Seite - 292 - in Differential Geometrical Theory of Statistics
Bild der Seite - 292 -
Text der Seite - 292 -
Entropy2016,18, 442
We call these bounds CELB and CEUB for Combinatorial Envelope Lower and Upper
Bounds,respectively.
2.1. TighterAdaptiveBounds
We shall now consider shape-dependent bounds improving over the additive logk+ logk′
non-adaptive bounds. This ismadepossible by adecomposition of the lse function explained as
follows. Let ti(x1,. . . ,xk)= log (
∑kj=1e xj−xi )
. By translation identityof the lse function,
lse(x1,. . . ,xk)= xi+ ti(x1,. . . ,xk) (13)
for all i∈ [k]. Since exj−xi = 1 if j= i, and exj−xi > 0,wehavenecessarily ti(x1,. . . ,xk)> 0 for any
i ∈ [k]. SinceEquation (13) is an identity for all i ∈ [k], weminimize the residual ti(x1,. . . ,xk)by
maximizing xi. Denotingby x(1), . . . ,x(k) the sequenceofnumbers sorted innon-decreasingorder,
thedecomposition
lse(x1,. . . ,xk)= x(k)+ t(k)(x1,. . . ,xk) (14)
yields thesmallest residual. Sincex(j)−x(k)≤0forall j∈ [k],wehave
t(k) (x1,. . . ,xk)= log (
1+ k−1
∑
j=1 ex(j)−x(k) )
≤ logk.
Thisshowstheboundsintroducedearliercanindeedbeimprovedbyamoreaccuratecomputation
of theresidual term t(k) (x1,. . . ,xk).
When considering 1D GMMs, let us now bound t(k)(x1,. . . ,xk) in a combinatorial range
Ir=(ar,ar+1). Letδ= δ(r) denote the index of the dominating weighted component in this
range.Then,
∀x∈ Ir,∀i, exp (
−logσi− (x−μi)
2
2σ2i + logwi )
≤exp (
−logσδ− (x−μδ)
2
2σ2δ + logwδ )
.
Thuswehave:
logm(x)= log wδ
σδ √
2π − (x−μδ) 2
2σ2δ + log (
1+∑
i =δ exp (
−(x−μi) 2
2σ2i + log wi
σi + (x−μδ)2
2σ2δ − logwδ
σδ ))
.
Nowconsider theratio term:
ρi,δ(x)= exp (
−(x−μi) 2
2σ2i + log wiσδ
wδσi + (x−μδ)2
2σ2δ )
.
It ismaximizedin Ir=(ar,ar+1)bymaximizingequivalently the followingquadraticequation:
li,δ(x)=−(x−μi) 2
2σ2i + log wiσδ
wδσi + (x−μδ)2
2σ2δ .
Setting thederivative tozero (l′i,δ(x)=0),weget theroot (whenσi =σδ)
xi,δ= (
μδ
σ2δ − μi
σ2i )/( 1
σ2δ − 1
σ2i )
.
Ifxi,δ∈ Ir, theratioρi,δ(x)canbeboundedin theslab Irbyconsidering theextremevaluesof the
threeelementset{ρi,δ(ar),ρi,δ(xi,δ),ρi,δ(ar+1)}.Otherwiseρi,δ(x) ismonotonic in Ir, itsbounds in Ir
292
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