 1
σ log (∫
Rm exp((−σ+Bm)|r|)dr1 · · ·drm )
(21)
However, forsomeconstantCmonlydependingonm,
∫
Rm exp((−σ+Bm)|r|)dr1 · · ·drm=Cm ∫ ∞
0 exp((−σ+Bm)u)um−1du
≤ (m−1)!Cm ∫ ∞
0 exp((−σ+Bm+1)u)du= (m−1)!Cm
σ−Bm−1
CombiningthisboundandEquation(21)yields limσ→+∞ 1σ log(ζm( 1
σ))=0.
371
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